Re: [PATCH rcu 13/14] workqueue: Make queue_rcu_work() use call_rcu_flush()

From: Paul E. McKenney
Date: Mon Oct 24 2022 - 13:07:02 EST

On Mon, Oct 24, 2022 at 02:31:15PM +0000, Joel Fernandes wrote:
> On Mon, Oct 24, 2022 at 02:23:39PM +0200, Uladzislau Rezki wrote:
> > > > On Sun, Oct 23, 2022 at 08:36:00PM -0400, Joel Fernandes wrote:
> > > > > On Wed, Oct 19, 2022 at 6:51 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > From: Uladzislau Rezki <urezki@xxxxxxxxx>
> > > > > >
> > > > > > call_rcu() changes to save power will slow down RCU workqueue items
> > > > > > queued via queue_rcu_work(). This may not be an issue, however we cannot
> > > > > > assume that workqueue users are OK with long delays. Use
> > > > > > call_rcu_flush() API instead which reverts to the old behavio
> > > > >
> > > > > On ChromeOS, I can see that queue_rcu_work() is pretty noisy and the
> > > > > batching is much better if we can just keep it as call_rcu() instead
> > > > > of call_rcu_flush().
> > > > >
> > > > > Is there really any reason to keep it as call_rcu_flush() ? If I
> > > > > recall, the real reason Vlad's system was slowing down was because of
> > > > > scsi and the queue_rcu_work() conversion was really a red herring.
> > > >
> > > <snip>
> > > *** drivers/acpi/osl.c:
> > > acpi_os_drop_map_ref[401] queue_rcu_work(system_wq, &map->track.rwork);
> > >
> > > *** drivers/gpu/drm/i915/gt/intel_execlists_submission.c:
> > > virtual_context_destroy[3653] queue_rcu_work(system_wq, &ve->rcu);
> > >
> > > *** fs/aio.c:
> > > free_ioctx_reqs[632] queue_rcu_work(system_wq, &ctx->free_rwork);
> > >
> > > *** fs/fs-writeback.c:
> > > inode_switch_wbs[604] queue_rcu_work(isw_wq, &isw->work);
> > > cleanup_offline_cgwb[676] queue_rcu_work(isw_wq, &isw->work);
> > >
> > > *** include/linux/workqueue.h:
> > > __printf[446] extern bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork);
> > >
> > > *** kernel/cgroup/cgroup.c:
> > > css_release_work_fn[5253] queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
> > > css_create[5384] queue_rcu_work(cgroup_destroy_wq, &css->destroy_rwork);
> > >
> > > *** kernel/rcu/tree.c:
> > > kfree_rcu_monitor[3192] queue_rcu_work(system_wq, &krwp->rcu_work);
> > >
> > > *** net/core/skmsg.c:
> > > sk_psock_drop[852] queue_rcu_work(system_wq, &psock->rwork);
> > >
> > > *** net/sched/act_ct.c:
> > > tcf_ct_flow_table_put[355] queue_rcu_work(act_ct_wq, &ct_ft->rwork);
> > >
> > > *** net/sched/cls_api.c:
> > > tcf_queue_work[225] return queue_rcu_work(tc_filter_wq, rwork);
> > > <snip>
> > > There are 9 users of the queue_rcu_work() functions. I think there can be
> > > a side effect if we keep it as lazy variant. Please note that i have not
> > > checked all those users.
> > >
> > > > There are less than 20 invocations of queue_rcu_work(), so it should
> > > > be possible look through each. The low-risk approach is of course to
> > > > have queue_rcu_work() use call_rcu_flush().
>
> Yes, once I get to a device (tomorrow), I'll look more. Last I checked it was
> kvfree_rcu() -- this was a few weeks/months ago though.
>
> > > > The next approach might be to have a Kconfig option and/or kernel
> > > > boot parameter that allowed a per-system choice.
> > > >
> > > > But it would not hurt to double-check on Android.
> > > >
> > > I did not see such noise but i will come back some data on 5.10 kernel
> > > today.
