2009-06-05 06:42:42 -06:00
|
|
|
/*
|
|
|
|
* Copyright 2008 Jerome Glisse.
|
|
|
|
* All Rights Reserved.
|
|
|
|
*
|
|
|
|
* 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
|
|
|
|
* PRECISION INSIGHT AND/OR ITS SUPPLIERS 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:
|
|
|
|
* Jerome Glisse <glisse@freedesktop.org>
|
|
|
|
*/
|
2014-03-01 16:56:20 -07:00
|
|
|
#include <linux/list_sort.h>
|
2012-10-02 11:01:07 -06:00
|
|
|
#include <drm/drmP.h>
|
|
|
|
#include <drm/radeon_drm.h>
|
2009-06-05 06:42:42 -06:00
|
|
|
#include "radeon_reg.h"
|
|
|
|
#include "radeon.h"
|
2013-09-07 10:29:01 -06:00
|
|
|
#include "radeon_trace.h"
|
2009-06-05 06:42:42 -06:00
|
|
|
|
2014-03-01 16:56:21 -07:00
|
|
|
#define RADEON_CS_MAX_PRIORITY 32u
|
|
|
|
#define RADEON_CS_NUM_BUCKETS (RADEON_CS_MAX_PRIORITY + 1)
|
|
|
|
|
|
|
|
/* This is based on the bucket sort with O(n) time complexity.
|
|
|
|
* An item with priority "i" is added to bucket[i]. The lists are then
|
|
|
|
* concatenated in descending order.
|
|
|
|
*/
|
|
|
|
struct radeon_cs_buckets {
|
|
|
|
struct list_head bucket[RADEON_CS_NUM_BUCKETS];
|
|
|
|
};
|
|
|
|
|
|
|
|
static void radeon_cs_buckets_init(struct radeon_cs_buckets *b)
|
|
|
|
{
|
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++)
|
|
|
|
INIT_LIST_HEAD(&b->bucket[i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void radeon_cs_buckets_add(struct radeon_cs_buckets *b,
|
|
|
|
struct list_head *item, unsigned priority)
|
|
|
|
{
|
|
|
|
/* Since buffers which appear sooner in the relocation list are
|
|
|
|
* likely to be used more often than buffers which appear later
|
|
|
|
* in the list, the sort mustn't change the ordering of buffers
|
|
|
|
* with the same priority, i.e. it must be stable.
|
|
|
|
*/
|
|
|
|
list_add_tail(item, &b->bucket[min(priority, RADEON_CS_MAX_PRIORITY)]);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void radeon_cs_buckets_get_list(struct radeon_cs_buckets *b,
|
|
|
|
struct list_head *out_list)
|
|
|
|
{
|
|
|
|
unsigned i;
|
|
|
|
|
|
|
|
/* Connect the sorted buckets in the output list. */
|
|
|
|
for (i = 0; i < RADEON_CS_NUM_BUCKETS; i++) {
|
|
|
|
list_splice(&b->bucket[i], out_list);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-08-31 11:43:50 -06:00
|
|
|
static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
|
2009-06-05 06:42:42 -06:00
|
|
|
{
|
|
|
|
struct radeon_cs_chunk *chunk;
|
2014-03-01 16:56:21 -07:00
|
|
|
struct radeon_cs_buckets buckets;
|
2014-12-03 07:46:49 -07:00
|
|
|
unsigned i;
|
|
|
|
bool need_mmap_lock = false;
|
drm/radeon: add userptr support v8
This patch adds an IOCTL for turning a pointer supplied by
userspace into a buffer object.
It imposes several restrictions upon the memory being mapped:
1. It must be page aligned (both start/end addresses, i.e ptr and size).
2. It must be normal system memory, not a pointer into another map of IO
space (e.g. it must not be a GTT mmapping of another object).
3. The BO is mapped into GTT, so the maximum amount of memory mapped at
all times is still the GTT limit.
4. The BO is only mapped readonly for now, so no write support.
5. List of backing pages is only acquired once, so they represent a
snapshot of the first use.
Exporting and sharing as well as mapping of buffer objects created by
this function is forbidden and results in an -EPERM.
v2: squash all previous changes into first public version
v3: fix tabs, map readonly, don't use MM callback any more
v4: set TTM_PAGE_FLAG_SG so that TTM never messes with the pages,
pin/unpin pages on bind/unbind instead of populate/unpopulate
v5: rebased on 3.17-wip, IOCTL renamed to userptr, reject any unknown
flags, better handle READONLY flag, improve permission check
v6: fix ptr cast warning, use set_page_dirty/mark_page_accessed on unpin
v7: add warning about it's availability in the API definition
v8: drop access_ok check, fix VM mapping bits
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v4)
Reviewed-by: Jérôme Glisse <jglisse@redhat.com> (v4)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2014-08-07 01:36:00 -06:00
|
|
|
int r;
|
2009-06-05 06:42:42 -06:00
|
|
|
|
2014-12-03 07:53:24 -07:00
|
|
|
if (p->chunk_relocs == NULL) {
|
2009-06-05 06:42:42 -06:00
|
|
|
return 0;
|
|
|
|
}
|
2014-12-03 07:53:24 -07:00
|
|
|
chunk = p->chunk_relocs;
|
2011-11-18 08:19:47 -07:00
|
|
|
p->dma_reloc_idx = 0;
|
2009-06-05 06:42:42 -06:00
|
|
|
/* FIXME: we assume that each relocs use 4 dwords */
|
|
|
|
p->nrelocs = chunk->length_dw / 4;
|
2017-05-17 06:23:12 -06:00
|
|
|
p->relocs = kvmalloc_array(p->nrelocs, sizeof(struct radeon_bo_list),
|
|
|
|
GFP_KERNEL | __GFP_ZERO);
|
2009-06-05 06:42:42 -06:00
|
|
|
if (p->relocs == NULL) {
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
2014-03-01 16:56:21 -07:00
|
|
|
|
|
|
|
radeon_cs_buckets_init(&buckets);
|
|
|
|
|
2009-06-05 06:42:42 -06:00
|
|
|
for (i = 0; i < p->nrelocs; i++) {
|
|
|
|
struct drm_radeon_cs_reloc *r;
|
2014-11-27 06:48:40 -07:00
|
|
|
struct drm_gem_object *gobj;
|
2014-03-01 16:56:21 -07:00
|
|
|
unsigned priority;
|
2009-06-05 06:42:42 -06:00
|
|
|
|
|
|
|
r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
|
2016-05-09 04:04:54 -06:00
|
|
|
gobj = drm_gem_object_lookup(p->filp, r->handle);
|
2014-11-27 06:48:40 -07:00
|
|
|
if (gobj == NULL) {
|
2013-04-08 04:41:28 -06:00
|
|
|
DRM_ERROR("gem object lookup failed 0x%x\n",
|
|
|
|
r->handle);
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
2014-11-27 06:48:40 -07:00
|
|
|
p->relocs[i].robj = gem_to_radeon_bo(gobj);
|
2014-03-01 16:56:21 -07:00
|
|
|
|
|
|
|
/* The userspace buffer priorities are from 0 to 15. A higher
|
|
|
|
* number means the buffer is more important.
|
|
|
|
* Also, the buffers used for write have a higher priority than
|
|
|
|
* the buffers used for read only, which doubles the range
|
|
|
|
* to 0 to 31. 32 is reserved for the kernel driver.
