linux/arch/x86/events/amd/power.c
<<
>>
Prefs
   1/*
   2 * Performance events - AMD Processor Power Reporting Mechanism
   3 *
   4 * Copyright (C) 2016 Advanced Micro Devices, Inc.
   5 *
   6 * Author: Huang Rui <ray.huang@amd.com>
   7 *
   8 * This program is free software; you can redistribute it and/or modify
   9 * it under the terms of the GNU General Public License version 2 as
  10 * published by the Free Software Foundation.
  11 */
  12
  13#include <linux/module.h>
  14#include <linux/slab.h>
  15#include <linux/perf_event.h>
  16#include <asm/cpu_device_id.h>
  17#include "../perf_event.h"
  18
  19#define MSR_F15H_CU_PWR_ACCUMULATOR     0xc001007a
  20#define MSR_F15H_CU_MAX_PWR_ACCUMULATOR 0xc001007b
  21#define MSR_F15H_PTSC                   0xc0010280
  22
  23/* Event code: LSB 8 bits, passed in attr->config any other bit is reserved. */
  24#define AMD_POWER_EVENT_MASK            0xFFULL
  25
  26/*
  27 * Accumulated power status counters.
  28 */
  29#define AMD_POWER_EVENTSEL_PKG          1
  30
  31/*
  32 * The ratio of compute unit power accumulator sample period to the
  33 * PTSC period.
  34 */
  35static unsigned int cpu_pwr_sample_ratio;
  36
  37/* Maximum accumulated power of a compute unit. */
  38static u64 max_cu_acc_power;
  39
  40static struct pmu pmu_class;
  41
  42/*
  43 * Accumulated power represents the sum of each compute unit's (CU) power
  44 * consumption. On any core of each CU we read the total accumulated power from
  45 * MSR_F15H_CU_PWR_ACCUMULATOR. cpu_mask represents CPU bit map of all cores
  46 * which are picked to measure the power for the CUs they belong to.
  47 */
  48static cpumask_t cpu_mask;
  49
  50static void event_update(struct perf_event *event)
  51{
  52        struct hw_perf_event *hwc = &event->hw;
  53        u64 prev_pwr_acc, new_pwr_acc, prev_ptsc, new_ptsc;
  54        u64 delta, tdelta;
  55
  56        prev_pwr_acc = hwc->pwr_acc;
  57        prev_ptsc = hwc->ptsc;
  58        rdmsrl(MSR_F15H_CU_PWR_ACCUMULATOR, new_pwr_acc);
  59        rdmsrl(MSR_F15H_PTSC, new_ptsc);
  60
  61        /*
  62         * Calculate the CU power consumption over a time period, the unit of
  63         * final value (delta) is micro-Watts. Then add it to the event count.
  64         */
  65        if (new_pwr_acc < prev_pwr_acc) {
  66                delta = max_cu_acc_power + new_pwr_acc;
  67                delta -= prev_pwr_acc;
  68        } else
  69                delta = new_pwr_acc - prev_pwr_acc;
  70
  71        delta *= cpu_pwr_sample_ratio * 1000;
  72        tdelta = new_ptsc - prev_ptsc;
  73
  74        do_div(delta, tdelta);
  75        local64_add(delta, &event->count);
  76}
  77
  78static void __pmu_event_start(struct perf_event *event)
  79{
  80        if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
  81                return;
  82
  83        event->hw.state = 0;
  84
  85        rdmsrl(MSR_F15H_PTSC, event->hw.ptsc);
  86        rdmsrl(MSR_F15H_CU_PWR_ACCUMULATOR, event->hw.pwr_acc);
  87}
  88
  89static void pmu_event_start(struct perf_event *event, int mode)
  90{
  91        __pmu_event_start(event);
  92}
  93
  94static void pmu_event_stop(struct perf_event *event, int mode)
  95{
  96        struct hw_perf_event *hwc = &event->hw;
  97
  98        /* Mark event as deactivated and stopped. */
  99        if (!(hwc->state & PERF_HES_STOPPED))
 100                hwc->state |= PERF_HES_STOPPED;
 101
 102        /* Check if software counter update is necessary. */
 103        if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
 104                /*
 105                 * Drain the remaining delta count out of an event
 106                 * that we are disabling:
 107                 */
 108                event_update(event);
 109                hwc->state |= PERF_HES_UPTODATE;
 110        }
 111}
 112
 113static int pmu_event_add(struct perf_event *event, int mode)
 114{
 115        struct hw_perf_event *hwc = &event->hw;
 116
 117        hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 118
 119        if (mode & PERF_EF_START)
 120                __pmu_event_start(event);
 121
 122        return 0;
 123}
 124
 125static void pmu_event_del(struct perf_event *event, int flags)
 126{
 127        pmu_event_stop(event, PERF_EF_UPDATE);
 128}
 129
 130static int pmu_event_init(struct perf_event *event)
 131{
 132        u64 cfg = event->attr.config & AMD_POWER_EVENT_MASK;
 133
 134        /* Only look at AMD power events. */
 135        if (event->attr.type != pmu_class.type)
 136                return -ENOENT;
 137
 138        /* Unsupported modes and filters. */
 139        if (event->attr.sample_period)
 140                return -EINVAL;
 141
 142        if (cfg != AMD_POWER_EVENTSEL_PKG)
 143                return -EINVAL;
 144
 145        return 0;
 146}
 147
 148static void pmu_event_read(struct perf_event *event)
 149{
 150        event_update(event);
 151}
 152
 153static ssize_t
 154get_attr_cpumask(struct device *dev, struct device_attribute *attr, char *buf)
 155{
 156        return cpumap_print_to_pagebuf(true, buf, &cpu_mask);
 157}
 158
 159static DEVICE_ATTR(cpumask, S_IRUGO, get_attr_cpumask, NULL);
 160
 161static struct attribute *pmu_attrs[] = {
 162        &dev_attr_cpumask.attr,
 163        NULL,
 164};
 165
 166static struct attribute_group pmu_attr_group = {
 167        .attrs = pmu_attrs,
 168};
 169
 170/*
 171 * Currently it only supports to report the power of each
 172 * processor/package.
