bably can merge this one by itself ahead of the rest of the series.
With that:
Reviewed-by: Matthew Brost
> ---
> drivers/gpu/drm/ttm/ttm_bo_util.c | 13 +++--
> 1 file changed, 7 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c
> b/
On Mon, Jun 23, 2025 at 01:37:35PM +0100, Matthew Auld wrote:
> +Matt B who is adding clear-on-free support in xe. I'm not sure if we might
> also see something like this.
>
Thanks for the heads up.
> On 23/06/2025 06:52, Arunpravin Paneer Selvam wrote:
> > - Added a handler in DRM buddy manager
On Mon, Jun 16, 2025 at 03:07:22PM +0200, Christian König wrote:
> The TTM refcount is about to be removed.
>
> Signed-off-by: Christian König
> ---
> drivers/gpu/drm/xe/xe_bo.c | 2 --
> 1 file changed, 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> i
On Mon, Jun 16, 2025 at 12:42:53PM +0200, Simona Vetter wrote:
> On Mon, Jun 16, 2025 at 09:24:38AM +0200, Christian König wrote:
> > On 6/13/25 21:11, Matthew Brost wrote:
> > > On Fri, Jun 13, 2025 at 07:26:22PM +0200, Christian König wrote:
> > >> On 6/13
On Wed, Jun 18, 2025 at 05:26:48PM +0200, Thomas Hellström wrote:
> Hi, Matt,
>
> On Wed, 2025-06-18 at 08:26 -0700, Matthew Brost wrote:
> > On Wed, Jun 18, 2025 at 04:21:22PM +0530, Ghimiray, Himal Prasad
> > wrote:
> > >
> > >
> > > On 18-06-
On Wed, Jun 18, 2025 at 04:21:22PM +0530, Ghimiray, Himal Prasad wrote:
>
>
> On 18-06-2025 14:56, Thomas Hellström wrote:
> > From: Matthew Brost
> >
> > The migration functionality and track-keeping of per-pagemap VRAM
> > mapped to the CPU mm is not p
On Tue, Jun 17, 2025 at 09:37:28PM +0200, Thomas Hellström wrote:
> On Tue, 2025-06-17 at 10:04 -0700, Matthew Brost wrote:
> > On Tue, Jun 17, 2025 at 04:55:26PM +0200, Thomas Hellström wrote:
> > > On Tue, 2025-06-17 at 20:17 +0530, Ghimiray, Himal Prasad wrote:
> > >
*/
> if (err) {
> if (migrate_try_count || !ctx.devmem_only) {
> @@ -1054,6 +1047,29 @@ int xe_svm_range_get_pages(struct xe_vm *vm, struct
> xe_svm_range *range,
>
> #if IS_ENABLED(CONFIG_DRM_XE_PAGEMAP)
>
> +/**
> + * xe_svm_alloc
off-by: Thomas Hellström
> Reviewed-by: Matthew Brost #v1
Review is still valid on this rev.
Matt
> ---
> drivers/gpu/drm/drm_gpusvm.c | 7 ++--
> drivers/gpu/drm/drm_pagemap.c | 67 ---
> include/drm/drm_pagemap.h | 34 ++
&
> > >
> > > >
> > > > On 13-06-2025 19:32, Thomas Hellström wrote:
> > > > > From: Matthew Brost
> > > > >
> > > > > The migration functionality and track-keeping of per-pagemap
> > > > > VRAM
> >
On Mon, Jun 16, 2025 at 09:24:38AM +0200, Christian König wrote:
> On 6/13/25 21:11, Matthew Brost wrote:
> > On Fri, Jun 13, 2025 at 07:26:22PM +0200, Christian König wrote:
> >> On 6/13/25 19:01, Matthew Brost wrote:
> >>> All,
> >>>
> >>> A
ue to scheduler.")
