Re: Finally starting on short RCU grace periods, but...

[Dmitry Vyukov]

On Thu, Aug 6, 2020 at 12:31 PM Marco Elver <elver@xxxxxxxxxx> wrote:
>
> +Cc kasan-dev
>
> On Thu, 6 Aug 2020 at 01:08, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >
> > Hello!
> >
> > If I remember correctly, one of you asked for a way to shorten RCU
> > grace periods so that KASAN would have a better chance of detecting bugs
> > such as pointers being leaked out of RCU read-side critical sections.
> > I am finally starting entering and testing code for this, but realized
> > that I had forgotten a couple of things:
> >
> > 1.      I don't remember exactly who asked, but I suspect that it was
> >         Kostya.  I am using his Reported-by as a placeholder for the
> >         moment, but please let me know if this should be adjusted.
>
> It certainly was not me.
>
> > 2.      Although this work is necessary to detect situtions where
> >         call_rcu() is used to initiate a grace period, there already
> >         exists a way to make short grace periods that are initiated by
> >         synchronize_rcu(), namely, the rcupdate.rcu_expedited kernel
> >         boot parameter.  This will cause all calls to synchronize_rcu()
> >         to act like synchronize_rcu_expedited(), resulting in about 2-3
> >         orders of magnitude reduction in grace-period latency on small
> >         systems (say 16 CPUs).
> >
> > In addition, I plan to make a few other adjustments that will
> > increase the probability of KASAN spotting a pointer leak even in the
> > rcupdate.rcu_expedited case.
>
> Thank you, that'll be useful I think.
>
> > But if you would like to start this sort of testing on current mainline,
> > rcupdate.rcu_expedited is your friend!

Hi Paul,

This is great!

I understand it's not a sufficiently challenging way of tracking
things, but it's simply here ;)
https://bugzilla.kernel.org/show_bug.cgi?id=208299
(now we also know who asked for this, +Jann)

I've tested on the latest mainline and with rcupdate.rcu_expedited=1
it boots to ssh successfully and I see:
[    0.369258][    T0] All grace periods are expedited (rcu_expedited).

I have created https://github.com/google/syzkaller/pull/2021 to enable
it on syzbot.
On syzbot we generally use only 2-4 CPUs per VM, so it should be even better.

> Do any of you remember some bugs we missed due to this? Can we find
> them if we add this option?

The problem is that it's hard to remember bugs that were not caught :)
Here is an approximation of UAFs with free in rcu callback:
https://groups.google.com/forum/#!searchin/syzkaller-bugs/KASAN$20use-after-free$20rcu_do_batch%7Csort:date
The ones with low hit count are the ones that we almost did not catch.
That's the best estimation I can think of. Also potentially we can get
reproducers for such bugs without reproducers.
Maybe we will be able to correlate some bugs/reproducers that appear
soon with this change.