[Bug 252913] [tcp] Using RACK leaks mbufs

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=252913

--- Comment #23 from Michael Tuexen <tuexen at freebsd.org> ---
OK. In the tracefile you provided, shows that:

* the TCP connection does not use time stamps
* the peer is not available anymore, since there are every 3 seconds 3 ICMP
messages indicating that the host is not reachable.
* the RACK stack sends out the same TCP segments once per ms.

The question is why the stack gets into this condition.

Two questions:

1. When the stack has been in such a condition, can you provide the output of
sysctl net.inet.tcp.rack Maybe the stats counters there provide some hints.

2. Looking at the graphic you provided seems to indicate that the traffic
volume increases when you enable RACK. I don't see a reason for a substantial
increase. So I guess that some connections get into the state we are interested
soon after you enable RACK. So could you start without enabling RACK, start
capturing traffic with tcpdump, enable RACK, ensure that some connection hit
the bug, and stop the capturing. It would be great, if I could get the .pcap
file for one of the connections getting in the buggy state. Based on the .pcap
file I can write a packetdrill script to try to reproduce what is happening.

-- 
You are receiving this mail because:
You are on the CC list for the bug.