Re: Printing backtrace of postgres processes

On Thu, Feb 08, 2024 at 12:25:18PM +0900, Michael Paquier wrote:
> In HandleLogBacktraceInterrupt(), we don't use backtrace_symbols() and
> rely on backtrace_symbols_fd() to avoid doing malloc() in the signal
> handler as mentioned in [1] back in 2022. Perhaps the part about the
> fact that we don't use backtrace_symbols() should be mentioned
> explicitely in a comment rather than silently implied? That's
> a very important point.

This has been itching me, so I have spent more time reading about
that, and while browsing signal(7) and signal-safety(7), I've first
noticed that this is not safe in the patch:
+ write_stderr("logging current backtrace of process with PID %d:\n",
+ MyProcPid);

Note that there's a write_stderr_signal_safe().

Anyway, I've been digging around the signal-safety of backtrace(3)
(even looking a bit at some GCC code, brrr), and I am under the
impression that backtrace() is just by nature not safe and also
dangerous in signal handlers. One example of issue I've found:
https://github.com/gperftools/gperftools/issues/838

This looks like enough ground to me to reject the patch.
--
Michael