1 // SPDX-License-Identifier: GPL-2.0-only
3 * Support Intel/AMD RAPL energy consumption counters
4 * Copyright (C) 2013 Google, Inc., Stephane Eranian
6 * Intel RAPL interface is specified in the IA-32 Manual Vol3b
7 * section 14.7.1 (September 2013)
9 * AMD RAPL interface for Fam17h is described in the public PPR:
10 * https://bugzilla.kernel.org/show_bug.cgi?id=206537
12 * RAPL provides more controls than just reporting energy consumption
13 * however here we only expose the 3 energy consumption free running
14 * counters (pp0, pkg, dram).
16 * Each of those counters increments in a power unit defined by the
17 * RAPL_POWER_UNIT MSR. On SandyBridge, this unit is 1/(2^16) Joules
20 * Counter to rapl events mappings:
22 * pp0 counter: consumption of all physical cores (power plane 0)
23 * event: rapl_energy_cores
26 * pkg counter: consumption of the whole processor package
27 * event: rapl_energy_pkg
30 * dram counter: consumption of the dram domain (servers only)
31 * event: rapl_energy_dram
34 * gpu counter: consumption of the builtin-gpu domain (client only)
35 * event: rapl_energy_gpu
38 * psys counter: consumption of the builtin-psys domain (client only)
39 * event: rapl_energy_psys
42 * We manage those counters as free running (read-only). They may be
43 * use simultaneously by other tools, such as turbostat.
45 * The events only support system-wide mode counting. There is no
46 * sampling support because it does not make sense and is not
47 * supported by the RAPL hardware.
49 * Because we want to avoid floating-point operations in the kernel,
50 * the events are all reported in fixed point arithmetic (32.32).
51 * Tools must adjust the counts to convert them to Watts using
52 * the duration of the measurement. Tools may use a function such as
53 * ldexp(raw_count, -32);
56 #define pr_fmt(fmt) "RAPL PMU: " fmt
58 #include <linux/module.h>
59 #include <linux/slab.h>
60 #include <linux/perf_event.h>
61 #include <linux/nospec.h>
62 #include <asm/cpu_device_id.h>
63 #include <asm/intel-family.h>
64 #include "perf_event.h"
67 MODULE_LICENSE("GPL");
70 * RAPL energy status counters
72 enum perf_rapl_events {
73 PERF_RAPL_PP0 = 0, /* all cores */
74 PERF_RAPL_PKG, /* entire package */
75 PERF_RAPL_RAM, /* DRAM */
76 PERF_RAPL_PP1, /* gpu */
77 PERF_RAPL_PSYS, /* psys */
80 NR_RAPL_DOMAINS = PERF_RAPL_MAX,
83 static const char *const rapl_domain_names[NR_RAPL_DOMAINS] __initconst = {
92 * event code: LSB 8 bits, passed in attr->config
93 * any other bit is reserved
95 #define RAPL_EVENT_MASK 0xFFULL
97 #define DEFINE_RAPL_FORMAT_ATTR(_var, _name, _format) \
98 static ssize_t __rapl_##_var##_show(struct kobject *kobj, \
99 struct kobj_attribute *attr, \
102 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
103 return sprintf(page, _format "\n"); \
105 static struct kobj_attribute format_attr_##_var = \
106 __ATTR(_name, 0444, __rapl_##_var##_show, NULL)
108 #define RAPL_CNTR_WIDTH 32
110 #define RAPL_EVENT_ATTR_STR(_name, v, str) \
111 static struct perf_pmu_events_attr event_attr_##v = { \
112 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
121 struct list_head active_list;
123 ktime_t timer_interval;
124 struct hrtimer hrtimer;
130 struct rapl_pmu *pmus[];
133 enum rapl_unit_quirk {
134 RAPL_UNIT_QUIRK_NONE,
135 RAPL_UNIT_QUIRK_INTEL_HSW,
136 RAPL_UNIT_QUIRK_INTEL_SPR,
140 struct perf_msr *rapl_msrs;
141 unsigned long events;
142 unsigned int msr_power_unit;
143 enum rapl_unit_quirk unit_quirk;
146 /* 1/2^hw_unit Joule */
147 static int rapl_hw_unit[NR_RAPL_DOMAINS] __read_mostly;
148 static struct rapl_pmus *rapl_pmus;
149 static cpumask_t rapl_cpu_mask;
150 static unsigned int rapl_cntr_mask;
151 static u64 rapl_timer_ms;
152 static struct perf_msr *rapl_msrs;
154 static inline struct rapl_pmu *cpu_to_rapl_pmu(unsigned int cpu)
156 unsigned int dieid = topology_logical_die_id(cpu);
159 * The unsigned check also catches the '-1' return value for non
160 * existent mappings in the topology map.
