Re: [PATCH 00/12] x86: Rewrite 64-bit syscall code

From: Ingo Molnar <mingo@kernel.org>
To: Andy Lutomirski <luto@amacapital.net>
Cc: "Andy Lutomirski" <luto@kernel.org>, "X86 ML" <x86@kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"Brian Gerst" <brgerst@gmail.com>,
	"Borislav Petkov" <bp@alien8.de>,
	"Frédéric Weisbecker" <fweisbec@gmail.com>,
	"Denys Vlasenko" <dvlasenk@redhat.com>,
	"Linus Torvalds" <torvalds@linux-foundation.org>
Subject: Re: [PATCH 00/12] x86: Rewrite 64-bit syscall code
Date: Tue, 8 Dec 2015 08:00:03 +0100	[thread overview]
Message-ID: <20151208070003.GA26154@gmail.com> (raw)
In-Reply-To: <CALCETrVfvqhgyjO3CLiDdOoKo30aQkQ1imGSJDP-UE0LHqqMPA@mail.gmail.com>

* Andy Lutomirski <luto@amacapital.net> wrote:

> On Mon, Dec 7, 2015 at 8:42 PM, Ingo Molnar <mingo@kernel.org> wrote:
> >
> > * Andy Lutomirski <luto@amacapital.net> wrote:
> >
> >> On Mon, Dec 7, 2015 at 1:51 PM, Andy Lutomirski <luto@kernel.org> wrote:
> >>
> >> > This is kind of like the 32-bit and compat code, except that I preserved the
> >> > fast path this time.  I was unable to measure any significant performance
> >> > change on my laptop in the fast path.
> >> >
> >> > What do you all think?
> >>
> >> For completeness, if I zap the fast path entirely (see attached), I lose 20
> >> cycles (148 cycles vs 128 cycles) on Skylake.  Switching between movq and pushq
> >> for stack setup makes no difference whatsoever, interestingly.  I haven't tried
> >> to figure out exactly where those 20 cycles go.
> >
> > So I asked for this before, and I'll do so again: could you please stick the cycle
> > granular system call performance test into a 'perf bench' variant so that:
> >
> >  1) More people can run it all on various pieces of hardware and help out quantify
> >     the patches.
> >
> >  2) We can keep an eye on not regressing base system call performance in the
> >     future, with a good in-tree testcase.
> >
> 
> Is it okay if it's not particularly shiny or modular? [...]

Absolutely!

> [...]  The tool I'm using is here:
> 
> https://git.kernel.org/cgit/linux/kernel/git/luto/misc-tests.git/tree/tight_loop/perf_self_monitor.c
> 
> and I can certainly stick it into 'perf bench' pretty easily.  Can I
> leave making it into a proper library to some future contributor?

Sure - 'perf bench' tests aren't librarized generally - the goal is to make it 
easy to create a new measurement.

> It's actually decently fancy.  It allocates a perf self-monitoring
> instance that counts cycles, and then it takes a bunch of samples and
> discards any that flagged a context switch.  It does some very
> rudimentary statistics on the rest.  It's utterly devoid of a fancy
> UI, though.
> 
> It works very well on native, and it works better than I had expected
> under KVM.  (KVM traps RDPMC because neither Intel nor AMD has seen
> fit to provide any sensible way to virtualize RDPMC without exiting.)

Sounds fantastic to me!

Thanks,

	Ingo