2 * perf_event_intel_rapl.c: support Intel RAPL energy consumption counters
3 * Copyright (C) 2013 Google, Inc., Stephane Eranian
5 * Intel RAPL interface is specified in the IA-32 Manual Vol3b
6 * section 14.7.1 (September 2013)
8 * RAPL provides more controls than just reporting energy consumption
9 * however here we only expose the 3 energy consumption free running
10 * counters (pp0, pkg, dram).
12 * Each of those counters increments in a power unit defined by the
13 * RAPL_POWER_UNIT MSR. On SandyBridge, this unit is 1/(2^16) Joules
16 * Counter to rapl events mappings:
18 * pp0 counter: consumption of all physical cores (power plane 0)
19 * event: rapl_energy_cores
22 * pkg counter: consumption of the whole processor package
23 * event: rapl_energy_pkg
26 * dram counter: consumption of the dram domain (servers only)
27 * event: rapl_energy_dram
30 * dram counter: consumption of the builtin-gpu domain (client only)
31 * event: rapl_energy_gpu
34 * We manage those counters as free running (read-only). They may be
35 * use simultaneously by other tools, such as turbostat.
37 * The events only support system-wide mode counting. There is no
38 * sampling support because it does not make sense and is not
39 * supported by the RAPL hardware.
41 * Because we want to avoid floating-point operations in the kernel,
42 * the events are all reported in fixed point arithmetic (32.32).
43 * Tools must adjust the counts to convert them to Watts using
44 * the duration of the measurement. Tools may use a function such as
45 * ldexp(raw_count, -32);
47 #include <linux/module.h>
48 #include <linux/slab.h>
49 #include <linux/perf_event.h>
50 #include <asm/cpu_device_id.h>
51 #include "perf_event.h"
54 * RAPL energy status counters
56 #define RAPL_IDX_PP0_NRG_STAT 0 /* all cores */
57 #define INTEL_RAPL_PP0 0x1 /* pseudo-encoding */
58 #define RAPL_IDX_PKG_NRG_STAT 1 /* entire package */
59 #define INTEL_RAPL_PKG 0x2 /* pseudo-encoding */
60 #define RAPL_IDX_RAM_NRG_STAT 2 /* DRAM */
61 #define INTEL_RAPL_RAM 0x3 /* pseudo-encoding */
62 #define RAPL_IDX_PP1_NRG_STAT 3 /* gpu */
63 #define INTEL_RAPL_PP1 0x4 /* pseudo-encoding */
65 #define NR_RAPL_DOMAINS 0x4
66 static const char *rapl_domain_names[NR_RAPL_DOMAINS] __initconst = {
73 /* Clients have PP0, PKG */
74 #define RAPL_IDX_CLN (1<<RAPL_IDX_PP0_NRG_STAT|\
75 1<<RAPL_IDX_PKG_NRG_STAT|\
76 1<<RAPL_IDX_PP1_NRG_STAT)
78 /* Servers have PP0, PKG, RAM */
79 #define RAPL_IDX_SRV (1<<RAPL_IDX_PP0_NRG_STAT|\
80 1<<RAPL_IDX_PKG_NRG_STAT|\
81 1<<RAPL_IDX_RAM_NRG_STAT)
83 /* Servers have PP0, PKG, RAM, PP1 */
84 #define RAPL_IDX_HSW (1<<RAPL_IDX_PP0_NRG_STAT|\
85 1<<RAPL_IDX_PKG_NRG_STAT|\
86 1<<RAPL_IDX_RAM_NRG_STAT|\
87 1<<RAPL_IDX_PP1_NRG_STAT)
89 /* Knights Landing has PKG, RAM */
90 #define RAPL_IDX_KNL (1<<RAPL_IDX_PKG_NRG_STAT|\
91 1<<RAPL_IDX_RAM_NRG_STAT)
94 * event code: LSB 8 bits, passed in attr->config
95 * any other bit is reserved
97 #define RAPL_EVENT_MASK 0xFFULL
