Re: BUG: MAX_LOCKDEP_CHAINS too low!

[Taehee Yoo]

On Thu, 16 Jan 2020 at 06:53, Cong Wang <xiyou.wangcong@xxxxxxxxx> wrote:
>

Hi Cong,

> On Mon, Jan 13, 2020 at 3:11 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> > +Taehee, Cong,
> >
> > In the other thread Taehee mentioned the creation of dynamic keys for
> > net devices that was added recently and that they are subject to some
> > limits.
> > syzkaller creates lots of net devices for isolation (several dozens
> > per test process, but then these can be created and destroyed
> > periodically). I wonder if it's the root cause of the lockdep limits
> > problems?
>
> Very possibly. In current code base, there are 4 lockdep keys
> per netdev:
>
>         struct lock_class_key   qdisc_tx_busylock_key;
>         struct lock_class_key   qdisc_running_key;
>         struct lock_class_key   qdisc_xmit_lock_key;
>         struct lock_class_key   addr_list_lock_key;
>
> so the number of lockdep keys is at least 4x number of network
> devices.
>
> I think only addr_list_lock_key is necessary as it has a nested
> locking use case, all the rest are not. Taehee, do you agree?
>
> I plan to remove at least qdisc_xmit_lock_key for net-next
> after the fix for net gets merged.
>

Yes, I fully agree with this.
If we calculate the subclass for lock_nested() very well, I think we
might use static lockdep key for addr_list_lock_key too. I think
"dev->upper_level" and "dev->lower_level" might be used as subclass.
These values are updated recursively in master/nomaster operation.

Thank you
Taehee Yoo

> Thanks!