ocfs2: update inode size after zeroing the hole
authorJunxiao Bi <junxiao.bi@oracle.com>
Mon, 10 Feb 2014 22:25:53 +0000 (14:25 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 11 Feb 2014 00:01:43 +0000 (16:01 -0800)
fs-writeback will release the dirty pages without page lock whose offset
are over inode size, the release happens at
block_write_full_page_endio().  If not update, dirty pages in file holes
may be released before flushed to the disk, then file holes will contain
some non-zero data, this will cause sparse file md5sum error.

To reproduce the bug, find a big sparse file with many holes, like vm
image file, its actual size should be bigger than available mem size to
make writeback work more frequently, tar it with -S option, then keep
untar it and check its md5sum again and again until you get a wrong
md5sum.

Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Younger Liu <younger.liu@huawei.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/ocfs2/file.c

index 9148353..8450262 100644 (file)
@@ -716,7 +716,8 @@ leave:
  * While a write will already be ordering the data, a truncate will not.
  * Thus, we need to explicitly order the zeroed pages.
  */
-static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
+static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode,
+                                               struct buffer_head *di_bh)
 {
        struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
        handle_t *handle = NULL;
@@ -733,7 +734,14 @@ static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode)
        }
 
        ret = ocfs2_jbd2_file_inode(handle, inode);
-       if (ret < 0)
+       if (ret < 0) {
+               mlog_errno(ret);
+               goto out;
+       }
+
+       ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
+                                     OCFS2_JOURNAL_ACCESS_WRITE);
+       if (ret)
                mlog_errno(ret);
 
 out:
@@ -749,7 +757,7 @@ out:
  * to be too fragile to do exactly what we need without us having to
  * worry about recursive locking in ->write_begin() and ->write_end(). */
 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
-                                u64 abs_to)
+                                u64 abs_to, struct buffer_head *di_bh)
 {
        struct address_space *mapping = inode->i_mapping;
        struct page *page;
@@ -757,6 +765,7 @@ static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
        handle_t *handle = NULL;
        int ret = 0;
        unsigned zero_from, zero_to, block_start, block_end;
+       struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 
        BUG_ON(abs_from >= abs_to);
        BUG_ON(abs_to > (((u64)index + 1) << PAGE_CACHE_SHIFT));
@@ -799,7 +808,8 @@ static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
                }
 
                if (!handle) {
-                       handle = ocfs2_zero_start_ordered_transaction(inode);
+                       handle = ocfs2_zero_start_ordered_transaction(inode,
+                                                                     di_bh);
                        if (IS_ERR(handle)) {
                                ret = PTR_ERR(handle);
                                handle = NULL;
@@ -816,8 +826,22 @@ static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
                        ret = 0;
        }
 
-       if (handle)
+       if (handle) {
+               /*
+                * fs-writeback will release the dirty pages without page lock
+                * whose offset are over inode size, the release happens at
+                * block_write_full_page_endio().
+                */
+               i_size_write(inode, abs_to);
+               inode->i_blocks = ocfs2_inode_sector_count(inode);
+               di->i_size = cpu_to_le64((u64)i_size_read(inode));
+               inode->i_mtime = inode->i_ctime = CURRENT_TIME;
+               di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
+               di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
+               di->i_mtime_nsec = di->i_ctime_nsec;
+               ocfs2_journal_dirty(handle, di_bh);
                ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
+       }
 
 out_unlock:
        unlock_page(page);
@@ -913,7 +937,7 @@ out:
  * has made sure that the entire range needs zeroing.
  */
 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
-                                  u64 range_end)
+                                  u64 range_end, struct buffer_head *di_bh)
 {
        int rc = 0;
        u64 next_pos;
@@ -929,7 +953,7 @@ static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
                next_pos = (zero_pos & PAGE_CACHE_MASK) + PAGE_CACHE_SIZE;
                if (next_pos > range_end)
                        next_pos = range_end;
-               rc = ocfs2_write_zero_page(inode, zero_pos, next_pos);
+               rc = ocfs2_write_zero_page(inode, zero_pos, next_pos, di_bh);
                if (rc < 0) {
                        mlog_errno(rc);
                        break;
@@ -975,7 +999,7 @@ int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
                        range_end = zero_to_size;
 
                ret = ocfs2_zero_extend_range(inode, range_start,
-                                             range_end);
+                                             range_end, di_bh);
                if (ret) {
                        mlog_errno(ret);
                        break;