 173 */
 174EVENT_ATTR_STR(power-pkg, power_pkg, "event=0x01");
 175
 176EVENT_ATTR_STR(power-pkg.unit, power_pkg_unit, "mWatts");
 177
 178/* Convert the count from micro-Watts to milli-Watts. */
 179EVENT_ATTR_STR(power-pkg.scale, power_pkg_scale, "1.000000e-3");
 180
 181static struct attribute *events_attr[] = {
 182        EVENT_PTR(power_pkg),
 183        EVENT_PTR(power_pkg_unit),
 184        EVENT_PTR(power_pkg_scale),
 185        NULL,
 186};
 187
 188static struct attribute_group pmu_events_group = {
 189        .name   = "events",
 190        .attrs  = events_attr,
 191};
 192
 193PMU_FORMAT_ATTR(event, "config:0-7");
 194
 195static struct attribute *formats_attr[] = {
 196        &format_attr_event.attr,
 197        NULL,
 198};
 199
 200static struct attribute_group pmu_format_group = {
 201        .name   = "format",
 202        .attrs  = formats_attr,
 203};
 204
 205static const struct attribute_group *attr_groups[] = {
 206        &pmu_attr_group,
 207        &pmu_format_group,
 208        &pmu_events_group,
 209        NULL,
 210};
 211
 212static struct pmu pmu_class = {
 213        .attr_groups    = attr_groups,
 214        /* system-wide only */
 215        .task_ctx_nr    = perf_invalid_context,
 216        .event_init     = pmu_event_init,
 217        .add            = pmu_event_add,
 218        .del            = pmu_event_del,
 219        .start          = pmu_event_start,
 220        .stop           = pmu_event_stop,
 221        .read           = pmu_event_read,
 222        .capabilities   = PERF_PMU_CAP_NO_EXCLUDE,
 223};
 224
 225static int power_cpu_exit(unsigned int cpu)
 226{
 227        int target;
 228
 229        if (!cpumask_test_and_clear_cpu(cpu, &cpu_mask))
 230                return 0;
 231
 232        /*
 233         * Find a new CPU on the same compute unit, if was set in cpumask
 234         * and still some CPUs on compute unit. Then migrate event and
 235         * context to new CPU.
 236         */
 237        target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
 238        if (target < nr_cpumask_bits) {
 239                cpumask_set_cpu(target, &cpu_mask);
 240                perf_pmu_migrate_context(&pmu_class, cpu, target);
 241        }
 242        return 0;
 243}
 244
 245static int power_cpu_init(unsigned int cpu)
 246{
 247        int target;
 248
 249        /*
 250         * 1) If any CPU is set at cpu_mask in the same compute unit, do
 251         * nothing.
 252         * 2) If no CPU is set at cpu_mask in the same compute unit,
 253         * set current ONLINE CPU.
 254         *
 255         * Note: if there is a CPU aside of the new one already in the
 256         * sibling mask, then it is also in cpu_mask.
 257         */
 258        target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
 259        if (target >= nr_cpumask_bits)
 260                cpumask_set_cpu(cpu, &cpu_mask);
 261        return 0;
 262}
 263
 264static const struct x86_cpu_id cpu_match[] = {
 265        X86_MATCH_VENDOR_FAM(AMD, 0x15, NULL),
 266        {},
 267};
 268
 269static int __init amd_power_pmu_init(void)
 270{
 271        int ret;
 272
 273        if (!x86_match_cpu(cpu_match))
 274                return -ENODEV;
 275
 276        if (!boot_cpu_has(X86_FEATURE_ACC_POWER))
 277                return -ENODEV;
 278
 279        cpu_pwr_sample_ratio = cpuid_ecx(0x80000007);
 280
 281        if (rdmsrl_safe(MSR_F15H_CU_MAX_PWR_ACCUMULATOR, &max_cu_acc_power)) {
 282                pr_err("Failed to read max compute unit power accumulator MSR\n");
 283                return -ENODEV;
 284        }
 285
 286
 287        cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_POWER_ONLINE,
 288                          "perf/x86/amd/power:online",
 289                          power_cpu_init, power_cpu_exit);
 290
 291        ret = perf_pmu_register(&pmu_class, "power", -1);
 292        if (WARN_ON(ret)) {
 293                pr_warn("AMD Power PMU registration failed\n");
 294                return ret;
 295        }
 296
 297        pr_info("AMD Power PMU detected\n");
 298        return ret;
 299}
 300module_init(amd_power_pmu_init);
 301
 302static void __exit amd_power_pmu_exit(void)
 303{
 304        cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_AMD_POWER_ONLINE);
 305        perf_pmu_unregister(&pmu_class);
 306}
 307module_exit(amd_power_pmu_exit);
 308
 309MODULE_AUTHOR("Huang Rui <ray.huang@amd.com>");
 310MODULE_DESCRIPTION("AMD Processor Power Reporting Mechanism");
 311MODULE_LICENSE("GPL v2");
 312