xfs
[Top] [All Lists]

Re: use-after-free on log replay failure

To: Brian Foster <bfoster@xxxxxxxxxx>
Subject: Re: use-after-free on log replay failure
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Thu, 14 Aug 2014 16:14:44 +1000
Cc: Alex Lyakas <alex@xxxxxxxxxxxxxxxxx>, xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20140813232135.GB8456@xxxxxxxxxxxxxx>
References: <20140806152042.GB39990@xxxxxxxxxxxxxxx> <CAOcd+r3bC59m7Rh-3tmjrnWnF5XoPQfE=U+=hz78NcAGu+Ou1g@xxxxxxxxxxxxxx> <20140811132057.GA1186@xxxxxxxxxxxxxxx> <20140811215207.GS20518@dastard> <20140812120341.GA46654@xxxxxxxxxxxxxxx> <DC407F6E8F8C4EE1AF7117D7D6ABF282@alyakaslap> <20140812235615.GB20518@dastard> <20140813125932.GA4605@xxxxxxxxxxxxxxx> <20140813205929.GE20518@dastard> <20140813232135.GB8456@xxxxxxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Wed, Aug 13, 2014 at 07:21:35PM -0400, Brian Foster wrote:
> On Thu, Aug 14, 2014 at 06:59:29AM +1000, Dave Chinner wrote:
> > On Wed, Aug 13, 2014 at 08:59:32AM -0400, Brian Foster wrote:
> > > Perhaps I'm missing some context... I don't follow how removing the
> > > error check doesn't solve the problem. It clearly closes the race and
> > > perhaps there are other means of doing the same thing, but what part of
> > > the problem does that leave unresolved?
> > 
> > Anything that does:
> > 
> >     xfs_buf_iorequest(bp);
> >     if (bp->b_error)
> >             xfs_buf_relse(bp);
> > 
> > is susceptible to the same race condition. based on bp->b_error
> > being set asynchronously and before the buffer IO completion
> > processing is complete.
> > 
> 
> Understood, by why would anything do that (as opposed to
> xfs_buf_iowait())? I don't see that we do that anywhere today
> (the check buried within xfs_buf_iowait() notwithstanding of course).

"Why" is not important - the fact is the caller *owns* the buffer
and so the above fragment of code is valid behaviour. If there is
an error on the buffer after xfs_buf_iorequest() request returns on
a synchronous IO, then it's a bug if there is still IO in progress
on that buffer.

We can't run IO completion synchronously from xfs_buf_bio_end_io in
this async dispatch error case - we cannot detect it as any
different from IO completion in interrupt context - and so we need
to have some kind of reference protecting the buffer from being
freed from under the completion.

i.e. the bug is that a synchronous buffer has no active reference
while it is sitting on the completion workqueue - it's references
are owned by other contexts that can drop them without regard to
the completion status of the buffer.

For async IO we transfer a reference and the lock to the IO context,
which gets dropped in xfs_buf_iodone_work when all the IO is
complete. Synchronous IO needs this protection, too.

As a proof of concept, adding this to the start of
xfs_buf_iorequest():

+       /*
+        * synchronous IO needs it's own reference count. async IO
+        * inherits the submitter's reference count.
+        */
+       if (!(bp->b_flags & XBF_ASYNC))
+               xfs_buf_hold(bp);

And this to the synchronous IO completion case for
xfs_buf_iodone_work():

        else {
                ASSERT(read && bp->b_ops);
                complete(&bp->b_iowait);
+               xfs_buf_rele(bp);
        }

Should ensure that all IO carries a reference count and the buffer
cannot be freed until all IO processing has been completed.

This means it does not matter what the buffer owner does after
xfs_buf_iorequest() - even unconditionally calling xfs_buf_relse()
will not result in use-after-free as the b_hold count will not go to
zero until the IO completion processing has been finalised.

Fixing the rest of the mess (i.e. determining how to deal with
submission/completion races) is going to require more effort and
thought. For the moment, though, correctly reference counting
buffers will solve the use-after-free without changing any
other behaviour.

> From what I can see, all it really guarantees is that the submission has
> either passed/failed the write verifier, yes?

No.  It can also mean it wasn't rejected by the lower layersi as
they process the bio passed by submit_bio(). e.g.  ENODEV, because
the underlying device has been hot-unplugged, EIO because the
buffer is beyond the end of the device, etc.

> > > It looks like submit_bio() manages this by providing the error through
> > > the callback (always). It also doesn't look like submission path is
> > > guaranteed to be synchronous either (consider md, which appears to use
> > > workqueues and kernel threads)), so I'm not sure that '...;
> > > xfs_buf_iorequest(bp); if (bp->b_error)' is really safe anywhere unless
> > > you're explicitly looking for a write verifier error or something and
> > > do nothing further on the buf contingent on completion (e.g., freeing it
> > > or something it depends on).
> > 
> > My point remains that it *should be safe*, and the intent is that
> > the caller should be able to check for submission errors without
> > being exposed to a use after free situation. That's the bug we need
> > to fix, not say "you can't check for submission errors on
> > synchronous IO" to avoid the race condition.....
> > 
> 
> Well, technically you can check for submission errors on sync I/O, just
> use the code you posted above. :) What we can't currently do is find out
> when the I/O subsystem is done with the buffer.

By definition, a buffer marked with an error after submission
processing is complete. It should not need to be waited on, and
there-in lies the bug.

> Perhaps the point here is around the semantics of xfs_buf_iowait(). With
> a mechanism that is fundamentally async, the sync variant obviously
> becomes the async mechanism + some kind of synchronization. I'd expect
> that synchronization to not necessarily just tell me whether an error
> occurred, but also tell me when the I/O subsystem is done with the
> object I've passed (e.g., so I'm free to chuck it, scribble over it, put
> it back where I got it, whatever).
>
> My impression is that's the purpose of the b_iowait mechanism.
> Otherwise, what's the point of the whole
> bio_end_io->buf_ioend->b_iodone->buf_ioend round trip dance?

Yes, that's exactly what xfs_buf_iorequest/xfs_buf_iowait() provides
and the b_error indication is an integral part of that
synchronisation mechanism.  Unfortunately, that is also the part of
the mechanism that is racy and causing problems.

Cheers,

Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx

<Prev in Thread] Current Thread [Next in Thread>