netdev - Re: [net-next 3/3] tcp: keep tcp_collapse controllable even after processing starts

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: <1507995970.31614.59.camel@edumazet-glaptop3.roam.corp.google.com>
Date:   Sat, 14 Oct 2017 08:46:10 -0700
From:   Eric Dumazet <eric.dumazet@...il.com>
To:     Koichiro Den <den@...ipeden.com>
Cc:     netdev@...r.kernel.org, davem@...emloft.net, kuznet@....inr.ac.ru,
        yoshfuji@...ux-ipv6.org
Subject: Re: [net-next 3/3] tcp: keep tcp_collapse controllable even after
 processing starts

On Sat, 2017-10-14 at 16:27 +0900, Koichiro Den wrote:
> Combining actual collapsing with reasoning for deciding the starting
> point, we can apply its logic in a consistent manner such that we can
> avoid costly yet not much useful collapsing. When collapsing to be
> triggered, it's not rare that most of the skbs in the receive or ooo
> queue are large ones without much metadata overhead. This also
> simplifies code and makes it easier to apply logic in a fair manner.
> 
> Subtle subsidiary changes included:
> - When the end_seq of the skb we are trying to collapse was larger than
>   the 'end' argument provided, we would end up copying to the 'end'
>   even though we couldn't collapse the original one. Current users of
>   tcp_collapse does not require such reserves so redefines it as the
>   point over which skbs whose seq passes guranteed not to be collapsed.
> - Naturally tcp_collapse_ofo_queue shapes up and we no longer need
>   'tail' argument.


I am not inclined to review such a large change, without you providing
actual numbers.

We have a problem in TCP right now, that receiver announces a too big
window, and that is the main reason we trigger collapsing.

I would rather fix the root cause.