Re: [PATCH v1 5/9] block: loop: convert to blk-mq

[Maxim Patlasov <mpatlasov@parallels.com>]

On 08/28/2014 06:06 AM, Ming Lei wrote:
> On 8/28/14, Maxim Patlasov <mpatlasov@parallels.com> wrote:
>> On 08/21/2014 09:44 AM, Ming Lei wrote:
>>> On Wed, Aug 20, 2014 at 4:50 AM, Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>>> Reworked a bit more:
>>>>
>>>> http://git.kernel.dk/?p=linux-block.git;a=commit;h=a323185a761b9a54dc340d383695b4205ea258b6
>>> One big problem of the commit is that it is basically a serialized
>>> workqueue
>>> because of single &hctx->run_work, and per-req work_struct has to be
>>> used for concurrent implementation.  So looks the approach isn't flexible
>>> enough compared with doing that in driver, or any idea about how to fix
>>> that?
>>>
>> I'm interested what's the price of handling requests in a separate
>> thread at large. I used the following fio script:
>>
>>       [global]
>>       direct=1
>>       bsrange=512-512
>>       timeout=10
>>       numjobs=1
>>       ioengine=sync
>>
>>       filename=/dev/loop0 # or /dev/nullb0
>>
>>       [f1]
>>       rw=randwrite
>>
>> to compare the performance of:
>>
>> 1) /dev/loop0 of 3.17.0-rc1 with Ming's patches applied -- 11K iops
> If you enable BLK_MQ_F_WQ_CONTEXT, it isn't strange to see this
> result since blk-mq implements a serialized workqueue.

BLK_MQ_F_WQ_CONTEXT is not in 3.17.0-rc1, so I couldn't enable it.

>
>> 2) the same as above, but call loop_queue_work() directly from
>> loop_queue_rq() -- 270K iops
>> 3) /dev/nullb0 of 3.17.0-rc1 -- 380K iops
> In my recent investigation and discussion with Jens, using workqueue
> may introduce some regression for cases like loop over null_blk, tmpfs.
>
> And 270K vs. 380K is a bit similar with my result, and it was observed that
> context switch is increased by more than 50% with introducing workqueue.

The figures are similar, but the comparison is not. Both 270K and 380K 
refer to configurations where no extra context switch involved.

>
> I will post V3 which will use previous kthread, with blk-mq & kernel aio, which
> should make full use of blk-mq and kernel aio, and won't introduce regression
> for cases like above.

That would be great!

>
>> Taking into account so big difference (11K vs. 270K), would it be worthy
>> to implement pure non-blocking version of aio_kernel_submit() returning
>> error if blocking needed? Then loop driver (or any other in-kernel user)
> The kernel aio submit is very similar with user space's implementation,
> except for block plug&unplug usage in user space aio submit path.
>
> If it is blocked in aio_kernel_submit(), you should observe similar thing
> with io_submit() too.

Yes, I agree. My point was that there is a room for optimization as my 
experiments demonstrate. The question is whether it's worthy to 
sophisticate kernel aio (and fs-specific code too) for the sake of that 
optimization.

In fact, in a simple case like block fs on top of loopback device on top 
of a file on another block fs, what kernel aio does for loopback driver 
is a subtle way of converting incoming bio-s to outgoing bio-s. In case 
you know where the image file is placed (e.g. by fiemap), such a 
conversion may be done with zero overhead and anything that makes the 
overhead noticeable is suspicious. And it is easy to imagine other 
use-cases when that extra context switch is avoidable.

Thanks,
Maxim