Skip to contentSkip to conversation

drm/sched: Documentation and refcount improvements

PATCH Superseded Philipp Stanner 5 participants 13 messages Jan 9, 2025 – Jan 20, 2025

Subsystems (4)

v1 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Philipp Stanner

author maintainer
#0 lore raw

cover drm/sched: Documentation and refcount improvements

This is as series succeeding my previous patch [1].

I recognized that we are still referring to a non-existing function and a deprecated one in the callback docu. We should probably also point out the important distinction between hardware and firmware schedulers more cleanly.

Please give me feedback, especially on the RFC comments in patch3.

(This series still fires docu-build-warnings. I want to gather feedback on the opion questions first and will solve them in v2.)

Thank you,
Philipp

[1] https://lore.kernel.org/all/[email protected]/

Philipp Stanner (3):
drm/sched: Document run_job() refcount hazard drm/sched: Adjust outdated docu for run_job() drm/sched: Update timedout_job()'s documentation

 drivers/gpu/drm/scheduler/sched_main.c |  10 ++-
 include/drm/gpu_scheduler.h            | 105 ++++++++++++++++---------
 2 files changed, 77 insertions(+), 38 deletions(-)
Signature · 1 line
2.47.1
author maintainer
#1 lore raw

PATCH 1/3 drm/sched: Document run_job() refcount hazard

From: Philipp Stanner <[email protected]>

drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler. That fence is signalled by the driver once the hardware completed the associated job. The scheduler does not increment the reference count on that fence, but implicitly expects to inherit this fence from run_job().

This is relatively subtle and prone to misunderstandings.

This implies that, to keep a reference for itself, a driver needs to call dma_fence_get() in addition to dma_fence_init() in that callback.

It's further complicated by the fact that the scheduler even decrements the refcount in drm_sched_run_job_work() since it created a new reference in drm_sched_fence_scheduled(). It does, however, still use its pointer to the fence after calling dma_fence_put() - which is safe because of the aforementioned new reference, but actually still violates the refcounting rules.

Improve the explanatory comment for that decrement.

Move the call to dma_fence_put() to the position behind the last usage of the fence.

Document the necessity to increment the reference count in drm_sched_backend_ops.run_job().

---
 drivers/gpu/drm/scheduler/sched_main.c | 10 +++++++---
 include/drm/gpu_scheduler.h            | 19 +++++++++++++++----
 2 files changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +7 -3
index 57da84908752..5f46c01eb01e 100644
@@ -1218,15 +1218,19 @@ static void drm_sched_run_job_work(struct work_struct *w)
 	drm_sched_fence_scheduled(s_fence, fence);
 
 	if (!IS_ERR_OR_NULL(fence)) {
-		/* Drop for original kref_init of the fence */
-		dma_fence_put(fence);
-
 		r = dma_fence_add_callback(fence, &sched_job->cb,
 					   drm_sched_job_done_cb);
 		if (r == -ENOENT)
 			drm_sched_job_done(sched_job, fence->error);
 		else if (r)
 			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
+
+		/*
+		 * s_fence took a new reference to fence in the call to
+		 * drm_sched_fence_scheduled() above. The reference passed by
+		 * run_job() above is now not needed any longer. Drop it.
+		 */
+		dma_fence_put(fence);
 	} else {
 		drm_sched_job_done(sched_job, IS_ERR(fence) ?
 				   PTR_ERR(fence) : 0);
include/drm/gpu_scheduler.h +15 -4
index 95e17504e46a..d5cd2a78f27c 100644
@@ -420,10 +420,21 @@ struct drm_sched_backend_ops {
 					 struct drm_sched_entity *s_entity);
 
 	/**
-         * @run_job: Called to execute the job once all of the dependencies
-         * have been resolved.  This may be called multiple times, if
-	 * timedout_job() has happened and drm_sched_job_recovery()
-	 * decides to try it again.
+	 * @run_job: Called to execute the job once all of the dependencies
+	 * have been resolved. This may be called multiple times, if
+	 * timedout_job() has happened and drm_sched_job_recovery() decides to
+	 * try it again.
+	 *
+	 * @sched_job: the job to run
+	 *
+	 * Returns: dma_fence the driver must signal once the hardware has
+	 *	completed the job ("hardware fence").
+	 *
+	 * Note that the scheduler expects to 'inherit' its own reference to
+	 * this fence from the callback. It does not invoke an extra
+	 * dma_fence_get() on it. Consequently, this callback must take a
+	 * reference for the scheduler, and additional ones for the driver's
+	 * respective needs.
 	 */
 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
 
Signature · 1 line
2.47.1
author maintainer
#8 lore raw

PATCH 2/3 drm/sched: Adjust outdated docu for run_job()

In reply to Philipp Stanner

The documentation for drm_sched_backend_ops.run_job() mentions a certain function called drm_sched_job_recovery(). This function does not exist. What's actually meant is drm_sched_resubmit_jobs(), which is by now also deprecated.

Remove the mention of the deprecated function.

Discourage the behavior of drm_sched_backend_ops.run_job() being called multiple times for the same job.

---
 include/drm/gpu_scheduler.h | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
include/drm/gpu_scheduler.h +6 -3
index d5cd2a78f27c..c4e65f9f7f22 100644
@@ -421,9 +421,12 @@ struct drm_sched_backend_ops {
 
 	/**
 	 * @run_job: Called to execute the job once all of the dependencies
-	 * have been resolved. This may be called multiple times, if
-	 * timedout_job() has happened and drm_sched_job_recovery() decides to
-	 * try it again.
+	 * have been resolved.
+	 *
+	 * The deprecated drm_sched_resubmit_jobs() (called from
+	 * drm_sched_backend_ops.timedout_job()) can invoke this again with the
+	 * same parameters. Doing this is strongly discouraged because it
+	 * violates dma_fence rules.
 	 *
 	 * @sched_job: the job to run
 	 *
Signature · 1 line
2.47.1
author maintainer
#10 lore raw

PATCH 3/3 drm/sched: Update timedout_job()'s documentation

In reply to Philipp Stanner

drm_sched_backend_ops.timedout_job()'s documentation is outdated. It mentions the deprecated function drm_sched_resubmit_job(). Furthermore, it does not point out the important distinction between hardware and firmware schedulers.

Since firmware schedulers tyipically only use one entity per scheduler, timeout handling is significantly more simple because the entity the faulted job came from can just be killed without affecting innocent processes.

Update the documentation with that distinction and other details.

Reformat the docstring to work to a unified style with the other handles.

