[Openib-windows] Added ability to change IPoIB mtu size

[Yossi Leybovich]

> -----Original Message-----
> From: Fabian Tillier [mailto:ftillier.sst at gmail.com] 
> Sent: Wednesday, November 22, 2006 9:13 AM
> To: Anatoly Lisenko
> Cc: Yossi Leybovich; openib-windows at openib.org
> Subject: Re: [Openib-windows] Added ability to change IPoIB mtu size
> 
> Hi folks,
> 
> On 11/19/06, Anatoly Lisenko <anatolyl at voltaire.com> wrote:
> >
> > Hi Yossi,
> >
> > This patch add the ability to change MTU size in IPoIB.
> >
> > Please review.
> >
> > Thanks,
> >
> > Anatoly
> 
> It would be nice to have an explanation of what problem 
> patches are trying to solve.  I assume this is done to enable 
> virtualization over IPoIB, which requires network packets to 
> be ethernet MTU-sized.
> 
> I didn't have a chance to look at the patch, but have these thoughts.
> Limiting the MTU reported to the OS works fine for outbound 
> packets, but care must be exercised for inbound packets.
> 
> 1. The buffers posted to the receive queue must be the full 
> 2048 byte IB MTU so that the IPoIB driver can receive any 
> packets sent to it without hitting a protection error due to 
> overrunning the receive buffers.
> 2. If a buffer larger than the configured MTU is received, it 
> must be reported to the OS somehow.  I don't know what's 
> required to do this, whether you create a new IP header for 
> the data beyond the first, or whatnot, but it must be done in 
> such a way that subsequent packets don't cause issues with 
> sequence numbers.  This doesn't seem like a simple thing to 
> do right, let alone in a performant way.
> 
Cant we just drop the packets and add that to the statiscs of packet
discards?
Wny how I will check that in the code.

> Cheers,
> 
> - Fab
>