drm/scheduler: Fix lockup in drm_sched_entity_kill()
authorDmitry Osipenko <dmitry.osipenko@collabora.com>
Wed, 23 Nov 2022 00:13:03 +0000 (03:13 +0300)
committerDmitry Osipenko <dmitry.osipenko@collabora.com>
Tue, 3 Jan 2023 11:49:59 +0000 (14:49 +0300)
The drm_sched_entity_kill() is invoked twice by drm_sched_entity_destroy()
while userspace process is exiting or being killed. First time it's invoked
when sched entity is flushed and second time when entity is released. This
causes a lockup within wait_for_completion(entity_idle) due to how completion
API works.

Calling wait_for_completion() more times than complete() was invoked is a
error condition that causes lockup because completion internally uses
counter for complete/wait calls. The complete_all() must be used instead
in such cases.

This patch fixes lockup of Panfrost driver that is reproducible by killing
any application in a middle of 3d drawing operation.

Fixes: 2fdb8a8f07c2 ("drm/scheduler: rework entity flush, kill and fini")
Signed-off-by: Dmitry Osipenko <dmitry.osipenko@collabora.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Tested-by: Guilherme G. Piccoli <gpiccoli@igalia.com> # Steam Deck
Link: https://patchwork.freedesktop.org/patch/msgid/20221123001303.533968-1-dmitry.osipenko@collabora.com
drivers/gpu/drm/scheduler/sched_entity.c
drivers/gpu/drm/scheduler/sched_main.c

index fe09e5b..15d04a0 100644 (file)
@@ -81,7 +81,7 @@ int drm_sched_entity_init(struct drm_sched_entity *entity,
        init_completion(&entity->entity_idle);
 
        /* We start in an idle state. */
-       complete(&entity->entity_idle);
+       complete_all(&entity->entity_idle);
 
        spin_lock_init(&entity->rq_lock);
        spsc_queue_init(&entity->job_queue);
index 31f3a12..fd22d75 100644 (file)
@@ -987,7 +987,7 @@ static int drm_sched_main(void *param)
                sched_job = drm_sched_entity_pop_job(entity);
 
                if (!sched_job) {
-                       complete(&entity->entity_idle);
+                       complete_all(&entity->entity_idle);
                        continue;
                }
 
@@ -998,7 +998,7 @@ static int drm_sched_main(void *param)
 
                trace_drm_run_job(sched_job, entity);
                fence = sched->ops->run_job(sched_job);
-               complete(&entity->entity_idle);
+               complete_all(&entity->entity_idle);
                drm_sched_fence_scheduled(s_fence);
 
                if (!IS_ERR_OR_NULL(fence)) {