Hi.
Here's a variation of the earlier patch that follows all links in
PGDATA. Does this look more like what you had in mind?
-- Abhijit
>From d86888b0d2f5a3a57027d26ce050a3bbb58670d3 Mon Sep 17 00:00:00 2001
From: Abhijit Menon-Sen <a...@2ndquadrant.com>
Date: Thu, 6 Nov 2014 00:45:56 +0530
Subject: Recursively fsync PGDATA at startup if needed
It's needed if we need to perform crash recovery or if fsync was
disabled at some point while the server was running earlier (and
we must store that information in the control file).
This is so that we don't lose older unflushed writes in the event of
a power failure after crash recovery, where more recent writes are
preserved.
See 20140918083148.ga17...@alap3.anarazel.de for details.
---
src/backend/access/transam/xlog.c | 168 ++++++++++++++++++++++++++++++++
src/bin/pg_controldata/pg_controldata.c | 2 +
src/include/catalog/pg_control.h | 8 +-
3 files changed, 177 insertions(+), 1 deletion(-)
diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c
index 2580996..263d2d0 100644
--- a/src/backend/access/transam/xlog.c
+++ b/src/backend/access/transam/xlog.c
@@ -835,6 +835,8 @@ static void WALInsertLockAcquireExclusive(void);
static void WALInsertLockRelease(void);
static void WALInsertLockUpdateInsertingAt(XLogRecPtr insertingAt);
+static void fsync_pgdata(char *datadir);
+
/*
* Insert an XLOG record represented by an already-constructed chain of data
* chunks. This is a low-level routine; to construct the WAL record header
@@ -4811,6 +4813,7 @@ BootStrapXLOG(void)
ControlFile->checkPoint = checkPoint.redo;
ControlFile->checkPointCopy = checkPoint;
ControlFile->unloggedLSN = 1;
+ ControlFile->fsync_disabled = false;
/* Set important parameter values for use when replaying WAL */
ControlFile->MaxConnections = MaxConnections;
@@ -5868,6 +5871,27 @@ StartupXLOG(void)
ereport(FATAL,
(errmsg("control file contains invalid data")));
+ /*
+ * If we need to perform crash recovery, we issue an fsync on the
+ * data directory and its contents to try to ensure that any data
+ * written before the crash are flushed to disk. Otherwise a power
+ * failure in the near future might cause earlier unflushed writes
+ * to be lost, even though more recent data written to disk from
+ * here on would be persisted.
+ *
+ * We also do this if the control file indicates that fsync was
+ * disabled at some point while the server was running earlier.
+ */
+
+ if (enableFsync &&
+ (ControlFile->fsync_disabled ||
+ (ControlFile->state != DB_SHUTDOWNED &&
+ ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)))
+ {
+ fsync_pgdata(data_directory);
+ ControlFile->fsync_disabled = false;
+ }
+
if (ControlFile->state == DB_SHUTDOWNED)
{
/* This is the expected case, so don't be chatty in standalone mode */
@@ -8236,6 +8260,8 @@ CreateCheckPoint(int flags)
/* crash recovery should always recover to the end of WAL */
ControlFile->minRecoveryPoint = InvalidXLogRecPtr;
ControlFile->minRecoveryPointTLI = 0;
+ if (!enableFsync)
+ ControlFile->fsync_disabled = true;
/*
* Persist unloggedLSN value. It's reset on crash recovery, so this goes
@@ -11107,3 +11133,145 @@ SetWalWriterSleeping(bool sleeping)
XLogCtl->WalWriterSleeping = sleeping;
SpinLockRelease(&XLogCtl->info_lck);
}
+
+/*
+ * Hint to the OS that it should get ready to fsync() this file.
+ *
+ * Adapted from pre_sync_fname in initdb.c
+ */
+static void
+pre_sync_fname(char *fname, bool isdir)
+{
+ int fd;
+
+ fd = open(fname, O_RDONLY | PG_BINARY);
+
+ /*
+ * Some OSs don't allow us to open directories at all (Windows returns
+ * EACCES)
+ */
+ if (fd < 0 && isdir && (errno == EISDIR || errno == EACCES))
+ return;
+
+ if (fd < 0)
+ ereport(FATAL,
+ (errmsg("could not open file \"%s\" before fsync",
+ fname)));
+
+ pg_flush_data(fd, 0, 0);
+
+ close(fd);
+}
+
+/*
+ * walkdir: recursively walk a directory, applying the action to each
+ * regular file and directory (including the named directory itself)
+ * and following symbolic links.
