.... and Intel's SDM actually does also define MSRs for some models for the TSC frequency as well... which gives the correct value for my home box...

i threw those together, but let me see how many of these from the SDM are in the lab

gist.github.com/jasonbking/744fdf69493219b717269c0ed76012f6 if anyone wants to check out a userland tester of it

(might need to exclude VMs)

ahh.. so actually it was a rounding error because the host's tsc frequency was measured

at least the vmm kernel module on illumos actually emulates the MSR with the host's tsc frequency

(so when exact, the guest gets an exact value)

jbk: you have Nehalem and Ivy/Sandy Bridge in there but not Westmere?

added INTC_MODEL_WESTMERE* to the INTC_MODEL_NEHALEM* and came up with a number that matched the calibration..

ahh... the SDM is annoying in how it documents them...

I was looking at all the platforms that documented the MSR_PLATFORM_INFO msr

except westmere is 'we also do all these others in this other table'

but that does appear to be valid for them as well..

at least 0x1, 1e, 1f, 2e (need to look at x86_archext.h to see what those are)

My understanding is that Westmere is more or less a process shrink of Nehalem so not surprising the docs are a little vaguer on it.

oh, wait, off by a factor of 1000. Your program (with Westmere added) said "frequency = 2533270 MHz" (should be kHz); system calibration said 2533260631 Hz

terahertz!

oh grr..

let me update and fix...

(that was just to try to make it a bit easier to read)

sommerfeld: gist updated

i went through all the families and verified the 'see <other table>' and added all of those

(which really is i guess everything recent)

separate from this though, I'm wondering if the missing model #defines should be a separate ticket

(and if the names i used are correct or if there's better ones -- it feels like the model specifics can be a bit scattered in the docs)

frequency = 2533270000 Hz (2.533 GHz)

does that sound correct for your CPU?

that matches the PIT calibration on the Westmere.

ok

i guess i'll go ahead and create a ticket

though might to a little more discovery before I put up a PR

(trying that on various intel systems)

(posted above: PIT calibratio was 2533260631 Hz which is within 3ppm of the nominal value).

on the other hand, on the much newer i5-1235U, the MSRs say 2500000000 Hz while CPUID says 2496000000 Hz and HPET calibration says 2495993673 Hz

which is more like 1600ppm error, well outside of what NTP can tolerate.

so one datapoint says cpuid should be preferred to MSR.

given the CPUID based approach is newer (it also supports > 8 bits of precision), that seems reasonable...

i should probably have it output the family value

just to make it easier..

oh that's rather annoying

i had typed up everything for the issue, clicked create

and the AI spam checker popped up, and apparently ignored everthing, losing all of the text...

(I blame AI for needing that)

i guess i'll try later... need to do a few things outside before it rains

worked that time

and partial victory for my other stuff :)

also been quite a while since I've crimped RJ45..a bit rusty :P