---
 include/drm/gpu_scheduler.h | 83 +++++++++++++++++++++++--------------
 1 file changed, 52 insertions(+), 31 deletions(-)
include/drm/gpu_scheduler.h +52 -31
index c4e65f9f7f22..380b8840c591 100644
@@ -445,43 +445,64 @@ struct drm_sched_backend_ops {
 	 * @timedout_job: Called when a job has taken too long to execute,
 	 * to trigger GPU recovery.
 	 *
-	 * This method is called in a workqueue context.
+	 * @sched_job: The job that has timed out
 	 *
-	 * Drivers typically issue a reset to recover from GPU hangs, and this
-	 * procedure usually follows the following workflow:
+	 * Returns:
+	 * - DRM_GPU_SCHED_STAT_NOMINAL, on success, i.e., if the underlying
+	 *   driver has started or completed recovery.
+	 * - DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
+	 *   available, i.e., has been unplugged.
 	 *
-	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
-	 *    scheduler thread and cancel the timeout work, guaranteeing that
-	 *    nothing is queued while we reset the hardware queue
-	 * 2. Try to gracefully stop non-faulty jobs (optional)
-	 * 3. Issue a GPU reset (driver-specific)
-	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
+	 * Drivers typically issue a reset to recover from GPU hangs.
+	 * This procedure looks very different depending on whether a firmware
+	 * or a hardware scheduler is being used.
+	 *
+	 * For a FIRMWARE SCHEDULER, each (pseudo-)ring has one scheduler, and
+	 * each scheduler (typically) has one entity. Hence, you typically
+	 * follow those steps:
+	 *
+	 * 1. Stop the scheduler using drm_sched_stop(). This will pause the
+	 *    scheduler workqueues and cancel the timeout work, guaranteeing
+	 *    that nothing is queued while we reset the hardware queue.
+	 * 2. Try to gracefully stop non-faulty jobs (optional).
+	 * TODO: RFC ^ Folks, should we remove this step? What does it even mean
+	 * precisely to "stop" those jobs? Is that even helpful to userspace in
+	 * any way?
+	 * 3. Issue a GPU reset (driver-specific).
+	 * 4. Kill the entity the faulted job stems from, and the associated
+	 *    scheduler.
 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
-	 *    jobs can be queued, and the scheduler thread is unblocked
+	 *    jobs can be queued, and the scheduler workqueues awake again.
+	 *
+	 * For a HARDWARE SCHEDULER, each ring also has one scheduler, but each
+	 * scheduler typically has many attached entities. This implies that you
+	 * cannot tear down all entities associated with the affected scheduler,
+	 * because this would effectively also kill innocent userspace processes
+	 * which did not submit faulty jobs (for example).
+	 *
+	 * Consequently, the procedure to recover with a hardware scheduler
+	 * should look like this:
+	 *
+	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop().
+	 * 2. Figure out to which entity the faulted job belongs.
+	 * 3. Try to gracefully stop non-faulty jobs (optional).
+	 * TODO: RFC ^ Folks, should we remove this step? What does it even mean
+	 * precisely to "stop" those jobs? Is that even helpful to userspace in
+	 * any way?
+	 * 4. Kill that entity.
+	 * 5. Issue a GPU reset on all faulty rings (driver-specific).
+	 * 6. Re-submit jobs on all schedulers impacted by re-submitting them to
+	 *    the entities which are still alive.
+	 * 7. Restart all schedulers that were stopped in step #1 using
+	 *    drm_sched_start().
 	 *
 	 * Note that some GPUs have distinct hardware queues but need to reset
 	 * the GPU globally, which requires extra synchronization between the
-	 * timeout handler of the different &drm_gpu_scheduler. One way to
-	 * achieve this synchronization is to create an ordered workqueue
-	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
-	 * queue to drm_sched_init(), to guarantee that timeout handlers are
-	 * executed sequentially. The above workflow needs to be slightly
-	 * adjusted in that case:
-	 *
-	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
-	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
-	 *    the reset (optional)
-	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
-	 * 4. Re-submit jobs on all schedulers impacted by the reset using
-	 *    drm_sched_resubmit_jobs()
-	 * 5. Restart all schedulers that were stopped in step #1 using
-	 *    drm_sched_start()
-	 *
-	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
-	 * and the underlying driver has started or completed recovery.
-	 *
-	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
-	 * available, i.e. has been unplugged.
+	 * timeout handlers of different schedulers. One way to achieve this
+	 * synchronization is to create an ordered workqueue (using
+	 * alloc_ordered_workqueue()) at the driver level, and pass this queue
+	 * as drm_sched_init()'s @timeout_wq parameter. This will guarantee
+	 * that timeout handlers are executed sequentially.
 	 */
 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
 
Signature · 1 line
2.47.1
#2 lore raw

Re: [PATCH 1/3] drm/sched: Document run_job() refcount hazard

In reply to Philipp Stanner
Quoted 77 lines · Philipp Stanner

From: Philipp Stanner <[email protected]>

drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler. That fence is signalled by the driver once the hardware completed the associated job. The scheduler does not increment the reference count on that fence, but implicitly expects to inherit this fence from run_job().

This is relatively subtle and prone to misunderstandings.

This implies that, to keep a reference for itself, a driver needs to call dma_fence_get() in addition to dma_fence_init() in that callback.

It's further complicated by the fact that the scheduler even decrements the refcount in drm_sched_run_job_work() since it created a new reference in drm_sched_fence_scheduled(). It does, however, still use its pointer to the fence after calling dma_fence_put() - which is safe because of the aforementioned new reference, but actually still violates the refcounting rules.

Improve the explanatory comment for that decrement.

Move the call to dma_fence_put() to the position behind the last usage of the fence.

Document the necessity to increment the reference count in drm_sched_backend_ops.run_job().

---
  drivers/gpu/drm/scheduler/sched_main.c | 10 +++++++---
  include/drm/gpu_scheduler.h            | 19 +++++++++++++++----
  2 files changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +7 -3
index 57da84908752..5f46c01eb01e 100644
@@ -1218,15 +1218,19 @@ static void drm_sched_run_job_work(struct work_struct *w)
  	drm_sched_fence_scheduled(s_fence, fence);
 
  	if (!IS_ERR_OR_NULL(fence)) {
-		/* Drop for original kref_init of the fence */
-		dma_fence_put(fence);
-
  		r = dma_fence_add_callback(fence, &sched_job->cb,
  					   drm_sched_job_done_cb);
  		if (r == -ENOENT)
  			drm_sched_job_done(sched_job, fence->error);
  		else if (r)
  			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
+
+		/*
+		 * s_fence took a new reference to fence in the call to
+		 * drm_sched_fence_scheduled() above. The reference passed by
+		 * run_job() above is now not needed any longer. Drop it.
+		 */
+		dma_fence_put(fence);
  	} else {
  		drm_sched_job_done(sched_job, IS_ERR(fence) ?
  				   PTR_ERR(fence) : 0);
include/drm/gpu_scheduler.h +5 -4
index 95e17504e46a..d5cd2a78f27c 100644
@@ -420,10 +420,21 @@ struct drm_sched_backend_ops {
  					 struct drm_sched_entity *s_entity);
 
  	/**
-         * @run_job: Called to execute the job once all of the dependencies
-         * have been resolved.  This may be called multiple times, if
-	 * timedout_job() has happened and drm_sched_job_recovery()
-	 * decides to try it again.
+	 * @run_job: Called to execute the job once all of the dependencies
+	 * have been resolved. This may be called multiple times, if
+	 * timedout_job() has happened and drm_sched_job_recovery() decides to
+	 * try it again.
+	 *

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 437

Quoted 10 lines
include/drm/gpu_scheduler.h +10 -0
+	 * @sched_job: the job to run
+	 *
+	 * Returns: dma_fence the driver must signal once the hardware has
+	 *	completed the job ("hardware fence").
+	 *
+	 * Note that the scheduler expects to 'inherit' its own reference to
+	 * this fence from the callback. It does not invoke an extra
+	 * dma_fence_get() on it. Consequently, this callback must take a
+	 * reference for the scheduler, and additional ones for the driver's
+	 * respective needs.