|
|
|
|
*/
|
2014-08-15 03:52:53 -06:00
|
|
|
priority = (r->flags & RADEON_RELOC_PRIO_MASK) * 2
|
|
|
|
+ !!r->write_domain;
|
2013-04-08 04:41:28 -06:00
|
|
|
|
2017-04-11 11:20:20 -06:00
|
|
|
/* The first reloc of an UVD job is the msg and that must be in
|
|
|
|
* VRAM, the second reloc is the DPB and for WMV that must be in
|
|
|
|
* VRAM as well. Also put everything into VRAM on AGP cards and older
|
|
|
|
* IGP chips to avoid image corruptions
|
|
|
|
*/
|
2013-09-15 05:31:28 -06:00
|
|
|
if (p->ring == R600_RING_TYPE_UVD_INDEX &&
|
2017-04-11 11:20:20 -06:00
|
|
|
(i <= 0 || pci_find_capability(p->rdev->ddev->pdev,
|
2017-01-24 23:26:52 -07:00
|
|
|
PCI_CAP_ID_AGP) ||
|
2013-04-16 07:41:25 -06:00
|
|
|
p->rdev->family == CHIP_RS780 ||
|
|
|
|
p->rdev->family == CHIP_RS880)) {
|
|
|
|
|
2013-10-15 12:12:03 -06:00
|
|
|
/* TODO: is this still needed for NI+ ? */
|
2017-08-08 05:50:46 -06:00
|
|
|
p->relocs[i].preferred_domains =
|
2013-04-08 04:41:29 -06:00
|
|
|
RADEON_GEM_DOMAIN_VRAM;
|
|
|
|
|
2014-06-02 09:33:07 -06:00
|
|
|
p->relocs[i].allowed_domains =
|
2013-04-08 04:41:29 -06:00
|
|
|
RADEON_GEM_DOMAIN_VRAM;
|
|
|
|
|
2014-03-01 16:56:21 -07:00
|
|
|
/* prioritize this over any other relocation */
|
|
|
|
priority = RADEON_CS_MAX_PRIORITY;
|
2013-04-08 04:41:29 -06:00
|
|
|
} else {
|
|
|
|
uint32_t domain = r->write_domain ?
|
|
|
|
r->write_domain : r->read_domains;
|
|
|
|
|
2014-05-26 18:56:36 -06:00
|
|
|
if (domain & RADEON_GEM_DOMAIN_CPU) {
|
|
|
|
DRM_ERROR("RADEON_GEM_DOMAIN_CPU is not valid "
|
|
|
|
"for command submission\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2017-08-08 05:50:46 -06:00
|
|
|
p->relocs[i].preferred_domains = domain;
|
2013-04-08 04:41:29 -06:00
|
|
|
if (domain == RADEON_GEM_DOMAIN_VRAM)
|
|
|
|
domain |= RADEON_GEM_DOMAIN_GTT;
|
2014-06-02 09:33:07 -06:00
|
|
|
p->relocs[i].allowed_domains = domain;
|
2013-04-08 04:41:29 -06:00
|
|
|
}
|
2013-04-08 04:41:28 -06:00
|
|
|
|
drm/radeon: add userptr support v8
This patch adds an IOCTL for turning a pointer supplied by
userspace into a buffer object.
It imposes several restrictions upon the memory being mapped:
1. It must be page aligned (both start/end addresses, i.e ptr and size).
2. It must be normal system memory, not a pointer into another map of IO
space (e.g. it must not be a GTT mmapping of another object).
3. The BO is mapped into GTT, so the maximum amount of memory mapped at
all times is still the GTT limit.
4. The BO is only mapped readonly for now, so no write support.
5. List of backing pages is only acquired once, so they represent a
snapshot of the first use.
Exporting and sharing as well as mapping of buffer objects created by
this function is forbidden and results in an -EPERM.
v2: squash all previous changes into first public version
v3: fix tabs, map readonly, don't use MM callback any more
v4: set TTM_PAGE_FLAG_SG so that TTM never messes with the pages,
pin/unpin pages on bind/unbind instead of populate/unpopulate
v5: rebased on 3.17-wip, IOCTL renamed to userptr, reject any unknown
flags, better handle READONLY flag, improve permission check
v6: fix ptr cast warning, use set_page_dirty/mark_page_accessed on unpin
v7: add warning about it's availability in the API definition
v8: drop access_ok check, fix VM mapping bits
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v4)
Reviewed-by: Jérôme Glisse <jglisse@redhat.com> (v4)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2014-08-07 01:36:00 -06:00
|
|
|
if (radeon_ttm_tt_has_userptr(p->relocs[i].robj->tbo.ttm)) {
|
2017-08-08 05:50:46 -06:00
|
|
|
uint32_t domain = p->relocs[i].preferred_domains;
|
drm/radeon: add userptr support v8
This patch adds an IOCTL for turning a pointer supplied by
userspace into a buffer object.
It imposes several restrictions upon the memory being mapped:
1. It must be page aligned (both start/end addresses, i.e ptr and size).
2. It must be normal system memory, not a pointer into another map of IO
space (e.g. it must not be a GTT mmapping of another object).
3. The BO is mapped into GTT, so the maximum amount of memory mapped at
all times is still the GTT limit.
4. The BO is only mapped readonly for now, so no write support.
5. List of backing pages is only acquired once, so they represent a
snapshot of the first use.
Exporting and sharing as well as mapping of buffer objects created by
this function is forbidden and results in an -EPERM.
v2: squash all previous changes into first public version
v3: fix tabs, map readonly, don't use MM callback any more
v4: set TTM_PAGE_FLAG_SG so that TTM never messes with the pages,
pin/unpin pages on bind/unbind instead of populate/unpopulate
v5: rebased on 3.17-wip, IOCTL renamed to userptr, reject any unknown
flags, better handle READONLY flag, improve permission check
v6: fix ptr cast warning, use set_page_dirty/mark_page_accessed on unpin
v7: add warning about it's availability in the API definition
v8: drop access_ok check, fix VM mapping bits
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v4)
Reviewed-by: Jérôme Glisse <jglisse@redhat.com> (v4)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2014-08-07 01:36:00 -06:00
|
|
|
if (!(domain & RADEON_GEM_DOMAIN_GTT)) {
|
|
|
|
DRM_ERROR("Only RADEON_GEM_DOMAIN_GTT is "
|
|
|
|
"allowed for userptr BOs\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
need_mmap_lock = true;
|
|
|
|
domain = RADEON_GEM_DOMAIN_GTT;
|
2017-08-08 05:50:46 -06:00
|
|
|
p->relocs[i].preferred_domains = domain;
|
drm/radeon: add userptr support v8
This patch adds an IOCTL for turning a pointer supplied by
userspace into a buffer object.
It imposes several restrictions upon the memory being mapped:
1. It must be page aligned (both start/end addresses, i.e ptr and size).
2. It must be normal system memory, not a pointer into another map of IO
space (e.g. it must not be a GTT mmapping of another object).
3. The BO is mapped into GTT, so the maximum amount of memory mapped at
all times is still the GTT limit.
4. The BO is only mapped readonly for now, so no write support.
5. List of backing pages is only acquired once, so they represent a
snapshot of the first use.
