2017-01-16 15:39-0200, Marcelo Tosatti:
> On Mon, Jan 16, 2017 at 06:27:58PM +0100, Radim Krcmar wrote:
>> 2017-01-16 15:08-0200, Marcelo Tosatti:
>> > On Mon, Jan 16, 2017 at 05:54:11PM +0100, Radim Krcmar wrote:
>> >> 2017-01-16 17:26+0100, Radim Krcmar:
>> >> > 2017-01-13 15:40-0200, Marcelo Tosatti:
>> >> >> On Fri, Jan 13, 2017 at 04:56:58PM +0100, Radim Krcmar wrote:
>> >> >> > 2017-01-13 10:01-0200, Marcelo Tosatti:
>> >> >>> > + version = pvclock_read_begin(src);
>> >> >>> > +
>> >> >>> > + ret = kvm_hypercall2(KVM_HC_CLOCK_OFFSET,
>> >> >>> > + clock_off_gpa,
>> >> >>> > + KVM_CLOCK_OFFSET_WALLCLOCK);
>> >> >>> > + if (ret != 0) {
>> >> >>> > + pr_err("clock offset hypercall ret %lu\n", ret);
>> >> >>> > + spin_unlock(&kvm_ptp_lock);
>> >> >>> > + preempt_enable_notrace();
>> >> >>> > + return -EOPNOTSUPP;
>> >> >>> > + }
>> >> >>> > +
>> >> >>> > + tspec.tv_sec = clock_off.sec;
>> >> >>> > + tspec.tv_nsec = clock_off.nsec;
>> >> >>> > +
>> >> >>> > + delta = rdtsc_ordered() - clock_off.tsc;
>> >> >>> > +
>> >> >>> > + offset = pvclock_scale_delta(delta,
>> >> >>> > src->tsc_to_system_mul,
>> >> >>> > + src->tsc_shift);
>> >> >>> > +
>> >> >>> > + } while (pvclock_read_retry(src, version));
>> >> >>> > +
>> >> >>> > + preempt_enable_notrace();
>> >> >>> > +
>> >> >>> > + tspec.tv_nsec = tspec.tv_nsec + offset;
>> >> >>> > +
>> >> >>> > + spin_unlock(&kvm_ptp_lock);
>> >> >>> > +
>> >> >>> > + if (tspec.tv_nsec >= NSEC_PER_SEC) {
>> >> >>> > + u64 secs = tspec.tv_nsec;
>> >> >>> > +
>> >> >>> > + tspec.tv_nsec = do_div(secs, NSEC_PER_SEC);
>> >> >>> > + tspec.tv_sec += secs;
>> >> >>> > + }
>> >> >>> > +
>> >> >>> > + memcpy(ts, &tspec, sizeof(struct timespec64));
>> >> >>>
>> >> >>> But the whole idea is of improving the time by reading tsc a bit later
>> >> >>> is just weird ... why is it better to provide
>> >> >>>
>> >> >>> tsc + x, time + tsc_delta_to_time(x)
>> >> >>>
>> >> >>> than just
>> >> >>>
>> >> >>> tsc, time
>> >> >>>
>> >> >>> ?
>> >> >>
>> >> >> Because you want to calculate the value of the host realtime clock
>> >> >> at the moment of ptp_kvm_gettime.
>> >> >>
>> >> >> We do:
>> >> >>
>> >> >> 1. kvm_hypercall.
>> >> >> 2. get {sec, nsec, guest_tsc}.
>> >> >> 3. kvm_hypercall returns.
>> >> >> 4. delay = rdtsc() - guest_tsc.
>> >> >>
>> >> >> Where delay is the delta (measured with the TSC) between points 2 and
>> >> >> 4.
>> >> >
>> >> > I see now ... the PTP interface is just not good for our purposes.
>> >>
>> >> There is getcrosststamp() callback in PTP, which seems to be exactly
>> >> what we want when pairing with TSC, so the pvclock delay fixup can be
>> >> dropped when using it.
>> >
>> > What pvclock delay fixup you refer to? The "rdtsc() - clock_offset.tsc"
>> > part?
>>
>> Yes.
>>
>> > You can't drop it, because if you do then your "host realtime
>> > clock read" will be behind by "rdtsc() - clock_offset.tsc" TSC cycles.
>>
>> The TSC read will be some cycles old when the hypercall ends, but that
>> doesn't matter, because we will pass {sec, nsec, guest_tsc} to PTP and
>> PTP should plug them into kernel's realtime clock roughly like this:
>>
>> sec/nsec + (rdtsc() - guest_tsc) * tsc_freq
>>
>> Adding delay to guest_tsc and sec/nsec cannot improve precision.
>> (And will likely degrade it as kvmclock's frequency is incorrect.)
>>
>> > We want the highest precision as possible.
>>
>> I agree, which is why we don't want to lose precision in the delay
>> guesswork because of gettime64().
>
> Sorry the clock difference is 10ns now. So the guest clock is off by _10 ns_
> of the host clock.
That is pretty good.
> You are suggesting to use getcrosststamp instead, to drop the (rdtsc() -
> guest_tsc) part ?
Yes, it results in simpler code, doesn't create dependency on the
dreaded kvmclock, and is the best we can currently do wrt. precision.
Thanks.
