From: Yukimasa Sugizaki ysugi@idein.jp
The V3D scheduler wants a timed-out job to continue running if it made progress. However, the current DRM scheduler removes the timed-out job from ring_mirror_list and thus the timer is not restarted automatically, resulting in an infinite timeout. We need stop and restart the DRM scheduler to rearm the timer.
Fixes: 135517d3565b ("drm/scheduler: Avoid accessing freed bad job.") Signed-off-by: Yukimasa Sugizaki ysugi@idein.jp --- drivers/gpu/drm/v3d/v3d_sched.c | 27 +++++++++++++++++++++++++++ 1 file changed, 27 insertions(+)
diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c index 001216f22017..feef0c749e68 100644 --- a/drivers/gpu/drm/v3d/v3d_sched.c +++ b/drivers/gpu/drm/v3d/v3d_sched.c @@ -312,9 +312,24 @@ v3d_cl_job_timedout(struct drm_sched_job *sched_job, enum v3d_queue q, u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(q)); u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(q));
+ /* If we've made progress, skip reset and let the timer get + * rearmed. + */ if (*timedout_ctca != ctca || *timedout_ctra != ctra) { *timedout_ctca = ctca; *timedout_ctra = ctra; + + /* Because the timed-out job has been removed from + * ring_mirror_list in drm_sched_job_timedout(), we need to + * stop and restart the scheduler to rearm the timer. + * Holding the reset_lock is necessary for concurrent + * v3d_gpu_reset_for_timeout(). + */ + mutex_lock(&v3d->reset_lock); + drm_sched_stop(sched_job->sched, sched_job); + drm_sched_start(sched_job->sched, sched_job); + mutex_unlock(&v3d->reset_lock); + return; }
@@ -359,6 +374,18 @@ v3d_csd_job_timedout(struct drm_sched_job *sched_job) */ if (job->timedout_batches != batches) { job->timedout_batches = batches; + + /* Because the timed-out job has been removed from + * ring_mirror_list in drm_sched_job_timedout(), we need to + * stop and restart the scheduler to rearm the timer. + * Holding the reset_lock is necessary for concurrent + * v3d_gpu_reset_for_timeout(). + */ + mutex_lock(&v3d->reset_lock); + drm_sched_stop(sched_job->sched, sched_job); + drm_sched_start(sched_job->sched, sched_job); + mutex_unlock(&v3d->reset_lock); + return; }
-- 2.7.4