Re: [PATCH 4/9] srpt: chain RDMA READ/WRITE requests

[Sagi Grimberg <sagig@dev.mellanox.co.il>]

On 18/11/2015 03:17, Bart Van Assche wrote:
> On 11/13/2015 05:46 AM, Christoph Hellwig wrote:
>> -        ret = ib_post_send(ch->qp, &wr.wr, &bad_wr);
>> -        if (ret)
>> -            break;
>> +        if (i == n_rdma - 1) {
>> +            /* only get completion event for the last rdma read */
>> +            if (dir == DMA_TO_DEVICE)
>> +                wr->wr.send_flags = IB_SEND_SIGNALED;
>> +            wr->wr.next = NULL;
>> +        } else {
>> +            wr->wr.next = &ioctx->rdma_ius[i + 1].wr;
>> +        }
>>       }
>>
>> +    ret = ib_post_send(ch->qp, &ioctx->rdma_ius->wr, &bad_wr);
>>       if (ret)
>>           pr_err("%s[%d]: ib_post_send() returned %d for %d/%d\n",
>>                    __func__, __LINE__, ret, i, n_rdma);
>
> Hello Christoph,

Hi Bart,

>
> Chaining RDMA requests is a great idea. But it seems to me that this
> patch is based on the assumption that posting multiple RDMA requests
> either succeeds as a whole or fails as a whole. Sorry but I'm not sure
> that the verbs API guarantees this. In the ib_srpt driver a QP can be
> changed at any time into the error state and there might be drivers that
> report an immediate failure in that case.

I'm not so sure it actually matters if some WRs succeeded. In the normal
flow when srpt has enough available work requests (sq_wr_avail) they
should all succeed otherwise we're in trouble. If the QP transitioned
to ERROR state, then some failed, but those that succeeded will
generate flush completions, and srpt should handle it correctly
shouldn't it?

> I think even when chaining
> RDMA requests that we still need a mechanism to wait until ongoing RDMA
> transfers have finished if some but not all RDMA requests have been posted.

I'm not an expert on srpt, can you explain how this mechanism will help?