alistair23-linux/drivers/gpu/drm/i915/i915_gpu_error.c

1521 lines
41 KiB
C
Raw Normal View History

/*
* Copyright (c) 2008 Intel Corporation
*
* 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 (including the next
* paragraph) 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 AUTHORS OR COPYRIGHT HOLDERS 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:
* Eric Anholt <eric@anholt.net>
* Keith Packard <keithp@keithp.com>
* Mika Kuoppala <mika.kuoppala@intel.com>
*
*/
#include <generated/utsrelease.h>
#include "i915_drv.h"
static const char *engine_str(int engine)
{
switch (engine) {
case RCS: return "render";
case VCS: return "bsd";
case BCS: return "blt";
case VECS: return "vebox";
case VCS2: return "bsd2";
default: return "";
}
}
static const char *pin_flag(int pinned)
{
if (pinned > 0)
return " P";
else if (pinned < 0)
return " p";
else
return "";
}
static const char *tiling_flag(int tiling)
{
switch (tiling) {
default:
case I915_TILING_NONE: return "";
case I915_TILING_X: return " X";
case I915_TILING_Y: return " Y";
}
}
static const char *dirty_flag(int dirty)
{
return dirty ? " dirty" : "";
}
static const char *purgeable_flag(int purgeable)
{
return purgeable ? " purgeable" : "";
}
static bool __i915_error_ok(struct drm_i915_error_state_buf *e)
{
if (!e->err && WARN(e->bytes > (e->size - 1), "overflow")) {
e->err = -ENOSPC;
return false;
}
if (e->bytes == e->size - 1 || e->err)
return false;
return true;
}
static bool __i915_error_seek(struct drm_i915_error_state_buf *e,
unsigned len)
{
if (e->pos + len <= e->start) {
e->pos += len;
return false;
}
/* First vsnprintf needs to fit in its entirety for memmove */
if (len >= e->size) {
e->err = -EIO;
return false;
}
return true;
}
static void __i915_error_advance(struct drm_i915_error_state_buf *e,
unsigned len)
{
/* If this is first printf in this window, adjust it so that
* start position matches start of the buffer
*/
if (e->pos < e->start) {
const size_t off = e->start - e->pos;
/* Should not happen but be paranoid */
if (off > len || e->bytes) {
e->err = -EIO;
return;
}
memmove(e->buf, e->buf + off, len - off);
e->bytes = len - off;
e->pos = e->start;
return;
}
e->bytes += len;
e->pos += len;
}
static void i915_error_vprintf(struct drm_i915_error_state_buf *e,
const char *f, va_list args)
{
unsigned len;
if (!__i915_error_ok(e))
return;
/* Seek the first printf which is hits start position */
if (e->pos < e->start) {
va_list tmp;
va_copy(tmp, args);
len = vsnprintf(NULL, 0, f, tmp);
va_end(tmp);
if (!__i915_error_seek(e, len))
return;
}
len = vsnprintf(e->buf + e->bytes, e->size - e->bytes, f, args);
if (len >= e->size - e->bytes)
len = e->size - e->bytes - 1;
__i915_error_advance(e, len);
}
static void i915_error_puts(struct drm_i915_error_state_buf *e,
const char *str)
{
unsigned len;
if (!__i915_error_ok(e))
return;
len = strlen(str);
/* Seek the first printf which is hits start position */
if (e->pos < e->start) {
if (!__i915_error_seek(e, len))
return;
}
if (len >= e->size - e->bytes)
len = e->size - e->bytes - 1;
memcpy(e->buf + e->bytes, str, len);
__i915_error_advance(e, len);
}
#define err_printf(e, ...) i915_error_printf(e, __VA_ARGS__)
#define err_puts(e, s) i915_error_puts(e, s)
static void print_error_buffers(struct drm_i915_error_state_buf *m,
const char *name,
struct drm_i915_error_buffer *err,
int count)
{
drm/i915: Implement inter-engine read-read optimisations Currently, we only track the last request globally across all engines. This prevents us from issuing concurrent read requests on e.g. the RCS and BCS engines (or more likely the render and media engines). Without semaphores, we incur costly stalls as we synchronise between rings - greatly impacting the current performance of Broadwell versus Haswell in certain workloads (like video decode). With the introduction of reference counted requests, it is much easier to track the last request per ring, as well as the last global write request so that we can optimise inter-engine read read requests (as well as better optimise certain CPU waits). v2: Fix inverted readonly condition for nonblocking waits. v3: Handle non-continguous engine array after waits v4: Rebase, tidy, rewrite ring list debugging v5: Use obj->active as a bitfield, it looks cool v6: Micro-optimise, mostly involving moving code around v7: Fix retire-requests-upto for execlists (and multiple rq->ringbuf) v8: Rebase v9: Refactor i915_gem_object_sync() to allow the compiler to better optimise it. Benchmark: igt/gem_read_read_speed hsw:gt3e (with semaphores): Before: Time to read-read 1024k: 275.794µs After: Time to read-read 1024k: 123.260µs hsw:gt3e (w/o semaphores): Before: Time to read-read 1024k: 230.433µs After: Time to read-read 1024k: 124.593µs bdw-u (w/o semaphores): Before After Time to read-read 1x1: 26.274µs 10.350µs Time to read-read 128x128: 40.097µs 21.366µs Time to read-read 256x256: 77.087µs 42.608µs Time to read-read 512x512: 281.999µs 181.155µs Time to read-read 1024x1024: 1196.141µs 1118.223µs Time to read-read 2048x2048: 5639.072µs 5225.837µs Time to read-read 4096x4096: 22401.662µs 21137.067µs Time to read-read 8192x8192: 89617.735µs 85637.681µs Testcase: igt/gem_concurrent_blit (read-read and friends) Cc: Lionel Landwerlin <lionel.g.landwerlin@linux.intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> [v8] [danvet: s/\<rq\>/req/g] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-04-27 06:41:17 -06:00
int i;
err_printf(m, " %s [%d]:\n", name, count);
while (count--) {
err_printf(m, " %08x_%08x %8u %02x %02x [ ",
upper_32_bits(err->gtt_offset),
lower_32_bits(err->gtt_offset),
err->size,
err->read_domains,
drm/i915: Implement inter-engine read-read optimisations Currently, we only track the last request globally across all engines. This prevents us from issuing concurrent read requests on e.g. the RCS and BCS engines (or more likely the render and media engines). Without semaphores, we incur costly stalls as we synchronise between rings - greatly impacting the current performance of Broadwell versus Haswell in certain workloads (like video decode). With the introduction of reference counted requests, it is much easier to track the last request per ring, as well as the last global write request so that we can optimise inter-engine read read requests (as well as better optimise certain CPU waits). v2: Fix inverted readonly condition for nonblocking waits. v3: Handle non-continguous engine array after waits v4: Rebase, tidy, rewrite ring list debugging v5: Use obj->active as a bitfield, it looks cool v6: Micro-optimise, mostly involving moving code around v7: Fix retire-requests-upto for execlists (and multiple rq->ringbuf) v8: Rebase v9: Refactor i915_gem_object_sync() to allow the compiler to better optimise it. Benchmark: igt/gem_read_read_speed hsw:gt3e (with semaphores): Before: Time to read-read 1024k: 275.794µs After: Time to read-read 1024k: 123.260µs hsw:gt3e (w/o semaphores): Before: Time to read-read 1024k: 230.433µs After: Time to read-read 1024k: 124.593µs bdw-u (w/o semaphores): Before After Time to read-read 1x1: 26.274µs 10.350µs Time to read-read 128x128: 40.097µs 21.366µs Time to read-read 256x256: 77.087µs 42.608µs Time to read-read 512x512: 281.999µs 181.155µs Time to read-read 1024x1024: 1196.141µs 1118.223µs Time to read-read 2048x2048: 5639.072µs 5225.837µs Time to read-read 4096x4096: 22401.662µs 21137.067µs Time to read-read 8192x8192: 89617.735µs 85637.681µs Testcase: igt/gem_concurrent_blit (read-read and friends) Cc: Lionel Landwerlin <lionel.g.landwerlin@linux.intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> [v8] [danvet: s/\<rq\>/req/g] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-04-27 06:41:17 -06:00
err->write_domain);
for (i = 0; i < I915_NUM_ENGINES; i++)
