| Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame] | 1 | /* |
| 2 | * x86 APERF/MPERF KHz calculation for |
| 3 | * /sys/.../cpufreq/scaling_cur_freq |
| 4 | * |
| 5 | * Copyright (C) 2017 Intel Corp. |
| 6 | * Author: Len Brown <len.brown@intel.com> |
| 7 | * |
| 8 | * This file is licensed under GPLv2. |
| 9 | */ |
| 10 | |
| 11 | #include <linux/delay.h> |
| 12 | #include <linux/ktime.h> |
| 13 | #include <linux/math64.h> |
| 14 | #include <linux/percpu.h> |
| 15 | #include <linux/smp.h> |
| 16 | |
| 17 | #include "cpu.h" |
| 18 | |
| 19 | struct aperfmperf_sample { |
| 20 | unsigned int khz; |
| 21 | ktime_t time; |
| 22 | u64 aperf; |
| 23 | u64 mperf; |
| 24 | }; |
| 25 | |
| 26 | static DEFINE_PER_CPU(struct aperfmperf_sample, samples); |
| 27 | |
| 28 | #define APERFMPERF_CACHE_THRESHOLD_MS 10 |
| 29 | #define APERFMPERF_REFRESH_DELAY_MS 10 |
| 30 | #define APERFMPERF_STALE_THRESHOLD_MS 1000 |
| 31 | |
| 32 | /* |
| 33 | * aperfmperf_snapshot_khz() |
| 34 | * On the current CPU, snapshot APERF, MPERF, and jiffies |
| 35 | * unless we already did it within 10ms |
| 36 | * calculate kHz, save snapshot |
| 37 | */ |
| 38 | static void aperfmperf_snapshot_khz(void *dummy) |
| 39 | { |
| 40 | u64 aperf, aperf_delta; |
| 41 | u64 mperf, mperf_delta; |
| 42 | struct aperfmperf_sample *s = this_cpu_ptr(&samples); |
| 43 | unsigned long flags; |
| 44 | |
| 45 | local_irq_save(flags); |
| 46 | rdmsrl(MSR_IA32_APERF, aperf); |
| 47 | rdmsrl(MSR_IA32_MPERF, mperf); |
| 48 | local_irq_restore(flags); |
| 49 | |
| 50 | aperf_delta = aperf - s->aperf; |
| 51 | mperf_delta = mperf - s->mperf; |
| 52 | |
| 53 | /* |
| 54 | * There is no architectural guarantee that MPERF |
| 55 | * increments faster than we can read it. |
| 56 | */ |
| 57 | if (mperf_delta == 0) |
| 58 | return; |
| 59 | |
| 60 | s->time = ktime_get(); |
| 61 | s->aperf = aperf; |
| 62 | s->mperf = mperf; |
| 63 | s->khz = div64_u64((cpu_khz * aperf_delta), mperf_delta); |
| 64 | } |
| 65 | |
| 66 | static bool aperfmperf_snapshot_cpu(int cpu, ktime_t now, bool wait) |
| 67 | { |
| 68 | s64 time_delta = ktime_ms_delta(now, per_cpu(samples.time, cpu)); |
| 69 | |
| 70 | /* Don't bother re-computing within the cache threshold time. */ |
| 71 | if (time_delta < APERFMPERF_CACHE_THRESHOLD_MS) |
| 72 | return true; |
| 73 | |
| 74 | smp_call_function_single(cpu, aperfmperf_snapshot_khz, NULL, wait); |
| 75 | |
| 76 | /* Return false if the previous iteration was too long ago. */ |
| 77 | return time_delta <= APERFMPERF_STALE_THRESHOLD_MS; |
| 78 | } |
| 79 | |
| 80 | unsigned int aperfmperf_get_khz(int cpu) |
| 81 | { |
| 82 | if (!cpu_khz) |
| 83 | return 0; |
| 84 | |
| 85 | if (!static_cpu_has(X86_FEATURE_APERFMPERF)) |
| 86 | return 0; |
| 87 | |
| 88 | aperfmperf_snapshot_cpu(cpu, ktime_get(), true); |
| 89 | return per_cpu(samples.khz, cpu); |
| 90 | } |
| 91 | |
| 92 | void arch_freq_prepare_all(void) |
| 93 | { |
| 94 | ktime_t now = ktime_get(); |
| 95 | bool wait = false; |
| 96 | int cpu; |
| 97 | |
| 98 | if (!cpu_khz) |
| 99 | return; |
| 100 | |
| 101 | if (!static_cpu_has(X86_FEATURE_APERFMPERF)) |
| 102 | return; |
| 103 | |
| 104 | for_each_online_cpu(cpu) |
| 105 | if (!aperfmperf_snapshot_cpu(cpu, now, false)) |
| 106 | wait = true; |
| 107 | |
| 108 | if (wait) |
| 109 | msleep(APERFMPERF_REFRESH_DELAY_MS); |
| 110 | } |
| 111 | |
| 112 | unsigned int arch_freq_get_on_cpu(int cpu) |
| 113 | { |
| 114 | if (!cpu_khz) |
| 115 | return 0; |
| 116 | |
| 117 | if (!static_cpu_has(X86_FEATURE_APERFMPERF)) |
| 118 | return 0; |
| 119 | |
| 120 | if (aperfmperf_snapshot_cpu(cpu, ktime_get(), true)) |
| 121 | return per_cpu(samples.khz, cpu); |
| 122 | |
| 123 | msleep(APERFMPERF_REFRESH_DELAY_MS); |
| 124 | smp_call_function_single(cpu, aperfmperf_snapshot_khz, NULL, 1); |
| 125 | |
| 126 | return per_cpu(samples.khz, cpu); |
| 127 | } |