Am 16.09.19 um 16:24 schrieb Daniel Vetter:

On Mon, Sep 16, 2019 at 10:11 AM Koenig, Christian Christian.Koenig@amd.com wrote:

...

Hi Steven,

the problem seems to be than panfrost is trying to sleep while freeing a job. E.g. it tries to take a mutex.

That is not allowed any more since we need to free the jobs from atomic and even interrupt context.

Your suggestion wouldn't work because this way jobs are not freed when there isn't a new one to be scheduled.

One fix would be to make sure that any that any calls to drm_sched_cleanup_jobs are atomic, by putting preempt_disable/enable or local_irq_disable/enable in there, at least when lockdep or sleep debugging is enabled. That should help catch these reliable, instead of just once every blue moon.

Yeah, thought about that as well. But I would also prefer a solution where drm_sched_cleanup_jobs() is never called in an atomic context.

The problem is that I don't see how that can be possible without delaying freeing of jobs.

Regards, Christian.

-Daniel

...

Regards, Christian.

Am 13.09.19 um 16:50 schrieb Steven Price:

...

Hi,

I hit the below splat randomly with panfrost. From what I can tell this is a more general issue which would affect other drivers.

----8<----- [58604.913130] ------------[ cut here ]------------ [58604.918590] WARNING: CPU: 1 PID: 1758 at kernel/sched/core.c:6556 __might_sleep+0x74/0x98 [58604.927965] do not call blocking ops when !TASK_RUNNING; state=1 set at [<0c590494>] prepare_to_wait_event+0x104/0x164 [58604.940047] Modules linked in: panfrost gpu_sched [58604.945370] CPU: 1 PID: 1758 Comm: pan_js Not tainted 5.3.0-rc1+ #13 [58604.952500] Hardware name: Rockchip (Device Tree) [58604.957815] [<c0111150>] (unwind_backtrace) from [<c010c99c>] (show_stack+0x10/0x14) [58604.966521] [<c010c99c>] (show_stack) from [<c07adbb4>] (dump_stack+0x9c/0xd4) [58604.974639] [<c07adbb4>] (dump_stack) from [<c0121da8>] (__warn+0xe8/0x104) [58604.982462] [<c0121da8>] (__warn) from [<c0121e08>] (warn_slowpath_fmt+0x44/0x6c) [58604.990867] [<c0121e08>] (warn_slowpath_fmt) from [<c014eccc>] (__might_sleep+0x74/0x98) [58604.999973] [<c014eccc>] (__might_sleep) from [<c07c73d8>] (__mutex_lock+0x38/0x948) [58605.008690] [<c07c73d8>] (__mutex_lock) from [<c07c7d00>] (mutex_lock_nested+0x18/0x20) [58605.017841] [<c07c7d00>] (mutex_lock_nested) from [<bf00b54c>] (panfrost_gem_free_object+0x60/0x10c [panfrost]) [58605.029430] [<bf00b54c>] (panfrost_gem_free_object [panfrost]) from [<bf00cecc>] (panfrost_job_put+0x138/0x150 [panfrost]) [58605.042076] [<bf00cecc>] (panfrost_job_put [panfrost]) from [<bf00121c>] (drm_sched_cleanup_jobs+0xc8/0xe0 [gpu_sched]) [58605.054417] [<bf00121c>] (drm_sched_cleanup_jobs [gpu_sched]) from [<bf001300>] (drm_sched_main+0xcc/0x26c [gpu_sched]) [58605.066620] [<bf001300>] (drm_sched_main [gpu_sched]) from [<c0146cfc>] (kthread+0x13c/0x154) [58605.076226] [<c0146cfc>] (kthread) from [<c01010b4>] (ret_from_fork+0x14/0x20) [58605.084346] Exception stack(0xe959bfb0 to 0xe959bff8) [58605.090046] bfa0: 00000000 00000000 00000000 00000000 [58605.099250] bfc0: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 [58605.108480] bfe0: 00000000 00000000 00000000 00000000 00000013 00000000 [58605.116210] irq event stamp: 179 [58605.119955] hardirqs last enabled at (187): [<c017f7e4>] console_unlock+0x564/0x5c4 [58605.128935] hardirqs last disabled at (202): [<c017f308>] console_unlock+0x88/0x5c4 [58605.137788] softirqs last enabled at (216): [<c0102334>] __do_softirq+0x18c/0x548 [58605.146543] softirqs last disabled at (227): [<c0129528>] irq_exit+0xc4/0x10c [58605.154618] ---[ end trace f65bdbd9ea9adfc0 ]--- ----8<-----

