cpufreq for apple silicon macs

[shoenig]

Previously on darwin, it was safe to assume that hw.cpufrequency
would be defined in sysctl. On the new M1 macs instead use the base
frequency hw.tbfrequency and kern.clockrate[hz]. The computed
value appears to be 2.4 GHz, which is I believe is the correct base
clock.

Before:

 m1  cpu  master % go test
--- FAIL: TestCpuInfo (0.00s)
    cpu_test.go:117: error no such file or directory
    cpu_test.go:120: could not get CPU Info
FAIL
exit status 1
FAIL	github.com/shirou/gopsutil/v3/cpu	0.371s

After:

 m1  cpu  m1freq % go test
PASS
ok  	github.com/shirou/gopsutil/v3/cpu	0.248s

[Lomanic]

Thanks for this PR, I didn't test it yet (hoping hw.tbfrequency and kern.clockrate are also available on amd64 as I don't have access to any Apple ARM machine).

[tmm1]

hoping hw.tbfrequency and kern.clockrate are also available on amd64

Yes they appear to be.

[Lomanic]

Having tested with the following small script on an amd64 hackintosh VM, I'm just confused by the result

package main

import (
        "fmt"

        "golang.org/x/sys/unix"
)

func getFrequencyIntel() (uint64, error) {
        // This works for Intel macs.
        return unix.SysctlUint64("hw.cpufrequency")
}

func getFrequencyArm() (uint64, error) {
        // On Apple Silicon fallback to hw.tbfrequency and kern.clockrate[hz]
        tbFreq, err := unix.SysctlUint64("hw.tbfrequency")
        if err != nil {
                return 0, err
        }

        kernClockrate, err := unix.SysctlClockinfo("kern.clockrate")
        if err != nil {
                return 0, err
        }

        return uint64(kernClockrate.Hz) * tbFreq, nil
}

func main() {
        fmt.Println(getFrequencyIntel())
        fmt.Println(getFrequencyArm())
}

As it returns

2500000000 <nil>                                                                                                                                                                                                                                                               
100000000000 <nil>

Also, is python psutil also failing with psutil.cpu_freq() on Apple M1?

[shoenig]

Yeah @Lomanic, I'm still trying to dig through and make sense of what Apple is doing here. There is probably a more correct way to get this value that is consistent across Intel/ARM macs but maybe only with cgo. If anyone else wants to play along, https://github.com/knightsc/darwin-xnu/blob/master/pexpert/arm/pe_identify_machine.c#L45

[shoenig]

Also, is python psutil also failing with psutil.cpu_freq() on Apple M1?

I don't have Python setup, but I would suspect it fails since they use the absent sysctl values.
https://github.com/giampaolo/psutil/blob/master/psutil/_psutil_osx.c#L693

[Lomanic]

Darwin kernel is under the Apple Public Source License which I understand is not compatible with ours so linking against is not an option. Porting its code (in Go?) looks heavily involved.

Reported the issue for psutil in giampaolo/psutil#1892.

[giampaolo]

Also see giampaolo/psutil#1892 (comment).

[shirou]

FYI: psutil fixed by using sysctl. I don't have an Apple M1. So I hope some other kind person make a PR.

[Lomanic]

FYI: psutil fixed by using sysctl.

For the record, it is not fixed in psutil, see giampaolo/psutil#1892 (comment). And @giampaolo concluded the same thing as #999 (comment) in giampaolo/psutil#1892 (comment): these values are not linked and don't seem to represent the same thing.

For the moment, it seems impossible to get the nominal CPU frequency on Apple M1 devices.

[shoenig]

Indeed, I'll go ahead and close this PR