Exporting and sharing as well as mapping of buffer objects created by
this function is forbidden and results in an -EPERM.
v2: squash all previous changes into first public version
v3: fix tabs, map readonly, don't use MM callback any more
v4: set TTM_PAGE_FLAG_SG so that TTM never messes with the pages,
pin/unpin pages on bind/unbind instead of populate/unpopulate
v5: rebased on 3.17-wip, IOCTL renamed to userptr, reject any unknown
flags, better handle READONLY flag, improve permission check
v6: fix ptr cast warning, use set_page_dirty/mark_page_accessed on unpin
v7: add warning about it's availability in the API definition
v8: drop access_ok check, fix VM mapping bits
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v4)
Reviewed-by: Jérôme Glisse <jglisse@redhat.com> (v4)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2014-08-07 01:36:00 -06:00
|
|
|
p->relocs[i].allowed_domains = domain;
|
|
|
|
}
|
|
|
|
|
2017-04-02 21:35:23 -06:00
|
|
|
/* Objects shared as dma-bufs cannot be moved to VRAM */
|
|
|
|
if (p->relocs[i].robj->prime_shared_count) {
|
|
|
|
p->relocs[i].allowed_domains &= ~RADEON_GEM_DOMAIN_VRAM;
|
|
|
|
if (!p->relocs[i].allowed_domains) {
|
|
|
|
DRM_ERROR("BO associated with dma-buf cannot "
|
|
|
|
"be moved to VRAM\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-03-03 04:38:08 -07:00
|
|
|
p->relocs[i].tv.bo = &p->relocs[i].robj->tbo;
|
2014-09-04 12:01:54 -06:00
|
|
|
p->relocs[i].tv.shared = !r->write_domain;
|
2013-04-08 04:41:28 -06:00
|
|
|
|
2014-03-03 04:38:08 -07:00
|
|
|
radeon_cs_buckets_add(&buckets, &p->relocs[i].tv.head,
|
2014-03-01 16:56:21 -07:00
|
|
|
priority);
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
2014-03-01 16:56:21 -07:00
|
|
|
|
|
|
|
radeon_cs_buckets_get_list(&buckets, &p->validated);
|
|
|
|
|
2014-02-20 05:42:17 -07:00
|
|
|
if (p->cs_flags & RADEON_CS_USE_VM)
|
|
|
|
p->vm_bos = radeon_vm_get_bos(p->rdev, p->ib.vm,
|
|
|
|
&p->validated);
|
drm/radeon: add userptr support v8
This patch adds an IOCTL for turning a pointer supplied by
userspace into a buffer object.
It imposes several restrictions upon the memory being mapped:
1. It must be page aligned (both start/end addresses, i.e ptr and size).
2. It must be normal system memory, not a pointer into another map of IO
space (e.g. it must not be a GTT mmapping of another object).
3. The BO is mapped into GTT, so the maximum amount of memory mapped at
all times is still the GTT limit.
4. The BO is only mapped readonly for now, so no write support.
5. List of backing pages is only acquired once, so they represent a
snapshot of the first use.
Exporting and sharing as well as mapping of buffer objects created by
this function is forbidden and results in an -EPERM.
v2: squash all previous changes into first public version
v3: fix tabs, map readonly, don't use MM callback any more
v4: set TTM_PAGE_FLAG_SG so that TTM never messes with the pages,
pin/unpin pages on bind/unbind instead of populate/unpopulate
v5: rebased on 3.17-wip, IOCTL renamed to userptr, reject any unknown
flags, better handle READONLY flag, improve permission check
v6: fix ptr cast warning, use set_page_dirty/mark_page_accessed on unpin
v7: add warning about it's availability in the API definition
v8: drop access_ok check, fix VM mapping bits
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v4)
Reviewed-by: Jérôme Glisse <jglisse@redhat.com> (v4)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2014-08-07 01:36:00 -06:00
|
|
|
if (need_mmap_lock)
|
|
|
|
down_read(¤t->mm->mmap_sem);
|
|
|
|
|
|
|
|
r = radeon_bo_list_validate(p->rdev, &p->ticket, &p->validated, p->ring);
|
|
|
|
|
|
|
|
if (need_mmap_lock)
|
|
|
|
up_read(¤t->mm->mmap_sem);
|
2014-02-20 05:42:17 -07:00
|
|
|
|
drm/radeon: add userptr support v8
This patch adds an IOCTL for turning a pointer supplied by
userspace into a buffer object.
It imposes several restrictions upon the memory being mapped:
1. It must be page aligned (both start/end addresses, i.e ptr and size).
2. It must be normal system memory, not a pointer into another map of IO
space (e.g. it must not be a GTT mmapping of another object).
3. The BO is mapped into GTT, so the maximum amount of memory mapped at
all times is still the GTT limit.
4. The BO is only mapped readonly for now, so no write support.
5. List of backing pages is only acquired once, so they represent a
snapshot of the first use.
Exporting and sharing as well as mapping of buffer objects created by
this function is forbidden and results in an -EPERM.
v2: squash all previous changes into first public version
v3: fix tabs, map readonly, don't use MM callback any more
v4: set TTM_PAGE_FLAG_SG so that TTM never messes with the pages,
pin/unpin pages on bind/unbind instead of populate/unpopulate
v5: rebased on 3.17-wip, IOCTL renamed to userptr, reject any unknown
flags, better handle READONLY flag, improve permission check
v6: fix ptr cast warning, use set_page_dirty/mark_page_accessed on unpin
v7: add warning about it's availability in the API definition
v8: drop access_ok check, fix VM mapping bits
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v4)
Reviewed-by: Jérôme Glisse <jglisse@redhat.com> (v4)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2014-08-07 01:36:00 -06:00
|
|
|
return r;
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
|
|
|
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
|
|
|
|
{
|
|
|
|
p->priority = priority;
|
|
|
|
|
|
|
|
switch (ring) {
|
|
|
|
default:
|
|
|
|
DRM_ERROR("unknown ring id: %d\n", ring);
|
|
|
|
return -EINVAL;
|
|
|
|
case RADEON_CS_RING_GFX:
|
|
|
|
p->ring = RADEON_RING_TYPE_GFX_INDEX;
|
|
|
|
break;
|
|
|
|
case RADEON_CS_RING_COMPUTE:
|
2013-06-26 15:37:11 -06:00
|
|
|
if (p->rdev->family >= CHIP_TAHITI) {
|
2012-03-20 15:18:24 -06:00
|
|
|
if (p->priority > 0)
|
|
|
|
p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
|
|
|
|
else
|
|
|
|
p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
|
|
|
|
} else
|
|
|
|
p->ring = RADEON_RING_TYPE_GFX_INDEX;
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
break;
|
2012-12-13 10:27:28 -07:00
|
|
|
case RADEON_CS_RING_DMA:
|
|
|
|
if (p->rdev->family >= CHIP_CAYMAN) {
|
|
|
|
if (p->priority > 0)
|
|
|
|
p->ring = R600_RING_TYPE_DMA_INDEX;
|
|
|
|
else
|
|
|
|
p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
|
2014-01-27 08:59:51 -07:00
|
|
|
} else if (p->rdev->family >= CHIP_RV770) {
|
2012-12-13 10:27:28 -07:00
|
|
|
p->ring = R600_RING_TYPE_DMA_INDEX;
|
|
|
|
} else {
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
break;
|
2013-04-08 04:41:29 -06:00
|
|
|
case RADEON_CS_RING_UVD:
|
|
|
|
p->ring = R600_RING_TYPE_UVD_INDEX;
|
|
|
|
break;
|
2013-05-23 04:10:04 -06:00
|
|
|
case RADEON_CS_RING_VCE:
|
|
|
|
/* TODO: only use the low priority ring for now */
|
|
|
|
p->ring = TN_RING_TYPE_VCE1_INDEX;
|
|
|
|
break;
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-09-25 04:39:38 -06:00
|
|
|
static int radeon_cs_sync_rings(struct radeon_cs_parser *p)
|
2012-01-05 20:11:06 -07:00
|
|
|
{
|
2014-11-27 06:48:42 -07:00
|
|
|
struct radeon_bo_list *reloc;
|
2014-11-26 08:29:33 -07:00
|
|
|
int r;
|
