Can we please have some actual commit message here, with detailed explanation of the race/bug/whatever, how you fix it and why this is the best option?

On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote:

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);

/* Protects against concurrent deletion in drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

Iow it's probably buggy, and an just unconditionally parking the kthread is probably the right thing to do. If it's not the right thing to do, there's a bug here for sure. -Daniel

• kthread_park(sched->thread);
  

• spin_lock(&sched->job_list_lock); job = list_first_entry_or_null(&sched->pending_list, struct drm_sched_job, list);

  if (job) {

• /*
  

•  * Remove the bad job so it cannot be freed by concurrent
  

•  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•  * is parked at which point it's safe.
  

•  */
  

• list_del_init(&job->list);
  

  spin_unlock(&sched->job_list_lock);

• /* vendor's timeout_job should call drm_sched_start() */
  

  status = job->sched->ops->timedout_job(job);

  /*

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread);

/*

• * Reinsert back the bad job here - now it's safe as
  

• * drm_sched_get_cleanup_job cannot race against us and release the
  

• * bad job at this point - we parked (waited for) any in progress
  

• * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
  

• * now until the scheduler thread is unparked.
  

• */
  

• if (bad && bad->sched == sched)

• /*
  

•  * Add at the head of the queue to reflect it was the earliest
  

•  * job extracted.
  

•  */
  

• list_add(&bad->list, &sched->pending_list);
  

• /*
  • Iterate the job list from later to earlier one and either deactive
  • their HW callbacks or remove them from pending list if they already
  • signaled.

-- 2.7.4

[AMD Official Use Only]

...

Also why don't we reuse the function drivers already have to stop a scheduler thread? We seem to have two kthread_park now, that's probably one too much.

Are you referring to drm_sched_stop ?

That's different, we don't need the logic from it, see that it go through pending list and remove all callbacks , etc... meanwhile vendor's timeout callback will call drm_sched_stop in a proper way, All we want in my patch is to simply park scheduler, Besides, even you call drm_sched_stop in job_timeout you still run into the warning issue I hit.

Thanks

------------------------------------------ Monk Liu | Cloud-GPU Core team ------------------------------------------

-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, August 31, 2021 9:02 PM To: Liu, Monk Monk.Liu@amd.com Cc: amd-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; Chen, Jingwen Jingwen.Chen@amd.com Subject: Re: [PATCH 2/2] drm/sched: serialize job_timeout and scheduler

On Tue, Aug 31, 2021 at 02:59:02PM +0200, Daniel Vetter wrote:

Can we please have some actual commit message here, with detailed explanation of the race/bug/whatever, how you fix it and why this is the best option?

On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote:

...

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);

/* Protects against concurrent deletion in drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

Iow it's probably buggy, and an just unconditionally parking the kthread is probably the right thing to do. If it's not the right thing to do, there's a bug here for sure.

Also why don't we reuse the function drivers already have to stop a scheduler thread? We seem to have two kthread_park now, that's probably one too much. -Daniel

...

• kthread_park(sched->thread);
  

• spin_lock(&sched->job_list_lock); job = list_first_entry_or_null(&sched->pending_list, struct drm_sched_job, list);

  if (job) {

• /*
  

•  * Remove the bad job so it cannot be freed by concurrent
  

•  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•  * is parked at which point it's safe.
  

•  */
  

• list_del_init(&job->list);
  

  spin_unlock(&sched->job_list_lock);

• /* vendor's timeout_job should call drm_sched_start() */
  

  status = job->sched->ops->timedout_job(job);

  /*

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread);

/*

• * Reinsert back the bad job here - now it's safe as
  

• * drm_sched_get_cleanup_job cannot race against us and release the
  

• * bad job at this point - we parked (waited for) any in progress
  

• * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
  

• * now until the scheduler thread is unparked.
  

• */
  

• if (bad && bad->sched == sched)

• /*
  

•  * Add at the head of the queue to reflect it was the earliest
  

•  * job extracted.
  

•  */
  

• list_add(&bad->list, &sched->pending_list);
  

• /*
  • Iterate the job list from later to earlier one and either deactive
  • their HW callbacks or remove them from pending list if they already
  • signaled.

-- 2.7.4

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog. ffwll.ch%2F&data=04%7C01%7CMonk.Liu%40amd.com%7C298815bea18f4fbf76 b308d96c7f7a8b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C6376601170 51194614%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiL CJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=QzgCU7%2BPdA0aWL5%2BJLg KeKbGaMMGqeGI9KE0P0LXlN4%3D&reserved=0

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll....

I stumbled across this thread when I ran into the same issue, while working out how to move drm/msm to use scheduler's retire + timeout/recovery (and get rid of our own mirror list of in-flight jobs). We already have hw error detection enabled, and it can signal quite fast, so assuming the first job on the list is the guilty job just won't work.

But I was considering a slightly different approach to fixing this, instead just handling it all in drm_sched_main() and getting rid of the complicated kthread parking gymnastics. Ie. something along the lines of:

--------------------- diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 67382621b429..4d6ce775c316 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -764,6 +764,45 @@ static bool drm_sched_blocked(struct drm_gpu_scheduler *sched) return false; }

+static bool handle_timeout(struct drm_gpu_scheduler *sched) +{ + struct drm_sched_job *bad; + + if (!sched->has_timeout) + return false; + + sched->has_timeout = false; + + spin_lock(&sched->job_list_lock); + bad = list_first_entry_or_null(&sched->pending_list, + struct drm_sched_job, list); + + if (!bad) { + spin_unlock(&sched->job_list_lock); + return false; + } + + spin_unlock(&sched->job_list_lock); + + if (sched->timeout_wq == system_wq) { + /* + * If driver has no specific requirements about serializing + * reset wrt. other engines, just call timedout_job() directly + */ + sched->ops->timedout_job(job); + } else { + /* + * Otherwise queue it on timeout_wq and wait for it to complete + */ + ... more typing needed here ... + } + + if (sched->free_guilty) { + sched->ops->free_job(job); + sched->free_guilty = false; + } +} + /** * drm_sched_main - main scheduler thread * @@ -787,6 +826,7 @@ static int drm_sched_main(void *param)

wait_event_interruptible(sched->wake_up_worker, (cleanup_job = drm_sched_get_cleanup_job(sched)) || + handle_timeout(sched) || (!drm_sched_blocked(sched) && (entity = drm_sched_select_entity(sched))) || kthread_should_stop()); ---------------------

drm_sched_fault() and the sw timeout handler would just set sched->has_timeout and kick sched->wake_up_worker.

And since we handle the timeout case after drm_sched_get_cleanup_job(), we know that all of the successfully completed jobs have already been popped off the list, and won't be unfairly maligned.

BR, -R

On Tue, Aug 31, 2021 at 6:29 PM Liu, Monk Monk.Liu@amd.com wrote:

[AMD Official Use Only]

Okay, I will reprepare this patch

Thanks

---

Monk Liu | Cloud-GPU Core team

-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, August 31, 2021 9:02 PM To: Liu, Monk Monk.Liu@amd.com Cc: amd-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; Chen, Jingwen Jingwen.Chen@amd.com Subject: Re: [PATCH 2/2] drm/sched: serialize job_timeout and scheduler

On Tue, Aug 31, 2021 at 02:59:02PM +0200, Daniel Vetter wrote:

...

Can we please have some actual commit message here, with detailed explanation of the race/bug/whatever, how you fix it and why this is the best option?

On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote:

...

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);

/* Protects against concurrent deletion in

drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

Iow it's probably buggy, and an just unconditionally parking the kthread is probably the right thing to do. If it's not the right thing to do, there's a bug here for sure.

Also why don't we reuse the function drivers already have to stop a scheduler thread? We seem to have two kthread_park now, that's probably one too much. -Daniel

...

...

•       kthread_park(sched->thread);
  

• spin_lock(&sched->job_list_lock); job = list_first_entry_or_null(&sched->pending_list, struct drm_sched_job, list);

  if (job) {

•       /*
  

•        * Remove the bad job so it cannot be freed by concurrent
  

•        * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•        * is parked at which point it's safe.
  

