kern/140597: Lost Retransmission Detection

[Richard Scheffenegger]

>Number:         140597
>Category:       kern
>Synopsis:       Lost Retransmission Detection
>Confidential:   no
>Severity:       serious
>Priority:       low
>Responsible:    freebsd-bugs
>State:          open
>Quarter:        
>Keywords:       
>Date-Required:
>Class:          sw-bug
>Submitter-Id:   current-users
>Arrival-Date:   Mon Nov 16 07:20:00 UTC 2009
>Closed-Date:
>Last-Modified:
>Originator:     Richard Scheffenegger
>Release:        8.0-RC2
>Organization:
NetApp
>Environment:
FreeBSD rsFreeBSD.vie.demo 8.0-RC2 FreeBSD 8.0-RC2 #16: Sat Nov 14 22:25:28 CET 2009     root at rsFreeBSD.vie.demo:/usr/obj/usr/src/sys/GENERIC  i386

>Description:
SACK TCP in FreeBSD does not do Lost Retransmission Detection.

Unlike linux, that functionality is not implemented, leading to a RTO when a single segment is lost twice. This reduces Goodput considerably during phases 
of higher congestion.

See http://ubinet.yonsei.ac.kr/v2/publication/hpmn_papaers/ij/29.pdf
And http://projects.itri.aist.go.jp/gnet/pfldnet08kodama.ppt


>How-To-Repeat:
Drop the same segment of a TCP Stram twice, when cwnd is open enough (3+ segments).
>Fix:
Simple fix is rather easy (doesn't work perfectly under all circumstances,
but doesn't need new state variable and runs in constant time):

When sending the last segment in a sackhole, store snd.max in rxmit.
During SACK Ack processing, check the value of rxmit from the lowest 
hole against snd.fack; if snd.fack <= rxmit of that hole, set rxmit to 
start, and sackhint.nexthole == sackhole[0]. 

Care must be taken during hole-splitting (if rxmit > end of lower hole, set
rxmit to snd.max), and all checks of sackhole[].rxmit need to deal with
the signal rxmit > end.

See http://www.ietf.org/mail-archive/web/tcpm/current/msg05003.html ff.

>Release-Note:
>Audit-Trail:
>Unformatted: