[ewg] [Bug 418] IPoIB CM causes large message IPv4 multicast to fail (was OFED 1.2 beta blocking bugs)

[Michael S. Tsirkin]

> Quoting Sean Hefty <mshefty at ichips.intel.com>:
> Subject: Re: [ofa-general] bug 418: was OFED 1.2 beta blocking bugs
> 
> >     ib0: failed send event (status=1, wrid=35 vend_err 69)
> 
> I believe that this is causing the QP to transition into the error state.
> 
> >     ib_mthca 0000:08:00.0: modify QP 3->3 returned status 10.
> 
> The mthca status of 0x10 indicates a bad QP state.  The transition from 3->3 is 
> RTS to RTS, but the QP is not in the RTS state, which makes sense given the 
> previous error.  The other receive side errors in the bug report are a fallout 
> from not recovering from the send error.

Errors on UD QP typically indicates a software problem.
It seems we are posting packets that exceed the MTU size.
But I do not see this problem here at the lab.
How to reproduce this problem?

> I don't know if this causes any problems, but at first glance it appears that 
> the IPoIB CM code begins listening for connection requests before the code has 
> had a chance to join the IPoIB broadcast group.  This allows a connection to 
> form before the broadcast traffic is ready.  Someone more familiar with the code 
> than I am will need to determine if this can lead to any undesirable race 
> conditions.

I don't see why is this a problem - I don't need to be a member of a broadcast group
to get incoming packets.

-- 
MST