udf: Fix race between write(2) and close(2)

Currently write(2) updating i_size and close(2) of the file can race in
such a way that udf_truncate_tail_extent() called from
udf_file_release() sees old i_size but already new extents added by the
running write call. This results in complaints like:
  UDF-fs: warning (device vdb2): udf_truncate_tail_extent: Too long extent
    after EOF in inode 877: i_size: 0 lbcount: 1073739776 extent 0+1073739776
  UDF-fs: error (device vdb2): udf_truncate_tail_extent: Extent after EOF
    in inode 877

Fix the problem by grabbing i_mutex in udf_file_release() to be sure
i_size is consistent with current state of extent list. Also avoid
truncating tail extent unnecessarily when the file is still open for
writing.

Signed-off-by: Jan Kara <jack@suse.cz>

diff --git a/fs/udf/file.c b/fs/udf/file.c
index 86c6743ec1feb7eb99b17d6557c2ed1a1a90c1b8..bb15771b92ae32ae02390179492647a1d87c9dc2 100644 (file)
--- a/fs/udf/file.c
+++ b/fs/udf/file.c
@@ -223,11 +223,18 @@ out:
 
 static int udf_release_file(struct inode *inode, struct file *filp)
 {
-       if (filp->f_mode & FMODE_WRITE) {
+       if (filp->f_mode & FMODE_WRITE &&
+           atomic_read(&inode->i_writecount) > 1) {
+               /*
+                * Grab i_mutex to avoid races with writes changing i_size
+                * while we are running.
+                */
+               mutex_lock(&inode->i_mutex);
                down_write(&UDF_I(inode)->i_data_sem);
                udf_discard_prealloc(inode);
                udf_truncate_tail_extent(inode);
                up_write(&UDF_I(inode)->i_data_sem);
+               mutex_unlock(&inode->i_mutex);
        }
        return 0;
 }