On Sat, Jan 25, 2025 at 12:03:58AM +0100, Frederic Weisbecker wrote:
> Le Fri, Dec 13, 2024 at 11:49:49AM -0800, Paul E. McKenney a écrit :
> > diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
> > index 2f9c9272cd486..d2a91f705a4ab 100644
> > --- a/kernel/rcu/rcu.h
> > +++ b/kernel/rcu/rcu.h
> > @@ -162,7 +162,7 @@ static inline bool rcu_seq_done_exact(unsigned long *sp, unsigned long s)
> > {
> > unsigned long cur_s = READ_ONCE(*sp);
> >
> > - return ULONG_CMP_GE(cur_s, s) || ULONG_CMP_LT(cur_s, s - (2 * RCU_SEQ_STATE_MASK + 1));
> > + return ULONG_CMP_GE(cur_s, s) || ULONG_CMP_LT(cur_s, s - (3 * RCU_SEQ_STATE_MASK + 1));
>
> This might need a comment.
Good point! Would you like to propose one?
> The way I understand it is that rcu_state.gp_seq might be seen started while
> root_rnp->gp_seq is not. So rcu_seq_snap() on the started rcu_state.gp_seq
> may return maximum 2 full GPs ahead of root_rnp->gp_seq. And therefore it takes below
> 2 GPs to safely deduce we wrapped around.
Exactly!
> Should it be ULONG_CMP_LT(cur_s, s - (2 * (RCU_SEQ_STATE_MASK + 1))) ?
Quite possibly. I freely admit that I allowed a bit of slop because
time was of the essence (holidays and all that) and also it does not
hurt much to lose a couple of counts out of a 2^32 cycle, to say nothing
of the common-case 2^64 cycle. It would not hurt to be exact, but it
would be necessary to convince ourselves that we were not off by one in
the wrong direction.
I would be happy to see a patch, as long as it was sufficiently
convincing.
> Or am I missing something?
Not that I can see. So the answer is probably "yes". ;-)
Thanx, Paul
