netdev - Re: [PATCH net] ipv6: Reflect MTU changes on PMTU of exceptions for MTU-less routes

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]

Date:   Fri, 2 Mar 2018 10:54:36 -0800
From:   Maciej Żenczykowski <zenczykowski@...il.com>
To:     Stefano Brivio <sbrivio@...hat.com>
Cc:     "David S . Miller" <davem@...emloft.net>,
        Wei Wang <weiwan@...gle.com>,
        Hideaki YOSHIFUJI <yoshfuji@...ux-ipv6.org>,
        Linux NetDev <netdev@...r.kernel.org>
Subject: Re: [PATCH net] ipv6: Reflect MTU changes on PMTU of exceptions for
 MTU-less routes

Conceptually this is right.

And I'm 100% fine with dev mtu change triggering pmtu decrease.

I'm not so sold on the pmtu increase.

PMTUD is one of those things that never ever works right in practice.
There's too many icmp blackholes, rate limits, overloaded management
cpus in switches,
misconfigurations, missing tcp mss clamps, icmps routed differently
then the flows due to ecmp hashing, middle boxes that don't affect the
icmp but change the tcp stream, etc.

In particular there's a lot of routing hardware that can handle
gigabits or terabits of traffic, but can generate only 10s-100s of
packet too big messages per second (ie. a tiny fraction of line rate
pps).  Worse yet, under overload it often falls back to simply
dropping and generating no icmp errors.

I spend a significant fraction of my time making sure we never rely on PMTUD.

Debugging MTU related blackholes is a constant bane of my existence.

[btw. we're considering adding a hack to always fragment UDP to
min(1280, dev/route/path mtu)...]

Basically: lower is always better because it's more likely to work...