Run checkpointer and bgwriter in crash recovery.
authorThomas Munro <tmunro@postgresql.org>
Mon, 2 Aug 2021 05:32:20 +0000 (17:32 +1200)
committerThomas Munro <tmunro@postgresql.org>
Mon, 2 Aug 2021 05:32:44 +0000 (17:32 +1200)
Start up the checkpointer and bgwriter during crash recovery (except in
--single mode), as we do for replication.  This wasn't done back in
commit cdd46c76 out of caution.  Now it seems like a better idea to make
the environment as similar as possible in both cases.  There may also be
some performance advantages.

Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Aleksander Alekseev <aleksander@timescale.com>
Tested-by: Jakub Wartak <Jakub.Wartak@tomtom.com>
Discussion: https://postgr.es/m/CA%2BhUKGJ8NRsqgkZEnsnRc2MFROBV-jCnacbYvtpptK2A9YYp9Q%40mail.gmail.com

src/backend/access/transam/xlog.c
src/backend/postmaster/bgwriter.c
src/backend/postmaster/checkpointer.c
src/backend/postmaster/postmaster.c
src/backend/storage/sync/sync.c
src/include/storage/sync.h

index efb3ca273ed9549adb72466920103b7d94353352..f84c0bb01eb8341e17f9ae8d9992a8044a1cc380 100644 (file)
@@ -869,9 +869,6 @@ bool        reachedConsistency = false;
 
 static bool InRedo = false;
 
-/* Have we launched bgwriter during recovery? */
-static bool bgwriterLaunched = false;
-
 /* For WALInsertLockAcquire/Release functions */
 static int MyLockNo = 0;
 static bool holdingAllLocks = false;
@@ -7311,25 +7308,15 @@ StartupXLOG(void)
        /* Also ensure XLogReceiptTime has a sane value */
        XLogReceiptTime = GetCurrentTimestamp();
 
+       /* Allow ProcSendSignal() to find us, for buffer pin wakeups. */
+       PublishStartupProcessInformation();
+
        /*
         * Let postmaster know we've started redo now, so that it can launch
-        * checkpointer to perform restartpoints.  We don't bother during
-        * crash recovery as restartpoints can only be performed during
-        * archive recovery.  And we'd like to keep crash recovery simple, to
-        * avoid introducing bugs that could affect you when recovering after
-        * crash.
-        *
-        * After this point, we can no longer assume that we're the only
-        * process in addition to postmaster!  Also, fsync requests are
-        * subsequently to be handled by the checkpointer, not locally.
+        * the archiver if necessary.
         */
-       if (ArchiveRecoveryRequested && IsUnderPostmaster)
-       {
-           PublishStartupProcessInformation();
-           EnableSyncRequestForwarding();
+       if (IsUnderPostmaster)
            SendPostmasterSignal(PMSIGNAL_RECOVERY_STARTED);
-           bgwriterLaunched = true;
-       }
 
        /*
         * Allow read-only connections immediately if we're consistent
@@ -7903,7 +7890,7 @@ StartupXLOG(void)
         * after we're fully out of recovery mode and already accepting
         * queries.
         */
-       if (bgwriterLaunched)
+       if (ArchiveRecoveryRequested && IsUnderPostmaster)
        {
            if (LocalPromoteIsTriggered)
            {
@@ -7927,7 +7914,11 @@ StartupXLOG(void)
                                  CHECKPOINT_WAIT);
        }
        else
-           CreateCheckPoint(CHECKPOINT_END_OF_RECOVERY | CHECKPOINT_IMMEDIATE);
+       {
+           RequestCheckpoint(CHECKPOINT_END_OF_RECOVERY |
+                             CHECKPOINT_IMMEDIATE |
+                             CHECKPOINT_WAIT);
+       }
    }
 
