Re: corrupted pvqspinlock in htab_map_update_elem

[Waiman Long]

On 2/1/21 1:09 PM, Dmitry Vyukov wrote:
> On Mon, Feb 1, 2021 at 6:54 PM Waiman Long <longman@xxxxxxxxxx> wrote:
> > On 2/1/21 6:23 AM, Peter Zijlstra wrote:
> > > On Mon, Feb 01, 2021 at 10:50:58AM +0100, Peter Zijlstra wrote:
> > >
> > > > >    queued_spin_unlock arch/x86/include/asm/qspinlock.h:56 [inline]
> > > > >    lockdep_unlock+0x10e/0x290 kernel/locking/lockdep.c:124
> > > > >    debug_locks_off_graph_unlock kernel/locking/lockdep.c:165 [inline]
> > > > >    print_usage_bug kernel/locking/lockdep.c:3710 [inline]
> > > > Ha, I think you hit a bug in lockdep.
> > > Something like so I suppose.
> > >
> > > ---
> > > Subject: locking/lockdep: Avoid unmatched unlock
> > > From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > > Date: Mon Feb 1 11:55:38 CET 2021
> > >
> > > Commit f6f48e180404 ("lockdep: Teach lockdep about "USED" <- "IN-NMI"
> > > inversions") overlooked that print_usage_bug() releases the graph_lock
> > > and called it without the graph lock held.
> > >
> > > Fixes: f6f48e180404 ("lockdep: Teach lockdep about "USED" <- "IN-NMI" inversions")
> > > Reported-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> > > Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > > ---
> > >    kernel/locking/lockdep.c |    3 ++-
> > >    1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > --- a/kernel/locking/lockdep.c
> > > +++ b/kernel/locking/lockdep.c
> > > @@ -3773,7 +3773,7 @@ static void
> > >    print_usage_bug(struct task_struct *curr, struct held_lock *this,
> > >                enum lock_usage_bit prev_bit, enum lock_usage_bit new_bit)
> > >    {
> > > -     if (!debug_locks_off_graph_unlock() || debug_locks_silent)
> > > +     if (!debug_locks_off() || debug_locks_silent)
> > >                return;
> > >
> > >        pr_warn("\n");
> > > @@ -3814,6 +3814,7 @@ valid_state(struct task_struct *curr, st
> > >            enum lock_usage_bit new_bit, enum lock_usage_bit bad_bit)
> > >    {
> > >        if (unlikely(hlock_class(this)->usage_mask & (1 << bad_bit))) {
> > > +             graph_unlock()
> > >                print_usage_bug(curr, this, bad_bit, new_bit);
> > >                return 0;
> > >        }
> > I have also suspected doing unlock without a corresponding lock. This
> > patch looks good to me.
> >
> > Acked-by: Waiman Long <longman@xxxxxxxxxx>
> Just so that it's not lost: there is still a bug related to bpf map lock, right?
That is right. This patch just fixes the bug in lockdep.

Cheers,
Longman