Cc: sta...@vger.kernel.org
Signed-off-by: Matthew Brost
---
include/drm/spsc_queue.h | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/include/drm/spsc_queue.h b/include/drm/spsc_queue.h
index 125f096c88cb..ee9df8cc67b7 100644
--- a/include/drm/spsc_q
On Fri, Jun 13, 2025 at 07:26:22PM +0200, Christian König wrote:
> On 6/13/25 19:01, Matthew Brost wrote:
> > All,
> >
> > After about six hours of debugging, I found an issue in a fairly
> > aggressive test case involving the DRM scheduler function
> > drm_sche
All,
After about six hours of debugging, I found an issue in a fairly
aggressive test case involving the DRM scheduler function
drm_sched_entity_push_job. The problem is that spsc_queue_push does not
correctly return first on a job push, causing the queue to fail to run
even though it is ready.
I
USVM
> config. (Matt B)
> - Always make .read_only match xe_vma_read_only() for the ctx. (Dafna)
>
> Signed-off-by: Matthew Auld
> Cc: Himal Prasad Ghimiray
> Cc: Thomas Hellström
> Cc: Dafna Hirschfeld
> Cc: Matthew Brost
Reviewed-by: Matthew Brost
> ---
t; runtime due to something not being initialised, even though it passed
> the build. Should be no functional change here.
>
> Signed-off-by: Matthew Auld
> Cc: Thomas Hellström
> Cc: Matthew Brost
> ---
> drivers/gpu/drm/xe/Makefile | 1 +
> drivers/gpu/drm/xe/xe
On Wed, Jun 04, 2025 at 11:35:34AM +0200, Thomas Hellström wrote:
> From: Matthew Brost
>
> The migration functionality and track-keeping of per-pagemap VRAM
> mapped to the CPU mm is not per GPU_vm, but rather per pagemap.
> This is also reflected by the functions not needing
On Wed, Jun 04, 2025 at 11:35:36AM +0200, Thomas Hellström wrote:
> Add runtime PM since we might call populate_mm on a foreign device.
> Also create the VRAM bos as ttm_bo_type_kernel. This avoids the
> initial clearing and the creation of an mmap handle.
>
I didn't read this part - skipping the
arly always available without storing
> > an extra copy for each tt in the system.
> >
> > Just noticed this while reading over the xe shrinker code.
> >
Reviewed-by: Matthew Brost
I'm not seeing a CI run for this on patchworks. Can you resend intel-xe?
You should
On Wed, Jun 04, 2025 at 11:35:35AM +0200, Thomas Hellström wrote:
> Add an operation to populate a part of a drm_mm with device
> private memory.
>
With the kernel doc fixed:
Reviewed-by: Matthew Brost
> Signed-off-by: Thomas Hellström
> ---
> drivers/gpu/drm/
On Wed, Jun 04, 2025 at 06:53:44PM +0200, Danilo Krummrich wrote:
> On Wed, Jun 04, 2025 at 09:45:00AM -0700, Matthew Brost wrote:
> > On Wed, Jun 04, 2025 at 05:07:15PM +0200, Simona Vetter wrote:
> > > We should definitely document this trick better though, I didn't fin
On Wed, Jun 04, 2025 at 05:07:15PM +0200, Simona Vetter wrote:
> On Wed, Jun 04, 2025 at 11:41:25AM +0200, Christian König wrote:
> > On 6/4/25 10:16, Philipp Stanner wrote:
> > > struct drm_sched_init_args provides the possibility of letting the
> > > scheduler use user-controlled workqueues, inst
On Wed, Jun 04, 2025 at 11:35:36AM +0200, Thomas Hellström wrote:
> Add runtime PM since we might call populate_mm on a foreign device.
I think taking a runtime PM will fix hard to hit splat [1] too.
[1] https://patchwork.freedesktop.org/patch/648954/?series=147849&rev=1
> Also create the VRAM b
On Wed, May 28, 2025 at 10:55:03PM -0600, Upadhyay, Tejas wrote:
>
>
> > -Original Message-
> > From: Intel-xe On Behalf Of Lucas
> > De Marchi
> > Sent: 29 May 2025 03:12
> > To: intel...@lists.freedesktop.org
> > Cc: Brost, Matthew ; dri-
> > de...@lists.freedesktop.org; De Marchi, Luc
the drm client_id in drm_sched_fence")
> Cc: Christian König
> Cc: Pierre-Eric Pelloux-Prayer
> Cc: Philipp Stanner
> Signed-off-by: Lucas De Marchi
Reviewed-by: Matthew Brost
> ---
> drivers/gpu/drm/xe/xe_sched_job.c | 2 +-
> 1 file changed, 1 insertion(+), 1 d
-by: Tvrtko Ursulin
This makes sense in the context of the series (e.g. assuming patch #9 lands).