drm/i915: Implement inter-engine read-read optimisations Currently, we only track the last request globally across all engines. This prevents us from issuing concurrent read requests on e.g. the RCS and BCS engines (or more likely the render and media engines). Without semaphores, we incur costly stalls as we synchronise between rings - greatly impacting the current performance of Broadwell versus Haswell in certain workloads (like video decode). With the introduction of reference counted requests, it is much easier to track the last request per ring, as well as the last global write request so that we can optimise inter-engine read read requests (as well as better optimise certain CPU waits). v2: Fix inverted readonly condition for nonblocking waits. v3: Handle non-continguous engine array after waits v4: Rebase, tidy, rewrite ring list debugging v5: Use obj->active as a bitfield, it looks cool v6: Micro-optimise, mostly involving moving code around v7: Fix retire-requests-upto for execlists (and multiple rq->ringbuf) v8: Rebase v9: Refactor i915_gem_object_sync() to allow the compiler to better optimise it. Benchmark: igt/gem_read_read_speed hsw:gt3e (with semaphores): Before: Time to read-read 1024k: 275.794µs After: Time to read-read 1024k: 123.260µs hsw:gt3e (w/o semaphores): Before: Time to read-read 1024k: 230.433µs After: Time to read-read 1024k: 124.593µs bdw-u (w/o semaphores): Before After Time to read-read 1x1: 26.274µs 10.350µs Time to read-read 128x128: 40.097µs 21.366µs Time to read-read 256x256: 77.087µs 42.608µs Time to read-read 512x512: 281.999µs 181.155µs Time to read-read 1024x1024: 1196.141µs 1118.223µs Time to read-read 2048x2048: 5639.072µs 5225.837µs Time to read-read 4096x4096: 22401.662µs 21137.067µs Time to read-read 8192x8192: 89617.735µs 85637.681µs Testcase: igt/gem_concurrent_blit (read-read and friends) Cc: Lionel Landwerlin <lionel.g.landwerlin@linux.intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> [v8] [danvet: s/\<rq\>/req/g] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-04-27 06:41:17 -06:00
err_printf(m, "%02x ", err->rseqno[i]);
err_printf(m, "] %02x", err->wseqno);
err_puts(m, pin_flag(err->pinned));
err_puts(m, tiling_flag(err->tiling));
err_puts(m, dirty_flag(err->dirty));
err_puts(m, purgeable_flag(err->purgeable));
drm/i915: Introduce mapping of user pages into video memory (userptr) ioctl By exporting the ability to map user address and inserting PTEs representing their backing pages into the GTT, we can exploit UMA in order to utilize normal application data as a texture source or even as a render target (depending upon the capabilities of the chipset). This has a number of uses, with zero-copy downloads to the GPU and efficient readback making the intermixed streaming of CPU and GPU operations fairly efficient. This ability has many widespread implications from faster rendering of client-side software rasterisers (chromium), mitigation of stalls due to read back (firefox) and to faster pipelining of texture data (such as pixel buffer objects in GL or data blobs in CL). v2: Compile with CONFIG_MMU_NOTIFIER v3: We can sleep while performing invalidate-range, which we can utilise to drop our page references prior to the kernel manipulating the vma (for either discard or cloning) and so protect normal users. v4: Only run the invalidate notifier if the range intercepts the bo. v5: Prevent userspace from attempting to GTT mmap non-page aligned buffers v6: Recheck after reacquire mutex for lost mmu. v7: Fix implicit padding of ioctl struct by rounding to next 64bit boundary. v8: Fix rebasing error after forwarding porting the back port. v9: Limit the userptr to page aligned entries. We now expect userspace to handle all the offset-in-page adjustments itself. v10: Prevent vma from being copied across fork to avoid issues with cow. v11: Drop vma behaviour changes -- locking is nigh on impossible. Use a worker to load user pages to avoid lock inversions. v12: Use get_task_mm()/mmput() for correct refcounting of mm. v13: Use a worker to release the mmu_notifier to avoid lock inversion v14: Decouple mmu_notifier from struct_mutex using a custom mmu_notifer with its own locking and tree of objects for each mm/mmu_notifier. v15: Prevent overlapping userptr objects, and invalidate all objects within the mmu_notifier range v16: Fix a typo for iterating over multiple objects in the range and rearrange error path to destroy the mmu_notifier locklessly. Also close a race between invalidate_range and the get_pages_worker. v17: Close a race between get_pages_worker/invalidate_range and fresh allocations of the same userptr range - and notice that struct_mutex was presumed to be held when during creation it wasn't. v18: Sigh. Fix the refactor of st_set_pages() to allocate enough memory for the struct sg_table and to clear it before reporting an error. v19: Always error out on read-only userptr requests as we don't have the hardware infrastructure to support them at the moment. v20: Refuse to implement read-only support until we have the required infrastructure - but reserve the bit in flags for future use. v21: use_mm() is not required for get_user_pages(). It is only meant to be used to fix up the kernel thread's current->mm for use with copy_user(). v22: Use sg_alloc_table_from_pages for that chunky feeling v23: Export a function for sanity checking dma-buf rather than encode userptr details elsewhere, and clean up comments based on suggestions by Bradley. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Akash Goel <akash.goel@intel.com> Cc: "Volkin, Bradley D" <bradley.d.volkin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Reviewed-by: Brad Volkin <bradley.d.volkin@intel.com> [danvet: Frob ioctl allocation to pick the next one - will cause a bit of fuss with create2 apparently, but such are the rules.] [danvet2: oops, forgot to git add after manual patch application] [danvet3: Appease sparse.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-05-16 07:22:37 -06:00
err_puts(m, err->userptr ? " userptr" : "");
err_puts(m, err->engine != -1 ? " " : "");
err_puts(m, engine_str(err->engine));
err_puts(m, i915_cache_level_str(m->i915, err->cache_level));
if (err->name)
err_printf(m, " (name: %d)", err->name);
if (err->fence_reg != I915_FENCE_REG_NONE)
err_printf(m, " (fence: %d)", err->fence_reg);
err_puts(m, "\n");
err++;
}
}
static const char *hangcheck_action_to_str(enum intel_engine_hangcheck_action a)
{
switch (a) {
case HANGCHECK_IDLE:
return "idle";
case HANGCHECK_WAIT:
return "wait";
case HANGCHECK_ACTIVE:
return "active";
case HANGCHECK_KICK:
return "kick";
case HANGCHECK_HUNG:
return "hung";
}
return "unknown";
}
static void error_print_engine(struct drm_i915_error_state_buf *m,
struct drm_i915_error_engine *ee)
{
err_printf(m, "%s command stream:\n", engine_str(ee->engine_id));
err_printf(m, " START: 0x%08x\n", ee->start);
err_printf(m, " HEAD: 0x%08x\n", ee->head);
err_printf(m, " TAIL: 0x%08x\n", ee->tail);
err_printf(m, " CTL: 0x%08x\n", ee->ctl);
err_printf(m, " HWS: 0x%08x\n", ee->hws);
err_printf(m, " ACTHD: 0x%08x %08x\n",
(u32)(ee->acthd>>32), (u32)ee->acthd);
err_printf(m, " IPEIR: 0x%08x\n", ee->ipeir);
err_printf(m, " IPEHR: 0x%08x\n", ee->ipehr);
err_printf(m, " INSTDONE: 0x%08x\n", ee->instdone);
if (INTEL_GEN(m->i915) >= 4) {
err_printf(m, " BBADDR: 0x%08x %08x\n",
(u32)(ee->bbaddr>>32), (u32)ee->bbaddr);
err_printf(m, " BB_STATE: 0x%08x\n", ee->bbstate);
err_printf(m, " INSTPS: 0x%08x\n", ee->instps);
}
err_printf(m, " INSTPM: 0x%08x\n", ee->instpm);
err_printf(m, " FADDR: 0x%08x %08x\n", upper_32_bits(ee->faddr),
lower_32_bits(ee->faddr));
if (INTEL_GEN(m->i915) >= 6) {
err_printf(m, " RC PSMI: 0x%08x\n", ee->rc_psmi);
err_printf(m, " FAULT_REG: 0x%08x\n", ee->fault_reg);
err_printf(m, " SYNC_0: 0x%08x [last synced 0x%08x]\n",
ee->semaphore_mboxes[0],
ee->semaphore_seqno[0]);
err_printf(m, " SYNC_1: 0x%08x [last synced 0x%08x]\n",
ee->semaphore_mboxes[1],
ee->semaphore_seqno[1]);
if (HAS_VEBOX(m->i915)) {
err_printf(m, " SYNC_2: 0x%08x [last synced 0x%08x]\n",
ee->semaphore_mboxes[2],
ee->semaphore_seqno[2]);
}
}
if (USES_PPGTT(m->i915)) {
err_printf(m, " GFX_MODE: 0x%08x\n", ee->vm_info.gfx_mode);
if (INTEL_GEN(m->i915) >= 8) {
int i;
for (i = 0; i < 4; i++)
err_printf(m, " PDP%d: 0x%016llx\n",
i, ee->vm_info.pdp[i]);
} else {
err_printf(m, " PP_DIR_BASE: 0x%08x\n",
ee->vm_info.pp_dir_base);
}
}
err_printf(m, " seqno: 0x%08x\n", ee->seqno);
err_printf(m, " last_seqno: 0x%08x\n", ee->last_seqno);
err_printf(m, " waiting: %s\n", yesno(ee->waiting));
err_printf(m, " ring->head: 0x%08x\n", ee->cpu_ring_head);
err_printf(m, " ring->tail: 0x%08x\n", ee->cpu_ring_tail);
err_printf(m, " hangcheck: %s [%d]\n",
hangcheck_action_to_str(ee->hangcheck_action),
ee->hangcheck_score);
}
void i915_error_printf(struct drm_i915_error_state_buf *e, const char *f, ...)
