Re: [PATCH 1/3] lib/vdso: Delay mask application in do_hres()
From: Andy Lutomirski
Date: Tue Jun 25 2019 - 16:15:35 EST
On Tue, Jun 25, 2019 at 11:27 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> On Tue, 25 Jun 2019, Thomas Gleixner wrote:
> > On Tue, 25 Jun 2019, Vincenzo Frascino wrote:
> > CC+ Andy
> > > do_hres() in the vDSO generic library masks the hw counter value
> > > immediately after reading it.
> > >
> > > Postpone the mask application after checking if the syscall fallback is
> > > enabled, in order to be able to detect a possible fallback for the
> > > architectures that have masks smaller than ULLONG_MAX.
> > Right. This only worked on x86 because the mask is there ULLONG_MAX for all
> > VDSO capable clocksources, i.e. that ever worked just by chance.
> > As we talked about that already yesterday, I tested this on a couple of
> > machines and as expected the outcome is uarch dependent. Minimal deviations
> > to both sides and some machines do not show any change at all. I doubt it's
> > possible to come up with a solution which makes all uarchs go faster
> > magically.
> > Though, thinking about it, we could remove the mask operation completely on
> > X86. /me runs tests
> Unsurprisingly the results vary. Two uarchs do not care, but they did not
> care about moving the mask either. The other two gain performance and the
> last one falls back to the state before moving the mask. So in general it
> looks like a worthwhile optimization.
At one point, I contemplated a different approach: have the "get the
counter" routine return 0 and then do if (unlikely(cycles <= last))
goto fallback. This will remove one branch from the hot path. I got
dubious results when I tried benchmarking it, probably because the
branch in question was always correctly predicted.