Re: regression: nvme rdma with bnxt_re0 broken

[Jason Gunthorpe <jgg@ziepe.ca>]

On Fri, Jul 12, 2019 at 09:59:38PM +0530, Selvin Xavier wrote:

> > bnxt guys: please just delete this duplicate detection code from the
> > driver. Every GID provided from the core must be programmed into the
> > given gid table index.
> 
> Jason,
>  This check is required in bnxt_re because the HW need only one entry
> in its table for RoCE V1 and RoCE v2 Gids.

The HW doesn't have a 'GID table' if it has this restriction. It
sounds like it has some 'IP table' or maybe 'IP and VLAN' table?

So the driver must provide a full emulated 'GID Table' with all the
normal semantics.

Which looks sort of like what the add side is doing, other than the
mis-naming it seems reasonable..

But then I see this in re_create_ah:

	/*
	 * If RoCE V2 is enabled, stack will have two entries for
	 * each GID entry. Avoiding this duplicte entry in HW. Dividing
	 * the GID index by 2 for RoCE V2
	 */
	ah->qplib_ah.sgid_index = grh->sgid_index / 2;

Which seems completely wrong - that is making assumptions about the
layout of the gid table that is just not allowed.

Surely it needs to access the per-gid driver context that add_gid set
and use the index into the sgid_table from there?

Jason