On Fri, Mar 19, 2021 at 11:10:40PM +0100, Frederic Weisbecker wrote:
> On Fri, Mar 19, 2021 at 10:51:16AM -0700, Paul E. McKenney wrote:
> > On Fri, Mar 19, 2021 at 02:58:54PM +0100, Frederic Weisbecker wrote:
> > > It's all a matter of personal taste but if I may suggest some namespace
> > > modifications:
> > >
> > > get_state_synchronize_rcu() -> synchronize_rcu_poll_start_raw()
> > > start_poll_synchronize_rcu() -> synchronize_rcu_poll_start()
> > > poll_state_synchronize_rcu() -> synchronize_rcu_poll()
> > > cond_synchronize_rcu() -> synchronize_rcu_cond()
> > >
> > > But it's up to you really.
> >
> > I am concerned about starting anything "synchronize_rcu" if that
> > thing doesn't unconditionally wait for a grace period. "What do
> > you mean that there was no grace period? Don't you see that call to
> > synchronize_rcu_poll_start_raw()???"
>
> I see, that could indeed be confusing.
>
> > This objection doesn't apply to cond_synchronize_rcu(), but it is
> > already in use, so any name change should be worked with the users.
> > All two of them. ;-)
>
> Probably not worth it. We have cond_resched() as a schedule() counterpart
> for a reference after all.
Good point!
> > > > /**
> > > > + * start_poll_state_synchronize_rcu - Snapshot and start RCU grace
> > > > period
> > > > + *
> > > > + * Returns a cookie that is used by a later call to
> > > > cond_synchronize_rcu()
> > >
> > > It may be worth noting that calling start_poll_synchronize_rcu() and then
> > > pass the cookie to cond_synchronize_rcu() soon after may end up waiting
> > > for
> > > one more grace period.
> >
> > You mean this sequence of events?
> >
> > 1. cookie = start_poll_synchronize_rcu()
> >
> > 2. The grace period corresponding to cookie is almost over...
> >
> > 3. cond_synchronize_rcu() checks the cookie and sees that the
> > grace period has not yet expired.
> >
> > 4. The grace period corresponding to cookie completes.
> >
> > 5. Someone else starts a grace period.
> >
> > 6. cond_synchronize_rcu() invokes synchronize_rcu(), which waits
> > for the just-started grace period plus another grace period.
> > Thus, there has been no fewer than three full grace periods
> > between the call to start_poll_synchronize_rcu() and the
> > return from cond_synchronize_rcu().
> >
> > Yes, this can happen! And it can be worse, for example, it is quite
> > possible that cond_synchronize_rcu() would be preempted for multiple
> > grace periods at step 5, in which case it would still wait for almost
> > two additional grace periods.
> >
> > Or are you thinking of something else?
>
> I didn't even think that far.
> My scenario was:
>
> 1. cookie = start_poll_synchronize_rcu()
>
>
> 2. cond_synchronize_rcu() checks the cookie and sees that the
> grace period has not yet expired. So it calls synchronize_rcu()
> which queues a callback.
>
> 3. The grace period for the cookie eventually completes.
>
> 4. The callback queued in 2. gets assigned a new grace period number.
> That new grace period starts.
>
> 5. The new grace period completes and synchronize_rcu() returns.
>
>
> But I think this is due to some deep misunderstanding from my end.
You mean like this?
oldstate = start_poll_synchronize_rcu();
// Why wait? Beat the rush!!!
cond_synchronize_rcu(oldstate);
This would be a bit silly (why not just call synchronize_rcu()?),
and yes, this would unconditionally get you an extra RCU grace period.
Then again, any call to cond_synchronize_rcu() before the desired grace
period has expired will get you an extra grace period, and maybe more.
So a given use case either needs to not care about the added latency
or have a high probability of invoking cond_synchronize_rcu() after
the desired grace period has expired.
> > > > + * If a full RCU grace period has elapsed since the earlier call from
> > > > + * which oldstate was obtained, return @true, otherwise return @false.
> > > > + * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
> > >
> > > Rephrase suggestion for the last sentence:
> > >
> > > "In case of failure, it's up to the caller to try polling again later or
> > > invoke synchronize_rcu() to wait for a new full grace period to complete."
> >
> > How about like this?
> >
> > /**
> > * poll_state_synchronize_rcu - Conditionally wait for an RCU grace period
> > *
> > * @oldstate: return from call to get_state_synchronize_rcu() or
> > start_poll_synchronize_rcu()
> > *
> > * If a full RCU grace period has elapsed since the earlier call from
> > * which oldstate was obtained, return @true, otherwise return @false.
> > * If @false is returned, it is the caller's responsibilty to invoke this
> > * function later on until it does return @true. Alternatively, the caller
> > * can explicitly wait for a grace period, for example, by passing @oldstate
> > * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
>
> Yes very nice!
You got it!
Thanx, Paul
