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
96 #define RAPL_CNTR_WIDTH 32
98 #define RAPL_EVENT_ATTR_STR(_name, v, str) \
99 static struct perf_pmu_events_attr event_attr_##v = { \
100 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
109 struct list_head active_list;
111 ktime_t timer_interval;
112 struct hrtimer hrtimer;
118 struct rapl_pmu *pmus[];
121 enum rapl_unit_quirk {
122 RAPL_UNIT_QUIRK_NONE,
123 RAPL_UNIT_QUIRK_INTEL_HSW,
124 RAPL_UNIT_QUIRK_INTEL_SPR,
128 struct perf_msr *rapl_msrs;
129 unsigned long events;
130 unsigned int msr_power_unit;
131 enum rapl_unit_quirk unit_quirk;
134 /* 1/2^hw_unit Joule */
135 static int rapl_hw_unit[NR_RAPL_DOMAINS] __read_mostly;
136 static struct rapl_pmus *rapl_pmus;
137 static cpumask_t rapl_cpu_mask;
138 static unsigned int rapl_cntr_mask;
139 static u64 rapl_timer_ms;
140 static struct perf_msr *rapl_msrs;
142 static inline struct rapl_pmu *cpu_to_rapl_pmu(unsigned int cpu)
144 unsigned int dieid = topology_logical_die_id(cpu);
147 * The unsigned check also catches the '-1' return value for non
148 * existent mappings in the topology map.
150 return dieid < rapl_pmus->maxdie ? rapl_pmus->pmus[dieid] : NULL;
153 static inline u64 rapl_read_counter(struct perf_event *event)
156 rdmsrl(event->hw.event_base, raw);
160 static inline u64 rapl_scale(u64 v, int cfg)
162 if (cfg > NR_RAPL_DOMAINS) {
163 pr_warn("Invalid domain %d, failed to scale data\n", cfg);
167 * scale delta to smallest unit (1/2^32)
168 * users must then scale back: count * 1/(1e9*2^32) to get Joules
169 * or use ldexp(count, -32).
170 * Watts = Joules/Time delta
172 return v << (32 - rapl_hw_unit[cfg - 1]);
175 static u64 rapl_event_update(struct perf_event *event)
177 struct hw_perf_event *hwc = &event->hw;
178 u64 prev_raw_count, new_raw_count;
180 int shift = RAPL_CNTR_WIDTH;
183 prev_raw_count = local64_read(&hwc->prev_count);
184 rdmsrl(event->hw.event_base, new_raw_count);
186 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
187 new_raw_count) != prev_raw_count) {
193 * Now we have the new raw value and have updated the prev
194 * timestamp already. We can now calculate the elapsed delta
195 * (event-)time and add that to the generic event.
197 * Careful, not all hw sign-extends above the physical width
200 delta = (new_raw_count << shift) - (prev_raw_count << shift);
203 sdelta = rapl_scale(delta, event->hw.config);
205 local64_add(sdelta, &event->count);
207 return new_raw_count;
210 static void rapl_start_hrtimer(struct rapl_pmu *pmu)
212 hrtimer_start(&pmu->hrtimer, pmu->timer_interval,
213 HRTIMER_MODE_REL_PINNED);
216 static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
218 struct rapl_pmu *pmu = container_of(hrtimer, struct rapl_pmu, hrtimer);
219 struct perf_event *event;
223 return HRTIMER_NORESTART;
225 raw_spin_lock_irqsave(&pmu->lock, flags);
227 list_for_each_entry(event, &pmu->active_list, active_entry)
228 rapl_event_update(event);
230 raw_spin_unlock_irqrestore(&pmu->lock, flags);
232 hrtimer_forward_now(hrtimer, pmu->timer_interval);
234 return HRTIMER_RESTART;
237 static void rapl_hrtimer_init(struct rapl_pmu *pmu)
239 struct hrtimer *hr = &pmu->hrtimer;
241 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
242 hr->function = rapl_hrtimer_handle;