•        */
  

•       list_del_init(&job->list);
        spin_unlock(&sched->job_list_lock);
  
  

•       /* vendor's timeout_job should call drm_sched_start() */
        status = job->sched->ops->timedout_job(job);
  
        /*
  

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread);

/*

• • Reinsert back the bad job here - now it's safe as
• • drm_sched_get_cleanup_job cannot race against us and release the
• • bad job at this point - we parked (waited for) any in progress
• • (earlier) cleanups and drm_sched_get_cleanup_job will not be called
• • now until the scheduler thread is unparked.
• */
• if (bad && bad->sched == sched)

•       /*
  

•        * Add at the head of the queue to reflect it was the earliest
  

•        * job extracted.
  

•        */
  

•       list_add(&bad->list, &sched->pending_list);
  

• /*
  • Iterate the job list from later to earlier one and either deactive
  • their HW callbacks or remove them from pending list if they already
  • signaled.

-- 2.7.4

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog. ffwll.ch%2F&data=04%7C01%7CMonk.Liu%40amd.com%7C298815bea18f4fbf76 b308d96c7f7a8b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C6376601170 51194614%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiL CJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=QzgCU7%2BPdA0aWL5%2BJLg KeKbGaMMGqeGI9KE0P0LXlN4%3D&reserved=0

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll....

On Tue, Nov 9, 2021 at 1:07 AM Daniel Vetter daniel@ffwll.ch wrote:

On Mon, Nov 08, 2021 at 03:39:17PM -0800, Rob Clark wrote:

...

I stumbled across this thread when I ran into the same issue, while working out how to move drm/msm to use scheduler's retire + timeout/recovery (and get rid of our own mirror list of in-flight jobs). We already have hw error detection enabled, and it can signal quite fast, so assuming the first job on the list is the guilty job just won't work.

But I was considering a slightly different approach to fixing this, instead just handling it all in drm_sched_main() and getting rid of the complicated kthread parking gymnastics. Ie. something along the lines of:

So handling timeouts in the main sched thread wont work as soon as you have multiple engines and reset that impacts across engines:

• Nothing is simplified since you still need to stop the other scheduler threads.

• You get deadlocks if 2 schedulers time out at the same time, and both want to stop the other one.

Hence workqueue. Now the rule for the wq is that you can only have one per reset domain, so

• single engine you just take the one drm/sched provides
• if reset affects all your engines in the chip, then you allocate on in the drm_device and pass that to all
• if you have a complex of gpus all interconnected (e.g. xgmi hive for amd), then it's one wq for the entire hive

_All_ reset related things must be run on that workqueue or things breaks, which means if you get hw fault that also needs to be run there. I guess we should either patch drm/sched to check you call that function from the right workqueue, or just handle it internally.

Hmm, ok.. I guess it would be useful to better document the reasoning for the current design, that would have steered me more towards the approach taken in this patch.

BR, -R

-Daniel

...

---

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 67382621b429..4d6ce775c316 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -764,6 +764,45 @@ static bool drm_sched_blocked(struct drm_gpu_scheduler *sched) return false; }

+static bool handle_timeout(struct drm_gpu_scheduler *sched) +{

•   struct drm_sched_job *bad;
  

•   if (!sched->has_timeout)
  

•           return false;
  

•   sched->has_timeout = false;
  

•   spin_lock(&sched->job_list_lock);
  

•   bad = list_first_entry_or_null(&sched->pending_list,
  

•                                  struct drm_sched_job, list);
  

•   if (!bad) {
  

•           spin_unlock(&sched->job_list_lock);
  

•           return false;
  

•   }
  

•   spin_unlock(&sched->job_list_lock);
  

•   if (sched->timeout_wq == system_wq) {
  

•           /*
  

•            * If driver has no specific requirements about serializing
  

•            * reset wrt. other engines, just call timedout_job() directly
  

•            */
  

•           sched->ops->timedout_job(job);
  

•   } else {
  

•           /*
  

•            * Otherwise queue it on timeout_wq and wait for it to complete
  

•            */
  

•           ... more typing needed here ...
  

•   }
  

•   if (sched->free_guilty) {
  

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

•           sched->free_guilty = false;
  

•   }
  

+}

/**

• drm_sched_main - main scheduler thread

@@ -787,6 +826,7 @@ static int drm_sched_main(void *param)

            wait_event_interruptible(sched->wake_up_worker,
                                     (cleanup_job =

drm_sched_get_cleanup_job(sched)) ||

•                                    handle_timeout(sched) ||
                                     (!drm_sched_blocked(sched) &&
                                      (entity =
  

drm_sched_select_entity(sched))) || kthread_should_stop());

---

drm_sched_fault() and the sw timeout handler would just set sched->has_timeout and kick sched->wake_up_worker.

And since we handle the timeout case after drm_sched_get_cleanup_job(), we know that all of the successfully completed jobs have already been popped off the list, and won't be unfairly maligned.

BR, -R

On Tue, Aug 31, 2021 at 6:29 PM Liu, Monk Monk.Liu@amd.com wrote:

...

[AMD Official Use Only]

Okay, I will reprepare this patch

Thanks

---

Monk Liu | Cloud-GPU Core team

-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, August 31, 2021 9:02 PM To: Liu, Monk Monk.Liu@amd.com Cc: amd-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; Chen, Jingwen Jingwen.Chen@amd.com Subject: Re: [PATCH 2/2] drm/sched: serialize job_timeout and scheduler

On Tue, Aug 31, 2021 at 02:59:02PM +0200, Daniel Vetter wrote:

...

Can we please have some actual commit message here, with detailed explanation of the race/bug/whatever, how you fix it and why this is the best option?

On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote:

...

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);

/* Protects against concurrent deletion in

drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

Iow it's probably buggy, and an just unconditionally parking the kthread is probably the right thing to do. If it's not the right thing to do, there's a bug here for sure.

Also why don't we reuse the function drivers already have to stop a scheduler thread? We seem to have two kthread_park now, that's probably one too much. -Daniel

...

...

•       kthread_park(sched->thread);
  

• spin_lock(&sched->job_list_lock); job = list_first_entry_or_null(&sched->pending_list, struct drm_sched_job, list);

  if (job) {

•       /*
  

•        * Remove the bad job so it cannot be freed by concurrent
  

•        * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•        * is parked at which point it's safe.
  

•        */
  

•       list_del_init(&job->list);
        spin_unlock(&sched->job_list_lock);
  
  

•       /* vendor's timeout_job should call drm_sched_start() */
        status = job->sched->ops->timedout_job(job);
  
        /*
  

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread);

/*

• • Reinsert back the bad job here - now it's safe as
• • drm_sched_get_cleanup_job cannot race against us and release the
• • bad job at this point - we parked (waited for) any in progress
• • (earlier) cleanups and drm_sched_get_cleanup_job will not be called
• • now until the scheduler thread is unparked.
• */
• if (bad && bad->sched == sched)

•       /*
  

•        * Add at the head of the queue to reflect it was the earliest
  

•        * job extracted.
  

•        */
  

•       list_add(&bad->list, &sched->pending_list);
  

• /*
  • Iterate the job list from later to earlier one and either deactive
  • their HW callbacks or remove them from pending list if they already
  • signaled.

-- 2.7.4

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog. ffwll.ch%2F&data=04%7C01%7CMonk.Liu%40amd.com%7C298815bea18f4fbf76 b308d96c7f7a8b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C6376601170 51194614%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiL CJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=QzgCU7%2BPdA0aWL5%2BJLg KeKbGaMMGqeGI9KE0P0LXlN4%3D&reserved=0

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll....

-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

Am 10.11.21 um 10:50 schrieb Daniel Vetter:

On Tue, Nov 09, 2021 at 08:17:01AM -0800, Rob Clark wrote:

...

On Tue, Nov 9, 2021 at 1:07 AM Daniel Vetter daniel@ffwll.ch wrote:

...

On Mon, Nov 08, 2021 at 03:39:17PM -0800, Rob Clark wrote:

...

