The ACPI Power Management Timer is "a 24-bit or 32-bit fixed rate free running
count-up timer that runs off a 3.579545 MHz clock" (ACPI specification, section
4.8.3.3; the rate is PMTMR_FREQ). vpmtmr_handler() returned the uptime divided
by that period held as an sbintime_t, and the period, 2^32 / 3579545 = 1199.864
sbintime units, is truncated to 1199 when stored, so the timer runs
1199.864 / 1199 - 1 = 720 ppm fast. A guest that keeps time by the PM timer
gains about a minute a day.

Compute the count as seconds times the frequency, which is exact: split the
32.32 fixed-point delta into its whole and fractional seconds and multiply
each by PMTMR_FREQ.

Signed-off-by: Wanpeng Qian <wanpengqian@gmail.com>
Sponsored by: keelos.dev

The error is deterministic and follows from the spec. The ACPI specification
(section 4.8.3.3, Power Management Timer) defines the timer as running off a
3.579545 MHz clock; FreeBSD's PMTMR_FREQ is 3579545. The old code stored the
period as bttosbt(FREQ2BT(PMTMR_FREQ)) = floor(2^32 / 3579545) = 1199 sbintime
units, dropping the true 1199.864, and divided the delta by it.
1199.864 / 1199 = 1.000720, i.e. 720 ppm fast. Multiplying by PMTMR_FREQ
instead is exact.

Effect on a guest, observed on real hardware (a bhyve host whose TSC is 3.600
GHz): Linux calibrates its TSC against the PM timer at boot. Before the fix it
reported "Refined TSC clocksource calibration: 3597.410 MHz" on every boot
(3600 / 1.000720 = 3597.41), and its clocksource watchdog then saw the TSC
about 346 us off per 0.5 s against the HPET, near the 500 us it tolerates; with
a little jitter the guest marked the TSC unstable and fell back to the HPET,
after which every clock read is a VM exit. After the fix the guest calibrates
to 3600.000 MHz and keeps the TSC as its clocksource.

A FreeBSD guest's ACPI-fast timecounter shows the same 720 ppm before and none
after. Only the returned count changes.