TCP/IP connections hang with TCP_RETRANS_COLLAPSE=1 - kernel bug?

TCP/IP connections hang with TCP_RETRANS_COLLAPSE=1 - kernel bug?

[Terence Haddock]

Dear All,

I am having TCP/IP connections "hang", and investigation seems to point to
the tcp_retrans_collapse feature as the culprit. I have the following
network configuration:

Web Server --Switch-> Proxy Server --Internet-> Client

The web server is Windows 2000 running IIS 5.0
The proxy server is Red Hat 7.1, kernel 2.4.5, running Squid 2.4-stable1.
The problem appears on clients running Linux (kernel 2.4.5), Solaris, and
Windows.

The problem is, if I load a page from the web server which is very large
(200k to 500k) and takes a long time to load (20-30 seconds), after some
amount of transmission (usually about 80-100k), the communication stops.
Netstat shows that the web server has sent the remaining data, but the
proxy has stopped sending data.

What I find in the TCP dumps is that at one point the proxy sends two small
(5-byte) packets. Those packets are lost somewhere between the proxy and
the client. After some more data is transmitted, the proxy re-sends the
missing data as one 10-byte packet. This packet is ignored by the client.
The client ignores and dropped any more incoming packets from the server.

>From snort, here are the packets which are lost:

=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+

07/17-14:34:18.391665 65.89.88.135:90 -> 216.158.32.38:10030
TCP TTL:64 TOS:0x0 ID:23851 IpLen:20 DgmLen:57 DF
***AP*** Seq: 0x499CEA40  Ack: 0xA509642A  Win: 0x43E0  TcpLen: 32
TCP Options (3) => NOP NOP TS: 7268336 1992429
3C 2F 44 4C 3E                                   </DL>

=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+

07/17-14:34:18.391665 65.89.88.135:90 -> 216.158.32.38:10030
TCP TTL:64 TOS:0x0 ID:23852 IpLen:20 DgmLen:57 DF
***AP*** Seq: 0x499CEA45  Ack: 0xA509642A  Win: 0x43E0  TcpLen: 32
TCP Options (3) => NOP NOP TS: 7268336 1992429
3C 2F 44 4C 3E                                   </DL>

And here is the re-send packet which is received by the client, but
ignored:

=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+

07/17-14:34:18.681665 65.89.88.135:90 -> 216.158.32.38:10030
TCP TTL:64 TOS:0x0 ID:23872 IpLen:20 DgmLen:62 DF
***AP*** Seq: 0x499CEA40  Ack: 0xA509642A  Win: 0x43E0  TcpLen: 32
TCP Options (3) => NOP NOP TS: 7268365 1992460
3C 2F 44 4C 3E 3C 2F 44 4C 3E                    </DL></DL>

--------------------------------------------------------------------------

If I set /proc/sys/net/ipv4/tcp_retrans_collapse to zero, the problem goes
away. Could this be a bug in the tcp_retrans_collapse() code in the kernel?

I am new to posting on the linux kernel newsgroups, please let me know if
any more information would be helpful, and please reply by E-mail since I
am not subscribed to Linux-kernel.

Sincerely,
Terence Haddock

Re: TCP/IP connections hang with TCP_RETRANS_COLLAPSE=1 - kernel bug?

[David S. Miller]

Terence Haddock writes:
 > I am having TCP/IP connections "hang", and investigation seems to point to
 > the tcp_retrans_collapse feature as the culprit. I have the following
 > network configuration:

Known checksum computation bug, fixed in current 2.4.7-preX kernels.

Later,
David S. Miller
davem@redhat.com