| Message ID | 20240716-linux-next-24-07-13-camss-fixes-v2-1-e60c9f6742f2@linaro.org (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | media: qcom: camss: Fix two CAMSS bugs found by dogfooding with SoftISP | expand |
On Tue, Jul 16, 2024 at 11:13:24PM +0100, Bryan O'Donoghue wrote: > The use_count check was introduced so that multiple concurrent Raw Data > Interfaces RDIs could be driven by different virtual channels VCs on the > CSIPHY input driving the video pipeline. > > This is an invalid use of use_count though as use_count pertains to the > number of times a video entity has been opened by user-space not the number > of active streams. > > If use_count and stream-on count don't agree then stop_streaming() will > break as is currently the case and has become apparent when using CAMSS > with libcamera's released softisp 0.3. > > The use of use_count like this is a bit hacky and right now breaks regular > usage of CAMSS for a single stream case. As an example the "qcam" > application in libcamera will fail with an -EBUSY result on stream stop and > cannot then subsequently be restarted. No, stopping qcam results in the splat below, and then it cannot be started again and any attempts to do so fails with -EBUSY. > The kernel log for this fault looks like this: > > [ 1265.509831] WARNING: CPU: 5 PID: 919 at drivers/media/common/videobuf2/videobuf2-core.c:2183 __vb2_queue_cancel+0x230/0x2c8 [videobuf2_common] > ... > [ 1265.510630] Call trace: > [ 1265.510636] __vb2_queue_cancel+0x230/0x2c8 [videobuf2_common] > [ 1265.510648] vb2_core_streamoff+0x24/0xcc [videobuf2_common] > [ 1265.510660] vb2_ioctl_streamoff+0x5c/0xa8 [videobuf2_v4l2] > [ 1265.510673] v4l_streamoff+0x24/0x30 [videodev] > [ 1265.510707] __video_do_ioctl+0x190/0x3f4 [videodev] > [ 1265.510732] video_usercopy+0x304/0x8c4 [videodev] > [ 1265.510757] video_ioctl2+0x18/0x34 [videodev] > [ 1265.510782] v4l2_ioctl+0x40/0x60 [videodev] > ... > [ 1265.510944] videobuf2_common: driver bug: stop_streaming operation is leaving buffer 0 in active state > [ 1265.511175] videobuf2_common: driver bug: stop_streaming operation is leaving buffer 1 in active state > [ 1265.511398] videobuf2_common: driver bug: stop_streaming operation is leaving buffer 2 in active st Johan
On 17/07/2024 10:06, Johan Hovold wrote: >> The use of use_count like this is a bit hacky and right now breaks regular >> usage of CAMSS for a single stream case. As an example the "qcam" >> application in libcamera will fail with an -EBUSY result on stream stop and >> cannot then subsequently be restarted. > No, stopping qcam results in the splat below, and then it cannot be > started again and any attempts to do so fails with -EBUSY. I thought that's what I said. Let me reword the commit log with your sentence included directly :) --- bod
On Wed, Jul 17, 2024 at 11:43:02AM +0100, Bryan O'Donoghue wrote: > On 17/07/2024 10:06, Johan Hovold wrote: > >> The use of use_count like this is a bit hacky and right now breaks regular > >> usage of CAMSS for a single stream case. As an example the "qcam" > >> application in libcamera will fail with an -EBUSY result on stream stop and > >> cannot then subsequently be restarted. > > No, stopping qcam results in the splat below, and then it cannot be > > started again and any attempts to do so fails with -EBUSY. > > I thought that's what I said. I read the above as if stopping the stream fails with -EBUSY, when it's really restarting the stream that fails that way after the first stop. > Let me reword the commit log with your sentence included directly :) Sounds good. Johan
