[PATCH 1/3] drm/v3d: Don't resubmit guilty CSD jobs

Yukimasa Sugizaki ysugi at idein.jp
Fri Sep 4 08:15:51 UTC 2020


The previous code misses a check for the timeout error set by
drm_sched_resubmit_jobs(), which results in an infinite GPU reset loop
if once a timeout occurs:

[  178.799106] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* Resetting GPU for hang.
[  178.807836] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* V3D_ERR_STAT: 0x00001000
[  179.839132] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* Resetting GPU for hang.
[  179.847865] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* V3D_ERR_STAT: 0x00001000
[  180.879146] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* Resetting GPU for hang.
[  180.887925] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* V3D_ERR_STAT: 0x00001000
[  181.919188] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* Resetting GPU for hang.
[  181.928002] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* V3D_ERR_STAT: 0x00001000
...

This commit adds the check for timeout as in v3d_{bin,render}_job_run():

[   66.408962] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* Resetting GPU for hang.
[   66.417734] v3d fec00000.v3d: [drm:v3d_reset [v3d]] *ERROR* V3D_ERR_STAT: 0x00001000
[   66.428296] [drm] Skipping CSD job resubmission due to previous error (-125)

, where -125 is -ECANCELED, though users currently have no way other
than inspecting the dmesg to check if the timeout has occurred.

Signed-off-by: Yukimasa Sugizaki <ysugi at idein.jp>
---
 drivers/gpu/drm/v3d/v3d_sched.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
index 0747614a78f0..001216f22017 100644
--- a/drivers/gpu/drm/v3d/v3d_sched.c
+++ b/drivers/gpu/drm/v3d/v3d_sched.c
@@ -226,6 +226,17 @@ v3d_csd_job_run(struct drm_sched_job *sched_job)
 	struct dma_fence *fence;
 	int i;

+	/* This error is set to -ECANCELED by drm_sched_resubmit_jobs() if this
+	 * job timed out more than sched_job->sched->hang_limit times.
+	 */
+	int error = sched_job->s_fence->finished.error;
+
+	if (unlikely(error < 0)) {
+		DRM_WARN("Skipping CSD job resubmission due to previous error (%d)\n",
+			 error);
+		return ERR_PTR(error);
+	}
+
 	v3d->csd_job = job;

 	v3d_invalidate_caches(v3d);
--
2.7.4



More information about the dri-devel mailing list