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;
MIME-Version: 1.0
References: <20220719165743.3409313-1-vschneid@redhat.com> <YthDjPq3CtWc+o0/@slm.duckdns.org>
 <CAJhGHyAso+JH+QMXcfKTyOqq4yBD9Vc19sBbebEXohHe5znL8Q@mail.gmail.com>
 <xhsmhv8rqfu6w.mognet@vschneid.remote.csb> <YtuEJLDkO/lATYeb@slm.duckdns.org>
 <xhsmh8rohfq6m.mognet@vschneid.remote.csb> <YuAkroXHF+Zg45KU@slm.duckdns.org>
 <xhsmhmtcvehmx.mognet@vschneid.remote.csb> <CAJhGHyA3gRmL=VCq2TAcc+TgBHCAizADJeukNRGmisa29U969g@mail.gmail.com>
In-Reply-To: <CAJhGHyA3gRmL=VCq2TAcc+TgBHCAizADJeukNRGmisa29U969g@mail.gmail.com>
From:   Lai Jiangshan <jiangshanlai@gmail.com>
Date:   Wed, 27 Jul 2022 14:30:07 +0800
Message-ID: <CAJhGHyAEHFbcqzVbBRHoFcRYJ+M9cf87WwV2u=V4=Acrgp-gkQ@mail.gmail.com>
Subject: Re: [RFC PATCH] workqueue: Unbind workers before sending them to exit()
To:     Valentin Schneider <vschneid@redhat.com>
Cc:     Tejun Heo <tj@kernel.org>, LKML <linux-kernel@vger.kernel.org>,
        Peter Zijlstra <peterz@infradead.org>,
        Frederic Weisbecker <frederic@kernel.org>,
        Juri Lelli <juri.lelli@redhat.com>,
        Phil Auld <pauld@redhat.com>,
        Marcelo Tosatti <mtosatti@redhat.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Jul 27, 2022 at 1:38 PM Lai Jiangshan <jiangshanlai@gmail.com> wrote:
>
> On Wed, Jul 27, 2022 at 4:36 AM Valentin Schneider <vschneid@redhat.com> wrote:
> >
> > On 26/07/22 07:30, Tejun Heo wrote:
> > > Hello,
> > >
> > > On Mon, Jul 25, 2022 at 11:21:37AM +0100, Valentin Schneider wrote:
> > >> Hm so my choice of words in the changelog wasn't great - "initial setup"
> > >> can be kernel init, but *also* setup of whatever workload is being deployed
> > >> onto the system.
> > >>
> > >> So you can be having "normal" background activity (I've seen some IRQs end
> > >> up with schedule_work() on isolated CPUs, they're not moved away at boot
> > >> time but rather shortly before launching the latency-sensitive app), some
> > >> preliminary stats collection / setup to make sure the CPU will be quiet
> > >> (e.g. refresh_vm_stats()), and *then* the application starts with
> > >> fresh-but-no-longer-required extra pcpu kworkers assigned to its CPU.
> > >
> > > Ah, I see. I guess we'll need to figure out how to unbind the workers then.
> > >
> >
> > I've been playing with different ways to unbind & wake the workers in a
> > sleepable context, but so far I haven't been happy with any of my
> > experiments.
>
>
> I'm writing code to handle the problems of cpu affinity and prematurely
> waking up of newly created worker.
>
> This work of unbinding the dying worker is also on the list.
> I haven't figured out a good solution.
>
> I was planning to add set_cpus_allowed_ptr_off_rq() which only set
> cpumasks to the task only if it is sleeping and returns -EBUSY otherwise.
> And it is ensured and documented as being usable in an atomic context
> and it is recommended to be used for dying tasks only.
>
> I can't really ensure it would be implemented as I'm expecting since
> it touches scheduler code.
>
> I'd better back off.
>
> >
> > What hasn't changed much between my attempts is transferring to-be-destroyed
> > kworkers from their pool->idle_list to a reaper_list which is walked by
> > *something* that does unbind+wakeup. AFAIA as long as the kworker is off
> > the pool->idle_list we can play with it (i.e. unbind+wake) off the
> > pool->lock.
> >
> > It's the *something* that's annoying to get right, I don't want it to be
> > overly complicated given most users are probably not impacted by what I'm
> > trying to fix, but I'm getting the feeling it should still be a per-pool
> > kthread. I toyed with a single reaper kthread but a central synchronization
> > for all the pools feels like a stupid overhead.
>
> I think fixing it in the workqueue.c is complicated.
>
> Nevertheless, I will also try to fix it inside workqueue only to see
> what will come up.

I'm going to kind of revert 3347fc9f36e7 ("workqueue: destroy worker
directly in the idle timeout handler"), so that we can have a sleepable
destroy_worker().

>
> >
> > If any of that sounds ludicrous please shout, otherwise I'm going to keep
> > tinkering :)
> >
> > > Thanks.
> > >
> > > --
> > > tejun
> >