2012-01-05 20:11:06 -07:00
|
|
|
|
2014-11-26 08:29:33 -07:00
|
|
|
list_for_each_entry(reloc, &p->validated, tv.head) {
|
2014-04-02 09:14:48 -06:00
|
|
|
struct reservation_object *resv;
|
|
|
|
|
2014-11-26 08:29:33 -07:00
|
|
|
resv = reloc->robj->tbo.resv;
|
2014-11-19 06:01:22 -07:00
|
|
|
r = radeon_sync_resv(p->rdev, &p->ib.sync, resv,
|
2014-11-26 08:29:33 -07:00
|
|
|
reloc->tv.shared);
|
2014-09-25 04:39:38 -06:00
|
|
|
if (r)
|
2014-11-26 08:29:33 -07:00
|
|
|
return r;
|
2012-05-02 07:11:18 -06:00
|
|
|
}
|
2014-11-26 08:29:33 -07:00
|
|
|
return 0;
|
2012-01-05 20:11:06 -07:00
|
|
|
}
|
|
|
|
|
2012-05-30 08:09:30 -06:00
|
|
|
/* XXX: note that this is called from the legacy UMS CS ioctl as well */
|
2009-06-05 06:42:42 -06:00
|
|
|
int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
|
|
|
|
{
|
|
|
|
struct drm_radeon_cs *cs = data;
|
|
|
|
uint64_t *chunk_array_ptr;
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
unsigned size, i;
|
|
|
|
u32 ring = RADEON_CS_RING_GFX;
|
|
|
|
s32 priority = 0;
|
2009-06-05 06:42:42 -06:00
|
|
|
|
2015-03-02 12:36:07 -07:00
|
|
|
INIT_LIST_HEAD(&p->validated);
|
|
|
|
|
2009-06-05 06:42:42 -06:00
|
|
|
if (!cs->num_chunks) {
|
|
|
|
return 0;
|
|
|
|
}
|
2015-03-02 12:36:07 -07:00
|
|
|
|
2009-06-05 06:42:42 -06:00
|
|
|
/* get chunks */
|
|
|
|
p->idx = 0;
|
2012-05-09 07:35:02 -06:00
|
|
|
p->ib.sa_bo = NULL;
|
|
|
|
p->const_ib.sa_bo = NULL;
|
2014-12-03 07:53:24 -07:00
|
|
|
p->chunk_ib = NULL;
|
|
|
|
p->chunk_relocs = NULL;
|
|
|
|
p->chunk_flags = NULL;
|
|
|
|
p->chunk_const_ib = NULL;
|
2009-06-05 06:42:42 -06:00
|
|
|
p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
|
|
|
|
if (p->chunks_array == NULL) {
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
|
2013-12-11 03:34:44 -07:00
|
|
|
if (copy_from_user(p->chunks_array, chunk_array_ptr,
|
2009-06-05 06:42:42 -06:00
|
|
|
sizeof(uint64_t)*cs->num_chunks)) {
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
p->cs_flags = 0;
|
2009-06-05 06:42:42 -06:00
|
|
|
p->nchunks = cs->num_chunks;
|
|
|
|
p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
|
|
|
|
if (p->chunks == NULL) {
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
for (i = 0; i < p->nchunks; i++) {
|
|
|
|
struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
|
|
|
|
struct drm_radeon_cs_chunk user_chunk;
|
|
|
|
uint32_t __user *cdata;
|
|
|
|
|
|
|
|
chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
|
2013-12-11 03:34:44 -07:00
|
|
|
if (copy_from_user(&user_chunk, chunk_ptr,
|
2009-06-05 06:42:42 -06:00
|
|
|
sizeof(struct drm_radeon_cs_chunk))) {
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
2009-06-29 19:47:14 -06:00
|
|
|
p->chunks[i].length_dw = user_chunk.length_dw;
|
2014-12-03 07:53:24 -07:00
|
|
|
if (user_chunk.chunk_id == RADEON_CHUNK_ID_RELOCS) {
|
|
|
|
p->chunk_relocs = &p->chunks[i];
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
2014-12-03 07:53:24 -07:00
|
|
|
if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
|
|
|
|
p->chunk_ib = &p->chunks[i];
|
2009-06-29 19:47:14 -06:00
|
|
|
/* zero length IB isn't useful */
|
|
|
|
if (p->chunks[i].length_dw == 0)
|
|
|
|
return -EINVAL;
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
2014-12-03 07:53:24 -07:00
|
|
|
if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB) {
|
|
|
|
p->chunk_const_ib = &p->chunks[i];
|
2012-03-20 15:18:14 -06:00
|
|
|
/* zero length CONST IB isn't useful */
|
|
|
|
if (p->chunks[i].length_dw == 0)
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
2014-12-03 07:53:24 -07:00
|
|
|
if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
|
|
|
|
p->chunk_flags = &p->chunks[i];
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
/* zero length flags aren't useful */
|
|
|
|
if (p->chunks[i].length_dw == 0)
|
|
|
|
return -EINVAL;
|
2011-10-24 17:38:45 -06:00
|
|
|
}
|
2009-06-29 19:47:14 -06:00
|
|
|
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
size = p->chunks[i].length_dw;
|
|
|
|
cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
|
|
|
|
p->chunks[i].user_ptr = cdata;
|
2014-12-03 07:53:24 -07:00
|
|
|
if (user_chunk.chunk_id == RADEON_CHUNK_ID_CONST_IB)
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
continue;
|
|
|
|
|
2014-12-03 07:53:24 -07:00
|
|
|
if (user_chunk.chunk_id == RADEON_CHUNK_ID_IB) {
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
if (!p->rdev || !(p->rdev->flags & RADEON_IS_AGP))
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
2017-05-17 06:23:12 -06:00
|
|
|
p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
size *= sizeof(uint32_t);
|
|
|
|
if (p->chunks[i].kdata == NULL) {
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
2013-12-11 03:34:44 -07:00
|
|
|
if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
return -EFAULT;
|
|
|
|
}
|
2014-12-03 07:53:24 -07:00
|
|
|
if (user_chunk.chunk_id == RADEON_CHUNK_ID_FLAGS) {
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
p->cs_flags = p->chunks[i].kdata[0];
|
|
|
|
if (p->chunks[i].length_dw > 1)
|
|
|
|
ring = p->chunks[i].kdata[1];
|
|
|
|
if (p->chunks[i].length_dw > 2)
|
|
|
|
priority = (s32)p->chunks[i].kdata[2];
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
|
|
|
}
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
|
2012-05-30 08:09:30 -06:00
|
|
|
/* these are KMS only */
|
|
|
|
if (p->rdev) {
|
|
|
|
if ((p->cs_flags & RADEON_CS_USE_VM) &&
|
|
|
|
!p->rdev->vm_manager.enabled) {
|
|
|
|
DRM_ERROR("VM not active on asic!\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
2012-03-20 15:18:16 -06:00
|
|
|
|
2013-04-08 04:41:27 -06:00
|
|
|
if (radeon_cs_get_ring(p, ring, priority))
|
2012-05-30 08:09:30 -06:00
|
|
|
return -EINVAL;
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
|
2013-04-08 04:41:27 -06:00
|
|
|
/* we only support VM on some SI+ rings */
|
2014-05-21 09:43:59 -06:00
|
|
|
if ((p->cs_flags & RADEON_CS_USE_VM) == 0) {
|
|
|
|
if (p->rdev->asic->ring[p->ring]->cs_parse == NULL) {
|
|
|
|
DRM_ERROR("Ring %d requires VM!\n", p->ring);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if (p->rdev->asic->ring[p->ring]->ib_parse == NULL) {
|
|
|
|
DRM_ERROR("VM not supported on ring %d!\n",
|
|
|
|
p->ring);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
2013-04-08 04:41:27 -06:00
|
|
|
}
|
2012-05-30 08:09:30 -06:00
|
|
|
}
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
|
2009-06-05 06:42:42 -06:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-03-01 16:56:20 -07:00
|
|
|
static int cmp_size_smaller_first(void *priv, struct list_head *a,
|
|
|
|
struct list_head *b)
|
|
|
|
{
|
2014-11-27 06:48:42 -07:00
|
|
|
struct radeon_bo_list *la = list_entry(a, struct radeon_bo_list, tv.head);
|
|
|
|
struct radeon_bo_list *lb = list_entry(b, struct radeon_bo_list, tv.head);
|
2014-03-01 16:56:20 -07:00
|
|
|
|
|
|
|
/* Sort A before B if A is smaller. */
|
2014-03-03 04:38:08 -07:00
|
|
|
return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
|
2014-03-01 16:56:20 -07:00
|
|
|
}
|
|
|
|
|
2009-06-05 06:42:42 -06:00
|
|
|
/**
|
|
|
|
* cs_parser_fini() - clean parser states
|
|
|
|
* @parser: parser structure holding parsing context.