{
va_list args;
va_start(args, f);
i915_error_vprintf(e, f, args);
va_end(args);
}
static void print_error_obj(struct drm_i915_error_state_buf *m,
struct drm_i915_error_object *obj)
{
int page, offset, elt;
for (page = offset = 0; page < obj->page_count; page++) {
for (elt = 0; elt < PAGE_SIZE/4; elt++) {
err_printf(m, "%08x : %08x\n", offset,
obj->pages[page][elt]);
offset += 4;
}
}
}
int i915_error_state_to_str(struct drm_i915_error_state_buf *m,
const struct i915_error_state_file_priv *error_priv)
{
struct drm_device *dev = error_priv->dev;
struct drm_i915_private *dev_priv = to_i915(dev);
struct drm_i915_error_state *error = error_priv->error;
struct drm_i915_error_object *obj;
int i, j, offset, elt;
int max_hangcheck_score;
if (!error) {
err_printf(m, "no error state collected\n");
goto out;
}
err_printf(m, "%s\n", error->error_msg);
err_printf(m, "Time: %ld s %ld us\n", error->time.tv_sec,
error->time.tv_usec);
err_printf(m, "Kernel: " UTS_RELEASE "\n");
max_hangcheck_score = 0;
for (i = 0; i < ARRAY_SIZE(error->engine); i++) {
if (error->engine[i].hangcheck_score > max_hangcheck_score)
max_hangcheck_score = error->engine[i].hangcheck_score;
}
for (i = 0; i < ARRAY_SIZE(error->engine); i++) {
if (error->engine[i].hangcheck_score == max_hangcheck_score &&
error->engine[i].pid != -1) {
err_printf(m, "Active process (on ring %s): %s [%d]\n",
engine_str(i),
error->engine[i].comm,
error->engine[i].pid);
}
}
err_printf(m, "Reset count: %u\n", error->reset_count);
err_printf(m, "Suspend count: %u\n", error->suspend_count);
err_printf(m, "PCI ID: 0x%04x\n", dev->pdev->device);
err_printf(m, "PCI Revision: 0x%02x\n", dev->pdev->revision);
err_printf(m, "PCI Subsystem: %04x:%04x\n",
dev->pdev->subsystem_vendor,
dev->pdev->subsystem_device);
err_printf(m, "IOMMU enabled?: %d\n", error->iommu);
if (HAS_CSR(dev)) {
struct intel_csr *csr = &dev_priv->csr;
err_printf(m, "DMC loaded: %s\n",
yesno(csr->dmc_payload != NULL));
err_printf(m, "DMC fw version: %d.%d\n",
CSR_VERSION_MAJOR(csr->version),
CSR_VERSION_MINOR(csr->version));
}
err_printf(m, "EIR: 0x%08x\n", error->eir);
err_printf(m, "IER: 0x%08x\n", error->ier);
if (INTEL_INFO(dev)->gen >= 8) {
for (i = 0; i < 4; i++)
err_printf(m, "GTIER gt %d: 0x%08x\n", i,
error->gtier[i]);
} else if (HAS_PCH_SPLIT(dev) || IS_VALLEYVIEW(dev))
err_printf(m, "GTIER: 0x%08x\n", error->gtier[0]);
err_printf(m, "PGTBL_ER: 0x%08x\n", error->pgtbl_er);
err_printf(m, "FORCEWAKE: 0x%08x\n", error->forcewake);
err_printf(m, "DERRMR: 0x%08x\n", error->derrmr);
err_printf(m, "CCID: 0x%08x\n", error->ccid);
err_printf(m, "Missed interrupts: 0x%08lx\n", dev_priv->gpu_error.missed_irq_rings);
for (i = 0; i < dev_priv->num_fence_regs; i++)
err_printf(m, " fence[%d] = %08llx\n", i, error->fence[i]);
for (i = 0; i < ARRAY_SIZE(error->extra_instdone); i++)
err_printf(m, " INSTDONE_%d: 0x%08x\n", i,
error->extra_instdone[i]);
if (INTEL_INFO(dev)->gen >= 6) {
err_printf(m, "ERROR: 0x%08x\n", error->error);
if (INTEL_INFO(dev)->gen >= 8)
err_printf(m, "FAULT_TLB_DATA: 0x%08x 0x%08x\n",
error->fault_data1, error->fault_data0);
err_printf(m, "DONE_REG: 0x%08x\n", error->done_reg);
}
if (IS_GEN7(dev))
err_printf(m, "ERR_INT: 0x%08x\n", error->err_int);
for (i = 0; i < ARRAY_SIZE(error->engine); i++) {
if (error->engine[i].engine_id != -1)
error_print_engine(m, &error->engine[i]);
}
for (i = 0; i < error->vm_count; i++) {
err_printf(m, "vm[%d]\n", i);
print_error_buffers(m, "Active",
error->active_bo[i],
error->active_bo_count[i]);
print_error_buffers(m, "Pinned",
error->pinned_bo[i],
error->pinned_bo_count[i]);
}
for (i = 0; i < ARRAY_SIZE(error->engine); i++) {
struct drm_i915_error_engine *ee = &error->engine[i];
obj = ee->batchbuffer;
if (obj) {
err_puts(m, dev_priv->engine[i].name);
if (ee->pid != -1)
err_printf(m, " (submitted by %s [%d])",
ee->comm,
ee->pid);
err_printf(m, " --- gtt_offset = 0x%08x %08x\n",
upper_32_bits(obj->gtt_offset),
lower_32_bits(obj->gtt_offset));
print_error_obj(m, obj);
}
obj = ee->wa_batchbuffer;
if (obj) {
err_printf(m, "%s (w/a) --- gtt_offset = 0x%08x\n",
dev_priv->engine[i].name,
lower_32_bits(obj->gtt_offset));
print_error_obj(m, obj);
}
if (ee->num_requests) {
err_printf(m, "%s --- %d requests\n",
dev_priv->engine[i].name,
ee->num_requests);
for (j = 0; j < ee->num_requests; j++) {
err_printf(m, " seqno 0x%08x, emitted %ld, tail 0x%08x\n",
ee->requests[j].seqno,
ee->requests[j].jiffies,
ee->requests[j].tail);
}
}
if (ee->num_waiters) {
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 10:23:15 -06:00
err_printf(m, "%s --- %d waiters\n",
dev_priv->engine[i].name,
ee->num_waiters);
for (j = 0; j < ee->num_waiters; j++) {
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 10:23:15 -06:00
err_printf(m, " seqno 0x%08x for %s [%d]\n",
ee->waiters[j].seqno,
ee->waiters[j].comm,
ee->waiters[j].pid);
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 10:23:15 -06:00
}
}
if ((obj = ee->ringbuffer)) {
err_printf(m, "%s --- ringbuffer = 0x%08x\n",
dev_priv->engine[i].name,
lower_32_bits(obj->gtt_offset));
print_error_obj(m, obj);
}
if ((obj = ee->hws_page)) {
u64 hws_offset = obj->gtt_offset;
u32 *hws_page = &obj->pages[0][0];
if (i915.enable_execlists) {
hws_offset += LRC_PPHWSP_PN * PAGE_SIZE;
hws_page = &obj->pages[LRC_PPHWSP_PN][0];
}
drm/i915: Integrate GuC-based command submission GuC-based submission is mostly the same as execlist mode, up to intel_logical_ring_advance_and_submit(), where the context being dispatched would be added to the execlist queue; at this point we submit the context to the GuC backend instead. There are, however, a few other changes also required, notably: 1. Contexts must be pinned at GGTT addresses accessible by the GuC i.e. NOT in the range [0..WOPCM_SIZE), so we have to add the PIN_OFFSET_BIAS flag to the relevant GGTT-pinning calls. 2. The GuC's TLB must be invalidated after a context is pinned at a new GGTT address. 3. GuC firmware uses the one page before Ring Context as shared data. Therefore, whenever driver wants to get base address of LRC, we will offset one page for it. LRC_PPHWSP_PN is defined as the page number of LRCA. 4. In the work queue used to pass requests to the GuC, the GuC firmware requires the ring-tail-offset to be represented as an 11-bit value, expressed in QWords. Therefore, the ringbuffer size must be reduced to the representable range (4 pages). v2: Defer adding #defines until needed [Chris Wilson] Rationalise type declarations [Chris Wilson] v4: Squashed kerneldoc patch into here [Daniel Vetter] v5: Update request->tail in code common to both GuC and execlist modes. Add a private version of lr_context_update(), as sharing the execlist version leads to race conditions when the CPU and the GuC both update TAIL in the context image. Conversion of error-captured HWS page to string must account for offset from start of object to actual HWS (LRC_PPHWSP_PN). Issue: VIZ-4884 Signed-off-by: Alex Dai <yu.dai@intel.com> Signed-off-by: Dave Gordon <david.s.gordon@intel.com> Reviewed-by: Tom O'Rourke <Tom.O'Rourke@intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-08-12 08:43:43 -06:00
err_printf(m, "%s --- HW Status = 0x%08llx\n",
dev_priv->engine[i].name, hws_offset);
offset = 0;
for (elt = 0; elt < PAGE_SIZE/16; elt += 4) {
err_printf(m, "[%04x] %08x %08x %08x %08x\n",
offset,
hws_page[elt],
hws_page[elt+1],
hws_page[elt+2],
hws_page[elt+3]);
offset += 16;
}
}
obj = ee->wa_ctx;
if (obj) {
u64 wa_ctx_offset = obj->gtt_offset;
u32 *wa_ctx_page = &obj->pages[0][0];
struct intel_engine_cs *engine = &dev_priv->engine[RCS];
u32 wa_ctx_size = (engine->wa_ctx.indirect_ctx.size +
engine->wa_ctx.per_ctx.size);
err_printf(m, "%s --- WA ctx batch buffer = 0x%08llx\n",
dev_priv->engine[i].name, wa_ctx_offset);
offset = 0;
for (elt = 0; elt < wa_ctx_size; elt += 4) {
err_printf(m, "[%04x] %08x %08x %08x %08x\n",
offset,
wa_ctx_page[elt + 0],
wa_ctx_page[elt + 1],
wa_ctx_page[elt + 2],
wa_ctx_page[elt + 3]);
offset += 16;
}
}
if ((obj = ee->ctx)) {
err_printf(m, "%s --- HW Context = 0x%08x\n",
dev_priv->engine[i].name,
lower_32_bits(obj->gtt_offset));
print_error_obj(m, obj);
}
}
if ((obj = error->semaphore_obj)) {
err_printf(m, "Semaphore page = 0x%08x\n",
lower_32_bits(obj->gtt_offset));
for (elt = 0; elt < PAGE_SIZE/16; elt += 4) {
err_printf(m, "[%04x] %08x %08x %08x %08x\n",
elt * 4,
obj->pages[0][elt],
obj->pages[0][elt+1],
obj->pages[0][elt+2],
obj->pages[0][elt+3]);
}
}
if (error->overlay)
intel_overlay_print_error_state(m, error->overlay);
if (error->display)
intel_display_print_error_state(m, dev, error->display);
out:
if (m->bytes == 0 && m->err)
return m->err;
return 0;
}
int i915_error_state_buf_init(struct drm_i915_error_state_buf *ebuf,
struct drm_i915_private *i915,
size_t count, loff_t pos)
{
memset(ebuf, 0, sizeof(*ebuf));
ebuf->i915 = i915;
/* We need to have enough room to store any i915_error_state printf
* so that we can move it to start position.
