Re: [RFC v2 4/5] rcu: Use for_each_leaf_node_cpu() in force_qs_rnp()

[Boqun Feng]

On Mon, Dec 19, 2016 at 09:09:13PM -0800, Paul E. McKenney wrote:
> On Mon, Dec 19, 2016 at 11:15:15PM +0800, Boqun Feng wrote:
> > On Thu, Dec 15, 2016 at 02:51:36PM +0000, Colin Ian King wrote:
> > > On 15/12/16 14:42, Boqun Feng wrote:
> > > > On Thu, Dec 15, 2016 at 12:04:59PM +0000, Mark Rutland wrote:
> > > >> On Thu, Dec 15, 2016 at 10:42:03AM +0800, Boqun Feng wrote:
> > > >>> ->qsmask of an RCU leaf node is usually more sparse than the
> > > >>> corresponding cpu_possible_mask. So replace the
> > > >>> for_each_leaf_node_possible_cpu() in force_qs_rnp() with
> > > >>> for_each_leaf_node_cpu() to save several checks.
> > > >>>
> > > >>> [Note we need to use "1UL << bit" instead of "1 << bit" to generate the
> > > >>> corresponding mask for a bit because @mask is unsigned long, this was
> > > >>> spotted by Colin Ian King <colin.king@xxxxxxxxxxxxx> and CoverityScan in
> > > >>> a previous version of this patch.]
> > > >>
> > > >> Nit: This note can go now that we use leaf_node_cpu_bit(). ;)
> > > >>
> > > > 
> > > > I kinda keep this here for honoring the effort of finding out this bug
> > > > from Colin, but yes, it's no longer needed here for the current code.
> > > 
> > > Yep, remove it.
> > > 
> > 
> > Paul, here is a modified version of this patch, what I only did is
> > removing this note.
> > 
> > Besides I rebased the whole series on the current rcu/dev branch of -rcu
> > tree, on this very commit:
> > 
> > 	8e9b2521b18a ("doc: Quick-Quiz answers are now inline")
> > 
> > And I put the latest version at
> > 
> > git://git.kernel.org/pub/scm/linux/kernel/git/boqun/linux.git leaf-node
> > 
> > If you thought it's better, I could send a v3 ;-)
> 
> I would feel better about this patchset if it reduced the number of lines
> of code rather than increasing them.  That said, part of the increase
> is a commment.  Still, I am not convinced that the extra level of macro
> is carrying its weight.
> 
> dbf18a2422e2 ("rcu: Introduce for_each_leaf_node_cpu()")
> 
> 	The commit log needs a bit of wordsmithing.
> 
> 	The added WARN_ON_ONCE(!cpu_possible(cpu)) still seems strange.
> 	What is its purpose, really?  What does its triggering tell you?
> 	What other checks did you consider as an alternative?
> 

The check is an over-case one, it's introduced because I'm worried about
some code outside the RCU code mis-sets the ->qsmask* or ->expmask* on
an "impossible" CPU. I will explanation later in more details.

> 	And if you are going to add checks of this type, should you
> 	also check for this being a leaf rcu_node structure?
> 

I don't think I want to check that, and I don't think check
cpu_possible(cpu) in the macro is similar to that.

> 3f0b4ba1fe94 ("rcu: Use for_each_leaf_node_cpu() in RCU stall checking")
> 
> 	This does look a bit nicer, but why the added blank lines?
> 	Are they really helping?
> 
> 	The commit log seems a bit misplaced.  This code is almost never
> 	executed (once per 21 seconds at the most), so performance really
> 	isn't a consideration.	The simpler-looking code might be.
> 
> fd799f1ac7b7 ("rcu: Use for_each_leaf_node_cpu() in ->expmask iteration")
> 
> 	Ditto on blank lines.
> 
> 	Again, this code is executed per expedited grace period, so
> 	performance really isn't a big deal.  More of a big deal than
> 	the stall-warning code, but we still are way off of any fastpath.
> 
> 69a1baedbf42 ("rcu: Use for_each_leaf_node_cpu() in force_qs_rnp()")
> 
> 	Ditto again on blank lines.
> 
> 	And on the commit log.  This code is executed about once
> 	per several jiffies, and on larger machines, per 20 jiffies
> 	or so.  Performance really isn't a consideration.
> 
> 7b00e50e3efb ("rcu: Use for_each_leaf_node_cpu() in online CPU iteration")
> 
> 	And another ditto on blank lines.
> 
> 	This code executes once per CPU-hotplug operation, so again isn't
> 	at all performance critical.
> 
> In short, if you are trying to sell this to me as a significant performance
> boost, I am not buying.  The added WARN_ON_ONCE() looks quite dubious,