245 static void __rapl_pmu_event_start(struct rapl_pmu *pmu,
246 struct perf_event *event)
248 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
253 list_add_tail(&event->active_entry, &pmu->active_list);
255 local64_set(&event->hw.prev_count, rapl_read_counter(event));
258 if (pmu->n_active == 1)
259 rapl_start_hrtimer(pmu);
262 static void rapl_pmu_event_start(struct perf_event *event, int mode)
264 struct rapl_pmu *pmu = event->pmu_private;
267 raw_spin_lock_irqsave(&pmu->lock, flags);
268 __rapl_pmu_event_start(pmu, event);
269 raw_spin_unlock_irqrestore(&pmu->lock, flags);
272 static void rapl_pmu_event_stop(struct perf_event *event, int mode)
274 struct rapl_pmu *pmu = event->pmu_private;
275 struct hw_perf_event *hwc = &event->hw;
278 raw_spin_lock_irqsave(&pmu->lock, flags);
280 /* mark event as deactivated and stopped */
281 if (!(hwc->state & PERF_HES_STOPPED)) {
282 WARN_ON_ONCE(pmu->n_active <= 0);
284 if (pmu->n_active == 0)
285 hrtimer_cancel(&pmu->hrtimer);
287 list_del(&event->active_entry);
289 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
290 hwc->state |= PERF_HES_STOPPED;
293 /* check if update of sw counter is necessary */
294 if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
296 * Drain the remaining delta count out of a event
297 * that we are disabling:
299 rapl_event_update(event);
300 hwc->state |= PERF_HES_UPTODATE;
303 raw_spin_unlock_irqrestore(&pmu->lock, flags);
306 static int rapl_pmu_event_add(struct perf_event *event, int mode)
308 struct rapl_pmu *pmu = event->pmu_private;
309 struct hw_perf_event *hwc = &event->hw;
312 raw_spin_lock_irqsave(&pmu->lock, flags);
314 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
316 if (mode & PERF_EF_START)
317 __rapl_pmu_event_start(pmu, event);
319 raw_spin_unlock_irqrestore(&pmu->lock, flags);
324 static void rapl_pmu_event_del(struct perf_event *event, int flags)
326 rapl_pmu_event_stop(event, PERF_EF_UPDATE);
329 static int rapl_pmu_event_init(struct perf_event *event)
331 u64 cfg = event->attr.config & RAPL_EVENT_MASK;
333 struct rapl_pmu *pmu;
335 /* only look at RAPL events */
336 if (event->attr.type != rapl_pmus->pmu.type)
339 /* check only supported bits are set */
340 if (event->attr.config & ~RAPL_EVENT_MASK)
346 event->event_caps |= PERF_EV_CAP_READ_ACTIVE_PKG;
348 if (!cfg || cfg >= NR_RAPL_DOMAINS + 1)
351 cfg = array_index_nospec((long)cfg, NR_RAPL_DOMAINS + 1);
354 /* check event supported */
355 if (!(rapl_cntr_mask & (1 << bit)))
358 /* unsupported modes and filters */
359 if (event->attr.sample_period) /* no sampling */
362 /* must be done before validate_group */
363 pmu = cpu_to_rapl_pmu(event->cpu);
366 event->cpu = pmu->cpu;
367 event->pmu_private = pmu;
368 event->hw.event_base = rapl_msrs[bit].msr;
369 event->hw.config = cfg;
375 static void rapl_pmu_event_read(struct perf_event *event)
377 rapl_event_update(event);
380 static ssize_t rapl_get_attr_cpumask(struct device *dev,
381 struct device_attribute *attr, char *buf)
383 return cpumap_print_to_pagebuf(true, buf, &rapl_cpu_mask);
386 static DEVICE_ATTR(cpumask, S_IRUGO, rapl_get_attr_cpumask, NULL);
388 static struct attribute *rapl_pmu_attrs[] = {
389 &dev_attr_cpumask.attr,
393 static struct attribute_group rapl_pmu_attr_group = {