*/
ebuf->size = count + 1 > PAGE_SIZE ? count + 1 : PAGE_SIZE;
ebuf->buf = kmalloc(ebuf->size,
GFP_TEMPORARY | __GFP_NORETRY | __GFP_NOWARN);
if (ebuf->buf == NULL) {
ebuf->size = PAGE_SIZE;
ebuf->buf = kmalloc(ebuf->size, GFP_TEMPORARY);
}
if (ebuf->buf == NULL) {
ebuf->size = 128;
ebuf->buf = kmalloc(ebuf->size, GFP_TEMPORARY);
}
if (ebuf->buf == NULL)
return -ENOMEM;
ebuf->start = pos;
return 0;
}
static void i915_error_object_free(struct drm_i915_error_object *obj)
{
int page;
if (obj == NULL)
return;
for (page = 0; page < obj->page_count; page++)
kfree(obj->pages[page]);
kfree(obj);
}
static void i915_error_state_free(struct kref *error_ref)
{
struct drm_i915_error_state *error = container_of(error_ref,
typeof(*error), ref);
int i;
for (i = 0; i < ARRAY_SIZE(error->engine); i++) {
struct drm_i915_error_engine *ee = &error->engine[i];
i915_error_object_free(ee->batchbuffer);
i915_error_object_free(ee->wa_batchbuffer);
i915_error_object_free(ee->ringbuffer);
i915_error_object_free(ee->hws_page);
i915_error_object_free(ee->ctx);
i915_error_object_free(ee->wa_ctx);
kfree(ee->requests);
kfree(ee->waiters);
}
i915_error_object_free(error->semaphore_obj);
drm/i915: Do not leak objects after capturing error state While running kmemleak chasing a different memleak, I saw that the capture_error_state function was leaking some objects, for example: unreferenced object 0xffff8800a9b72148 (size 8192): comm "kworker/u16:0", pid 1499, jiffies 4295201243 (age 990.096s) hex dump (first 32 bytes): 00 00 04 00 00 00 00 00 5d f4 ff ff 00 00 00 00 ........]....... 00 30 b0 01 00 00 00 00 37 00 00 00 00 00 00 00 .0......7....... backtrace: [<ffffffff811e5ae4>] create_object+0x104/0x2c0 [<ffffffff8178f50a>] kmemleak_alloc+0x7a/0xc0 [<ffffffff811cde4b>] __kmalloc+0xeb/0x220 [<ffffffffa038f1d9>] kcalloc.constprop.12+0x2d/0x2f [i915] [<ffffffffa0316064>] i915_capture_error_state+0x3f4/0x1660 [i915] [<ffffffffa03207df>] i915_handle_error+0x7f/0x660 [i915] [<ffffffffa03210f7>] i915_hangcheck_elapsed+0x2e7/0x470 [i915] [<ffffffff8108d574>] process_one_work+0x144/0x490 [<ffffffff8108dfbd>] worker_thread+0x11d/0x530 [<ffffffff81094079>] kthread+0xc9/0xe0 [<ffffffff817a2398>] ret_from_fork+0x58/0x90 [<ffffffffffffffff>] 0xffffffffffffffff The following objects are allocated in i915_gem_capture_buffers, but not released in i915_error_state_free: - error->active_bo_count - error->pinned_bo - error->pinned_bo_count - error->active_bo[vm_count] (allocated in i915_gem_capture_vm). The leaks were introduced by commit 95f5301dd880da2dea2c9a9c29750064536d426a Author: Ben Widawsky <ben@bwidawsk.net> Date: Wed Jul 31 17:00:15 2013 -0700 drm/i915: Update error capture for VMs v2: Reuse iterator and add culprit commit details (Chris) Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Michel Thierry <michel.thierry@intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-03-20 03:41:03 -06:00
for (i = 0; i < error->vm_count; i++)
kfree(error->active_bo[i]);
kfree(error->active_bo);
drm/i915: Do not leak objects after capturing error state While running kmemleak chasing a different memleak, I saw that the capture_error_state function was leaking some objects, for example: unreferenced object 0xffff8800a9b72148 (size 8192): comm "kworker/u16:0", pid 1499, jiffies 4295201243 (age 990.096s) hex dump (first 32 bytes): 00 00 04 00 00 00 00 00 5d f4 ff ff 00 00 00 00 ........]....... 00 30 b0 01 00 00 00 00 37 00 00 00 00 00 00 00 .0......7....... backtrace: [<ffffffff811e5ae4>] create_object+0x104/0x2c0 [<ffffffff8178f50a>] kmemleak_alloc+0x7a/0xc0 [<ffffffff811cde4b>] __kmalloc+0xeb/0x220 [<ffffffffa038f1d9>] kcalloc.constprop.12+0x2d/0x2f [i915] [<ffffffffa0316064>] i915_capture_error_state+0x3f4/0x1660 [i915] [<ffffffffa03207df>] i915_handle_error+0x7f/0x660 [i915] [<ffffffffa03210f7>] i915_hangcheck_elapsed+0x2e7/0x470 [i915] [<ffffffff8108d574>] process_one_work+0x144/0x490 [<ffffffff8108dfbd>] worker_thread+0x11d/0x530 [<ffffffff81094079>] kthread+0xc9/0xe0 [<ffffffff817a2398>] ret_from_fork+0x58/0x90 [<ffffffffffffffff>] 0xffffffffffffffff The following objects are allocated in i915_gem_capture_buffers, but not released in i915_error_state_free: - error->active_bo_count - error->pinned_bo - error->pinned_bo_count - error->active_bo[vm_count] (allocated in i915_gem_capture_vm). The leaks were introduced by commit 95f5301dd880da2dea2c9a9c29750064536d426a Author: Ben Widawsky <ben@bwidawsk.net> Date: Wed Jul 31 17:00:15 2013 -0700 drm/i915: Update error capture for VMs v2: Reuse iterator and add culprit commit details (Chris) Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Michel Thierry <michel.thierry@intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-03-20 03:41:03 -06:00
kfree(error->active_bo_count);
kfree(error->pinned_bo);
kfree(error->pinned_bo_count);
kfree(error->overlay);
kfree(error->display);
kfree(error);
}
static struct drm_i915_error_object *
i915_error_object_create(struct drm_i915_private *dev_priv,
struct drm_i915_gem_object *src,
struct i915_address_space *vm)
{
struct i915_ggtt *ggtt = &dev_priv->ggtt;
struct drm_i915_error_object *dst;
struct i915_vma *vma = NULL;
int num_pages;
bool use_ggtt;
int i = 0;
u64 reloc_offset;
if (src == NULL || src->pages == NULL)
return NULL;
num_pages = src->base.size >> PAGE_SHIFT;
dst = kmalloc(sizeof(*dst) + num_pages * sizeof(u32 *), GFP_ATOMIC);
if (dst == NULL)
return NULL;
if (i915_gem_obj_bound(src, vm))
dst->gtt_offset = i915_gem_obj_offset(src, vm);
else
dst->gtt_offset = -1;
reloc_offset = dst->gtt_offset;
if (i915_is_ggtt(vm))
vma = i915_gem_obj_to_ggtt(src);
use_ggtt = (src->cache_level == I915_CACHE_NONE &&
vma && (vma->bound & GLOBAL_BIND) &&
reloc_offset + num_pages * PAGE_SIZE <= ggtt->mappable_end);
/* Cannot access stolen address directly, try to use the aperture */
if (src->stolen) {
use_ggtt = true;
if (!(vma && vma->bound & GLOBAL_BIND))
goto unwind;
reloc_offset = i915_gem_obj_ggtt_offset(src);
if (reloc_offset + num_pages * PAGE_SIZE > ggtt->mappable_end)
goto unwind;
}
/* Cannot access snooped pages through the aperture */
if (use_ggtt && src->cache_level != I915_CACHE_NONE &&
!HAS_LLC(dev_priv))
goto unwind;
dst->page_count = num_pages;
while (num_pages--) {
unsigned long flags;
void *d;
d = kmalloc(PAGE_SIZE, GFP_ATOMIC);
if (d == NULL)
goto unwind;
local_irq_save(flags);
if (use_ggtt) {
void __iomem *s;
/* Simply ignore tiling or any overlapping fence.
* It's part of the error state, and this hopefully
* captures what the GPU read.
