read() returns ETIMEDOUT on steady TCP connection

[Mark Hills]

On Sun, 20 Apr 2008, Peter Jeremy wrote:

> Can you give some more detail about your hardware (speed, CPU,
> available RAM, UP or SMP) and the application (roughly what does the
> core of the code look like and is it single-threaded/multi-threaded
> and/or multi-process).

The current test is a Dell 2650, 2Gb, Quad Xeon with onboard bge.

The application is single threaded, non-blocking multiplexed I/O based on 
poll(). It's relatively simple at its core -- read() from an inbound 
connection and write() to outbound sockets.

>> As the number of outbound connections increases, the 'output drops'
>> increases to around 10% of the total packets sent and maintains that ratio.
>> There's no problems with network capacity.
>
> 'output drops' (ips_odropped) means that the kernel is unable to
> buffer the write (no mbufs or send queue full).  Userland should see
> ENOBUFS unless the error was triggered by a fragmentation request.

The app definitely isn't seeing ENOBUFS; this would be treated as a fatal 
condition and reported.

> I can't explain the problem but it definitely looks like a resource
> starvation issue within the kernel.

I've traced the source of the ETIMEDOUT within the kernel to 
tcp_timer_rexmt() in tcp_timer.c:

   if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) {
           tp->t_rxtshift = TCP_MAXRXTSHIFT;
           tcpstat.tcps_timeoutdrop++;
           tp = tcp_drop(tp, tp->t_softerror ?
                         tp->t_softerror : ETIMEDOUT);
           goto out;
   }

I'm new to FreeBSD, but it seems to implies that it's reaching a limit of 
a number of retransmits of sending ACKs on the TCP connection receiving 
the inbound data? But I checked this using tcpdump on the server and could 
see no retransmissions.

As a test, I ran a simulation with the necessary changes to increase 
TCP_MAXRXTSHIFT (including adding appropriate entries to 
tcp_sync_backoff[] and tcp_backoff[]) and it appeared I was able to reduce 
the frequency of the problem occurring, but not to a usable level.

With ACKs in mind, I took the test back to stock kernel and configuration, 
and went ahead with disabling sack on the server and the client which 
supplies the data (FreeBSD 6.1, not 7). This greatly reduced the 
'duplicate acks' metric, but didn't fix the problem. The next step was to 
switch off delayed_ack as well, and I didn't see the problem for some 
hours on the test system at 850mbit output. But hasn't eliminated it, as 
it happened again.

Perhaps someone with a greater knowledge can help to join the dots of all 
these symptoms?

Mark