394 .attrs = rapl_pmu_attrs,
397 RAPL_EVENT_ATTR_STR(energy-cores, rapl_cores, "event=0x01");
398 RAPL_EVENT_ATTR_STR(energy-pkg , rapl_pkg, "event=0x02");
399 RAPL_EVENT_ATTR_STR(energy-ram , rapl_ram, "event=0x03");
400 RAPL_EVENT_ATTR_STR(energy-gpu , rapl_gpu, "event=0x04");
401 RAPL_EVENT_ATTR_STR(energy-psys, rapl_psys, "event=0x05");
403 RAPL_EVENT_ATTR_STR(energy-cores.unit, rapl_cores_unit, "Joules");
404 RAPL_EVENT_ATTR_STR(energy-pkg.unit , rapl_pkg_unit, "Joules");
405 RAPL_EVENT_ATTR_STR(energy-ram.unit , rapl_ram_unit, "Joules");
406 RAPL_EVENT_ATTR_STR(energy-gpu.unit , rapl_gpu_unit, "Joules");
407 RAPL_EVENT_ATTR_STR(energy-psys.unit, rapl_psys_unit, "Joules");
410 * we compute in 0.23 nJ increments regardless of MSR
412 RAPL_EVENT_ATTR_STR(energy-cores.scale, rapl_cores_scale, "2.3283064365386962890625e-10");
413 RAPL_EVENT_ATTR_STR(energy-pkg.scale, rapl_pkg_scale, "2.3283064365386962890625e-10");
414 RAPL_EVENT_ATTR_STR(energy-ram.scale, rapl_ram_scale, "2.3283064365386962890625e-10");
415 RAPL_EVENT_ATTR_STR(energy-gpu.scale, rapl_gpu_scale, "2.3283064365386962890625e-10");
416 RAPL_EVENT_ATTR_STR(energy-psys.scale, rapl_psys_scale, "2.3283064365386962890625e-10");
419 * There are no default events, but we need to create
420 * "events" group (with empty attrs) before updating
421 * it with detected events.
423 static struct attribute *attrs_empty[] = {
427 static struct attribute_group rapl_pmu_events_group = {
429 .attrs = attrs_empty,
432 PMU_FORMAT_ATTR(event, "config:0-7");
433 static struct attribute *rapl_formats_attr[] = {
434 &format_attr_event.attr,
438 static struct attribute_group rapl_pmu_format_group = {
440 .attrs = rapl_formats_attr,
443 static const struct attribute_group *rapl_attr_groups[] = {
444 &rapl_pmu_attr_group,
445 &rapl_pmu_format_group,
446 &rapl_pmu_events_group,
450 static struct attribute *rapl_events_cores[] = {
451 EVENT_PTR(rapl_cores),
452 EVENT_PTR(rapl_cores_unit),
453 EVENT_PTR(rapl_cores_scale),
457 static struct attribute_group rapl_events_cores_group = {
459 .attrs = rapl_events_cores,
462 static struct attribute *rapl_events_pkg[] = {
464 EVENT_PTR(rapl_pkg_unit),
465 EVENT_PTR(rapl_pkg_scale),
469 static struct attribute_group rapl_events_pkg_group = {
471 .attrs = rapl_events_pkg,
474 static struct attribute *rapl_events_ram[] = {
476 EVENT_PTR(rapl_ram_unit),
477 EVENT_PTR(rapl_ram_scale),
481 static struct attribute_group rapl_events_ram_group = {
483 .attrs = rapl_events_ram,
486 static struct attribute *rapl_events_gpu[] = {
488 EVENT_PTR(rapl_gpu_unit),
489 EVENT_PTR(rapl_gpu_scale),
493 static struct attribute_group rapl_events_gpu_group = {
495 .attrs = rapl_events_gpu,
498 static struct attribute *rapl_events_psys[] = {
499 EVENT_PTR(rapl_psys),
500 EVENT_PTR(rapl_psys_unit),
501 EVENT_PTR(rapl_psys_scale),
505 static struct attribute_group rapl_events_psys_group = {
507 .attrs = rapl_events_psys,
510 static bool test_msr(int idx, void *data)
512 return test_bit(idx, (unsigned long *) data);
515 /* Only lower 32bits of the MSR represents the energy counter */
516 #define RAPL_MSR_MASK 0xFFFFFFFF
518 static struct perf_msr intel_rapl_msrs[] = {
519 [PERF_RAPL_PP0] = { MSR_PP0_ENERGY_STATUS, &rapl_events_cores_group, test_msr, false, RAPL_MSR_MASK },
