Martin Steghöfer authored Feb 05, 2015 and Thomas Daede committed Jan 21, 2018

The following interleaving can cause problems:
The interrupt for SIGINT could set sig_request.cancel right after the consumer thread
checks it ("buf->cancel_flag || sig_request.cancel"), but before the consumer
thread enters the mutex in order to wait for more data ("COND_WAIT"). If the producer
thread reacts to the new situation immediately by exiting (which it can do because the
consumer thread hasn't entered the mutex yet), the consumer thread initiates a COND_WAIT
that might never be fulfilled because the producer thread is gone. The producer's intention
to kill the consumer using buffer_thread_kill doesn't work because it was executed after
the consumer's check for buf->cancel_flag and before the consumer's COND_WAIT.

Fix the problem by putting the cancelation flag checks in the same mutex lock zone as
the "COND_WAIT".