When we're doing a state change from PLAYING to NULL, first we invoke
gst_rtspsrc_loop_send_cmd_and_wait (..., CMD_CLOSE, ...) during
PAUSED_TO_READY which will schedule a TEARDOWN to happen async on the
task thread.
The task thread will call gst_rtspsrc_close(), which will send the
TEARDOWN and once it's complete, it will call gst_rtspsrc_cleanup()
without taking any locks, which frees src->streams.
At the same time however, the state change in the app thread will
progress further and in READY_TO_NULL it will call gst_rtspsrc_stop()
which calls gst_rtspsrc_close() a second time, which accesses
src->streams (without a lock again), which leads to simultaneous
access of src->streams, and a segfault.
So the state change and the cleanup are racing, but they almost always
complete sequentially. Either the cleanup sets src->streams to NULL or
_stop() completes first. Very rarely, _stop() can start while
src->streams is being freed in a for loop. That causes the segfault.
This is unlocked access is unfixable with more locking, it just leads
to deadlocks. This pattern has been observed in rtspsrc a lot: state
changes and cleanup in the element are unfixably racy, and that
foundational issue is being addressed separately via a rewrite.
The bandage fix here is to prevent gst_rtspsrc_stop() from accessing
src->streams after it has already been freed by setting src->state to
INVALID.
Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/6302>
3bed35c342