nvme-tcp poll queue crash

[Sagi Grimberg]

>>> This was reported and is discussed in thread:
>>> nvme-tcp: kernel NULL pointer dereference, address: 0000000000000034
>>
>> Thanks, I couldn't remember this discussion.
> 
> It just happened late last week, so no worries.
> 
> I believe we can get rid of bio and bdev dependency, and I am working on
> detangling how to get the cookie through without these.

Keith,

I think we shouldn't have an issue *today* because the only non fs
I/O that can be polled is the fabrics connect, which has a payload
and hence has a bio. The minimal change to fix the regression afaict
is protecting against a polled request submission of a user command
that doesn't have a payload (just fail it), which again, shouldn't
happen afaict.