]> granicus.if.org Git - postgresql/commitdiff
Recursively fsync() the data directory after a crash.
authorRobert Haas <rhaas@postgresql.org>
Mon, 4 May 2015 18:13:53 +0000 (14:13 -0400)
committerRobert Haas <rhaas@postgresql.org>
Mon, 4 May 2015 18:13:53 +0000 (14:13 -0400)
Otherwise, if there's another crash, some writes from after the first
crash might make it to disk while writes from before the crash fail
to make it to disk.  This could lead to data corruption.

Back-patch to all supported versions.

Abhijit Menon-Sen, reviewed by Andres Freund and slightly revised
by me.

src/backend/access/transam/xlog.c
src/backend/storage/file/fd.c
src/include/storage/fd.h

index 6cf441534ccd07774c1c58d1b38d87055bbdbfc3..92822a192b0f5988e399b7542c6fc3f1163fe736 100644 (file)
@@ -845,6 +845,8 @@ 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
@@ -5910,6 +5912,18 @@ StartupXLOG(void)
                          (errmsg("database system was interrupted; last known up at %s",
                                          str_time(ControlFile->time))));
 
+       /*
+        * If we previously crashed, there might be data which we had written,
+        * intending to fsync it, but which we had not actually fsync'd yet.
+        * Therefore, 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.  To avoid that, fsync the entire
+        * data directory.
+        */
+       if (ControlFile->state != DB_SHUTDOWNED &&
+               ControlFile->state != DB_SHUTDOWNED_IN_RECOVERY)
+               fsync_pgdata(data_directory);
+
        /* This is just to allow attaching to startup process with a debugger */
 #ifdef XLOG_REPLAY_DELAY
        if (ControlFile->state != DB_SHUTDOWNED)
@@ -11123,3 +11137,31 @@ SetWalWriterSleeping(bool sleeping)
        XLogCtl->WalWriterSleeping = sleeping;
        SpinLockRelease(&XLogCtl->info_lck);
 }
+
+/*
+ * Issue fsync recursively on PGDATA and all its contents.
+ */
+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);
+}
index f7967178c694232c44015fa2831d2ab4f0db4684..99dac841ffa5f2f349db897406d91b7512b2548a 100644 (file)
@@ -2439,3 +2439,118 @@ looks_like_temp_rel_name(const char *name)
                return false;
        return true;
 }
+
+/*
+ * Hint to the OS that it should get ready to fsync() this file.
+ *
+ * Adapted from pre_sync_fname in initdb.c
+ */
+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.
+ *
+ * NB: There is another version of walkdir in initdb.c, but that version
+ * behaves differently with respect to symbolic links.  Caveat emptor!
+ */
+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);
+}
index 5e9571c9a1e7bb14df32ad54ca456f5db8c49f20..5b563a6a477c99d22d5162085f6577ef74707ef1 100644 (file)
@@ -114,6 +114,8 @@ extern int  pg_fsync_writethrough(int fd);
 extern int     pg_fdatasync(int fd);
 extern int     pg_flush_data(int fd, off_t offset, off_t amount);
 extern void fsync_fname(char *fname, bool isdir);
+extern void pre_sync_fname(char *fname, bool isdir);
+extern void walkdir(char *path, void (*action) (char *fname, bool isdir));
 
 /* Filename components for OpenTemporaryFile */
 #define PG_TEMP_FILES_DIR "pgsql_tmp"