With that:
Reviewed-by: Matthew Brost
> ---
> drivers/gpu/drm/xe/xe_guc_exec_queue_types.h | 2 ++
> drivers/gpu/drm/xe/xe_guc_submit.c | 7 ++-
> drivers/gpu/drm/xe
t from this series, to make the series completely about the
> new scheduling policy, not general other improvements.
>
>
> P.
>
> >
> > Signed-off-by: Tvrtko Ursulin
> > Cc: Christian König
> > Cc: Danilo Krummrich
> > Cc: Matthew Brost
> > Cc: Philipp Stann
On Fri, May 09, 2025 at 04:33:40PM +0100, Tvrtko Ursulin wrote:
> Replace open-coded helper with the subsystem one.
>
You probably can just send this one by itself as it good cleanup and
independent.
Reviewed-by: Matthew Brost
> Signed-off-by: Tvrtko Ursulin
> ---
> drivers/g
t. It's a large series so I will put most people on Cc only in
> the cover letter as a ping of a sort. Whoever is interested can for now find
> the
> series in the archives.
>
> 1)
> https://lore.kernel.org/dri-devel/20231024160727.282960-1-tvrtko.ursu...@linux.intel.com/
>
On Mon, May 05, 2025 at 07:41:09PM -0700, Matthew Brost wrote:
> On Sat, May 03, 2025 at 05:59:52PM -0300, Maíra Canal wrote:
> > When the DRM scheduler times out, it's possible that the GPU isn't hung;
> > instead, a job may still be running, and there may be no vali
On Sat, May 03, 2025 at 05:59:58PM -0300, Maíra Canal wrote:
> Xe can skip the reset if TDR has fired before the free job worker. Instead
> of using the scheduler internals to add the job to the pending list, use
> the DRM_GPU_SCHED_STAT_RUNNING status to skip the reset and rearm the
> timer.
>
>
e that the job
> should be reinserted into the pending list, and the driver will still
> signal its completion.
>
> Signed-off-by: Maíra Canal
Reviewed-by: Matthew Brost
> ---
> drivers/gpu/drm/scheduler/sched_main.c | 14 ++
> include/drm/gpu_scheduler.h
On Fri, May 02, 2025 at 03:01:01PM +0200, Thomas Hellström wrote:
> The docs were not properly updated from an earlier version of the code.
>
> Fixes: e7b5d23e5d47 ("drm/ttm: Provide a shmem backup implementation")
> Cc: Christian König
> Cc: Matthew Brost
Reviewed
On Fri, May 02, 2025 at 05:31:49AM +0100, Al Viro wrote:
> On Thu, May 01, 2025 at 09:26:25PM -0700, Matthew Brost wrote:
>
> > I;m fairly certain is just aliasing... but I do understand a file cannot
> > be embedded. Would comment help here indicating no other fields shou
On Fri, May 02, 2025 at 03:34:47AM +0100, Al Viro wrote:
> On Thu, May 01, 2025 at 07:13:12PM -0700, Matthew Brost wrote:
> > On Thu, May 01, 2025 at 05:24:38PM -0700, Kees Cook wrote:
> > > Casting through a "void *" isn't sufficient to convince the randstruct
&g
/ttm/ttm_backup.c: In function 'ttm_file_to_backup':
> drivers/gpu/drm/ttm/ttm_backup.c:21:16: note: randstruct: casting between
> randomized structure pointer types (ssa): 'struct ttm_backup' and 'struct
> file'
>21 | return (void *)file;
&g
> |^~~~
>
> Fixes: e7b5d23e5d47 ("drm/ttm: Provide a shmem backup implementation")
> Signed-off-by: Kees Cook
> ---
> Cc: Thomas Hellström
> Cc: Christian Koenig
> Cc: Somalapuram Amaranath
> Cc: Matthew Brost
What if we did somethi
On Tue, Apr 29, 2025 at 10:51:18AM -0700, Matthew Brost wrote:
> On Tue, Apr 29, 2025 at 09:34:00AM +0200, Arnd Bergmann wrote:
> > From: Arnd Bergmann
> >
> > The device core dumps are copied in 1.5GB chunks, which leads to a
> > link-time error on 32-bit builds bec
On Tue, Apr 29, 2025 at 09:34:00AM +0200, Arnd Bergmann wrote:
> From: Arnd Bergmann
>
> The device core dumps are copied in 1.5GB chunks, which leads to a
> link-time error on 32-bit builds because of the 64-bit division not
> getting trivially turned into mask and shift operations:
>
> ERROR:
On Mon, Apr 28, 2025 at 01:40:42PM +1000, Dave Airlie wrote:
> From: Dave Airlie
>
Thanks Dave, I already merge this fix. I guess our new CI policy
rejected this though - I'll follow up on allowing you trigger our CI for
anything going forward.