*/
s = io_mapping_map_atomic_wc(ggtt->mappable,
reloc_offset);
memcpy_fromio(d, s, PAGE_SIZE);
io_mapping_unmap_atomic(s);
} else {
struct page *page;
void *s;
page = i915_gem_object_get_page(src, i);
drm_clflush_pages(&page, 1);
s = kmap_atomic(page);
memcpy(d, s, PAGE_SIZE);
kunmap_atomic(s);
drm_clflush_pages(&page, 1);
}
local_irq_restore(flags);
dst->pages[i++] = d;
reloc_offset += PAGE_SIZE;
}
return dst;
unwind:
while (i--)
kfree(dst->pages[i]);
kfree(dst);
return NULL;
}
#define i915_error_ggtt_object_create(dev_priv, src) \
i915_error_object_create((dev_priv), (src), &(dev_priv)->ggtt.base)
static void capture_bo(struct drm_i915_error_buffer *err,
struct i915_vma *vma)
{
struct drm_i915_gem_object *obj = vma->obj;
struct intel_engine_cs *engine;
drm/i915: Implement inter-engine read-read optimisations Currently, we only track the last request globally across all engines. This prevents us from issuing concurrent read requests on e.g. the RCS and BCS engines (or more likely the render and media engines). Without semaphores, we incur costly stalls as we synchronise between rings - greatly impacting the current performance of Broadwell versus Haswell in certain workloads (like video decode). With the introduction of reference counted requests, it is much easier to track the last request per ring, as well as the last global write request so that we can optimise inter-engine read read requests (as well as better optimise certain CPU waits). v2: Fix inverted readonly condition for nonblocking waits. v3: Handle non-continguous engine array after waits v4: Rebase, tidy, rewrite ring list debugging v5: Use obj->active as a bitfield, it looks cool v6: Micro-optimise, mostly involving moving code around v7: Fix retire-requests-upto for execlists (and multiple rq->ringbuf) v8: Rebase v9: Refactor i915_gem_object_sync() to allow the compiler to better optimise it. Benchmark: igt/gem_read_read_speed hsw:gt3e (with semaphores): Before: Time to read-read 1024k: 275.794µs After: Time to read-read 1024k: 123.260µs hsw:gt3e (w/o semaphores): Before: Time to read-read 1024k: 230.433µs After: Time to read-read 1024k: 124.593µs bdw-u (w/o semaphores): Before After Time to read-read 1x1: 26.274µs 10.350µs Time to read-read 128x128: 40.097µs 21.366µs Time to read-read 256x256: 77.087µs 42.608µs Time to read-read 512x512: 281.999µs 181.155µs Time to read-read 1024x1024: 1196.141µs 1118.223µs Time to read-read 2048x2048: 5639.072µs 5225.837µs Time to read-read 4096x4096: 22401.662µs 21137.067µs Time to read-read 8192x8192: 89617.735µs 85637.681µs Testcase: igt/gem_concurrent_blit (read-read and friends) Cc: Lionel Landwerlin <lionel.g.landwerlin@linux.intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> [v8] [danvet: s/\<rq\>/req/g] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-04-27 06:41:17 -06:00
int i;
err->size = obj->base.size;
err->name = obj->base.name;
for (i = 0; i < I915_NUM_ENGINES; i++)
err->rseqno[i] = i915_gem_active_get_seqno(&obj->last_read[i]);
err->wseqno = i915_gem_active_get_seqno(&obj->last_write);
err->gtt_offset = vma->node.start;
err->read_domains = obj->base.read_domains;
err->write_domain = obj->base.write_domain;
err->fence_reg = obj->fence_reg;
err->pinned = 0;
if (i915_gem_obj_is_pinned(obj))
err->pinned = 1;
err->tiling = obj->tiling_mode;
err->dirty = obj->dirty;
err->purgeable = obj->madv != I915_MADV_WILLNEED;
drm/i915: Introduce mapping of user pages into video memory (userptr) ioctl By exporting the ability to map user address and inserting PTEs representing their backing pages into the GTT, we can exploit UMA in order to utilize normal application data as a texture source or even as a render target (depending upon the capabilities of the chipset). This has a number of uses, with zero-copy downloads to the GPU and efficient readback making the intermixed streaming of CPU and GPU operations fairly efficient. This ability has many widespread implications from faster rendering of client-side software rasterisers (chromium), mitigation of stalls due to read back (firefox) and to faster pipelining of texture data (such as pixel buffer objects in GL or data blobs in CL). v2: Compile with CONFIG_MMU_NOTIFIER v3: We can sleep while performing invalidate-range, which we can utilise to drop our page references prior to the kernel manipulating the vma (for either discard or cloning) and so protect normal users. v4: Only run the invalidate notifier if the range intercepts the bo. v5: Prevent userspace from attempting to GTT mmap non-page aligned buffers v6: Recheck after reacquire mutex for lost mmu. v7: Fix implicit padding of ioctl struct by rounding to next 64bit boundary. v8: Fix rebasing error after forwarding porting the back port. v9: Limit the userptr to page aligned entries. We now expect userspace to handle all the offset-in-page adjustments itself. v10: Prevent vma from being copied across fork to avoid issues with cow. v11: Drop vma behaviour changes -- locking is nigh on impossible. Use a worker to load user pages to avoid lock inversions. v12: Use get_task_mm()/mmput() for correct refcounting of mm. v13: Use a worker to release the mmu_notifier to avoid lock inversion v14: Decouple mmu_notifier from struct_mutex using a custom mmu_notifer with its own locking and tree of objects for each mm/mmu_notifier. v15: Prevent overlapping userptr objects, and invalidate all objects within the mmu_notifier range v16: Fix a typo for iterating over multiple objects in the range and rearrange error path to destroy the mmu_notifier locklessly. Also close a race between invalidate_range and the get_pages_worker. v17: Close a race between get_pages_worker/invalidate_range and fresh allocations of the same userptr range - and notice that struct_mutex was presumed to be held when during creation it wasn't. v18: Sigh. Fix the refactor of st_set_pages() to allocate enough memory for the struct sg_table and to clear it before reporting an error. v19: Always error out on read-only userptr requests as we don't have the hardware infrastructure to support them at the moment. v20: Refuse to implement read-only support until we have the required infrastructure - but reserve the bit in flags for future use. v21: use_mm() is not required for get_user_pages(). It is only meant to be used to fix up the kernel thread's current->mm for use with copy_user(). v22: Use sg_alloc_table_from_pages for that chunky feeling v23: Export a function for sanity checking dma-buf rather than encode userptr details elsewhere, and clean up comments based on suggestions by Bradley. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Akash Goel <akash.goel@intel.com> Cc: "Volkin, Bradley D" <bradley.d.volkin@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Reviewed-by: Brad Volkin <bradley.d.volkin@intel.com> [danvet: Frob ioctl allocation to pick the next one - will cause a bit of fuss with create2 apparently, but such are the rules.] [danvet2: oops, forgot to git add after manual patch application] [danvet3: Appease sparse.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-05-16 07:22:37 -06:00
err->userptr = obj->userptr.mm != NULL;
err->cache_level = obj->cache_level;
engine = i915_gem_active_get_engine(&obj->last_write);
err->engine = engine ? engine->id : -1;
}
static u32 capture_active_bo(struct drm_i915_error_buffer *err,
int count, struct list_head *head)
{
2013-07-31 18:00:14 -06:00
struct i915_vma *vma;
int i = 0;
list_for_each_entry(vma, head, vm_link) {
capture_bo(err++, vma);
if (++i == count)
break;
}
return i;
}
static u32 capture_pinned_bo(struct drm_i915_error_buffer *err,
int count, struct list_head *head,
struct i915_address_space *vm)
{
struct drm_i915_gem_object *obj;
struct drm_i915_error_buffer * const first = err;
struct drm_i915_error_buffer * const last = err + count;
list_for_each_entry(obj, head, global_list) {
struct i915_vma *vma;
if (err == last)
break;
list_for_each_entry(vma, &obj->vma_list, obj_link)
drm/i915: Infrastructure for supporting different GGTT views per object Things like reliable GGTT mappings and mirrored 2d-on-3d display will need to map objects into the same address space multiple times. Added a GGTT view concept and linked it with the VMA to distinguish between multiple instances per address space. New objects and GEM functions which do not take this new view as a parameter assume the default of zero (I915_GGTT_VIEW_NORMAL) which preserves the previous behaviour. This now means that objects can have multiple VMA entries so the code which assumed there will only be one also had to be modified. Alternative GGTT views are supposed to borrow DMA addresses from obj->pages which is DMA mapped on first VMA instantiation and unmapped on the last one going away. v2: * Removed per view special casing in i915_gem_ggtt_prepare / finish_object in favour of creating and destroying DMA mappings on first VMA instantiation and last VMA destruction. (Daniel Vetter) * Simplified i915_vma_unbind which does not need to count the GGTT views. (Daniel Vetter) * Also moved obj->map_and_fenceable reset under the same check. * Checkpatch cleanups. v3: * Only retire objects once the last VMA is unbound. v4: * Keep scatter-gather table for alternative views persistent for the lifetime of the VMA. * Propagate binding errors to callers and handle appropriately. v5: * Explicitly look for normal GGTT view in i915_gem_obj_bound to align usage in i915_gem_object_ggtt_unpin. (Michel Thierry) * Change to single if statement in i915_gem_obj_to_ggtt. (Michel Thierry) * Removed stray semi-colon in i915_gem_object_set_cache_level. For: VIZ-4544 Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Michel Thierry <michel.thierry@intel.com> [danvet: Drop hunk from i915_gem_shrink since it's just prettification but upsets a __must_check warning.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-12-10 10:27:58 -07:00
if (vma->vm == vm && vma->pin_count > 0)
capture_bo(err++, vma);
}
return err - first;
}
/* Generate a semi-unique error code. The code is not meant to have meaning, The
* code's only purpose is to try to prevent false duplicated bug reports by
* grossly estimating a GPU error state.