162 return dieid < rapl_pmus->maxdie ? rapl_pmus->pmus[dieid] : NULL;
165 static inline u64 rapl_read_counter(struct perf_event *event)
168 rdmsrl(event->hw.event_base, raw);
172 static inline u64 rapl_scale(u64 v, int cfg)
174 if (cfg > NR_RAPL_DOMAINS) {
175 pr_warn("Invalid domain %d, failed to scale data\n", cfg);
179 * scale delta to smallest unit (1/2^32)
180 * users must then scale back: count * 1/(1e9*2^32) to get Joules
181 * or use ldexp(count, -32).
182 * Watts = Joules/Time delta
184 return v << (32 - rapl_hw_unit[cfg - 1]);
187 static u64 rapl_event_update(struct perf_event *event)
189 struct hw_perf_event *hwc = &event->hw;
190 u64 prev_raw_count, new_raw_count;
192 int shift = RAPL_CNTR_WIDTH;
195 prev_raw_count = local64_read(&hwc->prev_count);
196 rdmsrl(event->hw.event_base, new_raw_count);
198 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
199 new_raw_count) != prev_raw_count) {
205 * Now we have the new raw value and have updated the prev
206 * timestamp already. We can now calculate the elapsed delta
207 * (event-)time and add that to the generic event.
209 * Careful, not all hw sign-extends above the physical width
212 delta = (new_raw_count << shift) - (prev_raw_count << shift);
215 sdelta = rapl_scale(delta, event->hw.config);
217 local64_add(sdelta, &event->count);
219 return new_raw_count;
222 static void rapl_start_hrtimer(struct rapl_pmu *pmu)
224 hrtimer_start(&pmu->hrtimer, pmu->timer_interval,
225 HRTIMER_MODE_REL_PINNED);
228 static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
230 struct rapl_pmu *pmu = container_of(hrtimer, struct rapl_pmu, hrtimer);
231 struct perf_event *event;
235 return HRTIMER_NORESTART;
237 raw_spin_lock_irqsave(&pmu->lock, flags);
239 list_for_each_entry(event, &pmu->active_list, active_entry)
240 rapl_event_update(event);
242 raw_spin_unlock_irqrestore(&pmu->lock, flags);
244 hrtimer_forward_now(hrtimer, pmu->timer_interval);
246 return HRTIMER_RESTART;
249 static void rapl_hrtimer_init(struct rapl_pmu *pmu)
251 struct hrtimer *hr = &pmu->hrtimer;
253 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
254 hr->function = rapl_hrtimer_handle;
257 static void __rapl_pmu_event_start(struct rapl_pmu *pmu,
258 struct perf_event *event)
260 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
265 list_add_tail(&event->active_entry, &pmu->active_list);
267 local64_set(&event->hw.prev_count, rapl_read_counter(event));
270 if (pmu->n_active == 1)
271 rapl_start_hrtimer(pmu);
274 static void rapl_pmu_event_start(struct perf_event *event, int mode)
276 struct rapl_pmu *pmu = event->pmu_private;
279 raw_spin_lock_irqsave(&pmu->lock, flags);
280 __rapl_pmu_event_start(pmu, event);
281 raw_spin_unlock_irqrestore(&pmu->lock, flags);
284 static void rapl_pmu_event_stop(struct perf_event *event, int mode)
286 struct rapl_pmu *pmu = event->pmu_private;