The problem is that drm_sched_main() calls drm_sched_cleanup_jobs() as part of the condition of wait_event_interruptible:

...

          wait_event_interruptible(sched->wake_up_worker,
                                   (drm_sched_cleanup_jobs(sched),
                                   (!drm_sched_blocked(sched) &&
                                    (entity = drm_sched_select_entity(sched))) ||
                                   kthread_should_stop()));

When drm_sched_cleanup_jobs() is called *after* a wait (i.e. after prepare_to_wait_event() has been called), then any might_sleep() will moan loudly about it. This doesn't seem to happen often (I've only triggered it once) because usually drm_sched_cleanup_jobs() either doesn't sleep or does the sleeping during the first call that wait_event_interruptible() makes (which is before the task state is set).

I don't really understand why drm_sched_cleanup_jobs() needs to be called here, a simple change like below 'fixes' it. But I presume there's some reason for the call being part of the wait_event_interruptible condition. Can anyone shed light on this?

The code was introduced in commit 5918045c4ed4 ("drm/scheduler: rework job destruction")

Steve

----8<----- diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 9a0ee74d82dc..528f295e3a31 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -699,11 +699,12 @@ static int drm_sched_main(void *param) struct drm_sched_job *sched_job; struct dma_fence *fence;

•         drm_sched_cleanup_jobs(sched);
  

•          wait_event_interruptible(sched->wake_up_worker,
  

•                                  (drm_sched_cleanup_jobs(sched),
                                    (!drm_sched_blocked(sched) &&
                                     (entity = drm_sched_select_entity(sched))) ||
  

•                                  kthread_should_stop()));
  

•                                  kthread_should_stop());
  
           if (!entity)
                   continue;
  

---

dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel

On 17/09/2019 09:42, Koenig, Christian wrote:

Hi Steven,

thought about that issue a bit more and I think I came up with a solution.

What you could do is to split up drm_sched_cleanup_jobs() into two functions.

One that checks if jobs to be cleaned up are present and one which does the actual cleanup.

This way we could call drm_sched_cleanup_jobs() outside of the wait_event_interruptible().

Yes that seems like a good solution - there doesn't seem to be a good reason why the actual job cleanup needs to be done within the wait_event_interruptible() condition. I did briefly attempt that before, but I couldn't work out exactly what the condition is which should cause the wake (my initial attempt caused continuous wake-ups).

I'm not back in the office until Monday, but I'll have another go then at spitting the checking and the actual freeing of the jobs.

Thanks,

Steve

Regards, Christian.

Am 13.09.19 um 16:50 schrieb Steven Price:

...

Hi,

I hit the below splat randomly with panfrost. From what I can tell this is a more general issue which would affect other drivers.