> > >
> > Home screen swipe:
> > <snip>
> > <...>-15 [003] d..1 202.142205: rcu_batch_start: rcu_preempt CBs=105 bl=10
> > <...>-55 [001] d..1 202.166174: rcu_batch_start: rcu_preempt CBs=135 bl=10
> > <...>-26 [001] d..1 202.402182: rcu_batch_start: rcu_preempt CBs=221 bl=10
> > rcuop/3-40 [003] d..1 202.650323: rcu_batch_start: rcu_preempt CBs=213 bl=10
> > rcuop/3-40 [000] d..1 203.210537: rcu_batch_start: rcu_preempt CBs=90 bl=10
> > rcuop/5-55 [001] d..1 204.675671: rcu_batch_start: rcu_preempt CBs=14 bl=10
> > rcuop/2-33 [002] d..1 205.162229: rcu_batch_start: rcu_preempt CBs=649 bl=10
> > rcuop/3-40 [000] d..1 205.418214: rcu_batch_start: rcu_preempt CBs=291 bl=10
> > rcuop/3-40 [003] d..1 206.134204: rcu_batch_start: rcu_preempt CBs=174 bl=10
> > rcuop/0-15 [003] d..1 206.726311: rcu_batch_start: rcu_preempt CBs=738 bl=10
> > rcuop/1-26 [001] d..1 206.814168: rcu_batch_start: rcu_preempt CBs=865 bl=10
> > rcuop/3-40 [003] d..1 207.278178: rcu_batch_start: rcu_preempt CBs=287 bl=10
> > rcuop/1-26 [001] d..1 208.826279: rcu_batch_start: rcu_preempt CBs=506 bl=10
> > <snip>
>
> This looks fine to me, but..
>
> > An app launch:
> > <snip>
> > rcuop/3-40 [002] d..1 322.118620: rcu_batch_start: rcu_preempt CBs=99 bl=10
> > rcuop/4-48 [005] dn.1 322.454052: rcu_batch_start: rcu_preempt CBs=270 bl=10
> > rcuop/5-55 [005] d..1 322.454109: rcu_batch_start: rcu_preempt CBs=91 bl=10
> > rcuop/5-55 [007] d..1 322.470054: rcu_batch_start: rcu_preempt CBs=106 bl=10
> > rcuop/6-62 [005] d..1 322.482120: rcu_batch_start: rcu_preempt CBs=231 bl=10
> > rcuop/4-48 [001] d..1 322.494150: rcu_batch_start: rcu_preempt CBs=227 bl=10
> > <...>-69 [002] d..1 322.502442: rcu_batch_start: rcu_preempt CBs=3350 bl=26
> > rcuop/1-26 [001] d..1 322.646099: rcu_batch_start: rcu_preempt CBs=1685 bl=13
> > rcuop/2-33 [001] d..1 322.670071: rcu_batch_start: rcu_preempt CBs=438 bl=10
> > rcuop/1-26 [001] d..1 322.674120: rcu_batch_start: rcu_preempt CBs=18 bl=10
> > rcuop/2-33 [003] d..1 322.690152: rcu_batch_start: rcu_preempt CBs=10 bl=10
> > rcuop/1-26 [002] d..1 322.698104: rcu_batch_start: rcu_preempt CBs=10 bl=10
> > rcuop/3-40 [002] d..1 322.706167: rcu_batch_start: rcu_preempt CBs=313 bl=10
> > rcuop/2-33 [003] d..1 322.710075: rcu_batch_start: rcu_preempt CBs=15 bl=10
>
> The above does not look fine to me (^^^) from a Lazu-RCU PoV. Here, RCU
> callbacks are being invoked every 10-20ms. The batching I seek is of the
> order of seconds, when the system is relatively idle.
>
> Why is Lazy-RCU not in effect for app launch? IOW, which callback is causing
> Lazy-RCU to not be lazy here? Could it be queue_rcu_work()? Whenever a
> non-lazy callback is queued, all the lazy ones are 'promoted' to non-lazy.
> That's why I am asking. Even if you queue one non-lazy callback at a high
> enough frequency, the lazy ones will no longer giving you batching or the
> benefits of laziness.
