ocfs2: Handle O_DIRECT when writing to a refcounted cluster.
authorTao Ma <tao.ma@oracle.com>
Thu, 3 Dec 2009 13:55:05 +0000 (21:55 +0800)
committerJoel Becker <joel.becker@oracle.com>
Thu, 31 Dec 2009 03:53:35 +0000 (19:53 -0800)
In case of writing to a refcounted cluster with O_DIRECT,
we need to fall back to buffer write. And when it is finished,
we need to flush the page and the journal as we did for other
O_DIRECT writes.

This patch fix oss bug 1191.
http://oss.oracle.com/bugzilla/show_bug.cgi?id=1191

Signed-off-by: Tao Ma <tao.ma@oracle.com>
Tested-by: Tristan Ye <tristan.ye@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
fs/ocfs2/file.c

index 3d30a1c974a89dbb4429d91d453a1c46d185bb55..06ccf6a86d35ca01a8f80e5a997a30ea8820f30a 100644 (file)
@@ -1772,7 +1772,8 @@ static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
                                         loff_t *ppos,
                                         size_t count,
                                         int appending,
-                                        int *direct_io)
+                                        int *direct_io,
+                                        int *has_refcount)
 {
        int ret = 0, meta_level = 0;
        struct inode *inode = dentry->d_inode;
@@ -1833,6 +1834,8 @@ static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
                                                               saved_pos,
                                                               count,
                                                               &meta_level);
+                       if (has_refcount)
+                               *has_refcount = 1;
                }
 
                if (ret < 0) {
@@ -1856,6 +1859,10 @@ static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
                        break;
                }
 
+               if (has_refcount && *has_refcount == 1) {
+                       *direct_io = 0;
+                       break;
+               }
                /*
                 * Allowing concurrent direct writes means
                 * i_size changes wouldn't be synchronized, so
@@ -1899,7 +1906,7 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
                                    loff_t pos)
 {
        int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
-       int can_do_direct;
+       int can_do_direct, has_refcount = 0;
        ssize_t written = 0;
        size_t ocount;          /* original count */
        size_t count;           /* after file limit checks */
@@ -1942,7 +1949,7 @@ relock:
        can_do_direct = direct_io;
        ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
                                            iocb->ki_left, appending,
-                                           &can_do_direct);
+                                           &can_do_direct, &has_refcount);
        if (ret < 0) {
                mlog_errno(ret);
                goto out;
@@ -2006,14 +2013,16 @@ out_dio:
        /* buffered aio wouldn't have proper lock coverage today */
        BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
 
-       if ((file->f_flags & O_DSYNC && !direct_io) || IS_SYNC(inode)) {
+       if ((file->f_flags & O_DSYNC && !direct_io) || IS_SYNC(inode) ||
+           (file->f_flags & O_DIRECT && has_refcount)) {
                ret = filemap_fdatawrite_range(file->f_mapping, pos,
                                               pos + count - 1);
                if (ret < 0)
                        written = ret;
 
                if (!ret && (old_size != i_size_read(inode) ||
-                   old_clusters != OCFS2_I(inode)->ip_clusters)) {
+                   old_clusters != OCFS2_I(inode)->ip_clusters ||
+                   has_refcount)) {
                        ret = jbd2_journal_force_commit(osb->journal->j_journal);
                        if (ret < 0)
                                written = ret;
@@ -2062,7 +2071,7 @@ static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
        int ret;
 
        ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
-                                           sd->total_len, 0, NULL);
+                                           sd->total_len, 0, NULL, NULL);
        if (ret < 0) {
                mlog_errno(ret);
                return ret;