Re: [PATCH tip/core/rcu 23/41] sched: Replace synchronize_sched() with synchronize_rcu()

From: Paul E. McKenney
Date: Mon Nov 12 2018 - 17:21:25 EST


On Mon, Nov 12, 2018 at 07:17:41PM +0100, Peter Zijlstra wrote:
> On Mon, Nov 12, 2018 at 05:28:52AM -0800, Paul E. McKenney wrote:
> > On Mon, Nov 12, 2018 at 10:00:47AM +0100, Peter Zijlstra wrote:
>
> > > Still, better safe than sorry. It was a rather big change in behaviour,
> > > so it wouldn't have been strange to call that out.
> >
> > This guy:
> >
> > 45975c7d21a1 ("rcu: Define RCU-sched API in terms of RCU for Tree RCU PREEMPT builds")
> >
> > Has a commit log that says:
> >
> > Now that RCU-preempt knows about preemption disabling, its
> > implementation of synchronize_rcu() works for synchronize_sched(),
> > and likewise for the other RCU-sched update-side API members.
> > This commit therefore confines the RCU-sched update-side code
> > to CONFIG_PREEMPT=n builds, and defines RCU-sched's update-side
> > API members in terms of those of RCU-preempt.
> >
> > That last phrase seems pretty explicit. What am I missing here?
>
> That does not explicitly state that because RCU-preempt
> synchornize_rcu() can take _much_ longer, the new synchronize_sched()
> can now take _much_ longer too.
>
> So when someone bisects a problem to this commit; and he reads the
> Changelog, he might get the impression that was unexpected.

Of course, a preempt_disable() section of code can still be preempted
by the underlying hypervisor, so in a surprisingly large fraction of
the installed base, there really isn't that much difference.

> > Not that it matters, given that I know of no way to change a mainlined
> > commit log. I suppose I could ask Jon if he would be willing to take
> > a 2018 RCU API LWN article, if that would help.
>
> Yes, it is water under the bridge; but Changelogs should be explicit
> about behavioural changes.
>
> And while the merged RCU has the semantic behaviour required, the timing
> behaviour did change significantly.

When running on bare metal, potentially. From what I see, preemption
of RCU read-side critical sections is the exception rather than the rule.
And again, when running on hypervisors, even irq-disable regions of code
can be preempted. (And yes, there is work in flight to allow RCU to deal
with this.)

> > > > > Again, the patch didn't say that.
> > > > >
> > > > > If the Changelog would've read something like:
> > > > >
> > > > > "Since synchronize_sched() is now equivalent to synchronize_rcu(),
> > > > > replace the synchronize_sched() usage such that we can eventually remove
> > > > > the interface."
> > > > >
> > > > > It would've been clear that the patch is a nop and what the purpose
> > > > > was.
> > > >
> > > > I can easily make that change.
> > >
> > > Please, sufficient doesn't imply necessary etc.. A changelog should
> > > always clarify why we do the patch.
> >
> > ??? Did you mean to say "necessary doesn't imply sufficient"? If so,
> > what else do you feel is missing?
>
> No, I meant to say that your original Changelog only states that
> sync_rcu now covers rcu-sched behaviour. Which means that the change is
> sufficient.
>
> It completely and utterly fails to explain _why_ you're doing the
> change. Ie. you do not address why it is necessary.
>
> A Changelog should always explain why the change is needed.
>
> In this case because you want to get rid of the sync_sched() api.

Right, which is stated in your suggested wording above. So I am still
not seeing what you want added to this:

"Since synchronize_sched() is now equivalent to synchronize_rcu(),
replace the synchronize_sched() usage such that we can eventually
remove the interface."

Thanx, Paul