netdev - Re: [PATCH net-next v2 1/1] driver: veth: Refine the statistics codes of veth driver

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: Windows password security audit tool. GUI, reports in PDF.

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

Message-ID: <1478183516.7065.440.camel@edumazet-glaptop3.roam.corp.google.com>
Date:   Thu, 03 Nov 2016 07:31:56 -0700
From:   Eric Dumazet <eric.dumazet@...il.com>
To:     Gao Feng <fgao@...ai8.com>
Cc:     "David S. Miller" <davem@...emloft.net>,
        Cong Wang <cwang@...pensource.com>,
        Vijay Pandurangan <vijayp@...ayp.ca>,
        Evan Jones <ej@...njones.ca>, pabeni@...hat.com,
        Linux Kernel Network Developers <netdev@...r.kernel.org>
Subject: Re: [PATCH net-next v2 1/1] driver: veth: Refine the statistics
 codes of veth driver

On Thu, 2016-11-03 at 21:39 +0800, Gao Feng wrote:
> Hi Eric,
> 
> On Thu, Nov 3, 2016 at 9:30 PM, Eric Dumazet <eric.dumazet@...il.com> wrote:
> > On Thu, 2016-11-03 at 21:03 +0800, fgao@...ai8.com wrote:
> >> From: Gao Feng <fgao@...ai8.com>
> >>
> >> The dropped count of veth is located in struct veth_priv, but other
> >> statistics like packets and bytes are in another struct pcpu_vstats.
> >> Now keep these three counters in the same struct.
> >>
> >> Signed-off-by: Gao Feng <fgao@...ai8.com>
> >> ---
> >>  v2: Use right "peer" instead of "dev";
> >>  v1: Initial version
> >
> > May I ask : Why ?
> 
> Just because I think statistics should be in the same struct.

That is not a good reason then.

> 
> >
> > We did that because there was no point making per-cpu requirements
> > bigger, for a counter that is hardly ever updated.
> >
> > Do you have a real case where performance dropping packets in a driver
> > is needed ?
> 
> No, I haven't met the performance issue now.

OK then kill this patch.

> 
> >
> > At some point we will have to stop dumb percpu explosion, when we have
> > 128+ cores per host. Folding all these percpu counters is taking a lot
> > of time too.
> >
> >
> >
> Ok, I get it. It is designed specially to put the dropped counter as
> atomic counter, not percpu.
> But I have one question that when put the counters as percpu, and when not?

Because the regular fast path needs to be fast ?

Try to _use_ veth without these percpu stats and be prepared to be
shocked.