*
* TODO Ideally, hashing the batchbuffer would be a very nice way to determine
* the hang if we could strip the GTT offset information from it.
*
* It's only a small step better than a random number in its current form.
*/
static uint32_t i915_error_generate_code(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error,
int *engine_id)
{
uint32_t error_code = 0;
int i;
/* IPEHR would be an ideal way to detect errors, as it's the gross
* measure of "the command that hung." However, has some very common
* synchronization commands which almost always appear in the case
* strictly a client bug. Use instdone to differentiate those some.
*/
for (i = 0; i < I915_NUM_ENGINES; i++) {
if (error->engine[i].hangcheck_action == HANGCHECK_HUNG) {
if (engine_id)
*engine_id = i;
return error->engine[i].ipehr ^ error->engine[i].instdone;
}
}
return error_code;
}
static void i915_gem_record_fences(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error)
{
int i;
if (IS_GEN3(dev_priv) || IS_GEN2(dev_priv)) {
for (i = 0; i < dev_priv->num_fence_regs; i++)
error->fence[i] = I915_READ(FENCE_REG(i));
} else if (IS_GEN5(dev_priv) || IS_GEN4(dev_priv)) {
for (i = 0; i < dev_priv->num_fence_regs; i++)
error->fence[i] = I915_READ64(FENCE_REG_965_LO(i));
} else if (INTEL_GEN(dev_priv) >= 6) {
for (i = 0; i < dev_priv->num_fence_regs; i++)
error->fence[i] = I915_READ64(FENCE_REG_GEN6_LO(i));
}
}
static void gen8_record_semaphore_state(struct drm_i915_error_state *error,
struct intel_engine_cs *engine,
struct drm_i915_error_engine *ee)
{
struct drm_i915_private *dev_priv = engine->i915;
struct intel_engine_cs *to;
enum intel_engine_id id;
if (!error->semaphore_obj)
return;
for_each_engine_id(to, dev_priv, id) {
int idx;
u16 signal_offset;
u32 *tmp;
if (engine == to)
continue;
signal_offset =
(GEN8_SIGNAL_OFFSET(engine, id) & (PAGE_SIZE - 1)) / 4;
tmp = error->semaphore_obj->pages[0];
idx = intel_engine_sync_index(engine, to);
ee->semaphore_mboxes[idx] = tmp[signal_offset];
ee->semaphore_seqno[idx] = engine->semaphore.sync_seqno[idx];
}
}
static void gen6_record_semaphore_state(struct intel_engine_cs *engine,
struct drm_i915_error_engine *ee)
{
struct drm_i915_private *dev_priv = engine->i915;
ee->semaphore_mboxes[0] = I915_READ(RING_SYNC_0(engine->mmio_base));
ee->semaphore_mboxes[1] = I915_READ(RING_SYNC_1(engine->mmio_base));
ee->semaphore_seqno[0] = engine->semaphore.sync_seqno[0];
ee->semaphore_seqno[1] = engine->semaphore.sync_seqno[1];
if (HAS_VEBOX(dev_priv)) {
ee->semaphore_mboxes[2] =
I915_READ(RING_SYNC_2(engine->mmio_base));
ee->semaphore_seqno[2] = engine->semaphore.sync_seqno[2];
}
}
static void error_record_engine_waiters(struct intel_engine_cs *engine,
struct drm_i915_error_engine *ee)
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 10:23:15 -06:00
{
struct intel_breadcrumbs *b = &engine->breadcrumbs;
struct drm_i915_error_waiter *waiter;
struct rb_node *rb;
int count;
ee->num_waiters = 0;
ee->waiters = NULL;
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 10:23:15 -06:00
spin_lock(&b->lock);
count = 0;
for (rb = rb_first(&b->waiters); rb != NULL; rb = rb_next(rb))
count++;
spin_unlock(&b->lock);
waiter = NULL;
if (count)
waiter = kmalloc_array(count,
sizeof(struct drm_i915_error_waiter),
GFP_ATOMIC);
if (!waiter)
return;
ee->waiters = waiter;
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 10:23:15 -06:00
spin_lock(&b->lock);
for (rb = rb_first(&b->waiters); rb; rb = rb_next(rb)) {
struct intel_wait *w = container_of(rb, typeof(*w), node);
strcpy(waiter->comm, w->tsk->comm);
waiter->pid = w->tsk->pid;
waiter->seqno = w->seqno;
waiter++;
if (++ee->num_waiters == count)
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 10:23:15 -06:00
break;
}
spin_unlock(&b->lock);
}
static void error_record_engine_registers(struct drm_i915_error_state *error,
struct intel_engine_cs *engine,
struct drm_i915_error_engine *ee)
{
struct drm_i915_private *dev_priv = engine->i915;
if (INTEL_GEN(dev_priv) >= 6) {
ee->rc_psmi = I915_READ(RING_PSMI_CTL(engine->mmio_base));
ee->fault_reg = I915_READ(RING_FAULT_REG(engine));
if (INTEL_GEN(dev_priv) >= 8)
gen8_record_semaphore_state(error, engine, ee);
else
gen6_record_semaphore_state(engine, ee);
}
if (INTEL_GEN(dev_priv) >= 4) {
ee->faddr = I915_READ(RING_DMA_FADD(engine->mmio_base));
ee->ipeir = I915_READ(RING_IPEIR(engine->mmio_base));
ee->ipehr = I915_READ(RING_IPEHR(engine->mmio_base));
ee->instdone = I915_READ(RING_INSTDONE(engine->mmio_base));
ee->instps = I915_READ(RING_INSTPS(engine->mmio_base));
ee->bbaddr = I915_READ(RING_BBADDR(engine->mmio_base));
if (INTEL_GEN(dev_priv) >= 8) {
ee->faddr |= (u64) I915_READ(RING_DMA_FADD_UDW(engine->mmio_base)) << 32;
ee->bbaddr |= (u64) I915_READ(RING_BBADDR_UDW(engine->mmio_base)) << 32;
}
ee->bbstate = I915_READ(RING_BBSTATE(engine->mmio_base));
} else {
ee->faddr = I915_READ(DMA_FADD_I8XX);
ee->ipeir = I915_READ(IPEIR);
ee->ipehr = I915_READ(IPEHR);
ee->instdone = I915_READ(GEN2_INSTDONE);
}
ee->waiting = intel_engine_has_waiter(engine);
ee->instpm = I915_READ(RING_INSTPM(engine->mmio_base));
ee->acthd = intel_engine_get_active_head(engine);
ee->seqno = intel_engine_get_seqno(engine);
ee->last_seqno = engine->last_submitted_seqno;
ee->start = I915_READ_START(engine);
ee->head = I915_READ_HEAD(engine);
ee->tail = I915_READ_TAIL(engine);
ee->ctl = I915_READ_CTL(engine);
if (I915_NEED_GFX_HWS(dev_priv)) {
drm/i915: Type safe register read/write Make I915_READ and I915_WRITE more type safe by wrapping the register offset in a struct. This should eliminate most of the fumbles we've had with misplaced parens. This only takes care of normal mmio registers. We could extend the idea to other register types and define each with its own struct. That way you wouldn't be able to accidentally pass the wrong thing to a specific register access function. The gpio_reg setup is probably the ugliest thing left. But I figure I'd just leave it for now, and wait for some divine inspiration to strike before making it nice. As for the generated code, it's actually a bit better sometimes. Eg. looking at i915_irq_handler(), we can see the following change: lea 0x70024(%rdx,%rax,1),%r9d mov $0x1,%edx - movslq %r9d,%r9 - mov %r9,%rsi - mov %r9,-0x58(%rbp) - callq *0xd8(%rbx) + mov %r9d,%esi + mov %r9d,-0x48(%rbp) callq *0xd8(%rbx) So previously gcc thought the register offset might be signed and decided to sign extend it, just in case. The rest appears to be mostly just minor shuffling of instructions. v2: i915_mmio_reg_{offset,equal,valid}() helpers added s/_REG/_MMIO/ in the register defines mo more switch statements left to worry about ring_emit stuff got sorted in a prep patch cmd parser, lrc context and w/a batch buildup also in prep patch vgpu stuff cleaned up and moved to a prep patch all other unrelated changes split out v3: Rebased due to BXT DSI/BLC, MOCS, etc. v4: Rebased due to churn, s/i915_mmio_reg_t/i915_reg_t/ Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Link: http://patchwork.freedesktop.org/patch/msgid/1447853606-2751-1-git-send-email-ville.syrjala@linux.intel.com
2015-11-18 06:33:26 -07:00
i915_reg_t mmio;
if (IS_GEN7(dev_priv)) {
switch (engine->id) {
default:
case RCS:
mmio = RENDER_HWS_PGA_GEN7;
break;
case BCS:
mmio = BLT_HWS_PGA_GEN7;
break;
case VCS:
mmio = BSD_HWS_PGA_GEN7;
break;
case VECS:
mmio = VEBOX_HWS_PGA_GEN7;
break;
}
} else if (IS_GEN6(engine->i915)) {
mmio = RING_HWS_PGA_GEN6(engine->mmio_base);
} else {
/* XXX: gen8 returns to sanity */
mmio = RING_HWS_PGA(engine->mmio_base);
}
ee->hws = I915_READ(mmio);
}
ee->hangcheck_score = engine->hangcheck.score;
ee->hangcheck_action = engine->hangcheck.action;
if (USES_PPGTT(dev_priv)) {
int i;
ee->vm_info.gfx_mode = I915_READ(RING_MODE_GEN7(engine));
if (IS_GEN6(dev_priv))
ee->vm_info.pp_dir_base =
I915_READ(RING_PP_DIR_BASE_READ(engine));
else if (IS_GEN7(dev_priv))
ee->vm_info.pp_dir_base =
I915_READ(RING_PP_DIR_BASE(engine));
else if (INTEL_GEN(dev_priv) >= 8)
for (i = 0; i < 4; i++) {
ee->vm_info.pdp[i] =
I915_READ(GEN8_RING_PDP_UDW(engine, i));
ee->vm_info.pdp[i] <<= 32;
ee->vm_info.pdp[i] |=
I915_READ(GEN8_RING_PDP_LDW(engine, i));
}
}
}
static void i915_gem_record_active_context(struct intel_engine_cs *engine,
struct drm_i915_error_state *error,
struct drm_i915_error_engine *ee)
{
struct drm_i915_private *dev_priv = engine->i915;
struct drm_i915_gem_object *obj;
/* Currently render ring is the only HW context user */
if (engine->id != RCS || !error->ccid)
return;
list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
