Avoid a race condition that could lead to a segmentation fault.

Let's assume the callback function of a callback job returns
JOB_REQUEUE_FAIR in one call and JOB_REQUEUE_NONE in the next. Before
this fix, the thread executing the callback job would requeue the job
before unregistering itself. If there was a context switch right after
the job got requeued, and if the thread that requeued the job never got
resumed until a second thread executed the job and, due to the return
value of JOB_REQUEUE_NONE, destroyed it, then when the first thread
eventually got resumed and tried to lock the mutex to unregister itself
the pointer wouldn't be valid anymore, thus resulting in a segmentation fault.

0 files changed, 0 insertions, 0 deletions