Matt
> Fixes: 55df7c0c62c1 ("drm/ttm/xe: drop unu
es in regs folder (Michal W)
> > - Move xe_guc_pagefault_desc to regs folder (Michal W)
> >
> > Bspec: 77412
> > Signed-off-by: Jonathan Cavitt
> > Reviewed-by: Shuicheng Lin
> > Acked-by: Matthew Brost
> > Cc: Michal Wajdeczko
> > ---
> > drivers
On Sun, Apr 27, 2025 at 06:51:30PM -0700, Matthew Brost wrote:
Ignore this one... I screwed up the reply list as intel-xe is not
include. Replying again to include intel-xe.
Matt
> On Sun, Apr 27, 2025 at 02:57:22PM +0300, Dafna Hirschfeld wrote:
> > On 24.04.2025 20:49, Jonathan Cav
es in regs folder (Michal W)
> > - Move xe_guc_pagefault_desc to regs folder (Michal W)
> >
> > Bspec: 77412
> > Signed-off-by: Jonathan Cavitt
> > Reviewed-by: Shuicheng Lin
> > Acked-by: Matthew Brost
> > Cc: Michal Wajdeczko
> > ---
> > drivers
On Wed, Apr 23, 2025 at 12:29:20PM +1000, Dave Airlie wrote:
> From: Dave Airlie
>
> This flag used to be used in the old memory tracking code, that
> code got migrated into the vmwgfx driver[1], and then got removed
> from the tree[2], but this piece got left behind.
>
> [1] f07069da6b4c ("drm/
On Fri, Apr 25, 2025 at 09:18:19AM +0200, Thomas Hellström wrote:
> On Thu, 2025-04-24 at 11:03 -0700, Matthew Brost wrote:
> > On Thu, Apr 24, 2025 at 04:39:21PM +0200, Thomas Hellström wrote:
> > > On Tue, 2025-04-22 at 10:04 -0700, Matthew Brost wrote:
> > > > M
ombine svm/userptr handling further down into op_check_svm_userptr.
>
> Suggested-by: Matthew Brost
The Kconfig issues pointed out in the previous patch need to be fixed
but implementation here look correct:
Reviewed-by: Matthew Brost
> Signed-off-by: Matthew Auld
> Cc: Himal
which is now provided by gpusvm.
>
> v2:
> - Reuse the same SVM vm struct we use for full SVM, that way we can
> use the same lock (Matt B & Himal)
> v3:
> - Re-use svm_init/fini for userptr.
>
> Signed-off-by: Matthew Auld
> Cc: Himal Prasad Ghimiray
On Thu, Apr 24, 2025 at 03:01:21PM -0700, Matthew Brost wrote:
> On Thu, Apr 24, 2025 at 01:18:33PM +0100, Matthew Auld wrote:
> > Export get/unmap/free pages API. We also need to tweak the SVM init to
> > allow skipping much of the unneeded parts.
> >
> > Signed-o
On Thu, Apr 24, 2025 at 01:18:33PM +0100, Matthew Auld wrote:
> Export get/unmap/free pages API. We also need to tweak the SVM init to
> allow skipping much of the unneeded parts.
