summaryrefslogtreecommitdiff
path: root/test/signal-threads.cpp
diff options
context:
space:
mode:
authorLaurent Pinchart <laurent.pinchart@ideasonboard.com>2021-08-27 18:15:13 +0300
committerLaurent Pinchart <laurent.pinchart@ideasonboard.com>2021-09-01 13:09:17 +0300
commitc0ca2cbdc2dfcca50d7bc71a93629ca8e02b1682 (patch)
tree820d9e27baac59f5bdf4cff7d3c10265c086c8e6 /test/signal-threads.cpp
parent29b372a56af2978b9f6bd428bdaaeff35bcca089 (diff)
libcamera: base: log: Don't crash when logging after Logger destruction
libcamera isn't supposed to log messages after the logger is destroyed, as the global logger instance is destroyed after the main() function returns, and the camera manager is supposed to have been stopped and destroyed before that. This rule is difficult to enforce in the V4L2 compat implementation, as there is no location where we can destroy the camera manager manually before the logger is destroyed. This results in a use-after-free condition when the camera manager gets stopped during destruction. Fix it by not trying to print log messages when the global logger instance has been destroyed. This is a bit of a hack, but hopefully not too bad. There could be race conditions when using a CameraManager instance that is destroyed as part of the destruction of global variables (like the V4L2 compat layer does, it wraps CameraManager in a singleton V4L2CompatManager class, and destroys it when V4L2CompatManager is destroyed) as the CameraManager thread will still be running when the logger gets destroyed, but this doesn't cause any regression as we destroy the logger without any safeguard measure today anyway. There are other options that could be considered. Forcing destruction of the logger after the camera manager in the V4L2 compat layer is one of them, but turned out to be difficult. For instance care would need to be taken *not* to log any message in the mmap() wrapper if the fd doesn't match a wrapped camera, as mmap() is called very early in the initialization process, before libcamera and the logger get initialized. The resulting implementation would likely be fairly complex. Another option could be to wrap the logger with a shared pointer, and keep a reference to it in CameraManager. That's more intrusive, and it's not clear if it would be worth it. Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> Reviewed-by: Paul Elder <paul.elder@ideasonboard.com>
Diffstat (limited to 'test/signal-threads.cpp')
0 files changed, 0 insertions, 0 deletions