On 1/31/2017 9:21 PM, Jan Kara wrote:

On Tue 31-01-17 20:08:57, Sahitya Tummala wrote:

quoted

Below is the synchronization issue between unmount and kjournald2
contexts, which results into use after free issue in kjournald2().
Fix this issue by using journal->j_state_lock to synchronize the
wait_event() done in journal_kill_thread() and the wake_up() done
in kjournald2().

TASK 1:
umount cmd:
    |--jbd2_journal_destroy() {
        |--journal_kill_thread() {
             write_lock(&journal->j_state_lock);
	    journal->j_flags |= JBD2_UNMOUNT;
	    ...
	    write_unlock(&journal->j_state_lock);
	    wake_up(&journal->j_wait_commit);	   TASK 2 wakes up here:
	    					   kjournald2() {
						     ...
						     checks JBD2_UNMOUNT flag and calls goto end-loop;
						     ...
						     end_loop:
						       write_unlock(&journal->j_state_lock);
						       journal->j_task = NULL; --> If this thread gets
						       pre-empted here, then TASK 1 wait_event will
						       exit even before this thread is completely
						       done.
	    wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
	    ...
	    write_lock(&journal->j_state_lock);
	    write_unlock(&journal->j_state_lock);
	  }
        |--kfree(journal);
      }
}
						       wake_up(&journal->j_wait_done_commit); --> this step
						       now results into use after free issue.
						   }

Signed-off-by: Sahitya Tummala <redacted>

Yeah, what you write looks possible (although rather unlikely). Thanks for
catching this. One small nit below:

Yes, it was observed only once and is very hard to reproduce.

quoted

diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c
index a097048..f5cd3c0 100644
--- a/fs/jbd2/journal.c
+++ b/fs/jbd2/journal.c

@@ -278,9 +278,11 @@ static int kjournald2(void *arg)
  end_loop:
  	write_unlock(&journal->j_state_lock);
  	del_timer_sync(&journal->j_commit_timer);
+	write_lock(&journal->j_state_lock);

There's no good reason to do del_timer_sync() outside of j_state_lock. This
is not performance critical code and commit_timeout is trivial and cannot
block on anything. So just keep j_state_lock locked upto the place where
you unlock it now...

Sure, I will update the patch.

								Honza

quoted

  	journal->j_task = NULL;
  	wake_up(&journal->j_wait_done_commit);
  	jbd_debug(1, "Journal thread exiting.\n");
+	write_unlock(&journal->j_state_lock);
  	return 0;
  }
  
-- 
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.

-- 
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.