I stumbled across this thread when I ran into the same issue, while working out how to move drm/msm to use scheduler's retire + timeout/recovery (and get rid of our own mirror list of in-flight jobs). We already have hw error detection enabled, and it can signal quite fast, so assuming the first job on the list is the guilty job just won't work.

But I was considering a slightly different approach to fixing this, instead just handling it all in drm_sched_main() and getting rid of the complicated kthread parking gymnastics. Ie. something along the lines of:

So handling timeouts in the main sched thread wont work as soon as you have multiple engines and reset that impacts across engines:

• Nothing is simplified since you still need to stop the other scheduler threads.

• You get deadlocks if 2 schedulers time out at the same time, and both want to stop the other one.

Hence workqueue. Now the rule for the wq is that you can only have one per reset domain, so

• single engine you just take the one drm/sched provides
• if reset affects all your engines in the chip, then you allocate on in the drm_device and pass that to all
• if you have a complex of gpus all interconnected (e.g. xgmi hive for amd), then it's one wq for the entire hive

_All_ reset related things must be run on that workqueue or things breaks, which means if you get hw fault that also needs to be run there. I guess we should either patch drm/sched to check you call that function from the right workqueue, or just handle it internally.

Hmm, ok.. I guess it would be useful to better document the reasoning for the current design, that would have steered me more towards the approach taken in this patch.

Maybe this was because you worked on an old kernel? Boris did update the kerneldoc as part of making gpu reset work for panfrost, which has this multi-engine reset problem. If that's not yet clear then we need to improve the docs further.

AMD's problem is even worse, because their reset domain is the entire xgmi hive, so multiple pci devices.

I'm pushing for quite a while that we get something like an amdgpu_reset_domain structure or similar for this, but we unfortunately don't have that yet.

Maybe it should be a good idea to have something like a drm_sched_domain or similar with all the necessary information for the inter scheduler handling.

E.g. a workqueue for reset etc...

Regards, Christian.

Also there might more issues in drm/sched ofc, e.g. I've looked a bit at ordering/barriers and I'm pretty sure a lot are still missing. Or at least we should have comments in the code explaining why it all works. -Daniel

...

BR, -R

...

-Daniel

...

---

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 67382621b429..4d6ce775c316 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -764,6 +764,45 @@ static bool drm_sched_blocked(struct drm_gpu_scheduler *sched) return false; }

+static bool handle_timeout(struct drm_gpu_scheduler *sched) +{

•   struct drm_sched_job *bad;
  

•   if (!sched->has_timeout)
  

•           return false;
  

•   sched->has_timeout = false;
  

•   spin_lock(&sched->job_list_lock);
  

•   bad = list_first_entry_or_null(&sched->pending_list,
  

•                                  struct drm_sched_job, list);
  

•   if (!bad) {
  

•           spin_unlock(&sched->job_list_lock);
  

•           return false;
  

•   }
  

•   spin_unlock(&sched->job_list_lock);
  

•   if (sched->timeout_wq == system_wq) {
  

•           /*
  

•            * If driver has no specific requirements about serializing
  

•            * reset wrt. other engines, just call timedout_job() directly
  

•            */
  

•           sched->ops->timedout_job(job);
  

•   } else {
  

•           /*
  

•            * Otherwise queue it on timeout_wq and wait for it to complete
  

•            */
  

•           ... more typing needed here ...
  

•   }
  

•   if (sched->free_guilty) {
  

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

•           sched->free_guilty = false;
  

•   }
  

+}

• /**
  • drm_sched_main - main scheduler thread

@@ -787,6 +826,7 @@ static int drm_sched_main(void *param)

             wait_event_interruptible(sched->wake_up_worker,
                                      (cleanup_job =

drm_sched_get_cleanup_job(sched)) ||

•                                    handle_timeout(sched) ||
                                      (!drm_sched_blocked(sched) &&
                                       (entity =
  

drm_sched_select_entity(sched))) || kthread_should_stop());

---

drm_sched_fault() and the sw timeout handler would just set sched->has_timeout and kick sched->wake_up_worker.

And since we handle the timeout case after drm_sched_get_cleanup_job(), we know that all of the successfully completed jobs have already been popped off the list, and won't be unfairly maligned.

BR, -R

On Tue, Aug 31, 2021 at 6:29 PM Liu, Monk Monk.Liu@amd.com wrote:

...

[AMD Official Use Only]

Okay, I will reprepare this patch

Thanks

---

Monk Liu | Cloud-GPU Core team

-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, August 31, 2021 9:02 PM To: Liu, Monk Monk.Liu@amd.com Cc: amd-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; Chen, Jingwen Jingwen.Chen@amd.com Subject: Re: [PATCH 2/2] drm/sched: serialize job_timeout and scheduler

On Tue, Aug 31, 2021 at 02:59:02PM +0200, Daniel Vetter wrote:

...

Can we please have some actual commit message here, with detailed explanation of the race/bug/whatever, how you fix it and why this is the best option?

On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote: > tested-by: jingwen chen jingwen.chen@amd.com > Signed-off-by: Monk Liu Monk.Liu@amd.com > Signed-off-by: jingwen chen jingwen.chen@amd.com > --- > drivers/gpu/drm/scheduler/sched_main.c | 24 > ++++-------------------- > 1 file changed, 4 insertions(+), 20 deletions(-) > > diff --git a/drivers/gpu/drm/scheduler/sched_main.c > b/drivers/gpu/drm/scheduler/sched_main.c > index ecf8140..894fdb24 100644 > --- a/drivers/gpu/drm/scheduler/sched_main.c > +++ b/drivers/gpu/drm/scheduler/sched_main.c > @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) > sched = container_of(work, struct drm_gpu_scheduler, > work_tdr.work); > > /* Protects against concurrent deletion in > drm_sched_get_cleanup_job */ > + if (!__kthread_should_park(sched->thread)) This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

Iow it's probably buggy, and an just unconditionally parking the kthread is probably the right thing to do. If it's not the right thing to do, there's a bug here for sure.

Also why don't we reuse the function drivers already have to stop a scheduler thread? We seem to have two kthread_park now, that's probably one too much. -Daniel

...

> + kthread_park(sched->thread); > + > spin_lock(&sched->job_list_lock); > job = list_first_entry_or_null(&sched->pending_list, > struct drm_sched_job, list); > > if (job) { > - /* > - * Remove the bad job so it cannot be freed by concurrent > - * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread > - * is parked at which point it's safe. > - */ > - list_del_init(&job->list); > spin_unlock(&sched->job_list_lock); > > + /* vendor's timeout_job should call drm_sched_start() */ > status = job->sched->ops->timedout_job(job); > > /* > @@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) > kthread_park(sched->thread); > > /* > - * Reinsert back the bad job here - now it's safe as > - * drm_sched_get_cleanup_job cannot race against us and release the > - * bad job at this point - we parked (waited for) any in progress > - * (earlier) cleanups and drm_sched_get_cleanup_job will not be called > - * now until the scheduler thread is unparked. > - */ > - if (bad && bad->sched == sched) > - /* > - * Add at the head of the queue to reflect it was the earliest > - * job extracted. > - */ > - list_add(&bad->list, &sched->pending_list); > - > - /* > * Iterate the job list from later to earlier one and either deactive > * their HW callbacks or remove them from pending list if they already > * signaled. > -- > 2.7.4

>

Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog. ffwll.ch%2F&data=04%7C01%7CMonk.Liu%40amd.com%7C298815bea18f4fbf76 b308d96c7f7a8b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C6376601170 51194614%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiL CJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=QzgCU7%2BPdA0aWL5%2BJLg KeKbGaMMGqeGI9KE0P0LXlN4%3D&reserved=0

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll....

-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

On Wed, Nov 10, 2021 at 11:09:50AM +0100, Christian König wrote:

Am 10.11.21 um 10:50 schrieb Daniel Vetter:

...

On Tue, Nov 09, 2021 at 08:17:01AM -0800, Rob Clark wrote:

...

On Tue, Nov 9, 2021 at 1:07 AM Daniel Vetter daniel@ffwll.ch wrote:

...

On Mon, Nov 08, 2021 at 03:39:17PM -0800, Rob Clark wrote:

...