Another thing which would be really good to document here is what other things run_job can return apart from the fence.

For instance amdgpu can return an ERR_PTR but I haven't looked into other drivers.

Regards,

Tvrtko

Quoted 3 lines
include/drm/gpu_scheduler.h +0 -0
  	 */
  	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
 
#3 lore raw

Re: [PATCH 1/3] drm/sched: Document run_job() refcount hazard

Quoted 125 lines · Tvrtko Ursulin
Quoted 117 lines · Philipp Stanner

From: Philipp Stanner <[email protected]>

drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler.
That fence is signalled by the driver once the hardware completed the
associated job. The scheduler does not increment the reference count on
that fence, but implicitly expects to inherit this fence from run_job().

This is relatively subtle and prone to misunderstandings.

This implies that, to keep a reference for itself, a driver needs to
call dma_fence_get() in addition to dma_fence_init() in that callback.

It's further complicated by the fact that the scheduler even decrements
the refcount in drm_sched_run_job_work() since it created a new reference in drm_sched_fence_scheduled(). It does, however, still use
its pointer to the fence after calling dma_fence_put() - which is safe
because of the aforementioned new reference, but actually still violates
the refcounting rules.

Improve the explanatory comment for that decrement.

Move the call to dma_fence_put() to the position behind the last usage
of the fence.

Document the necessity to increment the reference count in drm_sched_backend_ops.run_job().

---
  drivers/gpu/drm/scheduler/sched_main.c | 10 +++++++--- include/drm/gpu_scheduler.h            | 19 +++++++++++++++----

  2 files changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +0 -0
index 57da84908752..5f46c01eb01e 100644
@@ -1218,15 +1218,19 @@ static void drm_sched_run_job_work(struct
 work_struct *w)
   	drm_sched_fence_scheduled(s_fence, fence);


   if (!IS_ERR_OR_NULL(fence)) {

drivers/gpu/drm/scheduler/sched_main.c +0 -3
-		/* Drop for original kref_init of the fence */
-		dma_fence_put(fence);
-

   r = dma_fence_add_callback(fence, &sched_job->cb, drm_sched_job_done_cb); if (r == -ENOENT) drm_sched_job_done(sched_job, fence-

Quoted 1 line

error);

   else if (r) DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);

drivers/gpu/drm/scheduler/sched_main.c +3 -0
+
+		/*
+		 * s_fence took a new reference to fence in the

call to

drivers/gpu/drm/scheduler/sched_main.c +1 -0
+		 * drm_sched_fence_scheduled() above. The

reference passed by

drivers/gpu/drm/scheduler/sched_main.c +1 -0
+		 * run_job() above is now not needed any longer.

Drop it.

drivers/gpu/drm/scheduler/sched_main.c +2 -0
+		 */
+		dma_fence_put(fence);

   } else {
   drm_sched_job_done(sched_job, IS_ERR(fence) ? PTR_ERR(fence) : 0);

include/drm/gpu_scheduler.h +0 -0
index 95e17504e46a..d5cd2a78f27c 100644
@@ -420,10 +420,21 @@ struct drm_sched_backend_ops {

   struct drm_sched_entity *s_entity);

   /**
-         * @run_job: Called to execute the job once all of the dependencies
-         * have been resolved.  This may be called multiple times, if

include/drm/gpu_scheduler.h +0 -1
-	 * timedout_job() has happened and

drm_sched_job_recovery()

include/drm/gpu_scheduler.h +1 -1
-	 * decides to try it again.
+	 * @run_job: Called to execute the job once all of the

dependencies

include/drm/gpu_scheduler.h +1 -0
+	 * have been resolved. This may be called multiple times,

if

include/drm/gpu_scheduler.h +1 -0
+	 * timedout_job() has happened and

drm_sched_job_recovery() decides to

include/drm/gpu_scheduler.h +5 -0
+	 * try it again.
+	 *
+	 * @sched_job: the job to run
+	 *
+	 * Returns: dma_fence the driver must signal once the

hardware has

include/drm/gpu_scheduler.h +3 -0
+	 *	completed the job ("hardware fence").
+	 *
+	 * Note that the scheduler expects to 'inherit' its own

reference to

include/drm/gpu_scheduler.h +1 -0
+	 * this fence from the callback. It does not invoke an

extra

include/drm/gpu_scheduler.h +1 -0
+	 * dma_fence_get() on it. Consequently, this callback must

take a

include/drm/gpu_scheduler.h +1 -0
+	 * reference for the scheduler, and additional ones for

the driver's

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 437

Quoted 1 line
include/drm/gpu_scheduler.h +1 -0
+	 * respective needs.

Another thing which would be really good to document here is what other
things run_job can return apart from the fence.

Quoted 2 lines

For instance amdgpu can return an ERR_PTR but I haven't looked into other drivers.

ERR_PTR and NULL are both fine.

But good catch, I'll add that in v2

P.

Quoted 10 lines

Regards,

Tvrtko

Quoted 4 lines

   */
   struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);

#4 lore raw

Re: [PATCH 1/3] drm/sched: Document run_job() refcount hazard

In reply to Philipp Stanner

Am 09.01.25 um 14:37 schrieb Philipp Stanner:

Quoted 66 lines

From: Philipp Stanner <[email protected]>

drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler. That fence is signalled by the driver once the hardware completed the associated job. The scheduler does not increment the reference count on that fence, but implicitly expects to inherit this fence from run_job().

This is relatively subtle and prone to misunderstandings.

This implies that, to keep a reference for itself, a driver needs to call dma_fence_get() in addition to dma_fence_init() in that callback.

It's further complicated by the fact that the scheduler even decrements the refcount in drm_sched_run_job_work() since it created a new reference in drm_sched_fence_scheduled(). It does, however, still use its pointer to the fence after calling dma_fence_put() - which is safe because of the aforementioned new reference, but actually still violates the refcounting rules.

Improve the explanatory comment for that decrement.

Move the call to dma_fence_put() to the position behind the last usage of the fence.

Document the necessity to increment the reference count in drm_sched_backend_ops.run_job().

---
  drivers/gpu/drm/scheduler/sched_main.c | 10 +++++++---
  include/drm/gpu_scheduler.h            | 19 +++++++++++++++----
  2 files changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +7 -3
index 57da84908752..5f46c01eb01e 100644
@@ -1218,15 +1218,19 @@ static void drm_sched_run_job_work(struct work_struct *w)
  	drm_sched_fence_scheduled(s_fence, fence);
 
  	if (!IS_ERR_OR_NULL(fence)) {
-		/* Drop for original kref_init of the fence */
-		dma_fence_put(fence);
-
  		r = dma_fence_add_callback(fence, &sched_job->cb,
  					   drm_sched_job_done_cb);
  		if (r == -ENOENT)
  			drm_sched_job_done(sched_job, fence->error);
  		else if (r)
  			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
+
+		/*
+		 * s_fence took a new reference to fence in the call to
+		 * drm_sched_fence_scheduled() above. The reference passed by
+		 * run_job() above is now not needed any longer. Drop it.
+		 */
+		dma_fence_put(fence);
  	} else {
  		drm_sched_job_done(sched_job, IS_ERR(fence) ?
  				   PTR_ERR(fence) : 0);
include/drm/gpu_scheduler.h +0 -0
index 95e17504e46a..d5cd2a78f27c 100644
@@ -420,10 +420,21 @@ struct drm_sched_backend_ops {
  					 struct drm_sched_entity *s_entity);

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 426

Quoted 10 lines
include/drm/gpu_scheduler.h +4 -4
 
  	/**
-         * @run_job: Called to execute the job once all of the dependencies
-         * have been resolved.  This may be called multiple times, if
-	 * timedout_job() has happened and drm_sched_job_recovery()
-	 * decides to try it again.
+	 * @run_job: Called to execute the job once all of the dependencies
+	 * have been resolved. This may be called multiple times, if
+	 * timedout_job() has happened and drm_sched_job_recovery() decides to
+	 * try it again.

