Date:	Thu, 28 Jan 2016 16:21:19 +0100
From:	Michal Hocko <mhocko@...nel.org>
To:	Christoph Lameter <cl@...ux.com>
Cc:	Mike Galbraith <umgwanakikbuti@...il.com>,
	Peter Zijlstra <peterz@...radead.org>,
	LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] mm, vmstat: make quiet_vmstat lighter (was: Re: fast
 path cycle muncher (vmstat: make vmstat_updater deferrable) again and shut
 down on idle)

On Wed 27-01-16 12:26:16, Christoph Lameter wrote:
> On Wed, 27 Jan 2016, Michal Hocko wrote:
> 
> > This is caused by 0eb77e988032 ("vmstat: make vmstat_updater deferrable
> > again and shut down on idle") which has placed quiet_vmstat into
> > cpu_idle_loop. The main reason here seems to be that the idle entry has
> > to get over all zones and perform atomic operations for each vmstat
> > entry even though there might be no per cpu diffs. This is a pointless
> > overhead for _each_ idle entry.
> >
> > Make sure that quiet_vmstat is as light as possible.
> >
> >  First of all it doesn't make any sense to do any local sync if the
> > current cpu is already set in oncpu_stat_off because vmstat_update puts
> > itself there only if there is nothing to do.
> 
> But there might be something to do that happened in the meantime because
> counters were incremented. It needs to be checked. The shepherd can do
> that but it will delay the folding of diffs for awhile.

shepherd ticks with the same period so it should run relatively soon.
What is important here the test_and_set is the fast path which will
trigger a lot of the current CPU is mostly running in the userspace.

> > +void quiet_vmstat(void)
> > +{
> > +	if (system_state != SYSTEM_RUNNING)
> > +		return;
> > +
> > +	/*
> > +	 * If we are already in hands of the shepherd then there
> > +	 * is nothing for us to do here.
> > +	 */
> > +	if (cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
> > +		return;
> > +
> > +	if (!need_update(smp_processor_id()))
> > +		return;
> > +
> > +	/*
> > +	 * Just refresh counters and do not care about the pending delayed
> > +	 * vmstat_update. It doesn't fire that often to matter and canceling
> > +	 * it would be too expensive from this path.
> > +	 * vmstat_shepherd will take care about that for us.
> > +	 */
> > +	refresh_cpu_vm_stats(false);
> > +}
> 
> The problem here is that there will be an additional tick generated on
> idle. This is an issue for power because now the processor has to
> needlessly wake up again, do tick processing etc  just to effectively do a
> cancel_delayed_work().
> 
> The cancelling of the work request is required to avoid this additonal
> tick.

Yes but as mentioned in the changelog calling cancel_delayed_work is
quite expensive to be run on each idle entry slow path. In addition it
is RT unfriendly. So considering a single tick vs. other reasons I think
this is a reasonable compromise.

> > @@ -1470,11 +1487,14 @@ static void vmstat_shepherd(struct work_struct *w)
> >  	get_online_cpus();
> >  	/* Check processors whose vmstat worker threads have been disabled */
> >  	for_each_cpu(cpu, cpu_stat_off)
> > -		if (need_update(cpu) &&
> > -			cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
> > -
> > -			queue_delayed_work_on(cpu, vmstat_wq,
> > -				&per_cpu(vmstat_work, cpu), 0);
> > +		if (need_update(cpu)) {
> > +			if (cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
> > +				queue_delayed_work_on(cpu, vmstat_wq,
> > +					&per_cpu(vmstat_work, cpu), 0);
> > +		} else {
> > +			cpumask_set_cpu(cpu, cpu_stat_off);
> 
> Umm the flag is already set right? This just causes a bouncing cacheline
> since we are accessing a global set of cpus,. Drop this line?

Ohh, right you are. I will remove this.

> 
> > +			cancel_delayed_work(this_cpu_ptr(&vmstat_work));
> 
> Ok so this is to move the cancel into the shepherd.

Yeah, I have added a comment to explain that.

> Aside from the two issues mentioned this looks good.

Updated version:
---
>From 1d0febff044382b88ff44d245b5afe1a6e402c62 Mon Sep 17 00:00:00 2001
From: Michal Hocko <mhocko@...e.com>
Date: Wed, 27 Jan 2016 17:24:22 +0100
Subject: [PATCH] mm, vmstat: make quiet_vmstat lighter

Mike has reported a considerable overhead of refresh_cpu_vm_stats from
the idle entry during pipe test:
    12.89%  [kernel]       [k] refresh_cpu_vm_stats.isra.12
     4.75%  [kernel]       [k] __schedule
     4.70%  [kernel]       [k] mutex_unlock
     3.14%  [kernel]       [k] __switch_to

This is caused by 0eb77e988032 ("vmstat: make vmstat_updater deferrable
again and shut down on idle") which has placed quiet_vmstat into
cpu_idle_loop. The main reason here seems to be that the idle entry has
to get over all zones and perform atomic operations for each vmstat
entry even though there might be no per cpu diffs. This is a pointless
overhead for _each_ idle entry.

Make sure that quiet_vmstat is as light as possible.

 First of all it doesn't make any sense to do any local sync if the
current cpu is already set in oncpu_stat_off because vmstat_update puts
itself there only if there is nothing to do.

Then we can check need_update which should be a cheap way to check for
potential per-cpu diffs and only then do refresh_cpu_vm_stats.

