On Wed, Jan 05, 2022 at 01:00:42AM +0000, yangx.jy@xxxxxxxxxxx wrote:
> 1) In kernel, current SoftRoCE copies the content of struct rdma to RETH
> and copies the content of struct atomic to AtomicETH.
> 2) IBTA defines that RDMA Atomic Write uses RETH + payload.
> According to these two reasons, I perfer to tweak the existing struct rdma.
No this is basically meaningless
The wr struct is designed as a 'tagged union' where the op specified
which union is in effect.
It turns out that the op generally specifies the network headers as
well, but that is just a side effect.
> >> How about adding a member in struct rdma? for example:
> >> struct {
> >> uint64_t remote_addr;
> >> uint32_t rkey;
> >> uint64_t wr_value:
> >> } rdma;
> >
> > Yes, that's what Tomasz and I were suggesting - a new template for the
> > ATOMIC_WRITE request payload. The three fields are to be supplied by
> > the verb consumer when posting the work request.
>
> OK, I will update the patch in this way.
We are not extending the ib_send_wr anymore anyhow.
You should implement new ops inside struct ibv_qp_ex as function
calls.
Jason
