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

[Al Viro <viro@ZenIV.linux.org.uk>]

[-- Attachment #1: Type: text/plain, Size: 1547 bytes --]

On Fri, Jun 22, 2018 at 06:18:02PM +0200, Christoph Hellwig wrote:
> On Fri, Jun 22, 2018 at 05:28:50PM +0200, Christoph Hellwig wrote:
> > On Fri, Jun 22, 2018 at 04:14:09PM +0100, Al Viro wrote:
> > > >     http://git.infradead.org/users/hch/vfs.git/shortlog/refs/heads/remove-get-poll-head
> > > 
> > > See objections upthread re "fs,net: move poll busy loop handling into a
> > > separate method"; as for the next one... I'd like an ACK from networking
> > > folks.  The rest of queue makes sense.
> > 
> > I want to see basic results first before micro-optimizing.  After that
> > I'll send it out to the net folks for feedback.
> 
> I looked into this a bit, in the end sk_can_busy_loop does this:
> 
> 	return sk->sk_ll_usec && !signal_pending(current);
> 
> where sk_ll_usec defaults based on a sysctl that needs to be
> turned on, but can be overriden per socket.
> 
> While at the same time corect poll code already checks net_busy_loop_on
> to set POLL_BUSY_LOOP.  So except for sockets where people set the
> timeout to 0 the code already does the right thing as-is.  IMHO not
> really worth wasting a FMODE_* flag for it, but if you insist I'll add
> it.

It's not just that - there's also an issue of extra indirect call on the
fast path for sockets.  You get this method of yours + ->poll_mask(),
which hits another indirect to per-family ->poll_mask().  It might be
better to have these combined, sparing us an extra indirect call.

Just give it the same calling conventions as ->poll_mask() have...