I stumbled across this thread when I ran into the same issue, while working out how to move drm/msm to use scheduler's retire + timeout/recovery (and get rid of our own mirror list of in-flight jobs). We already have hw error detection enabled, and it can signal quite fast, so assuming the first job on the list is the guilty job just won't work.

But I was considering a slightly different approach to fixing this, instead just handling it all in drm_sched_main() and getting rid of the complicated kthread parking gymnastics. Ie. something along the lines of:

So handling timeouts in the main sched thread wont work as soon as you have multiple engines and reset that impacts across engines:

• Nothing is simplified since you still need to stop the other scheduler threads.

• You get deadlocks if 2 schedulers time out at the same time, and both want to stop the other one.

Hence workqueue. Now the rule for the wq is that you can only have one per reset domain, so

• single engine you just take the one drm/sched provides
• if reset affects all your engines in the chip, then you allocate on in the drm_device and pass that to all
• if you have a complex of gpus all interconnected (e.g. xgmi hive for amd), then it's one wq for the entire hive

_All_ reset related things must be run on that workqueue or things breaks, which means if you get hw fault that also needs to be run there. I guess we should either patch drm/sched to check you call that function from the right workqueue, or just handle it internally.

Hmm, ok.. I guess it would be useful to better document the reasoning for the current design, that would have steered me more towards the approach taken in this patch.

Maybe this was because you worked on an old kernel? Boris did update the kerneldoc as part of making gpu reset work for panfrost, which has this multi-engine reset problem. If that's not yet clear then we need to improve the docs further.

AMD's problem is even worse, because their reset domain is the entire xgmi hive, so multiple pci devices.

I'm pushing for quite a while that we get something like an amdgpu_reset_domain structure or similar for this, but we unfortunately don't have that yet.

Maybe it should be a good idea to have something like a drm_sched_domain or similar with all the necessary information for the inter scheduler handling.

E.g. a workqueue for reset etc...

Yeah I think as soon as we have more stuff than just the wq then a drm_sched_reset_domain sounds good.

But if it's just driver stuff (e.g. the xgmi locking you have in amdgpu reset comes to mind) then I think just a driver_reset_domain struct that bundles all that stuff up seems good enough.

E.g. on i915 I'm also pondering whether some of the fw requests should be processed by the reset wq, to avoid locking headaches, so I don't think hiding that work too much in abstractions is a good idea. -Daniel

Regards, Christian.

...

Also there might more issues in drm/sched ofc, e.g. I've looked a bit at ordering/barriers and I'm pretty sure a lot are still missing. Or at least we should have comments in the code explaining why it all works. -Daniel

...

BR, -R

...

-Daniel

...

---

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 67382621b429..4d6ce775c316 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -764,6 +764,45 @@ static bool drm_sched_blocked(struct drm_gpu_scheduler *sched) return false; }

+static bool handle_timeout(struct drm_gpu_scheduler *sched) +{

•   struct drm_sched_job *bad;
  

•   if (!sched->has_timeout)
  

•           return false;
  

•   sched->has_timeout = false;
  

•   spin_lock(&sched->job_list_lock);
  

•   bad = list_first_entry_or_null(&sched->pending_list,
  

•                                  struct drm_sched_job, list);
  

•   if (!bad) {
  

•           spin_unlock(&sched->job_list_lock);
  

•           return false;
  

•   }
  

•   spin_unlock(&sched->job_list_lock);
  

•   if (sched->timeout_wq == system_wq) {
  

•           /*
  

•            * If driver has no specific requirements about serializing
  

•            * reset wrt. other engines, just call timedout_job() directly
  

•            */
  

•           sched->ops->timedout_job(job);
  

•   } else {
  

•           /*
  

•            * Otherwise queue it on timeout_wq and wait for it to complete
  

•            */
  

•           ... more typing needed here ...
  

•   }
  

•   if (sched->free_guilty) {
  

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

•           sched->free_guilty = false;
  

•   }
  

+}

• /**
  • drm_sched_main - main scheduler thread

@@ -787,6 +826,7 @@ static int drm_sched_main(void *param)

             wait_event_interruptible(sched->wake_up_worker,
                                      (cleanup_job =

drm_sched_get_cleanup_job(sched)) ||

•                                    handle_timeout(sched) ||
                                      (!drm_sched_blocked(sched) &&
                                       (entity =
  

drm_sched_select_entity(sched))) || kthread_should_stop());

---

drm_sched_fault() and the sw timeout handler would just set sched->has_timeout and kick sched->wake_up_worker.

And since we handle the timeout case after drm_sched_get_cleanup_job(), we know that all of the successfully completed jobs have already been popped off the list, and won't be unfairly maligned.

BR, -R

On Tue, Aug 31, 2021 at 6:29 PM Liu, Monk Monk.Liu@amd.com wrote:

...

[AMD Official Use Only]

Okay, I will reprepare this patch

Thanks

---

Monk Liu | Cloud-GPU Core team

-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, August 31, 2021 9:02 PM To: Liu, Monk Monk.Liu@amd.com Cc: amd-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; Chen, Jingwen Jingwen.Chen@amd.com Subject: Re: [PATCH 2/2] drm/sched: serialize job_timeout and scheduler

On Tue, Aug 31, 2021 at 02:59:02PM +0200, Daniel Vetter wrote: > Can we please have some actual commit message here, with detailed > explanation of the race/bug/whatever, how you fix it and why this is > the best option? > > On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote: > > tested-by: jingwen chen jingwen.chen@amd.com > > Signed-off-by: Monk Liu Monk.Liu@amd.com > > Signed-off-by: jingwen chen jingwen.chen@amd.com > > --- > > drivers/gpu/drm/scheduler/sched_main.c | 24 > > ++++-------------------- > > 1 file changed, 4 insertions(+), 20 deletions(-) > > > > diff --git a/drivers/gpu/drm/scheduler/sched_main.c > > b/drivers/gpu/drm/scheduler/sched_main.c > > index ecf8140..894fdb24 100644 > > --- a/drivers/gpu/drm/scheduler/sched_main.c > > +++ b/drivers/gpu/drm/scheduler/sched_main.c > > @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) > > sched = container_of(work, struct drm_gpu_scheduler, > > work_tdr.work); > > > > /* Protects against concurrent deletion in > > drm_sched_get_cleanup_job */ > > + if (!__kthread_should_park(sched->thread)) > This is a __ function, i.e. considered internal, and it's lockless > atomic, i.e. unordered. And you're not explaining why this works. > > Iow it's probably buggy, and an just unconditionally parking the > kthread is probably the right thing to do. If it's not the right thing > to do, there's a bug here for sure. Also why don't we reuse the function drivers already have to stop a scheduler thread? We seem to have two kthread_park now, that's probably one too much. -Daniel

> > + kthread_park(sched->thread); > > + > > spin_lock(&sched->job_list_lock); > > job = list_first_entry_or_null(&sched->pending_list, > > struct drm_sched_job, list); > > > > if (job) { > > - /* > > - * Remove the bad job so it cannot be freed by concurrent > > - * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread > > - * is parked at which point it's safe. > > - */ > > - list_del_init(&job->list); > > spin_unlock(&sched->job_list_lock); > > > > + /* vendor's timeout_job should call drm_sched_start() */ > > status = job->sched->ops->timedout_job(job); > > > > /* > > @@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) > > kthread_park(sched->thread); > > > > /* > > - * Reinsert back the bad job here - now it's safe as > > - * drm_sched_get_cleanup_job cannot race against us and release the > > - * bad job at this point - we parked (waited for) any in progress > > - * (earlier) cleanups and drm_sched_get_cleanup_job will not be called > > - * now until the scheduler thread is unparked. > > - */ > > - if (bad && bad->sched == sched) > > - /* > > - * Add at the head of the queue to reflect it was the earliest > > - * job extracted. > > - */ > > - list_add(&bad->list, &sched->pending_list); > > - > > - /* > > * Iterate the job list from later to earlier one and either deactive > > * their HW callbacks or remove them from pending list if they already > > * signaled. > > -- > > 2.7.4 > > > -- > Daniel Vetter > Software Engineer, Intel Corporation > https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog. > ffwll.ch%2F&data=04%7C01%7CMonk.Liu%40amd.com%7C298815bea18f4fbf76 > b308d96c7f7a8b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C6376601170 > 51194614%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiL > CJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=QzgCU7%2BPdA0aWL5%2BJLg

> KeKbGaMMGqeGI9KE0P0LXlN4%3D&reserved=0

Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll....