|
|
|
|
* @error: error number
|
|
|
|
*
|
|
|
|
* If error is set than unvalidate buffer, otherwise just free memory
|
|
|
|
* used by parsing context.
|
|
|
|
**/
|
2013-06-27 05:48:17 -06:00
|
|
|
static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bool backoff)
|
2009-06-05 06:42:42 -06:00
|
|
|
{
|
|
|
|
unsigned i;
|
|
|
|
|
2012-08-06 10:32:21 -06:00
|
|
|
if (!error) {
|
2014-03-01 16:56:20 -07:00
|
|
|
/* Sort the buffer list from the smallest to largest buffer,
|
|
|
|
* which affects the order of buffers in the LRU list.
|
|
|
|
* This assures that the smallest buffers are added first
|
|
|
|
* to the LRU list, so they are likely to be later evicted
|
|
|
|
* first, instead of large buffers whose eviction is more
|
|
|
|
* expensive.
|
|
|
|
*
|
|
|
|
* This slightly lowers the number of bytes moved by TTM
|
|
|
|
* per frame under memory pressure.
|
|
|
|
*/
|
|
|
|
list_sort(NULL, &parser->validated, cmp_size_smaller_first);
|
|
|
|
|
2013-06-27 05:48:17 -06:00
|
|
|
ttm_eu_fence_buffer_objects(&parser->ticket,
|
|
|
|
&parser->validated,
|
2014-04-02 09:14:48 -06:00
|
|
|
&parser->ib.fence->base);
|
2013-06-27 05:48:17 -06:00
|
|
|
} else if (backoff) {
|
|
|
|
ttm_eu_backoff_reservation(&parser->ticket,
|
|
|
|
&parser->validated);
|
2012-08-06 10:32:21 -06:00
|
|
|
}
|
2010-11-17 05:38:32 -07:00
|
|
|
|
2010-03-19 01:44:33 -06:00
|
|
|
if (parser->relocs != NULL) {
|
|
|
|
for (i = 0; i < parser->nrelocs; i++) {
|
2014-11-27 06:48:40 -07:00
|
|
|
struct radeon_bo *bo = parser->relocs[i].robj;
|
|
|
|
if (bo == NULL)
|
|
|
|
continue;
|
|
|
|
|
2017-08-03 05:58:35 -06:00
|
|
|
drm_gem_object_put_unlocked(&bo->gem_base);
|
2010-03-19 01:44:33 -06:00
|
|
|
}
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
2009-09-15 09:09:32 -06:00
|
|
|
kfree(parser->track);
|
2017-05-17 06:23:12 -06:00
|
|
|
kvfree(parser->relocs);
|
|
|
|
kvfree(parser->vm_bos);
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
for (i = 0; i < parser->nchunks; i++)
|
2017-05-17 06:23:12 -06:00
|
|
|
kvfree(parser->chunks[i].kdata);
|
2009-06-05 06:42:42 -06:00
|
|
|
kfree(parser->chunks);
|
|
|
|
kfree(parser->chunks_array);
|
|
|
|
radeon_ib_free(parser->rdev, &parser->ib);
|
2012-05-09 07:35:02 -06:00
|
|
|
radeon_ib_free(parser->rdev, &parser->const_ib);
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
|
|
|
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
static int radeon_cs_ib_chunk(struct radeon_device *rdev,
|
|
|
|
struct radeon_cs_parser *parser)
|
|
|
|
{
|
|
|
|
int r;
|
|
|
|
|
2014-12-03 07:53:24 -07:00
|
|
|
if (parser->chunk_ib == NULL)
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (parser->cs_flags & RADEON_CS_USE_VM)
|
|
|
|
return 0;
|
|
|
|
|
2012-02-23 07:18:44 -07:00
|
|
|
r = radeon_cs_parse(rdev, parser->ring, parser);
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
if (r || parser->parser_error) {
|
|
|
|
DRM_ERROR("Invalid command stream !\n");
|
|
|
|
return r;
|
|
|
|
}
|
2013-07-24 10:12:49 -06:00
|
|
|
|
2014-09-25 04:39:38 -06:00
|
|
|
r = radeon_cs_sync_rings(parser);
|
|
|
|
if (r) {
|
|
|
|
if (r != -ERESTARTSYS)
|
|
|
|
DRM_ERROR("Failed to sync rings: %i\n", r);
|
|
|
|
return r;
|
|
|
|
}
|
|
|
|
|
2013-07-24 10:12:49 -06:00
|
|
|
if (parser->ring == R600_RING_TYPE_UVD_INDEX)
|
|
|
|
radeon_uvd_note_usage(rdev);
|
2013-08-23 09:56:26 -06:00
|
|
|
else if ((parser->ring == TN_RING_TYPE_VCE1_INDEX) ||
|
|
|
|
(parser->ring == TN_RING_TYPE_VCE2_INDEX))
|
|
|
|
radeon_vce_note_usage(rdev);
|
2013-07-24 10:12:49 -06:00
|
|
|
|
2014-08-18 02:34:55 -06:00
|
|
|
r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
if (r) {
|
|
|
|
DRM_ERROR("Failed to schedule IB !\n");
|
|
|
|
}
|
2012-07-03 06:05:41 -06:00
|
|
|
return r;
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
}
|
|
|
|
|
2014-02-20 05:42:17 -07:00
|
|
|
static int radeon_bo_vm_update_pte(struct radeon_cs_parser *p,
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
struct radeon_vm *vm)
|
|
|
|
{
|
2014-02-20 05:42:17 -07:00
|
|
|
struct radeon_device *rdev = p->rdev;
|
2014-07-18 00:56:40 -06:00
|
|
|
struct radeon_bo_va *bo_va;
|
2014-02-20 05:42:17 -07:00
|
|
|
int i, r;
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
|
2014-02-20 05:42:17 -07:00
|
|
|
r = radeon_vm_update_page_directory(rdev, vm);
|
|
|
|
if (r)
|
2012-08-13 10:07:33 -06:00
|
|
|
return r;
|
2014-02-20 05:42:17 -07:00
|
|
|
|
2014-07-18 00:56:40 -06:00
|
|
|
r = radeon_vm_clear_freed(rdev, vm);
|
|
|
|
if (r)
|
|
|
|
return r;
|
|
|
|
|
2014-07-18 05:48:10 -06:00
|
|
|
if (vm->ib_bo_va == NULL) {
|
2014-07-18 00:56:40 -06:00
|
|
|
DRM_ERROR("Tmp BO not in VM!\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2014-07-18 05:48:10 -06:00
|
|
|
r = radeon_vm_bo_update(rdev, vm->ib_bo_va,
|
|
|
|
&rdev->ring_tmp_bo.bo->tbo.mem);
|
2014-02-20 05:42:17 -07:00
|
|
|
if (r)
|
|
|
|
return r;
|
|
|
|
|
|
|
|
for (i = 0; i < p->nrelocs; i++) {
|
|
|
|
struct radeon_bo *bo;
|
|
|
|
|
|
|
|
bo = p->relocs[i].robj;
|
2014-07-18 00:56:40 -06:00
|
|
|
bo_va = radeon_vm_bo_find(vm, bo);
|
|
|
|
if (bo_va == NULL) {
|
|
|
|
dev_err(rdev->dev, "bo %p not in vm %p\n", bo, vm);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
r = radeon_vm_bo_update(rdev, bo_va, &bo->tbo.mem);
|
2014-02-20 05:42:17 -07:00
|
|
|
if (r)
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
return r;
|
2014-11-19 06:01:26 -07:00
|
|
|
|
|
|
|
radeon_sync_fence(&p->ib.sync, bo_va->last_pt_update);
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
}
|
2014-07-18 01:24:53 -06:00
|
|
|
|
|
|
|
return radeon_vm_clear_invalids(rdev, vm);
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
