1
0
Fork 0
alistair23-linux/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c

7699 lines
212 KiB
C
Raw Normal View History

/*
* Copyright 2015 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*
* Authors: AMD
*
*/
/* The caprices of the preprocessor require that this be declared right here */
#define CREATE_TRACE_POINTS
#include "dm_services_types.h"
#include "dc.h"
#include "dc/inc/core_types.h"
#include "dal_asic_id.h"
#include "vid.h"
#include "amdgpu.h"
#include "amdgpu_display.h"
#include "amdgpu_ucode.h"
#include "atom.h"
#include "amdgpu_dm.h"
#include "amdgpu_pm.h"
#include "amd_shared.h"
#include "amdgpu_dm_irq.h"
#include "dm_helpers.h"
#include "amdgpu_dm_mst_types.h"
#if defined(CONFIG_DEBUG_FS)
#include "amdgpu_dm_debugfs.h"
#endif
#include "ivsrcid/ivsrcid_vislands30.h"
#include <linux/module.h>
#include <linux/moduleparam.h>
#include <linux/version.h>
#include <linux/types.h>
drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail() So, unfortunately I recently made the discovery that in the upstream kernel, the only reason that amdgpu is not currently suffering from issues with runtime PM putting the GPU into suspend while it's driving displays is due to the fact that on most prime systems, we have sound devices associated with the GPU that hold their own runtime PM ref for the GPU. What this means however, is that in the event that there isn't any kind of sound device active (which can easily be reproduced by building a kernel with sound drivers disabled), the GPU will fall asleep even when there's displays active. This appears to be in part due to the fact that amdgpu has not actually ever relied on it's rpm_idle() function to be the only thing keeping it running, and normally grabs it's own power references whenever there are displays active (as can be seen with the original pre-DC codepath in amdgpu_display_crtc_set_config() in amdgpu_display.c). This means it's very likely that this bug was introduced during the switch over the DC. So to fix this, we start grabbing runtime PM references every time we enable a previously disabled CRTC in atomic_commit_tail(). This appears to be the correct solution, as it matches up with what i915 does in i915/intel_runtime_pm.c. The one sideaffect of this is that we ignore the variable that the pre-DC code used to use for tracking when it needed runtime PM refs, adev->have_disp_power_ref. This is mainly because there's no way for a driver to tell whether or not all of it's CRTCs are enabled or disabled when we've begun committing an atomic state, as there may be CRTC commits happening in parallel that aren't contained within the atomic state being committed. So, it's safer to just get/put a reference for each CRTC being enabled or disabled in the new atomic state. Signed-off-by: Lyude Paul <lyude@redhat.com> Acked-by: Christian König <christian.koenig@amd.com>. Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org
2018-06-04 13:35:03 -06:00
#include <linux/pm_runtime.h>
#include <linux/pci.h>
#include <linux/firmware.h>
#include <linux/component.h>
#include <drm/drm_atomic.h>
#include <drm/drm_atomic_uapi.h>
#include <drm/drm_atomic_helper.h>
#include <drm/drm_dp_mst_helper.h>
#include <drm/drm_fb_helper.h>
#include <drm/drm_fourcc.h>
#include <drm/drm_edid.h>
#include <drm/drm_vblank.h>
#include <drm/drm_audio_component.h>
#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
#include "ivsrcid/dcn/irqsrcs_dcn_1_0.h"
#include "dcn/dcn_1_0_offset.h"
#include "dcn/dcn_1_0_sh_mask.h"
#include "soc15_hw_ip.h"
#include "vega10_ip_offset.h"
#include "soc15_common.h"
#endif
#include "modules/inc/mod_freesync.h"
#include "modules/power/power_helpers.h"
#include "modules/inc/mod_info_packet.h"
#define FIRMWARE_RAVEN_DMCU "amdgpu/raven_dmcu.bin"
MODULE_FIRMWARE(FIRMWARE_RAVEN_DMCU);
/**
* DOC: overview
*
* The AMDgpu display manager, **amdgpu_dm** (or even simpler,
* **dm**) sits between DRM and DC. It acts as a liason, converting DRM
* requests into DC requests, and DC responses into DRM responses.
*
* The root control structure is &struct amdgpu_display_manager.
*/
/* basic init/fini API */
static int amdgpu_dm_init(struct amdgpu_device *adev);
static void amdgpu_dm_fini(struct amdgpu_device *adev);
/*
* initializes drm_device display related structures, based on the information
* provided by DAL. The drm strcutures are: drm_crtc, drm_connector,
* drm_encoder, drm_mode_config
*
* Returns 0 on success
*/
static int amdgpu_dm_initialize_drm_device(struct amdgpu_device *adev);
/* removes and deallocates the drm structures, created by the above function */
static void amdgpu_dm_destroy_drm_device(struct amdgpu_display_manager *dm);
static void
amdgpu_dm_update_connector_after_detect(struct amdgpu_dm_connector *aconnector);
static int amdgpu_dm_plane_init(struct amdgpu_display_manager *dm,
struct drm_plane *plane,
unsigned long possible_crtcs,
const struct dc_plane_cap *plane_cap);
static int amdgpu_dm_crtc_init(struct amdgpu_display_manager *dm,
struct drm_plane *plane,
uint32_t link_index);
static int amdgpu_dm_connector_init(struct amdgpu_display_manager *dm,
struct amdgpu_dm_connector *amdgpu_dm_connector,
uint32_t link_index,
struct amdgpu_encoder *amdgpu_encoder);
static int amdgpu_dm_encoder_init(struct drm_device *dev,
struct amdgpu_encoder *aencoder,
uint32_t link_index);
static int amdgpu_dm_connector_get_modes(struct drm_connector *connector);
static int amdgpu_dm_atomic_commit(struct drm_device *dev,
struct drm_atomic_state *state,
bool nonblock);
static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_state *state);
static int amdgpu_dm_atomic_check(struct drm_device *dev,
struct drm_atomic_state *state);
static void handle_cursor_update(struct drm_plane *plane,
struct drm_plane_state *old_plane_state);
/*
* dm_vblank_get_counter
*
* @brief
* Get counter for number of vertical blanks
*
* @param
* struct amdgpu_device *adev - [in] desired amdgpu device
* int disp_idx - [in] which CRTC to get the counter from
*
* @return
* Counter for vertical blanks
*/
static u32 dm_vblank_get_counter(struct amdgpu_device *adev, int crtc)
{
if (crtc >= adev->mode_info.num_crtc)
return 0;
else {
struct amdgpu_crtc *acrtc = adev->mode_info.crtcs[crtc];
struct dm_crtc_state *acrtc_state = to_dm_crtc_state(
acrtc->base.state);
if (acrtc_state->stream == NULL) {
DRM_ERROR("dc_stream_state is NULL for crtc '%d'!\n",
crtc);
return 0;
}
return dc_stream_get_vblank_counter(acrtc_state->stream);
}
}
static int dm_crtc_get_scanoutpos(struct amdgpu_device *adev, int crtc,
u32 *vbl, u32 *position)
{
uint32_t v_blank_start, v_blank_end, h_position, v_position;
if ((crtc < 0) || (crtc >= adev->mode_info.num_crtc))
return -EINVAL;
else {
struct amdgpu_crtc *acrtc = adev->mode_info.crtcs[crtc];
struct dm_crtc_state *acrtc_state = to_dm_crtc_state(
acrtc->base.state);
if (acrtc_state->stream == NULL) {
DRM_ERROR("dc_stream_state is NULL for crtc '%d'!\n",
crtc);
return 0;
}
/*
* TODO rework base driver to use values directly.
* for now parse it back into reg-format
*/
dc_stream_get_scanoutpos(acrtc_state->stream,
&v_blank_start,
&v_blank_end,
&h_position,
&v_position);
*position = v_position | (h_position << 16);
*vbl = v_blank_start | (v_blank_end << 16);
}
return 0;
}
static bool dm_is_idle(void *handle)
{
/* XXX todo */
return true;
}
static int dm_wait_for_idle(void *handle)
{
/* XXX todo */
return 0;
}
static bool dm_check_soft_reset(void *handle)
{
return false;
}
static int dm_soft_reset(void *handle)
{
/* XXX todo */
return 0;
}
static struct amdgpu_crtc *
get_crtc_by_otg_inst(struct amdgpu_device *adev,
int otg_inst)
{
struct drm_device *dev = adev->ddev;
struct drm_crtc *crtc;
struct amdgpu_crtc *amdgpu_crtc;
if (otg_inst == -1) {
WARN_ON(1);
return adev->mode_info.crtcs[0];
}
list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
amdgpu_crtc = to_amdgpu_crtc(crtc);
if (amdgpu_crtc->otg_inst == otg_inst)
return amdgpu_crtc;
}
return NULL;
}
static inline bool amdgpu_dm_vrr_active(struct dm_crtc_state *dm_state)
{
return dm_state->freesync_config.state == VRR_STATE_ACTIVE_VARIABLE ||
dm_state->freesync_config.state == VRR_STATE_ACTIVE_FIXED;
}
static void dm_pflip_high_irq(void *interrupt_params)
{
struct amdgpu_crtc *amdgpu_crtc;
struct common_irq_params *irq_params = interrupt_params;
struct amdgpu_device *adev = irq_params->adev;
unsigned long flags;
drm/amd/display: Make pageflip event delivery compatible with VRR. We want vblank counts and timestamps of flip completion as sent in pageflip completion events to be consistent with the vblank count and timestamp of the vblank of flip completion, like in non VRR mode. In VRR mode, drm_update_vblank_count() - and thereby vblank count and timestamp updates - must be delayed until after the end of front-porch of each vblank, as it is only safe to calculate vblank timestamps outside of the front-porch, when we actually know when the vblank will end or has ended. The function drm_update_vblank_count() which updates timestamps and counts gets called by drm_crtc_accurate_vblank_count() or by drm_crtc_handle_vblank(). Therefore we must make sure that pageflip events for a completed flip are only sent out after drm_crtc_accurate_vblank_count() or drm_crtc_handle_vblank() is executed, after end of front-porch for the vblank of flip completion. Two cases: a) Pageflip irq handler executes inside front-porch: In this case we must defer sending pageflip events until drm_crtc_handle_vblank() executes after end of front-porch, and thereby calculates proper vblank count and timestamp. Iow. the pflip irq handler must just arm a pageflip event to be sent out by drm_crtc_handle_vblank() later on. b) Pageflip irq handler executes after end of front-porch, e.g., after flip completion in back-porch or due to a massively delayed handler invocation into the active scanout of the new frame. In this case we can call drm_crtc_accurate_vblank_count() to safely force calculation of a proper vblank count and timestamp, and must send the pageflip completion event ourselves from the pageflip irq handler. This is the same behaviour as needed for standard fixed refresh rate mode. To decide from within pageflip handler if we are in case a) or b), we check the current scanout position against the boundary of front-porch. In non-VRR mode we just do what we did in the past. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-29 06:00:57 -06:00
struct drm_pending_vblank_event *e;
struct dm_crtc_state *acrtc_state;
uint32_t vpos, hpos, v_blank_start, v_blank_end;
bool vrr_active;
amdgpu_crtc = get_crtc_by_otg_inst(adev, irq_params->irq_src - IRQ_TYPE_PFLIP);
/* IRQ could occur when in initial stage */
/* TODO work and BO cleanup */
if (amdgpu_crtc == NULL) {
DRM_DEBUG_DRIVER("CRTC is null, returning.\n");
return;
}
spin_lock_irqsave(&adev->ddev->event_lock, flags);
if (amdgpu_crtc->pflip_status != AMDGPU_FLIP_SUBMITTED){
DRM_DEBUG_DRIVER("amdgpu_crtc->pflip_status = %d !=AMDGPU_FLIP_SUBMITTED(%d) on crtc:%d[%p] \n",
amdgpu_crtc->pflip_status,
AMDGPU_FLIP_SUBMITTED,
amdgpu_crtc->crtc_id,
amdgpu_crtc);
spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
return;
}
drm/amd/display: Make pageflip event delivery compatible with VRR. We want vblank counts and timestamps of flip completion as sent in pageflip completion events to be consistent with the vblank count and timestamp of the vblank of flip completion, like in non VRR mode. In VRR mode, drm_update_vblank_count() - and thereby vblank count and timestamp updates - must be delayed until after the end of front-porch of each vblank, as it is only safe to calculate vblank timestamps outside of the front-porch, when we actually know when the vblank will end or has ended. The function drm_update_vblank_count() which updates timestamps and counts gets called by drm_crtc_accurate_vblank_count() or by drm_crtc_handle_vblank(). Therefore we must make sure that pageflip events for a completed flip are only sent out after drm_crtc_accurate_vblank_count() or drm_crtc_handle_vblank() is executed, after end of front-porch for the vblank of flip completion. Two cases: a) Pageflip irq handler executes inside front-porch: In this case we must defer sending pageflip events until drm_crtc_handle_vblank() executes after end of front-porch, and thereby calculates proper vblank count and timestamp. Iow. the pflip irq handler must just arm a pageflip event to be sent out by drm_crtc_handle_vblank() later on. b) Pageflip irq handler executes after end of front-porch, e.g., after flip completion in back-porch or due to a massively delayed handler invocation into the active scanout of the new frame. In this case we can call drm_crtc_accurate_vblank_count() to safely force calculation of a proper vblank count and timestamp, and must send the pageflip completion event ourselves from the pageflip irq handler. This is the same behaviour as needed for standard fixed refresh rate mode. To decide from within pageflip handler if we are in case a) or b), we check the current scanout position against the boundary of front-porch. In non-VRR mode we just do what we did in the past. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-29 06:00:57 -06:00
/* page flip completed. */
e = amdgpu_crtc->event;
amdgpu_crtc->event = NULL;
drm/amd/display: Make pageflip event delivery compatible with VRR. We want vblank counts and timestamps of flip completion as sent in pageflip completion events to be consistent with the vblank count and timestamp of the vblank of flip completion, like in non VRR mode. In VRR mode, drm_update_vblank_count() - and thereby vblank count and timestamp updates - must be delayed until after the end of front-porch of each vblank, as it is only safe to calculate vblank timestamps outside of the front-porch, when we actually know when the vblank will end or has ended. The function drm_update_vblank_count() which updates timestamps and counts gets called by drm_crtc_accurate_vblank_count() or by drm_crtc_handle_vblank(). Therefore we must make sure that pageflip events for a completed flip are only sent out after drm_crtc_accurate_vblank_count() or drm_crtc_handle_vblank() is executed, after end of front-porch for the vblank of flip completion. Two cases: a) Pageflip irq handler executes inside front-porch: In this case we must defer sending pageflip events until drm_crtc_handle_vblank() executes after end of front-porch, and thereby calculates proper vblank count and timestamp. Iow. the pflip irq handler must just arm a pageflip event to be sent out by drm_crtc_handle_vblank() later on. b) Pageflip irq handler executes after end of front-porch, e.g., after flip completion in back-porch or due to a massively delayed handler invocation into the active scanout of the new frame. In this case we can call drm_crtc_accurate_vblank_count() to safely force calculation of a proper vblank count and timestamp, and must send the pageflip completion event ourselves from the pageflip irq handler. This is the same behaviour as needed for standard fixed refresh rate mode. To decide from within pageflip handler if we are in case a) or b), we check the current scanout position against the boundary of front-porch. In non-VRR mode we just do what we did in the past. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-29 06:00:57 -06:00
if (!e)
WARN_ON(1);
drm/amd/display: Make pageflip event delivery compatible with VRR. We want vblank counts and timestamps of flip completion as sent in pageflip completion events to be consistent with the vblank count and timestamp of the vblank of flip completion, like in non VRR mode. In VRR mode, drm_update_vblank_count() - and thereby vblank count and timestamp updates - must be delayed until after the end of front-porch of each vblank, as it is only safe to calculate vblank timestamps outside of the front-porch, when we actually know when the vblank will end or has ended. The function drm_update_vblank_count() which updates timestamps and counts gets called by drm_crtc_accurate_vblank_count() or by drm_crtc_handle_vblank(). Therefore we must make sure that pageflip events for a completed flip are only sent out after drm_crtc_accurate_vblank_count() or drm_crtc_handle_vblank() is executed, after end of front-porch for the vblank of flip completion. Two cases: a) Pageflip irq handler executes inside front-porch: In this case we must defer sending pageflip events until drm_crtc_handle_vblank() executes after end of front-porch, and thereby calculates proper vblank count and timestamp. Iow. the pflip irq handler must just arm a pageflip event to be sent out by drm_crtc_handle_vblank() later on. b) Pageflip irq handler executes after end of front-porch, e.g., after flip completion in back-porch or due to a massively delayed handler invocation into the active scanout of the new frame. In this case we can call drm_crtc_accurate_vblank_count() to safely force calculation of a proper vblank count and timestamp, and must send the pageflip completion event ourselves from the pageflip irq handler. This is the same behaviour as needed for standard fixed refresh rate mode. To decide from within pageflip handler if we are in case a) or b), we check the current scanout position against the boundary of front-porch. In non-VRR mode we just do what we did in the past. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-29 06:00:57 -06:00
acrtc_state = to_dm_crtc_state(amdgpu_crtc->base.state);
vrr_active = amdgpu_dm_vrr_active(acrtc_state);
/* Fixed refresh rate, or VRR scanout position outside front-porch? */
if (!vrr_active ||
!dc_stream_get_scanoutpos(acrtc_state->stream, &v_blank_start,
&v_blank_end, &hpos, &vpos) ||
(vpos < v_blank_start)) {
/* Update to correct count and vblank timestamp if racing with
* vblank irq. This also updates to the correct vblank timestamp
* even in VRR mode, as scanout is past the front-porch atm.
*/
drm_crtc_accurate_vblank_count(&amdgpu_crtc->base);
drm/amd/display: Make pageflip event delivery compatible with VRR. We want vblank counts and timestamps of flip completion as sent in pageflip completion events to be consistent with the vblank count and timestamp of the vblank of flip completion, like in non VRR mode. In VRR mode, drm_update_vblank_count() - and thereby vblank count and timestamp updates - must be delayed until after the end of front-porch of each vblank, as it is only safe to calculate vblank timestamps outside of the front-porch, when we actually know when the vblank will end or has ended. The function drm_update_vblank_count() which updates timestamps and counts gets called by drm_crtc_accurate_vblank_count() or by drm_crtc_handle_vblank(). Therefore we must make sure that pageflip events for a completed flip are only sent out after drm_crtc_accurate_vblank_count() or drm_crtc_handle_vblank() is executed, after end of front-porch for the vblank of flip completion. Two cases: a) Pageflip irq handler executes inside front-porch: In this case we must defer sending pageflip events until drm_crtc_handle_vblank() executes after end of front-porch, and thereby calculates proper vblank count and timestamp. Iow. the pflip irq handler must just arm a pageflip event to be sent out by drm_crtc_handle_vblank() later on. b) Pageflip irq handler executes after end of front-porch, e.g., after flip completion in back-porch or due to a massively delayed handler invocation into the active scanout of the new frame. In this case we can call drm_crtc_accurate_vblank_count() to safely force calculation of a proper vblank count and timestamp, and must send the pageflip completion event ourselves from the pageflip irq handler. This is the same behaviour as needed for standard fixed refresh rate mode. To decide from within pageflip handler if we are in case a) or b), we check the current scanout position against the boundary of front-porch. In non-VRR mode we just do what we did in the past. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-29 06:00:57 -06:00
/* Wake up userspace by sending the pageflip event with proper
* count and timestamp of vblank of flip completion.
*/
if (e) {
drm_crtc_send_vblank_event(&amdgpu_crtc->base, e);
/* Event sent, so done with vblank for this flip */
drm_crtc_vblank_put(&amdgpu_crtc->base);
}
} else if (e) {
/* VRR active and inside front-porch: vblank count and
* timestamp for pageflip event will only be up to date after
* drm_crtc_handle_vblank() has been executed from late vblank
* irq handler after start of back-porch (vline 0). We queue the
* pageflip event for send-out by drm_crtc_handle_vblank() with
* updated timestamp and count, once it runs after us.
*
* We need to open-code this instead of using the helper
* drm_crtc_arm_vblank_event(), as that helper would
* call drm_crtc_accurate_vblank_count(), which we must
* not call in VRR mode while we are in front-porch!
*/
/* sequence will be replaced by real count during send-out. */
e->sequence = drm_crtc_vblank_count(&amdgpu_crtc->base);
e->pipe = amdgpu_crtc->crtc_id;
list_add_tail(&e->base.link, &adev->ddev->vblank_event_list);
e = NULL;
}
/* Keep track of vblank of this flip for flip throttling. We use the
* cooked hw counter, as that one incremented at start of this vblank
* of pageflip completion, so last_flip_vblank is the forbidden count
* for queueing new pageflips if vsync + VRR is enabled.
*/
amdgpu_crtc->last_flip_vblank = amdgpu_get_vblank_counter_kms(adev->ddev,
amdgpu_crtc->crtc_id);
amdgpu_crtc->pflip_status = AMDGPU_FLIP_NONE;
spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
drm/amd/display: Make pageflip event delivery compatible with VRR. We want vblank counts and timestamps of flip completion as sent in pageflip completion events to be consistent with the vblank count and timestamp of the vblank of flip completion, like in non VRR mode. In VRR mode, drm_update_vblank_count() - and thereby vblank count and timestamp updates - must be delayed until after the end of front-porch of each vblank, as it is only safe to calculate vblank timestamps outside of the front-porch, when we actually know when the vblank will end or has ended. The function drm_update_vblank_count() which updates timestamps and counts gets called by drm_crtc_accurate_vblank_count() or by drm_crtc_handle_vblank(). Therefore we must make sure that pageflip events for a completed flip are only sent out after drm_crtc_accurate_vblank_count() or drm_crtc_handle_vblank() is executed, after end of front-porch for the vblank of flip completion. Two cases: a) Pageflip irq handler executes inside front-porch: In this case we must defer sending pageflip events until drm_crtc_handle_vblank() executes after end of front-porch, and thereby calculates proper vblank count and timestamp. Iow. the pflip irq handler must just arm a pageflip event to be sent out by drm_crtc_handle_vblank() later on. b) Pageflip irq handler executes after end of front-porch, e.g., after flip completion in back-porch or due to a massively delayed handler invocation into the active scanout of the new frame. In this case we can call drm_crtc_accurate_vblank_count() to safely force calculation of a proper vblank count and timestamp, and must send the pageflip completion event ourselves from the pageflip irq handler. This is the same behaviour as needed for standard fixed refresh rate mode. To decide from within pageflip handler if we are in case a) or b), we check the current scanout position against the boundary of front-porch. In non-VRR mode we just do what we did in the past. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-29 06:00:57 -06:00
DRM_DEBUG_DRIVER("crtc:%d[%p], pflip_stat:AMDGPU_FLIP_NONE, vrr[%d]-fp %d\n",
amdgpu_crtc->crtc_id, amdgpu_crtc,
vrr_active, (int) !e);
}
static void dm_vupdate_high_irq(void *interrupt_params)
{
struct common_irq_params *irq_params = interrupt_params;
struct amdgpu_device *adev = irq_params->adev;
struct amdgpu_crtc *acrtc;
struct dm_crtc_state *acrtc_state;
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
unsigned long flags;
acrtc = get_crtc_by_otg_inst(adev, irq_params->irq_src - IRQ_TYPE_VUPDATE);
if (acrtc) {
acrtc_state = to_dm_crtc_state(acrtc->base.state);
DRM_DEBUG_DRIVER("crtc:%d, vupdate-vrr:%d\n", acrtc->crtc_id,
amdgpu_dm_vrr_active(acrtc_state));
/* Core vblank handling is done here after end of front-porch in
* vrr mode, as vblank timestamping will give valid results
* while now done after front-porch. This will also deliver
* page-flip completion events that have been queued to us
* if a pageflip happened inside front-porch.
*/
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
if (amdgpu_dm_vrr_active(acrtc_state)) {
drm_crtc_handle_vblank(&acrtc->base);
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
/* BTR processing for pre-DCE12 ASICs */
if (acrtc_state->stream &&
adev->family < AMDGPU_FAMILY_AI) {
spin_lock_irqsave(&adev->ddev->event_lock, flags);
mod_freesync_handle_v_update(
adev->dm.freesync_module,
acrtc_state->stream,
&acrtc_state->vrr_params);
dc_stream_adjust_vmin_vmax(
adev->dm.dc,
acrtc_state->stream,
&acrtc_state->vrr_params.adjust);
spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
}
}
}
}
static void dm_crtc_high_irq(void *interrupt_params)
{
struct common_irq_params *irq_params = interrupt_params;
struct amdgpu_device *adev = irq_params->adev;
struct amdgpu_crtc *acrtc;
struct dm_crtc_state *acrtc_state;
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
unsigned long flags;
acrtc = get_crtc_by_otg_inst(adev, irq_params->irq_src - IRQ_TYPE_VBLANK);
if (acrtc) {
acrtc_state = to_dm_crtc_state(acrtc->base.state);
DRM_DEBUG_DRIVER("crtc:%d, vupdate-vrr:%d\n", acrtc->crtc_id,
amdgpu_dm_vrr_active(acrtc_state));
/* Core vblank handling at start of front-porch is only possible
* in non-vrr mode, as only there vblank timestamping will give
* valid results while done in front-porch. Otherwise defer it
* to dm_vupdate_high_irq after end of front-porch.
*/
if (!amdgpu_dm_vrr_active(acrtc_state))
drm_crtc_handle_vblank(&acrtc->base);
/* Following stuff must happen at start of vblank, for crc
* computation and below-the-range btr support in vrr mode.
*/
amdgpu_dm_crtc_handle_crc_irq(&acrtc->base);
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
if (acrtc_state->stream && adev->family >= AMDGPU_FAMILY_AI &&
acrtc_state->vrr_params.supported &&
acrtc_state->freesync_config.state == VRR_STATE_ACTIVE_VARIABLE) {
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
spin_lock_irqsave(&adev->ddev->event_lock, flags);
mod_freesync_handle_v_update(
adev->dm.freesync_module,
acrtc_state->stream,
&acrtc_state->vrr_params);
dc_stream_adjust_vmin_vmax(
adev->dm.dc,
acrtc_state->stream,
&acrtc_state->vrr_params.adjust);
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
}
}
}
static int dm_set_clockgating_state(void *handle,
enum amd_clockgating_state state)
{
return 0;
}
static int dm_set_powergating_state(void *handle,
enum amd_powergating_state state)
{
return 0;
}
/* Prototypes of private functions */
static int dm_early_init(void* handle);
/* Allocate memory for FBC compressed data */
static void amdgpu_dm_fbc_init(struct drm_connector *connector)
{
struct drm_device *dev = connector->dev;
struct amdgpu_device *adev = dev->dev_private;
struct dm_comressor_info *compressor = &adev->dm.compressor;
struct amdgpu_dm_connector *aconn = to_amdgpu_dm_connector(connector);
struct drm_display_mode *mode;
unsigned long max_size = 0;
if (adev->dm.dc->fbc_compressor == NULL)
return;
if (aconn->dc_link->connector_signal != SIGNAL_TYPE_EDP)
return;
if (compressor->bo_ptr)
return;
list_for_each_entry(mode, &connector->modes, head) {
if (max_size < mode->htotal * mode->vtotal)
max_size = mode->htotal * mode->vtotal;
}
if (max_size) {
int r = amdgpu_bo_create_kernel(adev, max_size * 4, PAGE_SIZE,
AMDGPU_GEM_DOMAIN_GTT, &compressor->bo_ptr,
&compressor->gpu_addr, &compressor->cpu_addr);
if (r)
DRM_ERROR("DM: Failed to initialize FBC\n");
else {
adev->dm.dc->ctx->fbc_gpu_addr = compressor->gpu_addr;
DRM_INFO("DM: FBC alloc %lu\n", max_size*4);
}
}
}
static int amdgpu_dm_audio_component_get_eld(struct device *kdev, int port,
int pipe, bool *enabled,
unsigned char *buf, int max_bytes)
{
struct drm_device *dev = dev_get_drvdata(kdev);
struct amdgpu_device *adev = dev->dev_private;
struct drm_connector *connector;
struct drm_connector_list_iter conn_iter;
struct amdgpu_dm_connector *aconnector;
int ret = 0;
*enabled = false;
mutex_lock(&adev->dm.audio_lock);
drm_connector_list_iter_begin(dev, &conn_iter);
drm_for_each_connector_iter(connector, &conn_iter) {
aconnector = to_amdgpu_dm_connector(connector);
if (aconnector->audio_inst != port)
continue;
*enabled = true;
ret = drm_eld_size(connector->eld);
memcpy(buf, connector->eld, min(max_bytes, ret));
break;
}
drm_connector_list_iter_end(&conn_iter);
mutex_unlock(&adev->dm.audio_lock);
DRM_DEBUG_KMS("Get ELD : idx=%d ret=%d en=%d\n", port, ret, *enabled);
return ret;
}
static const struct drm_audio_component_ops amdgpu_dm_audio_component_ops = {
.get_eld = amdgpu_dm_audio_component_get_eld,
};
static int amdgpu_dm_audio_component_bind(struct device *kdev,
struct device *hda_kdev, void *data)
{
struct drm_device *dev = dev_get_drvdata(kdev);
struct amdgpu_device *adev = dev->dev_private;
struct drm_audio_component *acomp = data;
acomp->ops = &amdgpu_dm_audio_component_ops;
acomp->dev = kdev;
adev->dm.audio_component = acomp;
return 0;
}
static void amdgpu_dm_audio_component_unbind(struct device *kdev,
struct device *hda_kdev, void *data)
{
struct drm_device *dev = dev_get_drvdata(kdev);
struct amdgpu_device *adev = dev->dev_private;
struct drm_audio_component *acomp = data;
acomp->ops = NULL;
acomp->dev = NULL;
adev->dm.audio_component = NULL;
}
static const struct component_ops amdgpu_dm_audio_component_bind_ops = {
.bind = amdgpu_dm_audio_component_bind,
.unbind = amdgpu_dm_audio_component_unbind,
};
static int amdgpu_dm_audio_init(struct amdgpu_device *adev)
{
int i, ret;
if (!amdgpu_audio)
return 0;
adev->mode_info.audio.enabled = true;
adev->mode_info.audio.num_pins = adev->dm.dc->res_pool->audio_count;
for (i = 0; i < adev->mode_info.audio.num_pins; i++) {
adev->mode_info.audio.pin[i].channels = -1;
adev->mode_info.audio.pin[i].rate = -1;
adev->mode_info.audio.pin[i].bits_per_sample = -1;
adev->mode_info.audio.pin[i].status_bits = 0;
adev->mode_info.audio.pin[i].category_code = 0;
adev->mode_info.audio.pin[i].connected = false;
adev->mode_info.audio.pin[i].id =
adev->dm.dc->res_pool->audios[i]->inst;
adev->mode_info.audio.pin[i].offset = 0;
}
ret = component_add(adev->dev, &amdgpu_dm_audio_component_bind_ops);
if (ret < 0)
return ret;
adev->dm.audio_registered = true;
return 0;
}
static void amdgpu_dm_audio_fini(struct amdgpu_device *adev)
{
if (!amdgpu_audio)
return;
if (!adev->mode_info.audio.enabled)
return;
if (adev->dm.audio_registered) {
component_del(adev->dev, &amdgpu_dm_audio_component_bind_ops);
adev->dm.audio_registered = false;
}
/* TODO: Disable audio? */
adev->mode_info.audio.enabled = false;
}
void amdgpu_dm_audio_eld_notify(struct amdgpu_device *adev, int pin)
{
struct drm_audio_component *acomp = adev->dm.audio_component;
if (acomp && acomp->audio_ops && acomp->audio_ops->pin_eld_notify) {
DRM_DEBUG_KMS("Notify ELD: %d\n", pin);
acomp->audio_ops->pin_eld_notify(acomp->audio_ops->audio_ptr,
pin, -1);
}
}
static int amdgpu_dm_init(struct amdgpu_device *adev)
{
struct dc_init_data init_data;
adev->dm.ddev = adev->ddev;
adev->dm.adev = adev;
/* Zero all the fields */
memset(&init_data, 0, sizeof(init_data));
mutex_init(&adev->dm.dc_lock);
mutex_init(&adev->dm.audio_lock);
if(amdgpu_dm_irq_init(adev)) {
DRM_ERROR("amdgpu: failed to initialize DM IRQ support.\n");
goto error;
}
init_data.asic_id.chip_family = adev->family;
init_data.asic_id.pci_revision_id = adev->rev_id;
init_data.asic_id.hw_internal_rev = adev->external_rev_id;
init_data.asic_id.vram_width = adev->gmc.vram_width;
/* TODO: initialize init_data.asic_id.vram_type here!!!! */
init_data.asic_id.atombios_base_address =
adev->mode_info.atom_context->bios;
init_data.driver = adev;
adev->dm.cgs_device = amdgpu_cgs_create_device(adev);
if (!adev->dm.cgs_device) {
DRM_ERROR("amdgpu: failed to create cgs device.\n");
goto error;
}
init_data.cgs_device = adev->dm.cgs_device;
init_data.dce_environment = DCE_ENV_PRODUCTION_DRV;
/*
* TODO debug why this doesn't work on Raven
*/
if (adev->flags & AMD_IS_APU &&
adev->asic_type >= CHIP_CARRIZO &&
adev->asic_type < CHIP_RAVEN)
init_data.flags.gpu_vm_support = true;
if (amdgpu_dc_feature_mask & DC_FBC_MASK)
init_data.flags.fbc_support = true;
if (amdgpu_dc_feature_mask & DC_MULTI_MON_PP_MCLK_SWITCH_MASK)
init_data.flags.multi_mon_pp_mclk_switch = true;
init_data.flags.power_down_display_on_boot = true;
#ifdef CONFIG_DRM_AMD_DC_DCN2_0
init_data.soc_bounding_box = adev->dm.soc_bounding_box;
#endif
/* Display Core create. */
adev->dm.dc = dc_create(&init_data);
if (adev->dm.dc) {
DRM_INFO("Display Core initialized with v%s!\n", DC_VER);
} else {
DRM_INFO("Display Core failed to initialize with v%s!\n", DC_VER);
goto error;
}
adev->dm.freesync_module = mod_freesync_create(adev->dm.dc);
if (!adev->dm.freesync_module) {
DRM_ERROR(
"amdgpu: failed to initialize freesync_module.\n");
} else
DRM_DEBUG_DRIVER("amdgpu: freesync_module init done %p.\n",
adev->dm.freesync_module);
amdgpu_dm_init_color_mod();
if (amdgpu_dm_initialize_drm_device(adev)) {
DRM_ERROR(
"amdgpu: failed to initialize sw for display support.\n");
goto error;
}
/* Update the actual used number of crtc */
adev->mode_info.num_crtc = adev->dm.display_indexes_num;
/* TODO: Add_display_info? */
/* TODO use dynamic cursor width */
adev->ddev->mode_config.cursor_width = adev->dm.dc->caps.max_cursor_size;
adev->ddev->mode_config.cursor_height = adev->dm.dc->caps.max_cursor_size;
if (drm_vblank_init(adev->ddev, adev->dm.display_indexes_num)) {
DRM_ERROR(
"amdgpu: failed to initialize sw for display support.\n");
goto error;
}
#if defined(CONFIG_DEBUG_FS)
if (dtn_debugfs_init(adev))
DRM_ERROR("amdgpu: failed initialize dtn debugfs support.\n");
#endif
DRM_DEBUG_DRIVER("KMS initialized.\n");
return 0;
error:
amdgpu_dm_fini(adev);
return -EINVAL;
}
static void amdgpu_dm_fini(struct amdgpu_device *adev)
{
amdgpu_dm_audio_fini(adev);
amdgpu_dm_destroy_drm_device(&adev->dm);
/* DC Destroy TODO: Replace destroy DAL */
if (adev->dm.dc)
dc_destroy(&adev->dm.dc);
/*
* TODO: pageflip, vlank interrupt
*
* amdgpu_dm_irq_fini(adev);
*/
if (adev->dm.cgs_device) {
amdgpu_cgs_destroy_device(adev->dm.cgs_device);
adev->dm.cgs_device = NULL;
}
if (adev->dm.freesync_module) {
mod_freesync_destroy(adev->dm.freesync_module);
adev->dm.freesync_module = NULL;
}
mutex_destroy(&adev->dm.audio_lock);
mutex_destroy(&adev->dm.dc_lock);
return;
}
static int load_dmcu_fw(struct amdgpu_device *adev)
{
const char *fw_name_dmcu = NULL;
int r;
const struct dmcu_firmware_header_v1_0 *hdr;
switch(adev->asic_type) {
case CHIP_BONAIRE:
case CHIP_HAWAII:
case CHIP_KAVERI:
case CHIP_KABINI:
case CHIP_MULLINS:
case CHIP_TONGA:
case CHIP_FIJI:
case CHIP_CARRIZO:
case CHIP_STONEY:
case CHIP_POLARIS11:
case CHIP_POLARIS10:
case CHIP_POLARIS12:
case CHIP_VEGAM:
case CHIP_VEGA10:
case CHIP_VEGA12:
case CHIP_VEGA20:
case CHIP_NAVI10:
case CHIP_NAVI14:
case CHIP_NAVI12:
case CHIP_RENOIR:
return 0;
case CHIP_RAVEN:
if (ASICREV_IS_PICASSO(adev->external_rev_id))
fw_name_dmcu = FIRMWARE_RAVEN_DMCU;
else if (ASICREV_IS_RAVEN2(adev->external_rev_id))
fw_name_dmcu = FIRMWARE_RAVEN_DMCU;
else
return 0;
break;
default:
DRM_ERROR("Unsupported ASIC type: 0x%X\n", adev->asic_type);
return -EINVAL;
}
if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
DRM_DEBUG_KMS("dm: DMCU firmware not supported on direct or SMU loading\n");
return 0;
}
r = request_firmware_direct(&adev->dm.fw_dmcu, fw_name_dmcu, adev->dev);
if (r == -ENOENT) {
/* DMCU firmware is not necessary, so don't raise a fuss if it's missing */
DRM_DEBUG_KMS("dm: DMCU firmware not found\n");
adev->dm.fw_dmcu = NULL;
return 0;
}
if (r) {
dev_err(adev->dev, "amdgpu_dm: Can't load firmware \"%s\"\n",
fw_name_dmcu);
return r;
}
r = amdgpu_ucode_validate(adev->dm.fw_dmcu);
if (r) {
dev_err(adev->dev, "amdgpu_dm: Can't validate firmware \"%s\"\n",
fw_name_dmcu);
release_firmware(adev->dm.fw_dmcu);
adev->dm.fw_dmcu = NULL;
return r;
}
hdr = (const struct dmcu_firmware_header_v1_0 *)adev->dm.fw_dmcu->data;
adev->firmware.ucode[AMDGPU_UCODE_ID_DMCU_ERAM].ucode_id = AMDGPU_UCODE_ID_DMCU_ERAM;
adev->firmware.ucode[AMDGPU_UCODE_ID_DMCU_ERAM].fw = adev->dm.fw_dmcu;
adev->firmware.fw_size +=
ALIGN(le32_to_cpu(hdr->header.ucode_size_bytes) - le32_to_cpu(hdr->intv_size_bytes), PAGE_SIZE);
adev->firmware.ucode[AMDGPU_UCODE_ID_DMCU_INTV].ucode_id = AMDGPU_UCODE_ID_DMCU_INTV;
adev->firmware.ucode[AMDGPU_UCODE_ID_DMCU_INTV].fw = adev->dm.fw_dmcu;
adev->firmware.fw_size +=
ALIGN(le32_to_cpu(hdr->intv_size_bytes), PAGE_SIZE);
adev->dm.dmcu_fw_version = le32_to_cpu(hdr->header.ucode_version);
DRM_DEBUG_KMS("PSP loading DMCU firmware\n");
return 0;
}
static int dm_sw_init(void *handle)
{
struct amdgpu_device *adev = (struct amdgpu_device *)handle;
return load_dmcu_fw(adev);
}
static int dm_sw_fini(void *handle)
{
struct amdgpu_device *adev = (struct amdgpu_device *)handle;
if(adev->dm.fw_dmcu) {
release_firmware(adev->dm.fw_dmcu);
adev->dm.fw_dmcu = NULL;
}
return 0;
}
static int detect_mst_link_for_all_connectors(struct drm_device *dev)
{
struct amdgpu_dm_connector *aconnector;
struct drm_connector *connector;
int ret = 0;
drm_modeset_lock(&dev->mode_config.connection_mutex, NULL);
list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
aconnector = to_amdgpu_dm_connector(connector);
if (aconnector->dc_link->type == dc_connection_mst_branch &&
aconnector->mst_mgr.aux) {
DRM_DEBUG_DRIVER("DM_MST: starting TM on aconnector: %p [id: %d]\n",
aconnector, aconnector->base.base.id);
ret = drm_dp_mst_topology_mgr_set_mst(&aconnector->mst_mgr, true);
if (ret < 0) {
DRM_ERROR("DM_MST: Failed to start MST\n");
((struct dc_link *)aconnector->dc_link)->type = dc_connection_single;
return ret;
}
}
}
drm_modeset_unlock(&dev->mode_config.connection_mutex);
return ret;
}
static int dm_late_init(void *handle)
{
struct amdgpu_device *adev = (struct amdgpu_device *)handle;
struct dmcu_iram_parameters params;
unsigned int linear_lut[16];
int i;
struct dmcu *dmcu = NULL;
bool ret;
if (!adev->dm.fw_dmcu)
return detect_mst_link_for_all_connectors(adev->ddev);
dmcu = adev->dm.dc->res_pool->dmcu;
for (i = 0; i < 16; i++)
linear_lut[i] = 0xFFFF * i / 15;
params.set = 0;
params.backlight_ramping_start = 0xCCCC;
params.backlight_ramping_reduction = 0xCCCCCCCC;
params.backlight_lut_array_size = 16;
params.backlight_lut_array = linear_lut;
/* Min backlight level after ABM reduction, Don't allow below 1%
* 0xFFFF x 0.01 = 0x28F
*/
params.min_abm_backlight = 0x28F;
ret = dmcu_load_iram(dmcu, params);
if (!ret)
return -EINVAL;
return detect_mst_link_for_all_connectors(adev->ddev);
}
static void s3_handle_mst(struct drm_device *dev, bool suspend)
{
struct amdgpu_dm_connector *aconnector;
struct drm_connector *connector;
struct drm_dp_mst_topology_mgr *mgr;
int ret;
bool need_hotplug = false;
drm_modeset_lock(&dev->mode_config.connection_mutex, NULL);
list_for_each_entry(connector, &dev->mode_config.connector_list,
head) {
aconnector = to_amdgpu_dm_connector(connector);
if (aconnector->dc_link->type != dc_connection_mst_branch ||
aconnector->mst_port)
continue;
mgr = &aconnector->mst_mgr;
if (suspend) {
drm_dp_mst_topology_mgr_suspend(mgr);
} else {
ret = drm_dp_mst_topology_mgr_resume(mgr);
if (ret < 0) {
drm_dp_mst_topology_mgr_set_mst(mgr, false);
need_hotplug = true;
}
}
}
drm_modeset_unlock(&dev->mode_config.connection_mutex);
if (need_hotplug)
drm_kms_helper_hotplug_event(dev);
}
/**
* dm_hw_init() - Initialize DC device
* @handle: The base driver device containing the amdpgu_dm device.
*
* Initialize the &struct amdgpu_display_manager device. This involves calling
* the initializers of each DM component, then populating the struct with them.
*
* Although the function implies hardware initialization, both hardware and
* software are initialized here. Splitting them out to their relevant init
* hooks is a future TODO item.
