MT: reduce interleaved backtraces in spawn unhandled exceptions

[ysbaddaden]

Instead of doing many individual writes to STDERR that may happen concurrently with other threads, we buffer the message in memory so we can write once.

I tried with 10 000 fibers and 16 threads and I couldn't get any interleaved message. That doesn't mean it can't happen, especially with other concurrent writes to STDERR —which an alternative solution to use a mutex wouldn't prevent it either.

Note: I chose to use a HEAP allocated buffer rather than a stack allocated one because the latter wouldn't be resizeable, and there is no guarantee that the backtrace would fit within 4KB (PIPE_BUF on Linux).

closes #8299

[beta-ziliani]

My only concern is that if the error comes from memory exhaustion, this would not help. OTOH, if the exception comes from stack exhaustion, then a stack allocated buffer wouldn't help either...

[ysbaddaden]

You're right @beta-ziliani. Hum, in case of stack exhaustion this will trigger a segfault that will be handled in a signal handler (no exception raised), but for OOM we will get an exception (which also allocates) that we might be able to catch, so we can use a fixed-size stack buffer.