The original patch also did cancel_delayed_work which we are not doing
here. There are two reasons for that. Firstly cancel_delayed_work from
idle context will blow up on RT kernels (reported by Mike):
[    2.279582] CPU: 1 PID: 0 Comm: swapper/1 Not tainted 4.5.0-rt3 #7
[    2.280444] Hardware name: MEDION MS-7848/MS-7848, BIOS M7848W08.20C 09/23/2013
[    2.281316]  ffff88040b00d640 ffff88040b01fe10 ffffffff812d20e2 0000000000000000
[    2.282202]  ffff88040b01fe30 ffffffff81081095 ffff88041ec4cee0 ffff88041ec501e0
[    2.283073]  ffff88040b01fe48 ffffffff815ff910 ffff88041ec4cee0 ffff88040b01fe88
[    2.283941] Call Trace:
[    2.284797]  [<ffffffff812d20e2>] dump_stack+0x49/0x67
[    2.285658]  [<ffffffff81081095>] ___might_sleep+0xf5/0x180
[    2.286521]  [<ffffffff815ff910>] rt_spin_lock+0x20/0x50
[    2.287382]  [<ffffffff81075919>] try_to_grab_pending+0x69/0x240
[    2.288239]  [<ffffffff81075b16>] cancel_delayed_work+0x26/0xe0
[    2.289094]  [<ffffffff8115ec05>] quiet_vmstat+0x75/0xa0
[    2.289949]  [<ffffffff8109ab38>] cpu_idle_loop+0x38/0x3e0
[    2.290800]  [<ffffffff8109aef3>] cpu_startup_entry+0x13/0x20
[    2.291647]  [<ffffffff81036164>] start_secondary+0x114/0x140

And secondly, even on !RT kernels it might add some non trivial overhead
which is not necessary. Even if the vmstat worker wakes up and preempts
idle then it will be most likely a single shot noop because the stats
were already synced and so it would end up on the oncpu_stat_off anyway.
We just need to teach both vmstat_shepherd and vmstat_update to stop
scheduling the worker if there is nothing to do.

Reported-by: Mike Galbraith <umgwanakikbuti@...il.com>
Signed-off-by: Michal Hocko <mhocko@...e.com>
---
 mm/vmstat.c | 64 ++++++++++++++++++++++++++++++++++++++++++-------------------
 1 file changed, 44 insertions(+), 20 deletions(-)

diff --git a/mm/vmstat.c b/mm/vmstat.c
index 40b2c74ddf16..eb30bf45bd55 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -1396,10 +1396,15 @@ static void vmstat_update(struct work_struct *w)
 		 * Counters were updated so we expect more updates
 		 * to occur in the future. Keep on running the
 		 * update worker thread.
+		 * If we were marked on cpu_stat_off clear the flag
+		 * so that vmstat_shepherd doesn't schedule us again.
 		 */
-		queue_delayed_work_on(smp_processor_id(), vmstat_wq,
-			this_cpu_ptr(&vmstat_work),
-			round_jiffies_relative(sysctl_stat_interval));
+		if (!cpumask_test_and_clear_cpu(smp_processor_id(),
+						cpu_stat_off)) {
+			queue_delayed_work_on(smp_processor_id(), vmstat_wq,
+				this_cpu_ptr(&vmstat_work),
+				round_jiffies_relative(sysctl_stat_interval));
+		}
 	} else {
 		/*
 		 * We did not update any counters so the app may be in
@@ -1417,18 +1422,6 @@ static void vmstat_update(struct work_struct *w)
  * until the diffs stay at zero. The function is used by NOHZ and can only be
  * invoked when tick processing is not active.
  */
-void quiet_vmstat(void)
-{
-	if (system_state != SYSTEM_RUNNING)
-		return;
-
-	do {
-		if (!cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
-			cancel_delayed_work(this_cpu_ptr(&vmstat_work));
-
-	} while (refresh_cpu_vm_stats(false));
-}
-
 /*
  * Check if the diffs for a certain cpu indicate that
  * an update is needed.
@@ -1452,6 +1445,30 @@ static bool need_update(int cpu)
 	return false;
 }
 
+void quiet_vmstat(void)
+{
+	if (system_state != SYSTEM_RUNNING)
+		return;
+
+	/*
+	 * If we are already in hands of the shepherd then there
+	 * is nothing for us to do here.
+	 */
+	if (cpumask_test_and_set_cpu(smp_processor_id(), cpu_stat_off))
+		return;
+
+	if (!need_update(smp_processor_id()))
+		return;
+
+	/*
+	 * Just refresh counters and do not care about the pending delayed
+	 * vmstat_update. It doesn't fire that often to matter and canceling
+	 * it would be too expensive from this path.
+	 * vmstat_shepherd will take care about that for us.
+	 */
+	refresh_cpu_vm_stats(false);
+}
+
 
 /*
  * Shepherd worker thread that checks the
@@ -1470,11 +1487,18 @@ static void vmstat_shepherd(struct work_struct *w)
 	get_online_cpus();
 	/* Check processors whose vmstat worker threads have been disabled */
 	for_each_cpu(cpu, cpu_stat_off)
-		if (need_update(cpu) &&
-			cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
-
-			queue_delayed_work_on(cpu, vmstat_wq,
-				&per_cpu(vmstat_work, cpu), 0);
+		if (need_update(cpu)) {
+			if (cpumask_test_and_clear_cpu(cpu, cpu_stat_off))
+				queue_delayed_work_on(cpu, vmstat_wq,
+					&per_cpu(vmstat_work, cpu), 0);
+		} else {
+			/*
+			 * Cancel the work if quiet_vmstat has put this
+			 * cpu on cpu_stat_off because the work item might
+			 * be still scheduled
+			 */
+			cancel_delayed_work(this_cpu_ptr(&vmstat_work));
+		}
 
 	put_online_cpus();
 
-- 
2.7.0.rc3


-- 
Michal Hocko
SUSE Labs

Powered by blists - more mailing lists