Btrfs: fix race between direct io and autodefrag
authorLiu Bo <liubo2009@cn.fujitsu.com>
Thu, 29 Mar 2012 13:57:44 +0000 (09:57 -0400)
committerChris Mason <chris.mason@oracle.com>
Thu, 29 Mar 2012 13:57:44 +0000 (09:57 -0400)
The bug is from running xfstests 209 with autodefrag.

The race is as follows:
       t1                       t2(autodefrag)
   direct IO
     invalidate pagecache
     dio(old data)             add_inode_defrag
     invalidate pagecache
   endio

   direct IO
     invalidate pagecache
                                run_defrag
                                  readpage(old data)
                                  set page dirty (old data)
     dio(new data, rewrite)
     invalidate pagecache (*)
     endio

t2(autodefrag) will get old data into pagecache via readpage and set
pagecache dirty.  Meanwhile, invalidate pagecache(*) will fail due to
dirty flags in pages.  So the old data may be flushed into disk by
flush thread, which will lead to data loss.

And so does the case of user defragment progs.

The patch fixes this race by holding i_mutex when we readpage and set page dirty.

Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com>
Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/ioctl.c

index a979ab7d396746413348fa7ad3ee72e32d59da0e..45910d4b8f659ddbc4bcc69696478dc28864cc38 100644 (file)
@@ -1137,12 +1137,16 @@ int btrfs_defrag_file(struct inode *inode, struct file *file,
                        ra_index += max_cluster;
                }
 
+               mutex_lock(&inode->i_mutex);
                ret = cluster_pages_for_defrag(inode, pages, i, cluster);
-               if (ret < 0)
+               if (ret < 0) {
+                       mutex_unlock(&inode->i_mutex);
                        goto out_ra;
+               }
 
                defrag_count += ret;
                balance_dirty_pages_ratelimited_nr(inode->i_mapping, ret);
+               mutex_unlock(&inode->i_mutex);
 
                if (newer_than) {
                        if (newer_off == (u64)-1)