[PATCH v3] kernel/watch_queue: Make pipe NULL while clearing watch_queue

[Eric Biggers]

On Wed, Aug 03, 2022 at 02:10:06PM +0530, Siddh Raman Pant wrote:
> On Wed, 03 Aug 2022 11:11:31 +0530  Eric Biggers <ebiggers at kernel.org> wrote:
> > I tested the syzbot reproducer
> > https://syzkaller.appspot.com/text?tag=ReproC&x=174ea97e080000, and it does
> > *not* trigger the bug on the latest upstream.  But, it does trigger the bug if I
> > recent Linus's recent watch_queue fixes.
> > 
> > So I don't currently see any evidence of an unfixed bug.  If, nevertheless, you
> > believe that a bug is still unfixed, then please provide a reproducer and a fix
> > patch that clearly explains what it is fixing. 
> > 
> > > There is a null check in post_one_notification for the pipe, most probably
> > > because it *expects* the pointer to be NULL'd. Also, there is no reason to have
> > > a dangling pointer stay, it's just a recipe for further bugs.
> > 
> > If you want to send a patch or patches to clean up the code, that is fine, but
> > please make it super clear what is a cleanup and what is a fix.
> > 
> > - Eric
> > 
> 
> I honestly feel like I am repeating myself yet again, but okay.

Well, you should try listening instead.  Because you are not listening.

> Of course, the race condition has been solved by a patch upstream, which I had
> myself mentioned earlier.
> 
> But what I am saying is that it did *not* address *what* that race condition
> had triggered, i.e. the visible cause of the UAF crash, which, among other
> things, is *because* there is a dangling pointer to the freed pipe, which
> *caused* the crash in post_one_notification() when it tried to access
> &pipe->rd.wait_lock as an argument to spin_lock_irq(), a path it reached
> after checking if wqueue->pipe is NULL and proceeded when it was not the case.
>
> And the upstream commit was made *after* I had posted this patch, hence this
> was a fix for the syzkaller issue. While I am *not* saying to accept it just
> because this was posted earlier, I am saying this patch addresses a parallel
> issue, i.e. the *actual use-after-free crash* which was reproduced by those
> reproducers, i.e., what was attempted to be used after getting freed and
> detected by KASAN.

Even if wqueue->pipe was set to NULL during free_pipe_info(), there would still
have been a use-after-free, as the real bug was the lack of synchronization
between post_one_notification() and free_pipe_info().  That is fixed now.

> 
> We don't need to wait for another similar syzbot report to pop up before doing
> this change, and say let's not fix a dangling pointer reference because now
> another commit apparately fixes the specific syzkaller issue, causing the given
> specific reproducer with its specific way of reproducing to fail, when we in
> fact now know it *can* be a valid problem in practice and doing this change
> too causes the specific reproducer under consideration to fail reproducing, as
> was reported by the reproducer itself.

To re-iterate, I encourage you to send a cleanup patch if you see an
opportunity.  It looks like the state wqueue->defunct==true could be replaced
with wqueue->pipe==NULL, which would be simpler, so how about doing that?  Just
don't claim that it is "fixing" something, unless it is, as that makes things
very confusing and difficult for everyone.

> 
> I really don't know how to create stress tests / reproducers like how syzkaller
> makes, so if a similar new reproducer is really required for showing this
> patch's validity disregarding any earlier reproducers, I unfortunately cannot
> make it due to skill issue as I just started in kernel dev, and I am deeply
> sorry for wasting the time of everyone, and I am thankful for your criticism of
> my patch.

A reproducer can just be written as a normal program, in C or another language.
The syzkaller reproducers are really hard to read as they are auto-generated, so
don't read too much into them -- they're certainly not examples of good code.

- Eric