Add TODO list (+ small docu change)
Subsystems (4)
Philipp Stanner
maintainerv2 cover Add TODO list (+ small docu change)
drm_sched has so many problems that we should have our own TODO list for contributors who might wanna help.
Changes in v2:
- Add generic TODO list example that can stay in the file forever.
Philipp Stanner (4):
drm/sched: Remove out of place resubmit docu drm/sched: Add a TODO list
drm/sched: Add TODO entry for resubmitting jobs drm/sched: Add TODO entry for missing runqueue locks
drivers/gpu/drm/scheduler/TODO | 51 ++++++++++++++++++++++++++++++++++ include/drm/gpu_scheduler.h | 10 ------- 2 files changed, 51 insertions(+), 10 deletions(-) create mode 100644 drivers/gpu/drm/scheduler/TODO
Signature · 1 line
2.49.0
Philipp Stanner
maintainerPATCH v2 1/4 drm/sched: Remove out of place resubmit docu
The documentation for drm_sched_backend_ops.run_job() details that that callback can be invoked multiple times by the deprecated function drm_sched_resubmit_jobs(). It also contains an unresolved TODO.
It is not useful to document side effects of a different, deprecated function in the docu of run_job(): Existing users won't re-evaluate their usage of the deprecated function by reading the non-deprecated one, and new users must not use the deprecated function in the first place.
Remove the out of place documentation.
--- include/drm/gpu_scheduler.h | 10 ---------- 1 file changed, 10 deletions(-)
include/drm/gpu_scheduler.h +0 -10
@@ -429,16 +429,6 @@ struct drm_sched_backend_ops {* * @sched_job: the job to run * -* The deprecated drm_sched_resubmit_jobs() (called by &struct -* drm_sched_backend_ops.timedout_job) can invoke this again with the -* same parameters. Using this is discouraged because it violates -* dma_fence rules, notably dma_fence_init() has to be called on -* already initialized fences for a second time. Moreover, this is -* dangerous because attempts to allocate memory might deadlock with -* memory management code waiting for the reset to complete. -* -* TODO: Document what drivers should do / use instead. -* * This method is called in a workqueue context - either from the * submit_wq the driver passed through drm_sched_init(), or, if the * driver passed NULL, a separate, ordered workqueue the scheduler
Signature · 1 line
2.49.0
Philipp Stanner
maintainerPATCH v2 2/4 drm/sched: Add a TODO list
In reply to Philipp StannerThe DRM GPU scheduler contains a huge number of problems. These should be documented for (new) contributors to pick up work.
Add a TODO list with an example entry.
--- drivers/gpu/drm/scheduler/TODO | 12 ++++++++++++ 1 file changed, 12 insertions(+) create mode 100644 drivers/gpu/drm/scheduler/TODO
drivers/gpu/drm/scheduler/TODO new file +12 -0
@@ -0,0 +1,12 @@ +=== drm_sched TODO list === + +* Example Entry + - Difficulty: hard + - Contact: + - Danilo Krummrich <[email protected]> + - Philipp Stanner <[email protected]> + - Description: + This is an example. + - Tasks: + 1. Read the example entry. + 2. Remove the entry once solved (never in this case)
Signature · 1 line
2.49.0
Philipp Stanner
maintainerPATCH v2 3/4 drm/sched: Add TODO entry for resubmitting jobs
In reply to Philipp StannerAdd the issue of a successor of drm_sched_resubmit_jobs() missing to the TODO file.
--- drivers/gpu/drm/scheduler/TODO | 26 ++++++++++++++++++++++++++ 1 file changed, 26 insertions(+)
drivers/gpu/drm/scheduler/TODO +26 -0
@@ -10,3 +10,29 @@ - Tasks: 1. Read the example entry. 2. Remove the entry once solved (never in this case) + +* GPU job resubmits + - Difficulty: hard + - Contact: + - Christian König <[email protected]> + - Philipp Stanner <[email protected]> + - Description: + drm_sched_resubmit_jobs() is deprecated. Main reason being that it leads to + reinitializing dma_fences. See that function's docu for details. The better + approach for valid resubmissions by amdgpu and Xe is (apparently) to figure + out which job (and, through association: which entity) caused the hang. Then, + the job's buffer data, together with all other jobs' buffer data currently + in the same hardware ring, must be invalidated. This can for example be done + by overwriting it. + amdgpu currently determines which jobs are in the ring and need to be + overwritten by keeping copies of the job. Xe obtains that information by + directly accessing drm_sched's pending_list. + - Tasks: + 1. implement scheduler functionality through which + the driver can obtain the information which *broken* jobs are currently in + the hardware ring. + 2. Such infrastructure would then typically be used in + drm_sched_backend_ops.timedout_job(). Document that. + 3. Port a driver as first user. + 3. Document the new alternative in the docu of deprecated + drm_sched_resubmit_jobs().
Signature · 1 line
2.49.0
Philipp Stanner
maintainerPATCH v2 4/4 drm/sched: Add TODO entry for missing runqueue locks
In reply to Philipp Stannerstruct drm_sched_rq is not being locked at many places throughout the scheduler, at least for readers. This was documented in a FIXME added in:
commit 981b04d96856 ("drm/sched: improve docs around drm_sched_entity")
Add a TODO entry for that problem.
--- drivers/gpu/drm/scheduler/TODO | 13 +++++++++++++ 1 file changed, 13 insertions(+)
drivers/gpu/drm/scheduler/TODO +13 -0
@@ -36,3 +36,16 @@ 3. Port a driver as first user. 3. Document the new alternative in the docu of deprecated drm_sched_resubmit_jobs(). + +* Unlocked readers for runqueues + - Difficulty: medium + - Contact: Philipp Stanner <[email protected]> + - Description: + There is an old FIXME by Sima in include/drm/gpu_scheduler.h. It details + that struct drm_sched_rq is read at many places without any locks, not even + with a READ_ONCE. At XDC 2025 no one could really tell why that is the case, + whether locks are needed and whether they could be added. (But for real, + that should probably be locked!). + - Tasks: + 1. Check whether locks for runqueue readers can be added. + 2. If yes, add the locks.
Signature · 1 line
2.49.0
Quoted 2 lines · Philipp Stanner
drm_schedhas so many problems that we should have our own TODO list for contributors who might wanna help.
Looks good,
Quoted 1 line
drivers/gpu/drm/scheduler/TODO | 51 ++++++++++++++++++++++++++++++++++
I'd move this into Documentation/, just like we did it for Nova [1].