Horton

Re: Weird failure with latches in curculio on v15

Thomas Munro <thomas.munro@gmail.com>

View thread View on postgr.es
From: Thomas Munro <thomas.munro@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Andres Freund <andres@anarazel.de>, Michael Paquier <michael@paquier.xyz>, Robert Haas <robertmhaas@gmail.com>, Fujii Masao <fujii@postgresql.org>, Nathan Bossart <nathandbossart@gmail.com>, Postgres hackers <pgsql-hackers@lists.postgresql.org>
Date: 2023-02-03T07:34:36Z
Lists: pgsql-hackers 
On Fri, Feb 3, 2023 at 8:24 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Andres Freund <andres@anarazel.de> writes:
> > Ugh, I think I might understand what's happening:
>
> > The signal arrives just after the fork() (within system()). Because we
> > have all our processes configure themselves as process group leaders,
> > and we signal the entire process group (c.f. signal_child()), both the
> > child process and the parent will process the signal. So we'll end up
> > doing a proc_exit() in both. As both are trying to remove themselves
> > from the same PGPROC etc entry, that doesn't end well.
>
> Ugh ...

Yuck, but yeah that makes sense.

> > I don't see how we can solve that properly as long as we use system().
>
> ... but I don't see how that's system()'s fault?  Doing the fork()
> ourselves wouldn't change anything about that.

What if we block signals, fork, then in the child, install the default
SIGTERM handler, then unblock, and then exec the shell?  If SIGTERM is
delivered either before or after exec (but before whatever is loaded
installs a new handler) then the child is terminated, but without
running the handler.  Isn't that what we want here?

Commits

  1. Avoid calling proc_exit() in processes forked by system().

  2. Move extra code out of the Pre/PostRestoreCommand() section.

  3. Revert refactoring of restore command code to shell_restore.c

  4. Refactor code in charge of running shell-based recovery commands

  5. Clean up inconsistent use of fflush().

  6. Report wait events for local shell commands like archive_command.