[PATCH for-next v5 4/4] nvme: wire up fixed buffer support for nvme passthrough

[Jens Axboe]

On 9/8/22 4:47 AM, Kanchan Joshi wrote:
> On Wed, Sep 07, 2022 at 02:51:31PM +0000, Chaitanya Kulkarni wrote:
>>
>>> ????? req = nvme_alloc_user_request(q, cmd, ubuffer, bufflen, meta_buffer,
>>> -??????????? meta_len, meta_seed, &meta, timeout, vec, 0, 0);
>>> +??????????? meta_len, meta_seed, &meta, timeout, vec, 0, 0, NULL, 0);
>>> ????? if (IS_ERR(req))
>>> ????????? return PTR_ERR(req);
>>
>> 14 Arguments to the function!
>>
>> Kanchan, I'm not pointing out to this patch it has happened over
>> the years, I think it is high time we find a way to trim this
>> down.
>>
>> Least we can do is to pass a structure member than 14 different
>> arguments, is everyone okay with it ?
>>
> Maybe it's just me, but there is something (unrelatedness) about these
> fields which makes packing all these into a single container feel bit
> unnatural (or do you already have a good name for such struct?).

I think the bigger question here would be "does it generate better
code?". Because it doesn't make the code any better at all, it just
potentially makes it more fragile. Packing into a struct is just a
work-around for the interface being pretty horrible, and it'd be a much
better idea to separate it out into separate functions instead rather
than have this behemoth of a function that does it all.

In any case, I think that's a separate cleanup that should be done, it
should not gate the change. It's already horrible.

> So how about we split the nvme_alloc_user_request into two.
> That will also serve the purpose. Here is a patch that creates
> - new nvme_alloc_user_request with 5 parameters
> - new nvme_map_user_request with 8 parameters

This is a good start though.

-- 
Jens Axboe