Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Fix fractional vacuum_cost_delay.
authorThomas Munro <tmunro@postgresql.org>
Wed, 15 Mar 2023 00:57:00 +0000 (13:57 +1300)
committerThomas Munro <tmunro@postgresql.org>
Wed, 15 Mar 2023 01:02:49 +0000 (14:02 +1300)
Commit 4753ef37 changed vacuum_delay_point() to use the WaitLatch() API,
to fix the problem that vacuum could keep running for a very long time
after the postmaster died.

Unfortunately, that broke commit caf626b2's support for fractional
vacuum_cost_delay, which shipped in PostgreSQL 12.  WaitLatch() works in
whole milliseconds.

For now, revert the change from commit 4753ef37, but add an explicit
check for postmaster death.  That's an extra system call on systems
other than Linux and FreeBSD, but that overhead doesn't matter much
considering that we willingly went to sleep and woke up again.  (In
later work, we might add higher resolution timeouts to the latch API so
that we could do this with our standard programming pattern, but that
wouldn't be back-patched.)

Back-patch to 14, where commit 4753ef37 arrived.

Reported-by: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://postgr.es/m/CAAKRu_b-q0hXCBUCAATh0Z4Zi6UkiC0k2DFgoD3nC-r3SkR3tg%40mail.gmail.com

src/backend/commands/vacuum.c

index dbdfe8bd2d419070d3ba07b04dbec45f113f9297..e32e7febf9e0501d55664063acc9b63d6f070fa5 100644 (file)
@@ -49,6 +49,7 @@
 #include "postmaster/bgworker_internals.h"
 #include "storage/bufmgr.h"
 #include "storage/lmgr.h"
+#include "storage/pmsignal.h"
 #include "storage/proc.h"
 #include "storage/procarray.h"
 #include "utils/acl.h"
@@ -2223,11 +2224,18 @@ vacuum_delay_point(void)
        if (msec > VacuumCostDelay * 4)
            msec = VacuumCostDelay * 4;
 
-       (void) WaitLatch(MyLatch,
-                        WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
-                        msec,
-                        WAIT_EVENT_VACUUM_DELAY);
-       ResetLatch(MyLatch);
+       pgstat_report_wait_start(WAIT_EVENT_VACUUM_DELAY);
+       pg_usleep(msec * 1000);
+       pgstat_report_wait_end();
+
+       /*
+        * We don't want to ignore postmaster death during very long vacuums
+        * with vacuum_cost_delay configured.  We can't use the usual
+        * WaitLatch() approach here because we want microsecond-based sleep
+        * durations above.
+        */
+       if (IsUnderPostmaster && !PostmasterIsAlive())
+           exit(1);
 
        VacuumCostBalance = 0;