[PATCH 1/1] loop: access lo_backing_file only when the loop device is Lo_bound

[PATCH 1/1] loop: access lo_backing_file only when the loop device is Lo_bound

[Dongli Zhang]

Commit 758a58d0bc67 ("loop: set GENHD_FL_NO_PART_SCAN after
blkdev_reread_part()") separates "lo->lo_backing_file = NULL" and
"lo->lo_state = Lo_unbound" into different critical regions protected by
loop_ctl_mutex.

However, there is below race that the NULL lo->lo_backing_file would be
accessed when the backend of a loop is another loop device, e.g., loop0's
backend is a file, while loop1's backend is loop0.

loop0's backend is file            loop1's backend is loop0

__loop_clr_fd()
  mutex_lock(&loop_ctl_mutex);
  lo->lo_backing_file = NULL; --> set to NULL
  mutex_unlock(&loop_ctl_mutex);
                                   loop_set_fd()
                                     mutex_lock_killable(&loop_ctl_mutex);
                                     loop_validate_file()
                                       f = l->lo_backing_file; --> NULL
                                         access if loop0 is not Lo_unbound
  mutex_lock(&loop_ctl_mutex);
  lo->lo_state = Lo_unbound;
  mutex_unlock(&loop_ctl_mutex);

lo->lo_backing_file should be accessed only when the loop device is
Lo_bound.

In fact, the problem has been introduced already in commit 7ccd0791d985
("loop: Push loop_ctl_mutex down into loop_clr_fd()") after which
loop_validate_file() could see devices in Lo_rundown state with which it
did not count. It was harmless at that point but still.

Fixes: 7ccd0791d985 ("loop: Push loop_ctl_mutex down into loop_clr_fd()")
Reported-by: syzbot+9bdc1adc1c55e7fe765b@syzkaller.appspotmail.com
Signed-off-by: Dongli Zhang <dongli.zhang@oracle.com>
Reviewed-by: Jan Kara <jack@suse.cz>
---
 drivers/block/loop.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/block/loop.c b/drivers/block/loop.c
index 1e6edd5..bf1c61c 100644
--- a/drivers/block/loop.c
+++ b/drivers/block/loop.c
@@ -656,7 +656,7 @@ static int loop_validate_file(struct file *file, struct block_device *bdev)
 			return -EBADF;
 
 		l = f->f_mapping->host->i_bdev->bd_disk->private_data;
-		if (l->lo_state == Lo_unbound) {
+		if (l->lo_state != Lo_bound) {
 			return -EINVAL;
 		}
 		f = l->lo_backing_file;
-- 
2.7.4

Re: [PATCH 1/1] loop: access lo_backing_file only when the loop device is Lo_bound

[Jens Axboe]

On 3/18/19 6:23 AM, Dongli Zhang wrote:
> Commit 758a58d0bc67 ("loop: set GENHD_FL_NO_PART_SCAN after
> blkdev_reread_part()") separates "lo->lo_backing_file = NULL" and
> "lo->lo_state = Lo_unbound" into different critical regions protected by
> loop_ctl_mutex.
> 
> However, there is below race that the NULL lo->lo_backing_file would be
> accessed when the backend of a loop is another loop device, e.g., loop0's
> backend is a file, while loop1's backend is loop0.
> 
> loop0's backend is file            loop1's backend is loop0
> 
> __loop_clr_fd()
>   mutex_lock(&loop_ctl_mutex);
>   lo->lo_backing_file = NULL; --> set to NULL
>   mutex_unlock(&loop_ctl_mutex);
>                                    loop_set_fd()
>                                      mutex_lock_killable(&loop_ctl_mutex);
>                                      loop_validate_file()
>                                        f = l->lo_backing_file; --> NULL
>                                          access if loop0 is not Lo_unbound
>   mutex_lock(&loop_ctl_mutex);
>   lo->lo_state = Lo_unbound;
>   mutex_unlock(&loop_ctl_mutex);
> 
> lo->lo_backing_file should be accessed only when the loop device is
> Lo_bound.
> 
> In fact, the problem has been introduced already in commit 7ccd0791d985
> ("loop: Push loop_ctl_mutex down into loop_clr_fd()") after which
> loop_validate_file() could see devices in Lo_rundown state with which it
> did not count. It was harmless at that point but still.

Thanks, applied.

-- 
Jens Axboe