520 [PERF_RAPL_PKG] = { MSR_PKG_ENERGY_STATUS, &rapl_events_pkg_group, test_msr, false, RAPL_MSR_MASK },
521 [PERF_RAPL_RAM] = { MSR_DRAM_ENERGY_STATUS, &rapl_events_ram_group, test_msr, false, RAPL_MSR_MASK },
522 [PERF_RAPL_PP1] = { MSR_PP1_ENERGY_STATUS, &rapl_events_gpu_group, test_msr, false, RAPL_MSR_MASK },
523 [PERF_RAPL_PSYS] = { MSR_PLATFORM_ENERGY_STATUS, &rapl_events_psys_group, test_msr, false, RAPL_MSR_MASK },
526 static struct perf_msr intel_rapl_spr_msrs[] = {
527 [PERF_RAPL_PP0] = { MSR_PP0_ENERGY_STATUS, &rapl_events_cores_group, test_msr, false, RAPL_MSR_MASK },
528 [PERF_RAPL_PKG] = { MSR_PKG_ENERGY_STATUS, &rapl_events_pkg_group, test_msr, false, RAPL_MSR_MASK },
529 [PERF_RAPL_RAM] = { MSR_DRAM_ENERGY_STATUS, &rapl_events_ram_group, test_msr, false, RAPL_MSR_MASK },
530 [PERF_RAPL_PP1] = { MSR_PP1_ENERGY_STATUS, &rapl_events_gpu_group, test_msr, false, RAPL_MSR_MASK },
531 [PERF_RAPL_PSYS] = { MSR_PLATFORM_ENERGY_STATUS, &rapl_events_psys_group, test_msr, true, RAPL_MSR_MASK },
535 * Force to PERF_RAPL_MAX size due to:
536 * - perf_msr_probe(PERF_RAPL_MAX)
537 * - want to use same event codes across both architectures
539 static struct perf_msr amd_rapl_msrs[PERF_RAPL_MAX] = {
540 [PERF_RAPL_PKG] = { MSR_AMD_PKG_ENERGY_STATUS, &rapl_events_pkg_group, test_msr },
544 static int rapl_cpu_offline(unsigned int cpu)
546 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
549 /* Check if exiting cpu is used for collecting rapl events */
550 if (!cpumask_test_and_clear_cpu(cpu, &rapl_cpu_mask))
554 /* Find a new cpu to collect rapl events */
555 target = cpumask_any_but(topology_die_cpumask(cpu), cpu);
557 /* Migrate rapl events to the new target */
558 if (target < nr_cpu_ids) {
559 cpumask_set_cpu(target, &rapl_cpu_mask);
561 perf_pmu_migrate_context(pmu->pmu, cpu, target);
566 static int rapl_cpu_online(unsigned int cpu)
568 struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
572 pmu = kzalloc_node(sizeof(*pmu), GFP_KERNEL, cpu_to_node(cpu));
576 raw_spin_lock_init(&pmu->lock);
577 INIT_LIST_HEAD(&pmu->active_list);
578 pmu->pmu = &rapl_pmus->pmu;
579 pmu->timer_interval = ms_to_ktime(rapl_timer_ms);
580 rapl_hrtimer_init(pmu);
582 rapl_pmus->pmus[topology_logical_die_id(cpu)] = pmu;
586 * Check if there is an online cpu in the package which collects rapl
589 target = cpumask_any_and(&rapl_cpu_mask, topology_die_cpumask(cpu));
590 if (target < nr_cpu_ids)
593 cpumask_set_cpu(cpu, &rapl_cpu_mask);
598 static int rapl_check_hw_unit(struct rapl_model *rm)
600 u64 msr_rapl_power_unit_bits;
603 /* protect rdmsrl() to handle virtualization */
604 if (rdmsrl_safe(rm->msr_power_unit, &msr_rapl_power_unit_bits))
606 for (i = 0; i < NR_RAPL_DOMAINS; i++)
607 rapl_hw_unit[i] = (msr_rapl_power_unit_bits >> 8) & 0x1FULL;
609 switch (rm->unit_quirk) {
611 * DRAM domain on HSW server and KNL has fixed energy unit which can be
612 * different than the unit from power unit MSR. See
613 * "Intel Xeon Processor E5-1600 and E5-2600 v3 Product Families, V2
614 * of 2. Datasheet, September 2014, Reference Number: 330784-001 "
616 case RAPL_UNIT_QUIRK_INTEL_HSW:
617 rapl_hw_unit[PERF_RAPL_RAM] = 16;
620 * SPR shares the same DRAM domain energy unit as HSW, plus it
621 * also has a fixed energy unit for Psys domain.
