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

[Peter Zijlstra]

On Sun, Nov 11, 2018 at 05:47:36PM -0800, Paul E. McKenney wrote:
> On Mon, Nov 12, 2018 at 01:53:29AM +0100, Peter Zijlstra wrote:
> > On Sun, Nov 11, 2018 at 04:45:28PM -0800, Paul E. McKenney wrote:
> > > On Mon, Nov 12, 2018 at 01:12:33AM +0100, Peter Zijlstra wrote:
> > > > On Sun, Nov 11, 2018 at 11:43:52AM -0800, Paul E. McKenney wrote:
> > > > > Now that synchronize_rcu() waits for preempt-disable regions of code
> > > > > as well as RCU read-side critical sections, synchronize_sched() can be
> > > > > replaced by synchronize_rcu().  This commit therefore makes this change.
> > > > 
> > > > Yes, but it also waits for an actual RCU quiestent state, which makes
> > > > synchoinize_rcu() potentially much more expensive than an actual
> > > > synchronize_sched().
> > > 
> > > None of the readers have changed.
> > > 
> > > For the updaters, if CONFIG_PREEMPT=n, synchronize_rcu() and
> > > synchronize_sched() always were one and the same.  When CONFIG_PREEMPT=y,
> > > synchronize_rcu() and synchronize_sched() are now one and the same.
> > 
> > The Changelog does not state this; and does the commit that makes that
> > happen state the regression potential?
> 
> The Changelog says this:
> 
> 	Now that synchronize_rcu() waits for preempt-disable
> 	regions of code as well as RCU read-side critical sections,
> 	synchronize_sched() can be replaced by synchronize_rcu().
> 	This commit therefore makes this change.
> 
> The "synchronize_rcu() waits for preempt-disable regions of code as
> well as RCU read-side critical sections" seems pretty unambiguous to me.
> Exactly what more are you wanting said there?

The quoted bit only states that synchronize_rcu() is sufficient; it does
not say it is equivalent and the patch is a nop. It also doesn't say
that the purpose is to get rid of the synchronize_sched() function.

> There were quite a few commits involved in making this happen.  Perhaps
> the most pertinent are these:
> 
> 3e3100989869 ("rcu: Defer reporting RCU-preempt quiescent states when disabled")
> 45975c7d21a1 ("rcu: Define RCU-sched API in terms of RCU for Tree RCU PREEMPT builds")

The latter; it does not mention that this will possible make
synchronize_sched() quite a bit more expensive on PREEMPT=y builds :/

> Normal grace periods are almost always quite long compared to typical
> read-side critical sections, preempt-disable regions of code, and so on.
> So in the common case this should be OK.  Or are you instead worried
> about synchronize_sched_expedited()?

No, I still feel expedited should not exist at all ;-)

But for PREEMPT=y synchronize_sched() can be quite a bit shorter than
synchronize_rcu(), since we don't have to wait for preempted read side
stuff.

> > > > So why are we doing this?
> > > 
> > > Given that synchronize_rcu() and synchronize_sched() are now always one
> > > and the same, this is a distinction without a difference.
> > 
> > The Changelog did not state a reason for the patch. Therefore it is a
> > bad patch.
> 
> ???  Here is the current definition of synchronize_sched() in mainline:
> 
> 	static inline void synchronize_sched(void)
> 	{
> 		synchronize_rcu();
> 	}

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.