On Wed, Apr 27, 2022 at 01:17:26PM +0200, Jan Kara wrote:
> CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.
>
>
>
> On Tue 26-04-22 11:31:24, Samuel Mendoza-Jonas wrote:
> > On Thu, Feb 10, 2022 at 09:07:11PM +0530, Ritesh Harjani wrote:
> > > jbd2_journal_wait_updates() is called with j_state_lock held. But if
> > > there is a commit in progress, then this transaction might get committed
> > > and freed via jbd2_journal_commit_transaction() ->
> > > jbd2_journal_free_transaction(), when we release j_state_lock.
> > > So check for journal->j_running_transaction everytime we release and
> > > acquire j_state_lock to avoid use-after-free issue.
> > >
> > > Fixes: 4f98186848707f53 ("jbd2: refactor wait logic for transaction updates into a common function")
> > > Reported-and-tested-by: syzbot+afa2ca5171d93e44b348@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Signed-off-by: Ritesh Harjani <riteshh@xxxxxxxxxxxxx>
> >
> > Hi Ritesh,
> >
> > Looking at the refactor in the commit this fixes, I believe the same
> > issue is present prior to the refactor, so this would apply before 5.17
> > as well.
> > I've posted a backport for 4.9-4.19 and 5.4-5.16 to stable here:
> > https://lore.kernel.org/stable/20220426182702.716304-1-samjonas@xxxxxxxxxx/T/#t
> >
> > Please have a look and let me know if you agree.
>
> Actually the refactor was indeed the cause for use-after-free. The original
> code in jbd2_journal_lock_updates() was like:
>
> /* Wait until there are no running updates */
> while (1) {
> transaction_t *transaction = journal->j_running_transaction;
>
> if (!transaction)
> break;
> spin_lock(&transaction->t_handle_lock);
> prepare_to_wait(&journal->j_wait_updates, &wait,
> TASK_UNINTERRUPTIBLE);
> if (!atomic_read(&transaction->t_updates)) {
> spin_unlock(&transaction->t_handle_lock);
> finish_wait(&journal->j_wait_updates, &wait);
> break;
> }
> spin_unlock(&transaction->t_handle_lock);
> write_unlock(&journal->j_state_lock);
> schedule();
> finish_wait(&journal->j_wait_updates, &wait);
> write_lock(&journal->j_state_lock);
> }
>
> So you can see the code was indeed careful enough to not touch
> t_handle_lock after sleeping. The code in jbd2_journal_commit_transaction()
> did touch t_handle_lock but there it didn't matter because nobody else
> besides the task running jbd2_journal_commit_transaction() can free the
> transaction...
>
Right you are, I misinterpreted the interaction with
jbd2_journal_commit_transaction(). Thanks for verifying, I'll follow up
stable to disregard those backports.
Cheers,
Sam
> Honza
> --
> Jan Kara <jack@xxxxxxxx>
> SUSE Labs, CR