>
> > rcuop/3-40 [002] d..1 322.742137: rcu_batch_start: rcu_preempt CBs=13 bl=10
> > rcuop/5-55 [000] d..1 322.754270: rcu_batch_start: rcu_preempt CBs=157 bl=10
> > rcuop/3-40 [000] d..1 322.762182: rcu_batch_start: rcu_preempt CBs=17 bl=10
> > rcuop/2-33 [003] d..1 322.774088: rcu_batch_start: rcu_preempt CBs=38 bl=10
> > rcuop/3-40 [000] d..1 322.778131: rcu_batch_start: rcu_preempt CBs=23 bl=10
> > rcuop/1-26 [002] d..1 322.790105: rcu_batch_start: rcu_preempt CBs=33 bl=10
> > rcuop/4-48 [001] d..1 322.798074: rcu_batch_start: rcu_preempt CBs=340 bl=10
> > rcuop/2-33 [002] d..1 322.806158: rcu_batch_start: rcu_preempt CBs=18 bl=10
> > rcuop/1-26 [002] d..1 322.814057: rcu_batch_start: rcu_preempt CBs=18 bl=10
> > rcuop/0-15 [001] d..1 322.822476: rcu_batch_start: rcu_preempt CBs=333 bl=10
> > rcuop/4-48 [003] d..1 322.830102: rcu_batch_start: rcu_preempt CBs=11 bl=10
> > rcuop/2-33 [001] d..1 322.846109: rcu_batch_start: rcu_preempt CBs=80 bl=10
> > rcuop/3-40 [001] d..1 322.854162: rcu_batch_start: rcu_preempt CBs=145 bl=10
> > rcuop/4-48 [003] d..1 322.874129: rcu_batch_start: rcu_preempt CBs=21 bl=10
> > rcuop/3-40 [001] d..1 322.878149: rcu_batch_start: rcu_preempt CBs=43 bl=10
> > rcuop/3-40 [001] d..1 322.906273: rcu_batch_start: rcu_preempt CBs=10 bl=10
> > rcuop/4-48 [001] d..1 322.918201: rcu_batch_start: rcu_preempt CBs=23 bl=10
> > rcuop/2-33 [001] d..1 322.926212: rcu_batch_start: rcu_preempt CBs=86 bl=10
> > rcuop/2-33 [001] d..1 322.946251: rcu_batch_start: rcu_preempt CBs=12 bl=10
> > rcuop/5-55 [003] d..1 322.954482: rcu_batch_start: rcu_preempt CBs=70 bl=10
> > rcuop/2-33 [003] d..1 322.978146: rcu_batch_start: rcu_preempt CBs=20 bl=10
> > rcuop/1-26 [002] d..1 323.014290: rcu_batch_start: rcu_preempt CBs=230 bl=10
> > rcuop/4-48 [001] d..1 323.026119: rcu_batch_start: rcu_preempt CBs=73 bl=10
> > rcuop/5-55 [003] d..1 323.026175: rcu_batch_start: rcu_preempt CBs=94 bl=10
> > rcuop/3-40 [001] d..1 323.035310: rcu_batch_start: rcu_preempt CBs=70 bl=10
> > rcuop/0-15 [001] d..1 323.046231: rcu_batch_start: rcu_preempt CBs=165 bl=10
> > rcuop/6-62 [005] d..1 323.066132: rcu_batch_start: rcu_preempt CBs=179 bl=10
> > rcuop/1-26 [002] d..1 323.174202: rcu_batch_start: rcu_preempt CBs=61 bl=10
> > rcuop/2-33 [003] d..1 323.190203: rcu_batch_start: rcu_preempt CBs=80 bl=10
> > rcuop/3-40 [003] d..1 323.206210: rcu_batch_start: rcu_preempt CBs=84 bl=10
> > rcuop/2-33 [003] d..1 323.226880: rcu_batch_start: rcu_preempt CBs=5 bl=10
>
> And for several seconds you have the same thing going ^^^.
>
> > <snip>
> >
> > It is on Android with 5.10 kernel running. I do not see that queue_rcu_work() makes
> > some noise.
>
> Your rcu_batch_start tracepoint output above does not really reveal much
> information about which callbacks are lazy and which are not.
> rcu_invoke_callback is better in the sense at least you have the name of the
> callback and can take a guess.
>
> > Joel Could you please post your batch_start trace point output where you see the noise?
>
> Sure, I'll do that once I get to a device.

You guys might need to agree on the definition of "good" here. Or maybe
understand the differences in your respective platforms' definitions of
"good". ;-)

Thanx, Paul