-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

On Wed, Nov 10, 2021 at 1:50 AM Daniel Vetter daniel@ffwll.ch wrote:

On Tue, Nov 09, 2021 at 08:17:01AM -0800, Rob Clark wrote:

...

On Tue, Nov 9, 2021 at 1:07 AM Daniel Vetter daniel@ffwll.ch wrote:

...

On Mon, Nov 08, 2021 at 03:39:17PM -0800, Rob Clark wrote:

...

I stumbled across this thread when I ran into the same issue, while working out how to move drm/msm to use scheduler's retire + timeout/recovery (and get rid of our own mirror list of in-flight jobs). We already have hw error detection enabled, and it can signal quite fast, so assuming the first job on the list is the guilty job just won't work.

But I was considering a slightly different approach to fixing this, instead just handling it all in drm_sched_main() and getting rid of the complicated kthread parking gymnastics. Ie. something along the lines of:

So handling timeouts in the main sched thread wont work as soon as you have multiple engines and reset that impacts across engines:

• Nothing is simplified since you still need to stop the other scheduler threads.

• You get deadlocks if 2 schedulers time out at the same time, and both want to stop the other one.

Hence workqueue. Now the rule for the wq is that you can only have one per reset domain, so

• single engine you just take the one drm/sched provides
• if reset affects all your engines in the chip, then you allocate on in the drm_device and pass that to all
• if you have a complex of gpus all interconnected (e.g. xgmi hive for amd), then it's one wq for the entire hive

_All_ reset related things must be run on that workqueue or things breaks, which means if you get hw fault that also needs to be run there. I guess we should either patch drm/sched to check you call that function from the right workqueue, or just handle it internally.

Hmm, ok.. I guess it would be useful to better document the reasoning for the current design, that would have steered me more towards the approach taken in this patch.

Maybe this was because you worked on an old kernel? Boris did update the kerneldoc as part of making gpu reset work for panfrost, which has this multi-engine reset problem. If that's not yet clear then we need to improve the docs further.

I saw that, and understood the ordered wq.. but missed the implication regarding having to park other scheduler kthreads

BR, -R

AMD's problem is even worse, because their reset domain is the entire xgmi hive, so multiple pci devices.

Also there might more issues in drm/sched ofc, e.g. I've looked a bit at ordering/barriers and I'm pretty sure a lot are still missing. Or at least we should have comments in the code explaining why it all works. -Daniel

...

BR, -R

...

-Daniel

...

---

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 67382621b429..4d6ce775c316 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -764,6 +764,45 @@ static bool drm_sched_blocked(struct drm_gpu_scheduler *sched) return false; }

+static bool handle_timeout(struct drm_gpu_scheduler *sched) +{

•   struct drm_sched_job *bad;
  

•   if (!sched->has_timeout)
  

•           return false;
  

•   sched->has_timeout = false;
  

•   spin_lock(&sched->job_list_lock);
  

•   bad = list_first_entry_or_null(&sched->pending_list,
  

•                                  struct drm_sched_job, list);
  

•   if (!bad) {
  

•           spin_unlock(&sched->job_list_lock);
  

•           return false;
  

•   }
  

•   spin_unlock(&sched->job_list_lock);
  

•   if (sched->timeout_wq == system_wq) {
  

•           /*
  

•            * If driver has no specific requirements about serializing
  

•            * reset wrt. other engines, just call timedout_job() directly
  

•            */
  

•           sched->ops->timedout_job(job);
  

•   } else {
  

•           /*
  

•            * Otherwise queue it on timeout_wq and wait for it to complete
  

•            */
  

•           ... more typing needed here ...
  

•   }
  

•   if (sched->free_guilty) {
  

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

•           sched->free_guilty = false;
  

•   }
  

+}

/**

• drm_sched_main - main scheduler thread

@@ -787,6 +826,7 @@ static int drm_sched_main(void *param)

            wait_event_interruptible(sched->wake_up_worker,
                                     (cleanup_job =

drm_sched_get_cleanup_job(sched)) ||

•                                    handle_timeout(sched) ||
                                     (!drm_sched_blocked(sched) &&
                                      (entity =
  

drm_sched_select_entity(sched))) || kthread_should_stop());

---

drm_sched_fault() and the sw timeout handler would just set sched->has_timeout and kick sched->wake_up_worker.

And since we handle the timeout case after drm_sched_get_cleanup_job(), we know that all of the successfully completed jobs have already been popped off the list, and won't be unfairly maligned.

BR, -R

On Tue, Aug 31, 2021 at 6:29 PM Liu, Monk Monk.Liu@amd.com wrote:

...

[AMD Official Use Only]

Okay, I will reprepare this patch

Thanks

---

Monk Liu | Cloud-GPU Core team

-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, August 31, 2021 9:02 PM To: Liu, Monk Monk.Liu@amd.com Cc: amd-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; Chen, Jingwen Jingwen.Chen@amd.com Subject: Re: [PATCH 2/2] drm/sched: serialize job_timeout and scheduler

On Tue, Aug 31, 2021 at 02:59:02PM +0200, Daniel Vetter wrote:

...

Can we please have some actual commit message here, with detailed explanation of the race/bug/whatever, how you fix it and why this is the best option?

On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote: > tested-by: jingwen chen jingwen.chen@amd.com > Signed-off-by: Monk Liu Monk.Liu@amd.com > Signed-off-by: jingwen chen jingwen.chen@amd.com > --- > drivers/gpu/drm/scheduler/sched_main.c | 24 > ++++-------------------- > 1 file changed, 4 insertions(+), 20 deletions(-) > > diff --git a/drivers/gpu/drm/scheduler/sched_main.c > b/drivers/gpu/drm/scheduler/sched_main.c > index ecf8140..894fdb24 100644 > --- a/drivers/gpu/drm/scheduler/sched_main.c > +++ b/drivers/gpu/drm/scheduler/sched_main.c > @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) > sched = container_of(work, struct drm_gpu_scheduler, > work_tdr.work); > > /* Protects against concurrent deletion in > drm_sched_get_cleanup_job */ > + if (!__kthread_should_park(sched->thread))

This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

Iow it's probably buggy, and an just unconditionally parking the kthread is probably the right thing to do. If it's not the right thing to do, there's a bug here for sure.

Also why don't we reuse the function drivers already have to stop a scheduler thread? We seem to have two kthread_park now, that's probably one too much. -Daniel

...

> + kthread_park(sched->thread); > + > spin_lock(&sched->job_list_lock); > job = list_first_entry_or_null(&sched->pending_list, > struct drm_sched_job, list); > > if (job) { > - /* > - * Remove the bad job so it cannot be freed by concurrent > - * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread > - * is parked at which point it's safe. > - */ > - list_del_init(&job->list); > spin_unlock(&sched->job_list_lock); > > + /* vendor's timeout_job should call drm_sched_start() */ > status = job->sched->ops->timedout_job(job); > > /* > @@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) > kthread_park(sched->thread); > > /* > - * Reinsert back the bad job here - now it's safe as > - * drm_sched_get_cleanup_job cannot race against us and release the > - * bad job at this point - we parked (waited for) any in progress > - * (earlier) cleanups and drm_sched_get_cleanup_job will not be called > - * now until the scheduler thread is unparked. > - */ > - if (bad && bad->sched == sched) > - /* > - * Add at the head of the queue to reflect it was the earliest > - * job extracted. > - */ > - list_add(&bad->list, &sched->pending_list); > - > - /* > * Iterate the job list from later to earlier one and either deactive > * their HW callbacks or remove them from pending list if they already > * signaled. > -- > 2.7.4 >

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog. ffwll.ch%2F&data=04%7C01%7CMonk.Liu%40amd.com%7C298815bea18f4fbf76 b308d96c7f7a8b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C6376601170 51194614%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiL CJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=QzgCU7%2BPdA0aWL5%2BJLg KeKbGaMMGqeGI9KE0P0LXlN4%3D&reserved=0

-- Daniel Vetter Software Engineer, Intel Corporation https://nam11.safelinks.protection.outlook.com/?url=http%3A%2F%2Fblog.ffwll....