99 #define DEFINE_RAPL_FORMAT_ATTR(_var, _name, _format) \
100 static ssize_t __rapl_##_var##_show(struct kobject *kobj, \
101 struct kobj_attribute *attr, \
104 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
105 return sprintf(page, _format "\n"); \
107 static struct kobj_attribute format_attr_##_var = \
108 __ATTR(_name, 0444, __rapl_##_var##_show, NULL)
110 #define RAPL_CNTR_WIDTH 32 /* 32-bit rapl counters */
112 #define RAPL_EVENT_ATTR_STR(_name, v, str) \
113 static struct perf_pmu_events_attr event_attr_##v = { \
114 .attr = __ATTR(_name, 0444, rapl_sysfs_show, NULL), \
121 int n_active; /* number of active events */
122 struct list_head active_list;
123 struct pmu *pmu; /* pointer to rapl_pmu_class */
124 ktime_t timer_interval; /* in ktime_t unit */
125 struct hrtimer hrtimer;
128 static int rapl_hw_unit[NR_RAPL_DOMAINS] __read_mostly; /* 1/2^hw_unit Joule */
129 static struct pmu rapl_pmu_class;
130 static cpumask_t rapl_cpu_mask;
131 static int rapl_cntr_mask;
133 static DEFINE_PER_CPU(struct rapl_pmu *, rapl_pmu);
134 static DEFINE_PER_CPU(struct rapl_pmu *, rapl_pmu_to_free);
136 static struct x86_pmu_quirk *rapl_quirks;
137 static inline u64 rapl_read_counter(struct perf_event *event)
140 rdmsrl(event->hw.event_base, raw);
144 #define rapl_add_quirk(func_) \
146 static struct x86_pmu_quirk __quirk __initdata = { \
149 __quirk.next = rapl_quirks; \
150 rapl_quirks = &__quirk; \
153 static inline u64 rapl_scale(u64 v, int cfg)
155 if (cfg > NR_RAPL_DOMAINS) {
156 pr_warn("invalid domain %d, failed to scale data\n", cfg);
160 * scale delta to smallest unit (1/2^32)
161 * users must then scale back: count * 1/(1e9*2^32) to get Joules
162 * or use ldexp(count, -32).
163 * Watts = Joules/Time delta
165 return v << (32 - rapl_hw_unit[cfg - 1]);
168 static u64 rapl_event_update(struct perf_event *event)
170 struct hw_perf_event *hwc = &event->hw;
171 u64 prev_raw_count, new_raw_count;
173 int shift = RAPL_CNTR_WIDTH;
176 prev_raw_count = local64_read(&hwc->prev_count);
177 rdmsrl(event->hw.event_base, new_raw_count);
179 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
180 new_raw_count) != prev_raw_count) {
186 * Now we have the new raw value and have updated the prev
187 * timestamp already. We can now calculate the elapsed delta
188 * (event-)time and add that to the generic event.
190 * Careful, not all hw sign-extends above the physical width
193 delta = (new_raw_count << shift) - (prev_raw_count << shift);
196 sdelta = rapl_scale(delta, event->hw.config);
198 local64_add(sdelta, &event->count);
200 return new_raw_count;
203 static void rapl_start_hrtimer(struct rapl_pmu *pmu)
205 hrtimer_start(&pmu->hrtimer, pmu->timer_interval,
206 HRTIMER_MODE_REL_PINNED);
209 static void rapl_stop_hrtimer(struct rapl_pmu *pmu)
211 hrtimer_cancel(&pmu->hrtimer);
214 static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
216 struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
217 struct perf_event *event;
221 return HRTIMER_NORESTART;
