TCP problems after 124 days of uptime?

[Andre Oppermann]

Julian Elischer wrote:
> Andre Oppermann wrote:
>> Gleb Smirnoff wrote:
>>>   Yesterday two of our web servers running 6.2-PRERELEASE experienced
>>> problems with accepting TCP connections. A small percentage of SYN
>>> packets was ignored. The packets were seen in tcpdump output, but not
>>> processed by TCP stack. No accept queue overflows occured, according
>>> to 'netstat -sp tcp'.
>>>
>>> Failing to find any clue I have rebooted one of them, and after
>>> reboot is started to work flawlessly. Reboot also helped the second
>>> one. Both servers were booted at the same time - 124 days ago.
>>>
>>> Any ideas, any similar reports?
>>
>> ticks is a 32 bit signed integer and TCP isn't really aware of it
>> going negative and rolling over.  This is something I'm working on
>> fixing in -current.  Haven't analyzed all potential cases there yet.
>>
> 
> 124 days is 32 bits at 400Hz
> or 30 bits at 100Hz
> 
> what is the granularity of the TCP timer?

ticks == HZ

This isn't necessarily a TCP timer issue.  It may just as well come
from timestamps or RTT calculation that is done based on ticks derived
values.

-- 
Andre