|
|
|
|
struct radeon_cs_parser *parser)
|
|
|
|
{
|
|
|
|
struct radeon_fpriv *fpriv = parser->filp->driver_priv;
|
|
|
|
struct radeon_vm *vm = &fpriv->vm;
|
|
|
|
int r;
|
|
|
|
|
2014-12-03 07:53:24 -07:00
|
|
|
if (parser->chunk_ib == NULL)
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
return 0;
|
|
|
|
if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
|
|
|
|
return 0;
|
|
|
|
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
if (parser->const_ib.length_dw) {
|
2012-05-09 07:35:02 -06:00
|
|
|
r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
|
2012-03-20 15:18:14 -06:00
|
|
|
if (r) {
|
|
|
|
return r;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-05-09 07:35:02 -06:00
|
|
|
r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
if (r) {
|
|
|
|
return r;
|
|
|
|
}
|
|
|
|
|
2013-07-24 10:12:49 -06:00
|
|
|
if (parser->ring == R600_RING_TYPE_UVD_INDEX)
|
|
|
|
radeon_uvd_note_usage(rdev);
|
|
|
|
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
mutex_lock(&vm->mutex);
|
|
|
|
r = radeon_bo_vm_update_pte(parser, vm);
|
|
|
|
if (r) {
|
|
|
|
goto out;
|
|
|
|
}
|
2014-09-25 04:39:38 -06:00
|
|
|
|
|
|
|
r = radeon_cs_sync_rings(parser);
|
|
|
|
if (r) {
|
|
|
|
if (r != -ERESTARTSYS)
|
|
|
|
DRM_ERROR("Failed to sync rings: %i\n", r);
|
|
|
|
goto out;
|
|
|
|
}
|
2012-07-13 05:06:00 -06:00
|
|
|
|
2012-03-20 15:18:14 -06:00
|
|
|
if ((rdev->family >= CHIP_TAHITI) &&
|
2014-12-03 07:53:24 -07:00
|
|
|
(parser->chunk_const_ib != NULL)) {
|
2014-08-18 02:34:55 -06:00
|
|
|
r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib, true);
|
2012-07-13 05:06:00 -06:00
|
|
|
} else {
|
2014-08-18 02:34:55 -06:00
|
|
|
r = radeon_ib_schedule(rdev, &parser->ib, NULL, true);
|
2012-03-20 15:18:14 -06:00
|
|
|
}
|
|
|
|
|
2012-08-09 08:21:08 -06:00
|
|
|
out:
|
2012-05-09 02:07:08 -06:00
|
|
|
mutex_unlock(&vm->mutex);
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
return r;
|
|
|
|
}
|
|
|
|
|
2012-05-02 07:11:19 -06:00
|
|
|
static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
|
|
|
|
{
|
|
|
|
if (r == -EDEADLK) {
|
|
|
|
r = radeon_gpu_reset(rdev);
|
|
|
|
if (!r)
|
|
|
|
r = -EAGAIN;
|
|
|
|
}
|
|
|
|
return r;
|
|
|
|
}
|
|
|
|
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
static int radeon_cs_ib_fill(struct radeon_device *rdev, struct radeon_cs_parser *parser)
|
|
|
|
{
|
|
|
|
struct radeon_cs_chunk *ib_chunk;
|
|
|
|
struct radeon_vm *vm = NULL;
|
|
|
|
int r;
|
|
|
|
|
2014-12-03 07:53:24 -07:00
|
|
|
if (parser->chunk_ib == NULL)
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
if (parser->cs_flags & RADEON_CS_USE_VM) {
|
|
|
|
struct radeon_fpriv *fpriv = parser->filp->driver_priv;
|
|
|
|
vm = &fpriv->vm;
|
|
|
|
|
|
|
|
if ((rdev->family >= CHIP_TAHITI) &&
|
2014-12-03 07:53:24 -07:00
|
|
|
(parser->chunk_const_ib != NULL)) {
|
|
|
|
ib_chunk = parser->chunk_const_ib;
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
|
|
|
|
DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
r = radeon_ib_get(rdev, parser->ring, &parser->const_ib,
|
|
|
|
vm, ib_chunk->length_dw * 4);
|
|
|
|
if (r) {
|
|
|
|
DRM_ERROR("Failed to get const ib !\n");
|
|
|
|
return r;
|
|
|
|
}
|
|
|
|
parser->const_ib.is_const_ib = true;
|
|
|
|
parser->const_ib.length_dw = ib_chunk->length_dw;
|
2013-12-11 03:34:44 -07:00
|
|
|
if (copy_from_user(parser->const_ib.ptr,
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
ib_chunk->user_ptr,
|
|
|
|
ib_chunk->length_dw * 4))
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
|
2014-12-03 07:53:24 -07:00
|
|
|
ib_chunk = parser->chunk_ib;
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
|
|
|
|
DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
}
|
2014-12-03 07:53:24 -07:00
|
|
|
ib_chunk = parser->chunk_ib;
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
|
|
|
|
r = radeon_ib_get(rdev, parser->ring, &parser->ib,
|
|
|
|
vm, ib_chunk->length_dw * 4);
|
|
|
|
if (r) {
|
|
|
|
DRM_ERROR("Failed to get ib !\n");
|
|
|
|
return r;
|
|
|
|
}
|
|
|
|
parser->ib.length_dw = ib_chunk->length_dw;
|
|
|
|
if (ib_chunk->kdata)
|
|
|
|
memcpy(parser->ib.ptr, ib_chunk->kdata, ib_chunk->length_dw * 4);
|
2013-12-11 03:34:44 -07:00
|
|
|
else if (copy_from_user(parser->ib.ptr, ib_chunk->user_ptr, ib_chunk->length_dw * 4))
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
return -EFAULT;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2009-06-05 06:42:42 -06:00
|
|
|
int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
|
|
|
|
{
|
|
|
|
struct radeon_device *rdev = dev->dev_private;
|
|
|
|
struct radeon_cs_parser parser;
|
|
|
|
int r;
|
|
|
|
|
2012-07-02 10:45:19 -06:00
|
|
|
down_read(&rdev->exclusive_lock);
|
2012-02-20 15:57:20 -07:00
|
|
|
if (!rdev->accel_working) {
|
2012-07-02 10:45:19 -06:00
|
|
|
up_read(&rdev->exclusive_lock);
|
2012-02-20 15:57:20 -07:00
|
|
|
return -EBUSY;
|
|
|
|
}
|
2014-08-27 14:45:18 -06:00
|
|
|
if (rdev->in_reset) {
|
|
|
|
up_read(&rdev->exclusive_lock);
|
|
|
|
r = radeon_gpu_reset(rdev);
|
|
|
|
if (!r)
|
|
|
|
r = -EAGAIN;
|
|
|
|
return r;
|
|
|
|
}
|
2009-06-05 06:42:42 -06:00
|
|
|
/* initialize parser */
|
|
|
|
memset(&parser, 0, sizeof(struct radeon_cs_parser));
|
|
|
|
parser.filp = filp;
|
|
|
|
parser.rdev = rdev;
|
2010-01-18 05:01:36 -07:00
|
|
|
parser.dev = rdev->dev;
|
2011-06-08 03:58:29 -06:00
|
|
|
parser.family = rdev->family;
|
2009-06-05 06:42:42 -06:00
|
|
|
r = radeon_cs_parser_init(&parser, data);
|
|
|
|
if (r) {
|
|
|
|
DRM_ERROR("Failed to initialize parser !\n");
|
2013-06-27 05:48:17 -06:00
|
|
|
radeon_cs_parser_fini(&parser, r, false);
|
2012-07-02 10:45:19 -06:00
|
|
|
