Re: [syzbot] [net?] INFO: rcu detected stall in unix_release

[Jamal Hadi Salim]

Hi Vladimir,

On Fri, Aug 18, 2023 at 12:07 PM Vladimir Oltean
<vladimir.oltean@xxxxxxx> wrote:
>
> Hi Jamal,
>
> On Fri, Aug 18, 2023 at 11:27:27AM -0400, Jamal Hadi Salim wrote:
> > Can you try the attached patchlet?
>
> Thanks for the patch. I've tried it, and it eliminates the code path
> (and thus the problem) exposed by the syzbot program, by responding to
> RTM_NEWQDISC messages having the NLM_F_CREATE|NLM_F_REPLACE|NLM_F_EXCL
> flags with "Error: Exclusivity flag on, cannot modify.".
>

Ok, that is more of the expected behavior.
Noone should ever send that mumbo-jumbo (I doubt there is a "legit"
control app that will do that).

> Actually, to be precise, the first such netlink message successfully
> creates the qdisc, but then the subsequent ones leave that qdisc alone
> (don't change it), by failing with this extack message.
>

Yes, the first one will succeed because the root qdisc hasnt been
grafted yet (and the only interesting bit is NLM_F_CREATE. everything
else is ignored).

> If that's the behavior that you intended, then I guess the answer is
> that it works. Thanks a lot.
>
> What would be an appropriate Fixes: tag?
>

This should have been from early days when we trusted that iproute2
would do the right thing. I will look.
I dont think this is a taprio only potential victim, it's just that
syzbot was able to aggravate taprio sooner (it probably would have got
to some other qdisc later in its adventures).

> Side note: I believe that we can now also revert commit be3618d96510
> ("net/sched: taprio: fix slab-out-of-bounds Read in taprio_dequeue_from_txq"),
> which was papering over an unknown (at the time) issue - the same as
> this one - without really even completely covering it, either.

Unfortunately the commit log is not helpful - i cant tell what
"replace" means and cant seem to find the repro either. If you revert
it and see the problem going away then we are good.
+Cc  Zhengchao Shao <shaozhengchao@xxxxxxxxxx>

>Hence
> this other syzbot report.
> https://lore.kernel.org/netdev/3b977f76-0289-270e-8310-179315ee927d@xxxxxxxxxx/T/
> https://lore.kernel.org/netdev/20230608062756.3626573-1-shaozhengchao@xxxxxxxxxx/

Makes sense.
BTW, thanks for your report - it made it faster to zone on the issue.
The comments above that code also need a bit of fixing to provide clarity.

cheers,
jamal