1/*
2 * perf_event_intel_rapl.c: support Intel RAPL energy consumption counters
3 * Copyright (C) 2013 Google, Inc., Stephane Eranian
4 *
5 * Intel RAPL interface is specified in the IA-32 Manual Vol3b
6 * section 14.7.1 (September 2013)
7 *
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).
11 *
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
14 * but it can vary.
15 *
16 * Counter to rapl events mappings:
17 *
18 *  pp0 counter: consumption of all physical cores (power plane 0)
19 * 	  event: rapl_energy_cores
20 *    perf code: 0x1
21 *
22 *  pkg counter: consumption of the whole processor package
23 *	  event: rapl_energy_pkg
24 *    perf code: 0x2
25 *
26 * dram counter: consumption of the dram domain (servers only)
27 *	  event: rapl_energy_dram
28 *    perf code: 0x3
29 *
30 * dram counter: consumption of the builtin-gpu domain (client only)
31 *	  event: rapl_energy_gpu
32 *    perf code: 0x4
33 *
34 * We manage those counters as free running (read-only). They may be
35 * use simultaneously by other tools, such as turbostat.
36 *
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.
40 *
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);
46 */
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"
52
53/*
54 * RAPL energy status counters
55 */
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 */
64
65#define NR_RAPL_DOMAINS         0x4
66static const char *rapl_domain_names[NR_RAPL_DOMAINS] __initconst = {
67	"pp0-core",
68	"package",
69	"dram",
70	"pp1-gpu",
71};
72
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)
77
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)
82
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)
88
89/* Knights Landing has PKG, RAM */
90#define RAPL_IDX_KNL	(1<<RAPL_IDX_PKG_NRG_STAT|\
91			 1<<RAPL_IDX_RAM_NRG_STAT)
92
93/*
94 * event code: LSB 8 bits, passed in attr->config
95 * any other bit is reserved
96 */
97#define RAPL_EVENT_MASK	0xFFULL
98
99#define DEFINE_RAPL_FORMAT_ATTR(_var, _name, _format)		\
100static ssize_t __rapl_##_var##_show(struct kobject *kobj,	\
101				struct kobj_attribute *attr,	\
102				char *page)			\
103{								\
104	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);		\
105	return sprintf(page, _format "\n");			\
106}								\
107static struct kobj_attribute format_attr_##_var =		\
108	__ATTR(_name, 0444, __rapl_##_var##_show, NULL)
109
110#define RAPL_CNTR_WIDTH 32 /* 32-bit rapl counters */
111
112#define RAPL_EVENT_ATTR_STR(_name, v, str)				\
113static struct perf_pmu_events_attr event_attr_##v = {			\
114	.attr		= __ATTR(_name, 0444, rapl_sysfs_show, NULL),	\
115	.id		= 0,						\
116	.event_str	= str,						\
117};
118
119struct rapl_pmu {
120	spinlock_t	 lock;
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;
126};
127
128static int rapl_hw_unit[NR_RAPL_DOMAINS] __read_mostly;  /* 1/2^hw_unit Joule */
129static struct pmu rapl_pmu_class;
130static cpumask_t rapl_cpu_mask;
131static int rapl_cntr_mask;
132
133static DEFINE_PER_CPU(struct rapl_pmu *, rapl_pmu);
134static DEFINE_PER_CPU(struct rapl_pmu *, rapl_pmu_to_free);
135
136static struct x86_pmu_quirk *rapl_quirks;
137static inline u64 rapl_read_counter(struct perf_event *event)
138{
139	u64 raw;
140	rdmsrl(event->hw.event_base, raw);
141	return raw;
142}
143
144#define rapl_add_quirk(func_)						\
145do {									\
146	static struct x86_pmu_quirk __quirk __initdata = {		\
147		.func = func_,						\
148	};								\
149	__quirk.next = rapl_quirks;					\
150	rapl_quirks = &__quirk;						\
151} while (0)
152
153static inline u64 rapl_scale(u64 v, int cfg)
154{
155	if (cfg > NR_RAPL_DOMAINS) {
156		pr_warn("invalid domain %d, failed to scale data\n", cfg);
157		return v;
158	}
159	/*
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
164	 */
165	return v << (32 - rapl_hw_unit[cfg - 1]);
166}
167
168static u64 rapl_event_update(struct perf_event *event)
169{
170	struct hw_perf_event *hwc = &event->hw;
171	u64 prev_raw_count, new_raw_count;
172	s64 delta, sdelta;
173	int shift = RAPL_CNTR_WIDTH;
174
175again:
176	prev_raw_count = local64_read(&hwc->prev_count);
177	rdmsrl(event->hw.event_base, new_raw_count);
178
179	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
180			    new_raw_count) != prev_raw_count) {
181		cpu_relax();
182		goto again;
183	}
184
185	/*
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.
