Date: Wed, 15 Mar 2017 10:24:29 -0700
From: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
To: Patrick Bellasi <patrick.bellasi@arm.com>
Cc: Joel Fernandes <joelaf@google.com>,
        "Joel Fernandes (Google)" <joel.opensrc@gmail.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        linux-pm@vger.kernel.org, Ingo Molnar <mingo@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>, Tejun Heo <tj@kernel.org>
Subject: Re: [RFC v3 1/5] sched/core: add capacity constraints to CPU
 controller
Reply-To: paulmck@linux.vnet.ibm.com
References: <1488292722-19410-1-git-send-email-patrick.bellasi@arm.com>
 <1488292722-19410-2-git-send-email-patrick.bellasi@arm.com>
 <CAEi0qNkAK_3a_2x9BgvyEC+bFvckGFoQX1fjFba6boC1ws_R5w@mail.gmail.com>
 <20170315112020.GA18557@e110439-lin>
 <CAJWu+or4AJ6_J8XJicKDJReMP6Dj7t7SbB7k78ia8FdhDdXXoA@mail.gmail.com>
 <20170315161048.GJ3637@linux.vnet.ibm.com>
 <20170315164439.GG18557@e110439-lin>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20170315164439.GG18557@e110439-lin>
User-Agent: Mutt/1.5.21 (2010-09-15)
Message-Id: <20170315172429.GK3637@linux.vnet.ibm.com>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 4347
Lines: 101

On Wed, Mar 15, 2017 at 04:44:39PM +0000, Patrick Bellasi wrote:
> On 15-Mar 09:10, Paul E. McKenney wrote:
> > On Wed, Mar 15, 2017 at 06:20:28AM -0700, Joel Fernandes wrote:
> > > On Wed, Mar 15, 2017 at 4:20 AM, Patrick Bellasi
> > > <patrick.bellasi@arm.com> wrote:
> > > > On 13-Mar 03:46, Joel Fernandes (Google) wrote:
> > > >> On Tue, Feb 28, 2017 at 6:38 AM, Patrick Bellasi
> > > >> <patrick.bellasi@arm.com> wrote:
> > > >> > The CPU CGroup controller allows to assign a specified (maximum)
> > > >> > bandwidth to tasks within a group, however it does not enforce any
> > > >> > constraint on how such bandwidth can be consumed.
> > > >> > With the integration of schedutil, the scheduler has now the proper
> > > >> > information about a task to select  the most suitable frequency to
> > > >> > satisfy tasks needs.
> > > >> [..]
> > > >>
> > > >> > +static u64 cpu_capacity_min_read_u64(struct cgroup_subsys_state *css,
> > > >> > +                                    struct cftype *cft)
> > > >> > +{
> > > >> > +       struct task_group *tg;
> > > >> > +       u64 min_capacity;
> > > >> > +
> > > >> > +       rcu_read_lock();
> > > >> > +       tg = css_tg(css);
> > > >> > +       min_capacity = tg->cap_clamp[CAP_CLAMP_MIN];
> > > >>
> > > >> Shouldn't the cap_clamp be accessed with READ_ONCE (and WRITE_ONCE in
> > > >> the write path) to avoid load-tearing?
> > > >
> > > > tg->cap_clamp is an "unsigned int" and thus I would expect a single
> > > > memory access to write/read it, isn't it? I mean: I do not expect the
> > > > compiler "to mess" with these accesses.
> > > 
> > > This depends on compiler and arch. I'm not sure if its in practice
> > > these days an issue, but see section on 'load tearing' in
> > > Documentation/memory-barriers.txt . If compiler decided to break down
> > > the access to multiple accesses due to some reason, then might be a
> > > problem.
> > 
> > The compiler might also be able to inline cpu_capacity_min_read_u64()
> > fuse the load from tg->cap_clamp[CAP_CLAMP_MIN] with other accesses.
> > If min_capacity is used several times in the ensuing code, the compiler
> > could reload multiple times from tg->cap_clamp[CAP_CLAMP_MIN], which at
> > best might be a bit confusing.
> 
> That's actually an interesting case, however I don't think it applies
> in this case since cpu_capacity_min_read_u64() is called only via
> a function poninter and thus it will never be inlined. isn't it?
> 
> > > Adding Paul for his expert opinion on the matter ;)
> > 
> > My personal approach is to use READ_ONCE() and WRITE_ONCE() unless
> > I can absolutely prove that the compiler cannot do any destructive
> > optimizations.  And I not-infrequently find unsuspected opportunities
> > for destructive optimization in my own code.  Your mileage may vary.  ;-)
> 
> I guess here the main concern from Joel is that such a pattern:
> 
>    u64 var = unsigned_int_value_from_memory;
> 
> can result is a couple of "load from memory" operations.

Indeed it can.  I first learned this the hard way in the early 1990s,
so 20-year-old compiler optimizations are quite capable of making this
sort of thing happen.

> In that case a similar:
> 
>   unsigned_int_left_value = new_unsigned_int_value;
> 
> executed on a different thread can overlap with the previous memory
> read operations and ending up in "var" containing a not consistent
> value.
> 
> Question is: can this really happen, given the data types in use?

So we have an updater changing the value of unsigned_int_left_value,
while readers in other threads are accessing it, correct?  And you
are asking whether the compiler can optimize the updater so as to
mess up the readers, right?

One such optimization would be a byte-wise write, though I have no
idea why a compiler would do such a thing assuming that the variable
is reasonably sized and aligned.  Another is that the compiler could
use the variable as temporary storage just before the assignment.
(You haven't told the compiler that anyone else is reading it, though
I don't know of this being done by production compilers.)  A third is
that the compiler could fuse successive stores, which might or might
not be a problem, depending.

Probably more, but that should be a start.  ;-)

							Thanx, Paul

> Thanks! ;-)
> 
> -- 
> #include <best/regards.h>
> 
> Patrick Bellasi
>