> > Is it possible for the page to be bigger than 4GB?  If so then yes you
 > > might be chopping off high-order bits or something.

 > Yes it is possible.
 > 
 > A MR can be created with an iov_base of say 0xffffffff00000000.
 > 
 > Then any sge.addr entries would be the iob_base + any offset.

But the code we're talking about is:

		/* to in the WQE == the offset into the page */
		wqe->recv.sgl[i].to = cpu_to_be64(((u32) wr->sg_list[i].addr) %
				(1UL << (12 + page_size[i])));

so it seems the top address bits don't matter unless page_size[i] is at
least 20 -- in which case using 1UL to shift overflows on 32 bits anyway...

 > So you are handling the offset patch that will make it u64 and remove
 > the mod usage, correct?

Yeah, I rolled the fix into the "offset needs to be u64" patch, it
should be in linux-next by now (or at least in my for-next branch).

 - R.