287 struct hw_perf_event *hwc = &event->hw;
290 raw_spin_lock_irqsave(&pmu->lock, flags);
292 /* mark event as deactivated and stopped */
293 if (!(hwc->state & PERF_HES_STOPPED)) {
294 WARN_ON_ONCE(pmu->n_active <= 0);
296 if (pmu->n_active == 0)
297 hrtimer_cancel(&pmu->hrtimer);
299 list_del(&event->active_entry);
301 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
302 hwc->state |= PERF_HES_STOPPED;
305 /* check if update of sw counter is necessary */
306 if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
308 * Drain the remaining delta count out of a event
309 * that we are disabling:
311 rapl_event_update(event);
312 hwc->state |= PERF_HES_UPTODATE;
315 raw_spin_unlock_irqrestore(&pmu->lock, flags);
318 static int rapl_pmu_event_add(struct perf_event *event, int mode)
320 struct rapl_pmu *pmu = event->pmu_private;
321 struct hw_perf_event *hwc = &event->hw;
324 raw_spin_lock_irqsave(&pmu->lock, flags);
326 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
328 if (mode & PERF_EF_START)
329 __rapl_pmu_event_start(pmu, event);
331 raw_spin_unlock_irqrestore(&pmu->lock, flags);
336 static void rapl_pmu_event_del(struct perf_event *event, int flags)
338 rapl_pmu_event_stop(event, PERF_EF_UPDATE);
341 static int rapl_pmu_event_init(struct perf_event *event)
343 u64 cfg = event->attr.config & RAPL_EVENT_MASK;
345 struct rapl_pmu *pmu;
347 /* only look at RAPL events */
348 if (event->attr.type != rapl_pmus->pmu.type)
351 /* check only supported bits are set */
352 if (event->attr.config & ~RAPL_EVENT_MASK)
358 event->event_caps |= PERF_EV_CAP_READ_ACTIVE_PKG;
360 if (!cfg || cfg >= NR_RAPL_DOMAINS + 1)
363 cfg = array_index_nospec((long)cfg, NR_RAPL_DOMAINS + 1);
366 /* check event supported */
367 if (!(rapl_cntr_mask & (1 << bit)))
370 /* unsupported modes and filters */
371 if (event->attr.sample_period) /* no sampling */
374 /* must be done before validate_group */
375 pmu = cpu_to_rapl_pmu(event->cpu);
378 event->cpu = pmu->cpu;
379 event->pmu_private = pmu;
380 event->hw.event_base = rapl_msrs[bit].msr;
381 event->hw.config = cfg;
387 static void rapl_pmu_event_read(struct perf_event *event)
389 rapl_event_update(event);
392 static ssize_t rapl_get_attr_cpumask(struct device *dev,
393 struct device_attribute *attr, char *buf)
395 return cpumap_print_to_pagebuf(true, buf, &rapl_cpu_mask);
398 static DEVICE_ATTR(cpumask, S_IRUGO, rapl_get_attr_cpumask, NULL);
400 static struct attribute *rapl_pmu_attrs[] = {
401 &dev_attr_cpumask.attr,
405 static struct attribute_group rapl_pmu_attr_group = {
406 .attrs = rapl_pmu_attrs,
409 RAPL_EVENT_ATTR_STR(energy-cores, rapl_cores, "event=0x01");
410 RAPL_EVENT_ATTR_STR(energy-pkg , rapl_pkg, "event=0x02");
411 RAPL_EVENT_ATTR_STR(energy-ram , rapl_ram, "event=0x03");
412 RAPL_EVENT_ATTR_STR(energy-gpu , rapl_gpu, "event=0x04");
