Re: [PATCH] locking/lockdep: Add debug_locks check in __lock_downgrade()

[Peter Zijlstra]

On Thu, Jan 10, 2019 at 11:21:13AM +0100, Dmitry Vyukov wrote:
> On Thu, Jan 10, 2019 at 5:04 AM Waiman Long <longman@xxxxxxxxxx> wrote:
> >
> > Tetsuo Handa had reported he saw an incorrect "downgrading a read lock"
> > warning right after a previous lockdep warning. It is likely that the
> > previous warning turned off lock debugging causing the lockdep to have
> > inconsistency states leading to the lock downgrade warning.
> >
> > Fix that by add a check for debug_locks at the beginning of
> > __lock_downgrade().
> >
> > Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
> > Reported-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
> 
> Please also add:
> 
> Reported-by: syzbot+53383ae265fb161ef488@xxxxxxxxxxxxxxxxxxxxxxxxx
> 
> for tracking purposes. But Tetsuo deserves lots of credit for debugging it.

I made that:

Reported-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
Debugged-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
Reported-by: syzbot+53383ae265fb161ef488@xxxxxxxxxxxxxxxxxxxxxxxxx

> > index 9593233..e805fe3 100644
> > --- a/kernel/locking/lockdep.c
> > +++ b/kernel/locking/lockdep.c
> > @@ -3535,6 +3535,9 @@ static int __lock_downgrade(struct lockdep_map *lock, unsigned long ip)
> >         unsigned int depth;
> >         int i;
> >
> > +       if (unlikely(!debug_locks))
> > +               return 0;
> > +
> 
> Are we sure this resolves the problem rather than makes the
> inconsistency window smaller?
> I don't understand all surrounding code, but looking just at this
> function it looks like it may just pepper over the problem. Say, we
> pass this check when lockdep was still turned on. Then this thread is
> preempted for some time (e.g. a virtual CPU), then another thread
> started reporting a warning, turned lockdep off, some information
> wasn't collected, and this this task resumes and reports a false
> warning.

Theoretically possible I suppose; but this is analogous to many of the
other lockdep hooks.

> Or we are holding the mutex here, and the fact that we are holding it
> ensures that no other task will take it and no information will be
> lost?

There is no lock here; for performance reasons we prefer not to acquire
a global spinlock on every lockdep hook, that would be horrific.