Michael Paquier wrote:
> Note that I am not convinced about adding the whole set of suggested
> wait events just based on the reason that they seem useful.

The ones in 0002 come from real cases.  Jakub's outage upthread was a
stuck remote syslog (SyslogWrite).  With Andrey's slow boot disk,
backends block on the syslogger pipe or on stderr, depending on
logging_collector (SysloggerWrite, StderrWrite).

> For one, I am not able to check the Windows-specific ones and if they
> actually make sense on their own.

Fair.  EventlogWrite has only been seen in CI runs, and WriteConsoleW
has never been exercised.  I'm fine with dropping 0003 and leaving the
Windows paths for later.

Andrey Borodin wrote:
> Could we reuse the debug nesting check to collect inner/outer call
> sites over a check-world run and identify most useful sites that can
> be converted to nested?

Yes.  A run in September found only two such pairs, but the check
aborts at the first one in each process.  I'll make it record the pairs
instead and post the list here.

I'll hold off on a new version until there are more opinions on the
swap approach.
--
Best regards,
Seongjun

Reply via email to