413 RAPL_EVENT_ATTR_STR(energy-psys, rapl_psys, "event=0x05");
415 RAPL_EVENT_ATTR_STR(energy-cores.unit, rapl_cores_unit, "Joules");
416 RAPL_EVENT_ATTR_STR(energy-pkg.unit , rapl_pkg_unit, "Joules");
417 RAPL_EVENT_ATTR_STR(energy-ram.unit , rapl_ram_unit, "Joules");
418 RAPL_EVENT_ATTR_STR(energy-gpu.unit , rapl_gpu_unit, "Joules");
419 RAPL_EVENT_ATTR_STR(energy-psys.unit, rapl_psys_unit, "Joules");
422 * we compute in 0.23 nJ increments regardless of MSR
424 RAPL_EVENT_ATTR_STR(energy-cores.scale, rapl_cores_scale, "2.3283064365386962890625e-10");
425 RAPL_EVENT_ATTR_STR(energy-pkg.scale, rapl_pkg_scale, "2.3283064365386962890625e-10");
426 RAPL_EVENT_ATTR_STR(energy-ram.scale, rapl_ram_scale, "2.3283064365386962890625e-10");
427 RAPL_EVENT_ATTR_STR(energy-gpu.scale, rapl_gpu_scale, "2.3283064365386962890625e-10");
428 RAPL_EVENT_ATTR_STR(energy-psys.scale, rapl_psys_scale, "2.3283064365386962890625e-10");
431 * There are no default events, but we need to create
432 * "events" group (with empty attrs) before updating
433 * it with detected events.
435 static struct attribute *attrs_empty[] = {
439 static struct attribute_group rapl_pmu_events_group = {
441 .attrs = attrs_empty,
444 DEFINE_RAPL_FORMAT_ATTR(event, event, "config:0-7");
445 static struct attribute *rapl_formats_attr[] = {
446 &format_attr_event.attr,
450 static struct attribute_group rapl_pmu_format_group = {
452 .attrs = rapl_formats_attr,
455 static const struct attribute_group *rapl_attr_groups[] = {
456 &rapl_pmu_attr_group,
457 &rapl_pmu_format_group,
458 &rapl_pmu_events_group,
462 static struct attribute *rapl_events_cores[] = {
463 EVENT_PTR(rapl_cores),
464 EVENT_PTR(rapl_cores_unit),
465 EVENT_PTR(rapl_cores_scale),
470 rapl_not_visible(struct kobject *kobj, struct attribute *attr, int i)
475 static struct attribute_group rapl_events_cores_group = {
477 .attrs = rapl_events_cores,
478 .is_visible = rapl_not_visible,
481 static struct attribute *rapl_events_pkg[] = {
483 EVENT_PTR(rapl_pkg_unit),
484 EVENT_PTR(rapl_pkg_scale),
488 static struct attribute_group rapl_events_pkg_group = {
490 .attrs = rapl_events_pkg,
491 .is_visible = rapl_not_visible,
494 static struct attribute *rapl_events_ram[] = {
496 EVENT_PTR(rapl_ram_unit),
497 EVENT_PTR(rapl_ram_scale),
501 static struct attribute_group rapl_events_ram_group = {
503 .attrs = rapl_events_ram,
504 .is_visible = rapl_not_visible,
507 static struct attribute *rapl_events_gpu[] = {
509 EVENT_PTR(rapl_gpu_unit),
510 EVENT_PTR(rapl_gpu_scale),
514 static struct attribute_group rapl_events_gpu_group = {
516 .attrs = rapl_events_gpu,
517 .is_visible = rapl_not_visible,
520 static struct attribute *rapl_events_psys[] = {
521 EVENT_PTR(rapl_psys),
522 EVENT_PTR(rapl_psys_unit),
523 EVENT_PTR(rapl_psys_scale),
527 static struct attribute_group rapl_events_psys_group = {
529 .attrs = rapl_events_psys,
