Re: [PATCH V4] block: optimize for small block size IO

[Jens Axboe <axboe@kernel.dk>]

On 11/4/19 7:46 PM, Ming Lei wrote:
> On Mon, Nov 04, 2019 at 09:30:02PM -0500, Kent Overstreet wrote:
>> On Tue, Nov 05, 2019 at 10:20:46AM +0800, Ming Lei wrote:
>>> On Mon, Nov 04, 2019 at 09:11:30PM -0500, Kent Overstreet wrote:
>>>> On Tue, Nov 05, 2019 at 09:11:35AM +0800, Ming Lei wrote:
>>>>> On Mon, Nov 04, 2019 at 01:42:17PM -0500, Kent Overstreet wrote:
>>>>>> On Mon, Nov 04, 2019 at 11:23:42AM -0700, Jens Axboe wrote:
>>>>>>> On 11/4/19 11:17 AM, Kent Overstreet wrote:
>>>>>>>> On Mon, Nov 04, 2019 at 10:15:41AM -0800, Christoph Hellwig wrote:
>>>>>>>>> On Mon, Nov 04, 2019 at 01:14:03PM -0500, Kent Overstreet wrote:
>>>>>>>>>> On Sat, Nov 02, 2019 at 03:29:11PM +0800, Ming Lei wrote:
>>>>>>>>>>> __blk_queue_split() may be a bit heavy for small block size(such as
>>>>>>>>>>> 512B, or 4KB) IO, so introduce one flag to decide if this bio includes
>>>>>>>>>>> multiple page. And only consider to try splitting this bio in case
>>>>>>>>>>> that the multiple page flag is set.
>>>>>>>>>>
>>>>>>>>>> So, back in the day I had an alternative approach in mind: get rid of
>>>>>>>>>> blk_queue_split entirely, by pushing splitting down to the request layer - when
>>>>>>>>>> we map the bio/request to sgl, just have it map as much as will fit in the sgl
>>>>>>>>>> and if it doesn't entirely fit bump bi_remaining and leave it on the request
>>>>>>>>>> queue.
>>>>>>>>>>
>>>>>>>>>> This would mean there'd be no need for counting segments at all, and would cut a
>>>>>>>>>> fair amount of code out of the io path.
>>>>>>>>>
>>>>>>>>> I thought about that to, but it will take a lot more effort.  Mostly
>>>>>>>>> because md/dm heavily rely on splitting as well.  I still think it is
>>>>>>>>> worthwhile, it will just take a significant amount of time and we
>>>>>>>>> should have the quick improvement now.
>>>>>>>>
>>>>>>>> We can do it one driver at a time - driver sets a flag to disable
>>>>>>>> blk_queue_split(). Obvious one to do first would be nvme since that's where it
>>>>>>>> shows up the most.
>>>>>>>>
>>>>>>>> And md/md do splitting internally, but I'm not so sure they need
>>>>>>>> blk_queue_split().
>>>>>>>
>>>>>>> I'm a big proponent of doing something like that instead, but it is a
>>>>>>> lot of work. I absolutely hate the splitting we're doing now, even
>>>>>>> though the original "let's work as hard as we add add page time to get
>>>>>>> things right" was pretty abysmal as well.
>>>>>>
>>>>>> Last I looked I don't think it was going to be that bad, just needed a bit of
>>>>>> finesse. We just need to be able to partially process a request in e.g.
>>>>>> nvme_map_data(), and blk_rq_map_sg() needs to be modified to only map as much as
>>>>>> will fit instead of popping an assertion.
>>>>>
>>>>> I think it may not be doable.
>>>>>
>>>>> blk_rq_map_sg() is called by drivers and has to work on single request, however
>>>>> more requests have to be involved if we delay the splitting to blk_rq_map_sg().
>>>>> Cause splitting means that two bios can't be submitted in single IO request.
>>>>
>>>> Of course it's doable, do I have to show you how?
>>>
>>> No, you don't have to, could you just point out where my above words is wrong?
>>
>> blk_rq_map_sg() _currently_ works on a single request, but as I said from the
>> start that this would involve changing it to only process as much of a request
>> as would fit on an sglist.
> 
>> Drivers will have to be modified, but the changes to driver code should be
>> pretty easy. What will be slightly trickier will be changing blk-mq to handle
>> requests that are only partially completed; that will be harder than it would
>> have been before blk-mq, since the old request queue code used to handle
>> partially completed requests - not much work would have to be done that code.
> 
> Looks you are suggesting partial request completion.
> 
> Then the biggest effect could be in performance, this change will cause the
> whole FS bio is handled part by part serially, instead of submitting all
> splitted bios(part) concurrently.
> 
> So sounds you are suggesting to fix one performance issue by causing new perf
> issue, is that doable?

It does seem like a rat hole of sorts. Because then you start adding code to
guesstimate how big the request could roughly be, and if you miss a bit,
you get a request that's tiny in between the normal sized ones.

Or you'd clone, and then you could still have them inflight in parallel.
But then you're paying the cost of that...

-- 
Jens Axboe