On Mon, Jul 31, 2017 at 3:58 PM, Mike Belopuhov <m...@belopuhov.com> wrote: > On Mon, Jul 31, 2017 at 09:48 +0800, Adam Steen wrote: >> Ted Unangst wrote: >> > we don't currently export this info, but we could add some sysctls. there's >> > some cpufeatures stuff there, but generally stuff isn't exported until >> > somebody finds a use for it... it shouldn't be too hard to add something to >> > amd64/machdep.c sysctl if you're interested. >> >> I am interested, as i need the info, i will look into it and hopefully >> come back with a patch. > > This is a bad idea because TSC as the time source is only usable > by OpenBSD on Skylake and Kaby Lake CPUs since they encode the TSC > frequency in the CPUID. All older CPUs have their TSCs measured > against the PIT. Currently the measurement done by the kernel isn't > very precise and if TSC is selected as a timecounter, the machine > would be gaining time on a pace that cannot be corrected by our NTP > daemon. (IIRC, about an hour a day on my Haswell running with NTP). > > To be able to use TSC as a timecounter source on OpenBSD or Solo5 > you'd have to improve the in-kernel measurement of the TSC frequency > first. I've tried to perform 10 measurements and take an average and > it does improve accuracy, however I believe we need to poach another > bit from Linux and re-calibrate TSC via HPET: > > > http://elixir.free-electrons.com/linux/v4.12.4/source/arch/x86/kernel/tsc.c#L409 > > I think this is the most sane thing we can do. Here's a complete > procedure that Linux kernel undertakes: > > > http://elixir.free-electrons.com/linux/v4.12.4/source/arch/x86/kernel/tsc.c#L751 > > Regards, > Mike
Hi Mike/All I would like to improve the accuracy of TSC frequency calibration as Mike B. describes above. I initially thought the calibration would take place at line 470 of amd64/identcpu.c (https://cvsweb.openbsd.org/cgi-bin/cvsweb/src/sys/arch/amd64/amd64/identcpu.c?annotate=1.87) But I looked into using the acpihpet directly but it is never exposed outside of acpihpet.c. Could someone point me to were if would be appropriate to complete this calibration and how to use the acpihpet? (Will it need to be exposed like i8254_delay/delay_func/delay in machdep.c and cpu.h) Lastly should the calibration be done using both delay(i8254 pit) and hpet timers similar to Linux described above or just using the hpet? Cheers Adam ps i have a diff in my tree exporting tsc_freq and invariant_tsc via sysctl which is what Ted is describing above too.