I just came to realize that this hack with calling run_job multiple times won't work any more with this patch here.

The background is that you can't allocate memory for a newly returned fence and as far as I know no driver pre allocates multiple HW fences for a job.

So at least amdgpu used to re-use the same HW fence as return over and over again, just re-initialized the reference count. I removed that hack from amdgpu, but just FYI it could be that other driver did the same.

Apart from that concern I think that this patch is really the right thing and that driver hacks relying on the order of dropping references are fundamentally broken approaches.

So feel free to add Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 14 lines
include/drm/gpu_scheduler.h +11 -0
+	 *
+	 * @sched_job: the job to run
+	 *
+	 * Returns: dma_fence the driver must signal once the hardware has
+	 *	completed the job ("hardware fence").
+	 *
+	 * Note that the scheduler expects to 'inherit' its own reference to
+	 * this fence from the callback. It does not invoke an extra
+	 * dma_fence_get() on it. Consequently, this callback must take a
+	 * reference for the scheduler, and additional ones for the driver's
+	 * respective needs.
  	 */
  	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
 
maintainer Reviewed-by on v1 1/3
#6 lore raw

Re: [PATCH 1/3] drm/sched: Document run_job() refcount hazard

In reply to Philipp Stanner
Quoted 44 lines · Philipp Stanner

From: Philipp Stanner <[email protected]>

drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler. That fence is signalled by the driver once the hardware completed the associated job. The scheduler does not increment the reference count on that fence, but implicitly expects to inherit this fence from run_job().

This is relatively subtle and prone to misunderstandings.

This implies that, to keep a reference for itself, a driver needs to call dma_fence_get() in addition to dma_fence_init() in that callback.

It's further complicated by the fact that the scheduler even decrements the refcount in drm_sched_run_job_work() since it created a new reference in drm_sched_fence_scheduled(). It does, however, still use its pointer to the fence after calling dma_fence_put() - which is safe because of the aforementioned new reference, but actually still violates the refcounting rules.

Improve the explanatory comment for that decrement.

Move the call to dma_fence_put() to the position behind the last usage of the fence.

Document the necessity to increment the reference count in drm_sched_backend_ops.run_job().

---
 drivers/gpu/drm/scheduler/sched_main.c | 10 +++++++---
 include/drm/gpu_scheduler.h            | 19 +++++++++++++++----
 2 files changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +0 -3
index 57da84908752..5f46c01eb01e 100644
@@ -1218,15 +1218,19 @@ static void drm_sched_run_job_work(struct work_struct *w)
 	drm_sched_fence_scheduled(s_fence, fence);
 
 	if (!IS_ERR_OR_NULL(fence)) {
-		/* Drop for original kref_init of the fence */
-		dma_fence_put(fence);
-

drivers/gpu/drm/scheduler/sched_main.c static void drm_sched_run_job_work(struct work_struct *w) · line 1230

Quoted 10 lines
drivers/gpu/drm/scheduler/sched_main.c +4 -0
 		r = dma_fence_add_callback(fence, &sched_job->cb,
 					   drm_sched_job_done_cb);
 		if (r == -ENOENT)
 			drm_sched_job_done(sched_job, fence->error);
 		else if (r)
 			DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);
+
+		/*
+		 * s_fence took a new reference to fence in the call to
+		 * drm_sched_fence_scheduled() above. The reference passed by

I think mentioning that in this context is a bit misleading. The reason we can put the fence here, is because we stop using the local fence pointer we have a reference for (from run_job()).

This has nothing to do with the fact that drm_sched_fence_scheduled() took its own reference when it stored a copy of this fence pointer in a separate data structure.

With that fixed,

Quoted 39 lines
drivers/gpu/drm/scheduler/sched_main.c +3 -0
+		 * run_job() above is now not needed any longer. Drop it.
+		 */
+		dma_fence_put(fence);
 	} else {
 		drm_sched_job_done(sched_job, IS_ERR(fence) ?
 				   PTR_ERR(fence) : 0);
include/drm/gpu_scheduler.h +15 -4
index 95e17504e46a..d5cd2a78f27c 100644
@@ -420,10 +420,21 @@ struct drm_sched_backend_ops {
 					 struct drm_sched_entity *s_entity);
 
 	/**
-         * @run_job: Called to execute the job once all of the dependencies
-         * have been resolved.  This may be called multiple times, if
-	 * timedout_job() has happened and drm_sched_job_recovery()
-	 * decides to try it again.
+	 * @run_job: Called to execute the job once all of the dependencies
+	 * have been resolved. This may be called multiple times, if
+	 * timedout_job() has happened and drm_sched_job_recovery() decides to
+	 * try it again.
+	 *
+	 * @sched_job: the job to run
+	 *
+	 * Returns: dma_fence the driver must signal once the hardware has
+	 *	completed the job ("hardware fence").
+	 *
+	 * Note that the scheduler expects to 'inherit' its own reference to
+	 * this fence from the callback. It does not invoke an extra
+	 * dma_fence_get() on it. Consequently, this callback must take a
+	 * reference for the scheduler, and additional ones for the driver's
+	 * respective needs.
 	 */
 	struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
 
Signature · 1 line
2.47.1
#7 lore raw

Re: [PATCH 1/3] drm/sched: Document run_job() refcount hazard

Quoted 79 lines · Danilo Krummrich
Quoted 70 lines · Philipp Stanner

From: Philipp Stanner <[email protected]>

drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler.
That fence is signalled by the driver once the hardware completed the
associated job. The scheduler does not increment the reference count on
that fence, but implicitly expects to inherit this fence from run_job().

This is relatively subtle and prone to misunderstandings.

This implies that, to keep a reference for itself, a driver needs to
call dma_fence_get() in addition to dma_fence_init() in that callback.

It's further complicated by the fact that the scheduler even decrements
the refcount in drm_sched_run_job_work() since it created a new reference in drm_sched_fence_scheduled(). It does, however, still use
its pointer to the fence after calling dma_fence_put() - which is safe
because of the aforementioned new reference, but actually still violates
the refcounting rules.

Improve the explanatory comment for that decrement.

Move the call to dma_fence_put() to the position behind the last usage
of the fence.

Document the necessity to increment the reference count in drm_sched_backend_ops.run_job().