>
> Signed-off-by: Matthew Auld
> Cc: Thomas Hellström
> Cc: Matthew Brost
> ---
> driv
On Thu, Apr 24, 2025 at 01:18:32PM +0100, Matthew Auld wrote:
> Refactor the core API of get/unmap/free pages to all operate on
> drm_gpusvm_pages. In the next patch we want to export a simplified core
> API without needing fully blown svm range etc.
>
> Suggested-by: Matthew Bros
erge request to this cover letter
>
> v19:
> - Perform kzalloc outside of lock (Auld)
>
> v20:
> - Fix inconsistent use of whitespace in defines
>
> v21:
> - Remove unnecessary size assertion (jcavitt)
>
> v22:
> - Fix xe_vm_fault_entry kernel docs (Shuichen
quot;)
> Signed-off-by: Jonathan Cavitt
> Suggested-by: Matthew Brost
Reviewed-by: Matthew Brost
> Reviewed-by: Shuicheng Lin
> ---
> drivers/gpu/drm/xe/xe_gt_pagefault.c | 5 +
> 1 file changed, 5 insertions(+)
>
> diff --git a/drivers/gpu/drm/xe/xe_gt_pagefau
4:
> - Map access type, fault type, and fault level to user macros (Matt
> Brost, Ivan)
>
> v15:
> - Remove unnecessary size assertion (jcavitt)
>
> Signed-off-by: Jonathan Cavitt
> Suggested-by: Matthew Brost
> Reviewed-by: Shuicheng Lin
> Cc: Jainxun Zhang
> Cc: Shui
On Thu, Apr 24, 2025 at 04:39:21PM +0200, Thomas Hellström wrote:
> On Tue, 2025-04-22 at 10:04 -0700, Matthew Brost wrote:
> > Mixing GPU and CPU atomics does not work unless a strict migration
> > policy of GPU atomics must be device memory. Enforce a policy of must
> >
counting
> > all
> > the involved objects has been rejected in the past. And since the problem
> > probably expands to all dma fences it certainly isn't easy.
> >
> > To be clear once more - lets not focus on how this does not fix it fully -
> > I am
pe, and fault level (Matt Brost,
> > Ivan)
> >
> > v6:
> > - Fix inconsistent use of whitespace in defines
> >
> > Signed-off-by: Jonathan Cavitt
> > Cc: Zhang Jianxun
> > Cc: Ivan Briano
> > Cc: Matthew Brost
> > ---
> > include/u
On Mon, Apr 07, 2025 at 08:29:11AM +0100, Matthew Auld wrote:
> On 04/04/2025 18:02, Matthew Brost wrote:
> > On Fri, Apr 04, 2025 at 09:19:34AM +0100, Matthew Auld wrote:
> > > On 03/04/2025 22:25, Matthew Brost wrote:
> > > > On Fri, Mar 28, 2025 at 06:10:
On Tue, Apr 22, 2025 at 10:51:44PM +0530, Ghimiray, Himal Prasad wrote:
>
>
> On 22-04-2025 22:34, Matthew Brost wrote:
> > Mixing GPU and CPU atomics does not work unless a strict migration
> > policy of GPU atomics must be device memory. Enforce a policy of must be
>
On Tue, Apr 22, 2025 at 10:04:11AM -0700, Matthew Brost wrote:
> From: Himal Prasad Ghimiray
>
> This commit adds a new flag, devmem_only, to the drm_gpusvm structure. The
> purpose of this flag is to ensure that the get_pages function allocates
> memory exclusively from the devi
1] f07069da6b4c ("drm/ttm: move memory accounting into vmwgfx v4")
> [2] 8aadeb8ad874 ("drm/vmwgfx: Remove the dedicated memory accounting")
>
> Cleanup the dead code.
