Re: [PATCH 4/4] rcu: Name internal polling flag

[Paul E. McKenney]

On Tue, Mar 22, 2022 at 11:32:24AM +0100, Frederic Weisbecker wrote:
> On Mon, Mar 21, 2022 at 07:11:07PM -0700, Paul E. McKenney wrote:
> > On Wed, Mar 16, 2022 at 03:42:55PM +0100, Frederic Weisbecker wrote:
> > > Give a proper self-explanatory name to the expedited grace period
> > > internal polling flag.
> > > 
> > > Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > > Cc: Neeraj Upadhyay <quic_neeraju@xxxxxxxxxxx>
> > > Cc: Boqun Feng <boqun.feng@xxxxxxxxx>
> > > Cc: Uladzislau Rezki <uladzislau.rezki@xxxxxxxx>
> > > Cc: Joel Fernandes <joel@xxxxxxxxxxxxxxxxx>
> > > ---
> > >  kernel/rcu/rcu.h      | 5 +++++
> > >  kernel/rcu/tree.c     | 2 +-
> > >  kernel/rcu/tree_exp.h | 9 +++++----
> > >  3 files changed, 11 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
> > > index eccbdbdaa02e..8a62bb416ba4 100644
> > > --- a/kernel/rcu/rcu.h
> > > +++ b/kernel/rcu/rcu.h
> > > @@ -30,6 +30,11 @@
> > >  #define RCU_GET_STATE_USE_NORMAL	0x2
> > >  #define RCU_GET_STATE_BAD_FOR_NORMAL	(RCU_GET_STATE_FROM_EXPEDITED | RCU_GET_STATE_USE_NORMAL)
> > >  
> > > +/*
> > > + * Low-order bit definitions for polled grace-period internals.
> > > + */
> > > +#define RCU_EXP_SEQ_POLL_DONE 0x1
> > > +
> > >  /*
> > >   * Return the counter portion of a sequence number previously returned
> > >   * by rcu_seq_snap() or rcu_seq_current().
> > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > index 5da381a3cbe5..b3223b365f9f 100644
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -4679,7 +4679,7 @@ static void __init rcu_init_one(void)
> > >  			spin_lock_init(&rnp->exp_lock);
> > >  			mutex_init(&rnp->boost_kthread_mutex);
> > >  			raw_spin_lock_init(&rnp->exp_poll_lock);
> > > -			rnp->exp_seq_poll_rq = 0x1;
> > > +			rnp->exp_seq_poll_rq = RCU_EXP_SEQ_POLL_DONE;
> > >  			INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
> > >  		}
> > >  	}
> > > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> > > index c4a19c6a83cf..7ccb909d6355 100644
> > > --- a/kernel/rcu/tree_exp.h
> > > +++ b/kernel/rcu/tree_exp.h
> > > @@ -910,14 +910,14 @@ static void sync_rcu_do_polled_gp(struct work_struct *wp)
> > >  	unsigned long s;
> > >  
> > >  	s = READ_ONCE(rnp->exp_seq_poll_rq);
> > > -	if (s & 0x1)
> > > +	if (s & RCU_EXP_SEQ_POLL_DONE)
> > >  		return;
> > >  	while (!sync_exp_work_done(s))
> > >  		__synchronize_rcu_expedited(true);
> > 
> > One additional question.  If we re-read rnp->exp_seq_poll_rq on each pass
> > through the loop, wouldn't we have less trouble with counter wrap?
> 
> We can indeed do that, though it won't eliminate the possibility of wrapping.

True.  But in conjunction with an exact check for expired grace-period
sequence number, it reduces the maximum addtional penalty for wrapping
to two grace periods.

							Thanx, Paul

> > >  	raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
> > >  	s = rnp->exp_seq_poll_rq;
> > > -	if (!(s & 0x1) && sync_exp_work_done(s))
> > > -		WRITE_ONCE(rnp->exp_seq_poll_rq, s | 0x1);
> > > +	if (!(s & RCU_EXP_SEQ_POLL_DONE) && sync_exp_work_done(s))
> > > +		WRITE_ONCE(rnp->exp_seq_poll_rq, s | RCU_EXP_SEQ_POLL_DONE);
> > >  	raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
> > >  }
> > >  
> > > @@ -946,7 +946,8 @@ unsigned long start_poll_synchronize_rcu_expedited(void)
> > >  	rnp = rdp->mynode;
> > >  	if (rcu_init_invoked())
> > >  		raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
> > > -	if ((rnp->exp_seq_poll_rq & 0x1) || ULONG_CMP_LT(rnp->exp_seq_poll_rq, s)) {
> > > +	if ((rnp->exp_seq_poll_rq & RCU_EXP_SEQ_POLL_DONE) ||
> > > +	    ULONG_CMP_LT(rnp->exp_seq_poll_rq, s)) {
> > >  		WRITE_ONCE(rnp->exp_seq_poll_rq, s);
> > >  		if (rcu_init_invoked())
> > >  			queue_work(rcu_gp_wq, &rnp->exp_poll_wq);
> > > -- 
> > > 2.25.1
> > >