Yep, it won't help the performance a lot, but it 

1)	helps the performance in theory, because it iterates less CPUs

2)	makes code cleaner. By "cleaner", I mean we can a) affort more
	blank lines to make loops separated from other code and b)
	descrease the indent levels for those loops. But, yes I should
	add those points in the commit log, because those are more
	visible effects.

> though perhaps I am misunderstanding its purpose.  My assumption is
> that you want to detect missing UL suffixes on bitmask constants, in
> which case I bet there is a better way.
> 

The WARN_ON_ONCE() is not for detecting missing UL suffixes on bitmask
constatns, and we don't need to check that, because we use
leaf_node_cpu_id() now. As I said, this is an over-case check, and we
can drop if we guarante that CPUs masked in ->qsmask* and ->expmask*
must be a "possible" CPU, IOW, ->qsmask* and ->expmask* are the subsets
(with offset fixed by ->grplo) of cpu_possible_mask.

Hmm.. and I just check the code, the initial values of ->qsmask* and
->expmask* are from ->qsmaskinitnext and ->expmaskinitnext, and the
latter two are set in rcu_cpu_starting() since commit

	7ec99de36f40 ("rcu: Provide exact CPU-online tracking for RCU")

, and rcu_cpu_starting() only set the corresponding bit of _this_ cpu in
a leaf node's ->qsmaskinitnext and ->expmaskinitnext. So looks like we
are safe to remove the WARN_ON_ONCE() check, because a ever-running CPU
must be a possible CPU, IIRC.

But this brings a side question, is the callsite of rcu_cpu_starting()
is correct? Given rcu_cpu_starting() ignores the @cpu parameter and only
set _this_ cpu's bit in a leaf node?

Regards,
Boqun

> Speaking of which, how do we know that this is free of bugs?
> 
> 							Thanx, Paul
> 
> > Regards,
> > Boqun
> > 
> > ------------------------>8
> > From: Boqun Feng <boqun.feng@xxxxxxxxx>
> > Date: Thu, 8 Dec 2016 23:21:11 +0800
> > Subject: [PATCH v2.1 4/5] rcu: Use for_each_leaf_node_cpu() in force_qs_rnp()
> > 
> > ->qsmask of an RCU leaf node is usually more sparse than the
> > corresponding cpu_possible_mask. So replace the
> > for_each_leaf_node_possible_cpu() in force_qs_rnp() with
> > for_each_leaf_node_cpu() to save several checks.
> > 
> > Signed-off-by: Boqun Feng <boqun.feng@xxxxxxxxx>
> > ---
> >  kernel/rcu/tree.c | 12 +++++-------
> >  1 file changed, 5 insertions(+), 7 deletions(-)
> > 
> > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > index 4ea4496f4ecc..c2b753fb7f09 100644
> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -3046,13 +3046,11 @@ static void force_qs_rnp(struct rcu_state *rsp,
> >  				continue;
> >  			}
> >  		}
> > -		for_each_leaf_node_possible_cpu(rnp, cpu) {
> > -			unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
> > -			if ((rnp->qsmask & bit) != 0) {
> > -				if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
> > -					mask |= bit;
> > -			}
> > -		}
> > +
> > +		for_each_leaf_node_cpu(rnp, rnp->qsmask, cpu)
> > +			if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
> > +				mask |= leaf_node_cpu_bit(rnp, cpu);
> > +
> >  		if (mask != 0) {
> >  			/* Idle/offline CPUs, report (releases rnp->lock. */
> >  			rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
> > -- 
> > 2.10.2
> > 
>

Attachment: signature.asc
Description: PGP signature