up_read(&rdev->exclusive_lock);
|
2012-05-02 07:11:19 -06:00
|
|
|
r = radeon_cs_handle_lockup(rdev, r);
|
2009-06-05 06:42:42 -06:00
|
|
|
return r;
|
|
|
|
}
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
|
|
|
|
r = radeon_cs_ib_fill(rdev, &parser);
|
|
|
|
if (!r) {
|
|
|
|
r = radeon_cs_parser_relocs(&parser);
|
|
|
|
if (r && r != -ERESTARTSYS)
|
2010-03-18 18:33:44 -06:00
|
|
|
DRM_ERROR("Failed to parse relocation %d!\n", r);
|
drm/radeon: fixup locking inversion between, mmap_sem and reservations
op 08-10-13 18:58, Thomas Hellstrom schreef:
> On 10/08/2013 06:47 PM, Jerome Glisse wrote:
>> On Tue, Oct 08, 2013 at 06:29:35PM +0200, Thomas Hellstrom wrote:
>>> On 10/08/2013 04:55 PM, Jerome Glisse wrote:
>>>> On Tue, Oct 08, 2013 at 04:45:18PM +0200, Christian König wrote:
>>>>> Am 08.10.2013 16:33, schrieb Jerome Glisse:
>>>>>> On Tue, Oct 08, 2013 at 04:14:40PM +0200, Maarten Lankhorst wrote:
>>>>>>> Allocate and copy all kernel memory before doing reservations. This prevents a locking
>>>>>>> inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
>>>>>>> in ttm_bo_vm_fault without upsetting lockdep.
>>>>>>>
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> I would say NAK. Current code only allocate temporary page in AGP case.
>>>>>> So AGP case is userspace -> temp page -> cs checker -> radeon ib.
>>>>>>
>>>>>> Non AGP is directly memcpy to radeon IB.
>>>>>>
>>>>>> Your patch allocate memory memcpy userspace to it and it will then be
>>>>>> memcpy to IB. Which means you introduce an extra memcpy in the process
>>>>>> not something we want.
>>>>> Totally agree. Additional to that there is no good reason to provide
>>>>> anything else than anonymous system memory to the CS ioctl, so the
>>>>> dependency between the mmap_sem and reservations are not really
>>>>> clear to me.
>>>>>
>>>>> Christian.
>>>> I think is that in other code path you take mmap_sem first then reserve
>>>> bo. But here we reserve bo and then we take mmap_sem because of copy
>>> >from user.
>>>> Cheers,
>>>> Jerome
>>>>
>>> Actually the log message is a little confusing. I think the mmap_sem
>>> locking inversion problem is orthogonal to what's being fixed here.
>>>
>>> This patch fixes the possible recursive bo::reserve caused by
>>> malicious user-space handing a pointer to ttm memory so that the ttm
>>> fault handler is called when bos are already reserved. That may
>>> cause a (possibly interruptible) livelock.
>>>
>>> Once that is fixed, we are free to choose the mmap_sem ->
>>> bo::reserve locking order. Currently it's bo::reserve->mmap_sem(),
>>> but the hack required in the ttm fault handler is admittedly a bit
>>> ugly. The plan is to change the locking order to
>>> mmap_sem->bo::reserve
>>>
>>> I'm not sure if it applies to this particular case, but it should be
>>> possible to make sure that copy_from_user_inatomic() will always
>>> succeed, by making sure the pages are present using
>>> get_user_pages(), and release the pages after
>>> copy_from_user_inatomic() is done. That way there's no need for a
>>> double memcpy slowpath, but if the copied data is very fragmented I
>>> guess the resulting code may look ugly. The get_user_pages()
>>> function will return an error if it hits TTM pages.
>>>
>>> /Thomas
>> get_user_pages + copy_from_user_inatomic is overkill. We should just
>> do get_user_pages which fails with ttm memory and then use copy_highpage
>> helper.
>>
>> Cheers,
>> Jerome
> Yeah, it may well be that that's the preferred solution.
>
> /Thomas
>
I still disagree, and shuffled radeon_ib_get around to be called sooner.
How does the patch below look?
8<-------
Allocate and copy all kernel memory before doing reservations. This prevents a locking
inversion between mmap_sem and reservation_class, and allows us to drop the trylocking
in ttm_bo_vm_fault without upsetting lockdep.
Changes since v1:
- Kill extra memcpy for !AGP case.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
2013-10-09 06:36:57 -06:00
|
|
|
}
|
|
|
|
|
|
|
|
if (r) {
|
2013-06-27 05:48:17 -06:00
|
|
|
radeon_cs_parser_fini(&parser, r, false);
|
2012-07-02 10:45:19 -06:00
|
|
|
up_read(&rdev->exclusive_lock);
|
2012-05-02 07:11:19 -06:00
|
|
|
r = radeon_cs_handle_lockup(rdev, r);
|
2009-06-05 06:42:42 -06:00
|
|
|
return r;
|
|
|
|
}
|
2013-04-18 07:25:59 -06:00
|
|
|
|
2013-09-07 10:29:01 -06:00
|
|
|
trace_radeon_cs(&parser);
|
|
|
|
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
r = radeon_cs_ib_chunk(rdev, &parser);
|
2009-06-05 06:42:42 -06:00
|
|
|
if (r) {
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
goto out;
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
r = radeon_cs_ib_vm_chunk(rdev, &parser);
|
2009-06-05 06:42:42 -06:00
|
|
|
if (r) {
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
goto out;
|
2009-06-05 06:42:42 -06:00
|
|
|
}
|
drm/radeon: GPU virtual memory support v22
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
2012-01-05 20:11:05 -07:00
|
|
|
out:
|
2013-06-27 05:48:17 -06:00
|
|
|
radeon_cs_parser_fini(&parser, r, true);
|
2012-07-02 10:45:19 -06:00
|
|
|
up_read(&rdev->exclusive_lock);
|
2012-05-02 07:11:19 -06:00
|
|
|
r = radeon_cs_handle_lockup(rdev, r);
|
2009-06-05 06:42:42 -06:00
|
|
|
return r;
|
|
|
|
}
|
2009-09-23 00:56:27 -06:00
|
|
|
|
2013-01-02 16:27:40 -07:00
|
|
|
/**
|
|
|
|
* radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
|
|
|
|
* @parser: parser structure holding parsing context.
|
|
|
|
* @pkt: where to store packet information
|
|
|
|
*
|
|
|
|
* Assume that chunk_ib_index is properly set. Will return -EINVAL
|
|
|
|
* if packet is bigger than remaining ib size. or if packets is unknown.