drm/i915/error: Check the potential ctx obj's vm The bound list is global (all objects which back the VMAs are stored here). Recently the BUG() in the offset lookup was demoted to a WARN, but the fault actually lies in the caller, here. This bug has existed since the initial introduction of PPGTT (however, it was fixed in unmerged patches to fix up the error state). Note: The reason for the BUG_ON to WARN_ON demotion was _not_ to duct-tape over this bug here but another but triggerable without ppgtt. See the commit for details: commit f25748ea7342bcc44866befcbf0222cd1d89d851 Author: Daniel Vetter <daniel.vetter@ffwll.ch> Date: Tue Jun 17 22:34:38 2014 +0200 drm/i915: Don't BUG_ON in i915_gem_obj_offset A WARN_ON is perfectly fine. The BUG in here seems to be the cause behind hard-hangs when I cat the i915_gem_pageflip debugfs file (which calls this from an irq spinlock). But only while running a full igt run after a while. I still need to root cause the underlying issue. I'll also start reject patches which add new BUG_ON but don't come with a really good justification for it. The general rule really should be to just WARN and hope the driver survives for long enough. v2: Make the WARN a bit more useful per Chris' suggestion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Ben Widawsky <ben@bwidawsk.net> [danvet: Clarfy that the WARN_ON (former BUG_ON) in ggtt_offset caught more than just this bug fixed in this patch here.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-07-01 12:17:41 -06:00
if (!i915_gem_obj_ggtt_bound(obj))
continue;
if ((error->ccid & PAGE_MASK) == i915_gem_obj_ggtt_offset(obj)) {
ee->ctx = i915_error_ggtt_object_create(dev_priv, obj);
break;
}
}
}
static void i915_gem_record_rings(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error)
{
struct i915_ggtt *ggtt = &dev_priv->ggtt;
struct drm_i915_gem_request *request;
int i, count;
if (dev_priv->semaphore_obj) {
error->semaphore_obj =
i915_error_ggtt_object_create(dev_priv,
dev_priv->semaphore_obj);
}
for (i = 0; i < I915_NUM_ENGINES; i++) {
struct intel_engine_cs *engine = &dev_priv->engine[i];
struct drm_i915_error_engine *ee = &error->engine[i];
ee->pid = -1;
ee->engine_id = -1;
if (!intel_engine_initialized(engine))
continue;
ee->engine_id = i;
error_record_engine_registers(error, engine, ee);
error_record_engine_waiters(engine, ee);
request = i915_gem_find_active_request(engine);
if (request) {
struct i915_address_space *vm;
struct intel_ring *ring;
vm = request->ctx->ppgtt ?
&request->ctx->ppgtt->base : &ggtt->base;
/* We need to copy these to an anonymous buffer
* as the simplest method to avoid being overwritten
* by userspace.
*/
ee->batchbuffer =
i915_error_object_create(dev_priv,
request->batch_obj,
vm);
if (HAS_BROKEN_CS_TLB(dev_priv))
ee->wa_batchbuffer =
i915_error_ggtt_object_create(dev_priv,
engine->scratch.obj);
if (request->pid) {
struct task_struct *task;
rcu_read_lock();
task = pid_task(request->pid, PIDTYPE_PID);
if (task) {
strcpy(ee->comm, task->comm);
ee->pid = task->pid;
}
rcu_read_unlock();
}
error->simulated |=
request->ctx->flags & CONTEXT_NO_ERROR_CAPTURE;
ring = request->ring;
ee->cpu_ring_head = ring->head;
ee->cpu_ring_tail = ring->tail;
ee->ringbuffer =
i915_error_ggtt_object_create(dev_priv,
ring->obj);
}
ee->hws_page =
i915_error_ggtt_object_create(dev_priv,
engine->status_page.obj);
ee->wa_ctx = i915_error_ggtt_object_create(dev_priv,
engine->wa_ctx.obj);
i915_gem_record_active_context(engine, error, ee);
count = 0;
list_for_each_entry(request, &engine->request_list, list)
count++;
ee->num_requests = count;
ee->requests =
kcalloc(count, sizeof(*ee->requests), GFP_ATOMIC);
if (!ee->requests) {
ee->num_requests = 0;
continue;
}
count = 0;
list_for_each_entry(request, &engine->request_list, list) {
struct drm_i915_error_request *erq;
if (count >= ee->num_requests) {
/*
* If the ring request list was changed in
* between the point where the error request
* list was created and dimensioned and this
* point then just exit early to avoid crashes.
*
* We don't need to communicate that the
* request list changed state during error
* state capture and that the error state is
* slightly incorrect as a consequence since we
* are typically only interested in the request
* list state at the point of error state
* capture, not in any changes happening during
* the capture.
*/
break;
}
erq = &ee->requests[count++];
erq->seqno = request->fence.seqno;
erq->jiffies = request->emitted_jiffies;
erq->tail = request->postfix;
}
}
}
/* FIXME: Since pin count/bound list is global, we duplicate what we capture per
* VM.
*/
static void i915_gem_capture_vm(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error,
struct i915_address_space *vm,
const int ndx)
{
struct drm_i915_error_buffer *active_bo = NULL, *pinned_bo = NULL;
struct drm_i915_gem_object *obj;
struct i915_vma *vma;
int i;
i = 0;
list_for_each_entry(vma, &vm->active_list, vm_link)
i++;
error->active_bo_count[ndx] = i;
list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
list_for_each_entry(vma, &obj->vma_list, obj_link)
drm/i915: Infrastructure for supporting different GGTT views per object Things like reliable GGTT mappings and mirrored 2d-on-3d display will need to map objects into the same address space multiple times. Added a GGTT view concept and linked it with the VMA to distinguish between multiple instances per address space. New objects and GEM functions which do not take this new view as a parameter assume the default of zero (I915_GGTT_VIEW_NORMAL) which preserves the previous behaviour. This now means that objects can have multiple VMA entries so the code which assumed there will only be one also had to be modified. Alternative GGTT views are supposed to borrow DMA addresses from obj->pages which is DMA mapped on first VMA instantiation and unmapped on the last one going away. v2: * Removed per view special casing in i915_gem_ggtt_prepare / finish_object in favour of creating and destroying DMA mappings on first VMA instantiation and last VMA destruction. (Daniel Vetter) * Simplified i915_vma_unbind which does not need to count the GGTT views. (Daniel Vetter) * Also moved obj->map_and_fenceable reset under the same check. * Checkpatch cleanups. v3: * Only retire objects once the last VMA is unbound. v4: * Keep scatter-gather table for alternative views persistent for the lifetime of the VMA. * Propagate binding errors to callers and handle appropriately. v5: * Explicitly look for normal GGTT view in i915_gem_obj_bound to align usage in i915_gem_object_ggtt_unpin. (Michel Thierry) * Change to single if statement in i915_gem_obj_to_ggtt. (Michel Thierry) * Removed stray semi-colon in i915_gem_object_set_cache_level. For: VIZ-4544 Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by: Michel Thierry <michel.thierry@intel.com> [danvet: Drop hunk from i915_gem_shrink since it's just prettification but upsets a __must_check warning.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-12-10 10:27:58 -07:00
if (vma->vm == vm && vma->pin_count > 0)
i++;
}
error->pinned_bo_count[ndx] = i - error->active_bo_count[ndx];
if (i) {
active_bo = kcalloc(i, sizeof(*active_bo), GFP_ATOMIC);
if (active_bo)
pinned_bo = active_bo + error->active_bo_count[ndx];
}
if (active_bo)
error->active_bo_count[ndx] =
capture_active_bo(active_bo,
error->active_bo_count[ndx],
&vm->active_list);
if (pinned_bo)
error->pinned_bo_count[ndx] =
capture_pinned_bo(pinned_bo,
error->pinned_bo_count[ndx],
&dev_priv->mm.bound_list, vm);
error->active_bo[ndx] = active_bo;
error->pinned_bo[ndx] = pinned_bo;
}
static void i915_gem_capture_buffers(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error)
{
struct i915_address_space *vm;
int cnt = 0, i = 0;
list_for_each_entry(vm, &dev_priv->vm_list, global_link)
cnt++;
error->active_bo = kcalloc(cnt, sizeof(*error->active_bo), GFP_ATOMIC);
error->pinned_bo = kcalloc(cnt, sizeof(*error->pinned_bo), GFP_ATOMIC);
error->active_bo_count = kcalloc(cnt, sizeof(*error->active_bo_count),
GFP_ATOMIC);
error->pinned_bo_count = kcalloc(cnt, sizeof(*error->pinned_bo_count),
GFP_ATOMIC);
if (error->active_bo == NULL ||
error->pinned_bo == NULL ||
error->active_bo_count == NULL ||
error->pinned_bo_count == NULL) {
kfree(error->active_bo);
kfree(error->active_bo_count);
kfree(error->pinned_bo);
kfree(error->pinned_bo_count);
error->active_bo = NULL;
error->active_bo_count = NULL;
error->pinned_bo = NULL;
error->pinned_bo_count = NULL;
} else {
list_for_each_entry(vm, &dev_priv->vm_list, global_link)
i915_gem_capture_vm(dev_priv, error, vm, i++);
error->vm_count = cnt;
}
}
/* Capture all registers which don't fit into another category. */
static void i915_capture_reg_state(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error)
{
struct drm_device *dev = &dev_priv->drm;
int i;
/* General organization
* 1. Registers specific to a single generation
* 2. Registers which belong to multiple generations
* 3. Feature specific registers.
