Re: [PATCH v1] drm/sched: fix deadlock in drm_sched_entity_kill_jobs_cb

[Pierre-Eric Pelloux-Prayer <pierre-eric@damsy.net>]

Le 30/10/2025 à 12:17, Philipp Stanner a écrit :
> On Wed, 2025-10-29 at 10:11 +0100, Pierre-Eric Pelloux-Prayer wrote:
>> https://gitlab.freedesktop.org/mesa/mesa/-/issues/13908 pointed out
> 
> This link should be moved to the tag section at the bottom at a Closes:
> tag. Optionally a Reported-by:, too.

The bug report is about a different issue. The potential deadlock being fixed by 
this patch was discovered while investigating it.
I'll add a Reported-by tag though.

> 
>> a possible deadlock:
>>
>> [ 1231.611031]  Possible interrupt unsafe locking scenario:
>>
>> [ 1231.611033]        CPU0                    CPU1
>> [ 1231.611034]        ----                    ----
>> [ 1231.611035]   lock(&xa->xa_lock#17);
>> [ 1231.611038]                                local_irq_disable();
>> [ 1231.611039]                                lock(&fence->lock);
>> [ 1231.611041]                                lock(&xa->xa_lock#17);
>> [ 1231.611044]   <Interrupt>
>> [ 1231.611045]     lock(&fence->lock);
>> [ 1231.611047]
>>                  *** DEADLOCK ***
>>
> 
> The commit message is lacking an explanation as to _how_ and _when_ the
> deadlock comes to be. That's a prerequisite for understanding why the
> below is the proper fix and solution.

I copy-pasted a small chunk of the full deadlock analysis report included in the 
ticket because it's 300+ lines long. Copying the full log isn't useful IMO, but 
I can add more context.

The problem is that a thread (CPU0 above) can lock the job's dependencies 
xa_array without disabling the interrupts.
If a fence signals while CPU0 holds this lock and drm_sched_entity_kill_jobs_cb 
is called, it will try to grab the xa_array lock which is not possible because 
CPU0 holds it already.


> 
> The issue seems to be that you cannot perform certain tasks from within
> that work item?
> 
>> My initial fix was to replace xa_erase by xa_erase_irq, but Christian
>> pointed out that calling dma_fence_add_callback from a callback can
>> also deadlock if the signalling fence and the one passed to
>> dma_fence_add_callback share the same lock.
>>
>> To fix both issues, the code iterating on dependencies and re-arming them
>> is moved out to drm_sched_entity_kill_jobs_work.
>>
>> Suggested-by: Christian König <christian.koenig@amd.com>
>> Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
>> ---
>>   drivers/gpu/drm/scheduler/sched_entity.c | 34 +++++++++++++-----------
>>   1 file changed, 19 insertions(+), 15 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/scheduler/sched_entity.c b/drivers/gpu/drm/scheduler/sched_entity.c
>> index c8e949f4a568..fe174a4857be 100644
>> --- a/drivers/gpu/drm/scheduler/sched_entity.c
>> +++ b/drivers/gpu/drm/scheduler/sched_entity.c
>> @@ -173,26 +173,15 @@ int drm_sched_entity_error(struct drm_sched_entity *entity)
>>   }
>>   EXPORT_SYMBOL(drm_sched_entity_error);
>>   
>> +static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
>> +					  struct dma_fence_cb *cb);
>> +
>>   static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
>>   {
>>   	struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
>> -
>> -	drm_sched_fence_scheduled(job->s_fence, NULL);
>> -	drm_sched_fence_finished(job->s_fence, -ESRCH);
>> -	WARN_ON(job->s_fence->parent);
>> -	job->sched->ops->free_job(job);
> 
> Can free_job() really not be called from within work item context?

It's still called from drm_sched_entity_kill_jobs_work but the diff is slightly 
confusing.

Pierre-Eric


> 
> 
> P.