[v2,0/3] x86/time: calibration rendezvous adjustments

Message

Jan Beulich Feb. 1, 2021, 12:41 p.m. UTC

The latter two patches are meant to address a regression reported on
the list under "Problems with APIC on versions 4.9 and later (4.8
works)". In the course of analyzing output from a debugging patch I
ran into another anomaly again, which I thought I should finally try
to address. Hence patch 1.

While looking closely at the corresponding debugging patch'es output I
noticed a suspicious drift between local and master stime: Measured not
very precisely, local was behind master by about 200ms in about half an
hour. Interestingly the recording of ->master_stime (and hence the not
really inexpensive invocation of read_platform_stime()) looks to be
pretty pointless when CONSTANT_TSC - I haven't been able to spot an
actual consumer. IOW the drift may not be a problem, and we might be
able to eliminate the platform timer reads. (When !CONSTANT_TSC, such
drift would get corrected anyway, by local_time_calibration().)

1: change initiation of the calibration timer
2: adjust time recording time_calibration_tsc_rendezvous()
3: don't move TSC backwards in time_calibration_tsc_rendezvous()

Jan