[dm-devel] [RFC PATCH 2/2] block: support to freeze bio based request queue

[JeffleXu]

On 4/19/21 9:50 PM, Ming Lei wrote:
> On Mon, Apr 19, 2021 at 08:05:46PM +0800, JeffleXu wrote:
>>
>>
>> On 4/15/21 6:33 PM, Ming Lei wrote:
>>> For bio based request queue, the queue usage refcnt is only grabbed
>>> during submission, which isn't consistent with request base queue.
>>>
>>> Queue freezing has been used widely, and turns out it is very useful
>>> to quiesce queue activity.
>>>
>>> Support to freeze bio based request queue by the following approach:
>>>
>>> 1) grab two queue usage refcount for blk-mq before submitting blk-mq
>>> bio, one is for bio, anther is for request;
>>
>>
>> Hi, I can't understand the sense of grabbing two refcounts on the
>> @q_usage_count of the underlying blk-mq device, while
>> @q_usage_count of the MD/DM device is kept untouched.
> 
> Follows the point:
> 
> 1) for blk-mq, we hold one refcount for bio and another for request, and
> release one after ending bio or completing request.

Blk-mq has already implemented queue freezing semantics, even without
this 'grabbing two refcount'. So is this just for the code consisdency
with the bio-based queue?


> 
> 2) for bio based queue, just holding one refcount for bio, and release it
> after the bio is ended.

OK.

> 
> As I mentioned to you, the current in-tree code only grabs the refcount
> during submitting bio for bio base queue, and the refcount is released
> after returning from submission, see __submit_bio().

Yes. I ignored that the refcount grabbed in the entry of bio submission
has been returned back when the submission completes for bio-based queue.

> 
>>
>> In the following calling stack
>>
>> ```
>> queue_poll_store
>> 	blk_mq_freeze_queue(q)
>> ```
>>
>> Is the input @q still the request queue of MD/DM device?
> 
> It can be either one after bio based io polling is supported,
> queue/io_poll is exposed for both blk-mq and bio based queue.
> 
> However, I guess bio based polling doesn't need such strict bio queue
> freezing, cause QUEUE_FLAG_POLL is only read in submission path, so
> looks current freezing just during submission is enough.

Not actually.

blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
 	struct blk_mq_hw_ctx *hctx;
 	long state;

-	if (!blk_qc_t_valid(cookie) || !blk_queue_poll(q))
+	if (!blk_queue_poll(q) || (queue_is_mq(q) && !blk_qc_t_valid(cookie)))

Here QUEUE_FLAG_POLL is still checked in blk_poll() for bio-based queue,
at least in your latest patch for bio-based polling.

-- 
Thanks,
Jeffle