| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <CAGsJ_4xcRFcDMpuC7vrpHe=aRbDpAnRd1F64aqh2EEcNgmZxCg@mail.gmail.com>
Date: Sun, 3 Oct 2021 20:34:49 +1300
From: Barry Song <21cnbao@...il.com>
To: Mike Galbraith <efault@....de>
Cc: Mel Gorman <mgorman@...hsingularity.net>,
Peter Zijlstra <peterz@...radead.org>,
Ingo Molnar <mingo@...nel.org>,
Vincent Guittot <vincent.guittot@...aro.org>,
Valentin Schneider <valentin.schneider@....com>,
Aubrey Li <aubrey.li@...ux.intel.com>,
Barry Song <song.bao.hua@...ilicon.com>,
Srikar Dronamraju <srikar@...ux.vnet.ibm.com>,
LKML <linux-kernel@...r.kernel.org>
Subject: Re: wakeup_affine_weight() is b0rked - was Re: [PATCH 2/2]
sched/fair: Scale wakeup granularity relative to nr_running
On Sun, Oct 3, 2021 at 4:11 PM Mike Galbraith <efault@....de> wrote:
>
> On Wed, 2021-09-22 at 07:22 +0200, Mike Galbraith wrote:
> > On Tue, 2021-09-21 at 11:36 +0100, Mel Gorman wrote:
> > > On Tue, Sep 21, 2021 at 05:52:32AM +0200, Mike Galbraith wrote:
> > >
> > >
> > > > Preemption does rapidly run into diminishing return as load climbs for
> > > > a lot of loads, but as you know, it's a rather sticky wicket because
> > > > even when over-committed, preventing light control threads from slicing
> > > > through (what can be a load's own work crew of) hogs can seriously
> > > > injure performance.
> > > >
> > >
> > > Turning this into a classic Rob Peter To Pay Paul problem. We don't know
> > > if there is a light control thread that needs to run or not that affects
> > > overall performance. It all depends on whether that control thread needs
> > > to make progress for the overall workload or whether there are a mix of
> > > workloads resulting in overloading.
> >
> > WRT overload, and our good buddies Peter and Paul :) I added...
> > if (gran >= sysctl_sched_latency >> 1)
> > trace_printk("runnable:%d preempt disabled\n",cfs_rq->nr_running);
> > ...to watch, and met the below when I.. logged in.
> >
> > homer:..debug/tracing # tail -20 trace
> > X-2229 [002] d..5. 60.690322: wakeup_gran: runnable:9 preempt disabled
> > X-2229 [002] d..5. 60.690325: wakeup_gran: runnable:10 preempt disabled
> > X-2229 [002] d..5. 60.690330: wakeup_gran: runnable:11 preempt disabled
> > X-2229 [002] d..5. 60.690363: wakeup_gran: runnable:13 preempt disabled
> > X-2229 [002] d..5. 60.690377: wakeup_gran: runnable:14 preempt disabled
> > X-2229 [002] d..5. 60.690390: wakeup_gran: runnable:15 preempt disabled
> > X-2229 [002] d..5. 60.690404: wakeup_gran: runnable:16 preempt disabled
> > X-2229 [002] d..5. 60.690425: wakeup_gran: runnable:9 preempt disabled
> > ksmserver-2694 [003] d..3. 60.690432: wakeup_gran: runnable:6 preempt disabled
> > ksmserver-2694 [003] d..3. 60.690436: wakeup_gran: runnable:7 preempt disabled
> > X-2229 [002] d..5. 60.690451: wakeup_gran: runnable:6 preempt disabled
> > X-2229 [002] d..5. 60.690465: wakeup_gran: runnable:7 preempt disabled
> > kmix-2736 [000] d..3. 60.690491: wakeup_gran: runnable:6 preempt disabled
> > X-2229 [004] d..5. 92.889635: wakeup_gran: runnable:6 preempt disabled
> > X-2229 [004] d..5. 92.889675: wakeup_gran: runnable:6 preempt disabled
> > X-2229 [004] d..5. 92.889863: wakeup_gran: runnable:6 preempt disabled
> > X-2229 [004] d..5. 92.889944: wakeup_gran: runnable:6 preempt disabled
> > X-2229 [004] d..5. 92.889957: wakeup_gran: runnable:7 preempt disabled
> > X-2229 [004] d..5. 92.889968: wakeup_gran: runnable:8 preempt disabled
> > QXcbEventQueue-2740 [000] d..4. 92.890025: wakeup_gran: runnable:6 preempt disabled
> > homer:..debug/tracing
> >
> > Watching 'while sleep 1; do clear;tail trace; done' with nothing but a
> > kbuild running is like watching top. There's enough stacking during
> > routine use of my desktop box that it runs into the tick granularity
> > wall pretty much continuously, so 'overload' may want redefining.
>
> I looked into that crazy stacking depth...
>
> static int
> wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
> int this_cpu, int prev_cpu, int sync)
> {
> s64 this_eff_load, prev_eff_load;
> unsigned long task_load;
>
> this_eff_load = cpu_load(cpu_rq(this_cpu));
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^ the butler didit!
>
> That's pretty darn busted as it sits. Between load updates, X, or any
> other waker of many, can stack wakees to a ludicrous depth. Tracing
> kbuild vs firefox playing a youtube clip, I watched X stack 20 of the
> zillion firefox minions while their previous CPUs all had 1 lousy task
> running but a cpu_load() higher than the cpu_load() of X's CPU. Most
> of those prev_cpus were where X had left them when it migrated. Each
> and every crazy depth migration was wake_affine_weight() deciding we
> should pull based on crappy data. As instantaneous load on the waker
> CPU blew through the roof in my trace snapshot, its cpu_load() did
> finally budge.. a tiny bit.. downward. No idea where the stack would
> have topped out, my tracing_off() limit was 20.
Mike, not quite sure I caught your point. It seems you mean x wakes up
many firefoxes within a short period, so it pulls them to the CPU where x
is running. Technically those pulling should increase cpu_load of x' CPU.
But due to some reason, the cpu_load is not increased in time on x' CPU,
So this makes a lot of firefoxes piled on x' CPU, but at that time, the load
of the cpu which firefox was running on is still larger than x' cpu with a lot
of firefoxes?
I am wondering if this should be the responsibility of wake_wide()?
>
> Hohum, my box grew a WA_INST companion to SIS_MIN_LAT cache cold task
> distribulator feature ;-) Not particularly lovely, but it knocks over
> the leaning tower of minions.
>
> -Mike
Thanks
barry
Powered by blists - more mailing lists