Re: [PATCH v3] drm/sched: Fix deadlock in drm_sched_entity_kill_jobs_cb

[Philipp Stanner <phasta@mailbox.org>]

On Tue, 2025-11-04 at 10:53 +0100, Pierre-Eric Pelloux-Prayer wrote:
> The Mesa issue referenced below pointed out 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 ***
> 
> In this example, CPU0 would be any function accessing job->dependencies
> through the xa_* functions that doesn't disable interrupts (eg:
> drm_sched_job_add_dependency, drm_sched_entity_kill_jobs_cb).
> 
> CPU1 is executing drm_sched_entity_kill_jobs_cb as a fence signalling
> callback so in an interrupt context. It will deadlock when trying to
> grab the xa_lock which is already held by CPU0.
> 
> Replacing all xa_* usage by their xa_*_irq counterparts would fix
> this issue, but Christian pointed out another issue: dma_fence_signal
> takes fence.lock and so does dma_fence_add_callback.
> 
>   dma_fence_signal() // locks f1.lock
>   -> drm_sched_entity_kill_jobs_cb()
>   -> foreach dependencies
>      -> dma_fence_add_callback() // locks f2.lock
> 
> This will deadlock if f1 and f2 share the same spinlock.
> 
> To fix both issues, the code iterating on dependencies and re-arming them
> is moved out to drm_sched_entity_kill_jobs_work.
> 
> v2: reworded commit message (Philipp)
> v3: added Fixes tag (Philipp)

Thx for the update.
In the future please put the changelog below between a pair of '---'

---
v2: …
v3: …
---

Some things I have unfortunately overlooked below.

> 
> Fixes: 2fdb8a8f07c2 ("drm/scheduler: rework entity flush, kill and fini")

We should +Cc stable. It's a deadlock after all.

> Link: https://gitlab.freedesktop.org/mesa/mesa/-/issues/13908
> Reported-by: Mikhail Gavrilov <mikhail.v.gavrilov@gmail.com>
> Suggested-by: Christian König <christian.koenig@amd.com>
> Reviewed-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);

It's far better to move the function up instead. Can you do that?

> +
> 

[…]

> +/* Signal the scheduler finished fence when the entity in question is killed. */
> +static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
> +					  struct dma_fence_cb *cb)
> +{
> +	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
> +						 finish_cb);
> +
> +	dma_fence_put(f);

It would be great if we knew what fence is being dropped here and why.
I know you're just moving the pre-existing code, but if you should
know, informing about that via comment would be great.

Optional.


Rest of the code looks good. No further objections.


P.