ocfs2: Commit journal on sync writes
authorMark Fasheh <mark.fasheh@oracle.com>
Thu, 18 Oct 2007 21:14:45 +0000 (14:14 -0700)
committerMark Fasheh <mark.fasheh@oracle.com>
Tue, 6 Nov 2007 23:32:00 +0000 (15:32 -0800)
We're missing a meta data commit for extending sync writes. In thoery, write
could return with the meta data required to read the data uncommitted to
disk. Fix that by detecting an allocating write and forcing a journal commit
in the sync case.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
fs/ocfs2/file.c

index f92fe91..bbac7cd 100644 (file)
@@ -1891,9 +1891,11 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
        ssize_t written = 0;
        size_t ocount;          /* original count */
        size_t count;           /* after file limit checks */
-       loff_t *ppos = &iocb->ki_pos;
+       loff_t old_size, *ppos = &iocb->ki_pos;
+       u32 old_clusters;
        struct file *file = iocb->ki_filp;
        struct inode *inode = file->f_path.dentry->d_inode;
+       struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 
        mlog_entry("(0x%p, %u, '%.*s')\n", file,
                   (unsigned int)nr_segs,
@@ -1949,6 +1951,13 @@ relock:
                goto relock;
        }
 
+       /*
+        * To later detect whether a journal commit for sync writes is
+        * necessary, we sample i_size, and cluster count here.
+        */
+       old_size = i_size_read(inode);
+       old_clusters = OCFS2_I(inode)->ip_clusters;
+
        /* communicate with ocfs2_dio_end_io */
        ocfs2_iocb_set_rw_locked(iocb, rw_level);
 
@@ -1978,6 +1987,21 @@ 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_SYNC && !direct_io) || IS_SYNC(inode)) {
+               /*
+                * The generic write paths have handled getting data
+                * to disk, but since we don't make use of the dirty
+                * inode list, a manual journal commit is necessary
+                * here.
+                */
+               if (old_size != i_size_read(inode) ||
+                   old_clusters != OCFS2_I(inode)->ip_clusters) {
+                       ret = journal_force_commit(osb->journal->j_journal);
+                       if (ret < 0)
+                               written = ret;
+               }
+       }
+
        /* 
         * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
         * function pointer which is called when o_direct io completes so that