*
* Some notable things that are initialized here:
*
* - Display Core, both software and hardware
* - DC modules that we need (freesync and color management)
* - DRM software states
* - Interrupt sources and handlers
* - Vblank support
* - Debug FS entries, if enabled
*/
static int dm_hw_init(void *handle)
{
struct amdgpu_device *adev = (struct amdgpu_device *)handle;
/* Create DAL display manager */
amdgpu_dm_init(adev);
amdgpu_dm_hpd_init(adev);
return 0;
}
/**
* dm_hw_fini() - Teardown DC device
* @handle: The base driver device containing the amdpgu_dm device.
*
* Teardown components within &struct amdgpu_display_manager that require
* cleanup. This involves cleaning up the DRM device, DC, and any modules that
* were loaded. Also flush IRQ workqueues and disable them.
*/
static int dm_hw_fini(void *handle)
{
struct amdgpu_device *adev = (struct amdgpu_device *)handle;
amdgpu_dm_hpd_fini(adev);
amdgpu_dm_irq_fini(adev);
amdgpu_dm_fini(adev);
return 0;
}
static int dm_suspend(void *handle)
{
struct amdgpu_device *adev = handle;
struct amdgpu_display_manager *dm = &adev->dm;
int ret = 0;
WARN_ON(adev->dm.cached_state);
adev->dm.cached_state = drm_atomic_helper_suspend(adev->ddev);
s3_handle_mst(adev->ddev, true);
amdgpu_dm_irq_suspend(adev);
dc_set_power_state(dm->dc, DC_ACPI_CM_POWER_STATE_D3);
return ret;
}
static struct amdgpu_dm_connector *
amdgpu_dm_find_first_crtc_matching_connector(struct drm_atomic_state *state,
struct drm_crtc *crtc)
{
uint32_t i;
struct drm_connector_state *new_con_state;
struct drm_connector *connector;
struct drm_crtc *crtc_from_state;
for_each_new_connector_in_state(state, connector, new_con_state, i) {
crtc_from_state = new_con_state->crtc;
if (crtc_from_state == crtc)
return to_amdgpu_dm_connector(connector);
}
return NULL;
}
static void emulated_link_detect(struct dc_link *link)
{
struct dc_sink_init_data sink_init_data = { 0 };
struct display_sink_capability sink_caps = { 0 };
enum dc_edid_status edid_status;
struct dc_context *dc_ctx = link->ctx;
struct dc_sink *sink = NULL;
struct dc_sink *prev_sink = NULL;
link->type = dc_connection_none;
prev_sink = link->local_sink;
if (prev_sink != NULL)
dc_sink_retain(prev_sink);
switch (link->connector_signal) {
case SIGNAL_TYPE_HDMI_TYPE_A: {
sink_caps.transaction_type = DDC_TRANSACTION_TYPE_I2C;
sink_caps.signal = SIGNAL_TYPE_HDMI_TYPE_A;
break;
}
case SIGNAL_TYPE_DVI_SINGLE_LINK: {
sink_caps.transaction_type = DDC_TRANSACTION_TYPE_I2C;
sink_caps.signal = SIGNAL_TYPE_DVI_SINGLE_LINK;
break;
}
case SIGNAL_TYPE_DVI_DUAL_LINK: {
sink_caps.transaction_type = DDC_TRANSACTION_TYPE_I2C;
sink_caps.signal = SIGNAL_TYPE_DVI_DUAL_LINK;
break;
}
case SIGNAL_TYPE_LVDS: {
sink_caps.transaction_type = DDC_TRANSACTION_TYPE_I2C;
sink_caps.signal = SIGNAL_TYPE_LVDS;
break;
}
case SIGNAL_TYPE_EDP: {
sink_caps.transaction_type =
DDC_TRANSACTION_TYPE_I2C_OVER_AUX;
sink_caps.signal = SIGNAL_TYPE_EDP;
break;
}
case SIGNAL_TYPE_DISPLAY_PORT: {
sink_caps.transaction_type =
DDC_TRANSACTION_TYPE_I2C_OVER_AUX;
sink_caps.signal = SIGNAL_TYPE_VIRTUAL;
break;
}
default:
DC_ERROR("Invalid connector type! signal:%d\n",
link->connector_signal);
return;
}
sink_init_data.link = link;
sink_init_data.sink_signal = sink_caps.signal;
sink = dc_sink_create(&sink_init_data);
if (!sink) {
DC_ERROR("Failed to create sink!\n");
return;
}
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
/* dc_sink_create returns a new reference */
link->local_sink = sink;
edid_status = dm_helpers_read_local_edid(
link->ctx,
link,
sink);
if (edid_status != EDID_OK)
DC_ERROR("Failed to read EDID");
}
static int dm_resume(void *handle)
{
struct amdgpu_device *adev = handle;
struct drm_device *ddev = adev->ddev;
struct amdgpu_display_manager *dm = &adev->dm;
struct amdgpu_dm_connector *aconnector;
struct drm_connector *connector;
struct drm_crtc *crtc;
struct drm_crtc_state *new_crtc_state;
struct dm_crtc_state *dm_new_crtc_state;
struct drm_plane *plane;
struct drm_plane_state *new_plane_state;
struct dm_plane_state *dm_new_plane_state;
struct dm_atomic_state *dm_state = to_dm_atomic_state(dm->atomic_obj.state);
enum dc_connection_type new_connection_type = dc_connection_none;
int i;
/* Recreate dc_state - DC invalidates it when setting power state to S3. */
dc_release_state(dm_state->context);
dm_state->context = dc_create_state(dm->dc);
/* TODO: Remove dc_state->dccg, use dc->dccg directly. */
dc_resource_state_construct(dm->dc, dm_state->context);
/* power on hardware */
dc_set_power_state(dm->dc, DC_ACPI_CM_POWER_STATE_D0);
/* program HPD filter */
dc_resume(dm->dc);
/* On resume we need to rewrite the MSTM control bits to enamble MST*/
s3_handle_mst(ddev, false);
/*
* early enable HPD Rx IRQ, should be done before set mode as short
* pulse interrupts are used for MST
*/
amdgpu_dm_irq_resume_early(adev);
/* Do detection*/
list_for_each_entry(connector, &ddev->mode_config.connector_list, head) {
aconnector = to_amdgpu_dm_connector(connector);
/*
* this is the case when traversing through already created
* MST connectors, should be skipped
*/
if (aconnector->mst_port)
continue;
mutex_lock(&aconnector->hpd_lock);
if (!dc_link_detect_sink(aconnector->dc_link, &new_connection_type))
DRM_ERROR("KMS: Failed to detect connector\n");
if (aconnector->base.force && new_connection_type == dc_connection_none)
emulated_link_detect(aconnector->dc_link);
else
dc_link_detect(aconnector->dc_link, DETECT_REASON_HPD);
if (aconnector->fake_enable && aconnector->dc_link->local_sink)
aconnector->fake_enable = false;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (aconnector->dc_sink)
dc_sink_release(aconnector->dc_sink);
aconnector->dc_sink = NULL;
amdgpu_dm_update_connector_after_detect(aconnector);
mutex_unlock(&aconnector->hpd_lock);
}
/* Force mode set in atomic commit */
for_each_new_crtc_in_state(dm->cached_state, crtc, new_crtc_state, i)
new_crtc_state->active_changed = true;
/*
* atomic_check is expected to create the dc states. We need to release
* them here, since they were duplicated as part of the suspend
* procedure.
*/
for_each_new_crtc_in_state(dm->cached_state, crtc, new_crtc_state, i) {
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
if (dm_new_crtc_state->stream) {
WARN_ON(kref_read(&dm_new_crtc_state->stream->refcount) > 1);
dc_stream_release(dm_new_crtc_state->stream);
dm_new_crtc_state->stream = NULL;
}
}
for_each_new_plane_in_state(dm->cached_state, plane, new_plane_state, i) {
dm_new_plane_state = to_dm_plane_state(new_plane_state);
if (dm_new_plane_state->dc_state) {
WARN_ON(kref_read(&dm_new_plane_state->dc_state->refcount) > 1);
dc_plane_state_release(dm_new_plane_state->dc_state);
dm_new_plane_state->dc_state = NULL;
}
}
drm/amdgpu: Don't fail resume process if resuming atomic state fails This is an ugly one unfortunately. Currently, all DRM drivers supporting atomic modesetting will save the state that userspace had set before suspending, then attempt to restore that state on resume. This probably worked very well at one point, like many other things, until DP MST came into the picture. While it's easy to restore state on normal display connectors that were disconnected during suspend regardless of their state post-resume, this can't really be done with MST because of the fact that setting up a downstream sink requires performing sideband transactions between the source and the MST hub, sending out the ACT packets, etc. Because of this, there isn't really a guarantee that we can restore the atomic state we had before suspend once we've resumed. This sucks pretty bad, but so far I haven't run into any compositors that this actually causes serious issues with. Most compositors will notice the hotplug we send afterwards, and then reprobe state. Since nouveau and i915 also don't fail the suspend/resume process due to failing to restore the atomic state, let's make amdgpu match this behavior. Better to resume the GPU properly, then to stop the process half way because of a potentially unavoidable atomic commit failure. Eventually, we'll have a real fix for this problem on the DRM level. But we've got some more important low-hanging fruit to deal with first. Signed-off-by: Lyude Paul <lyude@redhat.com> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Cc: Jerry Zuo <Jerry.Zuo@amd.com> Cc: <stable@vger.kernel.org> # v4.15+ Link: https://patchwork.freedesktop.org/patch/msgid/20190108211133.32564-3-lyude@redhat.com
2019-01-08 14:11:28 -07:00
drm_atomic_helper_resume(ddev, dm->cached_state);
dm->cached_state = NULL;
amdgpu_dm_irq_resume_late(adev);
drm/amdgpu: Don't fail resume process if resuming atomic state fails This is an ugly one unfortunately. Currently, all DRM drivers supporting atomic modesetting will save the state that userspace had set before suspending, then attempt to restore that state on resume. This probably worked very well at one point, like many other things, until DP MST came into the picture. While it's easy to restore state on normal display connectors that were disconnected during suspend regardless of their state post-resume, this can't really be done with MST because of the fact that setting up a downstream sink requires performing sideband transactions between the source and the MST hub, sending out the ACT packets, etc. Because of this, there isn't really a guarantee that we can restore the atomic state we had before suspend once we've resumed. This sucks pretty bad, but so far I haven't run into any compositors that this actually causes serious issues with. Most compositors will notice the hotplug we send afterwards, and then reprobe state. Since nouveau and i915 also don't fail the suspend/resume process due to failing to restore the atomic state, let's make amdgpu match this behavior. Better to resume the GPU properly, then to stop the process half way because of a potentially unavoidable atomic commit failure. Eventually, we'll have a real fix for this problem on the DRM level. But we've got some more important low-hanging fruit to deal with first. Signed-off-by: Lyude Paul <lyude@redhat.com> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Cc: Jerry Zuo <Jerry.Zuo@amd.com> Cc: <stable@vger.kernel.org> # v4.15+ Link: https://patchwork.freedesktop.org/patch/msgid/20190108211133.32564-3-lyude@redhat.com
2019-01-08 14:11:28 -07:00
return 0;
}
/**
* DOC: DM Lifecycle
*
* DM (and consequently DC) is registered in the amdgpu base driver as a IP
* block. When CONFIG_DRM_AMD_DC is enabled, the DM device IP block is added to
* the base driver's device list to be initialized and torn down accordingly.
*
* The functions to do so are provided as hooks in &struct amd_ip_funcs.
*/
static const struct amd_ip_funcs amdgpu_dm_funcs = {
.name = "dm",
.early_init = dm_early_init,
.late_init = dm_late_init,
.sw_init = dm_sw_init,
.sw_fini = dm_sw_fini,
.hw_init = dm_hw_init,
.hw_fini = dm_hw_fini,
.suspend = dm_suspend,
.resume = dm_resume,
.is_idle = dm_is_idle,
.wait_for_idle = dm_wait_for_idle,
.check_soft_reset = dm_check_soft_reset,
.soft_reset = dm_soft_reset,
.set_clockgating_state = dm_set_clockgating_state,
.set_powergating_state = dm_set_powergating_state,
};
const struct amdgpu_ip_block_version dm_ip_block =
{
.type = AMD_IP_BLOCK_TYPE_DCE,
.major = 1,
.minor = 0,
.rev = 0,
.funcs = &amdgpu_dm_funcs,
};
/**
* DOC: atomic
*
* *WIP*
*/
static const struct drm_mode_config_funcs amdgpu_dm_mode_funcs = {
.fb_create = amdgpu_display_user_framebuffer_create,
.output_poll_changed = drm_fb_helper_output_poll_changed,
.atomic_check = amdgpu_dm_atomic_check,
.atomic_commit = amdgpu_dm_atomic_commit,
};
static struct drm_mode_config_helper_funcs amdgpu_dm_mode_config_helperfuncs = {
.atomic_commit_tail = amdgpu_dm_atomic_commit_tail
};
static void
amdgpu_dm_update_connector_after_detect(struct amdgpu_dm_connector *aconnector)
{
struct drm_connector *connector = &aconnector->base;
struct drm_device *dev = connector->dev;
struct dc_sink *sink;
/* MST handled by drm_mst framework */
if (aconnector->mst_mgr.mst_state == true)
return;
sink = aconnector->dc_link->local_sink;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (sink)
dc_sink_retain(sink);
/*
* Edid mgmt connector gets first update only in mode_valid hook and then
* the connector sink is set to either fake or physical sink depends on link status.
* Skip if already done during boot.
*/
if (aconnector->base.force != DRM_FORCE_UNSPECIFIED
&& aconnector->dc_em_sink) {
/*
* For S3 resume with headless use eml_sink to fake stream
* because on resume connector->sink is set to NULL
*/
mutex_lock(&dev->mode_config.mutex);
if (sink) {
if (aconnector->dc_sink) {
amdgpu_dm_update_freesync_caps(connector, NULL);
/*
* retain and release below are used to
* bump up refcount for sink because the link doesn't point
* to it anymore after disconnect, so on next crtc to connector
* reshuffle by UMD we will get into unwanted dc_sink release
*/
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
dc_sink_release(aconnector->dc_sink);
}
aconnector->dc_sink = sink;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
dc_sink_retain(aconnector->dc_sink);
amdgpu_dm_update_freesync_caps(connector,
aconnector->edid);
} else {
amdgpu_dm_update_freesync_caps(connector, NULL);
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (!aconnector->dc_sink) {
aconnector->dc_sink = aconnector->dc_em_sink;
dc_sink_retain(aconnector->dc_sink);
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
}
}
mutex_unlock(&dev->mode_config.mutex);
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (sink)
dc_sink_release(sink);
return;
}
/*
* TODO: temporary guard to look for proper fix
* if this sink is MST sink, we should not do anything
*/
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (sink && sink->sink_signal == SIGNAL_TYPE_DISPLAY_PORT_MST) {
dc_sink_release(sink);
return;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
}
if (aconnector->dc_sink == sink) {
/*
* We got a DP short pulse (Link Loss, DP CTS, etc...).
* Do nothing!!
*/
DRM_DEBUG_DRIVER("DCHPD: connector_id=%d: dc_sink didn't change.\n",
aconnector->connector_id);
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (sink)
dc_sink_release(sink);
return;
}
DRM_DEBUG_DRIVER("DCHPD: connector_id=%d: Old sink=%p New sink=%p\n",
aconnector->connector_id, aconnector->dc_sink, sink);
mutex_lock(&dev->mode_config.mutex);
/*
* 1. Update status of the drm connector
* 2. Send an event and let userspace tell us what to do
*/
if (sink) {
/*
* TODO: check if we still need the S3 mode update workaround.
* If yes, put it here.
*/
if (aconnector->dc_sink)
amdgpu_dm_update_freesync_caps(connector, NULL);
aconnector->dc_sink = sink;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
dc_sink_retain(aconnector->dc_sink);
if (sink->dc_edid.length == 0) {
aconnector->edid = NULL;
if (aconnector->dc_link->aux_mode) {
drm_dp_cec_unset_edid(
&aconnector->dm_dp_aux.aux);
}
} else {
aconnector->edid =
(struct edid *)sink->dc_edid.raw_edid;
drm_connector_update_edid_property(connector,
aconnector->edid);
drm_add_edid_modes(connector, aconnector->edid);
if (aconnector->dc_link->aux_mode)
drm_dp_cec_set_edid(&aconnector->dm_dp_aux.aux,
aconnector->edid);
}
amdgpu_dm_update_freesync_caps(connector, aconnector->edid);
} else {
drm_dp_cec_unset_edid(&aconnector->dm_dp_aux.aux);
amdgpu_dm_update_freesync_caps(connector, NULL);
drm_connector_update_edid_property(connector, NULL);
aconnector->num_modes = 0;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
dc_sink_release(aconnector->dc_sink);
aconnector->dc_sink = NULL;
aconnector->edid = NULL;
}
mutex_unlock(&dev->mode_config.mutex);
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (sink)
dc_sink_release(sink);
}
static void handle_hpd_irq(void *param)
{
struct amdgpu_dm_connector *aconnector = (struct amdgpu_dm_connector *)param;
struct drm_connector *connector = &aconnector->base;
struct drm_device *dev = connector->dev;
enum dc_connection_type new_connection_type = dc_connection_none;
/*
* In case of failure or MST no need to update connector status or notify the OS
* since (for MST case) MST does this in its own context.
*/
mutex_lock(&aconnector->hpd_lock);
if (aconnector->fake_enable)
aconnector->fake_enable = false;
if (!dc_link_detect_sink(aconnector->dc_link, &new_connection_type))
DRM_ERROR("KMS: Failed to detect connector\n");
if (aconnector->base.force && new_connection_type == dc_connection_none) {
emulated_link_detect(aconnector->dc_link);
drm_modeset_lock_all(dev);
dm_restore_drm_connector_state(dev, connector);
drm_modeset_unlock_all(dev);
if (aconnector->base.force == DRM_FORCE_UNSPECIFIED)
drm_kms_helper_hotplug_event(dev);
} else if (dc_link_detect(aconnector->dc_link, DETECT_REASON_HPD)) {
amdgpu_dm_update_connector_after_detect(aconnector);
drm_modeset_lock_all(dev);
dm_restore_drm_connector_state(dev, connector);
drm_modeset_unlock_all(dev);
if (aconnector->base.force == DRM_FORCE_UNSPECIFIED)
drm_kms_helper_hotplug_event(dev);
}
mutex_unlock(&aconnector->hpd_lock);
}
static void dm_handle_hpd_rx_irq(struct amdgpu_dm_connector *aconnector)
{
uint8_t esi[DP_PSR_ERROR_STATUS - DP_SINK_COUNT_ESI] = { 0 };
uint8_t dret;
bool new_irq_handled = false;
int dpcd_addr;
int dpcd_bytes_to_read;
const int max_process_count = 30;
int process_count = 0;
const struct dc_link_status *link_status = dc_link_get_status(aconnector->dc_link);
if (link_status->dpcd_caps->dpcd_rev.raw < 0x12) {
dpcd_bytes_to_read = DP_LANE0_1_STATUS - DP_SINK_COUNT;
/* DPCD 0x200 - 0x201 for downstream IRQ */
dpcd_addr = DP_SINK_COUNT;
} else {
dpcd_bytes_to_read = DP_PSR_ERROR_STATUS - DP_SINK_COUNT_ESI;
/* DPCD 0x2002 - 0x2005 for downstream IRQ */
dpcd_addr = DP_SINK_COUNT_ESI;
}
dret = drm_dp_dpcd_read(
&aconnector->dm_dp_aux.aux,
dpcd_addr,
esi,
dpcd_bytes_to_read);
while (dret == dpcd_bytes_to_read &&
process_count < max_process_count) {
uint8_t retry;
dret = 0;
process_count++;
DRM_DEBUG_DRIVER("ESI %02x %02x %02x\n", esi[0], esi[1], esi[2]);
/* handle HPD short pulse irq */
if (aconnector->mst_mgr.mst_state)
drm_dp_mst_hpd_irq(
&aconnector->mst_mgr,
esi,
&new_irq_handled);
if (new_irq_handled) {
/* ACK at DPCD to notify down stream */
const int ack_dpcd_bytes_to_write =
dpcd_bytes_to_read - 1;
for (retry = 0; retry < 3; retry++) {
uint8_t wret;
wret = drm_dp_dpcd_write(
&aconnector->dm_dp_aux.aux,
dpcd_addr + 1,
&esi[1],
ack_dpcd_bytes_to_write);
if (wret == ack_dpcd_bytes_to_write)
break;
}
/* check if there is new irq to be handled */
dret = drm_dp_dpcd_read(
&aconnector->dm_dp_aux.aux,
dpcd_addr,
esi,
dpcd_bytes_to_read);
new_irq_handled = false;
} else {
break;
}
}
if (process_count == max_process_count)
DRM_DEBUG_DRIVER("Loop exceeded max iterations\n");
}
static void handle_hpd_rx_irq(void *param)
{
struct amdgpu_dm_connector *aconnector = (struct amdgpu_dm_connector *)param;
struct drm_connector *connector = &aconnector->base;
struct drm_device *dev = connector->dev;
struct dc_link *dc_link = aconnector->dc_link;
bool is_mst_root_connector = aconnector->mst_mgr.mst_state;
enum dc_connection_type new_connection_type = dc_connection_none;
/*
* TODO:Temporary add mutex to protect hpd interrupt not have a gpio
* conflict, after implement i2c helper, this mutex should be
* retired.
*/
if (dc_link->type != dc_connection_mst_branch)
mutex_lock(&aconnector->hpd_lock);
if (dc_link_handle_hpd_rx_irq(dc_link, NULL, NULL) &&
!is_mst_root_connector) {
/* Downstream Port status changed. */
if (!dc_link_detect_sink(dc_link, &new_connection_type))
DRM_ERROR("KMS: Failed to detect connector\n");
if (aconnector->base.force && new_connection_type == dc_connection_none) {
emulated_link_detect(dc_link);
if (aconnector->fake_enable)
aconnector->fake_enable = false;
amdgpu_dm_update_connector_after_detect(aconnector);
drm_modeset_lock_all(dev);
dm_restore_drm_connector_state(dev, connector);
drm_modeset_unlock_all(dev);
drm_kms_helper_hotplug_event(dev);
} else if (dc_link_detect(dc_link, DETECT_REASON_HPDRX)) {
if (aconnector->fake_enable)
aconnector->fake_enable = false;
amdgpu_dm_update_connector_after_detect(aconnector);
drm_modeset_lock_all(dev);
dm_restore_drm_connector_state(dev, connector);
drm_modeset_unlock_all(dev);
drm_kms_helper_hotplug_event(dev);
}
}
if ((dc_link->cur_link_settings.lane_count != LANE_COUNT_UNKNOWN) ||
(dc_link->type == dc_connection_mst_branch))
dm_handle_hpd_rx_irq(aconnector);
if (dc_link->type != dc_connection_mst_branch) {
drm_dp_cec_irq(&aconnector->dm_dp_aux.aux);
mutex_unlock(&aconnector->hpd_lock);
}
}
static void register_hpd_handlers(struct amdgpu_device *adev)
{
struct drm_device *dev = adev->ddev;
struct drm_connector *connector;
struct amdgpu_dm_connector *aconnector;
const struct dc_link *dc_link;
struct dc_interrupt_params int_params = {0};
int_params.requested_polarity = INTERRUPT_POLARITY_DEFAULT;
int_params.current_polarity = INTERRUPT_POLARITY_DEFAULT;
list_for_each_entry(connector,
&dev->mode_config.connector_list, head) {
aconnector = to_amdgpu_dm_connector(connector);
dc_link = aconnector->dc_link;
if (DC_IRQ_SOURCE_INVALID != dc_link->irq_source_hpd) {
int_params.int_context = INTERRUPT_LOW_IRQ_CONTEXT;
int_params.irq_source = dc_link->irq_source_hpd;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
handle_hpd_irq,
(void *) aconnector);
}
if (DC_IRQ_SOURCE_INVALID != dc_link->irq_source_hpd_rx) {
/* Also register for DP short pulse (hpd_rx). */
int_params.int_context = INTERRUPT_LOW_IRQ_CONTEXT;
int_params.irq_source = dc_link->irq_source_hpd_rx;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
handle_hpd_rx_irq,
(void *) aconnector);
}
}
}
/* Register IRQ sources and initialize IRQ callbacks */
static int dce110_register_irq_handlers(struct amdgpu_device *adev)
{
struct dc *dc = adev->dm.dc;
struct common_irq_params *c_irq_params;
struct dc_interrupt_params int_params = {0};
int r;
int i;
unsigned client_id = AMDGPU_IRQ_CLIENTID_LEGACY;
if (adev->asic_type >= CHIP_VEGA10)
client_id = SOC15_IH_CLIENTID_DCE;
int_params.requested_polarity = INTERRUPT_POLARITY_DEFAULT;
int_params.current_polarity = INTERRUPT_POLARITY_DEFAULT;
/*
* Actions of amdgpu_irq_add_id():
* 1. Register a set() function with base driver.
* Base driver will call set() function to enable/disable an
* interrupt in DC hardware.
* 2. Register amdgpu_dm_irq_handler().
* Base driver will call amdgpu_dm_irq_handler() for ALL interrupts
* coming from DC hardware.
* amdgpu_dm_irq_handler() will re-direct the interrupt to DC
* for acknowledging and handling. */
/* Use VBLANK interrupt */
for (i = VISLANDS30_IV_SRCID_D1_VERTICAL_INTERRUPT0; i <= VISLANDS30_IV_SRCID_D6_VERTICAL_INTERRUPT0; i++) {
r = amdgpu_irq_add_id(adev, client_id, i, &adev->crtc_irq);
if (r) {
DRM_ERROR("Failed to add crtc irq id!\n");
return r;
}
int_params.int_context = INTERRUPT_HIGH_IRQ_CONTEXT;
int_params.irq_source =
dc_interrupt_to_irq_source(dc, i, 0);
c_irq_params = &adev->dm.vblank_params[int_params.irq_source - DC_IRQ_SOURCE_VBLANK1];
c_irq_params->adev = adev;
c_irq_params->irq_src = int_params.irq_source;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
dm_crtc_high_irq, c_irq_params);
}
/* Use VUPDATE interrupt */
for (i = VISLANDS30_IV_SRCID_D1_V_UPDATE_INT; i <= VISLANDS30_IV_SRCID_D6_V_UPDATE_INT; i += 2) {
r = amdgpu_irq_add_id(adev, client_id, i, &adev->vupdate_irq);
if (r) {
DRM_ERROR("Failed to add vupdate irq id!\n");
return r;
}
int_params.int_context = INTERRUPT_HIGH_IRQ_CONTEXT;
int_params.irq_source =
dc_interrupt_to_irq_source(dc, i, 0);
c_irq_params = &adev->dm.vupdate_params[int_params.irq_source - DC_IRQ_SOURCE_VUPDATE1];
c_irq_params->adev = adev;
c_irq_params->irq_src = int_params.irq_source;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
dm_vupdate_high_irq, c_irq_params);
}
/* Use GRPH_PFLIP interrupt */
for (i = VISLANDS30_IV_SRCID_D1_GRPH_PFLIP;
i <= VISLANDS30_IV_SRCID_D6_GRPH_PFLIP; i += 2) {
r = amdgpu_irq_add_id(adev, client_id, i, &adev->pageflip_irq);
if (r) {
DRM_ERROR("Failed to add page flip irq id!\n");
return r;
}
int_params.int_context = INTERRUPT_HIGH_IRQ_CONTEXT;
int_params.irq_source =
dc_interrupt_to_irq_source(dc, i, 0);
c_irq_params = &adev->dm.pflip_params[int_params.irq_source - DC_IRQ_SOURCE_PFLIP_FIRST];
c_irq_params->adev = adev;
c_irq_params->irq_src = int_params.irq_source;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
dm_pflip_high_irq, c_irq_params);
}
/* HPD */
r = amdgpu_irq_add_id(adev, client_id,
VISLANDS30_IV_SRCID_HOTPLUG_DETECT_A, &adev->hpd_irq);
if (r) {
DRM_ERROR("Failed to add hpd irq id!\n");
return r;
}
register_hpd_handlers(adev);
return 0;
}
#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
/* Register IRQ sources and initialize IRQ callbacks */
static int dcn10_register_irq_handlers(struct amdgpu_device *adev)
{
struct dc *dc = adev->dm.dc;
struct common_irq_params *c_irq_params;
struct dc_interrupt_params int_params = {0};
int r;
int i;
int_params.requested_polarity = INTERRUPT_POLARITY_DEFAULT;
int_params.current_polarity = INTERRUPT_POLARITY_DEFAULT;
/*
* Actions of amdgpu_irq_add_id():
* 1. Register a set() function with base driver.
* Base driver will call set() function to enable/disable an
* interrupt in DC hardware.
* 2. Register amdgpu_dm_irq_handler().
* Base driver will call amdgpu_dm_irq_handler() for ALL interrupts
* coming from DC hardware.
* amdgpu_dm_irq_handler() will re-direct the interrupt to DC
* for acknowledging and handling.
*/
/* Use VSTARTUP interrupt */
for (i = DCN_1_0__SRCID__DC_D1_OTG_VSTARTUP;
i <= DCN_1_0__SRCID__DC_D1_OTG_VSTARTUP + adev->mode_info.num_crtc - 1;
i++) {
r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DCE, i, &adev->crtc_irq);
if (r) {
DRM_ERROR("Failed to add crtc irq id!\n");
return r;
}
int_params.int_context = INTERRUPT_HIGH_IRQ_CONTEXT;
int_params.irq_source =
dc_interrupt_to_irq_source(dc, i, 0);
c_irq_params = &adev->dm.vblank_params[int_params.irq_source - DC_IRQ_SOURCE_VBLANK1];
c_irq_params->adev = adev;
c_irq_params->irq_src = int_params.irq_source;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
dm_crtc_high_irq, c_irq_params);
}
/* Use VUPDATE_NO_LOCK interrupt on DCN, which seems to correspond to
* the regular VUPDATE interrupt on DCE. We want DC_IRQ_SOURCE_VUPDATEx
* to trigger at end of each vblank, regardless of state of the lock,
* matching DCE behaviour.
*/
for (i = DCN_1_0__SRCID__OTG0_IHC_V_UPDATE_NO_LOCK_INTERRUPT;
i <= DCN_1_0__SRCID__OTG0_IHC_V_UPDATE_NO_LOCK_INTERRUPT + adev->mode_info.num_crtc - 1;
i++) {
r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DCE, i, &adev->vupdate_irq);
if (r) {
DRM_ERROR("Failed to add vupdate irq id!\n");
return r;
}
int_params.int_context = INTERRUPT_HIGH_IRQ_CONTEXT;
int_params.irq_source =
dc_interrupt_to_irq_source(dc, i, 0);
c_irq_params = &adev->dm.vupdate_params[int_params.irq_source - DC_IRQ_SOURCE_VUPDATE1];
c_irq_params->adev = adev;
c_irq_params->irq_src = int_params.irq_source;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
dm_vupdate_high_irq, c_irq_params);
}
/* Use GRPH_PFLIP interrupt */
for (i = DCN_1_0__SRCID__HUBP0_FLIP_INTERRUPT;
i <= DCN_1_0__SRCID__HUBP0_FLIP_INTERRUPT + adev->mode_info.num_crtc - 1;
i++) {
r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DCE, i, &adev->pageflip_irq);
if (r) {
DRM_ERROR("Failed to add page flip irq id!\n");
return r;
}
int_params.int_context = INTERRUPT_HIGH_IRQ_CONTEXT;
int_params.irq_source =
dc_interrupt_to_irq_source(dc, i, 0);
c_irq_params = &adev->dm.pflip_params[int_params.irq_source - DC_IRQ_SOURCE_PFLIP_FIRST];
c_irq_params->adev = adev;
c_irq_params->irq_src = int_params.irq_source;
amdgpu_dm_irq_register_interrupt(adev, &int_params,
dm_pflip_high_irq, c_irq_params);
}
/* HPD */
r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DCE, DCN_1_0__SRCID__DC_HPD1_INT,
&adev->hpd_irq);
if (r) {
DRM_ERROR("Failed to add hpd irq id!\n");
return r;
}
register_hpd_handlers(adev);
return 0;
}
#endif
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
/*
* Acquires the lock for the atomic state object and returns
* the new atomic state.
*
* This should only be called during atomic check.
*/
static int dm_atomic_get_state(struct drm_atomic_state *state,
struct dm_atomic_state **dm_state)
{
struct drm_device *dev = state->dev;
struct amdgpu_device *adev = dev->dev_private;
struct amdgpu_display_manager *dm = &adev->dm;
struct drm_private_state *priv_state;
if (*dm_state)
return 0;
priv_state = drm_atomic_get_private_obj_state(state, &dm->atomic_obj);
if (IS_ERR(priv_state))
return PTR_ERR(priv_state);
*dm_state = to_dm_atomic_state(priv_state);
return 0;
}
struct dm_atomic_state *
dm_atomic_get_new_state(struct drm_atomic_state *state)
{
struct drm_device *dev = state->dev;
struct amdgpu_device *adev = dev->dev_private;
struct amdgpu_display_manager *dm = &adev->dm;
struct drm_private_obj *obj;
struct drm_private_state *new_obj_state;
int i;
for_each_new_private_obj_in_state(state, obj, new_obj_state, i) {
if (obj->funcs == dm->atomic_obj.funcs)
return to_dm_atomic_state(new_obj_state);
}
return NULL;
}
struct dm_atomic_state *
dm_atomic_get_old_state(struct drm_atomic_state *state)
{
struct drm_device *dev = state->dev;
struct amdgpu_device *adev = dev->dev_private;
struct amdgpu_display_manager *dm = &adev->dm;
struct drm_private_obj *obj;
struct drm_private_state *old_obj_state;
int i;
for_each_old_private_obj_in_state(state, obj, old_obj_state, i) {
if (obj->funcs == dm->atomic_obj.funcs)
return to_dm_atomic_state(old_obj_state);
}
return NULL;
}
static struct drm_private_state *
dm_atomic_duplicate_state(struct drm_private_obj *obj)
{
struct dm_atomic_state *old_state, *new_state;
new_state = kzalloc(sizeof(*new_state), GFP_KERNEL);
if (!new_state)
return NULL;
__drm_atomic_helper_private_obj_duplicate_state(obj, &new_state->base);
old_state = to_dm_atomic_state(obj->state);
if (old_state && old_state->context)
new_state->context = dc_copy_state(old_state->context);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (!new_state->context) {
kfree(new_state);
return NULL;
}
return &new_state->base;
}
static void dm_atomic_destroy_state(struct drm_private_obj *obj,
struct drm_private_state *state)
{
struct dm_atomic_state *dm_state = to_dm_atomic_state(state);
if (dm_state && dm_state->context)
dc_release_state(dm_state->context);
kfree(dm_state);
}
static struct drm_private_state_funcs dm_atomic_state_funcs = {
.atomic_duplicate_state = dm_atomic_duplicate_state,
.atomic_destroy_state = dm_atomic_destroy_state,
};
static int amdgpu_dm_mode_config_init(struct amdgpu_device *adev)
{
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct dm_atomic_state *state;
int r;
adev->mode_info.mode_config_initialized = true;
adev->ddev->mode_config.funcs = (void *)&amdgpu_dm_mode_funcs;
adev->ddev->mode_config.helper_private = &amdgpu_dm_mode_config_helperfuncs;
adev->ddev->mode_config.max_width = 16384;
adev->ddev->mode_config.max_height = 16384;
adev->ddev->mode_config.preferred_depth = 24;
adev->ddev->mode_config.prefer_shadow = 1;
/* indicates support for immediate flip */
adev->ddev->mode_config.async_page_flip = true;
adev->ddev->mode_config.fb_base = adev->gmc.aper_base;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
state = kzalloc(sizeof(*state), GFP_KERNEL);
if (!state)
return -ENOMEM;
state->context = dc_create_state(adev->dm.dc);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (!state->context) {
kfree(state);
return -ENOMEM;
}
dc_resource_state_copy_construct_current(adev->dm.dc, state->context);
drm-misc-next for 5.1: UAPI Changes: Cross-subsystem Changes: - Turn dma-buf fence sequence numbers into 64 bit numbers Core Changes: - Move to a common helper for the DP MST hotplug for radeon, i915 and amdgpu - i2c improvements for drm_dp_mst - Removal of drm_syncobj_cb - Introduction of an helper to create and attach the TV margin properties Driver Changes: - Improve cache flushes for v3d - Reflection support for vc4 - HDMI overscan support for vc4 - Add implicit fencing support for rockchip and sun4i - Switch to generic fbdev emulation for virtio -----BEGIN PGP SIGNATURE----- iHUEABYIAB0WIQRcEzekXsqa64kGDp7j7w1vZxhRxQUCXDOTqAAKCRDj7w1vZxhR xZ8QAQD4j8m9Ea3bzY5Rr8BYUx1k+Cjj6Y6abZmot2rSvdyOHwD+JzJFIFAPZjdd uOKhLnDlubaaoa6OGPDQShjl9p3gyQE= =WQGO -----END PGP SIGNATURE----- Merge tag 'drm-misc-next-2019-01-07-1' of git://anongit.freedesktop.org/drm/drm-misc into drm-next drm-misc-next for 5.1: UAPI Changes: Cross-subsystem Changes: - Turn dma-buf fence sequence numbers into 64 bit numbers Core Changes: - Move to a common helper for the DP MST hotplug for radeon, i915 and amdgpu - i2c improvements for drm_dp_mst - Removal of drm_syncobj_cb - Introduction of an helper to create and attach the TV margin properties Driver Changes: - Improve cache flushes for v3d - Reflection support for vc4 - HDMI overscan support for vc4 - Add implicit fencing support for rockchip and sun4i - Switch to generic fbdev emulation for virtio Signed-off-by: Dave Airlie <airlied@redhat.com> [airlied: applied amdgpu merge fixup] From: Maxime Ripard <maxime.ripard@bootlin.com> Link: https://patchwork.freedesktop.org/patch/msgid/20190107180333.amklwycudbsub3s5@flea
2019-01-09 12:53:51 -07:00
drm_atomic_private_obj_init(adev->ddev,
&adev->dm.atomic_obj,
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
&state->base,
&dm_atomic_state_funcs);
r = amdgpu_display_modeset_create_props(adev);
if (r) {
dc_release_state(state->context);
kfree(state);
return r;
}
r = amdgpu_dm_audio_init(adev);
if (r) {
dc_release_state(state->context);
kfree(state);
return r;
}
return 0;
}
#define AMDGPU_DM_DEFAULT_MIN_BACKLIGHT 12
#define AMDGPU_DM_DEFAULT_MAX_BACKLIGHT 255
#if defined(CONFIG_BACKLIGHT_CLASS_DEVICE) ||\
defined(CONFIG_BACKLIGHT_CLASS_DEVICE_MODULE)
static void amdgpu_dm_update_backlight_caps(struct amdgpu_display_manager *dm)
{
#if defined(CONFIG_ACPI)
struct amdgpu_dm_backlight_caps caps;
memset(&caps, 0, sizeof(caps));
if (dm->backlight_caps.caps_valid)
return;
amdgpu_acpi_get_backlight_caps(dm->adev, &caps);
if (caps.caps_valid) {
dm->backlight_caps.min_input_signal = caps.min_input_signal;
dm->backlight_caps.max_input_signal = caps.max_input_signal;
dm->backlight_caps.caps_valid = true;
} else {
dm->backlight_caps.min_input_signal =
AMDGPU_DM_DEFAULT_MIN_BACKLIGHT;
dm->backlight_caps.max_input_signal =
AMDGPU_DM_DEFAULT_MAX_BACKLIGHT;
}
#else
dm->backlight_caps.min_input_signal = AMDGPU_DM_DEFAULT_MIN_BACKLIGHT;
dm->backlight_caps.max_input_signal = AMDGPU_DM_DEFAULT_MAX_BACKLIGHT;
#endif
}
static int amdgpu_dm_backlight_update_status(struct backlight_device *bd)
{
struct amdgpu_display_manager *dm = bl_get_data(bd);
struct amdgpu_dm_backlight_caps caps;
uint32_t brightness = bd->props.brightness;
amdgpu_dm_update_backlight_caps(dm);
caps = dm->backlight_caps;
/*
* The brightness input is in the range 0-255
* It needs to be rescaled to be between the
* requested min and max input signal
*
* It also needs to be scaled up by 0x101 to
* match the DC interface which has a range of
* 0 to 0xffff
*/
brightness =
brightness
* 0x101
* (caps.max_input_signal - caps.min_input_signal)
/ AMDGPU_MAX_BL_LEVEL
+ caps.min_input_signal * 0x101;
if (dc_link_set_backlight_level(dm->backlight_link,
brightness, 0))
return 0;
else
return 1;
}
static int amdgpu_dm_backlight_get_brightness(struct backlight_device *bd)
{
struct amdgpu_display_manager *dm = bl_get_data(bd);
int ret = dc_link_get_backlight_level(dm->backlight_link);
if (ret == DC_ERROR_UNEXPECTED)
return bd->props.brightness;
return ret;
}
static const struct backlight_ops amdgpu_dm_backlight_ops = {
.options = BL_CORE_SUSPENDRESUME,
.get_brightness = amdgpu_dm_backlight_get_brightness,
.update_status = amdgpu_dm_backlight_update_status,
};
static void
amdgpu_dm_register_backlight_device(struct amdgpu_display_manager *dm)
{
char bl_name[16];
struct backlight_properties props = { 0 };
amdgpu_dm_update_backlight_caps(dm);
props.max_brightness = AMDGPU_MAX_BL_LEVEL;
props.brightness = AMDGPU_MAX_BL_LEVEL;
props.type = BACKLIGHT_RAW;
snprintf(bl_name, sizeof(bl_name), "amdgpu_bl%d",
dm->adev->ddev->primary->index);
dm->backlight_dev = backlight_device_register(bl_name,
dm->adev->ddev->dev,
dm,
&amdgpu_dm_backlight_ops,
&props);
if (IS_ERR(dm->backlight_dev))
DRM_ERROR("DM: Backlight registration failed!\n");
else
DRM_DEBUG_DRIVER("DM: Registered Backlight device: %s\n", bl_name);
}
#endif
static int initialize_plane(struct amdgpu_display_manager *dm,
2019-01-18 11:42:34 -07:00
struct amdgpu_mode_info *mode_info, int plane_id,
enum drm_plane_type plane_type,
const struct dc_plane_cap *plane_cap)
{
struct drm_plane *plane;
unsigned long possible_crtcs;
int ret = 0;
plane = kzalloc(sizeof(struct drm_plane), GFP_KERNEL);
if (!plane) {
DRM_ERROR("KMS: Failed to allocate plane\n");
return -ENOMEM;
}
2019-01-18 11:42:34 -07:00
plane->type = plane_type;
/*
2019-01-18 11:42:34 -07:00
* HACK: IGT tests expect that the primary plane for a CRTC
* can only have one possible CRTC. Only expose support for
* any CRTC if they're not going to be used as a primary plane
* for a CRTC - like overlay or underlay planes.
*/
possible_crtcs = 1 << plane_id;
if (plane_id >= dm->dc->caps.max_streams)
possible_crtcs = 0xff;
ret = amdgpu_dm_plane_init(dm, plane, possible_crtcs, plane_cap);
if (ret) {
DRM_ERROR("KMS: Failed to initialize plane\n");
kfree(plane);
return ret;
}
if (mode_info)
mode_info->planes[plane_id] = plane;
return ret;
}
static void register_backlight_device(struct amdgpu_display_manager *dm,
struct dc_link *link)
{
#if defined(CONFIG_BACKLIGHT_CLASS_DEVICE) ||\
defined(CONFIG_BACKLIGHT_CLASS_DEVICE_MODULE)
if ((link->connector_signal & (SIGNAL_TYPE_EDP | SIGNAL_TYPE_LVDS)) &&
link->type != dc_connection_none) {
/*
* Event if registration failed, we should continue with
* DM initialization because not having a backlight control
* is better then a black screen.