---
 drivers/gpu/drm/scheduler/sched_main.c | 10 +++++++--- include/drm/gpu_scheduler.h            | 19 +++++++++++++++----

 2 files changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +0 -0
index 57da84908752..5f46c01eb01e 100644
@@ -1218,15 +1218,19 @@ static void drm_sched_run_job_work(struct
 work_struct *w)
  	drm_sched_fence_scheduled(s_fence, fence);


  if (!IS_ERR_OR_NULL(fence)) {

drivers/gpu/drm/scheduler/sched_main.c +0 -3
-		/* Drop for original kref_init of the fence */
-		dma_fence_put(fence);
-

  r = dma_fence_add_callback(fence, &sched_job->cb, drm_sched_job_done_cb); if (r == -ENOENT) drm_sched_job_done(sched_job, fence-

Quoted 1 line

error);

  else if (r) DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);

drivers/gpu/drm/scheduler/sched_main.c +3 -0
+
+		/*
+		 * s_fence took a new reference to fence in the

call to

drivers/gpu/drm/scheduler/sched_main.c +1 -0
+		 * drm_sched_fence_scheduled() above. The
Quoted 1 line

reference passed by

I think mentioning that in this context is a bit misleading. The reason we can
put the fence here, is because we stop using the local fence pointer we have a
reference for (from run_job()).

Quoted 7 lines

This has nothing to do with the fact that drm_sched_fence_scheduled() took its
own reference when it stored a copy of this fence pointer in a separate data
structure.

With that fixed,

Then let's remove the comment completely I'd say.

Quoted 2 lines

And I forgot your SB. Will add.

Danke,
P.

Quoted 55 lines
Quoted 55 lines
drivers/gpu/drm/scheduler/sched_main.c +1 -0
+		 * run_job() above is now not needed any longer.

Drop it.

drivers/gpu/drm/scheduler/sched_main.c +2 -0
+		 */
+		dma_fence_put(fence);

  } else {
  drm_sched_job_done(sched_job, IS_ERR(fence) ? PTR_ERR(fence) : 0);

include/drm/gpu_scheduler.h +0 -0
index 95e17504e46a..d5cd2a78f27c 100644
@@ -420,10 +420,21 @@ struct drm_sched_backend_ops {

  struct drm_sched_entity *s_entity);

  /**
-         * @run_job: Called to execute the job once all of the dependencies
-         * have been resolved.  This may be called multiple times, if

include/drm/gpu_scheduler.h +0 -1
-	 * timedout_job() has happened and

drm_sched_job_recovery()

include/drm/gpu_scheduler.h +1 -1
-	 * decides to try it again.
+	 * @run_job: Called to execute the job once all of the

dependencies

include/drm/gpu_scheduler.h +1 -0
+	 * have been resolved. This may be called multiple times,

if

include/drm/gpu_scheduler.h +1 -0
+	 * timedout_job() has happened and

drm_sched_job_recovery() decides to

include/drm/gpu_scheduler.h +5 -0
+	 * try it again.
+	 *
+	 * @sched_job: the job to run
+	 *
+	 * Returns: dma_fence the driver must signal once the

hardware has

include/drm/gpu_scheduler.h +3 -0
+	 *	completed the job ("hardware fence").
+	 *
+	 * Note that the scheduler expects to 'inherit' its own

reference to

include/drm/gpu_scheduler.h +1 -0
+	 * this fence from the callback. It does not invoke an

extra

include/drm/gpu_scheduler.h +1 -0
+	 * dma_fence_get() on it. Consequently, this callback must

take a

include/drm/gpu_scheduler.h +1 -0
+	 * reference for the scheduler, and additional ones for

the driver's

include/drm/gpu_scheduler.h +1 -0
+	 * respective needs.

  */
  struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);

Signature · 1 line
2.47.1
maintainer
#9 lore raw

Re: [PATCH 2/3] drm/sched: Adjust outdated docu for run_job()

In reply to Philipp Stanner
Quoted 21 lines · Philipp Stanner

The documentation for drm_sched_backend_ops.run_job() mentions a certain function called drm_sched_job_recovery(). This function does not exist. What's actually meant is drm_sched_resubmit_jobs(), which is by now also deprecated.

Remove the mention of the deprecated function.

Discourage the behavior of drm_sched_backend_ops.run_job() being called multiple times for the same job.

---
 include/drm/gpu_scheduler.h | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
include/drm/gpu_scheduler.h +0 -0
index d5cd2a78f27c..c4e65f9f7f22 100644
@@ -421,9 +421,12 @@ struct drm_sched_backend_ops {
 

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 428

Quoted 10 lines
include/drm/gpu_scheduler.h +5 -3
 	/**
 	 * @run_job: Called to execute the job once all of the dependencies
-	 * have been resolved. This may be called multiple times, if
-	 * timedout_job() has happened and drm_sched_job_recovery() decides to
-	 * try it again.
+	 * have been resolved.
+	 *
+	 * The deprecated drm_sched_resubmit_jobs() (called from
+	 * drm_sched_backend_ops.timedout_job()) can invoke this again with the
+	 * same parameters. Doing this is strongly discouraged because it

Maybe "invoke this again for the same job"?

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 429

Quoted 1 line
include/drm/gpu_scheduler.h +1 -0
+	 * violates dma_fence rules.

Does it? AFAIU it puts certain expectations on the driver, before a driver can call this function, which likely leads to the driver to violate dma_fence rules, right?

Maybe we should also list the exact rules that are (likely to be) violated to allow drivers to fix it at their end more easily.

Quoted 6 lines
include/drm/gpu_scheduler.h +0 -0
 	 *
 	 * @sched_job: the job to run
 	 *
Signature · 1 line
2.47.1
maintainer
#11 lore raw

Re: [PATCH 3/3] drm/sched: Update timedout_job()'s documentation

In reply to Philipp Stanner
Quoted 28 lines · Philipp Stanner

drm_sched_backend_ops.timedout_job()'s documentation is outdated. It mentions the deprecated function drm_sched_resubmit_job(). Furthermore, it does not point out the important distinction between hardware and firmware schedulers.

Since firmware schedulers tyipically only use one entity per scheduler, timeout handling is significantly more simple because the entity the faulted job came from can just be killed without affecting innocent processes.

Update the documentation with that distinction and other details.

Reformat the docstring to work to a unified style with the other handles.

---
 include/drm/gpu_scheduler.h | 83 +++++++++++++++++++++++--------------
 1 file changed, 52 insertions(+), 31 deletions(-)
include/drm/gpu_scheduler.h +0 -0
index c4e65f9f7f22..380b8840c591 100644
@@ -445,43 +445,64 @@ struct drm_sched_backend_ops {
 	 * @timedout_job: Called when a job has taken too long to execute,
 	 * to trigger GPU recovery.
 	 *

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 454

Quoted 10 lines
include/drm/gpu_scheduler.h +6 -3
-	 * This method is called in a workqueue context.
+	 * @sched_job: The job that has timed out
 	 *
-	 * Drivers typically issue a reset to recover from GPU hangs, and this
-	 * procedure usually follows the following workflow:
+	 * Returns:
+	 * - DRM_GPU_SCHED_STAT_NOMINAL, on success, i.e., if the underlying
+	 *   driver has started or completed recovery.
+	 * - DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
+	 *   available, i.e., has been unplugged.

Maybe it'd be better to document this at the enum level and add a link.

