Use vectored I/O to fill new WAL segments.
authorThomas Munro <tmunro@postgresql.org>
Mon, 11 Jan 2021 01:41:39 +0000 (14:41 +1300)
committerThomas Munro <tmunro@postgresql.org>
Mon, 11 Jan 2021 02:28:31 +0000 (15:28 +1300)
Instead of making many block-sized write() calls to fill a new WAL file
with zeroes, make a smaller number of pwritev() calls (or various
emulations).  The actual number depends on the OS's IOV_MAX, which
PG_IOV_MAX currently caps at 32.  That means we'll write 256kB per call
on typical systems.  We may want to tune the number later with more
experience.

Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKGJA%2Bu-220VONeoREBXJ9P3S94Y7J%2BkqCnTYmahvZJwM%3Dg%40mail.gmail.com

src/backend/access/transam/xlog.c

index ede93ad7fdde22a3e0f03a9a9d855ea8f9803971..b18257c1980784385062de95119ea76875c9d9fd 100644 (file)
@@ -48,6 +48,7 @@
 #include "pg_trace.h"
 #include "pgstat.h"
 #include "port/atomics.h"
+#include "port/pg_iovec.h"
 #include "postmaster/bgwriter.h"
 #include "postmaster/startup.h"
 #include "postmaster/walwriter.h"
@@ -3270,7 +3271,6 @@ XLogFileInit(XLogSegNo logsegno, bool *use_existent, bool use_lock)
        XLogSegNo       installed_segno;
        XLogSegNo       max_segno;
        int                     fd;
-       int                     nbytes;
        int                     save_errno;
 
        XLogFilePath(path, ThisTimeLineID, logsegno, wal_segment_size);
@@ -3317,6 +3317,9 @@ XLogFileInit(XLogSegNo logsegno, bool *use_existent, bool use_lock)
        save_errno = 0;
        if (wal_init_zero)
        {
+               struct iovec iov[PG_IOV_MAX];
+               int                     blocks;
+
                /*
                 * Zero-fill the file.  With this setting, we do this the hard way to
                 * ensure that all the file space has really been allocated.  On
@@ -3326,15 +3329,28 @@ XLogFileInit(XLogSegNo logsegno, bool *use_existent, bool use_lock)
                 * indirect blocks are down on disk.  Therefore, fdatasync(2) or
                 * O_DSYNC will be sufficient to sync future writes to the log file.
                 */
-               for (nbytes = 0; nbytes < wal_segment_size; nbytes += XLOG_BLCKSZ)
+
+               /* Prepare to write out a lot of copies of our zero buffer at once. */
+               for (int i = 0; i < lengthof(iov); ++i)
                {
-                       errno = 0;
-                       if (write(fd, zbuffer.data, XLOG_BLCKSZ) != XLOG_BLCKSZ)
+                       iov[i].iov_base = zbuffer.data;
+                       iov[i].iov_len = XLOG_BLCKSZ;
+               }
+
+               /* Loop, writing as many blocks as we can for each system call. */
+               blocks = wal_segment_size / XLOG_BLCKSZ;
+               for (int i = 0; i < blocks;)
+               {
+                       int             iovcnt = Min(blocks - i, lengthof(iov));
+                       off_t           offset = i * XLOG_BLCKSZ;
+
+                       if (pg_pwritev_with_retry(fd, iov, iovcnt, offset) < 0)
                        {
-                               /* if write didn't set errno, assume no disk space */
-                               save_errno = errno ? errno : ENOSPC;
+                               save_errno = errno;
                                break;
                        }
+
+                       i += iovcnt;
                }
        }
        else