diff --git a/drivers/media/platform/qcom/camss/camss-video.c b/drivers/media/platform/qcom/camss/camss-video.c index cd72feca618c..3b8fc31d957c 100644 --- a/drivers/media/platform/qcom/camss/camss-video.c +++ b/drivers/media/platform/qcom/camss/camss-video.c @@ -297,12 +297,6 @@ static void video_stop_streaming(struct vb2_queue *q) ret = v4l2_subdev_call(subdev, video, s_stream, 0); - if (entity->use_count > 1) { - /* Don't stop if other instances of the pipeline are still running */ - dev_dbg(video->camss->dev, "Video pipeline still used, don't stop streaming.\n"); - return; - } - if (ret) { dev_err(video->camss->dev, "Video pipeline stop failed: %d\n", ret); return;
The use_count check was introduced so that multiple concurrent Raw Data Interfaces RDIs could be driven by different virtual channels VCs on the CSIPHY input driving the video pipeline. This is an invalid use of use_count though as use_count pertains to the number of times a video entity has been opened by user-space not the number of active streams. If use_count and stream-on count don't agree then stop_streaming() will break as is currently the case and has become apparent when using CAMSS with libcamera's released softisp 0.3. The use of use_count like this is a bit hacky and right now breaks regular usage of CAMSS for a single stream case. As an example the "qcam" application in libcamera will fail with an -EBUSY result on stream stop and cannot then subsequently be restarted. The kernel log for this fault looks like this: [ 1265.509831] WARNING: CPU: 5 PID: 919 at drivers/media/common/videobuf2/videobuf2-core.c:2183 __vb2_queue_cancel+0x230/0x2c8 [videobuf2_common] ... [ 1265.510630] Call trace: [ 1265.510636] __vb2_queue_cancel+0x230/0x2c8 [videobuf2_common] [ 1265.510648] vb2_core_streamoff+0x24/0xcc [videobuf2_common] [ 1265.510660] vb2_ioctl_streamoff+0x5c/0xa8 [videobuf2_v4l2] [ 1265.510673] v4l_streamoff+0x24/0x30 [videodev] [ 1265.510707] __video_do_ioctl+0x190/0x3f4 [videodev] [ 1265.510732] video_usercopy+0x304/0x8c4 [videodev] [ 1265.510757] video_ioctl2+0x18/0x34 [videodev] [ 1265.510782] v4l2_ioctl+0x40/0x60 [videodev] ... [ 1265.510944] videobuf2_common: driver bug: stop_streaming operation is leaving buffer 0 in active state [ 1265.511175] videobuf2_common: driver bug: stop_streaming operation is leaving buffer 1 in active state [ 1265.511398] videobuf2_common: driver bug: stop_streaming operation is leaving buffer 2 in active st One CAMSS specific way to handle multiple VCs on the same RDI might be: - Reference count each pipeline enable for CSIPHY, CSID, VFE and RDIx. - The video buffers are already associated with msm_vfeN_rdiX so release video buffers when told to do so by stop_streaming. - Only release the power-domains for the CSIPHY, CSID and VFE when their internal refcounts drop. Either way refusing to release video buffers based on use_count is erroneous and should be reverted. The silicon enabling code for selecting VCs is perfectly fine. Its a "known missing feature" that concurrent VCs won't work with CAMSS right now. Initial testing with this code didn't show an error but, SoftISP and "real" usage with Google Hangouts breaks the upstream code pretty quickly, we need to do a partial revert and take another pass at VCs. This commit partially reverts commit 89013969e232 ("media: camss: sm8250: Pipeline starting and stopping for multiple virtual channels") Fixes: 89013969e232 ("media: camss: sm8250: Pipeline starting and stopping for multiple virtual channels") Reported-by: Johan Hovold <johan+linaro@kernel.org> Closes: https://lore.kernel.org/lkml/ZoVNHOTI0PKMNt4_@hovoldconsulting.com/ Tested-by: Johan Hovold <johan+linaro@kernel.org> Cc: <stable@vger.kernel.org> Signed-off-by: Bryan O'Donoghue <bryan.odonoghue@linaro.org> --- drivers/media/platform/qcom/camss/camss-video.c | 6 ------ 1 file changed, 6 deletions(-)