Hi Melissa,
On 25/08/25 17:15, Melissa Wen wrote:
On 15/08/2025 16:58, Maíra Canal wrote:
The per-fd reset counter tracks GPU resets caused by jobs submitted
through a specific file descriptor. However, there's a race condition
where the file descriptor can be closed while jobs are still running,
leading to potential access to freed memory when updating the reset
counter.
Ensure that the per-fd reset counter is only updated when the file
descriptor is still valid, preventing use-after-free scenarios during
GPU reset handling.
Signed-off-by: Maíra Canal <mca...@igalia.com>
Reviewed-by: Iago Toral Quiroga <ito...@igalia.com>
---
drivers/gpu/drm/v3d/v3d_sched.c | 28 ++++++++++++++++------------
1 file changed, 16 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/
v3d_sched.c
index
d2045d0db12e63bc67bac6a47cabcf746189ea88..05b7c2eae3a5dbd34620f2666ffb69364a5136d5 100644
--- a/drivers/gpu/drm/v3d/v3d_sched.c
+++ b/drivers/gpu/drm/v3d/v3d_sched.c
@@ -722,17 +722,19 @@ v3d_cache_clean_job_run(struct drm_sched_job
*sched_job)
}
static enum drm_gpu_sched_stat
-v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job
*sched_job)
+v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job
*sched_job,
+ enum v3d_queue q)
{
struct v3d_job *job = to_v3d_job(sched_job);
struct v3d_file_priv *v3d_priv = job->file_priv;
- enum v3d_queue q;
+ unsigned long irqflags;
+ enum v3d_queue i;
mutex_lock(&v3d->reset_lock);
/* block scheduler */
- for (q = 0; q < V3D_MAX_QUEUES; q++)
- drm_sched_stop(&v3d->queue[q].sched, sched_job);
+ for (i = 0; i < V3D_MAX_QUEUES; i++)
+ drm_sched_stop(&v3d->queue[i].sched, sched_job);
if (sched_job)
drm_sched_increase_karma(sched_job);
@@ -741,15 +743,17 @@ v3d_gpu_reset_for_timeout(struct v3d_dev *v3d,
struct drm_sched_job *sched_job)
v3d_reset(v3d);
v3d->reset_counter++;
- v3d_priv->reset_counter++;
+ spin_lock_irqsave(&v3d->queue[q].queue_lock, irqflags);
+ if (v3d_priv)
+ v3d_priv->reset_counter++;
+ spin_unlock_irqrestore(&v3d->queue[q].queue_lock, irqflags);
- for (q = 0; q < V3D_MAX_QUEUES; q++)
+ for (i = 0; i < V3D_MAX_QUEUES; i++)
drm_sched_resubmit_jobs(&v3d->queue[q].sched);
I'm a bit confused here.
Why are you iterating over i but always resubmitting jobs of the same
queue?
:face_palm: Great catch! I'll address it in the next version.
/* Unblock schedulers and restart their jobs. */
- for (q = 0; q < V3D_MAX_QUEUES; q++) {
- drm_sched_start(&v3d->queue[q].sched, 0);
- }
+ for (i = 0; i < V3D_MAX_QUEUES; i++)
+ drm_sched_start(&v3d->queue[i].sched, 0);
Couldn't drm_sched_resumit_jobs() and drm_sched_start() calls be in the
same V3D_MAX_QUEUES loop?
Most likely, it would be harmless to do so. But I'd prefer to keep it as
it is, as the queues can have dependencies between them and
`drm_sched_resubmit_jobs()` is already a deprecated function that
doesn't seem to work 100% properly (check some discussions we had in the
mailing list earlier this year about it).
Best Regards,
- Maíra
mutex_unlock(&v3d->reset_lock);
@@ -777,7 +781,7 @@ v3d_cl_job_timedout(struct drm_sched_job
*sched_job, enum v3d_queue q,
return DRM_GPU_SCHED_STAT_NO_HANG;
}
- return v3d_gpu_reset_for_timeout(v3d, sched_job);
+ return v3d_gpu_reset_for_timeout(v3d, sched_job, q);
}
static enum drm_gpu_sched_stat
@@ -803,7 +807,7 @@ v3d_tfu_job_timedout(struct drm_sched_job *sched_job)
{
struct v3d_job *job = to_v3d_job(sched_job);
- return v3d_gpu_reset_for_timeout(job->v3d, sched_job);
+ return v3d_gpu_reset_for_timeout(job->v3d, sched_job, V3D_TFU);
}
static enum drm_gpu_sched_stat
@@ -822,7 +826,7 @@ v3d_csd_job_timedout(struct drm_sched_job *sched_job)
return DRM_GPU_SCHED_STAT_NO_HANG;
}
- return v3d_gpu_reset_for_timeout(v3d, sched_job);
+ return v3d_gpu_reset_for_timeout(v3d, sched_job, V3D_CSD);
}
static const struct drm_sched_backend_ops v3d_bin_sched_ops = {
