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/cpufreq.h> 16 #include <linux/smp.h> 17 18 #include "cpu.h" 19 20 struct aperfmperf_sample { 21 unsigned int khz; 22 ktime_t time; 23 u64 aperf; 24 u64 mperf; 25 }; 26 27 static DEFINE_PER_CPU(struct aperfmperf_sample, samples); 28 29 #define APERFMPERF_CACHE_THRESHOLD_MS 10 30 #define APERFMPERF_REFRESH_DELAY_MS 10 31 #define APERFMPERF_STALE_THRESHOLD_MS 1000 32 33 /* 34 * aperfmperf_snapshot_khz() 35 * On the current CPU, snapshot APERF, MPERF, and jiffies 36 * unless we already did it within 10ms 37 * calculate kHz, save snapshot 38 */ 39 static void aperfmperf_snapshot_khz(void *dummy) 40 { 41 u64 aperf, aperf_delta; 42 u64 mperf, mperf_delta; 43 struct aperfmperf_sample *s = this_cpu_ptr(&samples); 44 unsigned long flags; 45 46 local_irq_save(flags); 47 rdmsrl(MSR_IA32_APERF, aperf); 48 rdmsrl(MSR_IA32_MPERF, mperf); 49 local_irq_restore(flags); 50 51 aperf_delta = aperf - s->aperf; 52 mperf_delta = mperf - s->mperf; 53 54 /* 55 * There is no architectural guarantee that MPERF 56 * increments faster than we can read it. 57 */ 58 if (mperf_delta == 0) 59 return; 60 61 s->time = ktime_get(); 62 s->aperf = aperf; 63 s->mperf = mperf; 64 s->khz = div64_u64((cpu_khz * aperf_delta), mperf_delta); 65 } 66 67 static bool aperfmperf_snapshot_cpu(int cpu, ktime_t now, bool wait) 68 { 69 s64 time_delta = ktime_ms_delta(now, per_cpu(samples.time, cpu)); 70 71 /* Don't bother re-computing within the cache threshold time. */ 72 if (time_delta < APERFMPERF_CACHE_THRESHOLD_MS) 73 return true; 74 75 smp_call_function_single(cpu, aperfmperf_snapshot_khz, NULL, wait); 76 77 /* Return false if the previous iteration was too long ago. */ 78 return time_delta <= APERFMPERF_STALE_THRESHOLD_MS; 79 } 80 81 unsigned int aperfmperf_get_khz(int cpu) 82 { 83 if (!cpu_khz) 84 return 0; 85 86 if (!static_cpu_has(X86_FEATURE_APERFMPERF)) 87 return 0; 88 89 aperfmperf_snapshot_cpu(cpu, ktime_get(), true); 90 return per_cpu(samples.khz, cpu); 91 } 92 93 void arch_freq_prepare_all(void) 94 { 95 ktime_t now = ktime_get(); 96 bool wait = false; 97 int cpu; 98 99 if (!cpu_khz) 100 return; 101 102 if (!static_cpu_has(X86_FEATURE_APERFMPERF)) 103 return; 104 105 for_each_online_cpu(cpu) 106 if (!aperfmperf_snapshot_cpu(cpu, now, false)) 107 wait = true; 108 109 if (wait) 110 msleep(APERFMPERF_REFRESH_DELAY_MS); 111 } 112 113 unsigned int arch_freq_get_on_cpu(int cpu) 114 { 115 if (!cpu_khz) 116 return 0; 117 118 if (!static_cpu_has(X86_FEATURE_APERFMPERF)) 119 return 0; 120 121 if (aperfmperf_snapshot_cpu(cpu, ktime_get(), true)) 122 return per_cpu(samples.khz, cpu); 123 124 msleep(APERFMPERF_REFRESH_DELAY_MS); 125 smp_call_function_single(cpu, aperfmperf_snapshot_khz, NULL, 1); 126 127 return per_cpu(samples.khz, cpu); 128 } 129