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 461

Quoted 13 lines
include/drm/gpu_scheduler.h +6 -6
 	 *
-	 * 1. Stop the scheduler using drm_sched_stop(). This will park the
-	 *    scheduler thread and cancel the timeout work, guaranteeing that
-	 *    nothing is queued while we reset the hardware queue
-	 * 2. Try to gracefully stop non-faulty jobs (optional)
-	 * 3. Issue a GPU reset (driver-specific)
-	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
+	 * Drivers typically issue a reset to recover from GPU hangs.
+	 * This procedure looks very different depending on whether a firmware
+	 * or a hardware scheduler is being used.
+	 *
+	 * For a FIRMWARE SCHEDULER, each (pseudo-)ring has one scheduler, and
+	 * each scheduler (typically) has one entity. Hence, you typically

I think you can remove the first "typically" here. I don't think that for a firmware scheduler we ever have something else than a 1:1 relation, besides that having something else than a 1:1 relation (whatever that would be) would likely be a contradiction to the statement above.

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 466

Quoted 5 lines
include/drm/gpu_scheduler.h +5 -0
+	 * follow those steps:
+	 *
+	 * 1. Stop the scheduler using drm_sched_stop(). This will pause the
+	 *    scheduler workqueues and cancel the timeout work, guaranteeing
+	 *    that nothing is queued while we reset the hardware queue.

Rather reset / remove the software queue / ring.

(Besides: I think we should also define a distinct terminology, sometimes "queue" refers to the ring buffer, sometimes it refers to the entity, etc. At least we should be consequent within this commit, and then fix the rest.)

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 470

Quoted 4 lines
include/drm/gpu_scheduler.h +4 -0
+	 * 2. Try to gracefully stop non-faulty jobs (optional).
+	 * TODO: RFC ^ Folks, should we remove this step? What does it even mean
+	 * precisely to "stop" those jobs? Is that even helpful to userspace in
+	 * any way?

I think this means to prevent unrelated queues / jobs from being impacted by the subsequent GPU reset.

So, this is likely not applicable here, see below.

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 471

Quoted 1 line
include/drm/gpu_scheduler.h +1 -0
+	 * 3. Issue a GPU reset (driver-specific).

I'm not entirely sure it really applies to all GPUs that feature a FW scheduler, but I'd expect that the FW takes care of resetting the corresponding HW (including preventing impact on other FW rings) if the faulty FW ring is removed by the driver.

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 475

Quoted 5 lines
include/drm/gpu_scheduler.h +3 -1
+	 * 4. Kill the entity the faulted job stems from, and the associated
+	 *    scheduler.
 	 * 5. Restart the scheduler using drm_sched_start(). At that point, new
-	 *    jobs can be queued, and the scheduler thread is unblocked
+	 *    jobs can be queued, and the scheduler workqueues awake again.

How can we start the scheduler again after we've killed it? I think the most likely scenario is that the FW ring (including the scheduler structures) is removed entirely and simply re-created. So, I think we can probably remove 5.

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 478

Quoted 3 lines
include/drm/gpu_scheduler.h +3 -0
+	 *
+	 * For a HARDWARE SCHEDULER, each ring also has one scheduler, but each
+	 * scheduler typically has many attached entities. This implies that you

Maybe better "associated". For the second sentence, I'd rephrase it, e.g. "This implies that all entities associated with the affected scheduler can't be torn down, because [...]".

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 487

Quoted 9 lines
include/drm/gpu_scheduler.h +9 -0
+	 * cannot tear down all entities associated with the affected scheduler,
+	 * because this would effectively also kill innocent userspace processes
+	 * which did not submit faulty jobs (for example).
+	 *
+	 * Consequently, the procedure to recover with a hardware scheduler
+	 * should look like this:
+	 *
+	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop().
+	 * 2. Figure out to which entity the faulted job belongs.

"belongs to"

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 491

Quoted 4 lines
include/drm/gpu_scheduler.h +4 -0
+	 * 3. Try to gracefully stop non-faulty jobs (optional).
+	 * TODO: RFC ^ Folks, should we remove this step? What does it even mean
+	 * precisely to "stop" those jobs? Is that even helpful to userspace in
+	 * any way?

See above.

Quoted 41 lines
include/drm/gpu_scheduler.h +11 -21
+	 * 4. Kill that entity.
+	 * 5. Issue a GPU reset on all faulty rings (driver-specific).
+	 * 6. Re-submit jobs on all schedulers impacted by re-submitting them to
+	 *    the entities which are still alive.
+	 * 7. Restart all schedulers that were stopped in step #1 using
+	 *    drm_sched_start().
 	 *
 	 * Note that some GPUs have distinct hardware queues but need to reset
 	 * the GPU globally, which requires extra synchronization between the
-	 * timeout handler of the different &drm_gpu_scheduler. One way to
-	 * achieve this synchronization is to create an ordered workqueue
-	 * (using alloc_ordered_workqueue()) at the driver level, and pass this
-	 * queue to drm_sched_init(), to guarantee that timeout handlers are
-	 * executed sequentially. The above workflow needs to be slightly
-	 * adjusted in that case:
-	 *
-	 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
-	 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
-	 *    the reset (optional)
-	 * 3. Issue a GPU reset on all faulty queues (driver-specific)
-	 * 4. Re-submit jobs on all schedulers impacted by the reset using
-	 *    drm_sched_resubmit_jobs()
-	 * 5. Restart all schedulers that were stopped in step #1 using
-	 *    drm_sched_start()
-	 *
-	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
-	 * and the underlying driver has started or completed recovery.
-	 *
-	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
-	 * available, i.e. has been unplugged.
+	 * timeout handlers of different schedulers. One way to achieve this
+	 * synchronization is to create an ordered workqueue (using
+	 * alloc_ordered_workqueue()) at the driver level, and pass this queue
+	 * as drm_sched_init()'s @timeout_wq parameter. This will guarantee
+	 * that timeout handlers are executed sequentially.
 	 */
 	enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
 
Signature · 1 line
2.47.1
#5 lore raw

Re: [PATCH 1/3] drm/sched: Document run_job() refcount hazard

In reply to Christian König
Quoted 117 lines · Christian König

Am 09.01.25 um 14:37 schrieb Philipp Stanner:

Quoted 101 lines

From: Philipp Stanner <[email protected]>

drm_sched_backend_ops.run_job() returns a dma_fence for the scheduler.
That fence is signalled by the driver once the hardware completed the
associated job. The scheduler does not increment the reference count on
that fence, but implicitly expects to inherit this fence from run_job().

This is relatively subtle and prone to misunderstandings.

This implies that, to keep a reference for itself, a driver needs to
call dma_fence_get() in addition to dma_fence_init() in that callback.

It's further complicated by the fact that the scheduler even decrements
the refcount in drm_sched_run_job_work() since it created a new reference in drm_sched_fence_scheduled(). It does, however, still use
its pointer to the fence after calling dma_fence_put() - which is safe
because of the aforementioned new reference, but actually still violates
the refcounting rules.

Improve the explanatory comment for that decrement.

Move the call to dma_fence_put() to the position behind the last usage
of the fence.

Document the necessity to increment the reference count in drm_sched_backend_ops.run_job().

