Crash with -t switcher:excl_init + possible fix (switcher.c)

[armelvil]

I first noticed this issue after rebasing my fork from upstream March 13th -> May 28th.

Switching between inputs, when excl_init is enabled on the switcher, reliably results in the process running the encoder crashing.

After poking at the code for a few hours, I got to this:

After the init loop advances 'tmp' through all devices via vidcap_params_get_next(), tmp points to the NULL-driver stopper node, not the head of the params list. Passing it to vidcap_params_copy_all() stored only a single stopper node instead of the full device chain.

When a subsequent device switch occurred (via capture.data N), vidcap_params_get_nth() returned NULL for all devices, which was passed to initialize_video_capture() and caused a SEGV in vidcap_params_get_next(NULL).

Use the original 'params' pointer (head of list) instead of the exhausted loop variable.

After compilation with this patch and testing, the patched version seems to show expected behavior (switching works, no crash)

Note: I have to use excl_init on the encoder end, as multiple v4l2 inputs without that setting seem to result in some level of concatenation/overwrite between both devices when output to a decoder client. I haven't had time to work out why that is happening yet.

Code change is here:

master...armelvil:UltraGrid:fix/switcher-excl-init-crash

[MartinPulec]

Thanks for looking into it.

It looks to me that you've tracked it correctly. I didn't bisect but it seems to me that it is rather caused by a copy&paster error in 388e87e not the commit specified in the commit log message, but not sure (not so much important for me).

I've converted your code to a pull request and merged.

Note: I have to use excl_init on the encoder end, as multiple v4l2 inputs without that setting seem to result in some level of concatenation/overwrite between both devices when output to a decoder client. I haven't had time to work out why that is happening yet.

Let me know if you have some more info, I've tried with:

uv -t switcher -t testcard:pattern=gradient:size=200x200 -t testcard:size=300x300 -d gl

and it worked fine for me (regardless the :excl_init use, after your fix). I've switched simply by pressing '1' and '2' in the terminal.

I'd like to know, if it isn't v4l2 specific. Maybe your UG arguments may help? But, of course, it can be also elsewhere.

[armelvil]

Glad to be able to help!

On the other issue:

The testcards seem to be working just fine, however if i use two v4l2 inputs (note this is from a patch build from yday):

Cmdline:

--tool uv -t switcher -t v4l2:device=/dev/video0:codec=MJPG:size=2560x1440:fps=30:convert=RGB -t v4l2:device=/dev/video2:codec=MJPG:size=1920x1080:fps=30:convert=RGB -P 5004 -m 9000 192.168.1.32 --control-port 6162

I've played around with some parameters, and found ensuring the resolutions match doesn't seem to help.

UltraGrid 1.10+ (master rev 89a876f built Jun 11 2026 19:50:45)

Display device   : none
Capture device   : switcher
Audio capture    : none
Audio playback   : none
MTU              : 9000 B
Video compression: none
Audio codec      : PCM
Network protocol : UltraGrid RTP
Audio FEC        : none
Video FEC        : none

vidcap_v4l2_init
[V4L cap.] Capturing 2560x1440 @30.00p RGB from 
vidcap_v4l2_init
[V4L cap.] Capturing 1920x1080 @30.00p RGB from 
Control socket listening on port 6162
[switcher] Switched from device 1 to device 1 (v4l2:device=/dev/video0:codec=MJPG:size=2560x1440:fps=30:convert=RGB).
[V4L2 capture] 143 frames in 5.00719 seconds = 28.5589 FPS
[switcher] 152 frames in 5.0285 seconds = 30.2277 FPS
[switcher] Switched from device 1 to device 2 (v4l2:device=/dev/video2:codec=MJPG:size=1920x1080:fps=30:convert=RGB).
[V4L2 capture] 1 frames in 5.5641 seconds = 0.179724 FPS
[switcher] 152 frames in 5.00011 seconds = 30.3994 FPS
[V4L2 capture] 152 frames in 5.03456 seconds = 30.1913 FPS
[V4L cap.] Error converting video: v4l-convert: libjpeg error: End Of Image

[switcher] 149 frames in 5.00814 seconds = 29.7516 FPS
[V4L2 capture] 149 frames in 5.00353 seconds = 29.779 FPS
[V4L cap.] Error converting video: v4l-convert: libjpeg error: Corrupt JPEG data: premature end of data segment

[switcher] 149 frames in 5.00448 seconds = 29.7733 FPS
[V4L2 capture] 149 frames in 5.00741 seconds = 29.7559 FPS
[switcher] 150 frames in 5.00734 seconds = 29.956 FPS
[V4L2 capture] 150 frames in 5.00606 seconds = 29.9637 FPS
[V4L cap.] Error converting video: v4l-convert: libjpeg error: Corrupt JPEG data: premature end of data segment

Sometimes the process starts up just fine with the first input outputting normally, but attempting a switch to the second always results in some "interesting" output (attached bad phone photo)

You can just about make out the second input device's data being interleaved into the stream.

This manifests whether I stream to a client receiver or use direct display output with -d gl as in your example.

I've included a verbose output log during the switch from the "good" device to the "bad" device.
There seems to be a fair amount of

[1781275114.493] [lavc libdav1d @ 0x7f74e40045c0] Error parsing OBU data
[1781275114.493] [lavd] send - Error while decoding frame (rc == -1094995529): Invalid data found when processing input.

But i'm not sure where to look further for the v4l2 side.