223 raw_spin_lock_irqsave(&pmu->lock, flags);
225 list_for_each_entry(event, &pmu->active_list, active_entry) {
226 rapl_event_update(event);
229 raw_spin_unlock_irqrestore(&pmu->lock, flags);
231 hrtimer_forward_now(hrtimer, pmu->timer_interval);
233 return HRTIMER_RESTART;
236 static void rapl_hrtimer_init(struct rapl_pmu *pmu)
238 struct hrtimer *hr = &pmu->hrtimer;
240 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
241 hr->function = rapl_hrtimer_handle;
244 static void __rapl_pmu_event_start(struct rapl_pmu *pmu,
245 struct perf_event *event)
247 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
252 list_add_tail(&event->active_entry, &pmu->active_list);
254 local64_set(&event->hw.prev_count, rapl_read_counter(event));
257 if (pmu->n_active == 1)
258 rapl_start_hrtimer(pmu);
261 static void rapl_pmu_event_start(struct perf_event *event, int mode)
263 struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
266 raw_spin_lock_irqsave(&pmu->lock, flags);
267 __rapl_pmu_event_start(pmu, event);
268 raw_spin_unlock_irqrestore(&pmu->lock, flags);
271 static void rapl_pmu_event_stop(struct perf_event *event, int mode)
273 struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
274 struct hw_perf_event *hwc = &event->hw;
277 raw_spin_lock_irqsave(&pmu->lock, flags);
279 /* mark event as deactivated and stopped */
280 if (!(hwc->state & PERF_HES_STOPPED)) {
281 WARN_ON_ONCE(pmu->n_active <= 0);
283 if (pmu->n_active == 0)
284 rapl_stop_hrtimer(pmu);
286 list_del(&event->active_entry);
288 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
289 hwc->state |= PERF_HES_STOPPED;
292 /* check if update of sw counter is necessary */
293 if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
295 * Drain the remaining delta count out of a event
296 * that we are disabling:
298 rapl_event_update(event);
299 hwc->state |= PERF_HES_UPTODATE;
302 raw_spin_unlock_irqrestore(&pmu->lock, flags);
305 static int rapl_pmu_event_add(struct perf_event *event, int mode)
307 struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
308 struct hw_perf_event *hwc = &event->hw;
311 raw_spin_lock_irqsave(&pmu->lock, flags);
313 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
315 if (mode & PERF_EF_START)
316 __rapl_pmu_event_start(pmu, event);
318 raw_spin_unlock_irqrestore(&pmu->lock, flags);
323 static void rapl_pmu_event_del(struct perf_event *event, int flags)
325 rapl_pmu_event_stop(event, PERF_EF_UPDATE);
328 static int rapl_pmu_event_init(struct perf_event *event)
330 u64 cfg = event->attr.config & RAPL_EVENT_MASK;
331 int bit, msr, ret = 0;
333 /* only look at RAPL events */
334 if (event->attr.type != rapl_pmu_class.type)
337 /* check only supported bits are set */
338 if (event->attr.config & ~RAPL_EVENT_MASK)
342 * check event is known (determines counter)
346 bit = RAPL_IDX_PP0_NRG_STAT;
347 msr = MSR_PP0_ENERGY_STATUS;
350 bit = RAPL_IDX_PKG_NRG_STAT;
351 msr = MSR_PKG_ENERGY_STATUS;
354 bit = RAPL_IDX_RAM_NRG_STAT;