    if (ArchiveRecoveryRequested)
@@ -12182,7 +12173,7 @@ XLogPageRead(XLogReaderState *xlogreader, XLogRecPtr targetPagePtr, int reqLen,
         * Request a restartpoint if we've replayed too much xlog since the
         * last one.
         */
-       if (bgwriterLaunched)
+       if (ArchiveRecoveryRequested && IsUnderPostmaster)
        {
            if (XLogCheckpointNeeded(readSegNo))
            {
index 715d5195bb671e9c6a1df6dc98ba805b6d4be43a..5584f4bc241e041d6693127dfb1410cfeb1b7908 100644 (file)
@@ -12,9 +12,6 @@
  *
  * As of Postgres 9.2 the bgwriter no longer handles checkpoints.
  *
- * The bgwriter is started by the postmaster as soon as the startup subprocess
- * finishes, or as soon as recovery begins if we are doing archive recovery.
- * It remains alive until the postmaster commands it to terminate.
  * Normal termination is by SIGTERM, which instructs the bgwriter to exit(0).
  * Emergency termination is by SIGQUIT; like any backend, the bgwriter will
  * simply abort and exit on SIGQUIT.
index 75a95f3de7a8fe1cd0f5bb9cb0986d2112dbbb5a..bc9ac7ccfaf088bff13fe23dfd92dc1afb369bd8 100644 (file)
@@ -10,9 +10,6 @@
  * fill WAL segments; the checkpointer itself doesn't watch for the
  * condition.)
  *
- * The checkpointer is started by the postmaster as soon as the startup
- * subprocess finishes, or as soon as recovery begins if we are doing archive
- * recovery.  It remains alive until the postmaster commands it to terminate.
  * Normal termination is by SIGUSR2, which instructs the checkpointer to
  * execute a shutdown checkpoint and then exit(0).  (All backends must be
  * stopped before SIGUSR2 is issued!)  Emergency termination is by SIGQUIT;
index 122c2b05bdb70a5680dccfa8cb971bf5c0cf0b22..00d051d520e085ccdf4858ea110a07e51d6af29e 100644 (file)
@@ -1403,6 +1403,12 @@ PostmasterMain(int argc, char *argv[])
     */
    AddToDataDirLockFile(LOCK_FILE_LINE_PM_STATUS, PM_STATUS_STARTING);
 
+   /* Start bgwriter and checkpointer so they can help with recovery */
+   if (CheckpointerPID == 0)
+       CheckpointerPID = StartCheckpointer();
+   if (BgWriterPID == 0)
+       BgWriterPID = StartBackgroundWriter();
+
    /*
     * We're ready to rock and roll...
     */
@@ -1765,7 +1771,7 @@ ServerLoop(void)
         * fails, we'll just try again later.  Likewise for the checkpointer.
         */
        if (pmState == PM_RUN || pmState == PM_RECOVERY ||
-           pmState == PM_HOT_STANDBY)
+           pmState == PM_HOT_STANDBY || pmState == PM_STARTUP)
        {
            if (CheckpointerPID == 0)
                CheckpointerPID = StartCheckpointer();
@@ -5161,15 +5167,6 @@ sigusr1_handler(SIGNAL_ARGS)
        FatalError = false;
        AbortStartTime = 0;
 
-       /*
-        * Crank up the background tasks.  It doesn't matter if this fails,
-        * we'll just try again later.
-        */
-       Assert(CheckpointerPID == 0);
-       CheckpointerPID = StartCheckpointer();
-       Assert(BgWriterPID == 0);
-       BgWriterPID = StartBackgroundWriter();
-
        /*
         * Start the archiver if we're responsible for (re-)archiving received
         * files.
index bc3ceb27125d46204e2b5e605aa506a250c6ac48..4a2ed414b00ce38c7808432306f969dfa6efbafa 100644 (file)
@@ -129,10 +129,10 @@ InitSync(void)
 {
    /*
     * Create pending-operations hashtable if we need it.  Currently, we need
-    * it if we are standalone (not under a postmaster) or if we are a startup
-    * or checkpointer auxiliary process.
+    * it if we are standalone (not under a postmaster) or if we are a
+    * checkpointer auxiliary process.
     */
-   if (!IsUnderPostmaster || AmStartupProcess() || AmCheckpointerProcess())
+   if (!IsUnderPostmaster || AmCheckpointerProcess())
    {
        HASHCTL     hash_ctl;
 
@@ -589,27 +589,3 @@ RegisterSyncRequest(const FileTag *ftag, SyncRequestType type,
 
    return ret;
 }
-
-/*
- * In archive recovery, we rely on checkpointer to do fsyncs, but we will have
- * already created the pendingOps during initialization of the startup
- * process.  Calling this function drops the local pendingOps so that
- * subsequent requests will be forwarded to checkpointer.
- */
-void
-EnableSyncRequestForwarding(void)
-{
-   /* Perform any pending fsyncs we may have queued up, then drop table */
-   if (pendingOps)
-   {
-       ProcessSyncRequests();
-       hash_destroy(pendingOps);
-   }
-   pendingOps = NULL;
-
-   /*
-    * We should not have any pending unlink requests, since mdunlink doesn't
-    * queue unlink requests when isRedo.
-    */
-   Assert(pendingUnlinks == NIL);
-}
index fbdf34f7620b60963159e0e6fe2ca818defa9083..6fd50cfa7b7be30b7d96cb0f6dd36ddb50736c09 100644 (file)
@@ -60,7 +60,6 @@ extern void SyncPreCheckpoint(void);
 extern void SyncPostCheckpoint(void);
 extern void ProcessSyncRequests(void);
 extern void RememberSyncRequest(const FileTag *ftag, SyncRequestType type);
-extern void EnableSyncRequestForwarding(void);
 extern bool RegisterSyncRequest(const FileTag *ftag, SyncRequestType type,
                                bool retryOnError);