623 case RAPL_UNIT_QUIRK_INTEL_SPR:
624 rapl_hw_unit[PERF_RAPL_RAM] = 16;
625 rapl_hw_unit[PERF_RAPL_PSYS] = 0;
633 * Calculate the timer rate:
634 * Use reference of 200W for scaling the timeout to avoid counter
635 * overflows. 200W = 200 Joules/sec
636 * Divide interval by 2 to avoid lockstep (2 * 100)
637 * if hw unit is 32, then we use 2 ms 1/200/2
640 if (rapl_hw_unit[0] < 32) {
641 rapl_timer_ms = (1000 / (2 * 100));
642 rapl_timer_ms *= (1ULL << (32 - rapl_hw_unit[0] - 1));
647 static void __init rapl_advertise(void)
651 pr_info("API unit is 2^-32 Joules, %d fixed counters, %llu ms ovfl timer\n",
652 hweight32(rapl_cntr_mask), rapl_timer_ms);
654 for (i = 0; i < NR_RAPL_DOMAINS; i++) {
655 if (rapl_cntr_mask & (1 << i)) {
656 pr_info("hw unit of domain %s 2^-%d Joules\n",
657 rapl_domain_names[i], rapl_hw_unit[i]);
662 static void cleanup_rapl_pmus(void)
666 for (i = 0; i < rapl_pmus->maxdie; i++)
667 kfree(rapl_pmus->pmus[i]);
671 static const struct attribute_group *rapl_attr_update[] = {
672 &rapl_events_cores_group,
673 &rapl_events_pkg_group,
674 &rapl_events_ram_group,
675 &rapl_events_gpu_group,
676 &rapl_events_psys_group,
680 static int __init init_rapl_pmus(void)
682 int maxdie = topology_max_packages() * topology_max_die_per_package();
685 size = sizeof(*rapl_pmus) + maxdie * sizeof(struct rapl_pmu *);
686 rapl_pmus = kzalloc(size, GFP_KERNEL);
690 rapl_pmus->maxdie = maxdie;
691 rapl_pmus->pmu.attr_groups = rapl_attr_groups;
692 rapl_pmus->pmu.attr_update = rapl_attr_update;
693 rapl_pmus->pmu.task_ctx_nr = perf_invalid_context;
694 rapl_pmus->pmu.event_init = rapl_pmu_event_init;
695 rapl_pmus->pmu.add = rapl_pmu_event_add;
696 rapl_pmus->pmu.del = rapl_pmu_event_del;
697 rapl_pmus->pmu.start = rapl_pmu_event_start;
698 rapl_pmus->pmu.stop = rapl_pmu_event_stop;
699 rapl_pmus->pmu.read = rapl_pmu_event_read;
700 rapl_pmus->pmu.module = THIS_MODULE;
701 rapl_pmus->pmu.capabilities = PERF_PMU_CAP_NO_EXCLUDE;
705 static struct rapl_model model_snb = {
706 .events = BIT(PERF_RAPL_PP0) |
709 .msr_power_unit = MSR_RAPL_POWER_UNIT,
710 .rapl_msrs = intel_rapl_msrs,
713 static struct rapl_model model_snbep = {
714 .events = BIT(PERF_RAPL_PP0) |
717 .msr_power_unit = MSR_RAPL_POWER_UNIT,
718 .rapl_msrs = intel_rapl_msrs,
721 static struct rapl_model model_hsw = {
722 .events = BIT(PERF_RAPL_PP0) |
726 .msr_power_unit = MSR_RAPL_POWER_UNIT,
727 .rapl_msrs = intel_rapl_msrs,
730 static struct rapl_model model_hsx = {
731 .events = BIT(PERF_RAPL_PP0) |
734 .unit_quirk = RAPL_UNIT_QUIRK_INTEL_HSW,
735 .msr_power_unit = MSR_RAPL_POWER_UNIT,
736 .rapl_msrs = intel_rapl_msrs,
739 static struct rapl_model model_knl = {
740 .events = BIT(PERF_RAPL_PKG) |
742 .unit_quirk = RAPL_UNIT_QUIRK_INTEL_HSW,
743 .msr_power_unit = MSR_RAPL_POWER_UNIT,
744 .rapl_msrs = intel_rapl_msrs,
747 static struct rapl_model model_skl = {
748 .events = BIT(PERF_RAPL_PP0) |
753 .msr_power_unit = MSR_RAPL_POWER_UNIT,
754 .rapl_msrs = intel_rapl_msrs,
757 static struct rapl_model model_spr = {
758 .events = BIT(PERF_RAPL_PP0) |
762 .unit_quirk = RAPL_UNIT_QUIRK_INTEL_SPR,
763 .msr_power_unit = MSR_RAPL_POWER_UNIT,
764 .rapl_msrs = intel_rapl_spr_msrs,
767 static struct rapl_model model_amd_fam17h = {
768 .events = BIT(PERF_RAPL_PKG),
769 .msr_power_unit = MSR_AMD_RAPL_POWER_UNIT,
770 .rapl_msrs = amd_rapl_msrs,