355 msr = MSR_DRAM_ENERGY_STATUS;
358 bit = RAPL_IDX_PP1_NRG_STAT;
359 msr = MSR_PP1_ENERGY_STATUS;
364 /* check event supported */
365 if (!(rapl_cntr_mask & (1 << bit)))
368 /* unsupported modes and filters */
369 if (event->attr.exclude_user ||
370 event->attr.exclude_kernel ||
371 event->attr.exclude_hv ||
372 event->attr.exclude_idle ||
373 event->attr.exclude_host ||
374 event->attr.exclude_guest ||
375 event->attr.sample_period) /* no sampling */
378 /* must be done before validate_group */
379 event->hw.event_base = msr;
380 event->hw.config = cfg;
386 static void rapl_pmu_event_read(struct perf_event *event)
388 rapl_event_update(event);
391 static ssize_t rapl_get_attr_cpumask(struct device *dev,
392 struct device_attribute *attr, char *buf)
394 return cpumap_print_to_pagebuf(true, buf, &rapl_cpu_mask);
397 static DEVICE_ATTR(cpumask, S_IRUGO, rapl_get_attr_cpumask, NULL);
399 static struct attribute *rapl_pmu_attrs[] = {
400 &dev_attr_cpumask.attr,
404 static struct attribute_group rapl_pmu_attr_group = {
405 .attrs = rapl_pmu_attrs,
408 static ssize_t rapl_sysfs_show(struct device *dev,
409 struct device_attribute *attr,
412 struct perf_pmu_events_attr *pmu_attr = \
413 container_of(attr, struct perf_pmu_events_attr, attr);
415 if (pmu_attr->event_str)
416 return sprintf(page, "%s", pmu_attr->event_str);
421 RAPL_EVENT_ATTR_STR(energy-cores, rapl_cores, "event=0x01");
422 RAPL_EVENT_ATTR_STR(energy-pkg , rapl_pkg, "event=0x02");
423 RAPL_EVENT_ATTR_STR(energy-ram , rapl_ram, "event=0x03");
424 RAPL_EVENT_ATTR_STR(energy-gpu , rapl_gpu, "event=0x04");
426 RAPL_EVENT_ATTR_STR(energy-cores.unit, rapl_cores_unit, "Joules");
427 RAPL_EVENT_ATTR_STR(energy-pkg.unit , rapl_pkg_unit, "Joules");
428 RAPL_EVENT_ATTR_STR(energy-ram.unit , rapl_ram_unit, "Joules");
429 RAPL_EVENT_ATTR_STR(energy-gpu.unit , rapl_gpu_unit, "Joules");
432 * we compute in 0.23 nJ increments regardless of MSR
434 RAPL_EVENT_ATTR_STR(energy-cores.scale, rapl_cores_scale, "2.3283064365386962890625e-10");
435 RAPL_EVENT_ATTR_STR(energy-pkg.scale, rapl_pkg_scale, "2.3283064365386962890625e-10");
436 RAPL_EVENT_ATTR_STR(energy-ram.scale, rapl_ram_scale, "2.3283064365386962890625e-10");
437 RAPL_EVENT_ATTR_STR(energy-gpu.scale, rapl_gpu_scale, "2.3283064365386962890625e-10");
439 static struct attribute *rapl_events_srv_attr[] = {
440 EVENT_PTR(rapl_cores),
444 EVENT_PTR(rapl_cores_unit),
445 EVENT_PTR(rapl_pkg_unit),
446 EVENT_PTR(rapl_ram_unit),
448 EVENT_PTR(rapl_cores_scale),
449 EVENT_PTR(rapl_pkg_scale),
450 EVENT_PTR(rapl_ram_scale),
454 static struct attribute *rapl_events_cln_attr[] = {
455 EVENT_PTR(rapl_cores),
459 EVENT_PTR(rapl_cores_unit),
460 EVENT_PTR(rapl_pkg_unit),
461 EVENT_PTR(rapl_gpu_unit),
463 EVENT_PTR(rapl_cores_scale),
464 EVENT_PTR(rapl_pkg_scale),
465 EVENT_PTR(rapl_gpu_scale),
469 static struct attribute *rapl_events_hsw_attr[] = {
470 EVENT_PTR(rapl_cores),
475 EVENT_PTR(rapl_cores_unit),
476 EVENT_PTR(rapl_pkg_unit),
477 EVENT_PTR(rapl_gpu_unit),
