Re: [RFC PATCH v3 3/3] nvme: wire up support for async passthrough

[Jens Axboe <axboe@kernel.dk>]

On 3/17/21 10:59 AM, Christoph Hellwig wrote:
> On Wed, Mar 17, 2021 at 10:49:28AM -0600, Jens Axboe wrote:
>> I will post it soon, only reason I haven't reposted is that I'm not that
>> happy with how the sqe split is done (and that it's done in the first
>> place). But I'll probably just post the current version for comments,
>> and hopefully we can get it to where it needs to be soon.
> 
> Yes, I don't like that at all either.  I almost wonder if we should
> use an entirely different format after opcode and flags, although
> I suspect fd would be nice to have in the same spot as well.

Exactly - trying to think of how best to do this. It's somewhat a shame
that I didn't place user_data right after fd, or even at the end of the
struct. But oh well.

One idea would be to have io_uring_sqe_hdr and have that be
op/flags/prio/fd as we should have those for anything, and just embed
that at the top of both io_uring_sqe (our general command), and
io_uring_whatever which is what the passthrough stuff would use.

Not sure, I need to dabble in the code a bit and see how we can make it
the cleanest.

> On a related note: I think it really should have a generic cmd
> dispatching mechanism like ioctls have, preferably even enforcing
> the _IO* mechanism.

Yes, we could certainly do that. I don't want to turn it into a
free-for-all and the wild west of passthrough, some notion of coherent
definitions would be prudent.

-- 
Jens Axboe