-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch

[AMD Official Use Only]

...

This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

It's not a traditional habit from what I can see that put explain in code, but we can do that in mails , We want to park the scheduler in job_timeout to serialize the job accessing from both sched and TO handler , but inside vendor's callback timeout_job at least both panfrost and amd They both call drm_sched_stop() on all schedulers.

If we unconditionally call "kthread_park" in job_timedout then the bailing job's timedout will try to call "kthread_park" again on its scheduler and introduce "warning"

The scenario is : 1,Job1 on sched1 triggers timedout, and sched1 is parked, 2,vendor callback runs, it will usually stop all schedulers. 3,Job2 on sched2 triggers timedout, so the job_timedout also try to park sched2, but sched2 was stopped already by above step. (job2's timeout is introduced by job1, or by other VF) ---So there will be "warning" in kernel log from above step... after this "__kthread_should_park()" here we can avoid the warning, that's the only reason I need this __function. 4,Before vendor callback exit, it will unpark all schedulers.

From another hand if we don't do the kthread_park() and still delete the job here (drop deleting/reinserting the job from pending_list is what we want), we still have a small windows to hit the race issue:

That cleanup_job from sched thread is freeing the job while job is under processing by job_timedout or vendor's call back.

And the reason we want to avoid deleting/reinserting the timedout job is because we (amd vendor) have our own way to do a diagnostic on all jobs in pending list from all scheduler, we want to cherry-pick the real bad job

From all scheduler's pending list that causes this JOB TIMEOUT.

Besides, it is also much reasonable to park scheduler when job_timedout is running there, they should exclusively access those common members like sched_job. (due to spin_lock is off before running into vendor's calback)

Hope I explained ourselves well enough.

Thanks

------------------------------------------ Monk Liu | Cloud-GPU Core team ------------------------------------------

-----Original Message----- From: Daniel Vetter daniel@ffwll.ch Sent: Tuesday, August 31, 2021 8:59 PM To: Liu, Monk Monk.Liu@amd.com Cc: amd-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; Chen, Jingwen Jingwen.Chen@amd.com Subject: Re: [PATCH 2/2] drm/sched: serialize job_timeout and scheduler

Can we please have some actual commit message here, with detailed explanation of the race/bug/whatever, how you fix it and why this is the best option?

On Tue, Aug 31, 2021 at 06:35:39PM +0800, Monk Liu wrote:

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work);

/* Protects against concurrent deletion in drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

This is a __ function, i.e. considered internal, and it's lockless atomic, i.e. unordered. And you're not explaining why this works.

Iow it's probably buggy, and an just unconditionally parking the kthread is probably the right thing to do. If it's not the right thing to do, there's a bug here for sure. -Daniel

• kthread_park(sched->thread);
  

• spin_lock(&sched->job_list_lock); job = list_first_entry_or_null(&sched->pending_list, struct drm_sched_job, list);

  if (job) {

• /*
  

•  * Remove the bad job so it cannot be freed by concurrent
  

•  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•  * is parked at which point it's safe.
  

•  */
  

• list_del_init(&job->list);
  

  spin_unlock(&sched->job_list_lock);

• /* vendor's timeout_job should call drm_sched_start() */
  

  status = job->sched->ops->timedout_job(job);

  /*

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread);

/*

• * Reinsert back the bad job here - now it's safe as
  

• * drm_sched_get_cleanup_job cannot race against us and release the
  

• * bad job at this point - we parked (waited for) any in progress
  

• * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
  

• * now until the scheduler thread is unparked.
  

• */
  

• if (bad && bad->sched == sched)

• /*
  

•  * Add at the head of the queue to reflect it was the earliest
  

•  * job extracted.
  

•  */
  

• list_add(&bad->list, &sched->pending_list);
  

• /*
  • Iterate the job list from later to earlier one and either deactive
  • their HW callbacks or remove them from pending list if they already
  • signaled.

-- 2.7.4

On Tue, Aug 31, 2021 at 09:53:36AM -0400, Andrey Grodzovsky wrote:

It's says patch [2/2] but i can't find patch 1

On 2021-08-31 6:35 a.m., Monk Liu wrote:

...

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work); /* Protects against concurrent deletion in drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

• kthread_park(sched->thread);
  

As mentioned before, without serializing against other TDR handlers from other schedulers you just race here against them, e.g. you parked it now but another one in progress will unpark it as part of calling  drm_sched_start for other rings[1] Unless I am missing something since I haven't found patch [1/2]

[1] - https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/amd/amdgpu/am...

You need to have your own wq and run all your tdr work on the same wq if your reset has any cross-engine impact.

See

https://dri.freedesktop.org/docs/drm/gpu/drm-mm.html#c.drm_sched_backend_ops

for the ->timeout_job callback docs. I thought I brought this up already? -Daniel

Andrey

...

spin_lock(&sched->job_list_lock); job = list_first_entry_or_null(&sched->pending_list, struct drm_sched_job, list); if (job) {

• /*
  

•  * Remove the bad job so it cannot be freed by concurrent
  

•  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•  * is parked at which point it's safe.
  

•  */
  

• list_del_init(&job->list);
  

  spin_unlock(&sched->job_list_lock);

• /* vendor's timeout_job should call drm_sched_start() */
  

  status = job->sched->ops->timedout_job(job); /*

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread); /*

• * Reinsert back the bad job here - now it's safe as
  

• * drm_sched_get_cleanup_job cannot race against us and release the
  

• * bad job at this point - we parked (waited for) any in progress
  

• * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
  

• * now until the scheduler thread is unparked.
  

• */
  

• if (bad && bad->sched == sched)

• /*
  

•  * Add at the head of the queue to reflect it was the earliest
  

•  * job extracted.
  

•  */
  

• list_add(&bad->list, &sched->pending_list);
  

• /*
  • Iterate the job list from later to earlier one and either deactive
  • their HW callbacks or remove them from pending list if they already
  • signaled.

On Tue, Aug 31, 2021 at 10:20:40AM -0400, Andrey Grodzovsky wrote:

On 2021-08-31 10:03 a.m., Daniel Vetter wrote:

...

On Tue, Aug 31, 2021 at 09:53:36AM -0400, Andrey Grodzovsky wrote:

...

It's says patch [2/2] but i can't find patch 1

On 2021-08-31 6:35 a.m., Monk Liu wrote:

...

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work); /* Protects against concurrent deletion in drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

• kthread_park(sched->thread);
  

As mentioned before, without serializing against other TDR handlers from other schedulers you just race here against them, e.g. you parked it now but another one in progress will unpark it as part of calling  drm_sched_start for other rings[1] Unless I am missing something since I haven't found patch [1/2]

[1] - https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Felixir.boo...

You need to have your own wq and run all your tdr work on the same wq if your reset has any cross-engine impact.

IMHO what is problematic in serializing vs. locking (with trylock and bail out like we do in [1]) is for multiple TO events arising from same reason like maybe one job just waits for another and once first is hanged the second will also appear to be hanged triggering it's own TO event. In this case multiple TOs event will trigger multiple resets if we serialize but if we use lock with trylock the second one will quietly bail out.

[1] https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/amd/amdgpu/am...

Hm so I guess a single wq here, that will hold up all other TO. And they should recheck whether the job is moving meanwhile.

Also unless you use hw semaphores the job shouldn't even start before the deps are singalled, so not sure how this goes wrong?

The vm_id flush stuff can make things a bit more fun for your specific case, but in your specific case you have to run all TO handlers on the same ordered workqueue anyway (because trying to paper over this in other ways doesn't work imo).

So I think this should all work, no need for tricky cross-scheduler locking. -Daniel

Andrey

...

See

https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdri.freede...

for the ->timeout_job callback docs. I thought I brought this up already? -Daniel

Yes, this discussion is a continuation of your comment about serializing, I mentioned before that you proposed it.