>
> Signed-off-by: Dave Airlie
Reviewed-by: Matthew Brost
> ---
> drivers/gpu/drm/
On Tue, Apr 22, 2025 at 06:48:08PM -0700, Matthew Brost wrote:
> On Wed, Apr 23, 2025 at 09:35:35AM +1000, Dave Airlie wrote:
> > Hey,
> >
> > Been digging a bit into cgroups stuff and I found this force_alloc
> > flags in the TTM operation context that seemed like I s
On Wed, Apr 23, 2025 at 09:35:35AM +1000, Dave Airlie wrote:
> Hey,
>
> Been digging a bit into cgroups stuff and I found this force_alloc
> flags in the TTM operation context that seemed like I should dig
> deeper, but I can't see this flag being used anywhere, places set it
> but I can't spot th
rn an -EFAULT error.
v3:
- s/vram_only/devmem_only/
Signed-off-by: Matthew Brost
Signed-off-by: Himal Prasad Ghimiray
---
drivers/gpu/drm/drm_gpusvm.c | 5 +
include/drm/drm_gpusvm.h | 2 ++
2 files changed, 7 insertions(+)
diff --git a/drivers/gpu/drm/drm_gpusvm.c b/drivers/gp
Add timeslicing support to GPU SVM which will guarantee the GPU a
minimum execution time on piece of physical memory before migration back
to CPU. Intended to implement strict migration policies which require
memory to be in a certain placement for correct execution.
Signed-off-by: Matthew Brost
set vram_only on DGFX (Himal)
- Bail on get_pages failure if vram_only and retry count exceeded (Himal)
- s/vram_only/devmem_only
- Update xe_svm_range_is_valid to accept devmem_only argument
v4:
- Fix logic bug get_pages failure
Signed-off-by: Himal Prasad Ghimiray
Signed-off-by: Matthew Brost
Ensure GPU can make forward progress on an atomic SVM GPU fault by
giving the GPU a timeslice of 5ms
v2:
- Reduce timeslice to 5ms
- Double timeslice on retry
- Split out GPU SVM changes into independent patch
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_svm.c | 3 +++
1 file
Add some informal control for atomic SVM fault GPU timeslice to be able
to play around with values and tweak performance.
v2:
- Reduce timeslice default value to 5ms
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_debugfs.c | 38
drivers/gpu/drm/xe
Minimal set of patches to enable compute UMD SVM atomics.
Collaboration with Himal.
Matt
Himal Prasad Ghimiray (1):
drm/gpusvm: Introduce devmem_only flag for allocation
Matthew Brost (4):
drm/xe: Strict migration policy for atomic SVM faults
drm/gpusvm: Add timeslicing support to GPU
Add some informal control for atomic SVM fault GPU timeslice to be able
to play around with values and tweak performance.
v2:
- Reduce timeslice default value to 5ms
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_debugfs.c | 38
drivers/gpu/drm/xe
Minimal set of patches to enable compute UMD SVM atomics.
Collaboration with Himal.
Matt
Himal Prasad Ghimiray (1):
drm/gpusvm: Introduce devmem_only flag for allocation
Matthew Brost (4):
drm/xe: Strict migration policy for atomic SVM faults
drm/gpusvm: Add timeslicing support to GPU
Ensure GPU can make forward progress on an atomic SVM GPU fault by
giving the GPU a timeslice of 5ms
v2:
- Reduce timeslice to 5ms
- Double timeslice on retry
- Split out GPU SVM changes into independent patch
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_svm.c | 3 +++
1 file
Add timeslicing support to GPU SVM which will guarantee the GPU a
minimum execution time on piece of physical memory before migration back
to CPU. Intended to implement strict migration policies which require
memory to be in a certain placement for correct execution.
Signed-off-by: Matthew Brost
rn an -EFAULT error.
v3:
- s/vram_only/devmem_only/
Signed-off-by: Matthew Brost
Signed-off-by: Himal Prasad Ghimiray
---
drivers/gpu/drm/drm_gpusvm.c | 5 +
include/drm/drm_gpusvm.h | 2 ++
2 files changed, 7 insertions(+)
diff --git a/drivers/gpu/drm/drm_gpusvm.c b/drivers/gp
set vram_only on DGFX (Himal)
- Bail on get_pages failure if vram_only and retry count exceeded (Himal)
- s/vram_only/devmem_only
- Update xe_svm_range_is_valid to accept devmem_only argument
Signed-off-by: Himal Prasad Ghimiray
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_module.c
On Mon, Apr 21, 2025 at 12:09:15PM +0530, Ghimiray, Himal Prasad wrote:
>
>
> On 17-04-2025 09:43, Matthew Brost wrote:
> > Mixing GPU and CPU atomics does not work unless a strict migration
> > policy of GPU atomics must be device memory. Enforce a policy of must be
>
From: Himal Prasad Ghimiray
This commit adds a new flag, vram_only, to the drm_gpusvm structure. The
purpose of this flag is to ensure that the get_pages function allocates
memory exclusively from the device's VRAM. If the allocation from VRAM
fails, the function will return an -EFAULT error.
