On 1/25/20 9:41 PM, Andrzej Jakowski wrote:

In current implementation it is possible to perform IO polling if
underlying device is block-mq device. Is is not possible however to do
polled IO on stackable block devices like MD.

We have explored two paths for enabling IO polling on bio devices. First
idea revolved around rewriting MD to block-mq interface but it proved to
be complex. In the second idea we have built a prototype which
introduced new operation on request_queue - bio_poll_fn. bio_poll_fn if
provided by stackable block driver is called when user polls for IO
completion. bio_poll_fn approach was initially discussed and confirmed
with Jens.

We managed to collect performance data on RAID-0 volume built on top of
2xP4800X devices with polling on and off. Here are the results:
Polling      QD       Latency         IOPS
----------------------------------------------
off           1       12.03us         78800
off           2       13.27us        144000
off           4       15.83us        245000
off           8       31.14us        253000
off          16       63.03us        252000
off          32      128.89us        247000
off          64      259.10us        246000
off         128      517.27us        247000
on            1        9.00us        108000
on            2        9.07us        214000
on            4       12.00us        327000
on            8       21.43us        369000
on           16       43.18us        369000
on           32       85.75us        372000
on           64      169.87us        376000
on          128      346.15us        370000

blk_poll() used to be a pointer in the queue, but since we just had
one implementation, we got rid of it. Might make sense to
reintroduce that, and just make it an optimized indirect call. I
think that would be prettier than add the bio hack in the middle of
it, and you're having to add a queue pointer anyway.

Alternatively, do it like you did, but have it be:

if (q->poll_fn)
	return q->poll_fn(...);

instead of duplicating most of the core of the function with essentially
the same thing, just calling ->bio_poll_fn() instead of mq_ops->poll().

In other words, I like the feature, but I think the implementation
currently leaves a lot to be desired. It should be nicely integrated,
not some hack off to the side.

-- 
Jens Axboe