---
  drivers/gpu/drm/scheduler/sched_main.c | 10 +++++++--- include/drm/gpu_scheduler.h            | 19 +++++++++++++++----

  2 files changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/scheduler/sched_main.c +0 -0
index 57da84908752..5f46c01eb01e 100644
@@ -1218,15 +1218,19 @@ static void drm_sched_run_job_work(struct
 work_struct *w)
   	drm_sched_fence_scheduled(s_fence, fence);


   if (!IS_ERR_OR_NULL(fence)) {

drivers/gpu/drm/scheduler/sched_main.c +0 -3
-		/* Drop for original kref_init of the fence */
-		dma_fence_put(fence);
-

   r = dma_fence_add_callback(fence, &sched_job->cb, drm_sched_job_done_cb); if (r == -ENOENT) drm_sched_job_done(sched_job, fence-

Quoted 1 line

error);

   else if (r) DRM_DEV_ERROR(sched->dev, "fence add callback failed (%d)\n", r);

drivers/gpu/drm/scheduler/sched_main.c +3 -0
+
+		/*
+		 * s_fence took a new reference to fence in the

call to

drivers/gpu/drm/scheduler/sched_main.c +1 -0
+		 * drm_sched_fence_scheduled() above. The

reference passed by

drivers/gpu/drm/scheduler/sched_main.c +1 -0
+		 * run_job() above is now not needed any longer.

Drop it.

drivers/gpu/drm/scheduler/sched_main.c +2 -0
+		 */
+		dma_fence_put(fence);

   } else {
   drm_sched_job_done(sched_job, IS_ERR(fence) ? PTR_ERR(fence) : 0);

include/drm/gpu_scheduler.h +0 -0
index 95e17504e46a..d5cd2a78f27c 100644
@@ -420,10 +420,21 @@ struct drm_sched_backend_ops {

   struct drm_sched_entity *s_entity);

   /**
-         * @run_job: Called to execute the job once all of the dependencies
-         * have been resolved.  This may be called multiple times, if

include/drm/gpu_scheduler.h +0 -1
-	 * timedout_job() has happened and

drm_sched_job_recovery()

include/drm/gpu_scheduler.h +1 -1
-	 * decides to try it again.
+	 * @run_job: Called to execute the job once all of the

dependencies

include/drm/gpu_scheduler.h +1 -0
+	 * have been resolved. This may be called multiple times,

if

include/drm/gpu_scheduler.h +1 -0
+	 * timedout_job() has happened and

drm_sched_job_recovery() decides to

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 426

Quoted 1 line
include/drm/gpu_scheduler.h +1 -0
+	 * try it again.

I just came to realize that this hack with calling run_job multiple times won't work any more with this patch here.

The background is that you can't allocate memory for a newly returned fence and as far as I know no driver pre allocates multiple HW fences for a job.

So at least amdgpu used to re-use the same HW fence as return over and
over again, just re-initialized the reference count. I removed that hack
from amdgpu, but just FYI it could be that other driver did the same.

Quoted 4 lines

Apart from that concern I think that this patch is really the right thing and that driver hacks relying on the order of dropping references
are fundamentally broken approaches.

I don't see how a hack couldn't work anymore with this patch. All it does is add comments, and it moves the putting of the dma_fence to where it belongs. But we're in run_job_work() here, which executes sequentially.
Drivers reusing fences or modifying the refcount might be bad and hacky, but as long as we're not racing here it doesn't matter whether the scheduler decrements the refcount a few nanoseconds later than before.

Anyways, even if I'm wrong, it seems we agree that provoking such explosions would be worth it.

P.

Quoted 28 lines

So feel free to add Reviewed-by: Christian König
<[email protected]>.

Regards,
Christian.

Quoted 20 lines
include/drm/gpu_scheduler.h +4 -0
+	 *
+	 * @sched_job: the job to run
+	 *
+	 * Returns: dma_fence the driver must signal once the

hardware has

include/drm/gpu_scheduler.h +3 -0
+	 *	completed the job ("hardware fence").
+	 *
+	 * Note that the scheduler expects to 'inherit' its own

reference to

include/drm/gpu_scheduler.h +1 -0
+	 * this fence from the callback. It does not invoke an

extra

include/drm/gpu_scheduler.h +1 -0
+	 * dma_fence_get() on it. Consequently, this callback must

take a

include/drm/gpu_scheduler.h +1 -0
+	 * reference for the scheduler, and additional ones for

the driver's

include/drm/gpu_scheduler.h +1 -0
+	 * respective needs.

   */
   struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);

#12 lore raw

Re: [PATCH 3/3] drm/sched: Update timedout_job()'s documentation

In reply to Danilo Krummrich
Quoted 92 lines · Danilo Krummrich
Quoted 45 lines · Philipp Stanner

drm_sched_backend_ops.timedout_job()'s documentation is outdated. It
mentions the deprecated function drm_sched_resubmit_job(). Furthermore,
it does not point out the important distinction between hardware and
firmware schedulers.

Since firmware schedulers tyipically only use one entity per scheduler,
timeout handling is significantly more simple because the entity the
faulted job came from can just be killed without affecting innocent processes.

Update the documentation with that distinction and other details.

Reformat the docstring to work to a unified style with the other handles.

---
 include/drm/gpu_scheduler.h | 83 +++++++++++++++++++++++----------
----

 1 file changed, 52 insertions(+), 31 deletions(-)
include/drm/gpu_scheduler.h +1 -2
index c4e65f9f7f22..380b8840c591 100644
@@ -445,43 +445,64 @@ struct drm_sched_backend_ops {
  	 * @timedout_job: Called when a job has taken too long to
 execute,
  	 * to trigger GPU recovery.
  	 *
-	 * This method is called in a workqueue context.
+	 * @sched_job: The job that has timed out
  	 *
-	 * Drivers typically issue a reset to recover from GPU

hangs, and this

include/drm/gpu_scheduler.h +2 -1
-	 * procedure usually follows the following workflow:
+	 * Returns:
+	 * - DRM_GPU_SCHED_STAT_NOMINAL, on success, i.e., if the

underlying

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 454

Quoted 3 lines
include/drm/gpu_scheduler.h +3 -0
+	 *   driver has started or completed recovery.
+	 * - DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
+	 *   available, i.e., has been unplugged.

Maybe it'd be better to document this at the enum level and add a link.

Quoted 18 lines

  *

include/drm/gpu_scheduler.h +0 -1
-	 * 1. Stop the scheduler using drm_sched_stop(). This will

park the

include/drm/gpu_scheduler.h +0 -1
-	 *    scheduler thread and cancel the timeout work,

guaranteeing that

include/drm/gpu_scheduler.h +1 -4
-	 *    nothing is queued while we reset the hardware queue
-	 * 2. Try to gracefully stop non-faulty jobs (optional)
-	 * 3. Issue a GPU reset (driver-specific)
-	 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
+	 * Drivers typically issue a reset to recover from GPU

hangs.

include/drm/gpu_scheduler.h +1 -0
+	 * This procedure looks very different depending on

whether a firmware

include/drm/gpu_scheduler.h +3 -0
+	 * or a hardware scheduler is being used.
+	 *
+	 * For a FIRMWARE SCHEDULER, each (pseudo-)ring has one

scheduler, and

include/drm/gpu_scheduler.h +1 -0
+	 * each scheduler (typically) has one entity. Hence, you
Quoted 1 line

typically

I think you can remove the first "typically" here. I don't think that for a
firmware scheduler we ever have something else than a 1:1 relation, besides that
having something else than a 1:1 relation (whatever that would be) would likely
be a contradiction to the statement above.

Quoted 8 lines
include/drm/gpu_scheduler.h +3 -0
+	 * follow those steps:
+	 *
+	 * 1. Stop the scheduler using drm_sched_stop(). This will

pause the

include/drm/gpu_scheduler.h +1 -0
+	 *    scheduler workqueues and cancel the timeout work,

guaranteeing

include/drm/gpu_scheduler.h +1 -0
+	 *    that nothing is queued while we reset the hardware

queue.

