Bug: wait with no timeout with XNU times out after 72 minutes

[aittalam]

Contact Details

davide.eynard@gmail.com, davide@mozilla.ai

What happened?

While testing llamafile cosmocc builds, I realised the idle server was timing out after ~72 minutes.
I tested it both in older (0.6.2) versions, more recent (0.9.3) and mine (0.10.0) and saw the problem occurred only in the last two.

I wrapped up a report with Claude documenting my investigation. I provide that as reference but I realise it's quite verbose, so here's a short summary that is relevant to cosmopolitan: when cosmo_futex_wait is called with no timeout (as happens with e.g. pthread_cond_wait), the XNU code path sets the timeout to UINT32_MAX microseconds instead of 0 (see here). This results in not waiting for infinity, but for UINT32_MAX microseconds (which is ~72 minutes).

Version

4.0.2

What operating system are you seeing the problem on?

Mac

Relevant log output

$ (start=$SECONDS; while true; do elapsed=$((SECONDS - start)); printf "\r⏱ %d min %d sec elapsed" $((elapsed/60)) $((elapsed%60)); sleep 60; done) & timer_pid=$!
$ llamafile --model gemma-3-27b-it-Q5_K_M.gguf --server --verbose

...

⏱ 71 min 1 sec elapsedlibc++abi: terminating due to uncaught exception of type std::__1::system_error: condition_variable wait failed: Connection timed out