TCP socket timeout after IP address is removed

TCP socket timeout after IP address is removed

[Zoltan Kiss]

Hi,

We have a timout problem on TCP sockets, which probably has some trivial 
solution, but I couldn't figure it out after hours of staring at the 
code, so hopefully someone can shed some light on it here.
The scenario is the following:
- I start an ssh connection from host A to B, it works fine
- netstat shows the socket is using local address K
- then I remove local address K, and with DHCP I get a new one, L on a 
different interface (interfaces are always up during the whole time)
- the socket is still alive, but my SSH connection doesn't transmit 
anything, obviously, as netstat says it is still using IP address K
- ftrace showed me the data goes through the TCP layer and finally the 
IP layer drops it because the route is no longer valid

On 3.10 kernel it times out after ~15 minutes or so, which isn't too 
good for me. On 2.6.32 it times out after a couple of minutes, and I 
would like to achieve the same, but I couldn't figure out how. Could 
anyone give me some guidance?

Regards,

Zoltan Kiss

Re: TCP socket timeout after IP address is removed

[Yuchung Cheng]

On Fri, Jul 25, 2014 at 1:04 PM, Zoltan Kiss <zoltan.kiss@citrix.com> wrote:
> Hi,
>
> We have a timout problem on TCP sockets, which probably has some trivial
> solution, but I couldn't figure it out after hours of staring at the code,
> so hopefully someone can shed some light on it here.
> The scenario is the following:
> - I start an ssh connection from host A to B, it works fine
> - netstat shows the socket is using local address K
> - then I remove local address K, and with DHCP I get a new one, L on a
> different interface (interfaces are always up during the whole time)
> - the socket is still alive, but my SSH connection doesn't transmit
> anything, obviously, as netstat says it is still using IP address K
> - ftrace showed me the data goes through the TCP layer and finally the IP
> layer drops it because the route is no longer valid
>
> On 3.10 kernel it times out after ~15 minutes or so, which isn't too good
> for me. On 2.6.32 it times out after a couple of minutes, and I would like
> to achieve the same, but I couldn't figure out how. Could anyone give me
> some guidance?

Depending on your scenario, you can either try tweaking the sysctl
net.ipv4.tcp_retries2 or use setsockopt TCP_USER_TIMEOUT

man tcp or see http://patchwork.ozlabs.org/patch/62889/


>
> Regards,
>
> Zoltan Kiss
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: TCP socket timeout after IP address is removed

[Cong Wang]

On Fri, Jul 25, 2014 at 4:31 PM, Yuchung Cheng <ycheng@google.com> wrote:
> On Fri, Jul 25, 2014 at 1:04 PM, Zoltan Kiss <zoltan.kiss@citrix.com> wrote:
>> Hi,
>>
>> We have a timout problem on TCP sockets, which probably has some trivial
>> solution, but I couldn't figure it out after hours of staring at the code,
>> so hopefully someone can shed some light on it here.
>> The scenario is the following:
>> - I start an ssh connection from host A to B, it works fine
>> - netstat shows the socket is using local address K
>> - then I remove local address K, and with DHCP I get a new one, L on a
>> different interface (interfaces are always up during the whole time)
>> - the socket is still alive, but my SSH connection doesn't transmit
>> anything, obviously, as netstat says it is still using IP address K
>> - ftrace showed me the data goes through the TCP layer and finally the IP
>> layer drops it because the route is no longer valid
>>
>> On 3.10 kernel it times out after ~15 minutes or so, which isn't too good
>> for me. On 2.6.32 it times out after a couple of minutes, and I would like
>> to achieve the same, but I couldn't figure out how. Could anyone give me
>> some guidance?
>
> Depending on your scenario, you can either try tweaking the sysctl
> net.ipv4.tcp_retries2 or use setsockopt TCP_USER_TIMEOUT
>
> man tcp or see http://patchwork.ozlabs.org/patch/62889/
>

Besides that the cached dst by that socket should be obsolete as well,
since the old routing entry might not be valid any more. I mention this
because routing cache was removed between 2.6.32 and 3.10.