Re: TCP connection hang problem with 2.6.16.16, e1000.

[Ben Greear <greearb@candelatech.com>]

Ben Greear wrote:

> I haven't seen this problem on 2.6.13, so I'm now starting a manual bisect
> to see if I can narrow down where the problem appeared.

Turns out, I can reproduce it in 2.6.13, and 2.6.9.  I haven't tried anything
older.

I also tried to reproduce it using a simpler traffic generation tool, but could
not reproduce the problem with it.  That points to something wierd that my application
is doing, but I can't imagine what user-space could do to screw up a TCP connection
like this.

In all cases, there is a lot of data in the send-queue, but for whatever
reason, the connection will not make progress.  To user-space, it appears
that poll returns neither readable nor writable for the sockets.

I notice that if I increase the send-buffer-size while the connection is
in the hung state, my app will quickly fill the larger send buffer, but
still receives nothing new.

Starting a new connection on the same interfaces works for a few seconds
and then hangs as well, so the NICs can pass traffic.

Here is output from /proc/net/tcp and netstat from the 2.6.16.16 kernel.

netstat info:

tcp        0 5635368 172.1.5.169:33058       172.1.5.168:33057       ESTABLISHED
tcp        0 5987504 172.1.5.168:33057       172.1.5.169:33058       ESTABLISHED

/proc/net/tcp:

   20: A90501AC:8122 A80501AC:8121 01 0055FD28:00000000 01:00001A9F 0000000A
0        0 21309 2 f36d8580 120000 40 0 1 58
   21: A80501AC:8121 A90501AC:8122 01 005B5CB0:00000000 01:00001C9D 0000000A
0        0 21226 3 ef7bfa80 120000 40 0 1 35


-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com