[openib-general] Re: vapi versus openib imm_data

[Pete Wyckoff]

jackm at mellanox.co.il wrote on Mon, 22 May 2006 09:30 +0300:
> On Monday 22 May 2006 08:19, Michael S. Tsirkin wrote:
> > Quoting r. Pete Wyckoff <pw at osc.edu>:
> > > is there any good way to tell if the other side has put
> > > its imm_data in network byte order or not?
> >
> > gen2 always assumes imm_data is given in network byte order.
> 
> VAPI assumes that imm_data is given (i.e., supplied to the API) in host byte 
> order.  If the host is a little-endian host (as PCs are), the mlxhh (i.e., 
> inner) layer will convert immediate data to network byte order on the send, 
> and will convert received immediate data from Network byte order to host byte 
> order on receive -- and the VAPI caller will receive the immediate data in 
> host byte order.

Thanks both.  I'll solve this by adding htonl/ntohl, only on the
VAPI side, when sending immediate data and reading it back from
the CQ.  That should undo the little-endian-only byte swap
introduced by VAPI.  Much easier than trying to figure out what
the other side is doing with its immediate data.

		-- Pete