linux-kernel - Re: [PATCH] lib/flex_proportions.c: Use abs() when percpu

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

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20210518103928.GC28667@quack2.suse.cz>
Date:   Tue, 18 May 2021 12:39:28 +0200
From:   Jan Kara <jack@...e.cz>
To:     chi wu <wuchi.zero@...il.com>
Cc:     Jan Kara <jack@...e.cz>, Andrew Morton <akpm@...ux-foundation.org>,
        linux-kernel@...r.kernel.org, tan.hu@....com.cn
Subject: Re: [PATCH] lib/flex_proportions.c: Use abs() when percpu_counter is
 negative.

On Tue 18-05-21 18:22:05, chi wu wrote:
> Jan Kara <jack@...e.cz> 于2021年5月18日周二 下午4:59写道：
> >
> > On Tue 18-05-21 11:42:53, chi wu wrote:
> > > Chi Wu <wuchi.zero@...il.com> 于2021年5月17日周一 下午11:53写道：
> > > >
> > > > The value of percpu_counter_read() may become negative after
> > > > running percpu_counter_sum() in fprop_reflect_period_percpu().
> > > > The value of variable 'num' will be zero in fprop_fraction_percpu()
> > > > when using percpu_counter_read_positive(), but if using the abs of
> > > > percpu_counter_read() will be close to the correct value.
> > > >
> > >
> > > I realized that I was wrong as follow:
> > > (a) the decay rule is broken, the negative means the difference in
> > > decay here.
> > > (b) as the target event increasing, proportion of the event will
> > > decrease to 0 firstly and then it will increase. The logic is bad.
> > > 1. abs(-50) / abs(100) = 50%       //+50 to 2
> > > 2. abs(0) / abs(150) = 0 %           //+50 to 3
> > > 3. abs(50)/abs(200) = 25%
> > >
> > > Anyway, the percpu_counter_sum() had cost a lost performance,
> > > may be we could get a little benefits from that. So could we add a
> > > variable to stroe the decay value, we will get the value when
> > > percpu_counter_read() is negative?
> >
> > The result of percpu_counter_read() is inherently inexact (but fast! ;). It
> > can be upto number_of_cpus * counter_batch away from the real counter
> > value. But do you observe any practical problems with this inaccuracy on
> > your system? Sure, cache memory won't be split among devices exactly
> > according to writeout proportion but that usually does not matter.
> >
> >                                                                 Honza
> >
> 
> Thanks, Got it.
> Just try to optimize the fuse (with strictlimit feature)performance
> issue: The writing thread will be paused and runs slowly, when the
> proportion of fuse-bdi is 0.
> The issue is normal,and one of reasons is the characteristics of
> percpu_counter batch. Even the pages are writeout, we may be could not
> get the real proportion value due to side effects of counter
> performance. It's just a slight disappointment.

Well, you can tune 'min_ratio' of the fuse-bdi to avoid problems with these
near-zero states... To always give the bdi some breathing room for ramping
up.

								Honza

-- 
Jan Kara <jack@...e.com>
SUSE Labs, CR