Fix crash in autoprewarm.

nathan-bossart · nathan-bossart · commit bcd5b4bcbefc · 2024-01-23T14:20:14.000-06:00
Commit abb0b4f moved the shared state for autoprewarm to a dynamic shared memory (DSM) segment, but it left apw_detach_shmem() in the on_shmem_exit callback list for the autoprewarm leader process. This is a problem because shmem_exit() detaches all the DSM segments prior to calling the on_shmem_exit callbacks, thus producing segfaults in the exit path for the autoprewarm leader process. To fix, move apw_detach_shmem() to the before_shmem_exit callback list. This commit also adds a check to pg_prewarm's test that the server shut down normally. It might be worth making this a common check for all shutdowns in TAP tests, but that is left as a future exercise. Reported-by: Andres Freund Reviewed-by: Andres Freund, Álvaro Herrera Discussion: https://postgr.es/m/20240122204117.swton324xcoodnyi%40awork3.anarazel.de
diff --git a/contrib/pg_prewarm/autoprewarm.c b/contrib/pg_prewarm/autoprewarm.c
@@ -164,8 +164,14 @@ autoprewarm_main(Datum main_arg)
 	if (apw_init_shmem())
 		first_time = false;
 
-	/* Set on-detach hook so that our PID will be cleared on exit. */
-	on_shmem_exit(apw_detach_shmem, 0);
+	/*
+	 * Set on-detach hook so that our PID will be cleared on exit.
+	 *
+	 * NB: Autoprewarm's state is stored in a DSM segment, and DSM segments
+	 * are detached before calling the on_shmem_exit callbacks, so we must put
+	 * apw_detach_shmem in the before_shmem_exit callback list.
+	 */
+	before_shmem_exit(apw_detach_shmem, 0);
 
 	/*
 	 * Store our PID in the shared memory area --- unless there's already
diff --git a/contrib/pg_prewarm/t/001_basic.pl b/contrib/pg_prewarm/t/001_basic.pl
@@ -55,4 +55,10 @@
 
 $node->stop;
 
+# control file should indicate normal shut down
+command_like(
+	[ 'pg_controldata', $node->data_dir() ],
+	qr/Database cluster state:\s*shut down/,
+	'cluster shut down normally');
+
 done_testing();