Si
On Wed, Apr 16, 2025 at 12:59:08PM -0700, Matthew Brost wrote:
> On Wed, Apr 16, 2025 at 11:11:03AM -0700, Matthew Brost wrote:
> > Minimal set of patches to enable compute UMD SVM atomics.
> >
> > Collaboration with Himal.
> >
> > Sending as RFC to see if we
-by: Himal Prasad Ghimiray
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_module.c | 3 --
drivers/gpu/drm/xe/xe_module.h | 1 -
drivers/gpu/drm/xe/xe_svm.c| 57 ++
drivers/gpu/drm/xe/xe_svm.h| 5 ---
4 files changed, 44 insertions(+), 22 deletions
Ensure GPU can make forward progress on an atomic SVM GPU fault by
giving the GPU a timeslice of 5ms
v2:
- Reduce timeslice to 5ms
- Double timeslice on retry
- Split out GPU SVM changes into independent patch
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_svm.c | 2 ++
1 file
Add timeslicing support to GPU SVM which will guarantee the GPU a
minimum execution time on piece of physical memory before migration back
to CPU. Intended to implement strict migration policies which require
memory to be in a certain placement for correct execution.
Signed-off-by: Matthew Brost
Minimal set of patches to enable compute UMD SVM atomics.
Collaboration with Himal.
Matt
Himal Prasad Ghimiray (1):
drm/gpusvm: Introduce vram_only flag for VRAM allocation
Matthew Brost (4):
drm/xe: Strict migration policy for atomic SVM faults
drm/gpusvm: Add timeslicing support to GPU
Add some informal control for atomic SVM fault GPU timeslice to be able
to play around with values and tweak performance.
v2:
- Reduce timeslice default value to 5ms
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_debugfs.c | 38
drivers/gpu/drm/xe
On Wed, Apr 16, 2025 at 11:11:03AM -0700, Matthew Brost wrote:
> Minimal set of patches to enable compute UMD SVM atomics.
>
> Collaboration with Himal.
>
> Sending as RFC to see if we should pursue merging this series ASAP, the
> solution of timeslicing may not be the final
From: Himal Prasad Ghimiray
This commit adds a new flag, vram_only, to the drm_gpusvm structure. The
purpose of this flag is to ensure that the get_pages function allocates
memory exclusively from the device's VRAM. If the allocation from VRAM
fails, the function will return an -EFAULT error.
Si
Add some informal control for atomic SVM fault GPU timeslice to be able
to play around with values and tweak performance.
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_debugfs.c | 38
drivers/gpu/drm/xe/xe_device.c | 1 +
drivers/gpu/drm/xe
Mixing GPU and CPU atomics does not work unless a strict migration
policy of GPU atomics must be device memory. Enforce a policy of must be
in VRAM with a retry loop of 2 attempts, if retry loop fails abort
fault.
Signed-off-by: Himal Prasad Ghimiray
Signed-off-by: Matthew Brost
---
drivers
Ensure GPU can make forward progress on an atomic SVM GPU fault by
giving the GPU a timeslice of 10 MS.
