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

[Boqun Feng <boqun.feng@gmail.com>]

On Wed, Mar 26, 2025 at 11:39:49AM -0400, Waiman Long wrote:
[...]
> > > Anyway, that may work. The only problem that I see is the issue of nesting
> > > of an interrupt context on top of a task context. It is possible that the
> > > first use of a raw_spinlock may happen in an interrupt context. If the
> > > interrupt happens when the task has set the hazard pointer and iterating the
> > > hash list, the value of the hazard pointer may be overwritten. Alternatively
> > > we could have multiple slots for the hazard pointer, but that will make the
> > > code more complicated. Or we could disable interrupt before setting the
> > > hazard pointer.
> > Or we can use lockdep_recursion:
> > 
> > 	preempt_disable();
> > 	lockdep_recursion_inc();
> > 	barrier();
> > 
> > 	WRITE_ONCE(*hazptr, ...);
> > 
> > , it should prevent the re-entrant of lockdep in irq.
> That will probably work. Or we can disable irq. I am fine with both.

Disabling irq may not work in this case, because an NMI can also happen
and call register_lock_class().

I'm experimenting a new idea here, it might be better (for general
cases), and this has the similar spirit that we could move the
protection scope of a hazard pointer from a key to a hash_list: we can
introduce a wildcard address, and whenever we do a synchronize_hazptr(),
if the hazptr slot equal to wildcard, we treat as it matches to any ptr,
hence synchronize_hazptr() will still wait until it's zero'd. Not only
this could help in the nesting case, it can also be used if the users
want to protect multiple things with this simple hazard pointer
implementation.

Regards,
Boqun

[..]