[v3,0/2] media: qcom: camss: Fix two CAMSS bugs found by dogfooding with SoftISP

Message

Bryan O'Donoghue July 29, 2024, 12:42 p.m. UTC

v3:
- Amends the commit log for patch #1 per Johan's suggestion.
- Link to v2: https://lore.kernel.org/r/[e-mail]

v2:
- Updates commits with Johan's Review/Reported tags
- Adds Closes: https://lore.kernel.org/lkml/[e-mail]
- Cc's stable
- Adds in suggested kernel log to allow others to more easily match kernel
  log to fixes
- Link to v1: https://lore.kernel.org/r/[e-mail]

V1:
Dogfooding with SoftISP has uncovered two bugs in this series which I'm
posting fixes for.

- The first error:
  A simple race condition which to be honest I'm surprised I haven't found
  earlier nor has anybody else. Simply stated the order we typically
  end up loading CAMSS on boot has masked out the pm_runtime_enable() race
  condition that has been present in CAMSS for a long time.

  If you blacklist qcom-camss in modules.d and then modprobe after boot,
  the race condition shows up easily.

  Moving the pm_runtime_enable prior to subdevice registration fixes the
  problem.

The second error:
  Nomenclature:
    - CSIPHY: CSI Physical layer analogue to digital domain serialiser
    - CSID: CSI Decoder
    - VFE: Video Front End
    - RDI: Raw Data Interface
    - VC: Virtual Channel

  In order to support streaming multiple virtual-channels on the same RDI a
  V4L2 provided use_count variable is used to decide whether or not to actually
  terminate streaming and release buffers for 'msm_vfe_rdiX'.

  Unfortunately use_count indicates the number of times msm_vfe_rdiX has
  been opened by user-space not the number of concurrent streams on
  msm_vfe_rdiX.

  Simply stated use_count and stream_count are two different things.

  The silicon enabling code to select between VCs is valid but, a different
  solution needs to be found to support _concurrent_ VC streams.

  Right now the upstream use_count as-is is breaking the non concurrent VC
  case and I don't believe there are upstream users of concurrent VCs on
  CAMSS.

  This series implements a revert for the invalid use_count check,
  retaining the ability to select which VC is active on the RDI.

  Dogfooding with libcamera's SoftISP in Hangouts, Zoom and multiple runs
  of libcamera's "qcam" application is a very different test-case to the
  simple capture of frames we previously did when validating the
  'use_count' change.

  A partial revert in expectation of a renewed push to fixup that
  concurrent VC issue is included.

Signed-off-by: Bryan O'Donoghue <[e-mail]>
---
Bryan O'Donoghue (2):
      media: qcom: camss: Remove use_count guard in stop_streaming
      media: qcom: camss: Fix ordering of pm_runtime_enable

 drivers/media/platform/qcom/camss/camss-video.c | 6 ------
 drivers/media/platform/qcom/camss/camss.c       | 5 +++--
 2 files changed, 3 insertions(+), 8 deletions(-)
---
base-commit: c6ce8f9ab92edc9726996a0130bfc1c408132d47
change-id: 20240713-linux-next-24-07-13-camss-fixes-fa98c0965a5d

Best regards,