*/
amdgpu_dm_register_backlight_device(dm);
if (dm->backlight_dev)
dm->backlight_link = link;
}
#endif
}
/*
* In this architecture, the association
* connector -> encoder -> crtc
* id not really requried. The crtc and connector will hold the
* display_index as an abstraction to use with DAL component
*
* Returns 0 on success
*/
static int amdgpu_dm_initialize_drm_device(struct amdgpu_device *adev)
{
struct amdgpu_display_manager *dm = &adev->dm;
int32_t i;
struct amdgpu_dm_connector *aconnector = NULL;
struct amdgpu_encoder *aencoder = NULL;
struct amdgpu_mode_info *mode_info = &adev->mode_info;
uint32_t link_cnt;
int32_t primary_planes;
enum dc_connection_type new_connection_type = dc_connection_none;
const struct dc_plane_cap *plane;
link_cnt = dm->dc->caps.max_links;
if (amdgpu_dm_mode_config_init(dm->adev)) {
DRM_ERROR("DM: Failed to initialize mode config\n");
return -EINVAL;
}
2019-01-18 11:42:34 -07:00
/* There is one primary plane per CRTC */
primary_planes = dm->dc->caps.max_streams;
ASSERT(primary_planes <= AMDGPU_MAX_PLANES);
2019-01-18 11:42:34 -07:00
/*
* Initialize primary planes, implicit planes for legacy IOCTLS.
* Order is reversed to match iteration order in atomic check.
*/
for (i = (primary_planes - 1); i >= 0; i--) {
plane = &dm->dc->caps.planes[i];
2019-01-18 11:42:34 -07:00
if (initialize_plane(dm, mode_info, i,
DRM_PLANE_TYPE_PRIMARY, plane)) {
DRM_ERROR("KMS: Failed to initialize primary plane\n");
goto fail;
}
}
/*
* Initialize overlay planes, index starting after primary planes.
* These planes have a higher DRM index than the primary planes since
* they should be considered as having a higher z-order.
* Order is reversed to match iteration order in atomic check.
*
* Only support DCN for now, and only expose one so we don't encourage
* userspace to use up all the pipes.
*/
for (i = 0; i < dm->dc->caps.max_planes; ++i) {
struct dc_plane_cap *plane = &dm->dc->caps.planes[i];
if (plane->type != DC_PLANE_TYPE_DCN_UNIVERSAL)
continue;
if (!plane->blends_with_above || !plane->blends_with_below)
continue;
if (!plane->pixel_format_support.argb8888)
continue;
if (initialize_plane(dm, NULL, primary_planes + i,
DRM_PLANE_TYPE_OVERLAY, plane)) {
DRM_ERROR("KMS: Failed to initialize overlay plane\n");
goto fail;
}
/* Only create one overlay plane. */
break;
}
for (i = 0; i < dm->dc->caps.max_streams; i++)
if (amdgpu_dm_crtc_init(dm, mode_info->planes[i], i)) {
DRM_ERROR("KMS: Failed to initialize crtc\n");
goto fail;
}
dm->display_indexes_num = dm->dc->caps.max_streams;
/* loops over all connectors on the board */
for (i = 0; i < link_cnt; i++) {
struct dc_link *link = NULL;
if (i > AMDGPU_DM_MAX_DISPLAY_INDEX) {
DRM_ERROR(
"KMS: Cannot support more than %d display indexes\n",
AMDGPU_DM_MAX_DISPLAY_INDEX);
continue;
}
aconnector = kzalloc(sizeof(*aconnector), GFP_KERNEL);
if (!aconnector)
goto fail;
aencoder = kzalloc(sizeof(*aencoder), GFP_KERNEL);
if (!aencoder)
goto fail;
if (amdgpu_dm_encoder_init(dm->ddev, aencoder, i)) {
DRM_ERROR("KMS: Failed to initialize encoder\n");
goto fail;
}
if (amdgpu_dm_connector_init(dm, aconnector, i, aencoder)) {
DRM_ERROR("KMS: Failed to initialize connector\n");
goto fail;
}
link = dc_get_link_at_index(dm->dc, i);
if (!dc_link_detect_sink(link, &new_connection_type))
DRM_ERROR("KMS: Failed to detect connector\n");
if (aconnector->base.force && new_connection_type == dc_connection_none) {
emulated_link_detect(link);
amdgpu_dm_update_connector_after_detect(aconnector);
} else if (dc_link_detect(link, DETECT_REASON_BOOT)) {
amdgpu_dm_update_connector_after_detect(aconnector);
register_backlight_device(dm, link);
}
}
/* Software is initialized. Now we can register interrupt handlers. */
switch (adev->asic_type) {
case CHIP_BONAIRE:
case CHIP_HAWAII:
case CHIP_KAVERI:
case CHIP_KABINI:
case CHIP_MULLINS:
case CHIP_TONGA:
case CHIP_FIJI:
case CHIP_CARRIZO:
case CHIP_STONEY:
case CHIP_POLARIS11:
case CHIP_POLARIS10:
case CHIP_POLARIS12:
case CHIP_VEGAM:
case CHIP_VEGA10:
case CHIP_VEGA12:
case CHIP_VEGA20:
if (dce110_register_irq_handlers(dm->adev)) {
DRM_ERROR("DM: Failed to initialize IRQ\n");
goto fail;
}
break;
#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
case CHIP_RAVEN:
#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
case CHIP_NAVI12:
case CHIP_NAVI10:
case CHIP_NAVI14:
#endif
#if defined(CONFIG_DRM_AMD_DC_DCN2_1)
case CHIP_RENOIR:
#endif
if (dcn10_register_irq_handlers(dm->adev)) {
DRM_ERROR("DM: Failed to initialize IRQ\n");
goto fail;
}
break;
#endif
default:
DRM_ERROR("Unsupported ASIC type: 0x%X\n", adev->asic_type);
goto fail;
}
if (adev->asic_type != CHIP_CARRIZO && adev->asic_type != CHIP_STONEY)
dm->dc->debug.disable_stutter = amdgpu_pp_feature_mask & PP_STUTTER_MODE ? false : true;
return 0;
fail:
kfree(aencoder);
kfree(aconnector);
return -EINVAL;
}
static void amdgpu_dm_destroy_drm_device(struct amdgpu_display_manager *dm)
{
drm_mode_config_cleanup(dm->ddev);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
drm_atomic_private_obj_fini(&dm->atomic_obj);
return;
}
/******************************************************************************
* amdgpu_display_funcs functions
*****************************************************************************/
/*
* dm_bandwidth_update - program display watermarks
*
* @adev: amdgpu_device pointer
*
* Calculate and program the display watermarks and line buffer allocation.
*/
static void dm_bandwidth_update(struct amdgpu_device *adev)
{
/* TODO: implement later */
}
static const struct amdgpu_display_funcs dm_display_funcs = {
.bandwidth_update = dm_bandwidth_update, /* called unconditionally */
.vblank_get_counter = dm_vblank_get_counter,/* called unconditionally */
.backlight_set_level = NULL, /* never called for DC */
.backlight_get_level = NULL, /* never called for DC */
.hpd_sense = NULL,/* called unconditionally */
.hpd_set_polarity = NULL, /* called unconditionally */
.hpd_get_gpio_reg = NULL, /* VBIOS parsing. DAL does it. */
.page_flip_get_scanoutpos =
dm_crtc_get_scanoutpos,/* called unconditionally */
.add_encoder = NULL, /* VBIOS parsing. DAL does it. */
.add_connector = NULL, /* VBIOS parsing. DAL does it. */
};
#if defined(CONFIG_DEBUG_KERNEL_DC)
static ssize_t s3_debug_store(struct device *device,
struct device_attribute *attr,
const char *buf,
size_t count)
{
int ret;
int s3_state;
struct drm_device *drm_dev = dev_get_drvdata(device);
struct amdgpu_device *adev = drm_dev->dev_private;
ret = kstrtoint(buf, 0, &s3_state);
if (ret == 0) {
if (s3_state) {
dm_resume(adev);
drm_kms_helper_hotplug_event(adev->ddev);
} else
dm_suspend(adev);
}
return ret == 0 ? count : 0;
}
DEVICE_ATTR_WO(s3_debug);
#endif
static int dm_early_init(void *handle)
{
struct amdgpu_device *adev = (struct amdgpu_device *)handle;
switch (adev->asic_type) {
case CHIP_BONAIRE:
case CHIP_HAWAII:
adev->mode_info.num_crtc = 6;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 6;
break;
case CHIP_KAVERI:
adev->mode_info.num_crtc = 4;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 7;
break;
case CHIP_KABINI:
case CHIP_MULLINS:
adev->mode_info.num_crtc = 2;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 6;
break;
case CHIP_FIJI:
case CHIP_TONGA:
adev->mode_info.num_crtc = 6;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 7;
break;
case CHIP_CARRIZO:
adev->mode_info.num_crtc = 3;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 9;
break;
case CHIP_STONEY:
adev->mode_info.num_crtc = 2;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 9;
break;
case CHIP_POLARIS11:
case CHIP_POLARIS12:
adev->mode_info.num_crtc = 5;
adev->mode_info.num_hpd = 5;
adev->mode_info.num_dig = 5;
break;
case CHIP_POLARIS10:
case CHIP_VEGAM:
adev->mode_info.num_crtc = 6;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 6;
break;
case CHIP_VEGA10:
case CHIP_VEGA12:
case CHIP_VEGA20:
adev->mode_info.num_crtc = 6;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 6;
break;
#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
case CHIP_RAVEN:
adev->mode_info.num_crtc = 4;
adev->mode_info.num_hpd = 4;
adev->mode_info.num_dig = 4;
break;
#endif
#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
case CHIP_NAVI10:
case CHIP_NAVI12:
adev->mode_info.num_crtc = 6;
adev->mode_info.num_hpd = 6;
adev->mode_info.num_dig = 6;
break;
case CHIP_NAVI14:
adev->mode_info.num_crtc = 5;
adev->mode_info.num_hpd = 5;
adev->mode_info.num_dig = 5;
break;
#endif
#if defined(CONFIG_DRM_AMD_DC_DCN2_1)
case CHIP_RENOIR:
adev->mode_info.num_crtc = 4;
adev->mode_info.num_hpd = 4;
adev->mode_info.num_dig = 4;
break;
#endif
default:
DRM_ERROR("Unsupported ASIC type: 0x%X\n", adev->asic_type);
return -EINVAL;
}
amdgpu_dm_set_irq_funcs(adev);
if (adev->mode_info.funcs == NULL)
adev->mode_info.funcs = &dm_display_funcs;
/*
* Note: Do NOT change adev->audio_endpt_rreg and
* adev->audio_endpt_wreg because they are initialised in
* amdgpu_device_init()
*/
#if defined(CONFIG_DEBUG_KERNEL_DC)
device_create_file(
adev->ddev->dev,
&dev_attr_s3_debug);
#endif
return 0;
}
static bool modeset_required(struct drm_crtc_state *crtc_state,
struct dc_stream_state *new_stream,
struct dc_stream_state *old_stream)
{
if (!drm_atomic_crtc_needs_modeset(crtc_state))
return false;
if (!crtc_state->enable)
return false;
return crtc_state->active;
}
static bool modereset_required(struct drm_crtc_state *crtc_state)
{
if (!drm_atomic_crtc_needs_modeset(crtc_state))
return false;
return !crtc_state->enable || !crtc_state->active;
}
static void amdgpu_dm_encoder_destroy(struct drm_encoder *encoder)
{
drm_encoder_cleanup(encoder);
kfree(encoder);
}
static const struct drm_encoder_funcs amdgpu_dm_encoder_funcs = {
.destroy = amdgpu_dm_encoder_destroy,
};
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
static int fill_dc_scaling_info(const struct drm_plane_state *state,
struct dc_scaling_info *scaling_info)
{
int scale_w, scale_h;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
memset(scaling_info, 0, sizeof(*scaling_info));
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
/* Source is fixed 16.16 but we ignore mantissa for now... */
scaling_info->src_rect.x = state->src_x >> 16;
scaling_info->src_rect.y = state->src_y >> 16;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
scaling_info->src_rect.width = state->src_w >> 16;
if (scaling_info->src_rect.width == 0)
return -EINVAL;
scaling_info->src_rect.height = state->src_h >> 16;
if (scaling_info->src_rect.height == 0)
return -EINVAL;
scaling_info->dst_rect.x = state->crtc_x;
scaling_info->dst_rect.y = state->crtc_y;
if (state->crtc_w == 0)
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
return -EINVAL;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
scaling_info->dst_rect.width = state->crtc_w;
if (state->crtc_h == 0)
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
return -EINVAL;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
scaling_info->dst_rect.height = state->crtc_h;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
/* DRM doesn't specify clipping on destination output. */
scaling_info->clip_rect = scaling_info->dst_rect;
/* TODO: Validate scaling per-format with DC plane caps */
scale_w = scaling_info->dst_rect.width * 1000 /
scaling_info->src_rect.width;
if (scale_w < 250 || scale_w > 16000)
return -EINVAL;
scale_h = scaling_info->dst_rect.height * 1000 /
scaling_info->src_rect.height;
if (scale_h < 250 || scale_h > 16000)
return -EINVAL;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
/*
* The "scaling_quality" can be ignored for now, quality = 0 has DC
* assume reasonable defaults based on the format.
*/
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
return 0;
}
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
static int get_fb_info(const struct amdgpu_framebuffer *amdgpu_fb,
uint64_t *tiling_flags)
{
struct amdgpu_bo *rbo = gem_to_amdgpu_bo(amdgpu_fb->base.obj[0]);
int r = amdgpu_bo_reserve(rbo, false);
if (unlikely(r)) {
/* Don't show error message when returning -ERESTARTSYS */
if (r != -ERESTARTSYS)
DRM_ERROR("Unable to reserve buffer: %d\n", r);
return r;
}
if (tiling_flags)
amdgpu_bo_get_tiling_flags(rbo, tiling_flags);
amdgpu_bo_unreserve(rbo);
return r;
}
static inline uint64_t get_dcc_address(uint64_t address, uint64_t tiling_flags)
{
uint32_t offset = AMDGPU_TILING_GET(tiling_flags, DCC_OFFSET_256B);
return offset ? (address + offset * 256) : 0;
}
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
static int
fill_plane_dcc_attributes(struct amdgpu_device *adev,
const struct amdgpu_framebuffer *afb,
const enum surface_pixel_format format,
const enum dc_rotation_angle rotation,
const struct plane_size *plane_size,
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
const union dc_tiling_info *tiling_info,
const uint64_t info,
struct dc_plane_dcc_param *dcc,
struct dc_plane_address *address,
bool force_disable_dcc)
{
struct dc *dc = adev->dm.dc;
struct dc_dcc_surface_param input;
struct dc_surface_dcc_cap output;
uint32_t offset = AMDGPU_TILING_GET(info, DCC_OFFSET_256B);
uint32_t i64b = AMDGPU_TILING_GET(info, DCC_INDEPENDENT_64B) != 0;
uint64_t dcc_address;
memset(&input, 0, sizeof(input));
memset(&output, 0, sizeof(output));
if (force_disable_dcc)
return 0;
if (!offset)
return 0;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
if (format >= SURFACE_PIXEL_FORMAT_VIDEO_BEGIN)
return 0;
if (!dc->cap_funcs.get_dcc_compression_cap)
return -EINVAL;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
input.format = format;
input.surface_size.width = plane_size->surface_size.width;
input.surface_size.height = plane_size->surface_size.height;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
input.swizzle_mode = tiling_info->gfx9.swizzle;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
if (rotation == ROTATION_ANGLE_0 || rotation == ROTATION_ANGLE_180)
input.scan = SCAN_DIRECTION_HORIZONTAL;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
else if (rotation == ROTATION_ANGLE_90 || rotation == ROTATION_ANGLE_270)
input.scan = SCAN_DIRECTION_VERTICAL;
if (!dc->cap_funcs.get_dcc_compression_cap(dc, &input, &output))
return -EINVAL;
if (!output.capable)
return -EINVAL;
if (i64b == 0 && output.grph.rgb.independent_64b_blks != 0)
return -EINVAL;
dcc->enable = 1;
dcc->meta_pitch =
AMDGPU_TILING_GET(info, DCC_PITCH_MAX) + 1;
dcc->independent_64b_blks = i64b;
dcc_address = get_dcc_address(afb->address, info);
address->grph.meta_addr.low_part = lower_32_bits(dcc_address);
address->grph.meta_addr.high_part = upper_32_bits(dcc_address);
return 0;
}
static int
fill_plane_buffer_attributes(struct amdgpu_device *adev,
const struct amdgpu_framebuffer *afb,
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
const enum surface_pixel_format format,
const enum dc_rotation_angle rotation,
const uint64_t tiling_flags,
union dc_tiling_info *tiling_info,
struct plane_size *plane_size,
struct dc_plane_dcc_param *dcc,
struct dc_plane_address *address,
bool force_disable_dcc)
{
const struct drm_framebuffer *fb = &afb->base;
int ret;
memset(tiling_info, 0, sizeof(*tiling_info));
memset(plane_size, 0, sizeof(*plane_size));
memset(dcc, 0, sizeof(*dcc));
memset(address, 0, sizeof(*address));
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
if (format < SURFACE_PIXEL_FORMAT_VIDEO_BEGIN) {
plane_size->surface_size.x = 0;
plane_size->surface_size.y = 0;
plane_size->surface_size.width = fb->width;
plane_size->surface_size.height = fb->height;
plane_size->surface_pitch =
fb->pitches[0] / fb->format->cpp[0];
address->type = PLN_ADDR_TYPE_GRAPHICS;
address->grph.addr.low_part = lower_32_bits(afb->address);
address->grph.addr.high_part = upper_32_bits(afb->address);
} else if (format < SURFACE_PIXEL_FORMAT_INVALID) {
uint64_t chroma_addr = afb->address + fb->offsets[1];
plane_size->surface_size.x = 0;
plane_size->surface_size.y = 0;
plane_size->surface_size.width = fb->width;
plane_size->surface_size.height = fb->height;
plane_size->surface_pitch =
fb->pitches[0] / fb->format->cpp[0];
plane_size->chroma_size.x = 0;
plane_size->chroma_size.y = 0;
/* TODO: set these based on surface format */
plane_size->chroma_size.width = fb->width / 2;
plane_size->chroma_size.height = fb->height / 2;
plane_size->chroma_pitch =
fb->pitches[1] / fb->format->cpp[1];
address->type = PLN_ADDR_TYPE_VIDEO_PROGRESSIVE;
address->video_progressive.luma_addr.low_part =
lower_32_bits(afb->address);
address->video_progressive.luma_addr.high_part =
upper_32_bits(afb->address);
address->video_progressive.chroma_addr.low_part =
lower_32_bits(chroma_addr);
address->video_progressive.chroma_addr.high_part =
upper_32_bits(chroma_addr);
}
/* Fill GFX8 params */
if (AMDGPU_TILING_GET(tiling_flags, ARRAY_MODE) == DC_ARRAY_2D_TILED_THIN1) {
unsigned int bankw, bankh, mtaspect, tile_split, num_banks;
bankw = AMDGPU_TILING_GET(tiling_flags, BANK_WIDTH);
bankh = AMDGPU_TILING_GET(tiling_flags, BANK_HEIGHT);
mtaspect = AMDGPU_TILING_GET(tiling_flags, MACRO_TILE_ASPECT);
tile_split = AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT);
num_banks = AMDGPU_TILING_GET(tiling_flags, NUM_BANKS);
/* XXX fix me for VI */
tiling_info->gfx8.num_banks = num_banks;
tiling_info->gfx8.array_mode =
DC_ARRAY_2D_TILED_THIN1;
tiling_info->gfx8.tile_split = tile_split;
tiling_info->gfx8.bank_width = bankw;
tiling_info->gfx8.bank_height = bankh;
tiling_info->gfx8.tile_aspect = mtaspect;
tiling_info->gfx8.tile_mode =
DC_ADDR_SURF_MICRO_TILING_DISPLAY;
} else if (AMDGPU_TILING_GET(tiling_flags, ARRAY_MODE)
== DC_ARRAY_1D_TILED_THIN1) {
tiling_info->gfx8.array_mode = DC_ARRAY_1D_TILED_THIN1;
}
tiling_info->gfx8.pipe_config =
AMDGPU_TILING_GET(tiling_flags, PIPE_CONFIG);
if (adev->asic_type == CHIP_VEGA10 ||
adev->asic_type == CHIP_VEGA12 ||
adev->asic_type == CHIP_VEGA20 ||
#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
adev->asic_type == CHIP_NAVI10 ||
adev->asic_type == CHIP_NAVI14 ||
adev->asic_type == CHIP_NAVI12 ||
#endif
#if defined(CONFIG_DRM_AMD_DC_DCN2_1)
adev->asic_type == CHIP_RENOIR ||
#endif
adev->asic_type == CHIP_RAVEN) {
/* Fill GFX9 params */
tiling_info->gfx9.num_pipes =
adev->gfx.config.gb_addr_config_fields.num_pipes;
tiling_info->gfx9.num_banks =
adev->gfx.config.gb_addr_config_fields.num_banks;
tiling_info->gfx9.pipe_interleave =
adev->gfx.config.gb_addr_config_fields.pipe_interleave_size;
tiling_info->gfx9.num_shader_engines =
adev->gfx.config.gb_addr_config_fields.num_se;
tiling_info->gfx9.max_compressed_frags =
adev->gfx.config.gb_addr_config_fields.max_compress_frags;
tiling_info->gfx9.num_rb_per_se =
adev->gfx.config.gb_addr_config_fields.num_rb_per_se;
tiling_info->gfx9.swizzle =
AMDGPU_TILING_GET(tiling_flags, SWIZZLE_MODE);
tiling_info->gfx9.shaderEnable = 1;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
ret = fill_plane_dcc_attributes(adev, afb, format, rotation,
plane_size, tiling_info,
tiling_flags, dcc, address,
force_disable_dcc);
if (ret)
return ret;
}
return 0;
}
static void
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
fill_blending_from_plane_state(const struct drm_plane_state *plane_state,
bool *per_pixel_alpha, bool *global_alpha,
int *global_alpha_value)
{
*per_pixel_alpha = false;
*global_alpha = false;
*global_alpha_value = 0xff;
if (plane_state->plane->type != DRM_PLANE_TYPE_OVERLAY)
return;
if (plane_state->pixel_blend_mode == DRM_MODE_BLEND_PREMULTI) {
static const uint32_t alpha_formats[] = {
DRM_FORMAT_ARGB8888,
DRM_FORMAT_RGBA8888,
DRM_FORMAT_ABGR8888,
};
uint32_t format = plane_state->fb->format->format;
unsigned int i;
for (i = 0; i < ARRAY_SIZE(alpha_formats); ++i) {
if (format == alpha_formats[i]) {
*per_pixel_alpha = true;
break;
}
}
}
if (plane_state->alpha < 0xffff) {
*global_alpha = true;
*global_alpha_value = plane_state->alpha >> 8;
}
}
static int
fill_plane_color_attributes(const struct drm_plane_state *plane_state,
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
const enum surface_pixel_format format,
enum dc_color_space *color_space)
{
bool full_range;
*color_space = COLOR_SPACE_SRGB;
/* DRM color properties only affect non-RGB formats. */
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
if (format < SURFACE_PIXEL_FORMAT_VIDEO_BEGIN)
return 0;
full_range = (plane_state->color_range == DRM_COLOR_YCBCR_FULL_RANGE);
switch (plane_state->color_encoding) {
case DRM_COLOR_YCBCR_BT601:
if (full_range)
*color_space = COLOR_SPACE_YCBCR601;
else
*color_space = COLOR_SPACE_YCBCR601_LIMITED;
break;
case DRM_COLOR_YCBCR_BT709:
if (full_range)
*color_space = COLOR_SPACE_YCBCR709;
else
*color_space = COLOR_SPACE_YCBCR709_LIMITED;
break;
case DRM_COLOR_YCBCR_BT2020:
if (full_range)
*color_space = COLOR_SPACE_2020_YCBCR;
else
return -EINVAL;
break;
default:
return -EINVAL;
}
return 0;
}
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
static int
fill_dc_plane_info_and_addr(struct amdgpu_device *adev,
const struct drm_plane_state *plane_state,
const uint64_t tiling_flags,
struct dc_plane_info *plane_info,
struct dc_plane_address *address,
bool force_disable_dcc)
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
{
const struct drm_framebuffer *fb = plane_state->fb;
const struct amdgpu_framebuffer *afb =
to_amdgpu_framebuffer(plane_state->fb);
struct drm_format_name_buf format_name;
int ret;
memset(plane_info, 0, sizeof(*plane_info));
switch (fb->format->format) {
case DRM_FORMAT_C8:
plane_info->format =
SURFACE_PIXEL_FORMAT_GRPH_PALETA_256_COLORS;
break;
case DRM_FORMAT_RGB565:
plane_info->format = SURFACE_PIXEL_FORMAT_GRPH_RGB565;
break;
case DRM_FORMAT_XRGB8888:
case DRM_FORMAT_ARGB8888:
plane_info->format = SURFACE_PIXEL_FORMAT_GRPH_ARGB8888;
break;
case DRM_FORMAT_XRGB2101010:
case DRM_FORMAT_ARGB2101010:
plane_info->format = SURFACE_PIXEL_FORMAT_GRPH_ARGB2101010;
break;
case DRM_FORMAT_XBGR2101010:
case DRM_FORMAT_ABGR2101010:
plane_info->format = SURFACE_PIXEL_FORMAT_GRPH_ABGR2101010;
break;
case DRM_FORMAT_XBGR8888:
case DRM_FORMAT_ABGR8888:
plane_info->format = SURFACE_PIXEL_FORMAT_GRPH_ABGR8888;
break;
case DRM_FORMAT_NV21:
plane_info->format = SURFACE_PIXEL_FORMAT_VIDEO_420_YCbCr;
break;
case DRM_FORMAT_NV12:
plane_info->format = SURFACE_PIXEL_FORMAT_VIDEO_420_YCrCb;
break;
default:
DRM_ERROR(
"Unsupported screen format %s\n",
drm_get_format_name(fb->format->format, &format_name));
return -EINVAL;
}
switch (plane_state->rotation & DRM_MODE_ROTATE_MASK) {
case DRM_MODE_ROTATE_0:
plane_info->rotation = ROTATION_ANGLE_0;
break;
case DRM_MODE_ROTATE_90:
plane_info->rotation = ROTATION_ANGLE_90;
break;
case DRM_MODE_ROTATE_180:
plane_info->rotation = ROTATION_ANGLE_180;
break;
case DRM_MODE_ROTATE_270:
plane_info->rotation = ROTATION_ANGLE_270;
break;
default:
plane_info->rotation = ROTATION_ANGLE_0;
break;
}
plane_info->visible = true;
plane_info->stereo_format = PLANE_STEREO_FORMAT_NONE;
plane_info->layer_index = 0;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
ret = fill_plane_color_attributes(plane_state, plane_info->format,
&plane_info->color_space);
if (ret)
return ret;
ret = fill_plane_buffer_attributes(adev, afb, plane_info->format,
plane_info->rotation, tiling_flags,
&plane_info->tiling_info,
&plane_info->plane_size,
&plane_info->dcc, address,
force_disable_dcc);
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
if (ret)
return ret;
fill_blending_from_plane_state(
plane_state, &plane_info->per_pixel_alpha,
&plane_info->global_alpha, &plane_info->global_alpha_value);
return 0;
}
static int fill_dc_plane_attributes(struct amdgpu_device *adev,
struct dc_plane_state *dc_plane_state,
struct drm_plane_state *plane_state,
struct drm_crtc_state *crtc_state)
{
drm/amd/display: Rework CRTC color management [Why] To prepare for the upcoming DRM plane color management properties we need to correct a lot of wrong behavior and assumptions made for CRTC color management. The documentation added by this commit in amdgpu_dm_color explains how the HW color pipeline works and its limitations with the DRM interface. The current implementation does the following wrong: - Implicit sRGB DGM when no CRTC DGM is set - Implicit sRGB RGM when no CRTC RGM is set - No way to specify a non-linear DGM matrix that produces correct output - No way to specify a correct RGM when a linear DGM is used We had workarounds for passing kms_color tests but not all of the behavior we had wrong was covered by these tests (especially when it comes to non-linear DGM). Testing both DGM and RGM at the same time isn't something kms_color tests well either. [How] The specifics for how color management works in AMDGPU and the new behavior can be found by reading the documentation added to amdgpu_dm_color.c from this patch. All of the incorrect cases from the old implementation have been addressed for the atomic interface, but there still a few TODOs for the legacy one. Note: this does cause regressions for kms_color@pipe-a-ctm-* over HDMI. The result looks correct from visual inspection but the CRC no longer matches. For reference, the test was previously doing the following: linear degamma -> CTM -> sRGB regamma -> RGB to YUV (709) -> ... Now the test is doing: linear degamma -> CTM -> linear regamma -> RGB to YUV (709) -> ... Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-05-09 10:14:58 -06:00
struct dm_crtc_state *dm_crtc_state = to_dm_crtc_state(crtc_state);
const struct amdgpu_framebuffer *amdgpu_fb =
to_amdgpu_framebuffer(plane_state->fb);
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
struct dc_scaling_info scaling_info;
struct dc_plane_info plane_info;
uint64_t tiling_flags;
int ret;
bool force_disable_dcc = false;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
ret = fill_dc_scaling_info(plane_state, &scaling_info);
if (ret)
return ret;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
dc_plane_state->src_rect = scaling_info.src_rect;
dc_plane_state->dst_rect = scaling_info.dst_rect;
dc_plane_state->clip_rect = scaling_info.clip_rect;
dc_plane_state->scaling_quality = scaling_info.scaling_quality;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
ret = get_fb_info(amdgpu_fb, &tiling_flags);
if (ret)
return ret;
force_disable_dcc = adev->asic_type == CHIP_RAVEN && adev->in_suspend;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
ret = fill_dc_plane_info_and_addr(adev, plane_state, tiling_flags,
&plane_info,
&dc_plane_state->address,
force_disable_dcc);
if (ret)
return ret;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
dc_plane_state->format = plane_info.format;
dc_plane_state->color_space = plane_info.color_space;
dc_plane_state->format = plane_info.format;
dc_plane_state->plane_size = plane_info.plane_size;
dc_plane_state->rotation = plane_info.rotation;
dc_plane_state->horizontal_mirror = plane_info.horizontal_mirror;
dc_plane_state->stereo_format = plane_info.stereo_format;
dc_plane_state->tiling_info = plane_info.tiling_info;
dc_plane_state->visible = plane_info.visible;
dc_plane_state->per_pixel_alpha = plane_info.per_pixel_alpha;
dc_plane_state->global_alpha = plane_info.global_alpha;
dc_plane_state->global_alpha_value = plane_info.global_alpha_value;
dc_plane_state->dcc = plane_info.dcc;
dc_plane_state->layer_index = plane_info.layer_index; // Always returns 0
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
/*
* Always set input transfer function, since plane state is refreshed
* every time.
*/
drm/amd/display: Rework CRTC color management [Why] To prepare for the upcoming DRM plane color management properties we need to correct a lot of wrong behavior and assumptions made for CRTC color management. The documentation added by this commit in amdgpu_dm_color explains how the HW color pipeline works and its limitations with the DRM interface. The current implementation does the following wrong: - Implicit sRGB DGM when no CRTC DGM is set - Implicit sRGB RGM when no CRTC RGM is set - No way to specify a non-linear DGM matrix that produces correct output - No way to specify a correct RGM when a linear DGM is used We had workarounds for passing kms_color tests but not all of the behavior we had wrong was covered by these tests (especially when it comes to non-linear DGM). Testing both DGM and RGM at the same time isn't something kms_color tests well either. [How] The specifics for how color management works in AMDGPU and the new behavior can be found by reading the documentation added to amdgpu_dm_color.c from this patch. All of the incorrect cases from the old implementation have been addressed for the atomic interface, but there still a few TODOs for the legacy one. Note: this does cause regressions for kms_color@pipe-a-ctm-* over HDMI. The result looks correct from visual inspection but the CRC no longer matches. For reference, the test was previously doing the following: linear degamma -> CTM -> sRGB regamma -> RGB to YUV (709) -> ... Now the test is doing: linear degamma -> CTM -> linear regamma -> RGB to YUV (709) -> ... Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-05-09 10:14:58 -06:00
ret = amdgpu_dm_update_plane_color_mgmt(dm_crtc_state, dc_plane_state);
if (ret)
return ret;
drm/amd/display: Rework CRTC color management [Why] To prepare for the upcoming DRM plane color management properties we need to correct a lot of wrong behavior and assumptions made for CRTC color management. The documentation added by this commit in amdgpu_dm_color explains how the HW color pipeline works and its limitations with the DRM interface. The current implementation does the following wrong: - Implicit sRGB DGM when no CRTC DGM is set - Implicit sRGB RGM when no CRTC RGM is set - No way to specify a non-linear DGM matrix that produces correct output - No way to specify a correct RGM when a linear DGM is used We had workarounds for passing kms_color tests but not all of the behavior we had wrong was covered by these tests (especially when it comes to non-linear DGM). Testing both DGM and RGM at the same time isn't something kms_color tests well either. [How] The specifics for how color management works in AMDGPU and the new behavior can be found by reading the documentation added to amdgpu_dm_color.c from this patch. All of the incorrect cases from the old implementation have been addressed for the atomic interface, but there still a few TODOs for the legacy one. Note: this does cause regressions for kms_color@pipe-a-ctm-* over HDMI. The result looks correct from visual inspection but the CRC no longer matches. For reference, the test was previously doing the following: linear degamma -> CTM -> sRGB regamma -> RGB to YUV (709) -> ... Now the test is doing: linear degamma -> CTM -> linear regamma -> RGB to YUV (709) -> ... Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-05-09 10:14:58 -06:00
return 0;
}
static void update_stream_scaling_settings(const struct drm_display_mode *mode,
const struct dm_connector_state *dm_state,
struct dc_stream_state *stream)
{
enum amdgpu_rmx_type rmx_type;
struct rect src = { 0 }; /* viewport in composition space*/
struct rect dst = { 0 }; /* stream addressable area */
/* no mode. nothing to be done */
if (!mode)
return;
/* Full screen scaling by default */
src.width = mode->hdisplay;
src.height = mode->vdisplay;
dst.width = stream->timing.h_addressable;
dst.height = stream->timing.v_addressable;
if (dm_state) {
rmx_type = dm_state->scaling;
if (rmx_type == RMX_ASPECT || rmx_type == RMX_OFF) {
if (src.width * dst.height <
src.height * dst.width) {
/* height needs less upscaling/more downscaling */
dst.width = src.width *
dst.height / src.height;
} else {
/* width needs less upscaling/more downscaling */
dst.height = src.height *
dst.width / src.width;
}
} else if (rmx_type == RMX_CENTER) {
dst = src;
}
dst.x = (stream->timing.h_addressable - dst.width) / 2;
dst.y = (stream->timing.v_addressable - dst.height) / 2;
if (dm_state->underscan_enable) {
dst.x += dm_state->underscan_hborder / 2;
dst.y += dm_state->underscan_vborder / 2;
dst.width -= dm_state->underscan_hborder;
dst.height -= dm_state->underscan_vborder;
}
}
stream->src = src;
stream->dst = dst;
DRM_DEBUG_DRIVER("Destination Rectangle x:%d y:%d width:%d height:%d\n",
dst.x, dst.y, dst.width, dst.height);
}
static enum dc_color_depth
convert_color_depth_from_display_info(const struct drm_connector *connector,
const struct drm_connector_state *state)
{
uint8_t bpc = (uint8_t)connector->display_info.bpc;
/* Assume 8 bpc by default if no bpc is specified. */
bpc = bpc ? bpc : 8;
if (!state)
state = connector->state;
if (state) {
/*
* Cap display bpc based on the user requested value.
*
* The value for state->max_bpc may not correctly updated
* depending on when the connector gets added to the state
* or if this was called outside of atomic check, so it
* can't be used directly.