189	 *
190	 * Careful, not all hw sign-extends above the physical width
191	 * of the count.
192	 */
193	delta = (new_raw_count << shift) - (prev_raw_count << shift);
194	delta >>= shift;
195
196	sdelta = rapl_scale(delta, event->hw.config);
197
198	local64_add(sdelta, &event->count);
199
200	return new_raw_count;
201}
202
203static void rapl_start_hrtimer(struct rapl_pmu *pmu)
204{
205       hrtimer_start(&pmu->hrtimer, pmu->timer_interval,
206		     HRTIMER_MODE_REL_PINNED);
207}
208
209static void rapl_stop_hrtimer(struct rapl_pmu *pmu)
210{
211	hrtimer_cancel(&pmu->hrtimer);
212}
213
214static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
215{
216	struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
217	struct perf_event *event;
218	unsigned long flags;
219
220	if (!pmu->n_active)
221		return HRTIMER_NORESTART;
222
223	spin_lock_irqsave(&pmu->lock, flags);
224
225	list_for_each_entry(event, &pmu->active_list, active_entry) {
226		rapl_event_update(event);
227	}
228
229	spin_unlock_irqrestore(&pmu->lock, flags);
230
231	hrtimer_forward_now(hrtimer, pmu->timer_interval);
232
233	return HRTIMER_RESTART;
234}
235
236static void rapl_hrtimer_init(struct rapl_pmu *pmu)
237{
238	struct hrtimer *hr = &pmu->hrtimer;
239
240	hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
241	hr->function = rapl_hrtimer_handle;
242}
243
244static void __rapl_pmu_event_start(struct rapl_pmu *pmu,
245				   struct perf_event *event)
246{
247	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
248		return;
249
250	event->hw.state = 0;
251
252	list_add_tail(&event->active_entry, &pmu->active_list);
253
254	local64_set(&event->hw.prev_count, rapl_read_counter(event));
255
256	pmu->n_active++;
257	if (pmu->n_active == 1)
258		rapl_start_hrtimer(pmu);
259}
260
261static void rapl_pmu_event_start(struct perf_event *event, int mode)
262{
263	struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
264	unsigned long flags;
265
266	spin_lock_irqsave(&pmu->lock, flags);
267	__rapl_pmu_event_start(pmu, event);
268	spin_unlock_irqrestore(&pmu->lock, flags);
269}
270
271static void rapl_pmu_event_stop(struct perf_event *event, int mode)
272{
273	struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
274	struct hw_perf_event *hwc = &event->hw;
275	unsigned long flags;
276
277	spin_lock_irqsave(&pmu->lock, flags);
278
279	/* mark event as deactivated and stopped */
280	if (!(hwc->state & PERF_HES_STOPPED)) {
281		WARN_ON_ONCE(pmu->n_active <= 0);
282		pmu->n_active--;
283		if (pmu->n_active == 0)
284			rapl_stop_hrtimer(pmu);
285
286		list_del(&event->active_entry);
287
288		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
289		hwc->state |= PERF_HES_STOPPED;
290	}
291
292	/* check if update of sw counter is necessary */
293	if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
294		/*
295		 * Drain the remaining delta count out of a event
296		 * that we are disabling:
297		 */
298		rapl_event_update(event);
299		hwc->state |= PERF_HES_UPTODATE;
300	}
301
302	spin_unlock_irqrestore(&pmu->lock, flags);
303}
304
305static int rapl_pmu_event_add(struct perf_event *event, int mode)
306{
307	struct rapl_pmu *pmu = __this_cpu_read(rapl_pmu);
308	struct hw_perf_event *hwc = &event->hw;
309	unsigned long flags;
310
311	spin_lock_irqsave(&pmu->lock, flags);
312
313	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
314
315	if (mode & PERF_EF_START)
316		__rapl_pmu_event_start(pmu, event);
317
318	spin_unlock_irqrestore(&pmu->lock, flags);
319
320	return 0;
321}
322
