On Tue, 06 Jan 2015 13:44:01 +1100
Cyril Bur <cyril...@gmail.com> wrote:
> On Mon, 2015-01-05 at 14:10 -0800, Andrew Morton wrote:
> > On Mon, 22 Dec 2014 16:06:04 +1100 Cyril Bur <cyril...@gmail.com> wrote:
> >
> > > On POWER8 virtualised kernels the VTB register can be read to have a view
> > > of
> > > time that only increases while the guest is running. This will prevent
> > > guests
> > > from seeing time jump if a guest is paused for significant amounts of
> > > time.
> > >
> > > On POWER7 and below virtualised kernels stolen time is subtracted from
> > > sched_clock as a best effort approximation. This will not eliminate
> > > spurious
> > > warnings in the case of a suspended guest but may reduce the occurance in
> > > the
> > > case of softlockups due to host over commit.
> > >
> > > Bare metal kernels should avoid reading the VTB as KVM does not restore
> > > sane
> > > values when not executing. sched_clock is returned in this case.
> > >
> > > --- a/arch/powerpc/kernel/time.c
> > > +++ b/arch/powerpc/kernel/time.c
> > > @@ -621,6 +621,30 @@ unsigned long long sched_clock(void)
> > > return mulhdu(get_tb() - boot_tb, tb_to_ns_scale) << tb_to_ns_shift;
> > > }
> > >
> > > +unsigned long long running_clock(void)
> >
> > Non-kvm kernels don't need this code. Is there some appropriate
> > "#ifdef CONFIG_foo" we can wrap this in?
> CONFIG_PSERIES would work, having said that typical compilation for a
> powernv kernel almost always includes CONFIG_PSERIES (although it
> doesn't need to)... still, your point is valid, will add in v2.
> >
> >
> > > +{
> > > + /*
> > > + * Don't read the VTB as a host since KVM does not switch in host
> > > timebase
> > > + * into the VTB when it takes a guest off the CPU, reading the VTB would
> > > + * result in reading 'last switched out' guest VTB.
> > > + */
> > > +
> > > + if (firmware_has_feature(FW_FEATURE_LPAR)) {
> > > + if (cpu_has_feature(CPU_FTR_ARCH_207S))
> > > + return mulhdu(get_vtb() - boot_tb, tb_to_ns_scale) <<
> > > tb_to_ns_shift;
> > > +
> > > + /* This is a next best approximation without a VTB. */
> > > + return sched_clock() -
> > > cputime_to_nsecs(kcpustat_this_cpu->cpustat[CPUTIME_STEAL]);
> >
> > Why is this result dependent on FW_FEATURE_LPAR? It's all generic code.
> Good point, the reason it ended up there is because I wanted to avoid
> behaviour changes.
> >
> > In fact the kernel/sched/clock.c default implementation of
> > running_clock() could use this expression. Would that be good or bad? :)
> For power I'm almost certain it would be fine, on platforms which don't
> do stolen time cpustat[CPUTIME_STEAL] should always be zero and if not
> then the value should always be sane (although as mentioned in the
> comment, not as accurate as using the VTB).
>
> Putting it in the default implementation could cause behavioural changes
> for x86 and s390, I would want their views on doing that.
I would prefer to make sched_clock do all the work. We have been thinking
about steal time vs sched_clock as well, our solution would be to exchange
the time source. Right now sched_clock is based on the TOD clock, the code
that takes steal time into account would use the CPU timer instead.
With the subtraction of kcpustat_this_cpu->cpustat[CPUTIME_STEAL] in
common code we would have to add the same value in the sched_clock
implementation as the steal time is already included in the CPU timer
deltas.
--
blue skies,
Martin.
"Reality continues to ruin my life." - Calvin.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
