linux-kernel - Re: [BUG] skb corruption and kernel panic at forwarding with fragmentation

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: <CANn89iJ=09fL7NZhGGtt8+jgExckHkCFC=eFiNVEMiHoK5oOcA@mail.gmail.com>
Date:	Thu, 7 Jan 2016 14:47:06 -0500
From:	Eric Dumazet <edumazet@...gle.com>
To:	Konstantin Khlebnikov <koct9i@...il.com>
Cc:	Florian Westphal <fw@...len.de>,
	Thadeu Lima de Souza Cascardo <cascardo@...hat.com>,
	Cong Wang <xiyou.wangcong@...il.com>,
	Linux Kernel Network Developers <netdev@...r.kernel.org>,
	David Miller <davem@...emloft.net>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [BUG] skb corruption and kernel panic at forwarding with fragmentation

On Thu, Jan 7, 2016 at 2:35 PM, Konstantin Khlebnikov <koct9i@...il.com> wrote:
> On Thu, Jan 7, 2016 at 3:54 PM, Eric Dumazet <edumazet@...gle.com> wrote:
 want to look at git history to find out why it is done this way.
>>
>> TCP performance is critical for some of us, and doing such trick avoid
>> one cache miss per skb in some critical list traversals.
>
> Right. This way tcp stuff perfectly fits into leftovers of first cache line.
> Then probably it's better to put ipv4/ipv6 cb into second line from
> the beginning.

Then IP forwarding might be slower.

Look, each layer (TCP  , IP, ....) can organize its skb->cb[] as it wants.
Nobody tries to 'make universal room' for IPCB, since only IP layer wants it.

TCP could even find a way in the future to no longer hold a copy of
IPCB in the input skb,
if code is reorganized a bit.

Note that skbs for output path in TCP do not need IPCB at all.

Only when skb leaves TCP and enter IP, skb->cb[] content is scratched.