*/
bpc = min(bpc, state->max_requested_bpc);
/* Round down to the nearest even number. */
bpc = bpc - (bpc & 1);
}
switch (bpc) {
case 0:
/*
* Temporary Work around, DRM doesn't parse color depth for
* EDID revision before 1.4
* TODO: Fix edid parsing
*/
return COLOR_DEPTH_888;
case 6:
return COLOR_DEPTH_666;
case 8:
return COLOR_DEPTH_888;
case 10:
return COLOR_DEPTH_101010;
case 12:
return COLOR_DEPTH_121212;
case 14:
return COLOR_DEPTH_141414;
case 16:
return COLOR_DEPTH_161616;
default:
return COLOR_DEPTH_UNDEFINED;
}
}
static enum dc_aspect_ratio
get_aspect_ratio(const struct drm_display_mode *mode_in)
{
/* 1-1 mapping, since both enums follow the HDMI spec. */
return (enum dc_aspect_ratio) mode_in->picture_aspect_ratio;
}
static enum dc_color_space
get_output_color_space(const struct dc_crtc_timing *dc_crtc_timing)
{
enum dc_color_space color_space = COLOR_SPACE_SRGB;
switch (dc_crtc_timing->pixel_encoding) {
case PIXEL_ENCODING_YCBCR422:
case PIXEL_ENCODING_YCBCR444:
case PIXEL_ENCODING_YCBCR420:
{
/*
* 27030khz is the separation point between HDTV and SDTV
* according to HDMI spec, we use YCbCr709 and YCbCr601
* respectively
*/
if (dc_crtc_timing->pix_clk_100hz > 270300) {
if (dc_crtc_timing->flags.Y_ONLY)
color_space =
COLOR_SPACE_YCBCR709_LIMITED;
else
color_space = COLOR_SPACE_YCBCR709;
} else {
if (dc_crtc_timing->flags.Y_ONLY)
color_space =
COLOR_SPACE_YCBCR601_LIMITED;
else
color_space = COLOR_SPACE_YCBCR601;
}
}
break;
case PIXEL_ENCODING_RGB:
color_space = COLOR_SPACE_SRGB;
break;
default:
WARN_ON(1);
break;
}
return color_space;
}
static bool adjust_colour_depth_from_display_info(
struct dc_crtc_timing *timing_out,
const struct drm_display_info *info)
{
enum dc_color_depth depth = timing_out->display_color_depth;
int normalized_clk;
do {
normalized_clk = timing_out->pix_clk_100hz / 10;
/* YCbCr 4:2:0 requires additional adjustment of 1/2 */
if (timing_out->pixel_encoding == PIXEL_ENCODING_YCBCR420)
normalized_clk /= 2;
/* Adjusting pix clock following on HDMI spec based on colour depth */
switch (depth) {
case COLOR_DEPTH_888:
break;
case COLOR_DEPTH_101010:
normalized_clk = (normalized_clk * 30) / 24;
break;
case COLOR_DEPTH_121212:
normalized_clk = (normalized_clk * 36) / 24;
break;
case COLOR_DEPTH_161616:
normalized_clk = (normalized_clk * 48) / 24;
break;
default:
/* The above depths are the only ones valid for HDMI. */
return false;
}
if (normalized_clk <= info->max_tmds_clock) {
timing_out->display_color_depth = depth;
return true;
}
} while (--depth > COLOR_DEPTH_666);
return false;
}
static void fill_stream_properties_from_drm_display_mode(
struct dc_stream_state *stream,
const struct drm_display_mode *mode_in,
const struct drm_connector *connector,
const struct drm_connector_state *connector_state,
const struct dc_stream_state *old_stream)
{
struct dc_crtc_timing *timing_out = &stream->timing;
const struct drm_display_info *info = &connector->display_info;
memset(timing_out, 0, sizeof(struct dc_crtc_timing));
timing_out->h_border_left = 0;
timing_out->h_border_right = 0;
timing_out->v_border_top = 0;
timing_out->v_border_bottom = 0;
/* TODO: un-hardcode */
if (drm_mode_is_420_only(info, mode_in)
drm/amd/display: remove sink reference in dc_stream_state [why] dc_stream_state containing a pointer to sink is poor design. Sink describes the display, and the specifications or capabilities it has. That information is irrelevant for dc_stream_state, which describes hardware state, and is generally used for hardware programming. It could further be argued that dc_sink itself is just a convenience dc provides, and DC should be perfectly capable of programming hardware without any dc_sinks (for example, emulated sinks). [how] Phase 1: Deprecate use of dc_sink pointer in dc_stream. Most references are trivial to remove, but some call sites are risky (such as is_timing_changed) with no obvious logical replacement. These will be removed in follow up change. Add dc_link pointer to dc_stream. This is the typical reason DC really needed sink pointer, and most call sites are replaced with this. DMs also need minor updates, as all 3 DMs leverage stream->sink for some functionality. this is replaced instead by a pointer to private data inside dc_stream_state, which is used by DMs as a quality of life improvment for some key functionality. it allows DMs to set pointers have to their own objects which associate OS objects to dc_stream_states (such as DisplayTarget and amdgpu_dm_connector). Without the private pointer, DMs would be forced to perform a lookup for callbacks. Signed-off-by: Jun Lei <Jun.Lei@amd.com> Reviewed-by: Aric Cyr <Aric.Cyr@amd.com> Acked-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-09 07:21:21 -07:00
&& stream->signal == SIGNAL_TYPE_HDMI_TYPE_A)
timing_out->pixel_encoding = PIXEL_ENCODING_YCBCR420;
else if ((connector->display_info.color_formats & DRM_COLOR_FORMAT_YCRCB444)
drm/amd/display: remove sink reference in dc_stream_state [why] dc_stream_state containing a pointer to sink is poor design. Sink describes the display, and the specifications or capabilities it has. That information is irrelevant for dc_stream_state, which describes hardware state, and is generally used for hardware programming. It could further be argued that dc_sink itself is just a convenience dc provides, and DC should be perfectly capable of programming hardware without any dc_sinks (for example, emulated sinks). [how] Phase 1: Deprecate use of dc_sink pointer in dc_stream. Most references are trivial to remove, but some call sites are risky (such as is_timing_changed) with no obvious logical replacement. These will be removed in follow up change. Add dc_link pointer to dc_stream. This is the typical reason DC really needed sink pointer, and most call sites are replaced with this. DMs also need minor updates, as all 3 DMs leverage stream->sink for some functionality. this is replaced instead by a pointer to private data inside dc_stream_state, which is used by DMs as a quality of life improvment for some key functionality. it allows DMs to set pointers have to their own objects which associate OS objects to dc_stream_states (such as DisplayTarget and amdgpu_dm_connector). Without the private pointer, DMs would be forced to perform a lookup for callbacks. Signed-off-by: Jun Lei <Jun.Lei@amd.com> Reviewed-by: Aric Cyr <Aric.Cyr@amd.com> Acked-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-09 07:21:21 -07:00
&& stream->signal == SIGNAL_TYPE_HDMI_TYPE_A)
timing_out->pixel_encoding = PIXEL_ENCODING_YCBCR444;
else
timing_out->pixel_encoding = PIXEL_ENCODING_RGB;
timing_out->timing_3d_format = TIMING_3D_FORMAT_NONE;
timing_out->display_color_depth = convert_color_depth_from_display_info(
connector, connector_state);
timing_out->scan_type = SCANNING_TYPE_NODATA;
timing_out->hdmi_vic = 0;
if(old_stream) {
timing_out->vic = old_stream->timing.vic;
timing_out->flags.HSYNC_POSITIVE_POLARITY = old_stream->timing.flags.HSYNC_POSITIVE_POLARITY;
timing_out->flags.VSYNC_POSITIVE_POLARITY = old_stream->timing.flags.VSYNC_POSITIVE_POLARITY;
} else {
timing_out->vic = drm_match_cea_mode(mode_in);
if (mode_in->flags & DRM_MODE_FLAG_PHSYNC)
timing_out->flags.HSYNC_POSITIVE_POLARITY = 1;
if (mode_in->flags & DRM_MODE_FLAG_PVSYNC)
timing_out->flags.VSYNC_POSITIVE_POLARITY = 1;
}
timing_out->h_addressable = mode_in->crtc_hdisplay;
timing_out->h_total = mode_in->crtc_htotal;
timing_out->h_sync_width =
mode_in->crtc_hsync_end - mode_in->crtc_hsync_start;
timing_out->h_front_porch =
mode_in->crtc_hsync_start - mode_in->crtc_hdisplay;
timing_out->v_total = mode_in->crtc_vtotal;
timing_out->v_addressable = mode_in->crtc_vdisplay;
timing_out->v_front_porch =
mode_in->crtc_vsync_start - mode_in->crtc_vdisplay;
timing_out->v_sync_width =
mode_in->crtc_vsync_end - mode_in->crtc_vsync_start;
timing_out->pix_clk_100hz = mode_in->crtc_clock * 10;
timing_out->aspect_ratio = get_aspect_ratio(mode_in);
stream->output_color_space = get_output_color_space(timing_out);
stream->out_transfer_func->type = TF_TYPE_PREDEFINED;
stream->out_transfer_func->tf = TRANSFER_FUNCTION_SRGB;
if (stream->signal == SIGNAL_TYPE_HDMI_TYPE_A) {
if (!adjust_colour_depth_from_display_info(timing_out, info) &&
drm_mode_is_420_also(info, mode_in) &&
timing_out->pixel_encoding != PIXEL_ENCODING_YCBCR420) {
timing_out->pixel_encoding = PIXEL_ENCODING_YCBCR420;
adjust_colour_depth_from_display_info(timing_out, info);
}
}
}
static void fill_audio_info(struct audio_info *audio_info,
const struct drm_connector *drm_connector,
const struct dc_sink *dc_sink)
{
int i = 0;
int cea_revision = 0;
const struct dc_edid_caps *edid_caps = &dc_sink->edid_caps;
audio_info->manufacture_id = edid_caps->manufacturer_id;
audio_info->product_id = edid_caps->product_id;
cea_revision = drm_connector->display_info.cea_rev;
strscpy(audio_info->display_name,
edid_caps->display_name,
AUDIO_INFO_DISPLAY_NAME_SIZE_IN_CHARS);
if (cea_revision >= 3) {
audio_info->mode_count = edid_caps->audio_mode_count;
for (i = 0; i < audio_info->mode_count; ++i) {
audio_info->modes[i].format_code =
(enum audio_format_code)
(edid_caps->audio_modes[i].format_code);
audio_info->modes[i].channel_count =
edid_caps->audio_modes[i].channel_count;
audio_info->modes[i].sample_rates.all =
edid_caps->audio_modes[i].sample_rate;
audio_info->modes[i].sample_size =
edid_caps->audio_modes[i].sample_size;
}
}
audio_info->flags.all = edid_caps->speaker_flags;
/* TODO: We only check for the progressive mode, check for interlace mode too */
if (drm_connector->latency_present[0]) {
audio_info->video_latency = drm_connector->video_latency[0];
audio_info->audio_latency = drm_connector->audio_latency[0];
}
/* TODO: For DP, video and audio latency should be calculated from DPCD caps */
}
static void
copy_crtc_timing_for_drm_display_mode(const struct drm_display_mode *src_mode,
struct drm_display_mode *dst_mode)
{
dst_mode->crtc_hdisplay = src_mode->crtc_hdisplay;
dst_mode->crtc_vdisplay = src_mode->crtc_vdisplay;
dst_mode->crtc_clock = src_mode->crtc_clock;
dst_mode->crtc_hblank_start = src_mode->crtc_hblank_start;
dst_mode->crtc_hblank_end = src_mode->crtc_hblank_end;
dst_mode->crtc_hsync_start = src_mode->crtc_hsync_start;
dst_mode->crtc_hsync_end = src_mode->crtc_hsync_end;
dst_mode->crtc_htotal = src_mode->crtc_htotal;
dst_mode->crtc_hskew = src_mode->crtc_hskew;
dst_mode->crtc_vblank_start = src_mode->crtc_vblank_start;
dst_mode->crtc_vblank_end = src_mode->crtc_vblank_end;
dst_mode->crtc_vsync_start = src_mode->crtc_vsync_start;
dst_mode->crtc_vsync_end = src_mode->crtc_vsync_end;
dst_mode->crtc_vtotal = src_mode->crtc_vtotal;
}
static void
decide_crtc_timing_for_drm_display_mode(struct drm_display_mode *drm_mode,
const struct drm_display_mode *native_mode,
bool scale_enabled)
{
if (scale_enabled) {
copy_crtc_timing_for_drm_display_mode(native_mode, drm_mode);
} else if (native_mode->clock == drm_mode->clock &&
native_mode->htotal == drm_mode->htotal &&
native_mode->vtotal == drm_mode->vtotal) {
copy_crtc_timing_for_drm_display_mode(native_mode, drm_mode);
} else {
/* no scaling nor amdgpu inserted, no need to patch */
}
}
static struct dc_sink *
create_fake_sink(struct amdgpu_dm_connector *aconnector)
{
struct dc_sink_init_data sink_init_data = { 0 };
struct dc_sink *sink = NULL;
sink_init_data.link = aconnector->dc_link;
sink_init_data.sink_signal = aconnector->dc_link->connector_signal;
sink = dc_sink_create(&sink_init_data);
if (!sink) {
DRM_ERROR("Failed to create sink!\n");
return NULL;
}
sink->sink_signal = SIGNAL_TYPE_VIRTUAL;
return sink;
}
static void set_multisync_trigger_params(
struct dc_stream_state *stream)
{
if (stream->triggered_crtc_reset.enabled) {
stream->triggered_crtc_reset.event = CRTC_EVENT_VSYNC_RISING;
stream->triggered_crtc_reset.delay = TRIGGER_DELAY_NEXT_LINE;
}
}
static void set_master_stream(struct dc_stream_state *stream_set[],
int stream_count)
{
int j, highest_rfr = 0, master_stream = 0;
for (j = 0; j < stream_count; j++) {
if (stream_set[j] && stream_set[j]->triggered_crtc_reset.enabled) {
int refresh_rate = 0;
refresh_rate = (stream_set[j]->timing.pix_clk_100hz*100)/
(stream_set[j]->timing.h_total*stream_set[j]->timing.v_total);
if (refresh_rate > highest_rfr) {
highest_rfr = refresh_rate;
master_stream = j;
}
}
}
for (j = 0; j < stream_count; j++) {
if (stream_set[j])
stream_set[j]->triggered_crtc_reset.event_source = stream_set[master_stream];
}
}
static void dm_enable_per_frame_crtc_master_sync(struct dc_state *context)
{
int i = 0;
if (context->stream_count < 2)
return;
for (i = 0; i < context->stream_count ; i++) {
if (!context->streams[i])
continue;
/*
* TODO: add a function to read AMD VSDB bits and set
* crtc_sync_master.multi_sync_enabled flag
* For now it's set to false
*/
set_multisync_trigger_params(context->streams[i]);
}
set_master_stream(context->streams, context->stream_count);
}
static struct dc_stream_state *
create_stream_for_sink(struct amdgpu_dm_connector *aconnector,
const struct drm_display_mode *drm_mode,
const struct dm_connector_state *dm_state,
const struct dc_stream_state *old_stream)
{
struct drm_display_mode *preferred_mode = NULL;
struct drm_connector *drm_connector;
const struct drm_connector_state *con_state =
dm_state ? &dm_state->base : NULL;
struct dc_stream_state *stream = NULL;
struct drm_display_mode mode = *drm_mode;
bool native_mode_found = false;
bool scale = dm_state ? (dm_state->scaling != RMX_OFF) : false;
int mode_refresh;
int preferred_refresh = 0;
#ifdef CONFIG_DRM_AMD_DC_DSC_SUPPORT
struct dsc_dec_dpcd_caps dsc_caps;
uint32_t link_bandwidth_kbps;
#endif
struct dc_sink *sink = NULL;
if (aconnector == NULL) {
DRM_ERROR("aconnector is NULL!\n");
return stream;
}
drm_connector = &aconnector->base;
if (!aconnector->dc_sink) {
sink = create_fake_sink(aconnector);
if (!sink)
return stream;
} else {
sink = aconnector->dc_sink;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
dc_sink_retain(sink);
}
stream = dc_create_stream_for_sink(sink);
if (stream == NULL) {
DRM_ERROR("Failed to create stream for sink!\n");
goto finish;
}
drm/amd/display: remove sink reference in dc_stream_state [why] dc_stream_state containing a pointer to sink is poor design. Sink describes the display, and the specifications or capabilities it has. That information is irrelevant for dc_stream_state, which describes hardware state, and is generally used for hardware programming. It could further be argued that dc_sink itself is just a convenience dc provides, and DC should be perfectly capable of programming hardware without any dc_sinks (for example, emulated sinks). [how] Phase 1: Deprecate use of dc_sink pointer in dc_stream. Most references are trivial to remove, but some call sites are risky (such as is_timing_changed) with no obvious logical replacement. These will be removed in follow up change. Add dc_link pointer to dc_stream. This is the typical reason DC really needed sink pointer, and most call sites are replaced with this. DMs also need minor updates, as all 3 DMs leverage stream->sink for some functionality. this is replaced instead by a pointer to private data inside dc_stream_state, which is used by DMs as a quality of life improvment for some key functionality. it allows DMs to set pointers have to their own objects which associate OS objects to dc_stream_states (such as DisplayTarget and amdgpu_dm_connector). Without the private pointer, DMs would be forced to perform a lookup for callbacks. Signed-off-by: Jun Lei <Jun.Lei@amd.com> Reviewed-by: Aric Cyr <Aric.Cyr@amd.com> Acked-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-09 07:21:21 -07:00
stream->dm_stream_context = aconnector;
list_for_each_entry(preferred_mode, &aconnector->base.modes, head) {
/* Search for preferred mode */
if (preferred_mode->type & DRM_MODE_TYPE_PREFERRED) {
native_mode_found = true;
break;
}
}
if (!native_mode_found)
preferred_mode = list_first_entry_or_null(
&aconnector->base.modes,
struct drm_display_mode,
head);
mode_refresh = drm_mode_vrefresh(&mode);
if (preferred_mode == NULL) {
/*
* This may not be an error, the use case is when we have no
* usermode calls to reset and set mode upon hotplug. In this
* case, we call set mode ourselves to restore the previous mode
* and the modelist may not be filled in in time.
*/
DRM_DEBUG_DRIVER("No preferred mode found\n");
} else {
decide_crtc_timing_for_drm_display_mode(
&mode, preferred_mode,
dm_state ? (dm_state->scaling != RMX_OFF) : false);
preferred_refresh = drm_mode_vrefresh(preferred_mode);
}
if (!dm_state)
drm_mode_set_crtcinfo(&mode, 0);
/*
* If scaling is enabled and refresh rate didn't change
* we copy the vic and polarities of the old timings
*/
if (!scale || mode_refresh != preferred_refresh)
fill_stream_properties_from_drm_display_mode(stream,
&mode, &aconnector->base, con_state, NULL);
else
fill_stream_properties_from_drm_display_mode(stream,
&mode, &aconnector->base, con_state, old_stream);
#ifdef CONFIG_DRM_AMD_DC_DSC_SUPPORT
stream->timing.flags.DSC = 0;
if (aconnector->dc_link && sink->sink_signal == SIGNAL_TYPE_DISPLAY_PORT) {
dc_dsc_parse_dsc_dpcd(aconnector->dc_link->dpcd_caps.dsc_caps.dsc_basic_caps.raw,
aconnector->dc_link->dpcd_caps.dsc_caps.dsc_ext_caps.raw,
&dsc_caps);
link_bandwidth_kbps = dc_link_bandwidth_kbps(aconnector->dc_link,
dc_link_get_link_cap(aconnector->dc_link));
if (dsc_caps.is_dsc_supported)
if (dc_dsc_compute_config(aconnector->dc_link->ctx->dc,
&dsc_caps,
link_bandwidth_kbps,
&stream->timing,
&stream->timing.dsc_cfg))
stream->timing.flags.DSC = 1;
}
#endif
update_stream_scaling_settings(&mode, dm_state, stream);
fill_audio_info(
&stream->audio_info,
drm_connector,
sink);
drm/amd/display: remove sink reference in dc_stream_state [why] dc_stream_state containing a pointer to sink is poor design. Sink describes the display, and the specifications or capabilities it has. That information is irrelevant for dc_stream_state, which describes hardware state, and is generally used for hardware programming. It could further be argued that dc_sink itself is just a convenience dc provides, and DC should be perfectly capable of programming hardware without any dc_sinks (for example, emulated sinks). [how] Phase 1: Deprecate use of dc_sink pointer in dc_stream. Most references are trivial to remove, but some call sites are risky (such as is_timing_changed) with no obvious logical replacement. These will be removed in follow up change. Add dc_link pointer to dc_stream. This is the typical reason DC really needed sink pointer, and most call sites are replaced with this. DMs also need minor updates, as all 3 DMs leverage stream->sink for some functionality. this is replaced instead by a pointer to private data inside dc_stream_state, which is used by DMs as a quality of life improvment for some key functionality. it allows DMs to set pointers have to their own objects which associate OS objects to dc_stream_states (such as DisplayTarget and amdgpu_dm_connector). Without the private pointer, DMs would be forced to perform a lookup for callbacks. Signed-off-by: Jun Lei <Jun.Lei@amd.com> Reviewed-by: Aric Cyr <Aric.Cyr@amd.com> Acked-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-09 07:21:21 -07:00
update_stream_signal(stream, sink);
finish:
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
dc_sink_release(sink);
return stream;
}
static void amdgpu_dm_crtc_destroy(struct drm_crtc *crtc)
{
drm_crtc_cleanup(crtc);
kfree(crtc);
}
static void dm_crtc_destroy_state(struct drm_crtc *crtc,
struct drm_crtc_state *state)
{
struct dm_crtc_state *cur = to_dm_crtc_state(state);
/* TODO Destroy dc_stream objects are stream object is flattened */
if (cur->stream)
dc_stream_release(cur->stream);
__drm_atomic_helper_crtc_destroy_state(state);
kfree(state);
}
static void dm_crtc_reset_state(struct drm_crtc *crtc)
{
struct dm_crtc_state *state;
if (crtc->state)
dm_crtc_destroy_state(crtc, crtc->state);
state = kzalloc(sizeof(*state), GFP_KERNEL);
if (WARN_ON(!state))
return;
crtc->state = &state->base;
crtc->state->crtc = crtc;
}
static struct drm_crtc_state *
dm_crtc_duplicate_state(struct drm_crtc *crtc)
{
struct dm_crtc_state *state, *cur;
cur = to_dm_crtc_state(crtc->state);
if (WARN_ON(!crtc->state))
return NULL;
state = kzalloc(sizeof(*state), GFP_KERNEL);
if (!state)
return NULL;
__drm_atomic_helper_crtc_duplicate_state(crtc, &state->base);
if (cur->stream) {
state->stream = cur->stream;
dc_stream_retain(state->stream);
}
state->active_planes = cur->active_planes;
state->interrupts_enabled = cur->interrupts_enabled;
state->vrr_params = cur->vrr_params;
state->vrr_infopacket = cur->vrr_infopacket;
state->abm_level = cur->abm_level;
state->vrr_supported = cur->vrr_supported;
state->freesync_config = cur->freesync_config;
state->crc_src = cur->crc_src;
drm/amd/display: Rework CRTC color management [Why] To prepare for the upcoming DRM plane color management properties we need to correct a lot of wrong behavior and assumptions made for CRTC color management. The documentation added by this commit in amdgpu_dm_color explains how the HW color pipeline works and its limitations with the DRM interface. The current implementation does the following wrong: - Implicit sRGB DGM when no CRTC DGM is set - Implicit sRGB RGM when no CRTC RGM is set - No way to specify a non-linear DGM matrix that produces correct output - No way to specify a correct RGM when a linear DGM is used We had workarounds for passing kms_color tests but not all of the behavior we had wrong was covered by these tests (especially when it comes to non-linear DGM). Testing both DGM and RGM at the same time isn't something kms_color tests well either. [How] The specifics for how color management works in AMDGPU and the new behavior can be found by reading the documentation added to amdgpu_dm_color.c from this patch. All of the incorrect cases from the old implementation have been addressed for the atomic interface, but there still a few TODOs for the legacy one. Note: this does cause regressions for kms_color@pipe-a-ctm-* over HDMI. The result looks correct from visual inspection but the CRC no longer matches. For reference, the test was previously doing the following: linear degamma -> CTM -> sRGB regamma -> RGB to YUV (709) -> ... Now the test is doing: linear degamma -> CTM -> linear regamma -> RGB to YUV (709) -> ... Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-05-09 10:14:58 -06:00
state->cm_has_degamma = cur->cm_has_degamma;
state->cm_is_degamma_srgb = cur->cm_is_degamma_srgb;
/* TODO Duplicate dc_stream after objects are stream object is flattened */
return &state->base;
}
static inline int dm_set_vupdate_irq(struct drm_crtc *crtc, bool enable)
{
enum dc_irq_source irq_source;
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(crtc);
struct amdgpu_device *adev = crtc->dev->dev_private;
int rc;
irq_source = IRQ_TYPE_VUPDATE + acrtc->otg_inst;
rc = dc_interrupt_set(adev->dm.dc, irq_source, enable) ? 0 : -EBUSY;
DRM_DEBUG_DRIVER("crtc %d - vupdate irq %sabling: r=%d\n",
acrtc->crtc_id, enable ? "en" : "dis", rc);
return rc;
}
static inline int dm_set_vblank(struct drm_crtc *crtc, bool enable)
{
enum dc_irq_source irq_source;
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(crtc);
struct amdgpu_device *adev = crtc->dev->dev_private;
struct dm_crtc_state *acrtc_state = to_dm_crtc_state(crtc->state);
int rc = 0;
if (enable) {
/* vblank irq on -> Only need vupdate irq in vrr mode */
if (amdgpu_dm_vrr_active(acrtc_state))
rc = dm_set_vupdate_irq(crtc, true);
} else {
/* vblank irq off -> vupdate irq off */
rc = dm_set_vupdate_irq(crtc, false);
}
if (rc)
return rc;
irq_source = IRQ_TYPE_VBLANK + acrtc->otg_inst;
return dc_interrupt_set(adev->dm.dc, irq_source, enable) ? 0 : -EBUSY;
}
static int dm_enable_vblank(struct drm_crtc *crtc)
{
return dm_set_vblank(crtc, true);
}
static void dm_disable_vblank(struct drm_crtc *crtc)
{
dm_set_vblank(crtc, false);
}
/* Implemented only the options currently availible for the driver */
static const struct drm_crtc_funcs amdgpu_dm_crtc_funcs = {
.reset = dm_crtc_reset_state,
.destroy = amdgpu_dm_crtc_destroy,
.gamma_set = drm_atomic_helper_legacy_gamma_set,
.set_config = drm_atomic_helper_set_config,
.page_flip = drm_atomic_helper_page_flip,
.atomic_duplicate_state = dm_crtc_duplicate_state,
.atomic_destroy_state = dm_crtc_destroy_state,
.set_crc_source = amdgpu_dm_crtc_set_crc_source,
.verify_crc_source = amdgpu_dm_crtc_verify_crc_source,
.get_crc_sources = amdgpu_dm_crtc_get_crc_sources,
.enable_vblank = dm_enable_vblank,
.disable_vblank = dm_disable_vblank,
};
static enum drm_connector_status
amdgpu_dm_connector_detect(struct drm_connector *connector, bool force)
{
bool connected;
struct amdgpu_dm_connector *aconnector = to_amdgpu_dm_connector(connector);
/*
* Notes:
* 1. This interface is NOT called in context of HPD irq.
* 2. This interface *is called* in context of user-mode ioctl. Which
* makes it a bad place for *any* MST-related activity.
*/
if (aconnector->base.force == DRM_FORCE_UNSPECIFIED &&
!aconnector->fake_enable)
connected = (aconnector->dc_sink != NULL);
else
connected = (aconnector->base.force == DRM_FORCE_ON);
return (connected ? connector_status_connected :
connector_status_disconnected);
}
int amdgpu_dm_connector_atomic_set_property(struct drm_connector *connector,
struct drm_connector_state *connector_state,
struct drm_property *property,
uint64_t val)
{
struct drm_device *dev = connector->dev;
struct amdgpu_device *adev = dev->dev_private;
struct dm_connector_state *dm_old_state =
to_dm_connector_state(connector->state);
struct dm_connector_state *dm_new_state =
to_dm_connector_state(connector_state);
int ret = -EINVAL;
if (property == dev->mode_config.scaling_mode_property) {
enum amdgpu_rmx_type rmx_type;
switch (val) {
case DRM_MODE_SCALE_CENTER:
rmx_type = RMX_CENTER;
break;
case DRM_MODE_SCALE_ASPECT:
rmx_type = RMX_ASPECT;
break;
case DRM_MODE_SCALE_FULLSCREEN:
rmx_type = RMX_FULL;
break;
case DRM_MODE_SCALE_NONE:
default:
rmx_type = RMX_OFF;
break;
}
if (dm_old_state->scaling == rmx_type)
return 0;
dm_new_state->scaling = rmx_type;
ret = 0;
} else if (property == adev->mode_info.underscan_hborder_property) {
dm_new_state->underscan_hborder = val;
ret = 0;
} else if (property == adev->mode_info.underscan_vborder_property) {
dm_new_state->underscan_vborder = val;
ret = 0;
} else if (property == adev->mode_info.underscan_property) {
dm_new_state->underscan_enable = val;
ret = 0;
} else if (property == adev->mode_info.abm_level_property) {
dm_new_state->abm_level = val;
ret = 0;
}
return ret;
}
int amdgpu_dm_connector_atomic_get_property(struct drm_connector *connector,
const struct drm_connector_state *state,
struct drm_property *property,
uint64_t *val)
{
struct drm_device *dev = connector->dev;
struct amdgpu_device *adev = dev->dev_private;
struct dm_connector_state *dm_state =
to_dm_connector_state(state);
int ret = -EINVAL;
if (property == dev->mode_config.scaling_mode_property) {
switch (dm_state->scaling) {
case RMX_CENTER:
*val = DRM_MODE_SCALE_CENTER;
break;
case RMX_ASPECT:
*val = DRM_MODE_SCALE_ASPECT;
break;
case RMX_FULL:
*val = DRM_MODE_SCALE_FULLSCREEN;
break;
case RMX_OFF:
default:
*val = DRM_MODE_SCALE_NONE;
break;
}
ret = 0;
} else if (property == adev->mode_info.underscan_hborder_property) {
*val = dm_state->underscan_hborder;
ret = 0;
} else if (property == adev->mode_info.underscan_vborder_property) {
*val = dm_state->underscan_vborder;
ret = 0;
} else if (property == adev->mode_info.underscan_property) {
*val = dm_state->underscan_enable;
ret = 0;
} else if (property == adev->mode_info.abm_level_property) {
*val = dm_state->abm_level;
ret = 0;
}
return ret;
}
static void amdgpu_dm_connector_unregister(struct drm_connector *connector)
{
struct amdgpu_dm_connector *amdgpu_dm_connector = to_amdgpu_dm_connector(connector);
drm_dp_aux_unregister(&amdgpu_dm_connector->dm_dp_aux.aux);
}
static void amdgpu_dm_connector_destroy(struct drm_connector *connector)
{
struct amdgpu_dm_connector *aconnector = to_amdgpu_dm_connector(connector);
const struct dc_link *link = aconnector->dc_link;
struct amdgpu_device *adev = connector->dev->dev_private;
struct amdgpu_display_manager *dm = &adev->dm;
/*
* Call only if mst_mgr was iniitalized before since it's not done
* for all connector types.
*/
if (aconnector->mst_mgr.dev)
drm_dp_mst_topology_mgr_destroy(&aconnector->mst_mgr);
#if defined(CONFIG_BACKLIGHT_CLASS_DEVICE) ||\
defined(CONFIG_BACKLIGHT_CLASS_DEVICE_MODULE)
if ((link->connector_signal & (SIGNAL_TYPE_EDP | SIGNAL_TYPE_LVDS)) &&
link->type != dc_connection_none &&
dm->backlight_dev) {
backlight_device_unregister(dm->backlight_dev);
dm->backlight_dev = NULL;
}
#endif
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (aconnector->dc_em_sink)
dc_sink_release(aconnector->dc_em_sink);
aconnector->dc_em_sink = NULL;
if (aconnector->dc_sink)
dc_sink_release(aconnector->dc_sink);
aconnector->dc_sink = NULL;
drm_dp_cec_unregister_connector(&aconnector->dm_dp_aux.aux);
drm_connector_unregister(connector);
drm_connector_cleanup(connector);
if (aconnector->i2c) {
i2c_del_adapter(&aconnector->i2c->base);
kfree(aconnector->i2c);
}
kfree(connector);
}
void amdgpu_dm_connector_funcs_reset(struct drm_connector *connector)
{
struct dm_connector_state *state =
to_dm_connector_state(connector->state);
if (connector->state)
__drm_atomic_helper_connector_destroy_state(connector->state);
kfree(state);
state = kzalloc(sizeof(*state), GFP_KERNEL);
if (state) {
state->scaling = RMX_OFF;
state->underscan_enable = false;
state->underscan_hborder = 0;
state->underscan_vborder = 0;
state->base.max_requested_bpc = 8;
if (connector->connector_type == DRM_MODE_CONNECTOR_eDP)
state->abm_level = amdgpu_dm_abm_level;
__drm_atomic_helper_connector_reset(connector, &state->base);
}
}
struct drm_connector_state *
amdgpu_dm_connector_atomic_duplicate_state(struct drm_connector *connector)
{
struct dm_connector_state *state =
to_dm_connector_state(connector->state);
struct dm_connector_state *new_state =
kmemdup(state, sizeof(*state), GFP_KERNEL);
if (!new_state)
return NULL;
__drm_atomic_helper_connector_duplicate_state(connector, &new_state->base);
new_state->freesync_capable = state->freesync_capable;
new_state->abm_level = state->abm_level;
new_state->scaling = state->scaling;
new_state->underscan_enable = state->underscan_enable;
new_state->underscan_hborder = state->underscan_hborder;
new_state->underscan_vborder = state->underscan_vborder;
return &new_state->base;
}
static const struct drm_connector_funcs amdgpu_dm_connector_funcs = {
.reset = amdgpu_dm_connector_funcs_reset,
.detect = amdgpu_dm_connector_detect,
.fill_modes = drm_helper_probe_single_connector_modes,
.destroy = amdgpu_dm_connector_destroy,
.atomic_duplicate_state = amdgpu_dm_connector_atomic_duplicate_state,
.atomic_destroy_state = drm_atomic_helper_connector_destroy_state,
.atomic_set_property = amdgpu_dm_connector_atomic_set_property,
.atomic_get_property = amdgpu_dm_connector_atomic_get_property,
.early_unregister = amdgpu_dm_connector_unregister
};
static int get_modes(struct drm_connector *connector)
{
return amdgpu_dm_connector_get_modes(connector);
}
static void create_eml_sink(struct amdgpu_dm_connector *aconnector)
{
struct dc_sink_init_data init_params = {
.link = aconnector->dc_link,
.sink_signal = SIGNAL_TYPE_VIRTUAL
};
struct edid *edid;
if (!aconnector->base.edid_blob_ptr) {
DRM_ERROR("No EDID firmware found on connector: %s ,forcing to OFF!\n",
aconnector->base.name);
aconnector->base.force = DRM_FORCE_OFF;
aconnector->base.override_edid = false;
return;
}
edid = (struct edid *) aconnector->base.edid_blob_ptr->data;
aconnector->edid = edid;
aconnector->dc_em_sink = dc_link_add_remote_sink(
aconnector->dc_link,
(uint8_t *)edid,
(edid->extensions + 1) * EDID_LENGTH,
&init_params);
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
if (aconnector->base.force == DRM_FORCE_ON) {
aconnector->dc_sink = aconnector->dc_link->local_sink ?
aconnector->dc_link->local_sink :
aconnector->dc_em_sink;
drm/amd/display: Fix reference counting for struct dc_sink. Reference counting in amdgpu_dm_connector for amdgpu_dm_connector::dc_sink and amdgpu_dm_connector::dc_em_sink as well as in dc_link::local_sink seems to be out of shape. Thus make reference counting consistent for these members and just plain increment the reference count when the variable gets assigned and decrement when the pointer is set to zero or replaced. Also simplify reference counting in selected function sopes to be sure the reference is released in any case. In some cases add NULL pointer check before dereferencing. At a hand full of places a comment is placed to stat that the reference increment happened already somewhere else. This actually fixes the following kernel bug on my system when enabling display core in amdgpu. There are some more similar bug reports around, so it probably helps at more places. kernel BUG at mm/slub.c:294! invalid opcode: 0000 [#1] SMP PTI CPU: 9 PID: 1180 Comm: Xorg Not tainted 5.0.0-rc1+ #2 Hardware name: Supermicro X10DAi/X10DAI, BIOS 3.0a 02/05/2018 RIP: 0010:__slab_free+0x1e2/0x3d0 Code: 8b 54 24 30 48 89 4c 24 28 e8 da fb ff ff 4c 8b 54 24 28 85 c0 0f 85 67 fe ff ff 48 8d 65 d8 5b 41 5c 41 5d 41 5e 41 5f 5d c3 <0f> 0b 49 3b 5c 24 28 75 ab 48 8b 44 24 30 49 89 4c 24 28 49 89 44 RSP: 0018:ffffb0978589fa90 EFLAGS: 00010246 RAX: ffff92f12806c400 RBX: 0000000080200019 RCX: ffff92f12806c400 RDX: ffff92f12806c400 RSI: ffffdd6421a01a00 RDI: ffff92ed2f406e80 RBP: ffffb0978589fb40 R08: 0000000000000001 R09: ffffffffc0ee4748 R10: ffff92f12806c400 R11: 0000000000000001 R12: ffffdd6421a01a00 R13: ffff92f12806c400 R14: ffff92ed2f406e80 R15: ffffdd6421a01a20 FS: 00007f4170be0ac0(0000) GS:ffff92ed2fb40000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000562818aaa000 CR3: 000000045745a002 CR4: 00000000003606e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: ? drm_dbg+0x87/0x90 [drm] dc_stream_release+0x28/0x50 [amdgpu] amdgpu_dm_connector_mode_valid+0xb4/0x1f0 [amdgpu] drm_helper_probe_single_connector_modes+0x492/0x6b0 [drm_kms_helper] drm_mode_getconnector+0x457/0x490 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] drm_ioctl_kernel+0xa9/0xf0 [drm] drm_ioctl+0x201/0x3a0 [drm] ? drm_connector_property_set_ioctl+0x60/0x60 [drm] amdgpu_drm_ioctl+0x49/0x80 [amdgpu] do_vfs_ioctl+0xa4/0x630 ? __sys_recvmsg+0x83/0xa0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x16/0x20 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x44/0xa9 RIP: 0033:0x7f417110809b Code: 0f 1e fa 48 8b 05 ed bd 0c 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d bd bd 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffdd8d1c268 EFLAGS: 00000246 ORIG_RAX: 0000000000000010 RAX: ffffffffffffffda RBX: 0000562818a8ebc0 RCX: 00007f417110809b RDX: 00007ffdd8d1c2a0 RSI: 00000000c05064a7 RDI: 0000000000000012 RBP: 00007ffdd8d1c2a0 R08: 0000562819012280 R09: 0000000000000007 R10: 0000000000000000 R11: 0000000000000246 R12: 00000000c05064a7 R13: 0000000000000012 R14: 0000000000000012 R15: 00007ffdd8d1c2a0 Modules linked in: nfsv4 dns_resolver nfs lockd grace fscache fuse vfat fat amdgpu intel_rapl sb_edac x86_pkg_temp_thermal intel_powerclamp coretemp kvm_intel kvm irqbypass crct10dif_pclmul chash gpu_sched crc32_pclmul snd_hda_codec_realtek ghash_clmulni_intel amd_iommu_v2 iTCO_wdt iTCO_vendor_support ttm snd_hda_codec_generic snd_hda_codec_hdmi ledtrig_audio snd_hda_intel drm_kms_helper snd_hda_codec intel_cstate snd_hda_core drm snd_hwdep snd_seq snd_seq_device intel_uncore snd_pcm intel_rapl_perf snd_timer snd soundcore ioatdma pcspkr intel_wmi_thunderbolt mxm_wmi i2c_i801 lpc_ich pcc_cpufreq auth_rpcgss sunrpc igb crc32c_intel i2c_algo_bit dca wmi hid_cherry analog gameport joydev This patch is based on agd5f/drm-next-5.1-wip. This patch does not require all of that, but agd5f/drm-next-5.1-wip contains at least one more dc_sink counting fix that I could spot. Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-02-10 03:13:01 -07:00
dc_sink_retain(aconnector->dc_sink);
}
}
static void handle_edid_mgmt(struct amdgpu_dm_connector *aconnector)
{
struct dc_link *link = (struct dc_link *)aconnector->dc_link;
/*
* In case of headless boot with force on for DP managed connector
* Those settings have to be != 0 to get initial modeset
*/
if (link->connector_signal == SIGNAL_TYPE_DISPLAY_PORT) {
link->verified_link_cap.lane_count = LANE_COUNT_FOUR;
link->verified_link_cap.link_rate = LINK_RATE_HIGH2;
}
aconnector->base.override_edid = true;
create_eml_sink(aconnector);
}
enum drm_mode_status amdgpu_dm_connector_mode_valid(struct drm_connector *connector,
struct drm_display_mode *mode)
{
int result = MODE_ERROR;
struct dc_sink *dc_sink;
struct amdgpu_device *adev = connector->dev->dev_private;
/* TODO: Unhardcode stream count */
struct dc_stream_state *stream;
struct amdgpu_dm_connector *aconnector = to_amdgpu_dm_connector(connector);
enum dc_status dc_result = DC_OK;
if ((mode->flags & DRM_MODE_FLAG_INTERLACE) ||
(mode->flags & DRM_MODE_FLAG_DBLSCAN))
return result;
/*
* Only run this the first time mode_valid is called to initilialize
* EDID mgmt
*/
if (aconnector->base.force != DRM_FORCE_UNSPECIFIED &&
!aconnector->dc_em_sink)
handle_edid_mgmt(aconnector);
dc_sink = to_amdgpu_dm_connector(connector)->dc_sink;
if (dc_sink == NULL) {
DRM_ERROR("dc_sink is NULL!\n");
goto fail;
}
stream = create_stream_for_sink(aconnector, mode, NULL, NULL);
if (stream == NULL) {
DRM_ERROR("Failed to create stream for sink!\n");
goto fail;
}
dc_result = dc_validate_stream(adev->dm.dc, stream);
if (dc_result == DC_OK)
result = MODE_OK;
else
DRM_DEBUG_KMS("Mode %dx%d (clk %d) failed DC validation with error %d\n",
mode->vdisplay,
mode->hdisplay,
mode->clock,
dc_result);
dc_stream_release(stream);
fail:
/* TODO: error handling*/
return result;
}
static int fill_hdr_info_packet(const struct drm_connector_state *state,
struct dc_info_packet *out)
{
struct hdmi_drm_infoframe frame;
unsigned char buf[30]; /* 26 + 4 */
ssize_t len;
int ret, i;
memset(out, 0, sizeof(*out));
if (!state->hdr_output_metadata)
return 0;
ret = drm_hdmi_infoframe_set_hdr_metadata(&frame, state);
if (ret)
return ret;
len = hdmi_drm_infoframe_pack_only(&frame, buf, sizeof(buf));
if (len < 0)
return (int)len;
/* Static metadata is a fixed 26 bytes + 4 byte header. */
if (len != 30)
return -EINVAL;
/* Prepare the infopacket for DC. */
switch (state->connector->connector_type) {
case DRM_MODE_CONNECTOR_HDMIA:
out->hb0 = 0x87; /* type */
out->hb1 = 0x01; /* version */
out->hb2 = 0x1A; /* length */
out->sb[0] = buf[3]; /* checksum */
i = 1;
break;
case DRM_MODE_CONNECTOR_DisplayPort:
case DRM_MODE_CONNECTOR_eDP:
out->hb0 = 0x00; /* sdp id, zero */
out->hb1 = 0x87; /* type */
out->hb2 = 0x1D; /* payload len - 1 */
out->hb3 = (0x13 << 2); /* sdp version */
out->sb[0] = 0x01; /* version */
out->sb[1] = 0x1A; /* length */
i = 2;
break;
default:
return -EINVAL;
}
memcpy(&out->sb[i], &buf[4], 26);
out->valid = true;
print_hex_dump(KERN_DEBUG, "HDR SB:", DUMP_PREFIX_NONE, 16, 1, out->sb,
sizeof(out->sb), false);
return 0;
}
static bool
is_hdr_metadata_different(const struct drm_connector_state *old_state,
const struct drm_connector_state *new_state)
{
struct drm_property_blob *old_blob = old_state->hdr_output_metadata;
struct drm_property_blob *new_blob = new_state->hdr_output_metadata;
if (old_blob != new_blob) {
if (old_blob && new_blob &&
old_blob->length == new_blob->length)
return memcmp(old_blob->data, new_blob->data,
old_blob->length);
return true;
}
return false;
}
static int
amdgpu_dm_connector_atomic_check(struct drm_connector *conn,
struct drm_atomic_state *state)
{
struct drm_connector_state *new_con_state =
drm_atomic_get_new_connector_state(state, conn);
struct drm_connector_state *old_con_state =
drm_atomic_get_old_connector_state(state, conn);
struct drm_crtc *crtc = new_con_state->crtc;
struct drm_crtc_state *new_crtc_state;
int ret;
if (!crtc)
return 0;
if (is_hdr_metadata_different(old_con_state, new_con_state)) {
struct dc_info_packet hdr_infopacket;
ret = fill_hdr_info_packet(new_con_state, &hdr_infopacket);
if (ret)
return ret;
new_crtc_state = drm_atomic_get_crtc_state(state, crtc);
if (IS_ERR(new_crtc_state))
return PTR_ERR(new_crtc_state);
/*
* DC considers the stream backends changed if the
* static metadata changes. Forcing the modeset also
* gives a simple way for userspace to switch from
* 8bpc to 10bpc when setting the metadata to enter
* or exit HDR.
*
* Changing the static metadata after it's been
* set is permissible, however. So only force a
* modeset if we're entering or exiting HDR.
*/
new_crtc_state->mode_changed =
!old_con_state->hdr_output_metadata ||
!new_con_state->hdr_output_metadata;
}
return 0;
}
static const struct drm_connector_helper_funcs
amdgpu_dm_connector_helper_funcs = {
/*
* If hotplugging a second bigger display in FB Con mode, bigger resolution
* modes will be filtered by drm_mode_validate_size(), and those modes
* are missing after user start lightdm. So we need to renew modes list.
* in get_modes call back, not just return the modes count
*/
.get_modes = get_modes,
.mode_valid = amdgpu_dm_connector_mode_valid,
.atomic_check = amdgpu_dm_connector_atomic_check,
};
static void dm_crtc_helper_disable(struct drm_crtc *crtc)
{
}
static int count_crtc_active_planes(struct drm_crtc_state *new_crtc_state)
{
struct drm_atomic_state *state = new_crtc_state->state;
struct drm_plane *plane;
int num_active = 0;
drm_for_each_plane_mask(plane, state->dev, new_crtc_state->plane_mask) {
struct drm_plane_state *new_plane_state;
/* Cursor planes are "fake". */
if (plane->type == DRM_PLANE_TYPE_CURSOR)
continue;
new_plane_state = drm_atomic_get_new_plane_state(state, plane);
if (!new_plane_state) {
/*
* The plane is enable on the CRTC and hasn't changed
* state. This means that it previously passed
* validation and is therefore enabled.
*/
num_active += 1;
continue;
}
/* We need a framebuffer to be considered enabled. */
num_active += (new_plane_state->fb != NULL);
}
return num_active;
}
/*
* Sets whether interrupts should be enabled on a specific CRTC.
* We require that the stream be enabled and that there exist active
* DC planes on the stream.
*/
static void
dm_update_crtc_interrupt_state(struct drm_crtc *crtc,
struct drm_crtc_state *new_crtc_state)
{
struct dm_crtc_state *dm_new_crtc_state =
to_dm_crtc_state(new_crtc_state);
dm_new_crtc_state->active_planes = 0;
dm_new_crtc_state->interrupts_enabled = false;
if (!dm_new_crtc_state->stream)
return;
dm_new_crtc_state->active_planes =
count_crtc_active_planes(new_crtc_state);
dm_new_crtc_state->interrupts_enabled =
dm_new_crtc_state->active_planes > 0;
}
static int dm_crtc_helper_atomic_check(struct drm_crtc *crtc,
struct drm_crtc_state *state)
{
struct amdgpu_device *adev = crtc->dev->dev_private;
struct dc *dc = adev->dm.dc;
struct dm_crtc_state *dm_crtc_state = to_dm_crtc_state(state);
int ret = -EINVAL;
/*
* Update interrupt state for the CRTC. This needs to happen whenever
* the CRTC has changed or whenever any of its planes have changed.
* Atomic check satisfies both of these requirements since the CRTC
* is added to the state by DRM during drm_atomic_helper_check_planes.
*/
dm_update_crtc_interrupt_state(crtc, state);
if (unlikely(!dm_crtc_state->stream &&
modeset_required(state, NULL, dm_crtc_state->stream))) {
WARN_ON(1);
return ret;
}
/*
* We require the primary plane to be enabled whenever the CRTC is, otherwise
* drm_mode_cursor_universal may end up trying to enable the cursor plane while all other
* planes are disabled, which is not supported by the hardware. And there is legacy
* userspace which stops using the HW cursor altogether in response to the resulting EINVAL.
