On Fri Jun 26 10:22:30 2026 +0900, Jackson Lee wrote:
> Decoder instances sharing a VPU also share one v4l2_m2m job slot, released
> when the running context calls v4l2_m2m_job_finish(). While draining,
> device_run() defers job_finish() once EOS is sent (sent_eos), expecting a
> later finish_decode() (from a DEC_PIC completion IRQ) to release the slot.
>
> But the m2m core checks job_ready() only when a job is queued, not when it
> is dispatched. A job queued while draining can run after finish_decode()
> has already moved the instance to STOP and sent EOS. device_run() then runs
> in STOP, issues no DEC_PIC, yet still skips job_finish() - so no IRQ, no
> finish_decode(), and the shared slot is leaked, stalling every instance.
> With several v4l2h264dec instances in parallel, GStreamer hangs at EOS.
>
> Track whether the run actually queued a DEC_PIC (cmd_issued) and defer
> job_finish() only then. Otherwise finish the job immediately
>
> Fixes: a176ac5e701f ("media: chips-media: wave5: Improve performance of
> decoder")
> Cc: [email protected]
> Signed-off-by: Jackson Lee <[email protected]>
> Signed-off-by: Nas Chung <[email protected]>
> Reviewed-by: Nicolas Dufresne <[email protected]>
> Signed-off-by: Nicolas Dufresne <[email protected]>
> Signed-off-by: Hans Verkuil <[email protected]>
Patch committed.
Thanks,
Hans Verkuil
drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
---
diff --git a/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c
b/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c
index 93f7b724d86c..f33c00cb801b 100644
--- a/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c
+++ b/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c
@@ -1655,6 +1655,7 @@ static void wave5_vpu_dec_device_run(void *priv)
struct queue_status_info q_status;
u32 fail_res = 0;
int ret = 0;
+ bool cmd_issued = false;
dev_dbg(inst->dev->dev, "%s: Fill the ring buffer with new bitstream
data", __func__);
pm_runtime_resume_and_get(inst->dev->dev);
@@ -1752,6 +1753,7 @@ static void wave5_vpu_dec_device_run(void *priv)
inst->retry = false;
if (!inst->eos)
inst->queuing_num--;
+ cmd_issued = true;
}
break;
default:
@@ -1769,8 +1771,16 @@ finish_job_and_return:
* in power and CPU time.
* If EOS is passed, device_run will not call job_finish no more, it is
called
* only if HW is idle status in order to reduce overhead.
+ *
+ * Deferring job_finish() is only safe when this run actually queued a
+ * DEC_PIC command (cmd_issued): that guarantees a completion IRQ, and
+ * thus a later finish_decode(), will release the shared job slot. When
+ * device_run() is entered with no command to issue (e.g. a job that was
+ * queued while draining but reached the STOP state by the time it ran),
+ * no IRQ follows, so finish the job here to avoid leaking the slot and
+ * stalling every instance sharing the VPU.
*/
- if (!inst->sent_eos)
+ if (!inst->sent_eos || !cmd_issued)
v4l2_m2m_job_finish(inst->v4l2_m2m_dev, m2m_ctx);
}
_______________________________________________
linuxtv-commits mailing list -- [email protected]
To unsubscribe send an email to [email protected]
