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

[Matias Bjorling <m@bjorling.me>]

On 04/02/2014 12:46 AM, Christoph Hellwig wrote:
> On Tue, Apr 01, 2014 at 05:16:21PM -0700, Matias Bjorling wrote:
>> Hi Christoph,
>>
>> Can you rebase it on top of 3.14. I have trouble applying it for testing.
> 
> Hi Martin,
> 
> the series is based on top of Jens' for-next branch.  I've also pushed out a
> git tree to the blk-mq-share-tags.2 branch of
> 
> 	git://git.infradead.org/users/hch/scsi.git
> 
> to make testing and reviewing easier.
>

Thanks.

Regarding the tags API. I think the best approach is a struct
blk_mq_tags_reg. That'll make their parameters very visible in the
drivers. I'll send a patch with the change, using the nvme driver as an
example.

>> For nvme, there's need for two separate types of queues. The admin queue
>> (before initializing blk-mq) and the actual hardware queues.
>>
>> Should we allow the driver to get/put tags before initializing blk-mq?
>> Or let drivers implement their own framework?
> 
> What do you mean with initializing blk-mq?  We need to allocate data
> structures for sure, and I don't see much else in terms of initialization
> in blk-mq.
> 

For the nvme driver, there's a single admin queue, which is outside
blk-mq's control, and the X normal queues. Should we allow the shared
tags structure to be used (get/put) for the admin queue, without
initializing blk-mq? or should the drivers simply implement their own
tags for their admin queue?