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/drm_gpusvm.c | 9 +
drivers/gpu/drm/xe/xe_svm.c | 1 +
include/drm/drm_gpusvm.h | 5 +
3 files changed, 15 insertions
(1):
drm/gpusvm: Introduce vram_only flag for VRAM allocation
Matthew Brost (3):
drm/xe: Strict migration policy for atomic SVM faults
drm/xe: Timeslice GPU on atomic SVM fault
drm/xe: Add atomic_svm_timeslice_ms debugfs entry
drivers/gpu/drm/drm_gpusvm.c | 14 +++
drivers
On Thu, Apr 03, 2025 at 03:28:48PM -0600, Cavitt, Jonathan wrote:
> -Original Message-
> From: Intel-xe On Behalf Of Matthew
> Brost
> Sent: Thursday, April 3, 2025 1:27 PM
> To: intel...@lists.freedesktop.org
> Cc: dri-devel@lists.freedesktop.org
> Subject: [PATCH v
e next
> patch we want to extract get/unmap/free to operate on a different range
> type.
>
> Signed-off-by: Matthew Auld
> Cc: Thomas Hellström
> Cc: Matthew Brost
> ---
> drivers/gpu/drm/drm_gpusvm.c | 90 ++--
> 1 file change
Abort printing coredump in VM printer output if full. Helps speedup
large coredumps which need to walked multiple times in
xe_devcoredump_read.
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_vm.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers
Devcoredump were truncated at 2G, remove this restriction. While here,
add support for GPU copies of BOs to increase devcoredump speed.
v2:
- Fix build error
- Abort printing once printer if full
Matthew Brost (4):
drm/xe: Add devcoredump chunking
drm/xe: Update xe_ttm_access_memory to use
On Fri, Apr 04, 2025 at 09:19:34AM +0100, Matthew Auld wrote:
> On 03/04/2025 22:25, Matthew Brost wrote:
> > On Fri, Mar 28, 2025 at 06:10:36PM +, Matthew Auld wrote:
> > > We now use the same notifier lock for SVM and userptr, with that we can
> > > combine x
On Fri, Mar 28, 2025 at 06:10:36PM +, Matthew Auld wrote:
> We now use the same notifier lock for SVM and userptr, with that we can
> combine xe_pt_userptr_pre_commit and xe_pt_svm_pre_commit.
>
> Suggested-by: Matthew Brost
> Signed-off-by: Matthew Auld
> Cc: Himal Pra
On Thu, Apr 03, 2025 at 03:41:16PM -0700, John Harrison wrote:
> On 4/3/2025 3:32 PM, Matthew Brost wrote:
> > On Thu, Apr 03, 2025 at 02:39:16PM -0700, John Harrison wrote:
> > > On 4/3/2025 1:27 PM, Matthew Brost wrote:
> > > > Chunk devcoredump into 1.5G pieces
On Thu, Apr 03, 2025 at 02:39:16PM -0700, John Harrison wrote:
> On 4/3/2025 1:27 PM, Matthew Brost wrote:
> > Chunk devcoredump into 1.5G pieces to avoid hitting the kvmalloc limit
> > of 2G. Simple algorithm reads 1.5G at time in xe_devcoredump_read
> > callback as needed.
&
On Fri, Mar 28, 2025 at 06:10:36PM +, Matthew Auld wrote:
> We now use the same notifier lock for SVM and userptr, with that we can
> combine xe_pt_userptr_pre_commit and xe_pt_svm_pre_commit.
>
> Suggested-by: Matthew Brost
> Signed-off-by: Matthew Auld
> Cc: Himal Pra
s, which can then be re-used for stuff like userptr.
>
> v2:
> - Move seq into pages struct (Matt B)
>
> Suggested-by: Matthew Brost
Reviewed-by: Matthew Brost
> Signed-off-by: Matthew Auld
> Cc: Thomas Hellström
> ---
Add drm_printer_is_full which indicates if a drm printer's output is
full. Useful to short circuit coredump printing once printer's output is
full.
Signed-off-by: Matthew Brost
---
include/drm/drm_print.h | 17 +
1 file changed, 17 insertions(+)
diff --git a/i
Chunk devcoredump into 1.5G pieces to avoid hitting the kvmalloc limit
of 2G. Simple algorithm reads 1.5G at time in xe_devcoredump_read
callback as needed.
Signed-off-by: Matthew Brost
---
drivers/gpu/drm/xe/xe_devcoredump.c | 59 ++-
drivers/gpu/drm/xe
1 - 100 of 1952 matches
Mail list logo
