recursive locks (Duh).
Disable cancellation around the cond_wait() call, since that's also a
cancellation point. Arguably, that would be better handled with
pthread_cleanup_*(), but stubbing those for libc is difficult, and the
current non-exception-based implementation of cleanup handlers is
probably no faster than disabling and reenabling cancellation.
Finally, it only happens in the slow path where the thread is going to
sleep anyway...
takes an additional argument indicating whether this is an internal caller
taking the lock or an external (application) caller. When making an internal
lock, save the current thread cancellation state and disable cancellation
until the matching unlock. This should prevent canccelled threads from exiting
inside of stdio while holding a file lock and potentially leaving other
parts of the FILE structure in an inconsistent state.
trouble is caused by the memory allocation in the mutex initialization,
and uncontested mutexes and condition variables have become faster in the
meantime.
