Re: [PATCH v1 0/9] fix decoder corruption, stalls and seek issues

[Brandon Brnich <b-brnich@ti.com>]

Hi Jackson,

Thanks for sending this series.

On 10/6/26 20:59, Jackson.lee wrote:
> From: Jackson Lee <jackson.lee@chipsnmedia.com>
> 
> This series fixes a set of decoder problems in the wave5 driver that show
> up as corrupted output, stalled or crashed instances, and lost frames
> across seeks. All of them carry a Fixes: tag.

I have tested this series in multistream scenarios as well as seeking 
and looping cases. It has made the latter 2 far more robust and I no 
longer see corruption, stalls, or warnings from gstreamer. For the series:

Tested-by: Brandon Brnich <b-brnich@ti.com>

> 
> Locking and job ownership
>    1. Serialise removal of OUTPUT buffers between finish_decode(),
>       streamoff_output() and the start_decode() error path.
>    4. Acknowledge the interrupt only after it has been dispatched, so a
>       completion the VPU adds meanwhile is no longer dropped and does not
>       leave an instance waiting forever.
>    5. Finish each m2m job exactly once. A second v4l2_m2m_job_finish() from
>       the same context could clear curr_ctx under v4l2_m2m_try_run() and
>       oops in the m2m worker.
> 
> Decode scheduling
>    6. Issue DEC_PIC only when the ring buffer holds bitstream that no queued
>       command will consume. The current check compares two unrelated
>       counters, which sends commands against an empty ring or duplicates
>       one already in flight and corrupts the output. A delayed work restarts
>       an instance left idle with data still buffered.
> 
> Seek / flush
>    2. Drop the consumed-byte tally on OUTPUT streamoff, so it is not
>       credited to the first buffers queued after a seek.
>    3. Wait between FLUSH retries instead of spending the retry budget in
>       microseconds, and ignore dec_info when reading it failed.
>    8. Re-establish the display flags from the vb2 buffer states after a
>       flush, so buffers held by userspace are not decoded into.
>    9. Do not reset the frame buffers on a CAPTURE streamoff that is only
>       part of a seek.
> 
> Timestamps
>    7. Hand OUTPUT timestamps to decoded pictures in decode order. One
>       completion can consume several source buffers, and keeping only the
>       last timestamp lost the others ("Too old frames, bug in decoder" in
>       GStreamer).
> 
> Testing
> -------
> Tested on a TI J721S2 EVM (Wave521C, product code 0x521c, firmware
> revision 363254) with GStreamer 1.24.13 and the v4l2h264dec/v4l2h265dec
> elements.
> 
> fluster (GStreamer-H.264-V4L2-Gst1.0 / GStreamer-H.265-V4L2-Gst1.0):
> 
>    JVT-AVC_V1      77/135  (-j1 and -j3, identical per vector)
>    JCT-VC-HEVC_V1 133/147  (-j1 and -j3, identical per vector)
> 
> The remaining AVC failures are interlaced, MBAFF/PAFF, FMO and SP slice
> streams. 11 of the 14 HEVC failures are Main10 streams, which this SoC
> does not decode ("no support for 10 bit depth"); the others are
> PICSIZE_A_Bossen_1, PICSIZE_B_Bossen_1 and DELTAQP_A_BRCM_4.
> 
> No oops, call trace, workqueue warning or "invalid display frame index"
> message was logged during the runs.
> 
> v4l2-compliance 1.28.1-5233, SHA: fc15e229d9d3:
> 
>    Total for wave5-dec device /dev/video0: 47, Succeeded: 47, Failed: 0, Warnings: 2
>    Total for wave5-enc device /dev/video1: 47, Succeeded: 47, Failed: 0, Warnings: 0
> 
> Brandon Brnich (2):
>    media: chips-media: wave5: Ensure Atomic Access to src_buf list
>    media: chips-media: wave5: Stop FrameBuf Reset During Seek
> 
> Jackson Lee (7):
>    media: chips-media: wave5: drop the consumed-byte tally on OUTPUT
>      streamoff
>    media: chips-media: wave5: wait before retrying a refused flush
>    media: chips-media: wave5: ack the interrupt after dispatching it
>    media: chips-media: wave5: finish a job only once
>    media: chips-media: wave5: decode only when the ring holds unclaimed
>      bitstream
>    media: chips-media: wave5: stamp decoded pictures from a decode-order
>      queue
>    media: chips-media: wave5: restore the display flags after a flush
> 
>   .../chips-media/wave5/wave5-vpu-dec.c         | 312 +++++++++++++++---
>   .../platform/chips-media/wave5/wave5-vpu.c    |  28 +-
>   .../platform/chips-media/wave5/wave5-vpuapi.c |  18 +-
>   .../platform/chips-media/wave5/wave5-vpuapi.h |  13 +
>   4 files changed, 310 insertions(+), 61 deletions(-)
> 
> 
> base-commit: 58348f64125e9a3e44d3abb275ca7f4e6c9641e5

Best,
Brandon