Andrey

...

...

Andrey

...

spin_lock(&sched->job_list_lock);
job = list_first_entry_or_null(&sched->pending_list,
			       struct drm_sched_job, list);
if (job) {

• /*
  

•  * Remove the bad job so it cannot be freed by concurrent
  

•  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•  * is parked at which point it's safe.
  

•  */
  

• list_del_init(&job->list);
  spin_unlock(&sched->job_list_lock);
  

• /* vendor's timeout_job should call drm_sched_start() */
  status = job->sched->ops->timedout_job(job);
  /*
  

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread); /*

• * Reinsert back the bad job here - now it's safe as
  

• * drm_sched_get_cleanup_job cannot race against us and release the
  

• * bad job at this point - we parked (waited for) any in progress
  

• * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
  

• * now until the scheduler thread is unparked.
  

• */
  

• if (bad && bad->sched == sched)

• /*
  

•  * Add at the head of the queue to reflect it was the earliest
  

•  * job extracted.
  

•  */
  

• list_add(&bad->list, &sched->pending_list);
  

• /*
  • Iterate the job list from later to earlier one and either deactive
  • their HW callbacks or remove them from pending list if they already
  • signaled.

On 2021-08-31 11:23, Andrey Grodzovsky wrote:

On 2021-08-31 10:38 a.m., Daniel Vetter wrote:

...

On Tue, Aug 31, 2021 at 10:20:40AM -0400, Andrey Grodzovsky wrote:

...

On 2021-08-31 10:03 a.m., Daniel Vetter wrote:

...

On Tue, Aug 31, 2021 at 09:53:36AM -0400, Andrey Grodzovsky wrote:

...

It's says patch [2/2] but i can't find patch 1

On 2021-08-31 6:35 a.m., Monk Liu wrote:

...

tested-by: jingwen chen jingwen.chen@amd.com Signed-off-by: Monk Liu Monk.Liu@amd.com Signed-off-by: jingwen chen jingwen.chen@amd.com

---

drivers/gpu/drm/scheduler/sched_main.c | 24 ++++--------------------
1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index ecf8140..894fdb24 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work); /* Protects against concurrent deletion in drm_sched_get_cleanup_job */

• if (!__kthread_should_park(sched->thread))

• kthread_park(sched->thread);
  

As mentioned before, without serializing against other TDR handlers from other schedulers you just race here against them, e.g. you parked it now but another one in progress will unpark it as part of calling  drm_sched_start for other rings[1] Unless I am missing something since I haven't found patch [1/2]

[1] - https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Felixir.boo...

You need to have your own wq and run all your tdr work on the same wq if your reset has any cross-engine impact.

IMHO what is problematic in serializing vs. locking (with trylock and bail out like we do in [1]) is for multiple TO events arising from same reason like maybe one job just waits for another and once first is hanged the second will also appear to be hanged triggering it's own TO event. In this case multiple TOs event will trigger multiple resets if we serialize but if we use lock with trylock the second one will quietly bail out.

[1] https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Felixir.boo...

Hm so I guess a single wq here, that will hold up all other TO. And they should recheck whether the job is moving meanwhile.

Can you clarify about this ? What job should be moving ? The dependent job ?

...

Also unless you use hw semaphores the job shouldn't even start before the deps are singalled, so not sure how this goes wrong?

What about a simple example where we actually can submit a shader on one ring and a simple WAIT_REG_MEM packet on another to wait for the shader to write a specific value to specific memory location. Here you have both of them started in close proximity and no explicit dependencies involved (at the scheduler level) and yet if the shader hangs also the WAIT_REG_MEM job will hang.

...

The vm_id flush stuff can make things a bit more fun for your specific case, but in your specific case you have to run all TO handlers on the same ordered workqueue anyway (because trying to paper over this in other ways doesn't work imo).

I didn't get this one.

So, awhile back I tried to "serialize" this by moving timed-out jobs into their own timed-out-dedicated list, then freeing them asynchronously, but I never got it to work reliably due to races with low-level drivers and assumptions made way back.

My idea was to atomic-move timed-out jobs into their own list, at the time of timeout, and later asynchronously to free them (or better yet, inquire about their state, and free them or move them back--ideally the inquiry is atomic and done at timeout time before being moved to the timeout list). Anyway...

But I found out that all these knobs and levers weren't in place and I was getting problems with it and it never materialized.

The paradigm was loosely "let someone else do it", like, "on an event, move it to a list, and let someone else handle it", or "on an event, mark it, and let someone else handle it". (loosely borrowed from an iSCSI target I did many many years ago--it worked well and there were no races, even with out-of-order executions.)

If you guys have any ideas to that end, maybe we can try it out.

Regards, Luben

Andrey

...

So I think this should all work, no need for tricky cross-scheduler locking. -Daniel

...

Andrey

...

See

https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdri.freede...

for the ->timeout_job callback docs. I thought I brought this up already? -Daniel

Yes, this discussion is a continuation of your comment about serializing, I mentioned before that you proposed it.

Andrey

...

...

Andrey

...

	spin_lock(&sched->job_list_lock);
	job = list_first_entry_or_null(&sched->pending_list,
				       struct drm_sched_job, list);
	if (job) {

• /*
  

•  * Remove the bad job so it cannot be freed by concurrent
  

•  * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread
  

•  * is parked at which point it's safe.
  

•  */
  

• list_del_init(&job->list);
  	spin_unlock(&sched->job_list_lock);
  

• /* vendor's timeout_job should call drm_sched_start() */
  	status = job->sched->ops->timedout_job(job);
  	/*
  

@@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) kthread_park(sched->thread); /*

• * Reinsert back the bad job here - now it's safe as
  

• * drm_sched_get_cleanup_job cannot race against us and release the
  

• * bad job at this point - we parked (waited for) any in progress
  

• * (earlier) cleanups and drm_sched_get_cleanup_job will not be called
  

• * now until the scheduler thread is unparked.
  

• */
  

• if (bad && bad->sched == sched)

• /*
  

•  * Add at the head of the queue to reflect it was the earliest
  

•  * job extracted.
  

•  */
  

• list_add(&bad->list, &sched->pending_list);
  

• /* * Iterate the job list from later to earlier one and either deactive * their HW callbacks or remove them from pending list if they already * signaled.

On 2021-08-31 16:56, Andrey Grodzovsky wrote:

On 2021-08-31 12:01 p.m., Luben Tuikov wrote:

...

On 2021-08-31 11:23, Andrey Grodzovsky wrote:

...

On 2021-08-31 10:38 a.m., Daniel Vetter wrote:

...

On Tue, Aug 31, 2021 at 10:20:40AM -0400, Andrey Grodzovsky wrote:

...

On 2021-08-31 10:03 a.m., Daniel Vetter wrote:

...

On Tue, Aug 31, 2021 at 09:53:36AM -0400, Andrey Grodzovsky wrote: > It's says patch [2/2] but i can't find patch 1 > > On 2021-08-31 6:35 a.m., Monk Liu wrote: >> tested-by: jingwen chen jingwen.chen@amd.com >> Signed-off-by: Monk Liu Monk.Liu@amd.com >> Signed-off-by: jingwen chen jingwen.chen@amd.com >> --- >> drivers/gpu/drm/scheduler/sched_main.c | 24 ++++-------------------- >> 1 file changed, 4 insertions(+), 20 deletions(-) >> >> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c >> index ecf8140..894fdb24 100644 >> --- a/drivers/gpu/drm/scheduler/sched_main.c >> +++ b/drivers/gpu/drm/scheduler/sched_main.c >> @@ -319,19 +319,17 @@ static void drm_sched_job_timedout(struct work_struct *work) >> sched = container_of(work, struct drm_gpu_scheduler, work_tdr.work); >> /* Protects against concurrent deletion in drm_sched_get_cleanup_job */ >> + if (!__kthread_should_park(sched->thread)) >> + kthread_park(sched->thread); >> + > As mentioned before, without serializing against other TDR handlers from > other > schedulers you just race here against them, e.g. you parked it now but > another > one in progress will unpark it as part of calling  drm_sched_start for other > rings[1] > Unless I am missing something since I haven't found patch [1/2] > > [1] - https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Felixir.boo... You need to have your own wq and run all your tdr work on the same wq if your reset has any cross-engine impact.

