Re: [RFC] blk-mq: support for shared tags

[Matias Bjorling <m@bjorling.me>]

On 04/03/2014 11:01 AM, Christoph Hellwig wrote:
> On Thu, Apr 03, 2014 at 09:45:11AM -0700, Matias Bjorling wrote:
>>> I'd still create a request_queue for the internal queue, just not register
>>> a block device for it.  For example SCSI sets up queues for each LUN
>>> found, but only a subset actually is exposed as a block device.
>>>
>>
>> Ok. That is good enough for now. A little heavy on the overhead side, if
>> only the tag logic is needed.
>>
>> What about the following suggestions for shared tags:
>>
>> 1. Rename it from blk_mq_shared_tags to blk_mq_tag_group. A driver can
>> have several tag groups that it maintains.
> 
> I was going to rename it to tag_set, but tag_group sounds fine to me as well.
>

tag_set is shorter. tag_set it is.

>> 2. Instead of blk_mq_shared_tags structure in blk_mq_reg. Have  function
>> pointer for getting the tags structure during hctx initialization. This
>> is interesting for nvme, because it has as set of tags for each hardware
>> queue it exposes.
> 
> The current code also has an array of blk_mq_tags structures, one for
> each queue.  Do you need a more complicated mapping than that?
> 

No, that's great. Had misinterpreted the arrays, now that I look at it
again. Thanks

> Btw, I was also going to siply split out the tag allocation from the
> queue registration unconditionally.  While this adds a little more
> boilerplate to simple drivers it avoids unconditional code pathes and should
> make the model much easier to understand.  I should have a new version
> of the patches soon.
> 

ack, good idea.