Re: [RFC PATCH 0/2] RDMA/rxe: Add RDMA Atomic Write operation

["yangx.jy@fujitsu.com" <yangx.jy@fujitsu.com>]

On 2021/12/31 5:42, Tom Talpey wrote:
> On 12/30/2021 2:21 PM, Gromadzki, Tomasz wrote:
>> 1)
>>> rdma_post_atomic_writev(struct rdma_cm_id *id, void *context, struct 
>>> ibv_sge *sgl,
>>>             int nsge, int flags, uint64_t remote_addr, uint32_t rkey)
>>
>> Do we need this API at all?
>> Other atomic operations (compare_swap/add) do not use struct ibv_sge 
>> at all but have a dedicated place in
>> struct ib_send_wr {
>> ...
>>         struct {
>>             uint64_t    remote_addr;
>>             uint64_t    compare_add;
>>             uint64_t    swap;
>>             uint32_t    rkey;
>>         } atomic;
>> ...
>> }
>>
>> Would it be better to reuse (extend) any existing field?
>> i.e.
>>         struct {
>>             uint64_t    remote_addr;
>>             uint64_t    compare_add;
>>             uint64_t    swap_write;
>>             uint32_t    rkey;
>>         } atomic;
>
> Agreed. Passing the data to be written as an SGE is unnatural
> since it is always exactly 64 bits. Tweaking the existing atomic
> parameter block as Tomasz suggests is the best approach.
Hi Tomasz, Tom

Thanks for your quick reply.

If we want to pass the 8-byte value by tweaking struct atomic on user 
space, why don't we
tranfer the 8-byte value by ATOMIC Extended Transport Header (AtomicETH) 
on kernel space?
PS: IBTA defines that the 8-byte value is tranfered by RDMA Extended 
Transport Heade(RETH) + payload.

Is it inconsistent to use struct atomic on user space and RETH + payload 
on kernel space?

Best Regards,
Xiao Yang
>
> Tom.