+ */
+static void
+walkdir(char *path, void (*action) (char *fname, bool isdir))
+{
+ DIR *dir;
+ struct dirent *de;
+
+ dir = AllocateDir(path);
+ while ((de = ReadDir(dir, path)) != NULL)
+ {
+ char subpath[MAXPGPATH];
+ struct stat fst;
+
+ CHECK_FOR_INTERRUPTS();
+
+ if (strcmp(de->d_name, ".") == 0 ||
+ strcmp(de->d_name, "..") == 0)
+ continue;
+
+ snprintf(subpath, MAXPGPATH, "%s/%s", path, de->d_name);
+
+ if (lstat(subpath, &fst) < 0)
+ ereport(ERROR,
+ (errcode_for_file_access(),
+ errmsg("could not stat file \"%s\": %m", subpath)));
+
+ if (S_ISREG(fst.st_mode))
+ (*action) (subpath, false);
+ else if (S_ISDIR(fst.st_mode))
+ walkdir(subpath, action);
+#ifndef WIN32
+ else if (S_ISLNK(fst.st_mode))
+#else
+ else if (pg_win32_is_junction(subpath))
+#endif
+ {
+#if defined(HAVE_READLINK) || defined(WIN32)
+ char linkpath[MAXPGPATH];
+ int len;
+ struct stat lst;
+
+ len = readlink(subpath, linkpath, sizeof(linkpath)-1);
+ if (len < 0)
+ ereport(ERROR,
+ (errcode_for_file_access(),
+ errmsg("could not read symbolic link \"%s\": %m",
+ subpath)));
+
+ if (len >= sizeof(linkpath)-1)
+ ereport(ERROR,
+ (errmsg("symbolic link \"%s\" target is too long",
+ subpath)));
+
+ linkpath[len] = '\0';
+
+ if (lstat(linkpath, &lst) == 0)
+ {
+ if (S_ISREG(lst.st_mode))
+ (*action) (linkpath, false);
+ else if (S_ISDIR(lst.st_mode))
+ walkdir(subpath, action);
+ }
+ else if (errno != ENOENT)
+ ereport(ERROR,
+ (errcode_for_file_access(),
+ errmsg("could not stat file \"%s\": %m", linkpath)));
+#else
+ ereport(WARNING,
+ (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("this platform does not support symbolic links; ignoring \"%s\"",
+ subpath)));
+#endif
+ }
+ }
+ FreeDir(dir);
+
+ (*action) (path, true);
+}
+
+/*
+ * Issue fsync recursively on PGDATA and all its contents.
+ *
+ * Adapted from perform_fsync in initdb.c
+ */
+static void
+fsync_pgdata(char *datadir)
+{
+ if (!enableFsync)
+ return;
+
+ /*
+ * If possible, hint to the kernel that we're soon going to fsync
+ * the data directory and its contents.
+ */
+#if defined(HAVE_SYNC_FILE_RANGE) || \
+ (defined(USE_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED))
+ walkdir(datadir, pre_sync_fname);
+#endif
+
+ /*
+ * Now we do the fsync()s in the same order.
+ *
+ * It's important to fsync the destination directory itself as individual
+ * file fsyncs don't guarantee that the directory entry for the file is
+ * synced.
+ */
+ walkdir(datadir, fsync_fname);
+}
diff --git a/src/bin/pg_controldata/pg_controldata.c b/src/bin/pg_controldata/pg_controldata.c
index d8cfe5e..e99014f 100644
--- a/src/bin/pg_controldata/pg_controldata.c
+++ b/src/bin/pg_controldata/pg_controldata.c
@@ -290,6 +290,8 @@ main(int argc, char *argv[])
(uint32) ControlFile.backupEndPoint);
printf(_("End-of-backup record required: %s\n"),
ControlFile.backupEndRequired ? _("yes") : _("no"));
+ printf(_("Fsync disabled at runtime: %s\n"),
+ ControlFile.fsync_disabled ? _("yes") : _("no"));
printf(_("Current wal_level setting: %s\n"),
wal_level_str(ControlFile.wal_level));
printf(_("Current wal_log_hints setting: %s\n"),
diff --git a/src/include/catalog/pg_control.h b/src/include/catalog/pg_control.h
index 2e4c381..794acbc 100644
--- a/src/include/catalog/pg_control.h
+++ b/src/include/catalog/pg_control.h
@@ -21,7 +21,7 @@
/* Version identifier for this pg_control format */
-#define PG_CONTROL_VERSION 942
+#define PG_CONTROL_VERSION 943
/*
* Body of CheckPoint XLOG records. This is declared here because we keep
@@ -182,6 +182,12 @@ typedef struct ControlFileData
bool track_commit_timestamp;
/*
+ * Indicates whether fsync was ever disabled while the server was
+ * running. Tested and set at checkpoints, reset at startup.
+ */
+ bool fsync_disabled;
+
+ /*
* This data is used to check for hardware-architecture compatibility of
* the database and the backend executable. We need not check endianness
* explicitly, since the pg_control version will surely look wrong to a
--
1.9.1
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
