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

[Paul E. McKenney]

On Tue, Dec 20, 2016 at 01:59:14PM +0800, Boqun Feng wrote:
> 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.

Over-case check?

> > 	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.

If we are adding checks, they should be catching bugs.  This is of
course a trade-off -- too many checks makes the code less readable
and makes it more difficult to change the code.  Too few checks
makes bugs harder to pin down.

At this point, I don't really see the need for either check.  ;-)

> > 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.

#2 is the more important of the two, though you still have not
convinced me that those particular blank lines are helping.  Making
these functions longer isn't necessarily a good thing.

> > 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.

In which case it is not needed.

> 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?

The calls from notify_cpu_starting() are called from the various
start_kernel_secondary(), secondary_start_kernel(), and similarly
named functions.  These are called on the incoming CPU early in that
CPU's execution.  The call from rcu_init() is correct until such time
as more than one CPU can be running at rcu_init() time.  And that
day might be coming, so please see the untested patch below.

							Thanx, Paul

------------------------------------------------------------------------

commit 1e84402587173d6d4da8645689f0e24c877b3269
Author: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
Date:   Tue Dec 20 07:17:58 2016 -0800

    rcu: Make rcu_cpu_starting() use its "cpu" argument
    
    The rcu_cpu_starting() function uses this_cpu_ptr() to locate the
    incoming CPU's rcu_data structure.  This works for the boot CPU and for
    all CPUs onlined after rcu_init() executes (during very early boot).
    Currently, this is the full set of CPUs, so all is well.  But if
    anyone ever parallelizes boot before rcu_init() time, it will fail.
    This commit therefore substitutes the rcu_cpu_starting() function's
    this_cpu_pointer() for per_cpu_ptr(), future-proofing the code and
    (arguably) improving readability.
    
    Reported-by: Boqun Feng <boqun.feng@xxxxxxxxx>
    Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index b9d3c0e30935..083cb8a6299c 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -4017,7 +4017,7 @@ void rcu_cpu_starting(unsigned int cpu)
 	struct rcu_state *rsp;
 
 	for_each_rcu_flavor(rsp) {
-		rdp = this_cpu_ptr(rsp->rda);
+		rdp = per_cpu_ptr(rsp->rda, cpu);
 		rnp = rdp->mynode;
 		mask = rdp->grpmask;
 		raw_spin_lock_irqsave_rcu_node(rnp, flags);