Re: [PATCH v2] drm/sched: Fix racy access to drm_sched_entity.dependency

[Philipp Stanner <phasta@mailbox.org>]

On Tue, 2025-09-02 at 08:59 +0100, Tvrtko Ursulin wrote:
> 
> On 02/09/2025 08:27, Philipp Stanner wrote:
> > On Mon, 2025-09-01 at 14:40 +0200, Pierre-Eric Pelloux-Prayer wrote:
> > > The drm_sched_job_unschedulable trace point can access
> > > entity->dependency after it was cleared by the callback
> > > installed in drm_sched_entity_add_dependency_cb, causing:
> > > 
> > > BUG: kernel NULL pointer dereference, address: 0000000000000020
> > > [...]
> > > Workqueue: comp_1.1.0 drm_sched_run_job_work [gpu_sched]
> > > RIP: 0010:trace_event_raw_event_drm_sched_job_unschedulable+0x70/0xd0 [gpu_sched]
> > > 
> > > To fix this we either need to keep a reference to the fence before
> > > setting up the callbacks, or move the trace_drm_sched_job_unschedulable
> > > calls into drm_sched_entity_add_dependency_cb where they can be
> > > done earlier.
> > > 
> > > Fixes: 76d97c870f29 ("drm/sched: Trace dependencies for GPU jobs")
> > > 
> > > Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
> > > Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
> > 
> > Applied to drm-misc-next
> 
> Shouldn't it have been drm-misc-fixes?

I considered that, but thought not: the fixed commit is only in this rc
(v6.17), and in this case the committer guidelines say it should go to
misc-next:

https://drm.pages.freedesktop.org/maintainer-tools/committer/committer-drm-misc.html

Same reason we don't need to +Cc stable.

But correct me if I made a mistake.

P.


> 
> Regards,
> 
> Tvrtko
> 
> > > ---
> > >   drivers/gpu/drm/scheduler/sched_entity.c | 11 +++++++----
> > >   1 file changed, 7 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/scheduler/sched_entity.c b/drivers/gpu/drm/scheduler/sched_entity.c
> > > index 8867b95ab089..3d06f72531ba 100644
> > > --- a/drivers/gpu/drm/scheduler/sched_entity.c
> > > +++ b/drivers/gpu/drm/scheduler/sched_entity.c
> > > @@ -391,7 +391,8 @@ EXPORT_SYMBOL(drm_sched_entity_set_priority);
> > >    * Add a callback to the current dependency of the entity to wake up the
> > >    * scheduler when the entity becomes available.
> > >    */
> > > -static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
> > > +static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity,
> > > +					       struct drm_sched_job *sched_job)
> > >   {
> > >   	struct drm_gpu_scheduler *sched = entity->rq->sched;
> > >   	struct dma_fence *fence = entity->dependency;
> > > @@ -421,6 +422,10 @@ static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
> > >   		entity->dependency = fence;
> > >   	}
> > >   
> > > +	if (trace_drm_sched_job_unschedulable_enabled() &&
> > > +	    !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &entity->dependency->flags))
> > > +		trace_drm_sched_job_unschedulable(sched_job, entity->dependency);
> > > +
> > >   	if (!dma_fence_add_callback(entity->dependency, &entity->cb,
> > >   				    drm_sched_entity_wakeup))
> > >   		return true;
> > > @@ -461,10 +466,8 @@ struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
> > >   
> > >   	while ((entity->dependency =
> > >   			drm_sched_job_dependency(sched_job, entity))) {
> > > -		if (drm_sched_entity_add_dependency_cb(entity)) {
> > > -			trace_drm_sched_job_unschedulable(sched_job, entity->dependency);
> > > +		if (drm_sched_entity_add_dependency_cb(entity, sched_job))
> > >   			return NULL;
> > > -		}
> > >   	}
> > >   
> > >   	/* skip jobs from entity that marked guilty */
> > 
>