Re: [PATCH 2/3] rcu: Equip sleepable RCU with lockdep dependency graph checks

[Paul E. McKenney]

On Mon, Jan 16, 2023 at 09:54:32AM -0800, Boqun Feng wrote:
> On Mon, Jan 16, 2023 at 06:36:43PM +0100, Paolo Bonzini wrote:
> > On 1/13/23 20:11, Paul E. McKenney wrote:
> > > On Fri, Jan 13, 2023 at 10:05:22AM -0800, Boqun Feng wrote:
> > > > On Fri, Jan 13, 2023 at 03:29:49AM -0800, Paul E. McKenney wrote:
> > > > I prefer that the first two patches go through your tree, because it
> > > > reduces the synchronization among locking, rcu and KVM trees to the
> > > > synchronization betwen rcu and KVM trees.
> > > 
> > > Very well, I have queued and pushed these with the usual wordsmithing,
> > > thank you!
> > 
> > I'm worried about this case:
> > 
> > 	CPU 0				CPU 1
> > 	--------------------		------------------
> > 	lock A				srcu lock B
> > 	srcu lock B			lock A
> > 	srcu unlock B			unlock A
> > 	unlock A			srcu unlock B
> > 
> > While a bit unclean, there is nothing that downright forbids this; as long
> > as synchronize_srcu does not happen inside lock A, no deadlock can occur.
> > 
> 
> First, even with my change, lockdep won't report this as a deadlock
> because srcu_read_lock() is annotated as a recursive (unfair) read lock
> (the "read" parameter for lock_acquire() is 2) and in this case lockdep
> knows that it won't cause deadlock.
> 
> For SRCU, the annotation mapping that is 1) srcu_read_lock() is marked
> as recursive read lock and 2) synchronize_srcu() is marked as write lock
> sync, recursive read locks themselves cannot cause deadlocks and lockdep
> is aware of it.
> 
> Will add a selftest for this later.
> 
> > However, if srcu is replaced with an rwlock then lockdep should and does
> > report a deadlock.  Boqun, do you get a false positive or do your patches
> 
> To be more precise, to have a deadlock, the read lock on CPU 0 has to be
> a *fair* read lock (i.e. non-recursive reader, the "read" parameter for
> lock_acquire is 1)
> 
> > correctly suppress this?
> 
> I'm pretty sure that lockdep handles this ;-)

And lockdep agrees, refusing to complain about the following:

	idx = srcu_read_lock(&srcu1);
	mutex_lock(&mut1);
	mutex_unlock(&mut1);
	srcu_read_unlock(&srcu1, idx);

	mutex_lock(&mut1);
	idx = srcu_read_lock(&srcu1);
	srcu_read_unlock(&srcu1, idx);
	mutex_unlock(&mut1);

							Thanx, Paul