*/
if (state->enable &&
!(state->plane_mask & drm_plane_mask(crtc->primary)))
return -EINVAL;
/* In some use cases, like reset, no stream is attached */
if (!dm_crtc_state->stream)
return 0;
if (dc_validate_stream(dc, dm_crtc_state->stream) == DC_OK)
return 0;
return ret;
}
static bool dm_crtc_helper_mode_fixup(struct drm_crtc *crtc,
const struct drm_display_mode *mode,
struct drm_display_mode *adjusted_mode)
{
return true;
}
static const struct drm_crtc_helper_funcs amdgpu_dm_crtc_helper_funcs = {
.disable = dm_crtc_helper_disable,
.atomic_check = dm_crtc_helper_atomic_check,
.mode_fixup = dm_crtc_helper_mode_fixup
};
static void dm_encoder_helper_disable(struct drm_encoder *encoder)
{
}
static int dm_encoder_helper_atomic_check(struct drm_encoder *encoder,
struct drm_crtc_state *crtc_state,
struct drm_connector_state *conn_state)
{
return 0;
}
const struct drm_encoder_helper_funcs amdgpu_dm_encoder_helper_funcs = {
.disable = dm_encoder_helper_disable,
.atomic_check = dm_encoder_helper_atomic_check
};
static void dm_drm_plane_reset(struct drm_plane *plane)
{
struct dm_plane_state *amdgpu_state = NULL;
if (plane->state)
plane->funcs->atomic_destroy_state(plane, plane->state);
amdgpu_state = kzalloc(sizeof(*amdgpu_state), GFP_KERNEL);
WARN_ON(amdgpu_state == NULL);
if (amdgpu_state)
__drm_atomic_helper_plane_reset(plane, &amdgpu_state->base);
}
static struct drm_plane_state *
dm_drm_plane_duplicate_state(struct drm_plane *plane)
{
struct dm_plane_state *dm_plane_state, *old_dm_plane_state;
old_dm_plane_state = to_dm_plane_state(plane->state);
dm_plane_state = kzalloc(sizeof(*dm_plane_state), GFP_KERNEL);
if (!dm_plane_state)
return NULL;
__drm_atomic_helper_plane_duplicate_state(plane, &dm_plane_state->base);
if (old_dm_plane_state->dc_state) {
dm_plane_state->dc_state = old_dm_plane_state->dc_state;
dc_plane_state_retain(dm_plane_state->dc_state);
}
return &dm_plane_state->base;
}
void dm_drm_plane_destroy_state(struct drm_plane *plane,
struct drm_plane_state *state)
{
struct dm_plane_state *dm_plane_state = to_dm_plane_state(state);
if (dm_plane_state->dc_state)
dc_plane_state_release(dm_plane_state->dc_state);
drm_atomic_helper_plane_destroy_state(plane, state);
}
static const struct drm_plane_funcs dm_plane_funcs = {
.update_plane = drm_atomic_helper_update_plane,
.disable_plane = drm_atomic_helper_disable_plane,
.destroy = drm_primary_helper_destroy,
.reset = dm_drm_plane_reset,
.atomic_duplicate_state = dm_drm_plane_duplicate_state,
.atomic_destroy_state = dm_drm_plane_destroy_state,
};
static int dm_plane_helper_prepare_fb(struct drm_plane *plane,
struct drm_plane_state *new_state)
{
struct amdgpu_framebuffer *afb;
struct drm_gem_object *obj;
struct amdgpu_device *adev;
struct amdgpu_bo *rbo;
struct dm_plane_state *dm_plane_state_new, *dm_plane_state_old;
struct list_head list;
struct ttm_validate_buffer tv;
struct ww_acquire_ctx ticket;
uint64_t tiling_flags;
uint32_t domain;
int r;
bool force_disable_dcc = false;
dm_plane_state_old = to_dm_plane_state(plane->state);
dm_plane_state_new = to_dm_plane_state(new_state);
if (!new_state->fb) {
DRM_DEBUG_DRIVER("No FB bound\n");
return 0;
}
afb = to_amdgpu_framebuffer(new_state->fb);
obj = new_state->fb->obj[0];
rbo = gem_to_amdgpu_bo(obj);
adev = amdgpu_ttm_adev(rbo->tbo.bdev);
INIT_LIST_HEAD(&list);
tv.bo = &rbo->tbo;
tv.num_shared = 1;
list_add(&tv.head, &list);
r = ttm_eu_reserve_buffers(&ticket, &list, false, NULL, true);
if (r) {
dev_err(adev->dev, "fail to reserve bo (%d)\n", r);
return r;
}
if (plane->type != DRM_PLANE_TYPE_CURSOR)
domain = amdgpu_display_supported_domains(adev, rbo->flags);
else
domain = AMDGPU_GEM_DOMAIN_VRAM;
r = amdgpu_bo_pin(rbo, domain);
if (unlikely(r != 0)) {
if (r != -ERESTARTSYS)
DRM_ERROR("Failed to pin framebuffer with error %d\n", r);
ttm_eu_backoff_reservation(&ticket, &list);
return r;
}
r = amdgpu_ttm_alloc_gart(&rbo->tbo);
if (unlikely(r != 0)) {
amdgpu_bo_unpin(rbo);
ttm_eu_backoff_reservation(&ticket, &list);
DRM_ERROR("%p bind failed\n", rbo);
return r;
}
amdgpu_bo_get_tiling_flags(rbo, &tiling_flags);
ttm_eu_backoff_reservation(&ticket, &list);
afb->address = amdgpu_bo_gpu_offset(rbo);
amdgpu_bo_ref(rbo);
if (dm_plane_state_new->dc_state &&
dm_plane_state_old->dc_state != dm_plane_state_new->dc_state) {
struct dc_plane_state *plane_state = dm_plane_state_new->dc_state;
force_disable_dcc = adev->asic_type == CHIP_RAVEN && adev->in_suspend;
fill_plane_buffer_attributes(
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
adev, afb, plane_state->format, plane_state->rotation,
tiling_flags, &plane_state->tiling_info,
&plane_state->plane_size, &plane_state->dcc,
&plane_state->address,
force_disable_dcc);
}
return 0;
}
static void dm_plane_helper_cleanup_fb(struct drm_plane *plane,
struct drm_plane_state *old_state)
{
struct amdgpu_bo *rbo;
int r;
if (!old_state->fb)
return;
rbo = gem_to_amdgpu_bo(old_state->fb->obj[0]);
r = amdgpu_bo_reserve(rbo, false);
if (unlikely(r)) {
DRM_ERROR("failed to reserve rbo before unpin\n");
return;
}
amdgpu_bo_unpin(rbo);
amdgpu_bo_unreserve(rbo);
amdgpu_bo_unref(&rbo);
}
static int dm_plane_atomic_check(struct drm_plane *plane,
struct drm_plane_state *state)
{
struct amdgpu_device *adev = plane->dev->dev_private;
struct dc *dc = adev->dm.dc;
struct dm_plane_state *dm_plane_state;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
struct dc_scaling_info scaling_info;
int ret;
dm_plane_state = to_dm_plane_state(state);
if (!dm_plane_state->dc_state)
return 0;
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
ret = fill_dc_scaling_info(state, &scaling_info);
if (ret)
return ret;
if (dc_validate_plane(dc, dm_plane_state->dc_state) == DC_OK)
return 0;
return -EINVAL;
}
static int dm_plane_atomic_async_check(struct drm_plane *plane,
struct drm_plane_state *new_plane_state)
{
/* Only support async updates on cursor planes. */
if (plane->type != DRM_PLANE_TYPE_CURSOR)
return -EINVAL;
return 0;
}
static void dm_plane_atomic_async_update(struct drm_plane *plane,
struct drm_plane_state *new_state)
{
struct drm_plane_state *old_state =
drm_atomic_get_old_plane_state(new_state->state, plane);
swap(plane->state->fb, new_state->fb);
plane->state->src_x = new_state->src_x;
plane->state->src_y = new_state->src_y;
plane->state->src_w = new_state->src_w;
plane->state->src_h = new_state->src_h;
plane->state->crtc_x = new_state->crtc_x;
plane->state->crtc_y = new_state->crtc_y;
plane->state->crtc_w = new_state->crtc_w;
plane->state->crtc_h = new_state->crtc_h;
handle_cursor_update(plane, old_state);
}
static const struct drm_plane_helper_funcs dm_plane_helper_funcs = {
.prepare_fb = dm_plane_helper_prepare_fb,
.cleanup_fb = dm_plane_helper_cleanup_fb,
.atomic_check = dm_plane_atomic_check,
.atomic_async_check = dm_plane_atomic_async_check,
.atomic_async_update = dm_plane_atomic_async_update
};
/*
* TODO: these are currently initialized to rgb formats only.
* For future use cases we should either initialize them dynamically based on
* plane capabilities, or initialize this array to all formats, so internal drm
* check will succeed, and let DC implement proper check
*/
static const uint32_t rgb_formats[] = {
DRM_FORMAT_XRGB8888,
DRM_FORMAT_ARGB8888,
DRM_FORMAT_RGBA8888,
DRM_FORMAT_XRGB2101010,
DRM_FORMAT_XBGR2101010,
DRM_FORMAT_ARGB2101010,
DRM_FORMAT_ABGR2101010,
DRM_FORMAT_XBGR8888,
DRM_FORMAT_ABGR8888,
DRM_FORMAT_RGB565,
};
static const uint32_t overlay_formats[] = {
DRM_FORMAT_XRGB8888,
DRM_FORMAT_ARGB8888,
DRM_FORMAT_RGBA8888,
DRM_FORMAT_XBGR8888,
DRM_FORMAT_ABGR8888,
DRM_FORMAT_RGB565
};
static const u32 cursor_formats[] = {
DRM_FORMAT_ARGB8888
};
static int get_plane_formats(const struct drm_plane *plane,
const struct dc_plane_cap *plane_cap,
uint32_t *formats, int max_formats)
{
int i, num_formats = 0;
/*
* TODO: Query support for each group of formats directly from
* DC plane caps. This will require adding more formats to the
* caps list.
*/
switch (plane->type) {
case DRM_PLANE_TYPE_PRIMARY:
for (i = 0; i < ARRAY_SIZE(rgb_formats); ++i) {
if (num_formats >= max_formats)
break;
formats[num_formats++] = rgb_formats[i];
}
if (plane_cap && plane_cap->pixel_format_support.nv12)
formats[num_formats++] = DRM_FORMAT_NV12;
break;
case DRM_PLANE_TYPE_OVERLAY:
for (i = 0; i < ARRAY_SIZE(overlay_formats); ++i) {
if (num_formats >= max_formats)
break;
formats[num_formats++] = overlay_formats[i];
}
break;
case DRM_PLANE_TYPE_CURSOR:
for (i = 0; i < ARRAY_SIZE(cursor_formats); ++i) {
if (num_formats >= max_formats)
break;
formats[num_formats++] = cursor_formats[i];
}
break;
}
return num_formats;
}
static int amdgpu_dm_plane_init(struct amdgpu_display_manager *dm,
struct drm_plane *plane,
unsigned long possible_crtcs,
const struct dc_plane_cap *plane_cap)
{
uint32_t formats[32];
int num_formats;
int res = -EPERM;
num_formats = get_plane_formats(plane, plane_cap, formats,
ARRAY_SIZE(formats));
res = drm_universal_plane_init(dm->adev->ddev, plane, possible_crtcs,
&dm_plane_funcs, formats, num_formats,
NULL, plane->type, NULL);
if (res)
return res;
if (plane->type == DRM_PLANE_TYPE_OVERLAY &&
plane_cap && plane_cap->per_pixel_alpha) {
unsigned int blend_caps = BIT(DRM_MODE_BLEND_PIXEL_NONE) |
BIT(DRM_MODE_BLEND_PREMULTI);
drm_plane_create_alpha_property(plane);
drm_plane_create_blend_mode_property(plane, blend_caps);
}
if (plane->type == DRM_PLANE_TYPE_PRIMARY &&
plane_cap && plane_cap->pixel_format_support.nv12) {
/* This only affects YUV formats. */
drm_plane_create_color_properties(
plane,
BIT(DRM_COLOR_YCBCR_BT601) |
BIT(DRM_COLOR_YCBCR_BT709),
BIT(DRM_COLOR_YCBCR_LIMITED_RANGE) |
BIT(DRM_COLOR_YCBCR_FULL_RANGE),
DRM_COLOR_YCBCR_BT709, DRM_COLOR_YCBCR_LIMITED_RANGE);
}
drm_plane_helper_add(plane, &dm_plane_helper_funcs);
/* Create (reset) the plane state */
if (plane->funcs->reset)
plane->funcs->reset(plane);
return 0;
}
static int amdgpu_dm_crtc_init(struct amdgpu_display_manager *dm,
struct drm_plane *plane,
uint32_t crtc_index)
{
struct amdgpu_crtc *acrtc = NULL;
struct drm_plane *cursor_plane;
int res = -ENOMEM;
cursor_plane = kzalloc(sizeof(*cursor_plane), GFP_KERNEL);
if (!cursor_plane)
goto fail;
cursor_plane->type = DRM_PLANE_TYPE_CURSOR;
res = amdgpu_dm_plane_init(dm, cursor_plane, 0, NULL);
acrtc = kzalloc(sizeof(struct amdgpu_crtc), GFP_KERNEL);
if (!acrtc)
goto fail;
res = drm_crtc_init_with_planes(
dm->ddev,
&acrtc->base,
plane,
cursor_plane,
&amdgpu_dm_crtc_funcs, NULL);
if (res)
goto fail;
drm_crtc_helper_add(&acrtc->base, &amdgpu_dm_crtc_helper_funcs);
/* Create (reset) the plane state */
if (acrtc->base.funcs->reset)
acrtc->base.funcs->reset(&acrtc->base);
acrtc->max_cursor_width = dm->adev->dm.dc->caps.max_cursor_size;
acrtc->max_cursor_height = dm->adev->dm.dc->caps.max_cursor_size;
acrtc->crtc_id = crtc_index;
acrtc->base.enabled = false;
acrtc->otg_inst = -1;
dm->adev->mode_info.crtcs[crtc_index] = acrtc;
drm_crtc_enable_color_mgmt(&acrtc->base, MAX_COLOR_LUT_ENTRIES,
true, MAX_COLOR_LUT_ENTRIES);
drm_mode_crtc_set_gamma_size(&acrtc->base, MAX_COLOR_LEGACY_LUT_ENTRIES);
return 0;
fail:
kfree(acrtc);
kfree(cursor_plane);
return res;
}
static int to_drm_connector_type(enum signal_type st)
{
switch (st) {
case SIGNAL_TYPE_HDMI_TYPE_A:
return DRM_MODE_CONNECTOR_HDMIA;
case SIGNAL_TYPE_EDP:
return DRM_MODE_CONNECTOR_eDP;
case SIGNAL_TYPE_LVDS:
return DRM_MODE_CONNECTOR_LVDS;
case SIGNAL_TYPE_RGB:
return DRM_MODE_CONNECTOR_VGA;
case SIGNAL_TYPE_DISPLAY_PORT:
case SIGNAL_TYPE_DISPLAY_PORT_MST:
return DRM_MODE_CONNECTOR_DisplayPort;
case SIGNAL_TYPE_DVI_DUAL_LINK:
case SIGNAL_TYPE_DVI_SINGLE_LINK:
return DRM_MODE_CONNECTOR_DVID;
case SIGNAL_TYPE_VIRTUAL:
return DRM_MODE_CONNECTOR_VIRTUAL;
default:
return DRM_MODE_CONNECTOR_Unknown;
}
}
static struct drm_encoder *amdgpu_dm_connector_to_encoder(struct drm_connector *connector)
{
return drm_encoder_find(connector->dev, NULL, connector->encoder_ids[0]);
}
static void amdgpu_dm_get_native_mode(struct drm_connector *connector)
{
struct drm_encoder *encoder;
struct amdgpu_encoder *amdgpu_encoder;
encoder = amdgpu_dm_connector_to_encoder(connector);
if (encoder == NULL)
return;
amdgpu_encoder = to_amdgpu_encoder(encoder);
amdgpu_encoder->native_mode.clock = 0;
if (!list_empty(&connector->probed_modes)) {
struct drm_display_mode *preferred_mode = NULL;
list_for_each_entry(preferred_mode,
&connector->probed_modes,
head) {
if (preferred_mode->type & DRM_MODE_TYPE_PREFERRED)
amdgpu_encoder->native_mode = *preferred_mode;
break;
}
}
}
static struct drm_display_mode *
amdgpu_dm_create_common_mode(struct drm_encoder *encoder,
char *name,
int hdisplay, int vdisplay)
{
struct drm_device *dev = encoder->dev;
struct amdgpu_encoder *amdgpu_encoder = to_amdgpu_encoder(encoder);
struct drm_display_mode *mode = NULL;
struct drm_display_mode *native_mode = &amdgpu_encoder->native_mode;
mode = drm_mode_duplicate(dev, native_mode);
if (mode == NULL)
return NULL;
mode->hdisplay = hdisplay;
mode->vdisplay = vdisplay;
mode->type &= ~DRM_MODE_TYPE_PREFERRED;
strscpy(mode->name, name, DRM_DISPLAY_MODE_LEN);
return mode;
}
static void amdgpu_dm_connector_add_common_modes(struct drm_encoder *encoder,
struct drm_connector *connector)
{
struct amdgpu_encoder *amdgpu_encoder = to_amdgpu_encoder(encoder);
struct drm_display_mode *mode = NULL;
struct drm_display_mode *native_mode = &amdgpu_encoder->native_mode;
struct amdgpu_dm_connector *amdgpu_dm_connector =
to_amdgpu_dm_connector(connector);
int i;
int n;
struct mode_size {
char name[DRM_DISPLAY_MODE_LEN];
int w;
int h;
} common_modes[] = {
{ "640x480", 640, 480},
{ "800x600", 800, 600},
{ "1024x768", 1024, 768},
{ "1280x720", 1280, 720},
{ "1280x800", 1280, 800},
{"1280x1024", 1280, 1024},
{ "1440x900", 1440, 900},
{"1680x1050", 1680, 1050},
{"1600x1200", 1600, 1200},
{"1920x1080", 1920, 1080},
{"1920x1200", 1920, 1200}
};
n = ARRAY_SIZE(common_modes);
for (i = 0; i < n; i++) {
struct drm_display_mode *curmode = NULL;
bool mode_existed = false;
if (common_modes[i].w > native_mode->hdisplay ||
common_modes[i].h > native_mode->vdisplay ||
(common_modes[i].w == native_mode->hdisplay &&
common_modes[i].h == native_mode->vdisplay))
continue;
list_for_each_entry(curmode, &connector->probed_modes, head) {
if (common_modes[i].w == curmode->hdisplay &&
common_modes[i].h == curmode->vdisplay) {
mode_existed = true;
break;
}
}
if (mode_existed)
continue;
mode = amdgpu_dm_create_common_mode(encoder,
common_modes[i].name, common_modes[i].w,
common_modes[i].h);
drm_mode_probed_add(connector, mode);
amdgpu_dm_connector->num_modes++;
}
}
static void amdgpu_dm_connector_ddc_get_modes(struct drm_connector *connector,
struct edid *edid)
{
struct amdgpu_dm_connector *amdgpu_dm_connector =
to_amdgpu_dm_connector(connector);
if (edid) {
/* empty probed_modes */
INIT_LIST_HEAD(&connector->probed_modes);
amdgpu_dm_connector->num_modes =
drm_add_edid_modes(connector, edid);
/* sorting the probed modes before calling function
* amdgpu_dm_get_native_mode() since EDID can have
* more than one preferred mode. The modes that are
* later in the probed mode list could be of higher
* and preferred resolution. For example, 3840x2160
* resolution in base EDID preferred timing and 4096x2160
* preferred resolution in DID extension block later.
*/
drm_mode_sort(&connector->probed_modes);
amdgpu_dm_get_native_mode(connector);
} else {
amdgpu_dm_connector->num_modes = 0;
}
}
static int amdgpu_dm_connector_get_modes(struct drm_connector *connector)
{
struct amdgpu_dm_connector *amdgpu_dm_connector =
to_amdgpu_dm_connector(connector);
struct drm_encoder *encoder;
struct edid *edid = amdgpu_dm_connector->edid;
encoder = amdgpu_dm_connector_to_encoder(connector);
if (!edid || !drm_edid_is_valid(edid)) {
amdgpu_dm_connector->num_modes =
drm_add_modes_noedid(connector, 640, 480);
} else {
amdgpu_dm_connector_ddc_get_modes(connector, edid);
amdgpu_dm_connector_add_common_modes(encoder, connector);
}
amdgpu_dm_fbc_init(connector);
return amdgpu_dm_connector->num_modes;
}
void amdgpu_dm_connector_init_helper(struct amdgpu_display_manager *dm,
struct amdgpu_dm_connector *aconnector,
int connector_type,
struct dc_link *link,
int link_index)
{
struct amdgpu_device *adev = dm->ddev->dev_private;
/*
* Some of the properties below require access to state, like bpc.
* Allocate some default initial connector state with our reset helper.
*/
if (aconnector->base.funcs->reset)
aconnector->base.funcs->reset(&aconnector->base);
aconnector->connector_id = link_index;
aconnector->dc_link = link;
aconnector->base.interlace_allowed = false;
aconnector->base.doublescan_allowed = false;
aconnector->base.stereo_allowed = false;
aconnector->base.dpms = DRM_MODE_DPMS_OFF;
aconnector->hpd.hpd = AMDGPU_HPD_NONE; /* not used */
aconnector->audio_inst = -1;
mutex_init(&aconnector->hpd_lock);
/*
* configure support HPD hot plug connector_>polled default value is 0
* which means HPD hot plug not supported
*/
switch (connector_type) {
case DRM_MODE_CONNECTOR_HDMIA:
aconnector->base.polled = DRM_CONNECTOR_POLL_HPD;
aconnector->base.ycbcr_420_allowed =
link->link_enc->features.hdmi_ycbcr420_supported ? true : false;
break;
case DRM_MODE_CONNECTOR_DisplayPort:
aconnector->base.polled = DRM_CONNECTOR_POLL_HPD;
aconnector->base.ycbcr_420_allowed =
link->link_enc->features.dp_ycbcr420_supported ? true : false;
break;
case DRM_MODE_CONNECTOR_DVID:
aconnector->base.polled = DRM_CONNECTOR_POLL_HPD;
break;
default:
break;
}
drm_object_attach_property(&aconnector->base.base,
dm->ddev->mode_config.scaling_mode_property,
DRM_MODE_SCALE_NONE);
drm_object_attach_property(&aconnector->base.base,
adev->mode_info.underscan_property,
UNDERSCAN_OFF);
drm_object_attach_property(&aconnector->base.base,
adev->mode_info.underscan_hborder_property,
0);
drm_object_attach_property(&aconnector->base.base,
adev->mode_info.underscan_vborder_property,
0);
drm_connector_attach_max_bpc_property(&aconnector->base, 8, 16);
/* This defaults to the max in the range, but we want 8bpc. */
aconnector->base.state->max_bpc = 8;
aconnector->base.state->max_requested_bpc = 8;
if (connector_type == DRM_MODE_CONNECTOR_eDP &&
dc_is_dmcu_initialized(adev->dm.dc)) {
drm_object_attach_property(&aconnector->base.base,
adev->mode_info.abm_level_property, 0);
}
if (connector_type == DRM_MODE_CONNECTOR_HDMIA ||
connector_type == DRM_MODE_CONNECTOR_DisplayPort ||
connector_type == DRM_MODE_CONNECTOR_eDP) {
drm_object_attach_property(
&aconnector->base.base,
dm->ddev->mode_config.hdr_output_metadata_property, 0);
drm_connector_attach_vrr_capable_property(
&aconnector->base);
}
}
static int amdgpu_dm_i2c_xfer(struct i2c_adapter *i2c_adap,
struct i2c_msg *msgs, int num)
{
struct amdgpu_i2c_adapter *i2c = i2c_get_adapdata(i2c_adap);
struct ddc_service *ddc_service = i2c->ddc_service;
struct i2c_command cmd;
int i;
int result = -EIO;
cmd.payloads = kcalloc(num, sizeof(struct i2c_payload), GFP_KERNEL);
if (!cmd.payloads)
return result;
cmd.number_of_payloads = num;
cmd.engine = I2C_COMMAND_ENGINE_DEFAULT;
cmd.speed = 100;
for (i = 0; i < num; i++) {
cmd.payloads[i].write = !(msgs[i].flags & I2C_M_RD);
cmd.payloads[i].address = msgs[i].addr;
cmd.payloads[i].length = msgs[i].len;
cmd.payloads[i].data = msgs[i].buf;
}
if (dc_submit_i2c(
ddc_service->ctx->dc,
ddc_service->ddc_pin->hw_info.ddc_channel,
&cmd))
result = num;
kfree(cmd.payloads);
return result;
}
static u32 amdgpu_dm_i2c_func(struct i2c_adapter *adap)
{
return I2C_FUNC_I2C | I2C_FUNC_SMBUS_EMUL;
}
static const struct i2c_algorithm amdgpu_dm_i2c_algo = {
.master_xfer = amdgpu_dm_i2c_xfer,
.functionality = amdgpu_dm_i2c_func,
};
static struct amdgpu_i2c_adapter *
create_i2c(struct ddc_service *ddc_service,
int link_index,
int *res)
{
struct amdgpu_device *adev = ddc_service->ctx->driver_context;
struct amdgpu_i2c_adapter *i2c;
i2c = kzalloc(sizeof(struct amdgpu_i2c_adapter), GFP_KERNEL);
if (!i2c)
return NULL;
i2c->base.owner = THIS_MODULE;
i2c->base.class = I2C_CLASS_DDC;
i2c->base.dev.parent = &adev->pdev->dev;
i2c->base.algo = &amdgpu_dm_i2c_algo;
snprintf(i2c->base.name, sizeof(i2c->base.name), "AMDGPU DM i2c hw bus %d", link_index);
i2c_set_adapdata(&i2c->base, i2c);
i2c->ddc_service = ddc_service;
i2c->ddc_service->ddc_pin->hw_info.ddc_channel = link_index;
return i2c;
}
/*
* Note: this function assumes that dc_link_detect() was called for the
* dc_link which will be represented by this aconnector.
*/
static int amdgpu_dm_connector_init(struct amdgpu_display_manager *dm,
struct amdgpu_dm_connector *aconnector,
uint32_t link_index,
struct amdgpu_encoder *aencoder)
{
int res = 0;
int connector_type;
struct dc *dc = dm->dc;
struct dc_link *link = dc_get_link_at_index(dc, link_index);
struct amdgpu_i2c_adapter *i2c;
link->priv = aconnector;
DRM_DEBUG_DRIVER("%s()\n", __func__);
i2c = create_i2c(link->ddc, link->link_index, &res);
if (!i2c) {
DRM_ERROR("Failed to create i2c adapter data\n");
return -ENOMEM;
}
aconnector->i2c = i2c;
res = i2c_add_adapter(&i2c->base);
if (res) {
DRM_ERROR("Failed to register hw i2c %d\n", link->link_index);
goto out_free;
}
connector_type = to_drm_connector_type(link->connector_signal);
res = drm_connector_init(
dm->ddev,
&aconnector->base,
&amdgpu_dm_connector_funcs,
connector_type);
if (res) {
DRM_ERROR("connector_init failed\n");
aconnector->connector_id = -1;
goto out_free;
}
drm_connector_helper_add(
&aconnector->base,
&amdgpu_dm_connector_helper_funcs);
amdgpu_dm_connector_init_helper(
dm,
aconnector,
connector_type,
link,
link_index);
drm_connector_attach_encoder(
&aconnector->base, &aencoder->base);
drm_connector_register(&aconnector->base);
#if defined(CONFIG_DEBUG_FS)
connector_debugfs_init(aconnector);
aconnector->debugfs_dpcd_address = 0;
aconnector->debugfs_dpcd_size = 0;
#endif
if (connector_type == DRM_MODE_CONNECTOR_DisplayPort
|| connector_type == DRM_MODE_CONNECTOR_eDP)
amdgpu_dm_initialize_dp_connector(dm, aconnector);
out_free:
if (res) {
kfree(i2c);
aconnector->i2c = NULL;
}
return res;
}
int amdgpu_dm_get_encoder_crtc_mask(struct amdgpu_device *adev)
{
switch (adev->mode_info.num_crtc) {
case 1:
return 0x1;
case 2:
return 0x3;
case 3:
return 0x7;
case 4:
return 0xf;
case 5:
return 0x1f;
case 6:
default:
return 0x3f;
}
}
static int amdgpu_dm_encoder_init(struct drm_device *dev,
struct amdgpu_encoder *aencoder,
uint32_t link_index)
{
struct amdgpu_device *adev = dev->dev_private;
int res = drm_encoder_init(dev,
&aencoder->base,
&amdgpu_dm_encoder_funcs,
DRM_MODE_ENCODER_TMDS,
NULL);
aencoder->base.possible_crtcs = amdgpu_dm_get_encoder_crtc_mask(adev);
if (!res)
aencoder->encoder_id = link_index;
else
aencoder->encoder_id = -1;
drm_encoder_helper_add(&aencoder->base, &amdgpu_dm_encoder_helper_funcs);
return res;
}
static void manage_dm_interrupts(struct amdgpu_device *adev,
struct amdgpu_crtc *acrtc,
bool enable)
{
/*
* this is not correct translation but will work as soon as VBLANK
* constant is the same as PFLIP
*/
int irq_type =
amdgpu_display_crtc_idx_to_irq_type(
adev,
acrtc->crtc_id);
if (enable) {
drm_crtc_vblank_on(&acrtc->base);
amdgpu_irq_get(
adev,
&adev->pageflip_irq,
irq_type);
} else {
amdgpu_irq_put(
adev,
&adev->pageflip_irq,
irq_type);
drm_crtc_vblank_off(&acrtc->base);
}
}
static bool
is_scaling_state_different(const struct dm_connector_state *dm_state,
const struct dm_connector_state *old_dm_state)
{
if (dm_state->scaling != old_dm_state->scaling)
return true;
if (!dm_state->underscan_enable && old_dm_state->underscan_enable) {
if (old_dm_state->underscan_hborder != 0 && old_dm_state->underscan_vborder != 0)
return true;
} else if (dm_state->underscan_enable && !old_dm_state->underscan_enable) {
if (dm_state->underscan_hborder != 0 && dm_state->underscan_vborder != 0)
return true;
} else if (dm_state->underscan_hborder != old_dm_state->underscan_hborder ||
dm_state->underscan_vborder != old_dm_state->underscan_vborder)
return true;
return false;
}
static void remove_stream(struct amdgpu_device *adev,
struct amdgpu_crtc *acrtc,
struct dc_stream_state *stream)
{
/* this is the update mode case */
acrtc->otg_inst = -1;
acrtc->enabled = false;
}
static int get_cursor_position(struct drm_plane *plane, struct drm_crtc *crtc,
struct dc_cursor_position *position)
{
struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
int x, y;
int xorigin = 0, yorigin = 0;
position->enable = false;
position->x = 0;
position->y = 0;
if (!crtc || !plane->state->fb)
return 0;
if ((plane->state->crtc_w > amdgpu_crtc->max_cursor_width) ||
(plane->state->crtc_h > amdgpu_crtc->max_cursor_height)) {
DRM_ERROR("%s: bad cursor width or height %d x %d\n",
__func__,
plane->state->crtc_w,
plane->state->crtc_h);
return -EINVAL;
}
x = plane->state->crtc_x;
y = plane->state->crtc_y;
if (x <= -amdgpu_crtc->max_cursor_width ||
y <= -amdgpu_crtc->max_cursor_height)
return 0;
if (crtc->primary->state) {
/* avivo cursor are offset into the total surface */
x += crtc->primary->state->src_x >> 16;
y += crtc->primary->state->src_y >> 16;
}
if (x < 0) {
xorigin = min(-x, amdgpu_crtc->max_cursor_width - 1);
x = 0;
}
if (y < 0) {
yorigin = min(-y, amdgpu_crtc->max_cursor_height - 1);
y = 0;
}
position->enable = true;
position->x = x;
position->y = y;
position->x_hotspot = xorigin;
position->y_hotspot = yorigin;
return 0;
}
static void handle_cursor_update(struct drm_plane *plane,
struct drm_plane_state *old_plane_state)
{
struct amdgpu_device *adev = plane->dev->dev_private;
struct amdgpu_framebuffer *afb = to_amdgpu_framebuffer(plane->state->fb);
struct drm_crtc *crtc = afb ? plane->state->crtc : old_plane_state->crtc;
struct dm_crtc_state *crtc_state = crtc ? to_dm_crtc_state(crtc->state) : NULL;
struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
uint64_t address = afb ? afb->address : 0;
struct dc_cursor_position position;
struct dc_cursor_attributes attributes;
int ret;
if (!plane->state->fb && !old_plane_state->fb)
return;
DRM_DEBUG_DRIVER("%s: crtc_id=%d with size %d to %d\n",
__func__,
amdgpu_crtc->crtc_id,
plane->state->crtc_w,
plane->state->crtc_h);
ret = get_cursor_position(plane, crtc, &position);
if (ret)
return;
if (!position.enable) {
/* turn off cursor */
if (crtc_state && crtc_state->stream) {
mutex_lock(&adev->dm.dc_lock);
dc_stream_set_cursor_position(crtc_state->stream,
&position);
mutex_unlock(&adev->dm.dc_lock);
}
return;
}
amdgpu_crtc->cursor_width = plane->state->crtc_w;
amdgpu_crtc->cursor_height = plane->state->crtc_h;
memset(&attributes, 0, sizeof(attributes));
attributes.address.high_part = upper_32_bits(address);
attributes.address.low_part = lower_32_bits(address);
attributes.width = plane->state->crtc_w;
attributes.height = plane->state->crtc_h;
attributes.color_format = CURSOR_MODE_COLOR_PRE_MULTIPLIED_ALPHA;
attributes.rotation_angle = 0;
attributes.attribute_flags.value = 0;
attributes.pitch = attributes.width;
if (crtc_state->stream) {
mutex_lock(&adev->dm.dc_lock);
if (!dc_stream_set_cursor_attributes(crtc_state->stream,
&attributes))
DRM_ERROR("DC failed to set cursor attributes\n");
if (!dc_stream_set_cursor_position(crtc_state->stream,
&position))
DRM_ERROR("DC failed to set cursor position\n");
mutex_unlock(&adev->dm.dc_lock);
}
}
static void prepare_flip_isr(struct amdgpu_crtc *acrtc)
{
assert_spin_locked(&acrtc->base.dev->event_lock);
WARN_ON(acrtc->event);
acrtc->event = acrtc->base.state->event;
/* Set the flip status */
acrtc->pflip_status = AMDGPU_FLIP_SUBMITTED;
/* Mark this event as consumed */
acrtc->base.state->event = NULL;
DRM_DEBUG_DRIVER("crtc:%d, pflip_stat:AMDGPU_FLIP_SUBMITTED\n",
acrtc->crtc_id);
}
static void update_freesync_state_on_stream(
struct amdgpu_display_manager *dm,
struct dm_crtc_state *new_crtc_state,
struct dc_stream_state *new_stream,
struct dc_plane_state *surface,
u32 flip_timestamp_in_us)
{
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
struct mod_vrr_params vrr_params;
struct dc_info_packet vrr_infopacket = {0};
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
struct amdgpu_device *adev = dm->adev;
unsigned long flags;
if (!new_stream)
return;
/*
* TODO: Determine why min/max totals and vrefresh can be 0 here.
* For now it's sufficient to just guard against these conditions.
*/
if (!new_stream->timing.h_total || !new_stream->timing.v_total)
return;
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
spin_lock_irqsave(&adev->ddev->event_lock, flags);
vrr_params = new_crtc_state->vrr_params;
if (surface) {
mod_freesync_handle_preflip(
dm->freesync_module,
surface,
new_stream,
flip_timestamp_in_us,
&vrr_params);
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
if (adev->family < AMDGPU_FAMILY_AI &&
amdgpu_dm_vrr_active(new_crtc_state)) {
mod_freesync_handle_v_update(dm->freesync_module,
new_stream, &vrr_params);
/* Need to call this before the frame ends. */
dc_stream_adjust_vmin_vmax(dm->dc,
new_crtc_state->stream,
&vrr_params.adjust);
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
}
}
mod_freesync_build_vrr_infopacket(
dm->freesync_module,
new_stream,
&vrr_params,
PACKET_TYPE_VRR,
TRANSFER_FUNC_UNKNOWN,
&vrr_infopacket);
new_crtc_state->freesync_timing_changed |=
(memcmp(&new_crtc_state->vrr_params.adjust,
&vrr_params.adjust,
sizeof(vrr_params.adjust)) != 0);
new_crtc_state->freesync_vrr_info_changed |=
(memcmp(&new_crtc_state->vrr_infopacket,
&vrr_infopacket,
sizeof(vrr_infopacket)) != 0);
new_crtc_state->vrr_params = vrr_params;
new_crtc_state->vrr_infopacket = vrr_infopacket;
new_stream->adjust = new_crtc_state->vrr_params.adjust;
new_stream->vrr_infopacket = vrr_infopacket;
if (new_crtc_state->freesync_vrr_info_changed)
DRM_DEBUG_KMS("VRR packet update: crtc=%u enabled=%d state=%d",
new_crtc_state->base.crtc->base.id,
(int)new_crtc_state->base.vrr_enabled,
(int)vrr_params.state);
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
}
static void pre_update_freesync_state_on_stream(
struct amdgpu_display_manager *dm,
struct dm_crtc_state *new_crtc_state)
{
struct dc_stream_state *new_stream = new_crtc_state->stream;
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
struct mod_vrr_params vrr_params;
struct mod_freesync_config config = new_crtc_state->freesync_config;
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
struct amdgpu_device *adev = dm->adev;
unsigned long flags;
if (!new_stream)
return;
/*
* TODO: Determine why min/max totals and vrefresh can be 0 here.
* For now it's sufficient to just guard against these conditions.
*/
if (!new_stream->timing.h_total || !new_stream->timing.v_total)
return;
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
spin_lock_irqsave(&adev->ddev->event_lock, flags);
vrr_params = new_crtc_state->vrr_params;
if (new_crtc_state->vrr_supported &&
config.min_refresh_in_uhz &&
config.max_refresh_in_uhz) {
config.state = new_crtc_state->base.vrr_enabled ?
VRR_STATE_ACTIVE_VARIABLE :
VRR_STATE_INACTIVE;
} else {
config.state = VRR_STATE_UNSUPPORTED;
}
mod_freesync_build_vrr_params(dm->freesync_module,
new_stream,
&config, &vrr_params);
new_crtc_state->freesync_timing_changed |=
(memcmp(&new_crtc_state->vrr_params.adjust,
&vrr_params.adjust,
sizeof(vrr_params.adjust)) != 0);
new_crtc_state->vrr_params = vrr_params;
drm/amd/display: Compensate for pre-DCE12 BTR-VRR hw limitations. (v3) Pre-DCE12 needs special treatment for BTR / low framerate compensation for more stable behaviour: According to comments in the code and some testing on DCE-8 and DCE-11, DCE-11 and earlier only apply VTOTAL_MIN/MAX programming with a lag of one frame, so the special BTR hw programming for intermediate fixed duration frames must be done inside the current frame at flip submission in atomic commit tail, ie. one vblank earlier, and the fixed refresh intermediate frame mode must be also terminated one vblank earlier on pre-DCE12 display engines. To achieve proper termination on < DCE-12 shift the point when the switch-back from fixed vblank duration to variable vblank duration happens from the start of VBLANK (vblank irq, as done on DCE-12+) to back-porch or end of VBLANK (handled by vupdate irq handler). We must leave the switch-back code inside VBLANK irq for DCE12+, as before. Doing this, we get much better behaviour of BTR for up-sweeps, ie. going from short to long frame durations (~high to low fps) and for constant framerate flips, as tested on DCE-8 and DCE-11. Behaviour is still not quite as good as on DCN-1 though. On down-sweeps, going from long to short frame durations (low fps to high fps) < DCE-12 is a little bit improved, although by far not as much as for up-sweeps and constant fps. v2: Fix some wrong locking, as pointed out by Nicholas. v3: Simplify if-condition in vupdate-irq - nit by Nicholas. Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-26 15:40:16 -06:00
spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
}
static void amdgpu_dm_handle_vrr_transition(struct dm_crtc_state *old_state,
struct dm_crtc_state *new_state)
{
bool old_vrr_active = amdgpu_dm_vrr_active(old_state);
bool new_vrr_active = amdgpu_dm_vrr_active(new_state);
if (!old_vrr_active && new_vrr_active) {
/* Transition VRR inactive -> active:
* While VRR is active, we must not disable vblank irq, as a
* reenable after disable would compute bogus vblank/pflip
* timestamps if it likely happened inside display front-porch.
*
* We also need vupdate irq for the actual core vblank handling
* at end of vblank.
*/
dm_set_vupdate_irq(new_state->base.crtc, true);
drm_crtc_vblank_get(new_state->base.crtc);
DRM_DEBUG_DRIVER("%s: crtc=%u VRR off->on: Get vblank ref\n",
__func__, new_state->base.crtc->base.id);
} else if (old_vrr_active && !new_vrr_active) {
/* Transition VRR active -> inactive:
* Allow vblank irq disable again for fixed refresh rate.
*/
dm_set_vupdate_irq(new_state->base.crtc, false);
drm_crtc_vblank_put(new_state->base.crtc);
DRM_DEBUG_DRIVER("%s: crtc=%u VRR on->off: Drop vblank ref\n",
__func__, new_state->base.crtc->base.id);
}
}
static void amdgpu_dm_commit_cursors(struct drm_atomic_state *state)
{
struct drm_plane *plane;
struct drm_plane_state *old_plane_state, *new_plane_state;
int i;
/*
* TODO: Make this per-stream so we don't issue redundant updates for
* commits with multiple streams.
*/
for_each_oldnew_plane_in_state(state, plane, old_plane_state,
new_plane_state, i)
if (plane->type == DRM_PLANE_TYPE_CURSOR)
handle_cursor_update(plane, old_plane_state);
}
static void amdgpu_dm_commit_planes(struct drm_atomic_state *state,
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct dc_state *dc_state,
struct drm_device *dev,
struct amdgpu_display_manager *dm,
struct drm_crtc *pcrtc,
bool wait_for_vblank)
{
uint32_t i;
uint64_t timestamp_ns;
struct drm_plane *plane;
struct drm_plane_state *old_plane_state, *new_plane_state;
struct amdgpu_crtc *acrtc_attach = to_amdgpu_crtc(pcrtc);
struct drm_crtc_state *new_pcrtc_state =
drm_atomic_get_new_crtc_state(state, pcrtc);
struct dm_crtc_state *acrtc_state = to_dm_crtc_state(new_pcrtc_state);
struct dm_crtc_state *dm_old_crtc_state =
to_dm_crtc_state(drm_atomic_get_old_crtc_state(state, pcrtc));
int planes_count = 0, vpos, hpos;
long r;
unsigned long flags;
struct amdgpu_bo *abo;
uint64_t tiling_flags;
uint32_t target_vblank, last_flip_vblank;
bool vrr_active = amdgpu_dm_vrr_active(acrtc_state);
bool pflip_present = false;
struct {
struct dc_surface_update surface_updates[MAX_SURFACES];
struct dc_plane_info plane_infos[MAX_SURFACES];
struct dc_scaling_info scaling_infos[MAX_SURFACES];
struct dc_flip_addrs flip_addrs[MAX_SURFACES];
struct dc_stream_update stream_update;
} *bundle;
bundle = kzalloc(sizeof(*bundle), GFP_KERNEL);
if (!bundle) {
dm_error("Failed to allocate update bundle\n");
goto cleanup;
}
/*
* Disable the cursor first if we're disabling all the planes.
* It'll remain on the screen after the planes are re-enabled
* if we don't.
*/
if (acrtc_state->active_planes == 0)
amdgpu_dm_commit_cursors(state);
/* update planes when needed */
for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
struct drm_crtc *crtc = new_plane_state->crtc;
struct drm_crtc_state *new_crtc_state;
struct drm_framebuffer *fb = new_plane_state->fb;
bool plane_needs_flip;
struct dc_plane_state *dc_plane;
struct dm_plane_state *dm_new_plane_state = to_dm_plane_state(new_plane_state);
/* Cursor plane is handled after stream updates */
if (plane->type == DRM_PLANE_TYPE_CURSOR)
continue;
if (!fb || !crtc || pcrtc != crtc)
continue;
new_crtc_state = drm_atomic_get_new_crtc_state(state, crtc);
if (!new_crtc_state->active)
continue;
dc_plane = dm_new_plane_state->dc_state;
bundle->surface_updates[planes_count].surface = dc_plane;
if (new_pcrtc_state->color_mgmt_changed) {
bundle->surface_updates[planes_count].gamma = dc_plane->gamma_correction;
bundle->surface_updates[planes_count].in_transfer_func = dc_plane->in_transfer_func;
}
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
fill_dc_scaling_info(new_plane_state,
&bundle->scaling_infos[planes_count]);
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
bundle->surface_updates[planes_count].scaling_info =
&bundle->scaling_infos[planes_count];
plane_needs_flip = old_plane_state->fb && new_plane_state->fb;
pflip_present = pflip_present || plane_needs_flip;
if (!plane_needs_flip) {
planes_count += 1;
continue;
}
abo = gem_to_amdgpu_bo(fb->obj[0]);
/*
* Wait for all fences on this FB. Do limited wait to avoid
* deadlock during GPU reset when this fence will not signal
* but we hold reservation lock for the BO.
