Re: [lkp-robot] [fs] 3deb642f0d: will-it-scale.per_process_ops -8.8% regression

[Christoph Hellwig]

On Fri, Jun 22, 2018 at 06:25:45PM +0900, Linus Torvalds wrote:
> What was the alleged advantage of the new poll methods again? Because
> it sure isn't obvious - not from the numbers, and not from the commit
> messages.

The primary goal is that we can implement a race-free aio poll,
the primary benefit is that we can get rid of the currently racy
and bug prone way we do in-kernel poll-like calls for things like
eventfd.  The first is clearly is in 4.18-rc and provides massive
performance advantanges if used, the second is not there yet,
more on that below.

> I was assuming there  was a good reason for it, but looking closer I
> see absolutely nothing but negatives. The argument that keyed wake-ups
> somehow make multiple wake-queues irrelevant doesn't hold water when
> the code is more complex and apparently slower. It's not like anybody
> ever *had* to use multiple wait-queues, but the old code was both
> simpler and cleaner and *allowed* you to use multiple queues if you
> wanted to.

It wasn't cleaner at all if you aren't poll or select, and even
for those it isn't exactly clean, see the whole mess around ->qproc.

> The disadvantages are obvious: every poll event now causes *two*
> indirect branches to the low-level filesystem or driver - one to get
> he poll head, and one to get the mask. Add to that all the new "do we
> have the new-style or old sane poll interface" tests, and poll is
> obviously more complicated.

It already caused two, and now we have three thanks to ->qproc.  One
of the advantages of the new code is that we can eventually get rid
of ->qproc once all users of a non-default qproc are switched away
from vfs_poll.  Which requires a little more work, but I have the
patches for that to be posted soon.

> If we could get the poll head by just having a direct pointer in the
> 'struct file', maybe that would be one thing. As it is, this all
> literally just adds overhead for no obvious reason. It replaced one
> simple direct call with two dependent but separate ones.

People are doing weird things with their poll heads, so we can't do
that unconditionally.  We could however offer a waitqueue pointer
in struct file and most users would be very happy with that.

In the meantime below is an ugly patch that removes the _qproc
indirect for ->poll only (similar patch is possible for select
assuming the code uses select).  And for next merge window I plan
to kill it off entirely.

How can we get this thrown into the will it scale run?

---