[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <YpxzB3/HRN/EEHa8@pc638.lan>
Date: Sun, 5 Jun 2022 11:10:31 +0200
From: Uladzislau Rezki <urezki@...il.com>
To: "Paul E. McKenney" <paulmck@...nel.org>
Cc: "Uladzislau Rezki (Sony)" <urezki@...il.com>,
LKML <linux-kernel@...r.kernel.org>, RCU <rcu@...r.kernel.org>,
Frederic Weisbecker <frederic@...nel.org>,
Neeraj Upadhyay <neeraj.iitr10@...il.com>,
Joel Fernandes <joel@...lfernandes.org>,
Oleksiy Avramchenko <oleksiy.avramchenko@...y.com>
Subject: Re: [PATCH 2/2] rcu/kvfree: Introduce KFREE_DRAIN_JIFFIES_[MAX/MIN]
interval
> On Thu, Jun 02, 2022 at 10:06:44AM +0200, Uladzislau Rezki (Sony) wrote:
> > Currently the monitor work is scheduled with a fixed interval that
> > is HZ/20 or each 50 milliseconds. The drawback of such approach is
> > a low utilization of page slot in some scenarios. The page can store
> > up to 512 records. For example on Android system it can look like:
> >
> > <snip>
> > kworker/3:0-13872 [003] .... 11286.007048: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000026522604 nr_records=1
> > kworker/3:0-13872 [003] .... 11286.015638: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000095ed6fca nr_records=2
> > kworker/1:2-20434 [001] .... 11286.051230: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044872ffd nr_records=1
> > kworker/1:2-20434 [001] .... 11286.059322: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000026522604 nr_records=2
> > kworker/0:1-20052 [000] .... 11286.095295: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044872ffd nr_records=2
> > kworker/0:1-20052 [000] .... 11286.103418: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000cbcf05db nr_records=1
> > kworker/2:3-14372 [002] .... 11286.135155: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000095ed6fca nr_records=2
> > kworker/2:3-14372 [002] .... 11286.135198: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000044872ffd nr_records=1
> > kworker/1:2-20434 [001] .... 11286.155377: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000cbcf05db nr_records=5
> > kworker/2:3-14372 [002] .... 11286.167181: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000026522604 nr_records=5
> > kworker/1:2-20434 [001] .... 11286.179202: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x000000008ef95e14 nr_records=1
> > kworker/2:3-14372 [002] .... 11286.187398: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000c597d297 nr_records=6
> > kworker/3:0-13872 [003] .... 11286.187445: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000050bf92e2 nr_records=3
> > kworker/1:2-20434 [001] .... 11286.198975: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x00000000cbcf05db nr_records=4
> > kworker/1:2-20434 [001] .... 11286.207203: rcu_invoke_kfree_bulk_callback: rcu_preempt bulk=0x0000000095ed6fca nr_records=4
> > <snip>
> >
> > where a page only carries few records to reclaim a memory. In order to
> > improve batching and make utilization more efficient the patch introduces
> > a drain interval that can be set either to KFREE_DRAIN_JIFFIES_MAX or
> > KFREE_DRAIN_JIFFIES_MIN. It is adjusted if a flood is detected, in this
> > case a memory reclaim occurs more often whereas in mostly idle cases the
> > interval is set to its maximum timeout that improves the utilization of
> > page slots.
> >
> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@...il.com>
>
> That does look like a problem well worth solving!
>
Agree, better ideas make better final solution :)
>
> But I am missing one thing. If we are having a callback flood, why do we
> need a shorter timeout?
>
To offload faster, because otherwise we run into classical issue, it is a low
memory condition state resulting in OOM.
>
> Wouldn't a check on the number of blocks queued be simpler, more direct,
> and provide faster response to the start of a callback flood?
>
I rely on krcp->count because not always we can store the pointer in the page
slots. We can not allocate a page in the caller context thus we use page-cache
worker that fills the cache in normal context. While it populates the cache,
pointers temporary are queued to the linked-list.
Any thoughts?
--
Uladzislau Rezki
Powered by blists - more mailing lists