*/
r = dma_resv_wait_timeout_rcu(abo->tbo.base.resv, true,
false,
msecs_to_jiffies(5000));
if (unlikely(r <= 0))
DRM_ERROR("Waiting for fences timed out!");
/*
* TODO This might fail and hence better not used, wait
* explicitly on fences instead
* and in general should be called for
* blocking commit to as per framework helpers
*/
r = amdgpu_bo_reserve(abo, true);
if (unlikely(r != 0))
DRM_ERROR("failed to reserve buffer before flip\n");
amdgpu_bo_get_tiling_flags(abo, &tiling_flags);
amdgpu_bo_unreserve(abo);
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
fill_dc_plane_info_and_addr(
dm->adev, new_plane_state, tiling_flags,
&bundle->plane_infos[planes_count],
&bundle->flip_addrs[planes_count].address,
false);
DRM_DEBUG_DRIVER("plane: id=%d dcc_en=%d\n",
new_plane_state->plane->index,
bundle->plane_infos[planes_count].dcc.enable);
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
bundle->surface_updates[planes_count].plane_info =
&bundle->plane_infos[planes_count];
/*
* Only allow immediate flips for fast updates that don't
* change FB pitch, DCC state, rotation or mirroing.
*/
bundle->flip_addrs[planes_count].flip_immediate =
crtc->state->async_flip &&
acrtc_state->update_type == UPDATE_TYPE_FAST;
timestamp_ns = ktime_get_ns();
bundle->flip_addrs[planes_count].flip_timestamp_in_us = div_u64(timestamp_ns, 1000);
bundle->surface_updates[planes_count].flip_addr = &bundle->flip_addrs[planes_count];
bundle->surface_updates[planes_count].surface = dc_plane;
if (!bundle->surface_updates[planes_count].surface) {
DRM_ERROR("No surface for CRTC: id=%d\n",
acrtc_attach->crtc_id);
continue;
}
if (plane == pcrtc->primary)
update_freesync_state_on_stream(
dm,
acrtc_state,
acrtc_state->stream,
dc_plane,
bundle->flip_addrs[planes_count].flip_timestamp_in_us);
DRM_DEBUG_DRIVER("%s Flipping to hi: 0x%x, low: 0x%x\n",
__func__,
bundle->flip_addrs[planes_count].address.grph.addr.high_part,
bundle->flip_addrs[planes_count].address.grph.addr.low_part);
planes_count += 1;
}
if (pflip_present) {
drm/amd/display: Use vrr friendly pageflip throttling in DC. In VRR mode, keep track of the vblank count of the last completed pageflip in amdgpu_crtc->last_flip_vblank, as recorded in the pageflip completion handler after each completed flip. Use that count to prevent mmio programming a new pageflip within the same vblank in which the last pageflip completed, iow. to throttle pageflips to at most one flip per video frame, while at the same time allowing to request a flip not only before start of vblank, but also anywhere within vblank. The old logic did the same, and made sense for regular fixed refresh rate flipping, but in vrr mode it prevents requesting a flip anywhere inside the possibly huge vblank, thereby reducing framerate in vrr mode instead of improving it, by delaying a slightly delayed flip requests up to a maximum vblank duration + 1 scanout duration. This would limit VRR usefulness to only help applications with a very high GPU demand, which can submit the flip request before start of vblank, but then have to wait long for fences to complete. With this method a flip can be both requested and - after fences have completed - executed, ie. it doesn't matter if the request (amdgpu_dm_do_flip()) gets delayed until deep into the extended vblank due to cpu execution delays. This also allows clients which want to regulate framerate within the vrr range a much more fine-grained control of flip timing, a feature that might be useful for video playback, and is very useful for neuroscience/vision research applications. In regular non-VRR mode, retain the old flip submission behavior. This to keep flip scheduling for fullscreen X11/GLX OpenGL clients intact, if they use the GLX_OML_sync_control extensions glXSwapBufferMscOML(, ..., target_msc,...) function with a specific target_msc target vblank count. glXSwapBuffersMscOML() or DRI3/Present PresentPixmap() will not flip at the proper target_msc for a non-zero target_msc if VRR mode is active with this patch. They'd often flip one frame too early. However, this limitation should not matter much in VRR mode, as scheduling based on vblank counts is pretty futile/unusable under variable refresh duration anyway, so no real extra harm is done. According to some testing already done with this patch by Nicholas on top of my tests, IGT tests didn't report any problems. If fixes stuttering and flickering when flipping at rates below the minimum vrr refresh rate. Fixes: bb47de736661 ("drm/amdgpu: Set FreeSync state using drm VRR properties") Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Cc: <stable@vger.kernel.org> Cc: Harry Wentland <harry.wentland@amd.com> Cc: Alex Deucher <alexander.deucher@amd.com> Cc: Michel Dänzer <michel@daenzer.net> Tested-by: Bruno Filipe <bmilreu@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2019-02-08 23:52:55 -07:00
if (!vrr_active) {
/* Use old throttling in non-vrr fixed refresh rate mode
* to keep flip scheduling based on target vblank counts
* working in a backwards compatible way, e.g., for
* clients using the GLX_OML_sync_control extension or
* DRI3/Present extension with defined target_msc.
*/
last_flip_vblank = amdgpu_get_vblank_counter_kms(dm->ddev, acrtc_attach->crtc_id);
drm/amd/display: Use vrr friendly pageflip throttling in DC. In VRR mode, keep track of the vblank count of the last completed pageflip in amdgpu_crtc->last_flip_vblank, as recorded in the pageflip completion handler after each completed flip. Use that count to prevent mmio programming a new pageflip within the same vblank in which the last pageflip completed, iow. to throttle pageflips to at most one flip per video frame, while at the same time allowing to request a flip not only before start of vblank, but also anywhere within vblank. The old logic did the same, and made sense for regular fixed refresh rate flipping, but in vrr mode it prevents requesting a flip anywhere inside the possibly huge vblank, thereby reducing framerate in vrr mode instead of improving it, by delaying a slightly delayed flip requests up to a maximum vblank duration + 1 scanout duration. This would limit VRR usefulness to only help applications with a very high GPU demand, which can submit the flip request before start of vblank, but then have to wait long for fences to complete. With this method a flip can be both requested and - after fences have completed - executed, ie. it doesn't matter if the request (amdgpu_dm_do_flip()) gets delayed until deep into the extended vblank due to cpu execution delays. This also allows clients which want to regulate framerate within the vrr range a much more fine-grained control of flip timing, a feature that might be useful for video playback, and is very useful for neuroscience/vision research applications. In regular non-VRR mode, retain the old flip submission behavior. This to keep flip scheduling for fullscreen X11/GLX OpenGL clients intact, if they use the GLX_OML_sync_control extensions glXSwapBufferMscOML(, ..., target_msc,...) function with a specific target_msc target vblank count. glXSwapBuffersMscOML() or DRI3/Present PresentPixmap() will not flip at the proper target_msc for a non-zero target_msc if VRR mode is active with this patch. They'd often flip one frame too early. However, this limitation should not matter much in VRR mode, as scheduling based on vblank counts is pretty futile/unusable under variable refresh duration anyway, so no real extra harm is done. According to some testing already done with this patch by Nicholas on top of my tests, IGT tests didn't report any problems. If fixes stuttering and flickering when flipping at rates below the minimum vrr refresh rate. Fixes: bb47de736661 ("drm/amdgpu: Set FreeSync state using drm VRR properties") Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com> Cc: <stable@vger.kernel.org> Cc: Harry Wentland <harry.wentland@amd.com> Cc: Alex Deucher <alexander.deucher@amd.com> Cc: Michel Dänzer <michel@daenzer.net> Tested-by: Bruno Filipe <bmilreu@gmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Signed-off-by: Dave Airlie <airlied@redhat.com>
2019-02-08 23:52:55 -07:00
}
else {
/* For variable refresh rate mode only:
* Get vblank of last completed flip to avoid > 1 vrr
* flips per video frame by use of throttling, but allow
* flip programming anywhere in the possibly large
* variable vrr vblank interval for fine-grained flip
* timing control and more opportunity to avoid stutter
* on late submission of flips.
*/
spin_lock_irqsave(&pcrtc->dev->event_lock, flags);
last_flip_vblank = acrtc_attach->last_flip_vblank;
spin_unlock_irqrestore(&pcrtc->dev->event_lock, flags);
}
target_vblank = last_flip_vblank + wait_for_vblank;
/*
* Wait until we're out of the vertical blank period before the one
* targeted by the flip
*/
while ((acrtc_attach->enabled &&
(amdgpu_display_get_crtc_scanoutpos(dm->ddev, acrtc_attach->crtc_id,
0, &vpos, &hpos, NULL,
NULL, &pcrtc->hwmode)
& (DRM_SCANOUTPOS_VALID | DRM_SCANOUTPOS_IN_VBLANK)) ==
(DRM_SCANOUTPOS_VALID | DRM_SCANOUTPOS_IN_VBLANK) &&
(int)(target_vblank -
amdgpu_get_vblank_counter_kms(dm->ddev, acrtc_attach->crtc_id)) > 0)) {
usleep_range(1000, 1100);
}
if (acrtc_attach->base.state->event) {
drm_crtc_vblank_get(pcrtc);
spin_lock_irqsave(&pcrtc->dev->event_lock, flags);
WARN_ON(acrtc_attach->pflip_status != AMDGPU_FLIP_NONE);
prepare_flip_isr(acrtc_attach);
spin_unlock_irqrestore(&pcrtc->dev->event_lock, flags);
}
if (acrtc_state->stream) {
if (acrtc_state->freesync_vrr_info_changed)
bundle->stream_update.vrr_infopacket =
&acrtc_state->stream->vrr_infopacket;
}
}
/* Update the planes if changed or disable if we don't have any. */
if ((planes_count || acrtc_state->active_planes == 0) &&
acrtc_state->stream) {
if (new_pcrtc_state->mode_changed) {
bundle->stream_update.src = acrtc_state->stream->src;
bundle->stream_update.dst = acrtc_state->stream->dst;
}
drm/amd/display: Rework CRTC color management [Why] To prepare for the upcoming DRM plane color management properties we need to correct a lot of wrong behavior and assumptions made for CRTC color management. The documentation added by this commit in amdgpu_dm_color explains how the HW color pipeline works and its limitations with the DRM interface. The current implementation does the following wrong: - Implicit sRGB DGM when no CRTC DGM is set - Implicit sRGB RGM when no CRTC RGM is set - No way to specify a non-linear DGM matrix that produces correct output - No way to specify a correct RGM when a linear DGM is used We had workarounds for passing kms_color tests but not all of the behavior we had wrong was covered by these tests (especially when it comes to non-linear DGM). Testing both DGM and RGM at the same time isn't something kms_color tests well either. [How] The specifics for how color management works in AMDGPU and the new behavior can be found by reading the documentation added to amdgpu_dm_color.c from this patch. All of the incorrect cases from the old implementation have been addressed for the atomic interface, but there still a few TODOs for the legacy one. Note: this does cause regressions for kms_color@pipe-a-ctm-* over HDMI. The result looks correct from visual inspection but the CRC no longer matches. For reference, the test was previously doing the following: linear degamma -> CTM -> sRGB regamma -> RGB to YUV (709) -> ... Now the test is doing: linear degamma -> CTM -> linear regamma -> RGB to YUV (709) -> ... Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-05-09 10:14:58 -06:00
if (new_pcrtc_state->color_mgmt_changed) {
/*
* TODO: This isn't fully correct since we've actually
* already modified the stream in place.
*/
bundle->stream_update.gamut_remap =
&acrtc_state->stream->gamut_remap_matrix;
bundle->stream_update.output_csc_transform =
&acrtc_state->stream->csc_color_matrix;
bundle->stream_update.out_transfer_func =
acrtc_state->stream->out_transfer_func;
}
acrtc_state->stream->abm_level = acrtc_state->abm_level;
if (acrtc_state->abm_level != dm_old_crtc_state->abm_level)
bundle->stream_update.abm_level = &acrtc_state->abm_level;
/*
* If FreeSync state on the stream has changed then we need to
* re-adjust the min/max bounds now that DC doesn't handle this
* as part of commit.
*/
if (amdgpu_dm_vrr_active(dm_old_crtc_state) !=
amdgpu_dm_vrr_active(acrtc_state)) {
spin_lock_irqsave(&pcrtc->dev->event_lock, flags);
dc_stream_adjust_vmin_vmax(
dm->dc, acrtc_state->stream,
&acrtc_state->vrr_params.adjust);
spin_unlock_irqrestore(&pcrtc->dev->event_lock, flags);
}
mutex_lock(&dm->dc_lock);
dc_commit_updates_for_stream(dm->dc,
bundle->surface_updates,
planes_count,
acrtc_state->stream,
&bundle->stream_update,
dc_state);
mutex_unlock(&dm->dc_lock);
}
/*
* Update cursor state *after* programming all the planes.
* This avoids redundant programming in the case where we're going
* to be disabling a single plane - those pipes are being disabled.
*/
if (acrtc_state->active_planes)
amdgpu_dm_commit_cursors(state);
cleanup:
kfree(bundle);
}
static void amdgpu_dm_commit_audio(struct drm_device *dev,
struct drm_atomic_state *state)
{
struct amdgpu_device *adev = dev->dev_private;
struct amdgpu_dm_connector *aconnector;
struct drm_connector *connector;
struct drm_connector_state *old_con_state, *new_con_state;
struct drm_crtc_state *new_crtc_state;
struct dm_crtc_state *new_dm_crtc_state;
const struct dc_stream_status *status;
int i, inst;
/* Notify device removals. */
for_each_oldnew_connector_in_state(state, connector, old_con_state, new_con_state, i) {
if (old_con_state->crtc != new_con_state->crtc) {
/* CRTC changes require notification. */
goto notify;
}
if (!new_con_state->crtc)
continue;
new_crtc_state = drm_atomic_get_new_crtc_state(
state, new_con_state->crtc);
if (!new_crtc_state)
continue;
if (!drm_atomic_crtc_needs_modeset(new_crtc_state))
continue;
notify:
aconnector = to_amdgpu_dm_connector(connector);
mutex_lock(&adev->dm.audio_lock);
inst = aconnector->audio_inst;
aconnector->audio_inst = -1;
mutex_unlock(&adev->dm.audio_lock);
amdgpu_dm_audio_eld_notify(adev, inst);
}
/* Notify audio device additions. */
for_each_new_connector_in_state(state, connector, new_con_state, i) {
if (!new_con_state->crtc)
continue;
new_crtc_state = drm_atomic_get_new_crtc_state(
state, new_con_state->crtc);
if (!new_crtc_state)
continue;
if (!drm_atomic_crtc_needs_modeset(new_crtc_state))
continue;
new_dm_crtc_state = to_dm_crtc_state(new_crtc_state);
if (!new_dm_crtc_state->stream)
continue;
status = dc_stream_get_status(new_dm_crtc_state->stream);
if (!status)
continue;
aconnector = to_amdgpu_dm_connector(connector);
mutex_lock(&adev->dm.audio_lock);
inst = status->audio_inst;
aconnector->audio_inst = inst;
mutex_unlock(&adev->dm.audio_lock);
amdgpu_dm_audio_eld_notify(adev, inst);
}
}
drm/amd/display: Split enabling CRTC interrupts into two passes [Why] When disabling all the pipes for a CRTC the page-flip interrupt also gets disabled on Raven. We can't re-enable the page-flip interrupt unless we give DC at least one active DC plane. We currently enable interrupts after the call to dc_commit_state since there's currently no valid sequence that should disable all the planes or re-enable planes for a CRTC without first going through dc_commit_state. If we were to allow for a CRTC to be enabled with no primary plane this would not be the case - the call to dc_commit_updates_for_stream would enable the planes when going from zero to at least one active plane, but manage_dm_interrupts would have been called too early. This results in a page-flip timeout on any subsequent commits since we think the page-flip are now enabled when they're actually disabled. We need to enable interrupts after the call to dc_commit_updates_for_stream. [How] Split enabling interrupts into two passes. One pass before dc_commit_updates_for_stream and one after it. Shifting all the interrupts to be strictly below the call doesn't currently work even though it should in theory. We end up queuing off the vblank event to be handle by the flip handler before it's actually enabled in some cases, particularly: old_crtc_state->active = false -> new_crtc_state->active = true The framebuffer states haven't changed and we can technically still do a "pageflip" in this case and send back the event. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-08 09:18:31 -06:00
/*
* Enable interrupts on CRTCs that are newly active, undergone
* a modeset, or have active planes again.
*
* Done in two passes, based on the for_modeset flag:
* Pass 1: For CRTCs going through modeset
* Pass 2: For CRTCs going from 0 to n active planes
*
* Interrupts can only be enabled after the planes are programmed,
* so this requires a two-pass approach since we don't want to
* just defer the interrupts until after commit planes every time.
*/
static void amdgpu_dm_enable_crtc_interrupts(struct drm_device *dev,
struct drm_atomic_state *state,
bool for_modeset)
{
struct amdgpu_device *adev = dev->dev_private;
struct drm_crtc *crtc;
struct drm_crtc_state *old_crtc_state, *new_crtc_state;
int i;
#ifdef CONFIG_DEBUG_FS
enum amdgpu_dm_pipe_crc_source source;
#endif
drm/amd/display: Split enabling CRTC interrupts into two passes [Why] When disabling all the pipes for a CRTC the page-flip interrupt also gets disabled on Raven. We can't re-enable the page-flip interrupt unless we give DC at least one active DC plane. We currently enable interrupts after the call to dc_commit_state since there's currently no valid sequence that should disable all the planes or re-enable planes for a CRTC without first going through dc_commit_state. If we were to allow for a CRTC to be enabled with no primary plane this would not be the case - the call to dc_commit_updates_for_stream would enable the planes when going from zero to at least one active plane, but manage_dm_interrupts would have been called too early. This results in a page-flip timeout on any subsequent commits since we think the page-flip are now enabled when they're actually disabled. We need to enable interrupts after the call to dc_commit_updates_for_stream. [How] Split enabling interrupts into two passes. One pass before dc_commit_updates_for_stream and one after it. Shifting all the interrupts to be strictly below the call doesn't currently work even though it should in theory. We end up queuing off the vblank event to be handle by the flip handler before it's actually enabled in some cases, particularly: old_crtc_state->active = false -> new_crtc_state->active = true The framebuffer states haven't changed and we can technically still do a "pageflip" in this case and send back the event. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-08 09:18:31 -06:00
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state,
new_crtc_state, i) {
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(crtc);
struct dm_crtc_state *dm_new_crtc_state =
to_dm_crtc_state(new_crtc_state);
struct dm_crtc_state *dm_old_crtc_state =
to_dm_crtc_state(old_crtc_state);
bool modeset = drm_atomic_crtc_needs_modeset(new_crtc_state);
bool run_pass;
run_pass = (for_modeset && modeset) ||
(!for_modeset && !modeset &&
!dm_old_crtc_state->interrupts_enabled);
if (!run_pass)
continue;
if (!dm_new_crtc_state->interrupts_enabled)
continue;
manage_dm_interrupts(adev, acrtc, true);
#ifdef CONFIG_DEBUG_FS
/* The stream has changed so CRC capture needs to re-enabled. */
source = dm_new_crtc_state->crc_src;
if (amdgpu_dm_is_valid_crc_source(source)) {
amdgpu_dm_crtc_configure_crc_source(
crtc, dm_new_crtc_state,
dm_new_crtc_state->crc_src);
drm/amd/display: Split enabling CRTC interrupts into two passes [Why] When disabling all the pipes for a CRTC the page-flip interrupt also gets disabled on Raven. We can't re-enable the page-flip interrupt unless we give DC at least one active DC plane. We currently enable interrupts after the call to dc_commit_state since there's currently no valid sequence that should disable all the planes or re-enable planes for a CRTC without first going through dc_commit_state. If we were to allow for a CRTC to be enabled with no primary plane this would not be the case - the call to dc_commit_updates_for_stream would enable the planes when going from zero to at least one active plane, but manage_dm_interrupts would have been called too early. This results in a page-flip timeout on any subsequent commits since we think the page-flip are now enabled when they're actually disabled. We need to enable interrupts after the call to dc_commit_updates_for_stream. [How] Split enabling interrupts into two passes. One pass before dc_commit_updates_for_stream and one after it. Shifting all the interrupts to be strictly below the call doesn't currently work even though it should in theory. We end up queuing off the vblank event to be handle by the flip handler before it's actually enabled in some cases, particularly: old_crtc_state->active = false -> new_crtc_state->active = true The framebuffer states haven't changed and we can technically still do a "pageflip" in this case and send back the event. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-08 09:18:31 -06:00
}
#endif
}
}
/*
* amdgpu_dm_crtc_copy_transient_flags - copy mirrored flags from DRM to DC
* @crtc_state: the DRM CRTC state
* @stream_state: the DC stream state.
*
* Copy the mirrored transient state flags from DRM, to DC. It is used to bring
* a dc_stream_state's flags in sync with a drm_crtc_state's flags.
*/
static void amdgpu_dm_crtc_copy_transient_flags(struct drm_crtc_state *crtc_state,
struct dc_stream_state *stream_state)
{
stream_state->mode_changed = drm_atomic_crtc_needs_modeset(crtc_state);
}
static int amdgpu_dm_atomic_commit(struct drm_device *dev,
struct drm_atomic_state *state,
bool nonblock)
{
struct drm_crtc *crtc;
struct drm_crtc_state *old_crtc_state, *new_crtc_state;
struct amdgpu_device *adev = dev->dev_private;
int i;
/*
* We evade vblank and pflip interrupts on CRTCs that are undergoing
* a modeset, being disabled, or have no active planes.
*
* It's done in atomic commit rather than commit tail for now since
* some of these interrupt handlers access the current CRTC state and
* potentially the stream pointer itself.
*
* Since the atomic state is swapped within atomic commit and not within
* commit tail this would leave to new state (that hasn't been committed yet)
* being accesssed from within the handlers.
*
* TODO: Fix this so we can do this in commit tail and not have to block
* in atomic check.
*/
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
struct dm_crtc_state *dm_old_crtc_state = to_dm_crtc_state(old_crtc_state);
struct dm_crtc_state *dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(crtc);
if (dm_old_crtc_state->interrupts_enabled &&
(!dm_new_crtc_state->interrupts_enabled ||
drm_atomic_crtc_needs_modeset(new_crtc_state)))
manage_dm_interrupts(adev, acrtc, false);
}
/*
* Add check here for SoC's that support hardware cursor plane, to
* unset legacy_cursor_update
*/
return drm_atomic_helper_commit(dev, state, nonblock);
/*TODO Handle EINTR, reenable IRQ*/
}
/**
* amdgpu_dm_atomic_commit_tail() - AMDgpu DM's commit tail implementation.
* @state: The atomic state to commit
*
* This will tell DC to commit the constructed DC state from atomic_check,
* programming the hardware. Any failures here implies a hardware failure, since
* atomic check should have filtered anything non-kosher.
*/
static void amdgpu_dm_atomic_commit_tail(struct drm_atomic_state *state)
{
struct drm_device *dev = state->dev;
struct amdgpu_device *adev = dev->dev_private;
struct amdgpu_display_manager *dm = &adev->dm;
struct dm_atomic_state *dm_state;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct dc_state *dc_state = NULL, *dc_state_temp = NULL;
uint32_t i, j;
struct drm_crtc *crtc;
struct drm_crtc_state *old_crtc_state, *new_crtc_state;
unsigned long flags;
bool wait_for_vblank = true;
struct drm_connector *connector;
struct drm_connector_state *old_con_state, *new_con_state;
struct dm_crtc_state *dm_old_crtc_state, *dm_new_crtc_state;
drm/amdgpu: Count disabled CRTCs in commit tail earlier This fixes a regression I accidentally reduced that was picked up by kasan, where we were checking the CRTC atomic states after DRM's helpers had already freed them. Example: ================================================================== BUG: KASAN: use-after-free in amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] Read of size 1 at addr ffff8803a697b071 by task kworker/u16:0/7 CPU: 7 PID: 7 Comm: kworker/u16:0 Tainted: G O 4.18.0-rc1Lyude-Upstream+ #1 Hardware name: HP HP ZBook 15 G4/8275, BIOS P70 Ver. 01.21 05/02/2018 Workqueue: events_unbound commit_work [drm_kms_helper] Call Trace: dump_stack+0xc1/0x169 ? dump_stack_print_info.cold.1+0x42/0x42 ? kmsg_dump_rewind_nolock+0xd9/0xd9 ? printk+0x9f/0xc5 ? amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] print_address_description+0x6c/0x23c ? amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] kasan_report.cold.6+0x241/0x2fd amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] ? commit_planes_to_stream.constprop.45+0x13b0/0x13b0 [amdgpu] ? cpu_load_update_active+0x290/0x290 ? finish_task_switch+0x2bd/0x840 ? __switch_to_asm+0x34/0x70 ? read_word_at_a_time+0xe/0x20 ? strscpy+0x14b/0x460 ? drm_atomic_helper_wait_for_dependencies+0x47d/0x7e0 [drm_kms_helper] commit_tail+0x96/0xe0 [drm_kms_helper] process_one_work+0x88a/0x1360 ? create_worker+0x540/0x540 ? __sched_text_start+0x8/0x8 ? move_queued_task+0x760/0x760 ? call_rcu_sched+0x20/0x20 ? vsnprintf+0xcda/0x1350 ? wait_woken+0x1c0/0x1c0 ? mutex_unlock+0x1d/0x40 ? init_timer_key+0x190/0x230 ? schedule+0xea/0x390 ? __schedule+0x1ea0/0x1ea0 ? need_to_create_worker+0xe4/0x210 ? init_worker_pool+0x700/0x700 ? try_to_del_timer_sync+0xbf/0x110 ? del_timer+0x120/0x120 ? __mutex_lock_slowpath+0x10/0x10 worker_thread+0x196/0x11f0 ? flush_rcu_work+0x50/0x50 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __schedule+0x7d6/0x1ea0 ? migrate_swap_stop+0x850/0x880 ? __sched_text_start+0x8/0x8 ? save_stack+0x8c/0xb0 ? kasan_kmalloc+0xbf/0xe0 ? kmem_cache_alloc_trace+0xe4/0x190 ? kthread+0x98/0x390 ? ret_from_fork+0x35/0x40 ? ret_from_fork+0x35/0x40 ? deactivate_slab.isra.67+0x3c4/0x5c0 ? kthread+0x98/0x390 ? kthread+0x98/0x390 ? set_track+0x76/0x120 ? schedule+0xea/0x390 ? __schedule+0x1ea0/0x1ea0 ? wait_woken+0x1c0/0x1c0 ? kasan_unpoison_shadow+0x30/0x40 ? parse_args.cold.15+0x17a/0x17a ? flush_rcu_work+0x50/0x50 kthread+0x2d4/0x390 ? kthread_create_worker_on_cpu+0xc0/0xc0 ret_from_fork+0x35/0x40 Allocated by task 1124: kasan_kmalloc+0xbf/0xe0 kmem_cache_alloc_trace+0xe4/0x190 dm_crtc_duplicate_state+0x78/0x130 [amdgpu] drm_atomic_get_crtc_state+0x147/0x410 [drm] page_flip_common+0x57/0x230 [drm_kms_helper] drm_atomic_helper_page_flip+0xa6/0x110 [drm_kms_helper] drm_mode_page_flip_ioctl+0xc4b/0x10a0 [drm] drm_ioctl_kernel+0x1d4/0x260 [drm] drm_ioctl+0x433/0x920 [drm] amdgpu_drm_ioctl+0x11d/0x290 [amdgpu] do_vfs_ioctl+0x1a1/0x13d0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x6f/0xb0 do_syscall_64+0x147/0x440 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 1124: __kasan_slab_free+0x12e/0x180 kfree+0x92/0x1a0 drm_atomic_state_default_clear+0x315/0xc40 [drm] __drm_atomic_state_free+0x35/0xd0 [drm] drm_atomic_helper_update_plane+0xac/0x350 [drm_kms_helper] __setplane_internal+0x2d6/0x840 [drm] drm_mode_cursor_universal+0x41e/0xbe0 [drm] drm_mode_cursor_common+0x49f/0x880 [drm] drm_mode_cursor_ioctl+0xd8/0x130 [drm] drm_ioctl_kernel+0x1d4/0x260 [drm] drm_ioctl+0x433/0x920 [drm] amdgpu_drm_ioctl+0x11d/0x290 [amdgpu] do_vfs_ioctl+0x1a1/0x13d0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x6f/0xb0 do_syscall_64+0x147/0x440 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The buggy address belongs to the object at ffff8803a697b068 which belongs to the cache kmalloc-1024 of size 1024 The buggy address is located 9 bytes inside of 1024-byte region [ffff8803a697b068, ffff8803a697b468) The buggy address belongs to the page: page:ffffea000e9a5e00 count:1 mapcount:0 mapping:ffff88041e00efc0 index:0x0 compound_mapcount: 0 flags: 0x8000000000008100(slab|head) raw: 8000000000008100 ffffea000ecbc208 ffff88041e000c70 ffff88041e00efc0 raw: 0000000000000000 0000000000170017 00000001ffffffff 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff8803a697af00: fb fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ffff8803a697af80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc >ffff8803a697b000: fc fc fc fc fc fc fc fc fc fc fc fc fc fb fb fb ^ ffff8803a697b080: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff8803a697b100: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ================================================================== So, we fix this by counting the number of CRTCs this atomic commit disabled early on in the function before their atomic states have been freed, then use that count later to do the appropriate number of RPM puts at the end of the function. Acked-by: Michel Dänzer <michel.daenzer@amd.com> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Cc: stable@vger.kernel.org Fixes: 97028037a38ae ("drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail()") Signed-off-by: Lyude Paul <lyude@redhat.com> Cc: Michel Dänzer <michel@daenzer.net> Reported-by: Michel Dänzer <michel@daenzer.net> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-06-21 14:48:26 -06:00
int crtc_disable_count = 0;
drm_atomic_helper_update_legacy_modeset_state(dev, state);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
dm_state = dm_atomic_get_new_state(state);
if (dm_state && dm_state->context) {
dc_state = dm_state->context;
} else {
/* No state changes, retain current state. */
dc_state_temp = dc_create_state(dm->dc);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
ASSERT(dc_state_temp);
dc_state = dc_state_temp;
dc_resource_state_copy_construct_current(dm->dc, dc_state);
}
/* update changed items */
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(crtc);
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
dm_old_crtc_state = to_dm_crtc_state(old_crtc_state);
DRM_DEBUG_DRIVER(
"amdgpu_crtc id:%d crtc_state_flags: enable:%d, active:%d, "
"planes_changed:%d, mode_changed:%d,active_changed:%d,"
"connectors_changed:%d\n",
acrtc->crtc_id,
new_crtc_state->enable,
new_crtc_state->active,
new_crtc_state->planes_changed,
new_crtc_state->mode_changed,
new_crtc_state->active_changed,
new_crtc_state->connectors_changed);
/* Copy all transient state flags into dc state */
if (dm_new_crtc_state->stream) {
amdgpu_dm_crtc_copy_transient_flags(&dm_new_crtc_state->base,
dm_new_crtc_state->stream);
}
/* handles headless hotplug case, updating new_state and
* aconnector as needed
*/
if (modeset_required(new_crtc_state, dm_new_crtc_state->stream, dm_old_crtc_state->stream)) {
DRM_DEBUG_DRIVER("Atomic commit: SET crtc id %d: [%p]\n", acrtc->crtc_id, acrtc);
if (!dm_new_crtc_state->stream) {
/*
* this could happen because of issues with
* userspace notifications delivery.
* In this case userspace tries to set mode on
* display which is disconnected in fact.
* dc_sink is NULL in this case on aconnector.
* We expect reset mode will come soon.
*
* This can also happen when unplug is done
* during resume sequence ended
*
* In this case, we want to pretend we still
* have a sink to keep the pipe running so that
* hw state is consistent with the sw state
*/
DRM_DEBUG_DRIVER("%s: Failed to create new stream for crtc %d\n",
__func__, acrtc->base.base.id);
continue;
}
if (dm_old_crtc_state->stream)
remove_stream(adev, acrtc, dm_old_crtc_state->stream);
drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail() So, unfortunately I recently made the discovery that in the upstream kernel, the only reason that amdgpu is not currently suffering from issues with runtime PM putting the GPU into suspend while it's driving displays is due to the fact that on most prime systems, we have sound devices associated with the GPU that hold their own runtime PM ref for the GPU. What this means however, is that in the event that there isn't any kind of sound device active (which can easily be reproduced by building a kernel with sound drivers disabled), the GPU will fall asleep even when there's displays active. This appears to be in part due to the fact that amdgpu has not actually ever relied on it's rpm_idle() function to be the only thing keeping it running, and normally grabs it's own power references whenever there are displays active (as can be seen with the original pre-DC codepath in amdgpu_display_crtc_set_config() in amdgpu_display.c). This means it's very likely that this bug was introduced during the switch over the DC. So to fix this, we start grabbing runtime PM references every time we enable a previously disabled CRTC in atomic_commit_tail(). This appears to be the correct solution, as it matches up with what i915 does in i915/intel_runtime_pm.c. The one sideaffect of this is that we ignore the variable that the pre-DC code used to use for tracking when it needed runtime PM refs, adev->have_disp_power_ref. This is mainly because there's no way for a driver to tell whether or not all of it's CRTCs are enabled or disabled when we've begun committing an atomic state, as there may be CRTC commits happening in parallel that aren't contained within the atomic state being committed. So, it's safer to just get/put a reference for each CRTC being enabled or disabled in the new atomic state. Signed-off-by: Lyude Paul <lyude@redhat.com> Acked-by: Christian König <christian.koenig@amd.com>. Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org
2018-06-04 13:35:03 -06:00
pm_runtime_get_noresume(dev->dev);
acrtc->enabled = true;
acrtc->hw_mode = new_crtc_state->mode;
crtc->hwmode = new_crtc_state->mode;
} else if (modereset_required(new_crtc_state)) {
DRM_DEBUG_DRIVER("Atomic commit: RESET. crtc id %d:[%p]\n", acrtc->crtc_id, acrtc);
/* i.e. reset mode */
if (dm_old_crtc_state->stream)
remove_stream(adev, acrtc, dm_old_crtc_state->stream);
}
} /* for_each_crtc_in_state() */
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (dc_state) {
dm_enable_per_frame_crtc_master_sync(dc_state);
mutex_lock(&dm->dc_lock);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
WARN_ON(!dc_commit_state(dm->dc, dc_state));
mutex_unlock(&dm->dc_lock);
}
for_each_new_crtc_in_state(state, crtc, new_crtc_state, i) {
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(crtc);
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
if (dm_new_crtc_state->stream != NULL) {
const struct dc_stream_status *status =
dc_stream_get_status(dm_new_crtc_state->stream);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (!status)
status = dc_stream_get_status_from_state(dc_state,
dm_new_crtc_state->stream);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (!status)
DC_ERR("got no status for stream %p on acrtc%p\n", dm_new_crtc_state->stream, acrtc);
else
acrtc->otg_inst = status->primary_otg_inst;
}
}
/* Handle connector state changes */
for_each_oldnew_connector_in_state(state, connector, old_con_state, new_con_state, i) {
struct dm_connector_state *dm_new_con_state = to_dm_connector_state(new_con_state);
struct dm_connector_state *dm_old_con_state = to_dm_connector_state(old_con_state);
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(dm_new_con_state->base.crtc);
struct dc_surface_update dummy_updates[MAX_SURFACES];
struct dc_stream_update stream_update;
struct dc_info_packet hdr_packet;
struct dc_stream_status *status = NULL;
bool abm_changed, hdr_changed, scaling_changed;
memset(&dummy_updates, 0, sizeof(dummy_updates));
memset(&stream_update, 0, sizeof(stream_update));
if (acrtc) {
new_crtc_state = drm_atomic_get_new_crtc_state(state, &acrtc->base);
old_crtc_state = drm_atomic_get_old_crtc_state(state, &acrtc->base);
}
/* Skip any modesets/resets */
if (!acrtc || drm_atomic_crtc_needs_modeset(new_crtc_state))
continue;
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
dm_old_crtc_state = to_dm_crtc_state(old_crtc_state);
scaling_changed = is_scaling_state_different(dm_new_con_state,
dm_old_con_state);
abm_changed = dm_new_crtc_state->abm_level !=
dm_old_crtc_state->abm_level;
hdr_changed =
is_hdr_metadata_different(old_con_state, new_con_state);
if (!scaling_changed && !abm_changed && !hdr_changed)
continue;
if (scaling_changed) {
update_stream_scaling_settings(&dm_new_con_state->base.crtc->mode,
dm_new_con_state, (struct dc_stream_state *)dm_new_crtc_state->stream);
stream_update.src = dm_new_crtc_state->stream->src;
stream_update.dst = dm_new_crtc_state->stream->dst;
}
if (abm_changed) {
dm_new_crtc_state->stream->abm_level = dm_new_crtc_state->abm_level;
stream_update.abm_level = &dm_new_crtc_state->abm_level;
}
if (hdr_changed) {
fill_hdr_info_packet(new_con_state, &hdr_packet);
stream_update.hdr_static_metadata = &hdr_packet;
}
status = dc_stream_get_status(dm_new_crtc_state->stream);
WARN_ON(!status);
WARN_ON(!status->plane_count);
/*
* TODO: DC refuses to perform stream updates without a dc_surface_update.
* Here we create an empty update on each plane.
* To fix this, DC should permit updating only stream properties.
