Re: [PATCH 25/48] rcu: Mark writes to rcu_sync ->gp_count field

[Paul E. McKenney]

On Sun, May 12, 2024 at 12:53:06PM +0200, Oleg Nesterov wrote:
> On 05/10, Paul E. McKenney wrote:
> >
> > On Fri, May 10, 2024 at 01:31:49PM +0200, Oleg Nesterov wrote:
> >
> > > Why is that?
> >
> > Because I run KCSAN on RCU using Kconfig options that cause KCSAN
> > to be more strict.
> 
> Yes, I see now.
> 
> > > but how can KCSAN detect that all accesses to X are properly marked? I see nothing
> > > KCSAN-related in the definition of WRITE_ONCE() or READ_ONCE().
> >
> > The trick is that KCSAN sees the volatile cast that both READ_ONCE()
> > and WRITE_ONCE() use.
> 
> Hmm. grep-grep-grep... I seem to understand, DEFINE_TSAN_VOLATILE_READ_WRITE.
> So __tsan_volatile_readX() will use KCSAN_ACCESS_ATOMIC.
> 
> Thanks!

You got it!!!

> > > And what does the "all accesses" above actually mean? The 2nd version does
> > >
> > > 	WRITE_ONCE(X, X+1);
> > >
> > > but "X + 1" is the plain/unmarked access?
> >
> > ...
> >
> > In that case, the "X+1" cannot be involved in a data race, so KCSAN
> > won't complain.
> 
> Yes, yes, I understand now.
> 
> Paul, thanks for your explanations! and sorry for wasting your time by
> provoking the unnecessarily long discussion.

Not a problem and absolutely no need to apologize!  Of course, please do
pass this information on to anyone else needing it.

> I am going to send the trivial patch which moves these WARN_ON()'s under
> spin_lock(), this looks more clean to me. But I won't argue if you prefer
> your original patch.

Actually, I like your patch quite a bit better than I do my original.
In fact, I feel quite foolish that I did not think of this to begin with.
With your way, we have strict locking for that field and can therefore
just use plain C-language accesses for all accesses to it.  KCSAN will
then warn us of any buggy lockless access to that field, even if that
buggy access uses READ_ONCE().  Much much better your way!!!

So thank you for persisting on this!

							Thanx, Paul