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

[Christoph Hellwig <hch@lst.de>]

[-- Attachment #1: Type: text/plain, Size: 1597 bytes --]

On Fri, Jun 22, 2018 at 12:01:17PM +0100, Al Viro wrote:
> For fuck sake, if you want vfs_poll() inlined, *make* *it* *inlined*.

That is not going to help with de-virtualizing _qproc, which was
the whole idea of that change.  At least not without a compiler
way smarter than gcc.

But if you want it inline that is fine with me, it just seems little
large for inlining.

None that I plan to actually remove all calls except for poll and select
for vfs_poll in a pending series, at which point it would become static
anyway.

> Said that, you are not attacking the worst part of it - it's a static
> branch, not the considerably more costly indirect ones.  Remember when
> I asked you about the price of those?  Method calls are costly.

And back then it did not show up even in poll heavy workloads.  But
since then something new happened - spectre mitigations, which make
indirect calls exorbitantly more expensive.

> Now, ->sk_wq is modified only in sock_init_data() and sock_graft();
> the latter, IIRC, is ->accept() helper.  Do we ever call either of
> those on a sock of already opened file?  IOW, is there any real
> reason for socket ->get_poll_head() not to be constant, other
> than wanting to keep POLL_BUSY_LOOP handling out of ->poll_mask()?
> I agree that POLL_BUSY_LOOP is ugly as hell, but you *still* have
> sock_poll_mask() not free from it...

I'd have to defer to networking folks if busy looping after pollwait
is what they want, but I suspect the answer is no, by the time
we are already waiting for the queue busy waiting seems pointless.