----8<----- [58604.913130] ------------[ cut here ]------------ [58604.918590] WARNING: CPU: 1 PID: 1758 at kernel/sched/core.c:6556 __might_sleep+0x74/0x98 [58604.927965] do not call blocking ops when !TASK_RUNNING; state=1 set at [<0c590494>] prepare_to_wait_event+0x104/0x164 [58604.940047] Modules linked in: panfrost gpu_sched [58604.945370] CPU: 1 PID: 1758 Comm: pan_js Not tainted 5.3.0-rc1+ #13 [58604.952500] Hardware name: Rockchip (Device Tree) [58604.957815] [<c0111150>] (unwind_backtrace) from [<c010c99c>] (show_stack+0x10/0x14) [58604.966521] [<c010c99c>] (show_stack) from [<c07adbb4>] (dump_stack+0x9c/0xd4) [58604.974639] [<c07adbb4>] (dump_stack) from [<c0121da8>] (__warn+0xe8/0x104) [58604.982462] [<c0121da8>] (__warn) from [<c0121e08>] (warn_slowpath_fmt+0x44/0x6c) [58604.990867] [<c0121e08>] (warn_slowpath_fmt) from [<c014eccc>] (__might_sleep+0x74/0x98) [58604.999973] [<c014eccc>] (__might_sleep) from [<c07c73d8>] (__mutex_lock+0x38/0x948) [58605.008690] [<c07c73d8>] (__mutex_lock) from [<c07c7d00>] (mutex_lock_nested+0x18/0x20) [58605.017841] [<c07c7d00>] (mutex_lock_nested) from [<bf00b54c>] (panfrost_gem_free_object+0x60/0x10c [panfrost]) [58605.029430] [<bf00b54c>] (panfrost_gem_free_object [panfrost]) from [<bf00cecc>] (panfrost_job_put+0x138/0x150 [panfrost]) [58605.042076] [<bf00cecc>] (panfrost_job_put [panfrost]) from [<bf00121c>] (drm_sched_cleanup_jobs+0xc8/0xe0 [gpu_sched]) [58605.054417] [<bf00121c>] (drm_sched_cleanup_jobs [gpu_sched]) from [<bf001300>] (drm_sched_main+0xcc/0x26c [gpu_sched]) [58605.066620] [<bf001300>] (drm_sched_main [gpu_sched]) from [<c0146cfc>] (kthread+0x13c/0x154) [58605.076226] [<c0146cfc>] (kthread) from [<c01010b4>] (ret_from_fork+0x14/0x20) [58605.084346] Exception stack(0xe959bfb0 to 0xe959bff8) [58605.090046] bfa0: 00000000 00000000 00000000 00000000 [58605.099250] bfc0: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 [58605.108480] bfe0: 00000000 00000000 00000000 00000000 00000013 00000000 [58605.116210] irq event stamp: 179 [58605.119955] hardirqs last enabled at (187): [<c017f7e4>] console_unlock+0x564/0x5c4 [58605.128935] hardirqs last disabled at (202): [<c017f308>] console_unlock+0x88/0x5c4 [58605.137788] softirqs last enabled at (216): [<c0102334>] __do_softirq+0x18c/0x548 [58605.146543] softirqs last disabled at (227): [<c0129528>] irq_exit+0xc4/0x10c [58605.154618] ---[ end trace f65bdbd9ea9adfc0 ]--- ----8<-----

The problem is that drm_sched_main() calls drm_sched_cleanup_jobs() as part of the condition of wait_event_interruptible:

...

wait_event_interruptible(sched->wake_up_worker,
			 (drm_sched_cleanup_jobs(sched),
			 (!drm_sched_blocked(sched) &&
			  (entity = drm_sched_select_entity(sched))) ||
			 kthread_should_stop()));

When drm_sched_cleanup_jobs() is called *after* a wait (i.e. after prepare_to_wait_event() has been called), then any might_sleep() will moan loudly about it. This doesn't seem to happen often (I've only triggered it once) because usually drm_sched_cleanup_jobs() either doesn't sleep or does the sleeping during the first call that wait_event_interruptible() makes (which is before the task state is set).

I don't really understand why drm_sched_cleanup_jobs() needs to be called here, a simple change like below 'fixes' it. But I presume there's some reason for the call being part of the wait_event_interruptible condition. Can anyone shed light on this?

The code was introduced in commit 5918045c4ed4 ("drm/scheduler: rework job destruction")

Steve

----8<----- diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 9a0ee74d82dc..528f295e3a31 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -699,11 +699,12 @@ static int drm_sched_main(void *param) struct drm_sched_job *sched_job; struct dma_fence *fence;

• drm_sched_cleanup_jobs(sched);
  

• wait_event_interruptible(sched->wake_up_worker,
  

• 			 (drm_sched_cleanup_jobs(sched),
  			 (!drm_sched_blocked(sched) &&
  			  (entity = drm_sched_select_entity(sched))) ||
  

• 			 kthread_should_stop()));
  

• 			 kthread_should_stop());
  
  if (!entity)
  	continue;
  

---

dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel

On 24/09/2019 10:55, Koenig, Christian wrote:

Sorry for the delayed response, have been busy on other stuff last week.

Am 17.09.19 um 14:46 schrieb Steven Price:

...

On 17/09/2019 09:42, Koenig, Christian wrote:

...

Hi Steven,

thought about that issue a bit more and I think I came up with a solution.

What you could do is to split up drm_sched_cleanup_jobs() into two functions.

One that checks if jobs to be cleaned up are present and one which does the actual cleanup.

This way we could call drm_sched_cleanup_jobs() outside of the wait_event_interruptible().

Yes that seems like a good solution - there doesn't seem to be a good reason why the actual job cleanup needs to be done within the wait_event_interruptible() condition. I did briefly attempt that before, but I couldn't work out exactly what the condition is which should cause the wake (my initial attempt caused continuous wake-ups).