323static void rapl_pmu_event_del(struct perf_event *event, int flags)
324{
325	rapl_pmu_event_stop(event, PERF_EF_UPDATE);
326}
327
328static int rapl_pmu_event_init(struct perf_event *event)
329{
330	u64 cfg = event->attr.config & RAPL_EVENT_MASK;
331	int bit, msr, ret = 0;
332
333	/* only look at RAPL events */
334	if (event->attr.type != rapl_pmu_class.type)
335		return -ENOENT;
336
337	/* check only supported bits are set */
338	if (event->attr.config & ~RAPL_EVENT_MASK)
339		return -EINVAL;
340
341	/*
342	 * check event is known (determines counter)
343	 */
344	switch (cfg) {
345	case INTEL_RAPL_PP0:
346		bit = RAPL_IDX_PP0_NRG_STAT;
347		msr = MSR_PP0_ENERGY_STATUS;
348		break;
349	case INTEL_RAPL_PKG:
350		bit = RAPL_IDX_PKG_NRG_STAT;
351		msr = MSR_PKG_ENERGY_STATUS;
352		break;
353	case INTEL_RAPL_RAM:
354		bit = RAPL_IDX_RAM_NRG_STAT;
355		msr = MSR_DRAM_ENERGY_STATUS;
356		break;
357	case INTEL_RAPL_PP1:
358		bit = RAPL_IDX_PP1_NRG_STAT;
359		msr = MSR_PP1_ENERGY_STATUS;
360		break;
361	default:
362		return -EINVAL;
363	}
364	/* check event supported */
365	if (!(rapl_cntr_mask & (1 << bit)))
366		return -EINVAL;
367
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 */
376		return -EINVAL;
377
378	/* must be done before validate_group */
379	event->hw.event_base = msr;
380	event->hw.config = cfg;
381	event->hw.idx = bit;
382
383	return ret;
384}
385
386static void rapl_pmu_event_read(struct perf_event *event)
387{
388	rapl_event_update(event);
389}
390
391static ssize_t rapl_get_attr_cpumask(struct device *dev,
392				struct device_attribute *attr, char *buf)
393{
394	return cpumap_print_to_pagebuf(true, buf, &rapl_cpu_mask);
395}
396
397static DEVICE_ATTR(cpumask, S_IRUGO, rapl_get_attr_cpumask, NULL);
398
399static struct attribute *rapl_pmu_attrs[] = {
400	&dev_attr_cpumask.attr,
401	NULL,
402};
403
404static struct attribute_group rapl_pmu_attr_group = {
405	.attrs = rapl_pmu_attrs,
406};
407
408static ssize_t rapl_sysfs_show(struct device *dev,
409			       struct device_attribute *attr,
410			       char *page)
411{
412	struct perf_pmu_events_attr *pmu_attr = \
413		container_of(attr, struct perf_pmu_events_attr, attr);
414
415	if (pmu_attr->event_str)
416		return sprintf(page, "%s", pmu_attr->event_str);
417
418	return 0;
419}
420
421RAPL_EVENT_ATTR_STR(energy-cores, rapl_cores, "event=0x01");
422RAPL_EVENT_ATTR_STR(energy-pkg  ,   rapl_pkg, "event=0x02");
423RAPL_EVENT_ATTR_STR(energy-ram  ,   rapl_ram, "event=0x03");
424RAPL_EVENT_ATTR_STR(energy-gpu  ,   rapl_gpu, "event=0x04");
425
426RAPL_EVENT_ATTR_STR(energy-cores.unit, rapl_cores_unit, "Joules");
427RAPL_EVENT_ATTR_STR(energy-pkg.unit  ,   rapl_pkg_unit, "Joules");
428RAPL_EVENT_ATTR_STR(energy-ram.unit  ,   rapl_ram_unit, "Joules");
429RAPL_EVENT_ATTR_STR(energy-gpu.unit  ,   rapl_gpu_unit, "Joules");
430
431/*
432 * we compute in 0.23 nJ increments regardless of MSR
433 */
434RAPL_EVENT_ATTR_STR(energy-cores.scale, rapl_cores_scale, "2.3283064365386962890625e-10");
435RAPL_EVENT_ATTR_STR(energy-pkg.scale,     rapl_pkg_scale, "2.3283064365386962890625e-10");
436RAPL_EVENT_ATTR_STR(energy-ram.scale,     rapl_ram_scale, "2.3283064365386962890625e-10");
437RAPL_EVENT_ATTR_STR(energy-gpu.scale,     rapl_gpu_scale, "2.3283064365386962890625e-10");
438
439static struct attribute *rapl_events_srv_attr[] = {
440	EVENT_PTR(rapl_cores),
441	EVENT_PTR(rapl_pkg),
442	EVENT_PTR(rapl_ram),
443
444	EVENT_PTR(rapl_cores_unit),
445	EVENT_PTR(rapl_pkg_unit),
446	EVENT_PTR(rapl_ram_unit),
447