*/
for (j = 0; j < status->plane_count; j++)
dummy_updates[j].surface = status->plane_states[0];
mutex_lock(&dm->dc_lock);
dc_commit_updates_for_stream(dm->dc,
dummy_updates,
status->plane_count,
dm_new_crtc_state->stream,
&stream_update,
dc_state);
mutex_unlock(&dm->dc_lock);
}
drm/amd/display: Split enabling CRTC interrupts into two passes [Why] When disabling all the pipes for a CRTC the page-flip interrupt also gets disabled on Raven. We can't re-enable the page-flip interrupt unless we give DC at least one active DC plane. We currently enable interrupts after the call to dc_commit_state since there's currently no valid sequence that should disable all the planes or re-enable planes for a CRTC without first going through dc_commit_state. If we were to allow for a CRTC to be enabled with no primary plane this would not be the case - the call to dc_commit_updates_for_stream would enable the planes when going from zero to at least one active plane, but manage_dm_interrupts would have been called too early. This results in a page-flip timeout on any subsequent commits since we think the page-flip are now enabled when they're actually disabled. We need to enable interrupts after the call to dc_commit_updates_for_stream. [How] Split enabling interrupts into two passes. One pass before dc_commit_updates_for_stream and one after it. Shifting all the interrupts to be strictly below the call doesn't currently work even though it should in theory. We end up queuing off the vblank event to be handle by the flip handler before it's actually enabled in some cases, particularly: old_crtc_state->active = false -> new_crtc_state->active = true The framebuffer states haven't changed and we can technically still do a "pageflip" in this case and send back the event. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-08 09:18:31 -06:00
/* Count number of newly disabled CRTCs for dropping PM refs later. */
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state,
new_crtc_state, i) {
drm/amdgpu: Count disabled CRTCs in commit tail earlier This fixes a regression I accidentally reduced that was picked up by kasan, where we were checking the CRTC atomic states after DRM's helpers had already freed them. Example: ================================================================== BUG: KASAN: use-after-free in amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] Read of size 1 at addr ffff8803a697b071 by task kworker/u16:0/7 CPU: 7 PID: 7 Comm: kworker/u16:0 Tainted: G O 4.18.0-rc1Lyude-Upstream+ #1 Hardware name: HP HP ZBook 15 G4/8275, BIOS P70 Ver. 01.21 05/02/2018 Workqueue: events_unbound commit_work [drm_kms_helper] Call Trace: dump_stack+0xc1/0x169 ? dump_stack_print_info.cold.1+0x42/0x42 ? kmsg_dump_rewind_nolock+0xd9/0xd9 ? printk+0x9f/0xc5 ? amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] print_address_description+0x6c/0x23c ? amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] kasan_report.cold.6+0x241/0x2fd amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] ? commit_planes_to_stream.constprop.45+0x13b0/0x13b0 [amdgpu] ? cpu_load_update_active+0x290/0x290 ? finish_task_switch+0x2bd/0x840 ? __switch_to_asm+0x34/0x70 ? read_word_at_a_time+0xe/0x20 ? strscpy+0x14b/0x460 ? drm_atomic_helper_wait_for_dependencies+0x47d/0x7e0 [drm_kms_helper] commit_tail+0x96/0xe0 [drm_kms_helper] process_one_work+0x88a/0x1360 ? create_worker+0x540/0x540 ? __sched_text_start+0x8/0x8 ? move_queued_task+0x760/0x760 ? call_rcu_sched+0x20/0x20 ? vsnprintf+0xcda/0x1350 ? wait_woken+0x1c0/0x1c0 ? mutex_unlock+0x1d/0x40 ? init_timer_key+0x190/0x230 ? schedule+0xea/0x390 ? __schedule+0x1ea0/0x1ea0 ? need_to_create_worker+0xe4/0x210 ? init_worker_pool+0x700/0x700 ? try_to_del_timer_sync+0xbf/0x110 ? del_timer+0x120/0x120 ? __mutex_lock_slowpath+0x10/0x10 worker_thread+0x196/0x11f0 ? flush_rcu_work+0x50/0x50 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __schedule+0x7d6/0x1ea0 ? migrate_swap_stop+0x850/0x880 ? __sched_text_start+0x8/0x8 ? save_stack+0x8c/0xb0 ? kasan_kmalloc+0xbf/0xe0 ? kmem_cache_alloc_trace+0xe4/0x190 ? kthread+0x98/0x390 ? ret_from_fork+0x35/0x40 ? ret_from_fork+0x35/0x40 ? deactivate_slab.isra.67+0x3c4/0x5c0 ? kthread+0x98/0x390 ? kthread+0x98/0x390 ? set_track+0x76/0x120 ? schedule+0xea/0x390 ? __schedule+0x1ea0/0x1ea0 ? wait_woken+0x1c0/0x1c0 ? kasan_unpoison_shadow+0x30/0x40 ? parse_args.cold.15+0x17a/0x17a ? flush_rcu_work+0x50/0x50 kthread+0x2d4/0x390 ? kthread_create_worker_on_cpu+0xc0/0xc0 ret_from_fork+0x35/0x40 Allocated by task 1124: kasan_kmalloc+0xbf/0xe0 kmem_cache_alloc_trace+0xe4/0x190 dm_crtc_duplicate_state+0x78/0x130 [amdgpu] drm_atomic_get_crtc_state+0x147/0x410 [drm] page_flip_common+0x57/0x230 [drm_kms_helper] drm_atomic_helper_page_flip+0xa6/0x110 [drm_kms_helper] drm_mode_page_flip_ioctl+0xc4b/0x10a0 [drm] drm_ioctl_kernel+0x1d4/0x260 [drm] drm_ioctl+0x433/0x920 [drm] amdgpu_drm_ioctl+0x11d/0x290 [amdgpu] do_vfs_ioctl+0x1a1/0x13d0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x6f/0xb0 do_syscall_64+0x147/0x440 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 1124: __kasan_slab_free+0x12e/0x180 kfree+0x92/0x1a0 drm_atomic_state_default_clear+0x315/0xc40 [drm] __drm_atomic_state_free+0x35/0xd0 [drm] drm_atomic_helper_update_plane+0xac/0x350 [drm_kms_helper] __setplane_internal+0x2d6/0x840 [drm] drm_mode_cursor_universal+0x41e/0xbe0 [drm] drm_mode_cursor_common+0x49f/0x880 [drm] drm_mode_cursor_ioctl+0xd8/0x130 [drm] drm_ioctl_kernel+0x1d4/0x260 [drm] drm_ioctl+0x433/0x920 [drm] amdgpu_drm_ioctl+0x11d/0x290 [amdgpu] do_vfs_ioctl+0x1a1/0x13d0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x6f/0xb0 do_syscall_64+0x147/0x440 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The buggy address belongs to the object at ffff8803a697b068 which belongs to the cache kmalloc-1024 of size 1024 The buggy address is located 9 bytes inside of 1024-byte region [ffff8803a697b068, ffff8803a697b468) The buggy address belongs to the page: page:ffffea000e9a5e00 count:1 mapcount:0 mapping:ffff88041e00efc0 index:0x0 compound_mapcount: 0 flags: 0x8000000000008100(slab|head) raw: 8000000000008100 ffffea000ecbc208 ffff88041e000c70 ffff88041e00efc0 raw: 0000000000000000 0000000000170017 00000001ffffffff 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff8803a697af00: fb fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ffff8803a697af80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc >ffff8803a697b000: fc fc fc fc fc fc fc fc fc fc fc fc fc fb fb fb ^ ffff8803a697b080: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff8803a697b100: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ================================================================== So, we fix this by counting the number of CRTCs this atomic commit disabled early on in the function before their atomic states have been freed, then use that count later to do the appropriate number of RPM puts at the end of the function. Acked-by: Michel Dänzer <michel.daenzer@amd.com> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Cc: stable@vger.kernel.org Fixes: 97028037a38ae ("drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail()") Signed-off-by: Lyude Paul <lyude@redhat.com> Cc: Michel Dänzer <michel@daenzer.net> Reported-by: Michel Dänzer <michel@daenzer.net> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-06-21 14:48:26 -06:00
if (old_crtc_state->active && !new_crtc_state->active)
crtc_disable_count++;
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
dm_old_crtc_state = to_dm_crtc_state(old_crtc_state);
/* Update freesync active state. */
pre_update_freesync_state_on_stream(dm, dm_new_crtc_state);
/* Handle vrr on->off / off->on transitions */
amdgpu_dm_handle_vrr_transition(dm_old_crtc_state,
dm_new_crtc_state);
}
drm/amd/display: Split enabling CRTC interrupts into two passes [Why] When disabling all the pipes for a CRTC the page-flip interrupt also gets disabled on Raven. We can't re-enable the page-flip interrupt unless we give DC at least one active DC plane. We currently enable interrupts after the call to dc_commit_state since there's currently no valid sequence that should disable all the planes or re-enable planes for a CRTC without first going through dc_commit_state. If we were to allow for a CRTC to be enabled with no primary plane this would not be the case - the call to dc_commit_updates_for_stream would enable the planes when going from zero to at least one active plane, but manage_dm_interrupts would have been called too early. This results in a page-flip timeout on any subsequent commits since we think the page-flip are now enabled when they're actually disabled. We need to enable interrupts after the call to dc_commit_updates_for_stream. [How] Split enabling interrupts into two passes. One pass before dc_commit_updates_for_stream and one after it. Shifting all the interrupts to be strictly below the call doesn't currently work even though it should in theory. We end up queuing off the vblank event to be handle by the flip handler before it's actually enabled in some cases, particularly: old_crtc_state->active = false -> new_crtc_state->active = true The framebuffer states haven't changed and we can technically still do a "pageflip" in this case and send back the event. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-08 09:18:31 -06:00
/* Enable interrupts for CRTCs going through a modeset. */
amdgpu_dm_enable_crtc_interrupts(dev, state, true);
for_each_new_crtc_in_state(state, crtc, new_crtc_state, j)
if (new_crtc_state->async_flip)
wait_for_vblank = false;
/* update planes when needed per crtc*/
for_each_new_crtc_in_state(state, crtc, new_crtc_state, j) {
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
if (dm_new_crtc_state->stream)
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
amdgpu_dm_commit_planes(state, dc_state, dev,
dm, crtc, wait_for_vblank);
}
drm/amd/display: Split enabling CRTC interrupts into two passes [Why] When disabling all the pipes for a CRTC the page-flip interrupt also gets disabled on Raven. We can't re-enable the page-flip interrupt unless we give DC at least one active DC plane. We currently enable interrupts after the call to dc_commit_state since there's currently no valid sequence that should disable all the planes or re-enable planes for a CRTC without first going through dc_commit_state. If we were to allow for a CRTC to be enabled with no primary plane this would not be the case - the call to dc_commit_updates_for_stream would enable the planes when going from zero to at least one active plane, but manage_dm_interrupts would have been called too early. This results in a page-flip timeout on any subsequent commits since we think the page-flip are now enabled when they're actually disabled. We need to enable interrupts after the call to dc_commit_updates_for_stream. [How] Split enabling interrupts into two passes. One pass before dc_commit_updates_for_stream and one after it. Shifting all the interrupts to be strictly below the call doesn't currently work even though it should in theory. We end up queuing off the vblank event to be handle by the flip handler before it's actually enabled in some cases, particularly: old_crtc_state->active = false -> new_crtc_state->active = true The framebuffer states haven't changed and we can technically still do a "pageflip" in this case and send back the event. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-04-08 09:18:31 -06:00
/* Enable interrupts for CRTCs going from 0 to n active planes. */
amdgpu_dm_enable_crtc_interrupts(dev, state, false);
/* Update audio instances for each connector. */
amdgpu_dm_commit_audio(dev, state);
/*
* send vblank event on all events not handled in flip and
* mark consumed event for drm_atomic_helper_commit_hw_done
*/
spin_lock_irqsave(&adev->ddev->event_lock, flags);
for_each_new_crtc_in_state(state, crtc, new_crtc_state, i) {
if (new_crtc_state->event)
drm_send_event_locked(dev, &new_crtc_state->event->base);
new_crtc_state->event = NULL;
}
spin_unlock_irqrestore(&adev->ddev->event_lock, flags);
/* Signal HW programming completion */
drm_atomic_helper_commit_hw_done(state);
if (wait_for_vblank)
drm_atomic_helper_wait_for_flip_done(dev, state);
drm_atomic_helper_cleanup_planes(dev, state);
drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail() So, unfortunately I recently made the discovery that in the upstream kernel, the only reason that amdgpu is not currently suffering from issues with runtime PM putting the GPU into suspend while it's driving displays is due to the fact that on most prime systems, we have sound devices associated with the GPU that hold their own runtime PM ref for the GPU. What this means however, is that in the event that there isn't any kind of sound device active (which can easily be reproduced by building a kernel with sound drivers disabled), the GPU will fall asleep even when there's displays active. This appears to be in part due to the fact that amdgpu has not actually ever relied on it's rpm_idle() function to be the only thing keeping it running, and normally grabs it's own power references whenever there are displays active (as can be seen with the original pre-DC codepath in amdgpu_display_crtc_set_config() in amdgpu_display.c). This means it's very likely that this bug was introduced during the switch over the DC. So to fix this, we start grabbing runtime PM references every time we enable a previously disabled CRTC in atomic_commit_tail(). This appears to be the correct solution, as it matches up with what i915 does in i915/intel_runtime_pm.c. The one sideaffect of this is that we ignore the variable that the pre-DC code used to use for tracking when it needed runtime PM refs, adev->have_disp_power_ref. This is mainly because there's no way for a driver to tell whether or not all of it's CRTCs are enabled or disabled when we've begun committing an atomic state, as there may be CRTC commits happening in parallel that aren't contained within the atomic state being committed. So, it's safer to just get/put a reference for each CRTC being enabled or disabled in the new atomic state. Signed-off-by: Lyude Paul <lyude@redhat.com> Acked-by: Christian König <christian.koenig@amd.com>. Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org
2018-06-04 13:35:03 -06:00
/*
* Finally, drop a runtime PM reference for each newly disabled CRTC,
drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail() So, unfortunately I recently made the discovery that in the upstream kernel, the only reason that amdgpu is not currently suffering from issues with runtime PM putting the GPU into suspend while it's driving displays is due to the fact that on most prime systems, we have sound devices associated with the GPU that hold their own runtime PM ref for the GPU. What this means however, is that in the event that there isn't any kind of sound device active (which can easily be reproduced by building a kernel with sound drivers disabled), the GPU will fall asleep even when there's displays active. This appears to be in part due to the fact that amdgpu has not actually ever relied on it's rpm_idle() function to be the only thing keeping it running, and normally grabs it's own power references whenever there are displays active (as can be seen with the original pre-DC codepath in amdgpu_display_crtc_set_config() in amdgpu_display.c). This means it's very likely that this bug was introduced during the switch over the DC. So to fix this, we start grabbing runtime PM references every time we enable a previously disabled CRTC in atomic_commit_tail(). This appears to be the correct solution, as it matches up with what i915 does in i915/intel_runtime_pm.c. The one sideaffect of this is that we ignore the variable that the pre-DC code used to use for tracking when it needed runtime PM refs, adev->have_disp_power_ref. This is mainly because there's no way for a driver to tell whether or not all of it's CRTCs are enabled or disabled when we've begun committing an atomic state, as there may be CRTC commits happening in parallel that aren't contained within the atomic state being committed. So, it's safer to just get/put a reference for each CRTC being enabled or disabled in the new atomic state. Signed-off-by: Lyude Paul <lyude@redhat.com> Acked-by: Christian König <christian.koenig@amd.com>. Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org
2018-06-04 13:35:03 -06:00
* so we can put the GPU into runtime suspend if we're not driving any
* displays anymore
*/
drm/amdgpu: Count disabled CRTCs in commit tail earlier This fixes a regression I accidentally reduced that was picked up by kasan, where we were checking the CRTC atomic states after DRM's helpers had already freed them. Example: ================================================================== BUG: KASAN: use-after-free in amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] Read of size 1 at addr ffff8803a697b071 by task kworker/u16:0/7 CPU: 7 PID: 7 Comm: kworker/u16:0 Tainted: G O 4.18.0-rc1Lyude-Upstream+ #1 Hardware name: HP HP ZBook 15 G4/8275, BIOS P70 Ver. 01.21 05/02/2018 Workqueue: events_unbound commit_work [drm_kms_helper] Call Trace: dump_stack+0xc1/0x169 ? dump_stack_print_info.cold.1+0x42/0x42 ? kmsg_dump_rewind_nolock+0xd9/0xd9 ? printk+0x9f/0xc5 ? amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] print_address_description+0x6c/0x23c ? amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] kasan_report.cold.6+0x241/0x2fd amdgpu_dm_atomic_commit_tail.cold.50+0x13d/0x15a [amdgpu] ? commit_planes_to_stream.constprop.45+0x13b0/0x13b0 [amdgpu] ? cpu_load_update_active+0x290/0x290 ? finish_task_switch+0x2bd/0x840 ? __switch_to_asm+0x34/0x70 ? read_word_at_a_time+0xe/0x20 ? strscpy+0x14b/0x460 ? drm_atomic_helper_wait_for_dependencies+0x47d/0x7e0 [drm_kms_helper] commit_tail+0x96/0xe0 [drm_kms_helper] process_one_work+0x88a/0x1360 ? create_worker+0x540/0x540 ? __sched_text_start+0x8/0x8 ? move_queued_task+0x760/0x760 ? call_rcu_sched+0x20/0x20 ? vsnprintf+0xcda/0x1350 ? wait_woken+0x1c0/0x1c0 ? mutex_unlock+0x1d/0x40 ? init_timer_key+0x190/0x230 ? schedule+0xea/0x390 ? __schedule+0x1ea0/0x1ea0 ? need_to_create_worker+0xe4/0x210 ? init_worker_pool+0x700/0x700 ? try_to_del_timer_sync+0xbf/0x110 ? del_timer+0x120/0x120 ? __mutex_lock_slowpath+0x10/0x10 worker_thread+0x196/0x11f0 ? flush_rcu_work+0x50/0x50 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __switch_to_asm+0x34/0x70 ? __switch_to_asm+0x40/0x70 ? __schedule+0x7d6/0x1ea0 ? migrate_swap_stop+0x850/0x880 ? __sched_text_start+0x8/0x8 ? save_stack+0x8c/0xb0 ? kasan_kmalloc+0xbf/0xe0 ? kmem_cache_alloc_trace+0xe4/0x190 ? kthread+0x98/0x390 ? ret_from_fork+0x35/0x40 ? ret_from_fork+0x35/0x40 ? deactivate_slab.isra.67+0x3c4/0x5c0 ? kthread+0x98/0x390 ? kthread+0x98/0x390 ? set_track+0x76/0x120 ? schedule+0xea/0x390 ? __schedule+0x1ea0/0x1ea0 ? wait_woken+0x1c0/0x1c0 ? kasan_unpoison_shadow+0x30/0x40 ? parse_args.cold.15+0x17a/0x17a ? flush_rcu_work+0x50/0x50 kthread+0x2d4/0x390 ? kthread_create_worker_on_cpu+0xc0/0xc0 ret_from_fork+0x35/0x40 Allocated by task 1124: kasan_kmalloc+0xbf/0xe0 kmem_cache_alloc_trace+0xe4/0x190 dm_crtc_duplicate_state+0x78/0x130 [amdgpu] drm_atomic_get_crtc_state+0x147/0x410 [drm] page_flip_common+0x57/0x230 [drm_kms_helper] drm_atomic_helper_page_flip+0xa6/0x110 [drm_kms_helper] drm_mode_page_flip_ioctl+0xc4b/0x10a0 [drm] drm_ioctl_kernel+0x1d4/0x260 [drm] drm_ioctl+0x433/0x920 [drm] amdgpu_drm_ioctl+0x11d/0x290 [amdgpu] do_vfs_ioctl+0x1a1/0x13d0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x6f/0xb0 do_syscall_64+0x147/0x440 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Freed by task 1124: __kasan_slab_free+0x12e/0x180 kfree+0x92/0x1a0 drm_atomic_state_default_clear+0x315/0xc40 [drm] __drm_atomic_state_free+0x35/0xd0 [drm] drm_atomic_helper_update_plane+0xac/0x350 [drm_kms_helper] __setplane_internal+0x2d6/0x840 [drm] drm_mode_cursor_universal+0x41e/0xbe0 [drm] drm_mode_cursor_common+0x49f/0x880 [drm] drm_mode_cursor_ioctl+0xd8/0x130 [drm] drm_ioctl_kernel+0x1d4/0x260 [drm] drm_ioctl+0x433/0x920 [drm] amdgpu_drm_ioctl+0x11d/0x290 [amdgpu] do_vfs_ioctl+0x1a1/0x13d0 ksys_ioctl+0x60/0x90 __x64_sys_ioctl+0x6f/0xb0 do_syscall_64+0x147/0x440 entry_SYSCALL_64_after_hwframe+0x44/0xa9 The buggy address belongs to the object at ffff8803a697b068 which belongs to the cache kmalloc-1024 of size 1024 The buggy address is located 9 bytes inside of 1024-byte region [ffff8803a697b068, ffff8803a697b468) The buggy address belongs to the page: page:ffffea000e9a5e00 count:1 mapcount:0 mapping:ffff88041e00efc0 index:0x0 compound_mapcount: 0 flags: 0x8000000000008100(slab|head) raw: 8000000000008100 ffffea000ecbc208 ffff88041e000c70 ffff88041e00efc0 raw: 0000000000000000 0000000000170017 00000001ffffffff 0000000000000000 page dumped because: kasan: bad access detected Memory state around the buggy address: ffff8803a697af00: fb fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc ffff8803a697af80: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc >ffff8803a697b000: fc fc fc fc fc fc fc fc fc fc fc fc fc fb fb fb ^ ffff8803a697b080: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ffff8803a697b100: fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb fb ================================================================== So, we fix this by counting the number of CRTCs this atomic commit disabled early on in the function before their atomic states have been freed, then use that count later to do the appropriate number of RPM puts at the end of the function. Acked-by: Michel Dänzer <michel.daenzer@amd.com> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Cc: stable@vger.kernel.org Fixes: 97028037a38ae ("drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail()") Signed-off-by: Lyude Paul <lyude@redhat.com> Cc: Michel Dänzer <michel@daenzer.net> Reported-by: Michel Dänzer <michel@daenzer.net> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-06-21 14:48:26 -06:00
for (i = 0; i < crtc_disable_count; i++)
pm_runtime_put_autosuspend(dev->dev);
drm/amdgpu: Grab/put runtime PM references in atomic_commit_tail() So, unfortunately I recently made the discovery that in the upstream kernel, the only reason that amdgpu is not currently suffering from issues with runtime PM putting the GPU into suspend while it's driving displays is due to the fact that on most prime systems, we have sound devices associated with the GPU that hold their own runtime PM ref for the GPU. What this means however, is that in the event that there isn't any kind of sound device active (which can easily be reproduced by building a kernel with sound drivers disabled), the GPU will fall asleep even when there's displays active. This appears to be in part due to the fact that amdgpu has not actually ever relied on it's rpm_idle() function to be the only thing keeping it running, and normally grabs it's own power references whenever there are displays active (as can be seen with the original pre-DC codepath in amdgpu_display_crtc_set_config() in amdgpu_display.c). This means it's very likely that this bug was introduced during the switch over the DC. So to fix this, we start grabbing runtime PM references every time we enable a previously disabled CRTC in atomic_commit_tail(). This appears to be the correct solution, as it matches up with what i915 does in i915/intel_runtime_pm.c. The one sideaffect of this is that we ignore the variable that the pre-DC code used to use for tracking when it needed runtime PM refs, adev->have_disp_power_ref. This is mainly because there's no way for a driver to tell whether or not all of it's CRTCs are enabled or disabled when we've begun committing an atomic state, as there may be CRTC commits happening in parallel that aren't contained within the atomic state being committed. So, it's safer to just get/put a reference for each CRTC being enabled or disabled in the new atomic state. Signed-off-by: Lyude Paul <lyude@redhat.com> Acked-by: Christian König <christian.koenig@amd.com>. Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org
2018-06-04 13:35:03 -06:00
pm_runtime_mark_last_busy(dev->dev);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (dc_state_temp)
dc_release_state(dc_state_temp);
}
static int dm_force_atomic_commit(struct drm_connector *connector)
{
int ret = 0;
struct drm_device *ddev = connector->dev;
struct drm_atomic_state *state = drm_atomic_state_alloc(ddev);
struct amdgpu_crtc *disconnected_acrtc = to_amdgpu_crtc(connector->encoder->crtc);
struct drm_plane *plane = disconnected_acrtc->base.primary;
struct drm_connector_state *conn_state;
struct drm_crtc_state *crtc_state;
struct drm_plane_state *plane_state;
if (!state)
return -ENOMEM;
state->acquire_ctx = ddev->mode_config.acquire_ctx;
/* Construct an atomic state to restore previous display setting */
/*
* Attach connectors to drm_atomic_state
*/
conn_state = drm_atomic_get_connector_state(state, connector);
ret = PTR_ERR_OR_ZERO(conn_state);
if (ret)
goto err;
/* Attach crtc to drm_atomic_state*/
crtc_state = drm_atomic_get_crtc_state(state, &disconnected_acrtc->base);
ret = PTR_ERR_OR_ZERO(crtc_state);
if (ret)
goto err;
/* force a restore */
crtc_state->mode_changed = true;
/* Attach plane to drm_atomic_state */
plane_state = drm_atomic_get_plane_state(state, plane);
ret = PTR_ERR_OR_ZERO(plane_state);
if (ret)
goto err;
/* Call commit internally with the state we just constructed */
ret = drm_atomic_commit(state);
if (!ret)
return 0;
err:
DRM_ERROR("Restoring old state failed with %i\n", ret);
drm_atomic_state_put(state);
return ret;
}
/*
* This function handles all cases when set mode does not come upon hotplug.
* This includes when a display is unplugged then plugged back into the
* same port and when running without usermode desktop manager supprot
*/
void dm_restore_drm_connector_state(struct drm_device *dev,
struct drm_connector *connector)
{
struct amdgpu_dm_connector *aconnector = to_amdgpu_dm_connector(connector);
struct amdgpu_crtc *disconnected_acrtc;
struct dm_crtc_state *acrtc_state;
if (!aconnector->dc_sink || !connector->state || !connector->encoder)
return;
disconnected_acrtc = to_amdgpu_crtc(connector->encoder->crtc);
if (!disconnected_acrtc)
return;
acrtc_state = to_dm_crtc_state(disconnected_acrtc->base.state);
if (!acrtc_state->stream)
return;
/*
* If the previous sink is not released and different from the current,
* we deduce we are in a state where we can not rely on usermode call
* to turn on the display, so we do it here
*/
if (acrtc_state->stream->sink != aconnector->dc_sink)
dm_force_atomic_commit(&aconnector->base);
}
/*
* Grabs all modesetting locks to serialize against any blocking commits,
* Waits for completion of all non blocking commits.
*/
static int do_aquire_global_lock(struct drm_device *dev,
struct drm_atomic_state *state)
{
struct drm_crtc *crtc;
struct drm_crtc_commit *commit;
long ret;
/*
* Adding all modeset locks to aquire_ctx will
* ensure that when the framework release it the
* extra locks we are locking here will get released to
*/
ret = drm_modeset_lock_all_ctx(dev, state->acquire_ctx);
if (ret)
return ret;
list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
spin_lock(&crtc->commit_lock);
commit = list_first_entry_or_null(&crtc->commit_list,
struct drm_crtc_commit, commit_entry);
if (commit)
drm_crtc_commit_get(commit);
spin_unlock(&crtc->commit_lock);
if (!commit)
continue;
/*
* Make sure all pending HW programming completed and
* page flips done
*/
ret = wait_for_completion_interruptible_timeout(&commit->hw_done, 10*HZ);
if (ret > 0)
ret = wait_for_completion_interruptible_timeout(
&commit->flip_done, 10*HZ);
if (ret == 0)
DRM_ERROR("[CRTC:%d:%s] hw_done or flip_done "
"timed out\n", crtc->base.id, crtc->name);
drm_crtc_commit_put(commit);
}
return ret < 0 ? ret : 0;
}
static void get_freesync_config_for_crtc(
struct dm_crtc_state *new_crtc_state,
struct dm_connector_state *new_con_state)
{
struct mod_freesync_config config = {0};
struct amdgpu_dm_connector *aconnector =
to_amdgpu_dm_connector(new_con_state->base.connector);
struct drm_display_mode *mode = &new_crtc_state->base.mode;
int vrefresh = drm_mode_vrefresh(mode);
new_crtc_state->vrr_supported = new_con_state->freesync_capable &&
vrefresh >= aconnector->min_vfreq &&
vrefresh <= aconnector->max_vfreq;
if (new_crtc_state->vrr_supported) {
new_crtc_state->stream->ignore_msa_timing_param = true;
config.state = new_crtc_state->base.vrr_enabled ?
VRR_STATE_ACTIVE_VARIABLE :
VRR_STATE_INACTIVE;
config.min_refresh_in_uhz =
aconnector->min_vfreq * 1000000;
config.max_refresh_in_uhz =
aconnector->max_vfreq * 1000000;
config.vsif_supported = true;
config.btr = true;
}
new_crtc_state->freesync_config = config;
}
static void reset_freesync_config_for_crtc(
struct dm_crtc_state *new_crtc_state)
{
new_crtc_state->vrr_supported = false;
memset(&new_crtc_state->vrr_params, 0,
sizeof(new_crtc_state->vrr_params));
memset(&new_crtc_state->vrr_infopacket, 0,
sizeof(new_crtc_state->vrr_infopacket));
}
static int dm_update_crtc_state(struct amdgpu_display_manager *dm,
struct drm_atomic_state *state,
struct drm_crtc *crtc,
struct drm_crtc_state *old_crtc_state,
struct drm_crtc_state *new_crtc_state,
bool enable,
bool *lock_and_validation_needed)
{
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct dm_atomic_state *dm_state = NULL;
struct dm_crtc_state *dm_old_crtc_state, *dm_new_crtc_state;
drm/amd/display: Don't leak dc_stream_state. Noticed while playing “Valley”, which was causing some 8MB of leakage per second. kmemleak listed many entries looking like this: unreferenced object 0xffff8802c2951800 (size 1024): comm "Xorg", pid 2982, jiffies 4297410155 (age 392.787s) hex dump (first 32 bytes): 00 50 f9 0c 04 88 ff ff 98 08 00 00 00 00 00 00 .P.............. 80 07 00 00 00 00 00 00 58 00 00 00 2c 00 00 00 ........X...,... backtrace: [<ffffffff810cd4c3>] create_object+0x13c/0x261 [<ffffffff815abdc2>] kmemleak_alloc+0x20/0x3c [<ffffffff810cad1d>] slab_post_alloc_hook+0x42/0x52 [<ffffffff810cb8e0>] kmem_cache_alloc+0x67/0x76 [<ffffffff813bbb54>] dc_create_stream_for_sink+0x24/0x1cf [<ffffffff81373aaa>] create_stream_for_sink+0x6f/0x295 [<ffffffff81373dc2>] dm_update_crtcs_state+0xa6/0x268 [<ffffffff8137401e>] amdgpu_dm_atomic_check+0x9a/0x314 [<ffffffff812ac3dd>] drm_atomic_check_only+0x17a/0x42d [<ffffffff812ac6a3>] drm_atomic_commit+0x13/0x4b [<ffffffff812ad1a5>] drm_atomic_connector_commit_dpms+0xcb/0xe8 [<ffffffff812b1238>] drm_mode_obj_set_property_ioctl+0xe6/0x1e3 [<ffffffff812b027b>] drm_mode_connector_property_set_ioctl+0x2b/0x2d [<ffffffff8129f427>] drm_ioctl_kernel+0x64/0x9d [<ffffffff8129f6a2>] drm_ioctl+0x230/0x316 [<ffffffff812ca4d3>] amdgpu_drm_ioctl+0x4b/0x7d v2: also handle break statements. Signed-off-by: Darren Salt <devspam@moreofthesa.me.uk> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2017-09-12 10:10:25 -06:00
struct dc_stream_state *new_stream;
int ret = 0;
/*
* TODO Move this code into dm_crtc_atomic_check once we get rid of dc_validation_set
* update changed items
*/
struct amdgpu_crtc *acrtc = NULL;
struct amdgpu_dm_connector *aconnector = NULL;
struct drm_connector_state *drm_new_conn_state = NULL, *drm_old_conn_state = NULL;
struct dm_connector_state *dm_new_conn_state = NULL, *dm_old_conn_state = NULL;
new_stream = NULL;
drm/amd/display: Don't leak dc_stream_state. Noticed while playing “Valley”, which was causing some 8MB of leakage per second. kmemleak listed many entries looking like this: unreferenced object 0xffff8802c2951800 (size 1024): comm "Xorg", pid 2982, jiffies 4297410155 (age 392.787s) hex dump (first 32 bytes): 00 50 f9 0c 04 88 ff ff 98 08 00 00 00 00 00 00 .P.............. 80 07 00 00 00 00 00 00 58 00 00 00 2c 00 00 00 ........X...,... backtrace: [<ffffffff810cd4c3>] create_object+0x13c/0x261 [<ffffffff815abdc2>] kmemleak_alloc+0x20/0x3c [<ffffffff810cad1d>] slab_post_alloc_hook+0x42/0x52 [<ffffffff810cb8e0>] kmem_cache_alloc+0x67/0x76 [<ffffffff813bbb54>] dc_create_stream_for_sink+0x24/0x1cf [<ffffffff81373aaa>] create_stream_for_sink+0x6f/0x295 [<ffffffff81373dc2>] dm_update_crtcs_state+0xa6/0x268 [<ffffffff8137401e>] amdgpu_dm_atomic_check+0x9a/0x314 [<ffffffff812ac3dd>] drm_atomic_check_only+0x17a/0x42d [<ffffffff812ac6a3>] drm_atomic_commit+0x13/0x4b [<ffffffff812ad1a5>] drm_atomic_connector_commit_dpms+0xcb/0xe8 [<ffffffff812b1238>] drm_mode_obj_set_property_ioctl+0xe6/0x1e3 [<ffffffff812b027b>] drm_mode_connector_property_set_ioctl+0x2b/0x2d [<ffffffff8129f427>] drm_ioctl_kernel+0x64/0x9d [<ffffffff8129f6a2>] drm_ioctl+0x230/0x316 [<ffffffff812ca4d3>] amdgpu_drm_ioctl+0x4b/0x7d v2: also handle break statements. Signed-off-by: Darren Salt <devspam@moreofthesa.me.uk> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2017-09-12 10:10:25 -06:00
dm_old_crtc_state = to_dm_crtc_state(old_crtc_state);
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
acrtc = to_amdgpu_crtc(crtc);
aconnector = amdgpu_dm_find_first_crtc_matching_connector(state, crtc);
/* TODO This hack should go away */
if (aconnector && enable) {
/* Make sure fake sink is created in plug-in scenario */
drm_new_conn_state = drm_atomic_get_new_connector_state(state,
&aconnector->base);
drm_old_conn_state = drm_atomic_get_old_connector_state(state,
&aconnector->base);
if (IS_ERR(drm_new_conn_state)) {
ret = PTR_ERR_OR_ZERO(drm_new_conn_state);
goto fail;
}
dm_new_conn_state = to_dm_connector_state(drm_new_conn_state);
dm_old_conn_state = to_dm_connector_state(drm_old_conn_state);
if (!drm_atomic_crtc_needs_modeset(new_crtc_state))
goto skip_modeset;
new_stream = create_stream_for_sink(aconnector,
&new_crtc_state->mode,
dm_new_conn_state,
dm_old_crtc_state->stream);
/*
* we can have no stream on ACTION_SET if a display
* was disconnected during S3, in this case it is not an
* error, the OS will be updated after detection, and
* will do the right thing on next atomic commit
*/
if (!new_stream) {
DRM_DEBUG_DRIVER("%s: Failed to create new stream for crtc %d\n",
__func__, acrtc->base.base.id);
ret = -ENOMEM;
goto fail;
}
dm_new_crtc_state->abm_level = dm_new_conn_state->abm_level;
ret = fill_hdr_info_packet(drm_new_conn_state,
&new_stream->hdr_static_metadata);
if (ret)
goto fail;
/*
* If we already removed the old stream from the context
* (and set the new stream to NULL) then we can't reuse
* the old stream even if the stream and scaling are unchanged.
* We'll hit the BUG_ON and black screen.
*
* TODO: Refactor this function to allow this check to work
* in all conditions.
*/
if (dm_new_crtc_state->stream &&
dc_is_stream_unchanged(new_stream, dm_old_crtc_state->stream) &&
dc_is_stream_scaling_unchanged(new_stream, dm_old_crtc_state->stream)) {
new_crtc_state->mode_changed = false;
DRM_DEBUG_DRIVER("Mode change not required, setting mode_changed to %d",
new_crtc_state->mode_changed);
}
}
/* mode_changed flag may get updated above, need to check again */
if (!drm_atomic_crtc_needs_modeset(new_crtc_state))
goto skip_modeset;
DRM_DEBUG_DRIVER(
"amdgpu_crtc id:%d crtc_state_flags: enable:%d, active:%d, "
"planes_changed:%d, mode_changed:%d,active_changed:%d,"
"connectors_changed:%d\n",
acrtc->crtc_id,
new_crtc_state->enable,
new_crtc_state->active,
new_crtc_state->planes_changed,
new_crtc_state->mode_changed,
new_crtc_state->active_changed,
new_crtc_state->connectors_changed);
/* Remove stream for any changed/disabled CRTC */
if (!enable) {
if (!dm_old_crtc_state->stream)
goto skip_modeset;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
ret = dm_atomic_get_state(state, &dm_state);
if (ret)
goto fail;
DRM_DEBUG_DRIVER("Disabling DRM crtc: %d\n",
crtc->base.id);
/* i.e. reset mode */
if (dc_remove_stream_from_ctx(
dm->dc,
dm_state->context,
dm_old_crtc_state->stream) != DC_OK) {
ret = -EINVAL;
goto fail;
}
dc_stream_release(dm_old_crtc_state->stream);
dm_new_crtc_state->stream = NULL;
reset_freesync_config_for_crtc(dm_new_crtc_state);
*lock_and_validation_needed = true;
} else {/* Add stream for any updated/enabled CRTC */
/*
* Quick fix to prevent NULL pointer on new_stream when
* added MST connectors not found in existing crtc_state in the chained mode
* TODO: need to dig out the root cause of that
*/
if (!aconnector || (!aconnector->dc_sink && aconnector->mst_port))
goto skip_modeset;
if (modereset_required(new_crtc_state))
goto skip_modeset;
if (modeset_required(new_crtc_state, new_stream,
dm_old_crtc_state->stream)) {
WARN_ON(dm_new_crtc_state->stream);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
ret = dm_atomic_get_state(state, &dm_state);
if (ret)
goto fail;
dm_new_crtc_state->stream = new_stream;
dc_stream_retain(new_stream);
DRM_DEBUG_DRIVER("Enabling DRM crtc: %d\n",
crtc->base.id);
if (dc_add_stream_to_ctx(
dm->dc,
dm_state->context,
dm_new_crtc_state->stream) != DC_OK) {
ret = -EINVAL;
goto fail;
}
*lock_and_validation_needed = true;
}
}
skip_modeset:
/* Release extra reference */
if (new_stream)
dc_stream_release(new_stream);
/*
* We want to do dc stream updates that do not require a
* full modeset below.
*/
if (!(enable && aconnector && new_crtc_state->enable &&
new_crtc_state->active))
return 0;
/*
* Given above conditions, the dc state cannot be NULL because:
* 1. We're in the process of enabling CRTCs (just been added
* to the dc context, or already is on the context)
* 2. Has a valid connector attached, and
* 3. Is currently active and enabled.
* => The dc stream state currently exists.
*/
BUG_ON(dm_new_crtc_state->stream == NULL);
/* Scaling or underscan settings */
if (is_scaling_state_different(dm_old_conn_state, dm_new_conn_state))
update_stream_scaling_settings(
&new_crtc_state->mode, dm_new_conn_state, dm_new_crtc_state->stream);
/* ABM settings */
dm_new_crtc_state->abm_level = dm_new_conn_state->abm_level;
/*
* Color management settings. We also update color properties
* when a modeset is needed, to ensure it gets reprogrammed.
*/
if (dm_new_crtc_state->base.color_mgmt_changed ||
drm_atomic_crtc_needs_modeset(new_crtc_state)) {
drm/amd/display: Rework CRTC color management [Why] To prepare for the upcoming DRM plane color management properties we need to correct a lot of wrong behavior and assumptions made for CRTC color management. The documentation added by this commit in amdgpu_dm_color explains how the HW color pipeline works and its limitations with the DRM interface. The current implementation does the following wrong: - Implicit sRGB DGM when no CRTC DGM is set - Implicit sRGB RGM when no CRTC RGM is set - No way to specify a non-linear DGM matrix that produces correct output - No way to specify a correct RGM when a linear DGM is used We had workarounds for passing kms_color tests but not all of the behavior we had wrong was covered by these tests (especially when it comes to non-linear DGM). Testing both DGM and RGM at the same time isn't something kms_color tests well either. [How] The specifics for how color management works in AMDGPU and the new behavior can be found by reading the documentation added to amdgpu_dm_color.c from this patch. All of the incorrect cases from the old implementation have been addressed for the atomic interface, but there still a few TODOs for the legacy one. Note: this does cause regressions for kms_color@pipe-a-ctm-* over HDMI. The result looks correct from visual inspection but the CRC no longer matches. For reference, the test was previously doing the following: linear degamma -> CTM -> sRGB regamma -> RGB to YUV (709) -> ... Now the test is doing: linear degamma -> CTM -> linear regamma -> RGB to YUV (709) -> ... Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-05-09 10:14:58 -06:00
ret = amdgpu_dm_update_crtc_color_mgmt(dm_new_crtc_state);
if (ret)
goto fail;
}
/* Update Freesync settings. */
get_freesync_config_for_crtc(dm_new_crtc_state,
dm_new_conn_state);
return ret;
drm/amd/display: Don't leak dc_stream_state. Noticed while playing “Valley”, which was causing some 8MB of leakage per second. kmemleak listed many entries looking like this: unreferenced object 0xffff8802c2951800 (size 1024): comm "Xorg", pid 2982, jiffies 4297410155 (age 392.787s) hex dump (first 32 bytes): 00 50 f9 0c 04 88 ff ff 98 08 00 00 00 00 00 00 .P.............. 80 07 00 00 00 00 00 00 58 00 00 00 2c 00 00 00 ........X...,... backtrace: [<ffffffff810cd4c3>] create_object+0x13c/0x261 [<ffffffff815abdc2>] kmemleak_alloc+0x20/0x3c [<ffffffff810cad1d>] slab_post_alloc_hook+0x42/0x52 [<ffffffff810cb8e0>] kmem_cache_alloc+0x67/0x76 [<ffffffff813bbb54>] dc_create_stream_for_sink+0x24/0x1cf [<ffffffff81373aaa>] create_stream_for_sink+0x6f/0x295 [<ffffffff81373dc2>] dm_update_crtcs_state+0xa6/0x268 [<ffffffff8137401e>] amdgpu_dm_atomic_check+0x9a/0x314 [<ffffffff812ac3dd>] drm_atomic_check_only+0x17a/0x42d [<ffffffff812ac6a3>] drm_atomic_commit+0x13/0x4b [<ffffffff812ad1a5>] drm_atomic_connector_commit_dpms+0xcb/0xe8 [<ffffffff812b1238>] drm_mode_obj_set_property_ioctl+0xe6/0x1e3 [<ffffffff812b027b>] drm_mode_connector_property_set_ioctl+0x2b/0x2d [<ffffffff8129f427>] drm_ioctl_kernel+0x64/0x9d [<ffffffff8129f6a2>] drm_ioctl+0x230/0x316 [<ffffffff812ca4d3>] amdgpu_drm_ioctl+0x4b/0x7d v2: also handle break statements. Signed-off-by: Darren Salt <devspam@moreofthesa.me.uk> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2017-09-12 10:10:25 -06:00
fail:
if (new_stream)
dc_stream_release(new_stream);
return ret;
}
static bool should_reset_plane(struct drm_atomic_state *state,
struct drm_plane *plane,
struct drm_plane_state *old_plane_state,
struct drm_plane_state *new_plane_state)
{
struct drm_plane *other;
struct drm_plane_state *old_other_state, *new_other_state;
struct drm_crtc_state *new_crtc_state;
int i;
/*
* TODO: Remove this hack once the checks below are sufficient
* enough to determine when we need to reset all the planes on
* the stream.
*/
if (state->allow_modeset)
return true;
/* Exit early if we know that we're adding or removing the plane. */
if (old_plane_state->crtc != new_plane_state->crtc)
return true;
/* old crtc == new_crtc == NULL, plane not in context. */
if (!new_plane_state->crtc)
return false;
new_crtc_state =
drm_atomic_get_new_crtc_state(state, new_plane_state->crtc);
if (!new_crtc_state)
return true;
/* CRTC Degamma changes currently require us to recreate planes. */
if (new_crtc_state->color_mgmt_changed)
return true;
if (drm_atomic_crtc_needs_modeset(new_crtc_state))
return true;
/*
* If there are any new primary or overlay planes being added or
* removed then the z-order can potentially change. To ensure
* correct z-order and pipe acquisition the current DC architecture
* requires us to remove and recreate all existing planes.
*
* TODO: Come up with a more elegant solution for this.
*/
for_each_oldnew_plane_in_state(state, other, old_other_state, new_other_state, i) {
if (other->type == DRM_PLANE_TYPE_CURSOR)
continue;
if (old_other_state->crtc != new_plane_state->crtc &&
new_other_state->crtc != new_plane_state->crtc)
continue;
if (old_other_state->crtc != new_other_state->crtc)
return true;
/* TODO: Remove this once we can handle fast format changes. */
if (old_other_state->fb && new_other_state->fb &&
old_other_state->fb->format != new_other_state->fb->format)
return true;
}
return false;
}
static int dm_update_plane_state(struct dc *dc,
struct drm_atomic_state *state,
struct drm_plane *plane,
struct drm_plane_state *old_plane_state,
struct drm_plane_state *new_plane_state,
bool enable,
bool *lock_and_validation_needed)
{
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct dm_atomic_state *dm_state = NULL;
struct drm_crtc *new_plane_crtc, *old_plane_crtc;
struct drm_crtc_state *old_crtc_state, *new_crtc_state;
struct dm_crtc_state *dm_new_crtc_state, *dm_old_crtc_state;
struct dm_plane_state *dm_new_plane_state, *dm_old_plane_state;
struct amdgpu_crtc *new_acrtc;
bool needs_reset;
int ret = 0;
new_plane_crtc = new_plane_state->crtc;
old_plane_crtc = old_plane_state->crtc;
dm_new_plane_state = to_dm_plane_state(new_plane_state);
dm_old_plane_state = to_dm_plane_state(old_plane_state);
/*TODO Implement better atomic check for cursor plane */
if (plane->type == DRM_PLANE_TYPE_CURSOR) {
if (!enable || !new_plane_crtc ||
drm_atomic_plane_disabling(plane->state, new_plane_state))
return 0;
new_acrtc = to_amdgpu_crtc(new_plane_crtc);
if ((new_plane_state->crtc_w > new_acrtc->max_cursor_width) ||
(new_plane_state->crtc_h > new_acrtc->max_cursor_height)) {
DRM_DEBUG_ATOMIC("Bad cursor size %d x %d\n",
new_plane_state->crtc_w, new_plane_state->crtc_h);
return -EINVAL;
}
return 0;
}
needs_reset = should_reset_plane(state, plane, old_plane_state,
new_plane_state);
/* Remove any changed/removed planes */
if (!enable) {
if (!needs_reset)
return 0;
if (!old_plane_crtc)
return 0;
old_crtc_state = drm_atomic_get_old_crtc_state(
state, old_plane_crtc);
dm_old_crtc_state = to_dm_crtc_state(old_crtc_state);
if (!dm_old_crtc_state->stream)
return 0;
DRM_DEBUG_ATOMIC("Disabling DRM plane: %d on DRM crtc %d\n",
plane->base.id, old_plane_crtc->base.id);
ret = dm_atomic_get_state(state, &dm_state);
if (ret)
return ret;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (!dc_remove_plane_from_context(
dc,
dm_old_crtc_state->stream,
dm_old_plane_state->dc_state,
dm_state->context)) {
return -EINVAL;
}
dc_plane_state_release(dm_old_plane_state->dc_state);
dm_new_plane_state->dc_state = NULL;
*lock_and_validation_needed = true;
} else { /* Add new planes */
struct dc_plane_state *dc_new_plane_state;
if (drm_atomic_plane_disabling(plane->state, new_plane_state))
return 0;
if (!new_plane_crtc)
return 0;
new_crtc_state = drm_atomic_get_new_crtc_state(state, new_plane_crtc);
dm_new_crtc_state = to_dm_crtc_state(new_crtc_state);
if (!dm_new_crtc_state->stream)
return 0;
if (!needs_reset)
return 0;
WARN_ON(dm_new_plane_state->dc_state);
dc_new_plane_state = dc_create_plane_state(dc);
if (!dc_new_plane_state)
return -ENOMEM;
DRM_DEBUG_DRIVER("Enabling DRM plane: %d on DRM crtc %d\n",
plane->base.id, new_plane_crtc->base.id);
drm/amd/display: Rework DC plane filling and surface updates [Why] We currently don't do DC validation for medium or full updates where the plane state isn't created. There are some medium and full updates that can cause bandwidth or clock changes to occur resulting in underflow or corruption. We need to be able to fill surface and plane info updates during atomic commit for dm_determine_update_type for commit. Since we already do this during atomic commit tail it would be good if we had the same logic in both places for creating these structures. [How] Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr. These two functions cover the following three update structures: - struct dc_scaling_info - struct dc_plane_info - struct dc_plane_address Cleanup and adapter the existing fill_plane_* helpers to work with these functions. Update call sites that used most of these sub helpers directly to work with the new functions. The exception being prepare_fb - we just want the new buffer attributes specifically in the case where we're creating the plane. This is needed for dc_commit_state in the case where the FB hasn't been previously been used. This isn't quite a refactor, but functionally driver behavior should be mostly the smae as before. The one exception is that we now check the return code for fill_plane_buffer_attributes which means that commits will be rejected that try to enable DCC with erroneous parameters. Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: David Francis <David.Francis@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-03-28 12:45:19 -06:00
ret = fill_dc_plane_attributes(
new_plane_crtc->dev->dev_private,
dc_new_plane_state,
new_plane_state,
new_crtc_state);
if (ret) {
dc_plane_state_release(dc_new_plane_state);
return ret;
}
ret = dm_atomic_get_state(state, &dm_state);
if (ret) {
dc_plane_state_release(dc_new_plane_state);
return ret;
}
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
/*
* Any atomic check errors that occur after this will
* not need a release. The plane state will be attached
* to the stream, and therefore part of the atomic
* state. It'll be released when the atomic state is
* cleaned.
