1 files changed, 85 insertions, 1 deletions
diff --git a/fs/xfs/xfs_rw.c b/fs/xfs/xfs_rw.c
index d33e4f5808e..c55e28189ab 100644
--- a/fs/xfs/xfs_rw.c
+++ b/fs/xfs/xfs_rw.c
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
+ * Copyright (c) 2000-2006 Silicon Graphics, Inc.
  * All Rights Reserved.
  *
  * This program is free software; you can redistribute it and/or
@@ -92,6 +92,90 @@ xfs_write_clear_setuid(
 }
 
 /*
+ * Handle logging requirements of various synchronous types of write.
+ */
+int
+xfs_write_sync_logforce(
+	xfs_mount_t	*mp,
+	xfs_inode_t	*ip)
+{
+	int		error = 0;
+
+	/*
+	 * If we're treating this as O_DSYNC and we have not updated the
+	 * size, force the log.
+	 */
+	if (!(mp->m_flags & XFS_MOUNT_OSYNCISOSYNC) &&
+	    !(ip->i_update_size)) {
+		xfs_inode_log_item_t	*iip = ip->i_itemp;
+
+		/*
+		 * If an allocation transaction occurred
+		 * without extending the size, then we have to force
+		 * the log up the proper point to ensure that the
+		 * allocation is permanent.  We can't count on
+		 * the fact that buffered writes lock out direct I/O
+		 * writes - the direct I/O write could have extended
+		 * the size nontransactionally, then finished before
+		 * we started.  xfs_write_file will think that the file
+		 * didn't grow but the update isn't safe unless the
+		 * size change is logged.
+		 *
+		 * Force the log if we've committed a transaction
+		 * against the inode or if someone else has and
+		 * the commit record hasn't gone to disk (e.g.
+		 * the inode is pinned).  This guarantees that
+		 * all changes affecting the inode are permanent
+		 * when we return.
+		 */
+		if (iip && iip->ili_last_lsn) {
+			xfs_log_force(mp, iip->ili_last_lsn,
+					XFS_LOG_FORCE | XFS_LOG_SYNC);
+		} else if (xfs_ipincount(ip) > 0) {
+			xfs_log_force(mp, (xfs_lsn_t)0,
+					XFS_LOG_FORCE | XFS_LOG_SYNC);
+		}
+
+	} else {
+		xfs_trans_t	*tp;
+
+		/*
+		 * O_SYNC or O_DSYNC _with_ a size update are handled
+		 * the same way.
+		 *
+		 * If the write was synchronous then we need to make
+		 * sure that the inode modification time is permanent.
+		 * We'll have updated the timestamp above, so here
+		 * we use a synchronous transaction to log the inode.
+		 * It's not fast, but it's necessary.
+		 *
+		 * If this a dsync write and the size got changed
+		 * non-transactionally, then we need to ensure that
+		 * the size change gets logged in a synchronous
+		 * transaction.
+		 */
+		tp = xfs_trans_alloc(mp, XFS_TRANS_WRITE_SYNC);
+		if ((error = xfs_trans_reserve(tp, 0,
+						XFS_SWRITE_LOG_RES(mp),
+						0, 0, 0))) {
+			/* Transaction reserve failed */
+			xfs_trans_cancel(tp, 0);
+		} else {
+			/* Transaction reserve successful */
+			xfs_ilock(ip, XFS_ILOCK_EXCL);
+			xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
+			xfs_trans_ihold(tp, ip);
+			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
+			xfs_trans_set_sync(tp);
+			error = xfs_trans_commit(tp, 0, NULL);
+			xfs_iunlock(ip, XFS_ILOCK_EXCL);
+		}
+	}
+
+	return error;
+}
+
+/*
  * Force a shutdown of the filesystem instantly while keeping
  * the filesystem consistent. We don't do an unmount here; just shutdown
  * the shop, make sure that absolutely nothing persistent happens to