[PATCH v2 2/3] nvme-rdma: don't complete requests before a send work request has completed

[Sagi Grimberg]

>>> Both send and recv completed only ever go from not set to set on a live
>>> requests, so that's easy.  need_inval only goes from set to cleared
>>> so I'm not too worried about it either, but due to the lack of atomic
>>> bitops there it will need better memory barries, or just a switch
>>> to bitops..
>>>
>>> But looking at this in a little more detail I wonder if we just want
>>> a recount on the nvme_rdma_request, untested patch below.
>>
>> I'll give it a shot.
> 
> Btw, in our brave new world that should probably be a refcnt_t instead
> of an atomic_t, but that's a trivial search and replace.

Works like a charm :)

Sending a newer v3 soon.

Thanks Christoph!