448	EVENT_PTR(rapl_cores_scale),
449	EVENT_PTR(rapl_pkg_scale),
450	EVENT_PTR(rapl_ram_scale),
451	NULL,
452};
453
454static struct attribute *rapl_events_cln_attr[] = {
455	EVENT_PTR(rapl_cores),
456	EVENT_PTR(rapl_pkg),
457	EVENT_PTR(rapl_gpu),
458
459	EVENT_PTR(rapl_cores_unit),
460	EVENT_PTR(rapl_pkg_unit),
461	EVENT_PTR(rapl_gpu_unit),
462
463	EVENT_PTR(rapl_cores_scale),
464	EVENT_PTR(rapl_pkg_scale),
465	EVENT_PTR(rapl_gpu_scale),
466	NULL,
467};
468
469static struct attribute *rapl_events_hsw_attr[] = {
470	EVENT_PTR(rapl_cores),
471	EVENT_PTR(rapl_pkg),
472	EVENT_PTR(rapl_gpu),
473	EVENT_PTR(rapl_ram),
474
475	EVENT_PTR(rapl_cores_unit),
476	EVENT_PTR(rapl_pkg_unit),
477	EVENT_PTR(rapl_gpu_unit),
478	EVENT_PTR(rapl_ram_unit),
479
480	EVENT_PTR(rapl_cores_scale),
481	EVENT_PTR(rapl_pkg_scale),
482	EVENT_PTR(rapl_gpu_scale),
483	EVENT_PTR(rapl_ram_scale),
484	NULL,
485};
486
487static struct attribute *rapl_events_knl_attr[] = {
488	EVENT_PTR(rapl_pkg),
489	EVENT_PTR(rapl_ram),
490
491	EVENT_PTR(rapl_pkg_unit),
492	EVENT_PTR(rapl_ram_unit),
493
494	EVENT_PTR(rapl_pkg_scale),
495	EVENT_PTR(rapl_ram_scale),
496	NULL,
497};
498
499static struct attribute_group rapl_pmu_events_group = {
500	.name = "events",
501	.attrs = NULL, /* patched at runtime */
502};
503
504DEFINE_RAPL_FORMAT_ATTR(event, event, "config:0-7");
505static struct attribute *rapl_formats_attr[] = {
506	&format_attr_event.attr,
507	NULL,
508};
509
510static struct attribute_group rapl_pmu_format_group = {
511	.name = "format",
512	.attrs = rapl_formats_attr,
513};
514
515const struct attribute_group *rapl_attr_groups[] = {
516	&rapl_pmu_attr_group,
517	&rapl_pmu_format_group,
518	&rapl_pmu_events_group,
519	NULL,
520};
521
522static 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,
531};
532
533static void rapl_cpu_exit(int cpu)
534{
535	struct rapl_pmu *pmu = per_cpu(rapl_pmu, cpu);
536	int i, phys_id = topology_physical_package_id(cpu);
537	int target = -1;
538
539	/* find a new cpu on same package */
540	for_each_online_cpu(i) {
541		if (i == cpu)
542			continue;
543		if (phys_id == topology_physical_package_id(i)) {
544			target = i;
545			break;
546		}
547	}
548	/*
549	 * clear cpu from cpumask
550	 * if was set in cpumask and still some cpu on package,
551	 * then move to new cpu
552	 */
553	if (cpumask_test_and_clear_cpu(cpu, &rapl_cpu_mask) && target >= 0)
554		cpumask_set_cpu(target, &rapl_cpu_mask);
555
556	WARN_ON(cpumask_empty(&rapl_cpu_mask));
557	/*
558	 * migrate events and context to new cpu
559	 */
560	if (target >= 0)
561		perf_pmu_migrate_context(pmu->pmu, cpu, target);
562
563	/* cancel overflow polling timer for CPU */
564	rapl_stop_hrtimer(pmu);
565}
566
567static void rapl_cpu_init(int cpu)
568{
569	int i, phys_id = topology_physical_package_id(cpu);
570
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))
574			return;
575	}
576	/* was not found, so add it */
577	cpumask_set_cpu(cpu, &rapl_cpu_mask);
578}
579
580static __init void rapl_hsw_server_quirk(void)
581{
582	/*
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 "
587	 */
588	rapl_hw_unit[RAPL_IDX_RAM_NRG_STAT] = 16;
589}
590
591static int rapl_cpu_prepare(int cpu)
592{
593	struct rapl_pmu *pmu = per_cpu(rapl_pmu, cpu);
594	int phys_id = topology_physical_package_id(cpu);
595	u64 ms;
596
597	if (pmu)
598		return 0;
599
600	if (phys_id < 0)
601		return -1;
602
603	pmu = kzalloc_node(sizeof(*pmu), GFP_KERNEL, cpu_to_node(cpu));
604	if (!pmu)
605		return -1;
606	spin_lock_init(&pmu->lock);
607
