strange TCP issue on RELENG_7

[Julian Elischer]

Kip Macy wrote:
> On Sat, Aug 23, 2008 at 10:52 PM, Julian Elischer <julian at elischer.org> wrote:
>> Mike Tancsa wrote:
>>> At 10:16 PM 8/23/2008, Kip Macy wrote:
>>>> Can you help me out a bit with your workload?
>>>>
>>>> tcp_offload_connect(...) needs to determine which interface an address
>>>> corresponds to see if that interface supports TCP offload. The code
>>>> does the exact same thing as ip_output does except it doesn't have the
>>>> inpcb locked (which isn't used as part of the route lookup).
>>> This is the only RELENG_7 box that I have where it routes tcp packets
>>> asymmetrically, so that sounds like it might be the portion that is badly
>>> interacting. The server has just one default gateway, which is out em0, but
>>> clients all over the net will connect to IP addresses aliased on lo0 and to
>>> the one IP on em1.  But all connections exit out em0 other than connected
>>> routes of course.
>>>
>>>        ---Mike
>>>
>>>> Julian has worked in this code most recently, maybe he has some idea
>>>> what is going on.
>>>>
>> huh? wha?  I haven't been following this thread.. what's up?
>>
> Julian - see previous e-mails, the arp cache gets messed up as a
> result of calling rtalloc in tcp_offload.c - which is done to
> determine which interface will be used for connection. Any thoughts on
> why it may end up with dozens of bogus entries?
> 
> -Kip

has anyone tried the same scenario on -current?