svn commit: r235053 - stable/8/sys/netinet

[Gleb Smirnoff]

On Sat, May 05, 2012 at 10:05:13AM +0000, Gleb Smirnoff wrote:
T> Author: glebius
T> Date: Sat May  5 10:05:13 2012
T> New Revision: 235053
T> URL: http://svn.freebsd.org/changeset/base/235053
T> 
T> Log:
T>   Merge 234342 from head:
T>     When we receive an ICMP unreach need fragmentation datagram, we take
T>     proposed MTU value from it and update the TCP host cache. Then
T>     tcp_mss_update() is called on the corresponding tcpcb. It finds the
T>     just allocated entry in the TCP host cache and updates MSS on the
T>     tcpcb. And then we do a fast retransmit of what we have in the tcp
T>     send buffer.
T>   
T>     This sequence gets broken if the TCP host cache is exausted. In this
T>     case allocation fails, and later called tcp_mss_update() finds nothing
T>     in cache. The fast retransmit is done with not reduced MSS and is
T>     immidiately replied by remote host with new ICMP datagrams and the
T>     cycle repeats. This ping-pong can go up to wirespeed.
T>   
T>     To fix this:
T>     - tcp_mss_update() gets new parameter - mtuoffer, that is like
T>       offer, but needs to have min_protoh subtracted.
T>     - tcp_mtudisc() as notification method renamed to tcp_mtudisc_notify().
T>     - tcp_mtudisc() now accepts not a useless error argument, but proposed
T>       MTU value, that is passed to tcp_mss_update() as mtuoffer.
T>   
T>     Reported by:  az
T>     Reported by:  Andrey Zonov <andrey zonov.org>
T>     Reviewed by:  andre (previous version of patch)
T>   
T>   Tested by:  Andrey Zonov <andrey zonov.org>

Sorry, should have been: tested by az.

Andrej Zverev did extensive testing on stable/8.

-- 
Totus tuus, Glebius.