Use fadvise to prefetch WAL in xlogrecovery

Author: Jakub Wartak

src/backend/access/transam/xlogrecovery.c

@@ -3344,6 +3344,18 @@ XLogPageRead(XLogReaderState *xlogreader, XLogRecPtr targetPagePtr, int reqLen,
 	Assert(targetPageOff == readOff);
 	Assert(reqLen <= readLen);
 
+#if defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_WILLNEED)
+	/*
+	 *  Prefetch next wal blocks to avoid page misses on next read iterations.
+	 */
+#define RACHUNK (128*1024)
+	if (readOff % RACHUNK == 0) {
+		pgstat_report_wait_start(WAIT_EVENT_WAL_PREFETCH);
+		posix_fadvise(readFile, readOff + RACHUNK, RACHUNK, POSIX_FADV_WILLNEED);
+		pgstat_report_wait_end();
+	}
+#endif
+
 	xlogreader->seg.ws_tli = curFileTLI;
 
 	/*

src/backend/utils/activity/wait_event.c

@@ -747,6 +747,9 @@ pgstat_get_wait_io(WaitEventIO w)
 		case WAIT_EVENT_WAL_READ:
 			event_name = "WALRead";
 			break;
+		case WAIT_EVENT_WAL_PREFETCH:
+			event_name = "WALPrefetch";
+			break;
 		case WAIT_EVENT_WAL_SYNC:
 			event_name = "WALSync";
 			break;

src/include/utils/wait_event.h

@@ -232,6 +232,7 @@ typedef enum
 	WAIT_EVENT_WAL_INIT_SYNC,
 	WAIT_EVENT_WAL_INIT_WRITE,
 	WAIT_EVENT_WAL_READ,
+	WAIT_EVENT_WAL_PREFETCH,
 	WAIT_EVENT_WAL_SYNC,
 	WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN,
 	WAIT_EVENT_WAL_WRITE