Rather reset / remove the software queue / ring.

Quoted 5 lines

(Besides: I think we should also define a distinct terminology, sometimes "queue"
refers to the ring buffer, sometimes it refers to the entity, etc. At least we
should be consequent within this commit, and then fix the rest.)

Very good idea!

How about we, from now on, always just call it "ring" or "hardware ring"?

I think queue is very generic and, as you point out, often used for the entities and stuff like that.

Quoted 14 lines

include/drm/gpu_scheduler.h

Quoted 7 lines
include/drm/gpu_scheduler.h +2 -0
+	 * 2. Try to gracefully stop non-faulty jobs (optional).
+	 * TODO: RFC ^ Folks, should we remove this step? What

does it even mean

include/drm/gpu_scheduler.h +1 -0
+	 * precisely to "stop" those jobs? Is that even helpful to

userspace in

include/drm/gpu_scheduler.h +1 -0
+	 * any way?

I think this means to prevent unrelated queues / jobs from being impacted by the
subsequent GPU reset.

So, this is likely not applicable here, see below.

Quoted 9 lines

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 471

Quoted 1 line
include/drm/gpu_scheduler.h +1 -0
+	 * 3. Issue a GPU reset (driver-specific).

I'm not entirely sure it really applies to all GPUs that feature a FW scheduler,
but I'd expect that the FW takes care of resetting the corresponding HW
(including preventing impact on other FW rings) if the faulty FW ring is removed
by the driver.

@Christian: Agree? AMD is still purely a HW scheduler afaik.

Quoted 17 lines
Quoted 10 lines
include/drm/gpu_scheduler.h +1 -0
+	 * 4. Kill the entity the faulted job stems from, and the

associated

include/drm/gpu_scheduler.h +1 -0
+	 *    scheduler.

  * 5. Restart the scheduler using drm_sched_start(). At that point, new

include/drm/gpu_scheduler.h +0 -1
-	 *    jobs can be queued, and the scheduler thread is

unblocked

include/drm/gpu_scheduler.h +1 -0
+	 *    jobs can be queued, and the scheduler workqueues

awake again.

How can we start the scheduler again after we've killed it? I think the most
likely scenario is that the FW ring (including the scheduler structures) is
removed entirely and simply re-created. So, I think we can probably remove 5.

ACK

Quoted 29 lines
Quoted 6 lines
include/drm/gpu_scheduler.h +2 -0
+	 *
+	 * For a HARDWARE SCHEDULER, each ring also has one

scheduler, but each

include/drm/gpu_scheduler.h +1 -0
+	 * scheduler typically has many attached entities. This

implies that you

Maybe better "associated". For the second sentence, I'd rephrase it, e.g. "This
implies that all entities associated with the affected scheduler can't be torn
down, because [...]".

include/drm/gpu_scheduler.h struct drm_sched_backend_ops { · line 487

Quoted 13 lines
include/drm/gpu_scheduler.h +1 -0
+	 * cannot tear down all entities associated with the

affected scheduler,

include/drm/gpu_scheduler.h +1 -0
+	 * because this would effectively also kill innocent

userspace processes

include/drm/gpu_scheduler.h +3 -0
+	 * which did not submit faulty jobs (for example).
+	 *
+	 * Consequently, the procedure to recover with a hardware

scheduler

include/drm/gpu_scheduler.h +3 -0
+	 * should look like this:
+	 *
+	 * 1. Stop all schedulers impacted by the reset using

drm_sched_stop().

include/drm/gpu_scheduler.h +1 -0
+	 * 2. Figure out to which entity the faulted job belongs.

"belongs to"

Quoted 8 lines

include/drm/gpu_scheduler.h

Quoted 6 lines
include/drm/gpu_scheduler.h +2 -0
+	 * 3. Try to gracefully stop non-faulty jobs (optional).
+	 * TODO: RFC ^ Folks, should we remove this step? What

does it even mean

include/drm/gpu_scheduler.h +1 -0
+	 * precisely to "stop" those jobs? Is that even helpful to

userspace in

include/drm/gpu_scheduler.h +1 -0
+	 * any way?

See above.

Agree with all. Will fix those in v2.

Danke,
P.

Quoted 65 lines
Quoted 64 lines
include/drm/gpu_scheduler.h +2 -0
+	 * 4. Kill that entity.
+	 * 5. Issue a GPU reset on all faulty rings (driver-

specific).

include/drm/gpu_scheduler.h +1 -0
+	 * 6. Re-submit jobs on all schedulers impacted by re-

submitting them to

include/drm/gpu_scheduler.h +2 -0
+	 *    the entities which are still alive.
+	 * 7. Restart all schedulers that were stopped in step #1

using

include/drm/gpu_scheduler.h +1 -0
+	 *    drm_sched_start().

  *
  * Note that some GPUs have distinct hardware queues but need to reset
  * the GPU globally, which requires extra synchronization between the

include/drm/gpu_scheduler.h +0 -1
-	 * timeout handler of the different &drm_gpu_scheduler.

One way to

include/drm/gpu_scheduler.h +0 -1
-	 * achieve this synchronization is to create an ordered

workqueue

include/drm/gpu_scheduler.h +0 -1
-	 * (using alloc_ordered_workqueue()) at the driver level,

and pass this

include/drm/gpu_scheduler.h +0 -1
-	 * queue to drm_sched_init(), to guarantee that timeout

handlers are

include/drm/gpu_scheduler.h +0 -1
-	 * executed sequentially. The above workflow needs to be

slightly

include/drm/gpu_scheduler.h +0 -3
-	 * adjusted in that case:
-	 *
-	 * 1. Stop all schedulers impacted by the reset using

drm_sched_stop()

include/drm/gpu_scheduler.h +0 -1
-	 * 2. Try to gracefully stop non-faulty jobs on all queues

impacted by

include/drm/gpu_scheduler.h +0 -2
-	 *    the reset (optional)
-	 * 3. Issue a GPU reset on all faulty queues (driver-

specific)

include/drm/gpu_scheduler.h +0 -1
-	 * 4. Re-submit jobs on all schedulers impacted by the

reset using

include/drm/gpu_scheduler.h +0 -2
-	 *    drm_sched_resubmit_jobs()
-	 * 5. Restart all schedulers that were stopped in step #1

using

include/drm/gpu_scheduler.h +0 -4
-	 *    drm_sched_start()
-	 *
-	 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
-	 * and the underlying driver has started or completed

recovery.

include/drm/gpu_scheduler.h +0 -2
-	 *
-	 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no

longer

include/drm/gpu_scheduler.h +1 -1
-	 * available, i.e. has been unplugged.
+	 * timeout handlers of different schedulers. One way to

achieve this

include/drm/gpu_scheduler.h +1 -0
+	 * synchronization is to create an ordered workqueue

(using

include/drm/gpu_scheduler.h +1 -0
+	 * alloc_ordered_workqueue()) at the driver level, and

pass this queue

include/drm/gpu_scheduler.h +1 -0
+	 * as drm_sched_init()'s @timeout_wq parameter. This will

guarantee

include/drm/gpu_scheduler.h +1 -0
+	 * that timeout handlers are executed sequentially.

  */
  enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);

Signature · 1 line
2.47.1