Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 18 Jul 2018 22:39:34 -0700
From:      Kevin Bowling <kevin.bowling@kev009.com>
To:        Lev Serebryakov <lev@freebsd.org>
Cc:        Eugene Grosbein <eugen@grosbein.net>, FreeBSD Net <freebsd-net@freebsd.org>
Subject:   Re: [REGRESSION] Fresh CURRENT consume much more CPU on network traffic (vlans + routing + ipfw with NAT)
Message-ID:  <CAK7dMtCYO2ae7phBeHAk7w7TeB1_LxZVX6SeXCF2jxk9_du-0w@mail.gmail.com>
In-Reply-To: <eff19579-4c09-8530-1eaf-62185315be7d@FreeBSD.org>
References:  <6d8b301b-8ee9-0e8b-80b5-6aced15ca843@FreeBSD.org> <5B4DA0B0.8020302@grosbein.net> <eff19579-4c09-8530-1eaf-62185315be7d@FreeBSD.org>

next in thread | previous in thread | raw e-mail | index | archive | help
This sounds like a known quirk of the Atom CPU architecture and iflib
-- can you try this patch https://reviews.freebsd.org/D16302 it should
help specifically on your hardware.

Regards,
Kevin

On Tue, Jul 17, 2018 at 6:03 AM, Lev Serebryakov <lev@freebsd.org> wrote:
> On 17.07.2018 10:54, Eugene Grosbein wrote:
>
>>>  I have "SOHO" router on Atom D2500 with FreeBSD CURRENT. It runs
>>> CURRENT for very long time (from 11-CURRENT times), and recently it
>>> start to consume much more CPU on same traffic =E2=80=94 to the point w=
hen it
>>> becomes unresponsive in shell (via ssh, not local console).
>>>
>>>  I have rather complex ipfw ruleset, but this ruleset is the same for
>>> many years.
>>>
>>>  Revisions before r333989 worked normally, I never seen any problem wit=
h
>>> shell, no matter how much traffic is processed
>>>
>>>  Revision r334649 with same configuration, same firewall ruleset, etc.,
>>> becomes completely unresponsive under network load (pure transit traffi=
c).
>>>
>>>  when system is unresponsive I see this in `top -SH`
>>>
>>> 100083 root -76 - 0K   272K -       1 291.8H  95.31% kernel{if_io_tqg_1=
}
>>> 100082 root -76 - 0K   272K -       0 297.7H  95.20% kernel{if_io_tqg_0=
}
>>>
>>>  And it is new to me.
>>
>> I'm sure you will get it solved more quick if you perform bisection of r=
evision
>> even though it will take time.
>  I'll try latest version (seems here were a lot of commit to iflib after
> my revision) and after that try yo bisect.
>
> --
> // Lev Serebryakov
>



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CAK7dMtCYO2ae7phBeHAk7w7TeB1_LxZVX6SeXCF2jxk9_du-0w>