530 .is_visible = rapl_not_visible,
533 static bool test_msr(int idx, void *data)
535 return test_bit(idx, (unsigned long *) data);
538 static struct perf_msr intel_rapl_msrs[] = {
539 [PERF_RAPL_PP0] = { MSR_PP0_ENERGY_STATUS, &rapl_events_cores_group, test_msr },
540 [PERF_RAPL_PKG] = { MSR_PKG_ENERGY_STATUS, &rapl_events_pkg_group, test_msr },
541 [PERF_RAPL_RAM] = { MSR_DRAM_ENERGY_STATUS, &rapl_events_ram_group, test_msr },
542 [PERF_RAPL_PP1] = { MSR_PP1_ENERGY_STATUS, &rapl_events_gpu_group, test_msr },
543 [PERF_RAPL_PSYS] = { MSR_PLATFORM_ENERGY_STATUS, &rapl_events_psys_group, test_msr },
547 * Force to PERF_RAPL_MAX size due to:
548 * - perf_msr_probe(PERF_RAPL_MAX)
549 * - want to use same event codes across both architectures
551 static struct perf_msr amd_rapl_msrs[PERF_RAPL_MAX] = {
552 [PERF_RAPL_PKG] = { MSR_AMD_PKG_ENERGY_STATUS, &rapl_events_pkg_group, test_msr },
556 static int rapl_cpu_offline(unsigned int cpu)
558 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
561 /* Check if exiting cpu is used for collecting rapl events */
562 if (!cpumask_test_and_clear_cpu(cpu, &rapl_cpu_mask))
566 /* Find a new cpu to collect rapl events */
567 target = cpumask_any_but(topology_die_cpumask(cpu), cpu);
569 /* Migrate rapl events to the new target */
570 if (target < nr_cpu_ids) {
571 cpumask_set_cpu(target, &rapl_cpu_mask);
573 perf_pmu_migrate_context(pmu->pmu, cpu, target);
578 static int rapl_cpu_online(unsigned int cpu)
580 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
584 pmu = kzalloc_node(sizeof(*pmu), GFP_KERNEL, cpu_to_node(cpu));
588 raw_spin_lock_init(&pmu->lock);
589 INIT_LIST_HEAD(&pmu->active_list);
590 pmu->pmu = &rapl_pmus->pmu;
591 pmu->timer_interval = ms_to_ktime(rapl_timer_ms);
592 rapl_hrtimer_init(pmu);
594 rapl_pmus->pmus[topology_logical_die_id(cpu)] = pmu;
598 * Check if there is an online cpu in the package which collects rapl
601 target = cpumask_any_and(&rapl_cpu_mask, topology_die_cpumask(cpu));
602 if (target < nr_cpu_ids)
605 cpumask_set_cpu(cpu, &rapl_cpu_mask);
610 static int rapl_check_hw_unit(struct rapl_model *rm)
612 u64 msr_rapl_power_unit_bits;
615 /* protect rdmsrl() to handle virtualization */
616 if (rdmsrl_safe(rm->msr_power_unit, &msr_rapl_power_unit_bits))
618 for (i = 0; i < NR_RAPL_DOMAINS; i++)
619 rapl_hw_unit[i] = (msr_rapl_power_unit_bits >> 8) & 0x1FULL;
621 switch (rm->unit_quirk) {
623 * DRAM domain on HSW server and KNL has fixed energy unit which can be
624 * different than the unit from power unit MSR. See
625 * "Intel Xeon Processor E5-1600 and E5-2600 v3 Product Families, V2
626 * of 2. Datasheet, September 2014, Reference Number: 330784-001 "
628 case RAPL_UNIT_QUIRK_INTEL_HSW:
629 rapl_hw_unit[PERF_RAPL_RAM] = 16;
632 * SPR shares the same DRAM domain energy unit as HSW, plus it
633 * also has a fixed energy unit for Psys domain.