IMHO what is problematic in serializing vs. locking (with trylock and bail out like we do in [1]) is for multiple TO events arising from same reason like maybe one job just waits for another and once first is hanged the second will also appear to be hanged triggering it's own TO event. In this case multiple TOs event will trigger multiple resets if we serialize but if we use lock with trylock the second one will quietly bail out.

[1] https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Felixir.boo...

Hm so I guess a single wq here, that will hold up all other TO. And they should recheck whether the job is moving meanwhile.

Can you clarify about this ? What job should be moving ? The dependent job ?

...

Also unless you use hw semaphores the job shouldn't even start before the deps are singalled, so not sure how this goes wrong?

What about a simple example where we actually can submit a shader on one ring and a simple WAIT_REG_MEM packet on another to wait for the shader to write a specific value to specific memory location. Here you have both of them started in close proximity and no explicit dependencies involved (at the scheduler level) and yet if the shader hangs also the WAIT_REG_MEM job will hang.

...

The vm_id flush stuff can make things a bit more fun for your specific case, but in your specific case you have to run all TO handlers on the same ordered workqueue anyway (because trying to paper over this in other ways doesn't work imo).

I didn't get this one.

So, awhile back I tried to "serialize" this by moving timed-out jobs into their own timed-out-dedicated list, then freeing them asynchronously, but I never got it to work reliably due to races with low-level drivers and assumptions made way back.

My idea was to atomic-move timed-out jobs into their own list, at the time of timeout, and later asynchronously to free them (or better yet, inquire about their state, and free them or move them back--ideally the inquiry is atomic and done at timeout time before being moved to the timeout list). Anyway...

But I found out that all these knobs and levers weren't in place and I was getting problems with it and it never materialized.

The paradigm was loosely "let someone else do it", like, "on an event, move it to a list, and let someone else handle it", or "on an event, mark it, and let someone else handle it". (loosely borrowed from an iSCSI target I did many many years ago--it worked well and there were no races, even with out-of-order executions.)

If you guys have any ideas to that end, maybe we can try it out.

Regards, Luben

I wonder if we really need this serialization at all, if we do HW fence embedding at the drm_sched_job level instead of doing it only for amdgpu, and modifying all the drivers to support this we can both remove this hack and solve the race against concurrent drm_sched_cleanup_jobs job freeing just by taking reference to the hw fence of the job at the beginning of drm_sched_job_timedout

Andrey

This sounds like the right approach to me.

(Convincing the low-level drivers of this might might be a big task.)

Regards, Luben

...

...

Andrey

...

So I think this should all work, no need for tricky cross-scheduler locking. -Daniel

...

Andrey

...

See

https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdri.freede...

for the ->timeout_job callback docs. I thought I brought this up already? -Daniel

Yes, this discussion is a continuation of your comment about serializing, I mentioned before that you proposed it.

Andrey

...

> Andrey > > >> spin_lock(&sched->job_list_lock); >> job = list_first_entry_or_null(&sched->pending_list, >> struct drm_sched_job, list); >> if (job) { >> - /* >> - * Remove the bad job so it cannot be freed by concurrent >> - * drm_sched_cleanup_jobs. It will be reinserted back after sched->thread >> - * is parked at which point it's safe. >> - */ >> - list_del_init(&job->list); >> spin_unlock(&sched->job_list_lock); >> + /* vendor's timeout_job should call drm_sched_start() */ >> status = job->sched->ops->timedout_job(job); >> /* >> @@ -393,20 +391,6 @@ void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad) >> kthread_park(sched->thread); >> /* >> - * Reinsert back the bad job here - now it's safe as >> - * drm_sched_get_cleanup_job cannot race against us and release the >> - * bad job at this point - we parked (waited for) any in progress >> - * (earlier) cleanups and drm_sched_get_cleanup_job will not be called >> - * now until the scheduler thread is unparked. >> - */ >> - if (bad && bad->sched == sched) >> - /* >> - * Add at the head of the queue to reflect it was the earliest >> - * job extracted. >> - */ >> - list_add(&bad->list, &sched->pending_list); >> - >> - /* >> * Iterate the job list from later to earlier one and either deactive >> * their HW callbacks or remove them from pending list if they already >> * signaled.

What about removing

(kthread_should_park()) ? We decided it's useless as far as I remember.

Andrey ________________________________ From: amd-gfx amd-gfx-bounces@lists.freedesktop.org on behalf of Liu, Monk Monk.Liu@amd.com Sent: 31 August 2021 20:24 To: Liu, Monk Monk.Liu@amd.com; amd-gfx@lists.freedesktop.org amd-gfx@lists.freedesktop.org Cc: dri-devel@lists.freedesktop.org dri-devel@lists.freedesktop.org Subject: RE: [PATCH 1/2] drm/sched: fix the bug of time out calculation(v3)

[AMD Official Use Only]

Ping Christian, Andrey

Can we merge this patch first ? this is a standalone patch for the timer

Thanks

------------------------------------------ Monk Liu | Cloud-GPU Core team ------------------------------------------

-----Original Message----- From: Monk Liu Monk.Liu@amd.com Sent: Tuesday, August 31, 2021 6:36 PM To: amd-gfx@lists.freedesktop.org Cc: dri-devel@lists.freedesktop.org; Liu, Monk Monk.Liu@amd.com Subject: [PATCH 1/2] drm/sched: fix the bug of time out calculation(v3)

issue: in cleanup_job the cancle_delayed_work will cancel a TO timer even the its corresponding job is still running.

fix: do not cancel the timer in cleanup_job, instead do the cancelling only when the heading job is signaled, and if there is a "next" job we start_timeout again.

v2: further cleanup the logic, and do the TDR timer cancelling if the signaled job is the last one in its scheduler.

v3: change the issue description remove the cancel_delayed_work in the begining of the cleanup_job recover the implement of drm_sched_job_begin.

TODO: 1)introduce pause/resume scheduler in job_timeout to serial the handling of scheduler and job_timeout. 2)drop the bad job's del and insert in scheduler due to above serialization (no race issue anymore with the serialization)

tested-by: jingwen <jingwen.chen@@amd.com> Signed-off-by: Monk Liu Monk.Liu@amd.com --- drivers/gpu/drm/scheduler/sched_main.c | 25 ++++++++++--------------- 1 file changed, 10 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index a2a9536..ecf8140 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -676,13 +676,7 @@ drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched) { struct drm_sched_job *job, *next;

- /* - * Don't destroy jobs while the timeout worker is running OR thread - * is being parked and hence assumed to not touch pending_list - */ - if ((sched->timeout != MAX_SCHEDULE_TIMEOUT && - !cancel_delayed_work(&sched->work_tdr)) || - kthread_should_park()) + if (kthread_should_park()) return NULL;

spin_lock(&sched->job_list_lock); @@ -693,17 +687,21 @@ drm_sched_get_cleanup_job(struct drm_gpu_scheduler *sched) if (job && dma_fence_is_signaled(&job->s_fence->finished)) { /* remove job from pending_list */ list_del_init(&job->list); + + /* cancel this job's TO timer */ + cancel_delayed_work(&sched->work_tdr); /* make the scheduled timestamp more accurate */ next = list_first_entry_or_null(&sched->pending_list, typeof(*next), list); - if (next) + + if (next) { next->s_fence->scheduled.timestamp = job->s_fence->finished.timestamp; - + /* start TO timer for next job */ + drm_sched_start_timeout(sched); + } } else { job = NULL; - /* queue timeout for next job */ - drm_sched_start_timeout(sched); }

spin_unlock(&sched->job_list_lock); @@ -791,11 +789,8 @@ static int drm_sched_main(void *param) (entity = drm_sched_select_entity(sched))) || kthread_should_stop());

- if (cleanup_job) { + if (cleanup_job) sched->ops->free_job(cleanup_job); - /* queue timeout for next job */ - drm_sched_start_timeout(sched); - }

if (!entity) continue; -- 2.7.4