Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Mon, 12 Dec 2022 18:36:38 +0100
In-Reply-To: <cdcfcd64-c76f-0d2d-6653-0229c956f2bc@intel.com>
Mime-Version: 1.0
References: <cdcfcd64-c76f-0d2d-6653-0229c956f2bc@intel.com>
Message-ID: <20221212173638.1858573-1-peternewman@google.com>
Subject: Re: [PATCH v4 1/2] x86/resctrl: Update task closid/rmid with task_call_func()
From:   Peter Newman <peternewman@google.com>
To:     reinette.chatre@intel.com
Cc:     bp@alien8.de, derkling@google.com, eranian@google.com,
        fenghua.yu@intel.com, hpa@zytor.com, james.morse@arm.com,
        jannh@google.com, kpsingh@google.com, linux-kernel@vger.kernel.org,
        mingo@redhat.com, peternewman@google.com, tglx@linutronix.de,
        x86@kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

Hi Reinette,

On Sat, Dec 10, 2022 at 12:54 AM Reinette Chatre <reinette.chatre@intel.com=
> wrote:
> On 12/8/2022 2:30 PM, Peter Newman wrote:
> > Based on this, I'll just sketch out the first scenario below and drop
> > (2) from the changelog. This also implies that the group update cases
>
> ok, thank you for doing that analysis.
>
> > can use a single smp_mb() to provide all the necessary ordering, becaus=
e
> > there's a full barrier on context switch for it to pair with, so I don'=
t
> > need to broadcast IPI anymore. =C2=A0I don't know whether task_call_fun=
c() is
>
> This is not clear to me because rdt_move_group_tasks() seems to have the
> same code as shown below as vulnerable to re-ordering. Only difference
> is that it uses the "//false" checks to set a bit in the cpumask for a
> later IPI instead of an immediate IPI.

An smp_mb() between writing the new task_struct::{closid,rmid} and
calling task_curr() would prevent the reordering I described, but I
was worried about the cost of executing a full barrier for every
matching task.

I tried something like this:

for_each_process_thread(p, t) {
	if (!from || is_closid_match(t, from) ||
	    is_rmid_match(t, from)) {
		WRITE_ONCE(t->closid, to->closid);
		WRITE_ONCE(t->rmid, to->mon.rmid);
		/* group moves are serialized by rdt */
		t->resctrl_dirty =3D true;
	}
}
if (IS_ENABLED(CONFIG_SMP) && mask) {
	/* Order t->{closid,rmid} stores before loads in task_curr() */
	smp_mb();
	for_each_process_thread(p, t) {
		if (t->resctrl_dirty) {
			if (task_curr(t))
				cpumask_set_cpu(task_cpu(t), mask);
			t->resctrl_dirty =3D false;
		}
	}
}

I repeated the `perf bench sched messaging -g 40 -l100000 ` benchmark
from before[1] to compare this with the baseline, and found that it
only increased the time to delete the benchmark's group from 1.65ms to
1.66ms, so it's an alternative to what I last posted.

I could do something similar in the single-task move, but I don't think
it makes much of a performance difference in that case. It's only a win
for the group move because the synchronization cost doesn't grow with
the group size.

[1] https://lore.kernel.org/lkml/20221129111055.953833-3-peternewman@google=
.com/


>
> > faster than an smp_mb(). I'll take some measurements to see.
> >
> > The presumed behavior is __rdtgroup_move_task() not seeing t1 running
> > yet implies that it observes the updated values:
> >
> > CPU 0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 CPU 1
> > ----- =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 -----
> > (t1->{closid,rmid} -> {1,1}) =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0(=
rq->curr -> t0)
> >
> > __rdtgroup_move_task():
> > =C2=A0 t1->{closid,rmid} <- {2,2}
> > =C2=A0 curr <- t1->cpu->rq->curr
> > =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 __sch=
edule():
> > =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=
=A0 rq->curr <- t1
> > =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 resct=
rl_sched_in():
> > =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =
=C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=A0 =C2=
=A0 t1->{closid,rmid} -> {2,2}
> > =C2=A0 if (curr =3D=3D t1) // false
> > =C2=A0 =C2=A0 IPI(t1->cpu)
>
> I understand that the test is false when it may be expected to be true, b=
ut
> there does not seem to be a problem because of that. t1 was scheduled in =
with
> the correct CLOSID/RMID and its CPU did not get an unnecessary IPI.

Yes, this one was reminding the reader of the correct behavior. I can
just leave it out.

-Peter