635 case RAPL_UNIT_QUIRK_INTEL_SPR:
636 rapl_hw_unit[PERF_RAPL_RAM] = 16;
637 rapl_hw_unit[PERF_RAPL_PSYS] = 0;
645 * Calculate the timer rate:
646 * Use reference of 200W for scaling the timeout to avoid counter
647 * overflows. 200W = 200 Joules/sec
648 * Divide interval by 2 to avoid lockstep (2 * 100)
649 * if hw unit is 32, then we use 2 ms 1/200/2
652 if (rapl_hw_unit[0] < 32) {
653 rapl_timer_ms = (1000 / (2 * 100));
654 rapl_timer_ms *= (1ULL << (32 - rapl_hw_unit[0] - 1));
659 static void __init rapl_advertise(void)
663 pr_info("API unit is 2^-32 Joules, %d fixed counters, %llu ms ovfl timer\n",
664 hweight32(rapl_cntr_mask), rapl_timer_ms);
666 for (i = 0; i < NR_RAPL_DOMAINS; i++) {
667 if (rapl_cntr_mask & (1 << i)) {
668 pr_info("hw unit of domain %s 2^-%d Joules\n",
669 rapl_domain_names[i], rapl_hw_unit[i]);
674 static void cleanup_rapl_pmus(void)
678 for (i = 0; i < rapl_pmus->maxdie; i++)
679 kfree(rapl_pmus->pmus[i]);
683 static const struct attribute_group *rapl_attr_update[] = {
684 &rapl_events_cores_group,
685 &rapl_events_pkg_group,
686 &rapl_events_ram_group,
687 &rapl_events_gpu_group,
688 &rapl_events_psys_group,
692 static int __init init_rapl_pmus(void)
694 int maxdie = topology_max_packages() * topology_max_die_per_package();
697 size = sizeof(*rapl_pmus) + maxdie * sizeof(struct rapl_pmu *);
698 rapl_pmus = kzalloc(size, GFP_KERNEL);
702 rapl_pmus->maxdie = maxdie;
703 rapl_pmus->pmu.attr_groups = rapl_attr_groups;
704 rapl_pmus->pmu.attr_update = rapl_attr_update;
705 rapl_pmus->pmu.task_ctx_nr = perf_invalid_context;
706 rapl_pmus->pmu.event_init = rapl_pmu_event_init;
707 rapl_pmus->pmu.add = rapl_pmu_event_add;
708 rapl_pmus->pmu.del = rapl_pmu_event_del;
709 rapl_pmus->pmu.start = rapl_pmu_event_start;
710 rapl_pmus->pmu.stop = rapl_pmu_event_stop;
711 rapl_pmus->pmu.read = rapl_pmu_event_read;
712 rapl_pmus->pmu.module = THIS_MODULE;
713 rapl_pmus->pmu.capabilities = PERF_PMU_CAP_NO_EXCLUDE;
717 static struct rapl_model model_snb = {
718 .events = BIT(PERF_RAPL_PP0) |
721 .msr_power_unit = MSR_RAPL_POWER_UNIT,
722 .rapl_msrs = intel_rapl_msrs,
725 static struct rapl_model model_snbep = {
726 .events = BIT(PERF_RAPL_PP0) |
729 .msr_power_unit = MSR_RAPL_POWER_UNIT,
730 .rapl_msrs = intel_rapl_msrs,
733 static struct rapl_model model_hsw = {
734 .events = BIT(PERF_RAPL_PP0) |
738 .msr_power_unit = MSR_RAPL_POWER_UNIT,
739 .rapl_msrs = intel_rapl_msrs,
742 static struct rapl_model model_hsx = {
743 .events = BIT(PERF_RAPL_PP0) |
746 .unit_quirk = RAPL_UNIT_QUIRK_INTEL_HSW,
747 .msr_power_unit = MSR_RAPL_POWER_UNIT,
748 .rapl_msrs = intel_rapl_msrs,
751 static struct rapl_model model_knl = {
752 .events = BIT(PERF_RAPL_PKG) |
754 .unit_quirk = RAPL_UNIT_QUIRK_INTEL_HSW,
755 .msr_power_unit = MSR_RAPL_POWER_UNIT,
756 .rapl_msrs = intel_rapl_msrs,
759 static struct rapl_model model_skl = {
760 .events = BIT(PERF_RAPL_PP0) |
765 .msr_power_unit = MSR_RAPL_POWER_UNIT,
766 .rapl_msrs = intel_rapl_msrs,
769 static struct rapl_model model_spr = {
770 .events = BIT(PERF_RAPL_PP0) |
774 .unit_quirk = RAPL_UNIT_QUIRK_INTEL_SPR,
775 .msr_power_unit = MSR_RAPL_POWER_UNIT,
776 .rapl_msrs = intel_rapl_msrs,
779 static struct rapl_model model_amd_fam17h = {
780 .events = BIT(PERF_RAPL_PKG),
781 .msr_power_unit = MSR_AMD_RAPL_POWER_UNIT,