Basically you need something like the following:

1. Test is timeout worker is running:

if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&     !cancel_delayed_work(&sched->work_tdr))         return false;

2. Test if there is any job ready to be cleaned up.

job = list_first_entry_or_null(&sched->ring_mirror_list, struct drm_sched_job, node); if (!job || !dma_fence_is_signaled(&job->s_fence->finished))     return false;

That should basically do it.

Thanks for the pointers. I wasn't sure if the "queue timeout for next job" part was necessary or not if step 2 above returns false.

I've been testing the following patch which simply pulls the sched->ops->free_job() out of the wait_event_interruptible().

I'll try with just the tests you've described.

----8<-----

From 873c1816394beee72904e64aa2ee0f169e768d76 Mon Sep 17 00:00:00 2001

From: Steven Price steven.price@arm.com Date: Mon, 23 Sep 2019 11:08:50 +0100 Subject: [PATCH] drm: Don't free jobs in wait_event_interruptible()

drm_sched_cleanup_jobs() attempts to free finished jobs, however because it is called as the condition of wait_event_interruptible() it must not sleep. Unfortuantly some free callbacks (notibly for Panfrost) do sleep.

Instead let's rename drm_sched_cleanup_jobs() to drm_sched_get_cleanup_job() and simply return a job for processing if there is one. The caller can then call the free_job() callback outside the wait_event_interruptible() where sleeping is possible before re-checking and returning to sleep if necessary.

Signed-off-by: Steven Price steven.price@arm.com --- drivers/gpu/drm/scheduler/sched_main.c | 22 +++++++++++++++------- 1 file changed, 15 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 9a0ee74d82dc..bf9b4931ddfd 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -622,20 +622,21 @@ static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb) }

/** - * drm_sched_cleanup_jobs - destroy finished jobs + * drm_sched_get_cleanup_job - fetch the next finished job to be destroyed * * @sched: scheduler instance * - * Remove all finished jobs from the mirror list and destroy them. + * Returns the next finished job from the mirror list (if there is one) + * ready for it to be destroyed. */ -static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched) +static struct drm_sched_job *drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched) { unsigned long flags;

/* Don't destroy jobs while the timeout worker is running */ if (sched->timeout != MAX_SCHEDULE_TIMEOUT && !cancel_delayed_work(&sched->work_tdr)) - return; + return NULL;

while (!list_empty(&sched->ring_mirror_list)) { @@ -651,7 +652,7 @@ static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched) list_del_init(&job->node); spin_unlock_irqrestore(&sched->job_list_lock, flags);

- sched->ops->free_job(job); + return job; }

/* queue timeout for next job */ @@ -659,6 +660,7 @@ static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched) drm_sched_start_timeout(sched); spin_unlock_irqrestore(&sched->job_list_lock, flags);

+ return NULL; }

/** @@ -698,12 +700,18 @@ static int drm_sched_main(void *param) struct drm_sched_fence *s_fence; struct drm_sched_job *sched_job; struct dma_fence *fence; + struct drm_sched_job *cleanup_job = NULL;

wait_event_interruptible(sched->wake_up_worker, - (drm_sched_cleanup_jobs(sched), + (cleanup_job = drm_sched_get_cleanup_job(sched)) || (!drm_sched_blocked(sched) && (entity = drm_sched_select_entity(sched))) || - kthread_should_stop())); + kthread_should_stop()); + + while (cleanup_job) { + sched->ops->free_job(cleanup_job); + cleanup_job = drm_sched_get_cleanup_job(sched); + }

if (!entity) continue;

On 25/09/2019 15:12, Koenig, Christian wrote:

Am 25.09.19 um 16:06 schrieb Steven Price:

...

On 24/09/2019 10:55, Koenig, Christian wrote:

...

Sorry for the delayed response, have been busy on other stuff last week.

Am 17.09.19 um 14:46 schrieb Steven Price:

...

On 17/09/2019 09:42, Koenig, Christian wrote:

...

Hi Steven,

thought about that issue a bit more and I think I came up with a solution.

What you could do is to split up drm_sched_cleanup_jobs() into two functions.

One that checks if jobs to be cleaned up are present and one which does the actual cleanup.

This way we could call drm_sched_cleanup_jobs() outside of the wait_event_interruptible().

Yes that seems like a good solution - there doesn't seem to be a good reason why the actual job cleanup needs to be done within the wait_event_interruptible() condition. I did briefly attempt that before, but I couldn't work out exactly what the condition is which should cause the wake (my initial attempt caused continuous wake-ups).

