Message-ID: <YK9nMgamPsr9YsoY@elver.google.com>
Date:   Thu, 27 May 2021 11:32:34 +0200
From:   Marco Elver <elver@...gle.com>
To:     Pavel Begunkov <asml.silence@...il.com>
Cc:     Jens Axboe <axboe@...nel.dk>,
        syzbot <syzbot+73554e2258b7b8bf0bbf@...kaller.appspotmail.com>,
        io-uring@...r.kernel.org, LKML <linux-kernel@...r.kernel.org>,
        syzkaller-bugs <syzkaller-bugs@...glegroups.com>,
        Dmitry Vyukov <dvyukov@...gle.com>
Subject: Re: [syzbot] KCSAN: data-race in __io_uring_cancel /
 io_uring_try_cancel_requests

On Wed, May 26, 2021 at 09:31PM +0100, Pavel Begunkov wrote:
> On 5/26/21 5:36 PM, Marco Elver wrote:
> > On Wed, 26 May 2021 at 18:29, Pavel Begunkov <asml.silence@...il.com> wrote:
> >> On 5/26/21 4:52 PM, Marco Elver wrote:
> >>> Due to some moving around of code, the patch lost the actual fix (using
> >>> atomically read io_wq) -- so here it is again ... hopefully as intended.
> >>> :-)
> >>
> >> "fortify" damn it... It was synchronised with &ctx->uring_lock
> >> before, see io_uring_try_cancel_iowq() and io_uring_del_tctx_node(),
> >> so should not clear before *del_tctx_node()
> > 
> > Ah, so if I understand right, the property stated by the comment in
> > io_uring_try_cancel_iowq() was broken, and your patch below would fix
> > that, right?
> 
> "io_uring: fortify tctx/io_wq cleanup" broke it and the diff
> should fix it.
> 
> >> The fix should just move it after this sync point. Will you send
> >> it out as a patch?
> > 
> > Do you mean your move of write to io_wq goes on top of the patch I
> > proposed? (If so, please also leave your Signed-of-by so I can squash
> > it.)
> 
> No, only my diff, but you hinted on what has happened, so I would
> prefer you to take care of patching. If you want of course.
> 
> To be entirely fair, assuming that aligned ptr
> reads can't be torn, I don't see any _real_ problem. But surely
> the report is very helpful and the current state is too wonky, so
> should be patched.

In the current version, it is a problem if we end up with a double-read,
as it is in the current C code. The compiler might of course optimize
it into 1 read into a register.

Tangent: I avoid reasoning in terms of compiler optimizations where
I can. :-) It's is a slippery slope if the code in question isn't
tolerant to data races by design (examples are stats counting, or other
heuristics -- in the case here that's certainly not the case).
Therefore, my wish is that we really ought to resolve as many data races
as we can (+ mark intentional ones appropriately). Also, so that we're
left with only the interesting cases like in the case here.  (More
background if you're interested: https://lwn.net/Articles/816850/)

The problem here, however, has a nicer resolution as you suggested.

> TL;DR;
> The synchronisation goes as this: it's usually used by the owner
> task, and the owner task deletes it, so is mostly naturally
> synchronised. An exception is a worker (not only) that accesses
> it for cancellation purpose, but it uses it only under ->uring_lock,
> so if removal is also taking the lock it should be fine. see
> io_uring_del_tctx_node() locking.

Did you mean io_uring_del_task_file()? There is no
io_uring_del_tctx_node().

> > So if I understand right, we do in fact have 2 problems:
> > 1. the data race as I noted in my patch, and
> 
> Yes, and it deals with it
> 
> > 2. the fact that io_wq does not live long enough.
> 
> Nope, io_wq outlives them fine. 

I've sent:
https://lkml.kernel.org/r/20210527092547.2656514-1-elver@google.com

Thanks,
-- Marco

Powered by blists - more mailing lists