[PATCH 2/7] block: have ->poll_fn() return number of entries polled

[axboe@kernel.dk (Jens Axboe)]

On 11/25/18 3:52 AM, Sagi Grimberg wrote:
>> We currently only really support sync poll, ie poll with 1
>> IO in flight. This prepares us for supporting async poll.
> 
> Hey Jens,
> 
> So are we sure that this is fine to simply replace the
> poll functionality? you say that that we support poll
> with only 1 I/O inflight but is it entirely true?
> semantically speaking?

It is, unless you have multiple threads all doing polling.  Which is
pretty inefficient, as I'm sure you know.

> The interface is selective polling, which means that
> the user can have more than a single I/O inflight but wants
> to poll for a specific one that it really cares about.
> 
> Would this introduce a regression for users that rely
> on preadv2 to basically "complete *this* IO as fast as possible"?

No, it'll be exactly the same, and believe me, I've done plenty of
performance testing to ensure that it works well.  In fact, with the
changes queued up and this on top, polling is both faster and more
efficient than it ever has been before, for both the classic
preadv2/pwritev2 and the async polled IO.

> Note that I like the proposed direction, I'm merely questioning
> if it is OK to simply change how selective polling worked until
> today instead of introducing a separate blk_poll_all(q) (but with
> a better name) which would be wired up to aio polling and friends.

I think that would be useless. For SYNC type polling with one thread, it
doesn't matter if we're looking for a particular IO, or just any IO.
Once your into two threads, you're already wasting huge amounts of CPU,
just to get to QD=2. The poll loop handles finding each others IOs just
fine, so there's no worry on that side.

Polling was originally designed for the strictly SYNC interfaces, and
since we had the cookie anyway, it was tempting to look for a specific
IO. I think that was a mistake, as it assumed that we'd never do async
polling.

-- 
Jens Axboe