608	INIT_LIST_HEAD(&pmu->active_list);
609
610	pmu->pmu = &rapl_pmu_class;
611
612	/*
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
618	 */
619	if (rapl_hw_unit[0] < 32)
620		ms = (1000 / (2 * 100)) * (1ULL << (32 - rapl_hw_unit[0] - 1));
621	else
622		ms = 2;
623
624	pmu->timer_interval = ms_to_ktime(ms);
625
626	rapl_hrtimer_init(pmu);
627
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;
631
632	return 0;
633}
634
635static void rapl_cpu_kfree(int cpu)
636{
637	struct rapl_pmu *pmu = per_cpu(rapl_pmu_to_free, cpu);
638
639	kfree(pmu);
640
641	per_cpu(rapl_pmu_to_free, cpu) = NULL;
642}
643
644static int rapl_cpu_dying(int cpu)
645{
646	struct rapl_pmu *pmu = per_cpu(rapl_pmu, cpu);
647
648	if (!pmu)
649		return 0;
650
651	per_cpu(rapl_pmu, cpu) = NULL;
652
653	per_cpu(rapl_pmu_to_free, cpu) = pmu;
654
655	return 0;
656}
657
658static int rapl_cpu_notifier(struct notifier_block *self,
659			     unsigned long action, void *hcpu)
660{
661	unsigned int cpu = (long)hcpu;
662
663	switch (action & ~CPU_TASKS_FROZEN) {
664	case CPU_UP_PREPARE:
665		rapl_cpu_prepare(cpu);
666		break;
667	case CPU_STARTING:
668		rapl_cpu_init(cpu);
669		break;
670	case CPU_UP_CANCELED:
671	case CPU_DYING:
672		rapl_cpu_dying(cpu);
673		break;
674	case CPU_ONLINE:
675	case CPU_DEAD:
676		rapl_cpu_kfree(cpu);
677		break;
678	case CPU_DOWN_PREPARE:
679		rapl_cpu_exit(cpu);
680		break;
681	default:
682		break;
683	}
684
685	return NOTIFY_OK;
686}
687
688static int rapl_check_hw_unit(void)
689{
690	u64 msr_rapl_power_unit_bits;
691	int i;
692
693	/* protect rdmsrl() to handle virtualization */
694	if (rdmsrl_safe(MSR_RAPL_POWER_UNIT, &msr_rapl_power_unit_bits))
695		return -1;
696	for (i = 0; i < NR_RAPL_DOMAINS; i++)
697		rapl_hw_unit[i] = (msr_rapl_power_unit_bits >> 8) & 0x1FULL;
698
699	return 0;
700}
701
702static const struct x86_cpu_id rapl_cpu_match[] = {
703	[0] = { .vendor = X86_VENDOR_INTEL, .family = 6 },
704	[1] = {},
705};
706
707static int __init rapl_pmu_init(void)
708{
709	struct rapl_pmu *pmu;
710	int cpu, ret;
711	struct x86_pmu_quirk *quirk;
712	int i;
713
714	/*
715	 * check for Intel processor family 6
716	 */
717	if (!x86_match_cpu(rapl_cpu_match))
718		return 0;
719
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;
726		break;
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;
731		break;
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;
737		break;
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;
742		break;
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;
747
748	default:
749		/* unsupported */
750		return 0;
751	}
752	ret = rapl_check_hw_unit();
753	if (ret)
754		return ret;
755
756	/* run cpu model quirks */
757	for (quirk = rapl_quirks; quirk; quirk = quirk->next)
758		quirk->func();
759	cpu_notifier_register_begin();
760
761	for_each_online_cpu(cpu) {
762		ret = rapl_cpu_prepare(cpu);
763		if (ret)
764			goto out;
765		rapl_cpu_init(cpu);
766	}
767
768	__perf_cpu_notifier(rapl_cpu_notifier);
769
770	ret = perf_pmu_register(&rapl_pmu_class, "power", -1);
771	if (WARN_ON(ret)) {
772		pr_info("RAPL PMU detected, registration failed (%d), RAPL PMU disabled\n", ret);
773		cpu_notifier_register_done();
774		return -1;
775	}
776
777	pmu = __this_cpu_read(rapl_pmu);
778
779	pr_info("RAPL PMU detected,"
780		" API unit is 2^-32 Joules,"
781		" %d fixed counters"
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]);
789		}
790	}
791out:
792	cpu_notifier_register_done();
793
794	return 0;
795}
796device_initcall(rapl_pmu_init);
797