> From: "Ted Unangst" <t...@tedunangst.com>
> Date: Sat, 14 May 2022 20:23:39 -0400
>
> The cpu hz sensor is more accurate and updates faster than than the value
> currently used for hw.cpuspeed. So return that value (scaled).
>
> This doesn't set cpuspeed directly because the acpi does that and it's hard
> to create a whole system of priority overrides. I think it's simpler and
> maybe even better to track every value, and then return the best one
> available.
At this point, I don't want this. Right now, it is actually useful
that hw.cpuspeed represents the requested P-state whereas the sensors
give us the "effective" frequency the core is actually running at.
And that is useful for evaluating what is actually happening.
> Index: identcpu.c
> ===================================================================
> RCS file: /home/cvs/src/sys/arch/amd64/amd64/identcpu.c,v
> retrieving revision 1.124
> diff -u -p -r1.124 identcpu.c
> --- identcpu.c 26 Apr 2022 10:48:20 -0000 1.124
> +++ identcpu.c 15 May 2022 00:15:39 -0000
> @@ -64,6 +64,7 @@ void cpu_check_vmm_cap(struct cpu_info *
> /* sysctl wants this. */
> char cpu_model[48];
> int cpuspeed;
> +uint64_t sensorspeed;
>
> int amd64_has_xcrypt;
> #ifdef CRYPTO
> @@ -244,6 +245,8 @@ int
> cpu_amd64speed(int *freq)
> {
> *freq = cpuspeed;
> + if (sensorspeed)
> + *freq = sensorspeed / 1000000000000ULL;
> return (0);
> }
>
> @@ -337,6 +340,8 @@ cpu_hz_update_sensor(void *args)
> val = (adelta * 1000000) / mdelta * tsc_frequency;
> val = ((val + FREQ_50MHZ / 2) / FREQ_50MHZ) * FREQ_50MHZ;
> ci->ci_hz_sensor.value = val;
> + if (CPU_IS_PRIMARY(ci))
> + sensorspeed = val;
> }
>
> atomic_clearbits_int(&curproc->p_flag, P_CPUPEG);
>
>
