Usage of SA_ONSTACK may want to arrange for a large sigaltstack

[alexcrichton]

I was recently developing Rust support for coz since I was curious how it could help us analyze compile times (in rustc and cargo themselves), and in developing Rust support for coz one thing I was initially baffled by was that coz-run programs would always segfault very quickly! Digging in it looked like the first SIGPROF signal was segfaulting on the first stack access.

Rust programs, by default, configure a pretty small sigaltstack for SIGSEGV so they can print "you just overflowed the stack" if the stack was overflowed. It looks like coz's own usage of SA_ONSTACK for SIGPROF causes coz itself to run on the Rust standard library's tiny stack, which immediately stack overflows. For reference the Rust standard library allocates SIGSTKSZ which I think is around 8k, and the stack frame for the first function called by SIGPROF was around 8k, hence the segfault.

I managed to work around this by setting up a much larger sigaltstack whenever coz is used, but I figured it might be good to report this for anyone else who runs into it!

[ccurtsinger]

I don't see any harm in doing this for all programs run with coz, not just with rust.

[alexcrichton]

Thinking on this a bit more, is SA_ONSTACK still needed with the SIGPROF handler? From some local testing if SA_ONSTACK is used but no stack is configured with sigaltstack, I think it just uses the thread's current stack to deliver the signal, which should be appropriate in most cases for libcoz?