Re: [HACKERS] PATCH: Keep one postmaster monitoring pipe per process

Marco Pfatschbacher <marco_pfatschbacher@genua.de>

From: Marco Pfatschbacher <Marco_Pfatschbacher@genua.de>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Thomas Munro <thomas.munro@enterprisedb.com>, Pg Hackers <pgsql-hackers@postgresql.org>
Date: 2017-12-04T16:55:25Z
Lists: pgsql-hackers
On Thu, Sep 15, 2016 at 04:40:00PM -0400, Tom Lane wrote:
> Thomas Munro <thomas.munro@enterprisedb.com> writes:
> > Very interesting.  Perhaps that is why NetBSD shows a speedup with the
> > kqueue patch[1] but FreeBSD doesn't.  I guess that if I could get the
> > kqueue patch to perform better on large FreeBSD systems, it would also
> > be a solution to this problem.
> 
> I just noticed that kqueue appears to offer a solution to this problem,
> ie one of the things you can wait for is exit of another process (named
> by PID, looks like).  If that's portable to all kqueue platforms, then
> integrating a substitute for the postmaster death pipe might push that
> patch over the hump to being a net win.
> 
> 			regards, tom lane

Hi,

sorry for the long silence.
I forgot about this after being on vacation.

kqueue does indeed support EVFILT_PROC like you said.
But using this effectively, would need a separate monitoring process,
because we cannot use poll(2) and kevent(2) together within the same loop.

But how about this:
We just use getppid to see whether the process id of our parent
has changed to 1 (init).
Just like calling read on the pipe, that's just one systemcall.

I did not bother to weed the postmaster_alive_fds yet, but
this change works fine for me.

Regards,
   Marco

diff --git a/src/backend/storage/ipc/pmsignal.c b/src/backend/storage/ipc/pmsignal.c
index 85db6b21f8..a0596c61fd 100644
--- a/src/backend/storage/ipc/pmsignal.c
+++ b/src/backend/storage/ipc/pmsignal.c
@@ -272,21 +272,16 @@ bool
 PostmasterIsAlive(void)
 {
 #ifndef WIN32
-	char		c;
-	ssize_t		rc;
+	pid_t		ppid;
 
-	rc = read(postmaster_alive_fds[POSTMASTER_FD_WATCH], &c, 1);
-	if (rc < 0)
+	ppid = getppid();
+	if (ppid == 1)
 	{
-		if (errno == EAGAIN || errno == EWOULDBLOCK)
-			return true;
-		else
-			elog(FATAL, "read on postmaster death monitoring pipe failed: %m");
+		elog(FATAL, "postmaster died: our ppid changed to init");
+		return false;
 	}
-	else if (rc > 0)
-		elog(FATAL, "unexpected data in postmaster death monitoring pipe");
 
-	return false;
+	return true;
 #else							/* WIN32 */
 	return (WaitForSingleObject(PostmasterHandle, 0) == WAIT_TIMEOUT);
 #endif							/* WIN32 */


Commits

  1. Add WL_EXIT_ON_PM_DEATH pseudo-event.

  2. Use signals for postmaster death on Linux.