Re: [PATCH] lockdep: Speed up lockdep_unregister_key() with expedited RCU synchronization

[Eric Dumazet]

On Mon, Mar 24, 2025 at 1:23 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
>
> On Mon, Mar 24, 2025 at 1:12 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Mar 21, 2025 at 02:30:49AM -0700, Breno Leitao wrote:
> > > lockdep_unregister_key() is called from critical code paths, including
> > > sections where rtnl_lock() is held. For example, when replacing a qdisc
> > > in a network device, network egress traffic is disabled while
> > > __qdisc_destroy() is called for every network queue.
> > >
> > > If lockdep is enabled, __qdisc_destroy() calls lockdep_unregister_key(),
> > > which gets blocked waiting for synchronize_rcu() to complete.
> > >
> > > For example, a simple tc command to replace a qdisc could take 13
> > > seconds:
> > >
> > >   # time /usr/sbin/tc qdisc replace dev eth0 root handle 0x1: mq
> > >     real    0m13.195s
> > >     user    0m0.001s
> > >     sys     0m2.746s
> > >
> > > During this time, network egress is completely frozen while waiting for
> > > RCU synchronization.
> > >
> > > Use synchronize_rcu_expedited() instead to minimize the impact on
> > > critical operations like network connectivity changes.
> > >
> > > This improves 10x the function call to tc, when replacing the qdisc for
> > > a network card.
> > >
> > >    # time /usr/sbin/tc qdisc replace dev eth0 root handle 0x1: mq
> > >      real     0m1.789s
> > >      user     0m0.000s
> > >      sys      0m1.613s
> > >
> > > Reported-by: Erik Lundgren <elundgren@xxxxxxxx>
> > > Signed-off-by: Breno Leitao <leitao@xxxxxxxxxx>
> > > Reviewed-by: "Paul E. McKenney" <paulmck@xxxxxxxxxx>
> > > ---
> > >  kernel/locking/lockdep.c | 6 ++++--
> > >  1 file changed, 4 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
> > > index 4470680f02269..a79030ac36dd4 100644
> > > --- a/kernel/locking/lockdep.c
> > > +++ b/kernel/locking/lockdep.c
> > > @@ -6595,8 +6595,10 @@ void lockdep_unregister_key(struct lock_class_key *key)
> > >       if (need_callback)
> > >               call_rcu(&delayed_free.rcu_head, free_zapped_rcu);
> > >
> > > -     /* Wait until is_dynamic_key() has finished accessing k->hash_entry. */
> > > -     synchronize_rcu();
> > > +     /* Wait until is_dynamic_key() has finished accessing k->hash_entry.
> > > +      * This needs to be quick, since it is called in critical sections
> > > +      */
> > > +     synchronize_rcu_expedited();
> > >  }
> > >  EXPORT_SYMBOL_GPL(lockdep_unregister_key);
> >
> > So I fundamentally despise synchronize_rcu_expedited(), also your
> > comment style is broken.
> >
> > Why can't qdisc call this outside of the lock?
>
> Good luck with that, and anyway the time to call it 256 times would
> still hurt Breno use case.
>
> My suggestion was to change lockdep_unregister_key() contract, and use
> kfree_rcu() there
>
> > I think we should redesign lockdep_unregister_key() to work on a separately
> > allocated piece of memory,
> > then use kfree_rcu() in it.
> >
> > Ie not embed a "struct lock_class_key" in the struct Qdisc, but a pointer to
> >
> > struct ... {
> >      struct lock_class_key key;
> >      struct rcu_head  rcu;
> > }
>
> More work because it requires changing all lockdep_unregister_key() users.

Or add a new function, basically allow users to be converted when they need to.