478 EVENT_PTR(rapl_ram_unit),
480 EVENT_PTR(rapl_cores_scale),
481 EVENT_PTR(rapl_pkg_scale),
482 EVENT_PTR(rapl_gpu_scale),
483 EVENT_PTR(rapl_ram_scale),
487 static struct attribute *rapl_events_knl_attr[] = {
491 EVENT_PTR(rapl_pkg_unit),
492 EVENT_PTR(rapl_ram_unit),
494 EVENT_PTR(rapl_pkg_scale),
495 EVENT_PTR(rapl_ram_scale),
499 static struct attribute_group rapl_pmu_events_group = {
501 .attrs = NULL, /* patched at runtime */
504 DEFINE_RAPL_FORMAT_ATTR(event, event, "config:0-7");
505 static struct attribute *rapl_formats_attr[] = {
506 &format_attr_event.attr,
510 static struct attribute_group rapl_pmu_format_group = {
512 .attrs = rapl_formats_attr,
515 const struct attribute_group *rapl_attr_groups[] = {
516 &rapl_pmu_attr_group,
517 &rapl_pmu_format_group,
518 &rapl_pmu_events_group,
522 static struct pmu rapl_pmu_class = {
523 .attr_groups = rapl_attr_groups,
524 .task_ctx_nr = perf_invalid_context, /* system-wide only */
525 .event_init = rapl_pmu_event_init,
526 .add = rapl_pmu_event_add, /* must have */
527 .del = rapl_pmu_event_del, /* must have */
528 .start = rapl_pmu_event_start,
529 .stop = rapl_pmu_event_stop,
530 .read = rapl_pmu_event_read,
533 static void rapl_cpu_exit(int cpu)
535 struct rapl_pmu *pmu = per_cpu(rapl_pmu, cpu);
536 int i, phys_id = topology_physical_package_id(cpu);
539 /* find a new cpu on same package */
540 for_each_online_cpu(i) {
543 if (phys_id == topology_physical_package_id(i)) {
549 * clear cpu from cpumask
550 * if was set in cpumask and still some cpu on package,
551 * then move to new cpu
553 if (cpumask_test_and_clear_cpu(cpu, &rapl_cpu_mask) && target >= 0)
554 cpumask_set_cpu(target, &rapl_cpu_mask);
556 WARN_ON(cpumask_empty(&rapl_cpu_mask));
558 * migrate events and context to new cpu
561 perf_pmu_migrate_context(pmu->pmu, cpu, target);
563 /* cancel overflow polling timer for CPU */
564 rapl_stop_hrtimer(pmu);
567 static void rapl_cpu_init(int cpu)
569 int i, phys_id = topology_physical_package_id(cpu);
571 /* check if phys_is is already covered */
572 for_each_cpu(i, &rapl_cpu_mask) {
573 if (phys_id == topology_physical_package_id(i))
576 /* was not found, so add it */
577 cpumask_set_cpu(cpu, &rapl_cpu_mask);
580 static __init void rapl_hsw_server_quirk(void)
583 * DRAM domain on HSW server has fixed energy unit which can be
584 * different than the unit from power unit MSR.
585 * "Intel Xeon Processor E5-1600 and E5-2600 v3 Product Families, V2
586 * of 2. Datasheet, September 2014, Reference Number: 330784-001 "
588 rapl_hw_unit[RAPL_IDX_RAM_NRG_STAT] = 16;
591 static int rapl_cpu_prepare(int cpu)
593 struct rapl_pmu *pmu = per_cpu(rapl_pmu, cpu);
594 int phys_id = topology_physical_package_id(cpu);
603 pmu = kzalloc_node(sizeof(*pmu), GFP_KERNEL, cpu_to_node(cpu));
606 raw_spin_lock_init(&pmu->lock);
608 INIT_LIST_HEAD(&pmu->active_list);
610 pmu->pmu = &rapl_pmu_class;
613 * use reference of 200W for scaling the timeout
614 * to avoid missing counter overflows.