782 .rapl_msrs = amd_rapl_msrs,
785 static const struct x86_cpu_id rapl_model_match[] __initconst = {
786 X86_MATCH_INTEL_FAM6_MODEL(SANDYBRIDGE, &model_snb),
787 X86_MATCH_INTEL_FAM6_MODEL(SANDYBRIDGE_X, &model_snbep),
788 X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE, &model_snb),
789 X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE_X, &model_snbep),
790 X86_MATCH_INTEL_FAM6_MODEL(HASWELL, &model_hsw),
791 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, &model_hsx),
792 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_L, &model_hsw),
793 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_G, &model_hsw),
794 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL, &model_hsw),
795 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_G, &model_hsw),
796 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, &model_hsx),
797 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_D, &model_hsx),
798 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNL, &model_knl),
799 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNM, &model_knl),
800 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_L, &model_skl),
801 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE, &model_skl),
802 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_X, &model_hsx),
803 X86_MATCH_INTEL_FAM6_MODEL(KABYLAKE_L, &model_skl),
804 X86_MATCH_INTEL_FAM6_MODEL(KABYLAKE, &model_skl),
805 X86_MATCH_INTEL_FAM6_MODEL(CANNONLAKE_L, &model_skl),
806 X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT, &model_hsw),
807 X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT_D, &model_hsw),
808 X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT_PLUS, &model_hsw),
809 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_L, &model_skl),
810 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE, &model_skl),
811 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, &model_hsx),
812 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, &model_hsx),
813 X86_MATCH_INTEL_FAM6_MODEL(COMETLAKE_L, &model_skl),
814 X86_MATCH_INTEL_FAM6_MODEL(COMETLAKE, &model_skl),
815 X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, &model_spr),
816 X86_MATCH_VENDOR_FAM(AMD, 0x17, &model_amd_fam17h),
817 X86_MATCH_VENDOR_FAM(HYGON, 0x18, &model_amd_fam17h),
820 MODULE_DEVICE_TABLE(x86cpu, rapl_model_match);
822 static int __init rapl_pmu_init(void)
824 const struct x86_cpu_id *id;
825 struct rapl_model *rm;
828 id = x86_match_cpu(rapl_model_match);
832 rm = (struct rapl_model *) id->driver_data;
834 rapl_msrs = rm->rapl_msrs;
836 rapl_cntr_mask = perf_msr_probe(rapl_msrs, PERF_RAPL_MAX,
837 false, (void *) &rm->events);
839 ret = rapl_check_hw_unit(rm);
843 ret = init_rapl_pmus();
848 * Install callbacks. Core will call them for each online cpu.
850 ret = cpuhp_setup_state(CPUHP_AP_PERF_X86_RAPL_ONLINE,
851 "perf/x86/rapl:online",
852 rapl_cpu_online, rapl_cpu_offline);
856 ret = perf_pmu_register(&rapl_pmus->pmu, "power", -1);
864 cpuhp_remove_state(CPUHP_AP_PERF_X86_RAPL_ONLINE);
866 pr_warn("Initialization failed (%d), disabled\n", ret);
870 module_init(rapl_pmu_init);
872 static void __exit intel_rapl_exit(void)
874 cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_RAPL_ONLINE);
875 perf_pmu_unregister(&rapl_pmus->pmu);
878 module_exit(intel_rapl_exit);