773 static const struct x86_cpu_id rapl_model_match[] __initconst = {
774 X86_MATCH_INTEL_FAM6_MODEL(SANDYBRIDGE, &model_snb),
775 X86_MATCH_INTEL_FAM6_MODEL(SANDYBRIDGE_X, &model_snbep),
776 X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE, &model_snb),
777 X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE_X, &model_snbep),
778 X86_MATCH_INTEL_FAM6_MODEL(HASWELL, &model_hsw),
779 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, &model_hsx),
780 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_L, &model_hsw),
781 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_G, &model_hsw),
782 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL, &model_hsw),
783 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_G, &model_hsw),
784 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, &model_hsx),
785 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_D, &model_hsx),
786 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNL, &model_knl),
787 X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNM, &model_knl),
788 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_L, &model_skl),
789 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE, &model_skl),
790 X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_X, &model_hsx),
791 X86_MATCH_INTEL_FAM6_MODEL(KABYLAKE_L, &model_skl),
792 X86_MATCH_INTEL_FAM6_MODEL(KABYLAKE, &model_skl),
793 X86_MATCH_INTEL_FAM6_MODEL(CANNONLAKE_L, &model_skl),
794 X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT, &model_hsw),
795 X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT_D, &model_hsw),
796 X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT_PLUS, &model_hsw),
797 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_L, &model_skl),
798 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE, &model_skl),
799 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, &model_hsx),
800 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, &model_hsx),
801 X86_MATCH_INTEL_FAM6_MODEL(COMETLAKE_L, &model_skl),
802 X86_MATCH_INTEL_FAM6_MODEL(COMETLAKE, &model_skl),
803 X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, &model_spr),
804 X86_MATCH_VENDOR_FAM(AMD, 0x17, &model_amd_fam17h),
805 X86_MATCH_VENDOR_FAM(HYGON, 0x18, &model_amd_fam17h),
806 X86_MATCH_VENDOR_FAM(AMD, 0x19, &model_amd_fam17h),
809 MODULE_DEVICE_TABLE(x86cpu, rapl_model_match);
811 static int __init rapl_pmu_init(void)
813 const struct x86_cpu_id *id;
814 struct rapl_model *rm;
817 id = x86_match_cpu(rapl_model_match);
821 rm = (struct rapl_model *) id->driver_data;
823 rapl_msrs = rm->rapl_msrs;
825 rapl_cntr_mask = perf_msr_probe(rapl_msrs, PERF_RAPL_MAX,
826 false, (void *) &rm->events);
828 ret = rapl_check_hw_unit(rm);
832 ret = init_rapl_pmus();
837 * Install callbacks. Core will call them for each online cpu.
839 ret = cpuhp_setup_state(CPUHP_AP_PERF_X86_RAPL_ONLINE,
840 "perf/x86/rapl:online",
841 rapl_cpu_online, rapl_cpu_offline);
845 ret = perf_pmu_register(&rapl_pmus->pmu, "power", -1);
853 cpuhp_remove_state(CPUHP_AP_PERF_X86_RAPL_ONLINE);
855 pr_warn("Initialization failed (%d), disabled\n", ret);
859 module_init(rapl_pmu_init);
861 static void __exit intel_rapl_exit(void)
863 cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_RAPL_ONLINE);
864 perf_pmu_unregister(&rapl_pmus->pmu);
867 module_exit(intel_rapl_exit);