615 * 200W = 200 Joules/sec
616 * divide interval by 2 to avoid lockstep (2 * 100)
617 * if hw unit is 32, then we use 2 ms 1/200/2
619 if (rapl_hw_unit[0] < 32)
620 ms = (1000 / (2 * 100)) * (1ULL << (32 - rapl_hw_unit[0] - 1));
624 pmu->timer_interval = ms_to_ktime(ms);
626 rapl_hrtimer_init(pmu);
628 /* set RAPL pmu for this cpu for now */
629 per_cpu(rapl_pmu, cpu) = pmu;
630 per_cpu(rapl_pmu_to_free, cpu) = NULL;
635 static void rapl_cpu_kfree(int cpu)
637 struct rapl_pmu *pmu = per_cpu(rapl_pmu_to_free, cpu);
641 per_cpu(rapl_pmu_to_free, cpu) = NULL;
644 static int rapl_cpu_dying(int cpu)
646 struct rapl_pmu *pmu = per_cpu(rapl_pmu, cpu);
651 per_cpu(rapl_pmu, cpu) = NULL;
653 per_cpu(rapl_pmu_to_free, cpu) = pmu;
658 static int rapl_cpu_notifier(struct notifier_block *self,
659 unsigned long action, void *hcpu)
661 unsigned int cpu = (long)hcpu;
663 switch (action & ~CPU_TASKS_FROZEN) {
665 rapl_cpu_prepare(cpu);
670 case CPU_UP_CANCELED:
678 case CPU_DOWN_PREPARE:
688 static int rapl_check_hw_unit(void)
690 u64 msr_rapl_power_unit_bits;
693 /* protect rdmsrl() to handle virtualization */
694 if (rdmsrl_safe(MSR_RAPL_POWER_UNIT, &msr_rapl_power_unit_bits))
696 for (i = 0; i < NR_RAPL_DOMAINS; i++)
697 rapl_hw_unit[i] = (msr_rapl_power_unit_bits >> 8) & 0x1FULL;
702 static const struct x86_cpu_id rapl_cpu_match[] = {
703 [0] = { .vendor = X86_VENDOR_INTEL, .family = 6 },
707 static int __init rapl_pmu_init(void)
709 struct rapl_pmu *pmu;
711 struct x86_pmu_quirk *quirk;
715 * check for Intel processor family 6
717 if (!x86_match_cpu(rapl_cpu_match))
720 /* check supported CPU */
721 switch (boot_cpu_data.x86_model) {
722 case 42: /* Sandy Bridge */
723 case 58: /* Ivy Bridge */
724 rapl_cntr_mask = RAPL_IDX_CLN;
725 rapl_pmu_events_group.attrs = rapl_events_cln_attr;
727 case 63: /* Haswell-Server */
728 rapl_add_quirk(rapl_hsw_server_quirk);
729 rapl_cntr_mask = RAPL_IDX_SRV;
730 rapl_pmu_events_group.attrs = rapl_events_srv_attr;
732 case 60: /* Haswell */
733 case 69: /* Haswell-Celeron */
734 case 61: /* Broadwell */
735 rapl_cntr_mask = RAPL_IDX_HSW;
736 rapl_pmu_events_group.attrs = rapl_events_hsw_attr;
738 case 45: /* Sandy Bridge-EP */
739 case 62: /* IvyTown */
740 rapl_cntr_mask = RAPL_IDX_SRV;
741 rapl_pmu_events_group.attrs = rapl_events_srv_attr;
743 case 87: /* Knights Landing */
744 rapl_add_quirk(rapl_hsw_server_quirk);
745 rapl_cntr_mask = RAPL_IDX_KNL;
746 rapl_pmu_events_group.attrs = rapl_events_knl_attr;
752 ret = rapl_check_hw_unit();
756 /* run cpu model quirks */
757 for (quirk = rapl_quirks; quirk; quirk = quirk->next)
759 cpu_notifier_register_begin();
761 for_each_online_cpu(cpu) {
762 ret = rapl_cpu_prepare(cpu);
768 __perf_cpu_notifier(rapl_cpu_notifier);
770 ret = perf_pmu_register(&rapl_pmu_class, "power", -1);
772 pr_info("RAPL PMU detected, registration failed (%d), RAPL PMU disabled\n", ret);
773 cpu_notifier_register_done();
777 pmu = __this_cpu_read(rapl_pmu);
779 pr_info("RAPL PMU detected,"
780 " API unit is 2^-32 Joules,"
782 " %llu ms ovfl timer\n",
783 hweight32(rapl_cntr_mask),
784 ktime_to_ms(pmu->timer_interval));
785 for (i = 0; i < NR_RAPL_DOMAINS; i++) {
786 if (rapl_cntr_mask & (1 << i)) {
787 pr_info("hw unit of domain %s 2^-%d Joules\n",
788 rapl_domain_names[i], rapl_hw_unit[i]);
792 cpu_notifier_register_done();
796 device_initcall(rapl_pmu_init);