* 4. Everything else
* Please try to follow the order.
*/
/* 1: Registers specific to a single generation */
if (IS_VALLEYVIEW(dev)) {
error->gtier[0] = I915_READ(GTIER);
error->ier = I915_READ(VLV_IER);
error->forcewake = I915_READ_FW(FORCEWAKE_VLV);
}
if (IS_GEN7(dev))
error->err_int = I915_READ(GEN7_ERR_INT);
if (INTEL_INFO(dev)->gen >= 8) {
error->fault_data0 = I915_READ(GEN8_FAULT_TLB_DATA0);
error->fault_data1 = I915_READ(GEN8_FAULT_TLB_DATA1);
}
if (IS_GEN6(dev)) {
error->forcewake = I915_READ_FW(FORCEWAKE);
error->gab_ctl = I915_READ(GAB_CTL);
error->gfx_mode = I915_READ(GFX_MODE);
}
/* 2: Registers which belong to multiple generations */
if (INTEL_INFO(dev)->gen >= 7)
error->forcewake = I915_READ_FW(FORCEWAKE_MT);
if (INTEL_INFO(dev)->gen >= 6) {
error->derrmr = I915_READ(DERRMR);
error->error = I915_READ(ERROR_GEN6);
error->done_reg = I915_READ(DONE_REG);
}
/* 3: Feature specific registers */
if (IS_GEN6(dev) || IS_GEN7(dev)) {
error->gam_ecochk = I915_READ(GAM_ECOCHK);
error->gac_eco = I915_READ(GAC_ECO_BITS);
}
/* 4: Everything else */
if (HAS_HW_CONTEXTS(dev))
error->ccid = I915_READ(CCID);
if (INTEL_INFO(dev)->gen >= 8) {
error->ier = I915_READ(GEN8_DE_MISC_IER);
for (i = 0; i < 4; i++)
error->gtier[i] = I915_READ(GEN8_GT_IER(i));
} else if (HAS_PCH_SPLIT(dev)) {
error->ier = I915_READ(DEIER);
error->gtier[0] = I915_READ(GTIER);
} else if (IS_GEN2(dev)) {
error->ier = I915_READ16(IER);
} else if (!IS_VALLEYVIEW(dev)) {
error->ier = I915_READ(IER);
}
error->eir = I915_READ(EIR);
error->pgtbl_er = I915_READ(PGTBL_ER);
i915_get_extra_instdone(dev_priv, error->extra_instdone);
}
static void i915_error_capture_msg(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error,
u32 engine_mask,
const char *error_msg)
{
u32 ecode;
int engine_id = -1, len;
ecode = i915_error_generate_code(dev_priv, error, &engine_id);
len = scnprintf(error->error_msg, sizeof(error->error_msg),
"GPU HANG: ecode %d:%d:0x%08x",
INTEL_GEN(dev_priv), engine_id, ecode);
if (engine_id != -1 && error->engine[engine_id].pid != -1)
len += scnprintf(error->error_msg + len,
sizeof(error->error_msg) - len,
", in %s [%d]",
error->engine[engine_id].comm,
error->engine[engine_id].pid);
scnprintf(error->error_msg + len, sizeof(error->error_msg) - len,
", reason: %s, action: %s",
error_msg,
engine_mask ? "reset" : "continue");
}
static void i915_capture_gen_state(struct drm_i915_private *dev_priv,
struct drm_i915_error_state *error)
{
error->iommu = -1;
#ifdef CONFIG_INTEL_IOMMU
error->iommu = intel_iommu_gfx_mapped;
#endif
error->reset_count = i915_reset_count(&dev_priv->gpu_error);
error->suspend_count = dev_priv->suspend_count;
}
/**
* i915_capture_error_state - capture an error record for later analysis
* @dev: drm device
*
* Should be called when an error is detected (either a hang or an error
* interrupt) to capture error state from the time of the error. Fills
* out a structure which becomes available in debugfs for user level tools
* to pick up.
*/
void i915_capture_error_state(struct drm_i915_private *dev_priv,
u32 engine_mask,
const char *error_msg)
{
static bool warned;
struct drm_i915_error_state *error;
unsigned long flags;
if (READ_ONCE(dev_priv->gpu_error.first_error))
return;
/* Account for pipe specific data like PIPE*STAT */
error = kzalloc(sizeof(*error), GFP_ATOMIC);
if (!error) {
DRM_DEBUG_DRIVER("out of memory, not capturing error state\n");
return;
}
kref_init(&error->ref);
i915_capture_gen_state(dev_priv, error);
i915_capture_reg_state(dev_priv, error);
i915_gem_capture_buffers(dev_priv, error);
i915_gem_record_fences(dev_priv, error);
i915_gem_record_rings(dev_priv, error);
do_gettimeofday(&error->time);
error->overlay = intel_overlay_capture_error_state(dev_priv);
error->display = intel_display_capture_error_state(dev_priv);
i915_error_capture_msg(dev_priv, error, engine_mask, error_msg);
DRM_INFO("%s\n", error->error_msg);
if (!error->simulated) {
spin_lock_irqsave(&dev_priv->gpu_error.lock, flags);
if (!dev_priv->gpu_error.first_error) {
dev_priv->gpu_error.first_error = error;
error = NULL;
}
spin_unlock_irqrestore(&dev_priv->gpu_error.lock, flags);
}
if (error) {
i915_error_state_free(&error->ref);
return;
}
if (!warned) {
DRM_INFO("GPU hangs can indicate a bug anywhere in the entire gfx stack, including userspace.\n");
DRM_INFO("Please file a _new_ bug report on bugs.freedesktop.org against DRI -> DRM/Intel\n");
DRM_INFO("drm/i915 developers can then reassign to the right component if it's not a kernel issue.\n");
DRM_INFO("The gpu crash dump is required to analyze gpu hangs, so please always attach it.\n");
DRM_INFO("GPU crash dump saved to /sys/class/drm/card%d/error\n",
dev_priv->drm.primary->index);
warned = true;
}
}
void i915_error_state_get(struct drm_device *dev,
struct i915_error_state_file_priv *error_priv)
{
struct drm_i915_private *dev_priv = to_i915(dev);
spin_lock_irq(&dev_priv->gpu_error.lock);
error_priv->error = dev_priv->gpu_error.first_error;
if (error_priv->error)
kref_get(&error_priv->error->ref);
spin_unlock_irq(&dev_priv->gpu_error.lock);
}
void i915_error_state_put(struct i915_error_state_file_priv *error_priv)
{
if (error_priv->error)
kref_put(&error_priv->error->ref, i915_error_state_free);
}
void i915_destroy_error_state(struct drm_device *dev)
{
struct drm_i915_private *dev_priv = to_i915(dev);
struct drm_i915_error_state *error;
spin_lock_irq(&dev_priv->gpu_error.lock);
error = dev_priv->gpu_error.first_error;
dev_priv->gpu_error.first_error = NULL;
spin_unlock_irq(&dev_priv->gpu_error.lock);
if (error)
kref_put(&error->ref, i915_error_state_free);
}
const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
{
switch (type) {
case I915_CACHE_NONE: return " uncached";
case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
case I915_CACHE_L3_LLC: return " L3+LLC";
case I915_CACHE_WT: return " WT";
default: return "";
}
}
/* NB: please notice the memset */
void i915_get_extra_instdone(struct drm_i915_private *dev_priv,
uint32_t *instdone)
{
memset(instdone, 0, sizeof(*instdone) * I915_NUM_INSTDONE_REG);
if (IS_GEN2(dev_priv) || IS_GEN3(dev_priv))
instdone[0] = I915_READ(GEN2_INSTDONE);
else if (IS_GEN4(dev_priv) || IS_GEN5(dev_priv) || IS_GEN6(dev_priv)) {
instdone[0] = I915_READ(RING_INSTDONE(RENDER_RING_BASE));
instdone[1] = I915_READ(GEN4_INSTDONE1);
} else if (INTEL_GEN(dev_priv) >= 7) {
instdone[0] = I915_READ(RING_INSTDONE(RENDER_RING_BASE));
instdone[1] = I915_READ(GEN7_SC_INSTDONE);
instdone[2] = I915_READ(GEN7_SAMPLER_INSTDONE);
instdone[3] = I915_READ(GEN7_ROW_INSTDONE);
}
}