|
|
|
|
**/
|
|
|
|
int radeon_cs_packet_parse(struct radeon_cs_parser *p,
|
|
|
|
struct radeon_cs_packet *pkt,
|
|
|
|
unsigned idx)
|
|
|
|
{
|
2014-12-03 07:53:24 -07:00
|
|
|
struct radeon_cs_chunk *ib_chunk = p->chunk_ib;
|
2013-01-02 16:27:40 -07:00
|
|
|
struct radeon_device *rdev = p->rdev;
|
|
|
|
uint32_t header;
|
2015-02-18 08:15:10 -07:00
|
|
|
int ret = 0, i;
|
2013-01-02 16:27:40 -07:00
|
|
|
|
|
|
|
if (idx >= ib_chunk->length_dw) {
|
|
|
|
DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
|
|
|
|
idx, ib_chunk->length_dw);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
header = radeon_get_ib_value(p, idx);
|
|
|
|
pkt->idx = idx;
|
|
|
|
pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
|
|
|
|
pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
|
|
|
|
pkt->one_reg_wr = 0;
|
|
|
|
switch (pkt->type) {
|
|
|
|
case RADEON_PACKET_TYPE0:
|
|
|
|
if (rdev->family < CHIP_R600) {
|
|
|
|
pkt->reg = R100_CP_PACKET0_GET_REG(header);
|
|
|
|
pkt->one_reg_wr =
|
|
|
|
RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
|
|
|
|
} else
|
|
|
|
pkt->reg = R600_CP_PACKET0_GET_REG(header);
|
|
|
|
break;
|
|
|
|
case RADEON_PACKET_TYPE3:
|
|
|
|
pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
|
|
|
|
break;
|
|
|
|
case RADEON_PACKET_TYPE2:
|
|
|
|
pkt->count = -1;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
|
2015-02-18 08:15:10 -07:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto dump_ib;
|
2013-01-02 16:27:40 -07:00
|
|
|
}
|
|
|
|
if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
|
|
|
|
DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
|
|
|
|
pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
|
2015-02-18 08:15:10 -07:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto dump_ib;
|
2013-01-02 16:27:40 -07:00
|
|
|
}
|
|
|
|
return 0;
|
2015-02-18 08:15:10 -07:00
|
|
|
|
|
|
|
dump_ib:
|
|
|
|
for (i = 0; i < ib_chunk->length_dw; i++) {
|
|
|
|
if (i == idx)
|
|
|
|
printk("\t0x%08x <---\n", radeon_get_ib_value(p, i));
|
|
|
|
else
|
|
|
|
printk("\t0x%08x\n", radeon_get_ib_value(p, i));
|
|
|
|
}
|
|
|
|
return ret;
|
2013-01-02 16:27:40 -07:00
|
|
|
}
|
2013-01-02 16:27:42 -07:00
|
|
|
|
|
|
|
/**
|
|
|
|
* radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
|
|
|
|
* @p: structure holding the parser context.
|
|
|
|
*
|
|
|
|
* Check if the next packet is NOP relocation packet3.
|
|
|
|
**/
|
|
|
|
bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
|
|
|
|
{
|
|
|
|
struct radeon_cs_packet p3reloc;
|
|
|
|
int r;
|
|
|
|
|
|
|
|
r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
|
|
|
|
if (r)
|
|
|
|
return false;
|
|
|
|
if (p3reloc.type != RADEON_PACKET_TYPE3)
|
|
|
|
return false;
|
|
|
|
if (p3reloc.opcode != RADEON_PACKET3_NOP)
|
|
|
|
return false;
|
|
|
|
return true;
|
|
|
|
}
|
2013-01-02 16:27:45 -07:00
|
|
|
|
|
|
|
/**
|
|
|
|
* radeon_cs_dump_packet() - dump raw packet context
|
|
|
|
* @p: structure holding the parser context.
|
|
|
|
* @pkt: structure holding the packet.
|
|
|
|
*
|
|
|
|
* Used mostly for debugging and error reporting.
|
|
|
|
**/
|
|
|
|
void radeon_cs_dump_packet(struct radeon_cs_parser *p,
|
|
|
|
struct radeon_cs_packet *pkt)
|
|
|
|
{
|
|
|
|
volatile uint32_t *ib;
|
|
|
|
unsigned i;
|
|
|
|
unsigned idx;
|
|
|
|
|
|
|
|
ib = p->ib.ptr;
|
|
|
|
idx = pkt->idx;
|
|
|
|
for (i = 0; i <= (pkt->count + 1); i++, idx++)
|
|
|
|
DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
|
|
|
|
}
|
|
|
|
|
2013-01-02 16:27:46 -07:00
|
|
|
/**
|
|
|
|
* radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
|
|
|
|
* @parser: parser structure holding parsing context.
|
|
|
|
* @data: pointer to relocation data
|
|
|
|
* @offset_start: starting offset
|
|
|
|
* @offset_mask: offset mask (to align start offset on)
|
|
|
|
* @reloc: reloc informations
|
|
|
|
*
|
|
|
|
* Check if next packet is relocation packet3, do bo validation and compute
|
|
|
|
* GPU offset using the provided start.
|
|
|
|
**/
|
|
|
|
int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
|
2014-11-27 06:48:42 -07:00
|
|
|
struct radeon_bo_list **cs_reloc,
|
2013-01-02 16:27:46 -07:00
|
|
|
int nomm)
|
|
|
|
{
|
|
|
|
struct radeon_cs_chunk *relocs_chunk;
|
|
|
|
struct radeon_cs_packet p3reloc;
|
|
|
|
unsigned idx;
|
|
|
|
int r;
|
|
|
|
|
2014-12-03 07:53:24 -07:00
|
|
|
if (p->chunk_relocs == NULL) {
|
2013-01-02 16:27:46 -07:00
|
|
|
DRM_ERROR("No relocation chunk !\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
*cs_reloc = NULL;
|
2014-12-03 07:53:24 -07:00
|
|
|
relocs_chunk = p->chunk_relocs;
|
2013-01-02 16:27:46 -07:00
|
|
|
r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
|
|
|
|
if (r)
|
|
|
|
return r;
|
|
|
|
p->idx += p3reloc.count + 2;
|
|
|
|
if (p3reloc.type != RADEON_PACKET_TYPE3 ||
|
|
|
|
p3reloc.opcode != RADEON_PACKET3_NOP) {
|
|
|
|
DRM_ERROR("No packet3 for relocation for packet at %d.\n",
|
|
|
|
p3reloc.idx);
|
|
|
|
radeon_cs_dump_packet(p, &p3reloc);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
idx = radeon_get_ib_value(p, p3reloc.idx + 1);
|
|
|
|
if (idx >= relocs_chunk->length_dw) {
|
|
|
|
DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
|
|
|
|
idx, relocs_chunk->length_dw);
|
|
|
|
radeon_cs_dump_packet(p, &p3reloc);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
/* FIXME: we assume reloc size is 4 dwords */
|
|
|
|
if (nomm) {
|
|
|
|
*cs_reloc = p->relocs;
|
2014-03-03 04:38:08 -07:00
|
|
|
(*cs_reloc)->gpu_offset =
|
2013-01-02 16:27:46 -07:00
|
|
|
(u64)relocs_chunk->kdata[idx + 3] << 32;
|
2014-03-03 04:38:08 -07:00
|
|
|
(*cs_reloc)->gpu_offset |= relocs_chunk->kdata[idx + 0];
|
2013-01-02 16:27:46 -07:00
|
|
|
} else
|
2014-12-03 07:46:49 -07:00
|
|
|
*cs_reloc = &p->relocs[(idx / 4)];
|
2013-01-02 16:27:46 -07:00
|
|
|
return 0;
|
|
|
|
}
|