*/
if (!dc_add_plane_to_context(
dc,
dm_new_crtc_state->stream,
dc_new_plane_state,
dm_state->context)) {
dc_plane_state_release(dc_new_plane_state);
return -EINVAL;
}
dm_new_plane_state->dc_state = dc_new_plane_state;
/* Tell DC to do a full surface update every time there
* is a plane change. Inefficient, but works for now.
*/
dm_new_plane_state->dc_state->update_flags.bits.full_update = 1;
*lock_and_validation_needed = true;
}
return ret;
}
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
static int
dm_determine_update_type_for_commit(struct amdgpu_display_manager *dm,
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct drm_atomic_state *state,
enum surface_update_type *out_type)
{
struct dc *dc = dm->dc;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct dm_atomic_state *dm_state = NULL, *old_dm_state = NULL;
int i, j, num_plane, ret = 0;
struct drm_plane_state *old_plane_state, *new_plane_state;
struct dm_plane_state *new_dm_plane_state, *old_dm_plane_state;
struct drm_crtc *new_plane_crtc, *old_plane_crtc;
struct drm_plane *plane;
struct drm_crtc *crtc;
struct drm_crtc_state *new_crtc_state, *old_crtc_state;
struct dm_crtc_state *new_dm_crtc_state, *old_dm_crtc_state;
struct dc_stream_status *status = NULL;
struct dc_surface_update *updates;
enum surface_update_type update_type = UPDATE_TYPE_FAST;
updates = kcalloc(MAX_SURFACES, sizeof(*updates), GFP_KERNEL);
if (!updates) {
DRM_ERROR("Failed to allocate plane updates\n");
/* Set type to FULL to avoid crashing in DC*/
update_type = UPDATE_TYPE_FULL;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
goto cleanup;
}
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
struct dc_scaling_info scaling_info;
struct dc_stream_update stream_update;
memset(&stream_update, 0, sizeof(stream_update));
new_dm_crtc_state = to_dm_crtc_state(new_crtc_state);
old_dm_crtc_state = to_dm_crtc_state(old_crtc_state);
num_plane = 0;
if (new_dm_crtc_state->stream != old_dm_crtc_state->stream) {
update_type = UPDATE_TYPE_FULL;
goto cleanup;
}
if (!new_dm_crtc_state->stream)
continue;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, j) {
const struct amdgpu_framebuffer *amdgpu_fb =
to_amdgpu_framebuffer(new_plane_state->fb);
struct dc_plane_info plane_info;
struct dc_flip_addrs flip_addr;
uint64_t tiling_flags;
new_plane_crtc = new_plane_state->crtc;
old_plane_crtc = old_plane_state->crtc;
new_dm_plane_state = to_dm_plane_state(new_plane_state);
old_dm_plane_state = to_dm_plane_state(old_plane_state);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (plane->type == DRM_PLANE_TYPE_CURSOR)
continue;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (new_dm_plane_state->dc_state != old_dm_plane_state->dc_state) {
update_type = UPDATE_TYPE_FULL;
goto cleanup;
}
if (crtc != new_plane_crtc)
continue;
updates[num_plane].surface = new_dm_plane_state->dc_state;
if (new_crtc_state->mode_changed) {
stream_update.dst = new_dm_crtc_state->stream->dst;
stream_update.src = new_dm_crtc_state->stream->src;
}
if (new_crtc_state->color_mgmt_changed) {
updates[num_plane].gamma =
new_dm_plane_state->dc_state->gamma_correction;
updates[num_plane].in_transfer_func =
new_dm_plane_state->dc_state->in_transfer_func;
stream_update.gamut_remap =
&new_dm_crtc_state->stream->gamut_remap_matrix;
drm/amd/display: Rework CRTC color management [Why] To prepare for the upcoming DRM plane color management properties we need to correct a lot of wrong behavior and assumptions made for CRTC color management. The documentation added by this commit in amdgpu_dm_color explains how the HW color pipeline works and its limitations with the DRM interface. The current implementation does the following wrong: - Implicit sRGB DGM when no CRTC DGM is set - Implicit sRGB RGM when no CRTC RGM is set - No way to specify a non-linear DGM matrix that produces correct output - No way to specify a correct RGM when a linear DGM is used We had workarounds for passing kms_color tests but not all of the behavior we had wrong was covered by these tests (especially when it comes to non-linear DGM). Testing both DGM and RGM at the same time isn't something kms_color tests well either. [How] The specifics for how color management works in AMDGPU and the new behavior can be found by reading the documentation added to amdgpu_dm_color.c from this patch. All of the incorrect cases from the old implementation have been addressed for the atomic interface, but there still a few TODOs for the legacy one. Note: this does cause regressions for kms_color@pipe-a-ctm-* over HDMI. The result looks correct from visual inspection but the CRC no longer matches. For reference, the test was previously doing the following: linear degamma -> CTM -> sRGB regamma -> RGB to YUV (709) -> ... Now the test is doing: linear degamma -> CTM -> linear regamma -> RGB to YUV (709) -> ... Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com> Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-05-09 10:14:58 -06:00
stream_update.output_csc_transform =
&new_dm_crtc_state->stream->csc_color_matrix;
stream_update.out_transfer_func =
new_dm_crtc_state->stream->out_transfer_func;
}
ret = fill_dc_scaling_info(new_plane_state,
&scaling_info);
if (ret)
goto cleanup;
updates[num_plane].scaling_info = &scaling_info;
if (amdgpu_fb) {
ret = get_fb_info(amdgpu_fb, &tiling_flags);
if (ret)
goto cleanup;
memset(&flip_addr, 0, sizeof(flip_addr));
ret = fill_dc_plane_info_and_addr(
dm->adev, new_plane_state, tiling_flags,
&plane_info,
&flip_addr.address,
false);
if (ret)
goto cleanup;
updates[num_plane].plane_info = &plane_info;
updates[num_plane].flip_addr = &flip_addr;
}
num_plane++;
}
if (num_plane == 0)
continue;
ret = dm_atomic_get_state(state, &dm_state);
if (ret)
goto cleanup;
old_dm_state = dm_atomic_get_old_state(state);
if (!old_dm_state) {
ret = -EINVAL;
goto cleanup;
}
status = dc_stream_get_status_from_state(old_dm_state->context,
new_dm_crtc_state->stream);
/*
* TODO: DC modifies the surface during this call so we need
* to lock here - find a way to do this without locking.
*/
mutex_lock(&dm->dc_lock);
update_type = dc_check_update_surfaces_for_stream(dc, updates, num_plane,
&stream_update, status);
mutex_unlock(&dm->dc_lock);
if (update_type > UPDATE_TYPE_MED) {
update_type = UPDATE_TYPE_FULL;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
goto cleanup;
}
}
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
cleanup:
kfree(updates);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
*out_type = update_type;
return ret;
}
#if defined(CONFIG_DRM_AMD_DC_DCN)
static int add_affected_mst_dsc_crtcs(struct drm_atomic_state *state, struct drm_crtc *crtc)
{
struct drm_connector *connector;
struct drm_connector_state *conn_state;
struct amdgpu_dm_connector *aconnector = NULL;
int i;
for_each_new_connector_in_state(state, connector, conn_state, i) {
if (conn_state->crtc != crtc)
continue;
aconnector = to_amdgpu_dm_connector(connector);
if (!aconnector->port || !aconnector->mst_port)
aconnector = NULL;
else
break;
}
if (!aconnector)
return 0;
return drm_dp_mst_add_affected_dsc_crtcs(state, &aconnector->mst_port->mst_mgr);
}
#endif
/**
* amdgpu_dm_atomic_check() - Atomic check implementation for AMDgpu DM.
* @dev: The DRM device
* @state: The atomic state to commit
*
* Validate that the given atomic state is programmable by DC into hardware.
* This involves constructing a &struct dc_state reflecting the new hardware
* state we wish to commit, then querying DC to see if it is programmable. It's
* important not to modify the existing DC state. Otherwise, atomic_check
* may unexpectedly commit hardware changes.
*
* When validating the DC state, it's important that the right locks are
* acquired. For full updates case which removes/adds/updates streams on one
* CRTC while flipping on another CRTC, acquiring global lock will guarantee
* that any such full update commit will wait for completion of any outstanding
* flip using DRMs synchronization events. See
* dm_determine_update_type_for_commit()
*
* Note that DM adds the affected connectors for all CRTCs in state, when that
* might not seem necessary. This is because DC stream creation requires the
* DC sink, which is tied to the DRM connector state. Cleaning this up should
* be possible but non-trivial - a possible TODO item.
*
* Return: -Error code if validation failed.
*/
static int amdgpu_dm_atomic_check(struct drm_device *dev,
struct drm_atomic_state *state)
{
struct amdgpu_device *adev = dev->dev_private;
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
struct dm_atomic_state *dm_state = NULL;
struct dc *dc = adev->dm.dc;
struct drm_connector *connector;
struct drm_connector_state *old_con_state, *new_con_state;
struct drm_crtc *crtc;
struct drm_crtc_state *old_crtc_state, *new_crtc_state;
struct drm_plane *plane;
struct drm_plane_state *old_plane_state, *new_plane_state;
enum surface_update_type update_type = UPDATE_TYPE_FAST;
enum surface_update_type overall_update_type = UPDATE_TYPE_FAST;
int ret, i;
/*
* This bool will be set for true for any modeset/reset
* or plane update which implies non fast surface update.
*/
bool lock_and_validation_needed = false;
ret = drm_atomic_helper_check_modeset(dev, state);
if (ret)
goto fail;
/* Check connector changes */
for_each_oldnew_connector_in_state(state, connector, old_con_state, new_con_state, i) {
struct dm_connector_state *dm_old_con_state = to_dm_connector_state(old_con_state);
struct dm_connector_state *dm_new_con_state = to_dm_connector_state(new_con_state);
/* Skip connectors that are disabled or part of modeset already. */
if (!old_con_state->crtc && !new_con_state->crtc)
continue;
if (!new_con_state->crtc)
continue;
new_crtc_state = drm_atomic_get_crtc_state(state, new_con_state->crtc);
if (IS_ERR(new_crtc_state)) {
ret = PTR_ERR(new_crtc_state);
goto fail;
}
if (dm_old_con_state->abm_level !=
dm_new_con_state->abm_level)
new_crtc_state->connectors_changed = true;
}
#if defined(CONFIG_DRM_AMD_DC_DCN)
if (adev->asic_type >= CHIP_NAVI10) {
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
if (drm_atomic_crtc_needs_modeset(new_crtc_state)) {
ret = add_affected_mst_dsc_crtcs(state, crtc);
if (ret)
goto fail;
}
}
}
#endif
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
if (!drm_atomic_crtc_needs_modeset(new_crtc_state) &&
!new_crtc_state->color_mgmt_changed &&
old_crtc_state->vrr_enabled == new_crtc_state->vrr_enabled)
continue;
if (!new_crtc_state->enable)
continue;
ret = drm_atomic_add_affected_connectors(state, crtc);
if (ret)
return ret;
ret = drm_atomic_add_affected_planes(state, crtc);
if (ret)
goto fail;
}
/*
* Add all primary and overlay planes on the CRTC to the state
* whenever a plane is enabled to maintain correct z-ordering
* and to enable fast surface updates.
*/
drm_for_each_crtc(crtc, dev) {
bool modified = false;
for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
if (plane->type == DRM_PLANE_TYPE_CURSOR)
continue;
if (new_plane_state->crtc == crtc ||
old_plane_state->crtc == crtc) {
modified = true;
break;
}
}
if (!modified)
continue;
drm_for_each_plane_mask(plane, state->dev, crtc->state->plane_mask) {
if (plane->type == DRM_PLANE_TYPE_CURSOR)
continue;
new_plane_state =
drm_atomic_get_plane_state(state, plane);
if (IS_ERR(new_plane_state)) {
ret = PTR_ERR(new_plane_state);
goto fail;
}
}
}
/* Remove exiting planes if they are modified */
for_each_oldnew_plane_in_state_reverse(state, plane, old_plane_state, new_plane_state, i) {
ret = dm_update_plane_state(dc, state, plane,
old_plane_state,
new_plane_state,
false,
&lock_and_validation_needed);
if (ret)
goto fail;
}
/* Disable all crtcs which require disable */
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
ret = dm_update_crtc_state(&adev->dm, state, crtc,
old_crtc_state,
new_crtc_state,
false,
&lock_and_validation_needed);
if (ret)
goto fail;
}
/* Enable all crtcs which require enable */
for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
ret = dm_update_crtc_state(&adev->dm, state, crtc,
old_crtc_state,
new_crtc_state,
true,
&lock_and_validation_needed);
if (ret)
goto fail;
}
/* Add new/modified planes */
for_each_oldnew_plane_in_state_reverse(state, plane, old_plane_state, new_plane_state, i) {
ret = dm_update_plane_state(dc, state, plane,
old_plane_state,
new_plane_state,
true,
&lock_and_validation_needed);
if (ret)
goto fail;
}
/* Run this here since we want to validate the streams we created */
ret = drm_atomic_helper_check_planes(dev, state);
if (ret)
goto fail;
if (state->legacy_cursor_update) {
/*
* This is a fast cursor update coming from the plane update
* helper, check if it can be done asynchronously for better
* performance.
*/
state->async_update =
!drm_atomic_helper_async_check(dev, state);
/*
* Skip the remaining global validation if this is an async
* update. Cursor updates can be done without affecting
* state or bandwidth calcs and this avoids the performance
* penalty of locking the private state object and
* allocating a new dc_state.
*/
if (state->async_update)
return 0;
}
/* Check scaling and underscan changes*/
/* TODO Removed scaling changes validation due to inability to commit
* new stream into context w\o causing full reset. Need to
* decide how to handle.
*/
for_each_oldnew_connector_in_state(state, connector, old_con_state, new_con_state, i) {
struct dm_connector_state *dm_old_con_state = to_dm_connector_state(old_con_state);
struct dm_connector_state *dm_new_con_state = to_dm_connector_state(new_con_state);
struct amdgpu_crtc *acrtc = to_amdgpu_crtc(dm_new_con_state->base.crtc);
/* Skip any modesets/resets */
if (!acrtc || drm_atomic_crtc_needs_modeset(
drm_atomic_get_new_crtc_state(state, &acrtc->base)))
continue;
/* Skip any thing not scale or underscan changes */
if (!is_scaling_state_different(dm_new_con_state, dm_old_con_state))
continue;
overall_update_type = UPDATE_TYPE_FULL;
lock_and_validation_needed = true;
}
ret = dm_determine_update_type_for_commit(&adev->dm, state, &update_type);
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
if (ret)
goto fail;
if (overall_update_type < update_type)
overall_update_type = update_type;
/*
* lock_and_validation_needed was an old way to determine if we need to set
* the global lock. Leaving it in to check if we broke any corner cases
* lock_and_validation_needed true = UPDATE_TYPE_FULL or UPDATE_TYPE_MED
* lock_and_validation_needed false = UPDATE_TYPE_FAST
*/
if (lock_and_validation_needed && overall_update_type <= UPDATE_TYPE_FAST)
WARN(1, "Global lock should be Set, overall_update_type should be UPDATE_TYPE_MED or UPDATE_TYPE_FULL");
if (overall_update_type > UPDATE_TYPE_FAST) {
drm/amd/display: Use private obj helpers for dm_atomic_state [Why] Two non-blocking commits in succession can result in a sequence where the same dc->current_state is queried for both commits. 1. 1st commit -> check -> commit -> swaps atomic state -> queues work 2. 2nd commit -> check -> commit -> swaps atomic state -> queues work 3. 1st commit work finishes The issue with this sequence is that the same dc->current_state is read in both atomic checks. If the first commit modifies streams or planes those will be missing from the dc->current_state for the second atomic check. This result in many stream and plane errors in atomic commit tail. [How] The driver still needs to track old to new state to determine if the commit in its current implementation. Updating the dc_state in atomic tail is wrong since the dc_state swap should be happening as part of drm_atomic_helper_swap_state *before* the worker queue kicks its work off. The simplest replacement for the subclassing (which doesn't properly manage the old to new atomic state swap) is to use the drm private object helpers. While some of the dc_state members could be merged into dm_crtc_state or dm_plane_state and copied over that way it is easier for now to just treat the whole dc_state structure as a single private object. This allows amdgpu_dm to drop the dc->current_state copy from within atomic check. It's replaced by a copy from the current atomic state which is propagated correctly for the sequence described above. Since access to the dm_state private object is now locked this should also fix issues that could arise if submitting non-blocking commits from different threads. Cc: Harry Wentland <harry.wentland@amd.com> Cc: Leo Li <sunpeng.li@amd.com> Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Reviewed-by: Leo Li <sunpeng.li@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2018-11-22 10:34:36 -07:00
ret = dm_atomic_get_state(state, &dm_state);
if (ret)
goto fail;
ret = do_aquire_global_lock(dev, state);
if (ret)
goto fail;
if (dc_validate_global_state(dc, dm_state->context, false) != DC_OK) {
ret = -EINVAL;
goto fail;
}
drm/amd/display: Don't replace the dc_state for fast updates [Why] DRM private objects have no hw_done/flip_done fencing mechanism on their own and cannot be used to sequence commits accordingly. When issuing commits that don't touch the same set of hardware resources like page-flips on different CRTCs we can run into the issue below because of this: 1. Client requests non-blocking Commit #1, has a new dc_state #1, state is swapped, commit tail is deferred to work queue 2. Client requests non-blocking Commit #2, has a new dc_state #2, state is swapped, commit tail is deferred to work queue 3. Commit #2 work starts, commit tail finishes, atomic state is cleared, dc_state #1 is freed 4. Commit #1 work starts, commit tail encounters null pointer deref on dc_state #1 In order to change the DC state as in the private object we need to ensure that we wait for all outstanding commits to finish and that any other pending commits must wait for the current one to finish as well. We do this for MEDIUM and FULL updates. But not for FAST updates, nor would we want to since it would cause stuttering from the delays. FAST updates that go through dm_determine_update_type_for_commit always create a new dc_state and lock the DRM private object if there are any changed planes. We need the old state to validate, but we don't actually need the new state here. [How] If the commit isn't a full update then the use after free can be resolved by simply discarding the new state entirely and retaining the existing one instead. With this change the sequence above can be reexamined. Commit #2 will still free Commit #1's reference, but before this happens we actually added an additional reference as part of Commit #2. If an update comes in during this that needs to change the dc_state it will need to wait on Commit #1 and Commit #2 to finish. Then it'll swap the state, finish the work in commit tail and drop the last reference on Commit #2's dc_state. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204181 Fixes: 004b3938e637 ("drm/amd/display: Check scaling info when determing update type") Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Acked-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: David Francis <david.francis@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-07-31 08:33:54 -06:00
} else {
/*
drm/amd/display: Don't replace the dc_state for fast updates [Why] DRM private objects have no hw_done/flip_done fencing mechanism on their own and cannot be used to sequence commits accordingly. When issuing commits that don't touch the same set of hardware resources like page-flips on different CRTCs we can run into the issue below because of this: 1. Client requests non-blocking Commit #1, has a new dc_state #1, state is swapped, commit tail is deferred to work queue 2. Client requests non-blocking Commit #2, has a new dc_state #2, state is swapped, commit tail is deferred to work queue 3. Commit #2 work starts, commit tail finishes, atomic state is cleared, dc_state #1 is freed 4. Commit #1 work starts, commit tail encounters null pointer deref on dc_state #1 In order to change the DC state as in the private object we need to ensure that we wait for all outstanding commits to finish and that any other pending commits must wait for the current one to finish as well. We do this for MEDIUM and FULL updates. But not for FAST updates, nor would we want to since it would cause stuttering from the delays. FAST updates that go through dm_determine_update_type_for_commit always create a new dc_state and lock the DRM private object if there are any changed planes. We need the old state to validate, but we don't actually need the new state here. [How] If the commit isn't a full update then the use after free can be resolved by simply discarding the new state entirely and retaining the existing one instead. With this change the sequence above can be reexamined. Commit #2 will still free Commit #1's reference, but before this happens we actually added an additional reference as part of Commit #2. If an update comes in during this that needs to change the dc_state it will need to wait on Commit #1 and Commit #2 to finish. Then it'll swap the state, finish the work in commit tail and drop the last reference on Commit #2's dc_state. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204181 Fixes: 004b3938e637 ("drm/amd/display: Check scaling info when determing update type") Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Acked-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: David Francis <david.francis@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-07-31 08:33:54 -06:00
* The commit is a fast update. Fast updates shouldn't change
* the DC context, affect global validation, and can have their
* commit work done in parallel with other commits not touching
* the same resource. If we have a new DC context as part of
* the DM atomic state from validation we need to free it and
* retain the existing one instead.
drm/amd/display: Clear dm_state for fast updates commit fde9f39ac7f1ffd799a96ffa1e06b2051f0898f1 upstream. This patch fixes a race condition that causes a use-after-free during amdgpu_dm_atomic_commit_tail. This can occur when 2 non-blocking commits are requested and the second one finishes before the first. Essentially, this bug occurs when the following sequence of events happens: 1. Non-blocking commit #1 is requested w/ a new dm_state #1 and is deferred to the workqueue. 2. Non-blocking commit #2 is requested w/ a new dm_state #2 and is deferred to the workqueue. 3. Commit #2 starts before commit #1, dm_state #1 is used in the commit_tail and commit #2 completes, freeing dm_state #1. 4. Commit #1 starts after commit #2 completes, uses the freed dm_state 1 and dereferences a freelist pointer while setting the context. Since this bug has only been spotted with fast commits, this patch fixes the bug by clearing the dm_state instead of using the old dc_state for fast updates. In addition, since dm_state is only used for its dc_state and amdgpu_dm_atomic_commit_tail will retain the dc_state if none is found, removing the dm_state should not have any consequences in fast updates. This use-after-free bug has existed for a while now, but only caused a noticeable issue starting from 5.7-rc1 due to 3202fa62f ("slub: relocate freelist pointer to middle of object") moving the freelist pointer from dm_state->base (which was unused) to dm_state->context (which is dereferenced). Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=207383 Fixes: bd200d190f45 ("drm/amd/display: Don't replace the dc_state for fast updates") Reported-by: Duncan <1i5t5.duncan@cox.net> Signed-off-by: Mazin Rezk <mnrzk@protonmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2020-07-26 23:40:46 -06:00
*
* Furthermore, since the DM atomic state only contains the DC
* context and can safely be annulled, we can free the state
* and clear the associated private object now to free
* some memory and avoid a possible use-after-free later.
*/
drm/amd/display: Don't replace the dc_state for fast updates [Why] DRM private objects have no hw_done/flip_done fencing mechanism on their own and cannot be used to sequence commits accordingly. When issuing commits that don't touch the same set of hardware resources like page-flips on different CRTCs we can run into the issue below because of this: 1. Client requests non-blocking Commit #1, has a new dc_state #1, state is swapped, commit tail is deferred to work queue 2. Client requests non-blocking Commit #2, has a new dc_state #2, state is swapped, commit tail is deferred to work queue 3. Commit #2 work starts, commit tail finishes, atomic state is cleared, dc_state #1 is freed 4. Commit #1 work starts, commit tail encounters null pointer deref on dc_state #1 In order to change the DC state as in the private object we need to ensure that we wait for all outstanding commits to finish and that any other pending commits must wait for the current one to finish as well. We do this for MEDIUM and FULL updates. But not for FAST updates, nor would we want to since it would cause stuttering from the delays. FAST updates that go through dm_determine_update_type_for_commit always create a new dc_state and lock the DRM private object if there are any changed planes. We need the old state to validate, but we don't actually need the new state here. [How] If the commit isn't a full update then the use after free can be resolved by simply discarding the new state entirely and retaining the existing one instead. With this change the sequence above can be reexamined. Commit #2 will still free Commit #1's reference, but before this happens we actually added an additional reference as part of Commit #2. If an update comes in during this that needs to change the dc_state it will need to wait on Commit #1 and Commit #2 to finish. Then it'll swap the state, finish the work in commit tail and drop the last reference on Commit #2's dc_state. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204181 Fixes: 004b3938e637 ("drm/amd/display: Check scaling info when determing update type") Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Acked-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: David Francis <david.francis@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-07-31 08:33:54 -06:00
drm/amd/display: Clear dm_state for fast updates commit fde9f39ac7f1ffd799a96ffa1e06b2051f0898f1 upstream. This patch fixes a race condition that causes a use-after-free during amdgpu_dm_atomic_commit_tail. This can occur when 2 non-blocking commits are requested and the second one finishes before the first. Essentially, this bug occurs when the following sequence of events happens: 1. Non-blocking commit #1 is requested w/ a new dm_state #1 and is deferred to the workqueue. 2. Non-blocking commit #2 is requested w/ a new dm_state #2 and is deferred to the workqueue. 3. Commit #2 starts before commit #1, dm_state #1 is used in the commit_tail and commit #2 completes, freeing dm_state #1. 4. Commit #1 starts after commit #2 completes, uses the freed dm_state 1 and dereferences a freelist pointer while setting the context. Since this bug has only been spotted with fast commits, this patch fixes the bug by clearing the dm_state instead of using the old dc_state for fast updates. In addition, since dm_state is only used for its dc_state and amdgpu_dm_atomic_commit_tail will retain the dc_state if none is found, removing the dm_state should not have any consequences in fast updates. This use-after-free bug has existed for a while now, but only caused a noticeable issue starting from 5.7-rc1 due to 3202fa62f ("slub: relocate freelist pointer to middle of object") moving the freelist pointer from dm_state->base (which was unused) to dm_state->context (which is dereferenced). Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=207383 Fixes: bd200d190f45 ("drm/amd/display: Don't replace the dc_state for fast updates") Reported-by: Duncan <1i5t5.duncan@cox.net> Signed-off-by: Mazin Rezk <mnrzk@protonmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2020-07-26 23:40:46 -06:00
for (i = 0; i < state->num_private_objs; i++) {
struct drm_private_obj *obj = state->private_objs[i].ptr;
drm/amd/display: Don't replace the dc_state for fast updates [Why] DRM private objects have no hw_done/flip_done fencing mechanism on their own and cannot be used to sequence commits accordingly. When issuing commits that don't touch the same set of hardware resources like page-flips on different CRTCs we can run into the issue below because of this: 1. Client requests non-blocking Commit #1, has a new dc_state #1, state is swapped, commit tail is deferred to work queue 2. Client requests non-blocking Commit #2, has a new dc_state #2, state is swapped, commit tail is deferred to work queue 3. Commit #2 work starts, commit tail finishes, atomic state is cleared, dc_state #1 is freed 4. Commit #1 work starts, commit tail encounters null pointer deref on dc_state #1 In order to change the DC state as in the private object we need to ensure that we wait for all outstanding commits to finish and that any other pending commits must wait for the current one to finish as well. We do this for MEDIUM and FULL updates. But not for FAST updates, nor would we want to since it would cause stuttering from the delays. FAST updates that go through dm_determine_update_type_for_commit always create a new dc_state and lock the DRM private object if there are any changed planes. We need the old state to validate, but we don't actually need the new state here. [How] If the commit isn't a full update then the use after free can be resolved by simply discarding the new state entirely and retaining the existing one instead. With this change the sequence above can be reexamined. Commit #2 will still free Commit #1's reference, but before this happens we actually added an additional reference as part of Commit #2. If an update comes in during this that needs to change the dc_state it will need to wait on Commit #1 and Commit #2 to finish. Then it'll swap the state, finish the work in commit tail and drop the last reference on Commit #2's dc_state. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204181 Fixes: 004b3938e637 ("drm/amd/display: Check scaling info when determing update type") Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Acked-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: David Francis <david.francis@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-07-31 08:33:54 -06:00
drm/amd/display: Clear dm_state for fast updates commit fde9f39ac7f1ffd799a96ffa1e06b2051f0898f1 upstream. This patch fixes a race condition that causes a use-after-free during amdgpu_dm_atomic_commit_tail. This can occur when 2 non-blocking commits are requested and the second one finishes before the first. Essentially, this bug occurs when the following sequence of events happens: 1. Non-blocking commit #1 is requested w/ a new dm_state #1 and is deferred to the workqueue. 2. Non-blocking commit #2 is requested w/ a new dm_state #2 and is deferred to the workqueue. 3. Commit #2 starts before commit #1, dm_state #1 is used in the commit_tail and commit #2 completes, freeing dm_state #1. 4. Commit #1 starts after commit #2 completes, uses the freed dm_state 1 and dereferences a freelist pointer while setting the context. Since this bug has only been spotted with fast commits, this patch fixes the bug by clearing the dm_state instead of using the old dc_state for fast updates. In addition, since dm_state is only used for its dc_state and amdgpu_dm_atomic_commit_tail will retain the dc_state if none is found, removing the dm_state should not have any consequences in fast updates. This use-after-free bug has existed for a while now, but only caused a noticeable issue starting from 5.7-rc1 due to 3202fa62f ("slub: relocate freelist pointer to middle of object") moving the freelist pointer from dm_state->base (which was unused) to dm_state->context (which is dereferenced). Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=207383 Fixes: bd200d190f45 ("drm/amd/display: Don't replace the dc_state for fast updates") Reported-by: Duncan <1i5t5.duncan@cox.net> Signed-off-by: Mazin Rezk <mnrzk@protonmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2020-07-26 23:40:46 -06:00
if (obj->funcs == adev->dm.atomic_obj.funcs) {
int j = state->num_private_objs-1;
drm/amd/display: Don't replace the dc_state for fast updates [Why] DRM private objects have no hw_done/flip_done fencing mechanism on their own and cannot be used to sequence commits accordingly. When issuing commits that don't touch the same set of hardware resources like page-flips on different CRTCs we can run into the issue below because of this: 1. Client requests non-blocking Commit #1, has a new dc_state #1, state is swapped, commit tail is deferred to work queue 2. Client requests non-blocking Commit #2, has a new dc_state #2, state is swapped, commit tail is deferred to work queue 3. Commit #2 work starts, commit tail finishes, atomic state is cleared, dc_state #1 is freed 4. Commit #1 work starts, commit tail encounters null pointer deref on dc_state #1 In order to change the DC state as in the private object we need to ensure that we wait for all outstanding commits to finish and that any other pending commits must wait for the current one to finish as well. We do this for MEDIUM and FULL updates. But not for FAST updates, nor would we want to since it would cause stuttering from the delays. FAST updates that go through dm_determine_update_type_for_commit always create a new dc_state and lock the DRM private object if there are any changed planes. We need the old state to validate, but we don't actually need the new state here. [How] If the commit isn't a full update then the use after free can be resolved by simply discarding the new state entirely and retaining the existing one instead. With this change the sequence above can be reexamined. Commit #2 will still free Commit #1's reference, but before this happens we actually added an additional reference as part of Commit #2. If an update comes in during this that needs to change the dc_state it will need to wait on Commit #1 and Commit #2 to finish. Then it'll swap the state, finish the work in commit tail and drop the last reference on Commit #2's dc_state. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204181 Fixes: 004b3938e637 ("drm/amd/display: Check scaling info when determing update type") Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Acked-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: David Francis <david.francis@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-07-31 08:33:54 -06:00
drm/amd/display: Clear dm_state for fast updates commit fde9f39ac7f1ffd799a96ffa1e06b2051f0898f1 upstream. This patch fixes a race condition that causes a use-after-free during amdgpu_dm_atomic_commit_tail. This can occur when 2 non-blocking commits are requested and the second one finishes before the first. Essentially, this bug occurs when the following sequence of events happens: 1. Non-blocking commit #1 is requested w/ a new dm_state #1 and is deferred to the workqueue. 2. Non-blocking commit #2 is requested w/ a new dm_state #2 and is deferred to the workqueue. 3. Commit #2 starts before commit #1, dm_state #1 is used in the commit_tail and commit #2 completes, freeing dm_state #1. 4. Commit #1 starts after commit #2 completes, uses the freed dm_state 1 and dereferences a freelist pointer while setting the context. Since this bug has only been spotted with fast commits, this patch fixes the bug by clearing the dm_state instead of using the old dc_state for fast updates. In addition, since dm_state is only used for its dc_state and amdgpu_dm_atomic_commit_tail will retain the dc_state if none is found, removing the dm_state should not have any consequences in fast updates. This use-after-free bug has existed for a while now, but only caused a noticeable issue starting from 5.7-rc1 due to 3202fa62f ("slub: relocate freelist pointer to middle of object") moving the freelist pointer from dm_state->base (which was unused) to dm_state->context (which is dereferenced). Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=207383 Fixes: bd200d190f45 ("drm/amd/display: Don't replace the dc_state for fast updates") Reported-by: Duncan <1i5t5.duncan@cox.net> Signed-off-by: Mazin Rezk <mnrzk@protonmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2020-07-26 23:40:46 -06:00
dm_atomic_destroy_state(obj,
state->private_objs[i].state);
/* If i is not at the end of the array then the
* last element needs to be moved to where i was
* before the array can safely be truncated.
*/
if (i != j)
state->private_objs[i] =
state->private_objs[j];
drm/amd/display: Don't replace the dc_state for fast updates [Why] DRM private objects have no hw_done/flip_done fencing mechanism on their own and cannot be used to sequence commits accordingly. When issuing commits that don't touch the same set of hardware resources like page-flips on different CRTCs we can run into the issue below because of this: 1. Client requests non-blocking Commit #1, has a new dc_state #1, state is swapped, commit tail is deferred to work queue 2. Client requests non-blocking Commit #2, has a new dc_state #2, state is swapped, commit tail is deferred to work queue 3. Commit #2 work starts, commit tail finishes, atomic state is cleared, dc_state #1 is freed 4. Commit #1 work starts, commit tail encounters null pointer deref on dc_state #1 In order to change the DC state as in the private object we need to ensure that we wait for all outstanding commits to finish and that any other pending commits must wait for the current one to finish as well. We do this for MEDIUM and FULL updates. But not for FAST updates, nor would we want to since it would cause stuttering from the delays. FAST updates that go through dm_determine_update_type_for_commit always create a new dc_state and lock the DRM private object if there are any changed planes. We need the old state to validate, but we don't actually need the new state here. [How] If the commit isn't a full update then the use after free can be resolved by simply discarding the new state entirely and retaining the existing one instead. With this change the sequence above can be reexamined. Commit #2 will still free Commit #1's reference, but before this happens we actually added an additional reference as part of Commit #2. If an update comes in during this that needs to change the dc_state it will need to wait on Commit #1 and Commit #2 to finish. Then it'll swap the state, finish the work in commit tail and drop the last reference on Commit #2's dc_state. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204181 Fixes: 004b3938e637 ("drm/amd/display: Check scaling info when determing update type") Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Acked-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: David Francis <david.francis@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-07-31 08:33:54 -06:00
drm/amd/display: Clear dm_state for fast updates commit fde9f39ac7f1ffd799a96ffa1e06b2051f0898f1 upstream. This patch fixes a race condition that causes a use-after-free during amdgpu_dm_atomic_commit_tail. This can occur when 2 non-blocking commits are requested and the second one finishes before the first. Essentially, this bug occurs when the following sequence of events happens: 1. Non-blocking commit #1 is requested w/ a new dm_state #1 and is deferred to the workqueue. 2. Non-blocking commit #2 is requested w/ a new dm_state #2 and is deferred to the workqueue. 3. Commit #2 starts before commit #1, dm_state #1 is used in the commit_tail and commit #2 completes, freeing dm_state #1. 4. Commit #1 starts after commit #2 completes, uses the freed dm_state 1 and dereferences a freelist pointer while setting the context. Since this bug has only been spotted with fast commits, this patch fixes the bug by clearing the dm_state instead of using the old dc_state for fast updates. In addition, since dm_state is only used for its dc_state and amdgpu_dm_atomic_commit_tail will retain the dc_state if none is found, removing the dm_state should not have any consequences in fast updates. This use-after-free bug has existed for a while now, but only caused a noticeable issue starting from 5.7-rc1 due to 3202fa62f ("slub: relocate freelist pointer to middle of object") moving the freelist pointer from dm_state->base (which was unused) to dm_state->context (which is dereferenced). Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=207383 Fixes: bd200d190f45 ("drm/amd/display: Don't replace the dc_state for fast updates") Reported-by: Duncan <1i5t5.duncan@cox.net> Signed-off-by: Mazin Rezk <mnrzk@protonmail.com> Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com> Cc: stable@vger.kernel.org Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2020-07-26 23:40:46 -06:00
state->private_objs[j].ptr = NULL;
state->private_objs[j].state = NULL;
state->private_objs[j].old_state = NULL;
state->private_objs[j].new_state = NULL;
state->num_private_objs = j;
break;
}
drm/amd/display: Don't replace the dc_state for fast updates [Why] DRM private objects have no hw_done/flip_done fencing mechanism on their own and cannot be used to sequence commits accordingly. When issuing commits that don't touch the same set of hardware resources like page-flips on different CRTCs we can run into the issue below because of this: 1. Client requests non-blocking Commit #1, has a new dc_state #1, state is swapped, commit tail is deferred to work queue 2. Client requests non-blocking Commit #2, has a new dc_state #2, state is swapped, commit tail is deferred to work queue 3. Commit #2 work starts, commit tail finishes, atomic state is cleared, dc_state #1 is freed 4. Commit #1 work starts, commit tail encounters null pointer deref on dc_state #1 In order to change the DC state as in the private object we need to ensure that we wait for all outstanding commits to finish and that any other pending commits must wait for the current one to finish as well. We do this for MEDIUM and FULL updates. But not for FAST updates, nor would we want to since it would cause stuttering from the delays. FAST updates that go through dm_determine_update_type_for_commit always create a new dc_state and lock the DRM private object if there are any changed planes. We need the old state to validate, but we don't actually need the new state here. [How] If the commit isn't a full update then the use after free can be resolved by simply discarding the new state entirely and retaining the existing one instead. With this change the sequence above can be reexamined. Commit #2 will still free Commit #1's reference, but before this happens we actually added an additional reference as part of Commit #2. If an update comes in during this that needs to change the dc_state it will need to wait on Commit #1 and Commit #2 to finish. Then it'll swap the state, finish the work in commit tail and drop the last reference on Commit #2's dc_state. Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=204181 Fixes: 004b3938e637 ("drm/amd/display: Check scaling info when determing update type") Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com> Acked-by: Alex Deucher <alexander.deucher@amd.com> Reviewed-by: David Francis <david.francis@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2019-07-31 08:33:54 -06:00
}
}
/* Store the overall update type for use later in atomic check. */
for_each_new_crtc_in_state (state, crtc, new_crtc_state, i) {
struct dm_crtc_state *dm_new_crtc_state =
to_dm_crtc_state(new_crtc_state);
dm_new_crtc_state->update_type = (int)overall_update_type;
}
/* Must be success */
WARN_ON(ret);
return ret;
fail:
if (ret == -EDEADLK)
DRM_DEBUG_DRIVER("Atomic check stopped to avoid deadlock.\n");
else if (ret == -EINTR || ret == -EAGAIN || ret == -ERESTARTSYS)
DRM_DEBUG_DRIVER("Atomic check stopped due to signal.\n");
else
DRM_DEBUG_DRIVER("Atomic check failed with err: %d \n", ret);
return ret;
}
static bool is_dp_capable_without_timing_msa(struct dc *dc,
struct amdgpu_dm_connector *amdgpu_dm_connector)
{
uint8_t dpcd_data;
bool capable = false;
if (amdgpu_dm_connector->dc_link &&
dm_helpers_dp_read_dpcd(
NULL,
amdgpu_dm_connector->dc_link,
DP_DOWN_STREAM_PORT_COUNT,
&dpcd_data,
sizeof(dpcd_data))) {
capable = (dpcd_data & DP_MSA_TIMING_PAR_IGNORED) ? true:false;
}
return capable;
}
void amdgpu_dm_update_freesync_caps(struct drm_connector *connector,
struct edid *edid)
{
int i;
bool edid_check_required;
struct detailed_timing *timing;
struct detailed_non_pixel *data;
struct detailed_data_monitor_range *range;
struct amdgpu_dm_connector *amdgpu_dm_connector =
to_amdgpu_dm_connector(connector);
struct dm_connector_state *dm_con_state = NULL;
struct drm_device *dev = connector->dev;
struct amdgpu_device *adev = dev->dev_private;
bool freesync_capable = false;
if (!connector->state) {
DRM_ERROR("%s - Connector has no state", __func__);
goto update;
}
if (!edid) {
dm_con_state = to_dm_connector_state(connector->state);
amdgpu_dm_connector->min_vfreq = 0;
amdgpu_dm_connector->max_vfreq = 0;
amdgpu_dm_connector->pixel_clock_mhz = 0;
goto update;
}
dm_con_state = to_dm_connector_state(connector->state);
edid_check_required = false;
if (!amdgpu_dm_connector->dc_sink) {
DRM_ERROR("dc_sink NULL, could not add free_sync module.\n");
goto update;
}
if (!adev->dm.freesync_module)
goto update;
/*
* if edid non zero restrict freesync only for dp and edp
*/
if (edid) {
if (amdgpu_dm_connector->dc_sink->sink_signal == SIGNAL_TYPE_DISPLAY_PORT
|| amdgpu_dm_connector->dc_sink->sink_signal == SIGNAL_TYPE_EDP) {
edid_check_required = is_dp_capable_without_timing_msa(
adev->dm.dc,
amdgpu_dm_connector);
}
}
if (edid_check_required == true && (edid->version > 1 ||
(edid->version == 1 && edid->revision > 1))) {
for (i = 0; i < 4; i++) {
timing = &edid->detailed_timings[i];
data = &timing->data.other_data;
range = &data->data.range;
/*
* Check if monitor has continuous frequency mode
*/
if (data->type != EDID_DETAIL_MONITOR_RANGE)
continue;
/*
* Check for flag range limits only. If flag == 1 then
* no additional timing information provided.
* Default GTF, GTF Secondary curve and CVT are not
* supported
*/
if (range->flags != 1)
continue;
amdgpu_dm_connector->min_vfreq = range->min_vfreq;
amdgpu_dm_connector->max_vfreq = range->max_vfreq;
amdgpu_dm_connector->pixel_clock_mhz =
range->pixel_clock_mhz * 10;
break;
}
if (amdgpu_dm_connector->max_vfreq -
amdgpu_dm_connector->min_vfreq > 10) {
freesync_capable = true;
}
}
update:
if (dm_con_state)
dm_con_state->freesync_capable = freesync_capable;
if (connector->vrr_capable_property)
drm_connector_set_vrr_capable_property(connector,
freesync_capable);
}