On Wed, Jul 20, 2022 at 06:51:45PM -0700, Boqun Feng wrote:
> On Wed, Jul 20, 2022 at 06:04:55PM -0700, Paul E. McKenney wrote:
> [...]
> > > > @@ -3860,7 +3944,7 @@ unsigned long get_state_synchronize_rcu(void)
> > > > * before the load from ->gp_seq.
> > > > */
> > > > smp_mb(); /* ^^^ */
> > > > - return rcu_seq_snap(&rcu_state.gp_seq);
> > > > + return rcu_seq_snap(&rcu_state.gp_seq_polled);
> > >
> > > I happened to run into this. There is one usage of
> > > get_state_synchronize_rcu() in start_poll_synchronize_rcu(), in which
> > > the return value of get_state_synchronize_rcu() ("gp_seq") will be used
> > > for rcu_start_this_gp(). I don't think this is quite right, because
> > > after this change, rcu_state.gp_seq and rcu_state.gp_seq_polled are
> > > different values, in fact ->gp_seq_polled is greater than ->gp_seq
> > > by how many synchronize_rcu() is called in early boot.
> > >
> > > Am I missing something here?
> >
> > It does not appear that your are missing anything, sad to say!
> >
> > Does the following make it work better?
> >
> > Thanx, Paul
> >
> > ------------------------------------------------------------------------
> >
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index 2122359f0c862..cf2fd58a93a41 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -3571,7 +3571,7 @@ EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
> > unsigned long start_poll_synchronize_rcu(void)
> > {
> > unsigned long flags;
> > - unsigned long gp_seq = get_state_synchronize_rcu();
> > + unsigned long gp_seq = rcu_seq_snap(&rcu_state.gp_seq);
>
> get_state_synchronize_rcu() is still needed, because we need to return
> a cookie for polling for this function. Something like below maybe? Hope
> I didn't mess up the ordering ;-)
My thought is to combine your comment with my functionally equivalent
code that avoids the extra variable. If that works for you (and if it
works, for that matter), does Co-developed-by work for you?
Thanx, Paul
> Regards,
> Boqun
>
> ---------------
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 84d281776688..0f9134871289 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -3571,11 +3583,39 @@ EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
> unsigned long start_poll_synchronize_rcu(void)
> {
> unsigned long flags;
> - unsigned long gp_seq = get_state_synchronize_rcu();
> + unsigned long gp_seq_poll = get_state_synchronize_rcu();
> + unsigned long gp_seq;
> bool needwake;
> struct rcu_data *rdp;
> struct rcu_node *rnp;
>
> + /*
> + * Need to start a gp if no gp has been started yet.
> + *
> + * Note that we need to snapshot gp_seq after gp_seq_poll, otherwise
> + * consider the follow case:
> + *
> + * <no gp in progress> // gp# is 0
> + * snapshot gp_seq // gp #2 will be set as needed
> + * <a gp passed>
> + * // gp# is 1
> + * snapshot gp_seq_poll // polling gets ready until gp #3
> + *
> + * then the following rcu_start_this_gp() won't mark gp #3 as needed,
> + * and polling won't become ready if others don't start a gp.
> + *
> + * And the following case is fine:
> + *
> + * <no gp in progress> // gp# is 0
> + * snapshot gp_seq_poll // polling gets ready until gp #2
> + * <a gp passed>
> + * // gp# is 1
> + * snapshot gp_seq // gp #3 will be set as needed
> + *
> + * Also note, we rely on the smp_mb() in get_state_synchronize_rcu()
> + * to order the two snapshots.
> + */
> + gp_seq = rcu_seq_snap(&rcu_state.gp_seq);
> lockdep_assert_irqs_enabled();
> local_irq_save(flags);
> rdp = this_cpu_ptr(&rcu_data);
> @@ -3585,7 +3625,7 @@ unsigned long start_poll_synchronize_rcu(void)
> raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> if (needwake)
> rcu_gp_kthread_wake();
> - return gp_seq;
> + return gp_seq_poll;
> }
> EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
