Fix atomicity violation in try_acquire by using compare_exchange

[quantpoet]

Replaced the racy load-then-store pattern with atomic compare_exchange to eliminate the potential race condition.

This ensures the non-blocking flag is set atomically, preventing multiple threads from simultaneously calling set_nonblocking on the same file descriptor.

[tgross35]

This ensures the non-blocking flag is set atomically, preventing multiple threads from simultaneously calling set_nonblocking on the same file descriptor.

Why is this a problem? It looks like an idempotent fcntl call. Also, wouldn't swap be sufficient?

[tgross35]

Worse actually: with this PR, thread A can swap is_non_blocking from false to true and start fnctl. Thread B may observe is_non_blocking = true before thread A's syscall completes and assume the FIFO is O_NONBLOCK, even though that isn't correct.

The current implementation ensures O_NONBLOCK is set before continuing.

[quantpoet]

Worse actually: with this PR, thread A can swap is_non_blocking from false to true and start fnctl. Thread B may observe is_non_blocking = true before thread A's syscall completes and assume the FIFO is O_NONBLOCK, even though that isn't correct.

The current implementation ensures O_NONBLOCK is set before continuing.

Oh indeed. Thanks for your reply.