Basically you need something like the following:

1. Test is timeout worker is running:

if (sched->timeout != MAX_SCHEDULE_TIMEOUT &&     !cancel_delayed_work(&sched->work_tdr))         return false;

2. Test if there is any job ready to be cleaned up.

job = list_first_entry_or_null(&sched->ring_mirror_list, struct drm_sched_job, node); if (!job || !dma_fence_is_signaled(&job->s_fence->finished))     return false;

That should basically do it.

Thanks for the pointers. I wasn't sure if the "queue timeout for next job" part was necessary or not if step 2 above returns false.

I've been testing the following patch which simply pulls the sched->ops->free_job() out of the wait_event_interruptible().

I'll try with just the tests you've described.

It looks like it is necessary to queue the timeout for the next job even if there isn't a job to be cleaned up (i.e. even if we wouldn't exit from wait_event_interruptible(). So I'll go with a patch like below.

...

----8<----- From 873c1816394beee72904e64aa2ee0f169e768d76 Mon Sep 17 00:00:00 2001 From: Steven Price steven.price@arm.com Date: Mon, 23 Sep 2019 11:08:50 +0100 Subject: [PATCH] drm: Don't free jobs in wait_event_interruptible()

drm_sched_cleanup_jobs() attempts to free finished jobs, however because it is called as the condition of wait_event_interruptible() it must not sleep. Unfortuantly some free callbacks (notibly for Panfrost) do sleep.

Instead let's rename drm_sched_cleanup_jobs() to drm_sched_get_cleanup_job() and simply return a job for processing if there is one. The caller can then call the free_job() callback outside the wait_event_interruptible() where sleeping is possible before re-checking and returning to sleep if necessary.

Signed-off-by: Steven Price steven.price@arm.com

drivers/gpu/drm/scheduler/sched_main.c | 22 +++++++++++++++------- 1 file changed, 15 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 9a0ee74d82dc..bf9b4931ddfd 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -622,20 +622,21 @@ static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb) }

/**

• • drm_sched_cleanup_jobs - destroy finished jobs

• • drm_sched_get_cleanup_job - fetch the next finished job to be destroyed
  • @sched: scheduler instance

• • Remove all finished jobs from the mirror list and destroy them.

• • Returns the next finished job from the mirror list (if there is one)
• • ready for it to be destroyed.
  */

-static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched) +static struct drm_sched_job *drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched) { unsigned long flags;

/* Don't destroy jobs while the timeout worker is running */ if (sched->timeout != MAX_SCHEDULE_TIMEOUT && !cancel_delayed_work(&sched->work_tdr))

• return;
  

• return NULL;
  
  
  

  while (!list_empty(&sched->ring_mirror_list)) {

Yeah, you should probably clean that up a bit here, but apart from that this should work as well.

Good point - this function can be tidied up quite a bit, I'll post a new patch shortly.

Thanks,

Steve

Regards, Christian.

...

@@ -651,7 +652,7 @@ static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched) list_del_init(&job->node); spin_unlock_irqrestore(&sched->job_list_lock, flags);

• sched->ops->free_job(job);
  

• return job;
  

  }

  /* queue timeout for next job */

@@ -659,6 +660,7 @@ static void drm_sched_cleanup_jobs(struct drm_gpu_scheduler *sched) drm_sched_start_timeout(sched); spin_unlock_irqrestore(&sched->job_list_lock, flags);

• return NULL; }

  /**

@@ -698,12 +700,18 @@ static int drm_sched_main(void *param) struct drm_sched_fence *s_fence; struct drm_sched_job *sched_job; struct dma_fence *fence;

• struct drm_sched_job *cleanup_job = NULL;
  
  

  wait_event_interruptible(sched->wake_up_worker,

• 			 (drm_sched_cleanup_jobs(sched),
  

• 			 (cleanup_job = drm_sched_get_cleanup_job(sched)) ||
  		 (!drm_sched_blocked(sched) &&
  		  (entity = drm_sched_select_entity(sched))) ||
  

• 			 kthread_should_stop()));
  

• 			 kthread_should_stop());
  

• while (cleanup_job) {
  

• 	sched->ops->free_job(cleanup_job);
  

• 	cleanup_job = drm_sched_get_cleanup_job(sched);
  

• }
  
  

  if (!entity) continue;

---

dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel