1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Support Intel/AMD RAPL energy consumption counters
4  * Copyright (C) 2013 Google, Inc., Stephane Eranian
5  *
6  * Intel RAPL interface is specified in the IA-32 Manual Vol3b
7  * section 14.7.1 (September 2013)
8  *
9  * AMD RAPL interface for Fam17h is described in the public PPR:
10  * https://bugzilla.kernel.org/show_bug.cgi?id=206537
11  *
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).
15  *
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
18  * but it can vary.
19  *
20  * Counter to rapl events mappings:
21  *
22  *  pp0 counter: consumption of all physical cores (power plane 0)
23  * 	  event: rapl_energy_cores
24  *    perf code: 0x1
25  *
26  *  pkg counter: consumption of the whole processor package
27  *	  event: rapl_energy_pkg
28  *    perf code: 0x2
29  *
30  * dram counter: consumption of the dram domain (servers only)
31  *	  event: rapl_energy_dram
32  *    perf code: 0x3
33  *
34  * gpu counter: consumption of the builtin-gpu domain (client only)
35  *	  event: rapl_energy_gpu
36  *    perf code: 0x4
37  *
38  *  psys counter: consumption of the builtin-psys domain (client only)
39  *	  event: rapl_energy_psys
40  *    perf code: 0x5
41  *
42  * We manage those counters as free running (read-only). They may be
43  * use simultaneously by other tools, such as turbostat.
44  *
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.
48  *
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);
54  */
55 
56 #define pr_fmt(fmt) "RAPL PMU: " fmt
57 
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"
65 #include "probe.h"
66 
67 MODULE_LICENSE("GPL");
68 
69 /*
70  * RAPL energy status counters
71  */
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 */
78 
79 	PERF_RAPL_MAX,
80 	NR_RAPL_DOMAINS = PERF_RAPL_MAX,
81 };
82 
83 static const char *const rapl_domain_names[NR_RAPL_DOMAINS] __initconst = {
84 	"pp0-core",
85 	"package",
86 	"dram",
87 	"pp1-gpu",
88 	"psys",
89 };
90 
91 /*
92  * event code: LSB 8 bits, passed in attr->config
93  * any other bit is reserved
94  */
95 #define RAPL_EVENT_MASK	0xFFULL
96 #define RAPL_CNTR_WIDTH 32
97 
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),	\
101 	.id		= 0,							\
102 	.event_str	= str,							\
103 };
104 
105 struct rapl_pmu {
106 	raw_spinlock_t		lock;
107 	int			n_active;
108 	int			cpu;
109 	struct list_head	active_list;
110 	struct pmu		*pmu;
111 	ktime_t			timer_interval;
112 	struct hrtimer		hrtimer;
113 };
114 
115 struct rapl_pmus {
116 	struct pmu		pmu;
117 	unsigned int		maxdie;
118 	struct rapl_pmu		*pmus[];
119 };
120 
121 enum rapl_unit_quirk {
122 	RAPL_UNIT_QUIRK_NONE,
123 	RAPL_UNIT_QUIRK_INTEL_HSW,
124 	RAPL_UNIT_QUIRK_INTEL_SPR,
125 };
126 
127 struct rapl_model {
128 	struct perf_msr *rapl_msrs;
129 	unsigned long	events;
130 	unsigned int	msr_power_unit;
131 	enum rapl_unit_quirk	unit_quirk;
132 };
133 
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;
141 
cpu_to_rapl_pmu(unsigned int cpu)142 static inline struct rapl_pmu *cpu_to_rapl_pmu(unsigned int cpu)
143 {
144 	unsigned int dieid = topology_logical_die_id(cpu);
145 
146 	/*
147 	 * The unsigned check also catches the '-1' return value for non
148 	 * existent mappings in the topology map.
149 	 */
150 	return dieid < rapl_pmus->maxdie ? rapl_pmus->pmus[dieid] : NULL;
151 }
152 
rapl_read_counter(struct perf_event * event)153 static inline u64 rapl_read_counter(struct perf_event *event)
154 {
155 	u64 raw;
156 	rdmsrl(event->hw.event_base, raw);
157 	return raw;
158 }
159 
rapl_scale(u64 v,int cfg)160 static inline u64 rapl_scale(u64 v, int cfg)
161 {
162 	if (cfg > NR_RAPL_DOMAINS) {
163 		pr_warn("Invalid domain %d, failed to scale data\n", cfg);
164 		return v;
165 	}
166 	/*
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
171 	 */
172 	return v << (32 - rapl_hw_unit[cfg - 1]);
173 }
174 
rapl_event_update(struct perf_event * event)175 static u64 rapl_event_update(struct perf_event *event)
176 {
177 	struct hw_perf_event *hwc = &event->hw;
178 	u64 prev_raw_count, new_raw_count;
179 	s64 delta, sdelta;
180 	int shift = RAPL_CNTR_WIDTH;
181 
182 again:
183 	prev_raw_count = local64_read(&hwc->prev_count);
184 	rdmsrl(event->hw.event_base, new_raw_count);
185 
186 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
187 			    new_raw_count) != prev_raw_count) {
188 		cpu_relax();
189 		goto again;
190 	}
191 
192 	/*
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.
196 	 *
197 	 * Careful, not all hw sign-extends above the physical width
198 	 * of the count.
199 	 */
200 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
201 	delta >>= shift;
202 
203 	sdelta = rapl_scale(delta, event->hw.config);
204 
205 	local64_add(sdelta, &event->count);
206 
207 	return new_raw_count;
208 }
209 
rapl_start_hrtimer(struct rapl_pmu * pmu)210 static void rapl_start_hrtimer(struct rapl_pmu *pmu)
211 {
212        hrtimer_start(&pmu->hrtimer, pmu->timer_interval,
213 		     HRTIMER_MODE_REL_PINNED);
214 }
215 
rapl_hrtimer_handle(struct hrtimer * hrtimer)216 static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
217 {
218 	struct rapl_pmu *pmu = container_of(hrtimer, struct rapl_pmu, hrtimer);
219 	struct perf_event *event;
220 	unsigned long flags;
221 
222 	if (!pmu->n_active)
223 		return HRTIMER_NORESTART;
224 
225 	raw_spin_lock_irqsave(&pmu->lock, flags);
226 
227 	list_for_each_entry(event, &pmu->active_list, active_entry)
228 		rapl_event_update(event);
229 
230 	raw_spin_unlock_irqrestore(&pmu->lock, flags);
231 
232 	hrtimer_forward_now(hrtimer, pmu->timer_interval);
233 
234 	return HRTIMER_RESTART;
235 }
236 
rapl_hrtimer_init(struct rapl_pmu * pmu)237 static void rapl_hrtimer_init(struct rapl_pmu *pmu)
238 {
239 	struct hrtimer *hr = &pmu->hrtimer;
240 
241 	hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
242 	hr->function = rapl_hrtimer_handle;
243 }
244 
__rapl_pmu_event_start(struct rapl_pmu * pmu,struct perf_event * event)245 static void __rapl_pmu_event_start(struct rapl_pmu *pmu,
246 				   struct perf_event *event)
247 {
248 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
249 		return;
250 
251 	event->hw.state = 0;
252 
253 	list_add_tail(&event->active_entry, &pmu->active_list);
254 
255 	local64_set(&event->hw.prev_count, rapl_read_counter(event));
256 
257 	pmu->n_active++;
258 	if (pmu->n_active == 1)
259 		rapl_start_hrtimer(pmu);
260 }
261 
rapl_pmu_event_start(struct perf_event * event,int mode)262 static void rapl_pmu_event_start(struct perf_event *event, int mode)
263 {
264 	struct rapl_pmu *pmu = event->pmu_private;
265 	unsigned long flags;
266 
267 	raw_spin_lock_irqsave(&pmu->lock, flags);
268 	__rapl_pmu_event_start(pmu, event);
269 	raw_spin_unlock_irqrestore(&pmu->lock, flags);
270 }
271 
rapl_pmu_event_stop(struct perf_event * event,int mode)272 static void rapl_pmu_event_stop(struct perf_event *event, int mode)
273 {
274 	struct rapl_pmu *pmu = event->pmu_private;
275 	struct hw_perf_event *hwc = &event->hw;
276 	unsigned long flags;
277 
278 	raw_spin_lock_irqsave(&pmu->lock, flags);
279 
280 	/* mark event as deactivated and stopped */
281 	if (!(hwc->state & PERF_HES_STOPPED)) {
282 		WARN_ON_ONCE(pmu->n_active <= 0);
283 		pmu->n_active--;
284 		if (pmu->n_active == 0)
285 			hrtimer_cancel(&pmu->hrtimer);
286 
287 		list_del(&event->active_entry);
288 
289 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
290 		hwc->state |= PERF_HES_STOPPED;
291 	}
292 
293 	/* check if update of sw counter is necessary */
294 	if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
295 		/*
296 		 * Drain the remaining delta count out of a event
297 		 * that we are disabling:
298 		 */
299 		rapl_event_update(event);
300 		hwc->state |= PERF_HES_UPTODATE;
301 	}
302 
303 	raw_spin_unlock_irqrestore(&pmu->lock, flags);
304 }
305 
rapl_pmu_event_add(struct perf_event * event,int mode)306 static int rapl_pmu_event_add(struct perf_event *event, int mode)
307 {
308 	struct rapl_pmu *pmu = event->pmu_private;
309 	struct hw_perf_event *hwc = &event->hw;
310 	unsigned long flags;
311 
312 	raw_spin_lock_irqsave(&pmu->lock, flags);
313 
314 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
315 
316 	if (mode & PERF_EF_START)
317 		__rapl_pmu_event_start(pmu, event);
318 
319 	raw_spin_unlock_irqrestore(&pmu->lock, flags);
320 
321 	return 0;
322 }
323 
rapl_pmu_event_del(struct perf_event * event,int flags)324 static void rapl_pmu_event_del(struct perf_event *event, int flags)
325 {
326 	rapl_pmu_event_stop(event, PERF_EF_UPDATE);
327 }
328 
rapl_pmu_event_init(struct perf_event * event)329 static int rapl_pmu_event_init(struct perf_event *event)
330 {
331 	u64 cfg = event->attr.config & RAPL_EVENT_MASK;
332 	int bit, ret = 0;
333 	struct rapl_pmu *pmu;
334 
335 	/* only look at RAPL events */
336 	if (event->attr.type != rapl_pmus->pmu.type)
337 		return -ENOENT;
338 
339 	/* check only supported bits are set */
340 	if (event->attr.config & ~RAPL_EVENT_MASK)
341 		return -EINVAL;
342 
343 	if (event->cpu < 0)
344 		return -EINVAL;
345 
346 	event->event_caps |= PERF_EV_CAP_READ_ACTIVE_PKG;
347 
348 	if (!cfg || cfg >= NR_RAPL_DOMAINS + 1)
349 		return -EINVAL;
350 
351 	cfg = array_index_nospec((long)cfg, NR_RAPL_DOMAINS + 1);
352 	bit = cfg - 1;
353 
354 	/* check event supported */
355 	if (!(rapl_cntr_mask & (1 << bit)))
356 		return -EINVAL;
357 
358 	/* unsupported modes and filters */
359 	if (event->attr.sample_period) /* no sampling */
360 		return -EINVAL;
361 
362 	/* must be done before validate_group */
363 	pmu = cpu_to_rapl_pmu(event->cpu);
364 	if (!pmu)
365 		return -EINVAL;
366 	event->cpu = pmu->cpu;
367 	event->pmu_private = pmu;
368 	event->hw.event_base = rapl_msrs[bit].msr;
369 	event->hw.config = cfg;
370 	event->hw.idx = bit;
371 
372 	return ret;
373 }
374 
rapl_pmu_event_read(struct perf_event * event)375 static void rapl_pmu_event_read(struct perf_event *event)
376 {
377 	rapl_event_update(event);
378 }
379 
rapl_get_attr_cpumask(struct device * dev,struct device_attribute * attr,char * buf)380 static ssize_t rapl_get_attr_cpumask(struct device *dev,
381 				struct device_attribute *attr, char *buf)
382 {
383 	return cpumap_print_to_pagebuf(true, buf, &rapl_cpu_mask);
384 }
385 
386 static DEVICE_ATTR(cpumask, S_IRUGO, rapl_get_attr_cpumask, NULL);
387 
388 static struct attribute *rapl_pmu_attrs[] = {
389 	&dev_attr_cpumask.attr,
390 	NULL,
391 };
392 
393 static struct attribute_group rapl_pmu_attr_group = {
394 	.attrs = rapl_pmu_attrs,
395 };
396 
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");
402 
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");
408 
409 /*
410  * we compute in 0.23 nJ increments regardless of MSR
411  */
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");
417 
418 /*
419  * There are no default events, but we need to create
420  * "events" group (with empty attrs) before updating
421  * it with detected events.
422  */
423 static struct attribute *attrs_empty[] = {
424 	NULL,
425 };
426 
427 static struct attribute_group rapl_pmu_events_group = {
428 	.name = "events",
429 	.attrs = attrs_empty,
430 };
431 
432 PMU_FORMAT_ATTR(event, "config:0-7");
433 static struct attribute *rapl_formats_attr[] = {
434 	&format_attr_event.attr,
435 	NULL,
436 };
437 
438 static struct attribute_group rapl_pmu_format_group = {
439 	.name = "format",
440 	.attrs = rapl_formats_attr,
441 };
442 
443 static const struct attribute_group *rapl_attr_groups[] = {
444 	&rapl_pmu_attr_group,
445 	&rapl_pmu_format_group,
446 	&rapl_pmu_events_group,
447 	NULL,
448 };
449 
450 static struct attribute *rapl_events_cores[] = {
451 	EVENT_PTR(rapl_cores),
452 	EVENT_PTR(rapl_cores_unit),
453 	EVENT_PTR(rapl_cores_scale),
454 	NULL,
455 };
456 
457 static struct attribute_group rapl_events_cores_group = {
458 	.name  = "events",
459 	.attrs = rapl_events_cores,
460 };
461 
462 static struct attribute *rapl_events_pkg[] = {
463 	EVENT_PTR(rapl_pkg),
464 	EVENT_PTR(rapl_pkg_unit),
465 	EVENT_PTR(rapl_pkg_scale),
466 	NULL,
467 };
468 
469 static struct attribute_group rapl_events_pkg_group = {
470 	.name  = "events",
471 	.attrs = rapl_events_pkg,
472 };
473 
474 static struct attribute *rapl_events_ram[] = {
475 	EVENT_PTR(rapl_ram),
476 	EVENT_PTR(rapl_ram_unit),
477 	EVENT_PTR(rapl_ram_scale),
478 	NULL,
479 };
480 
481 static struct attribute_group rapl_events_ram_group = {
482 	.name  = "events",
483 	.attrs = rapl_events_ram,
484 };
485 
486 static struct attribute *rapl_events_gpu[] = {
487 	EVENT_PTR(rapl_gpu),
488 	EVENT_PTR(rapl_gpu_unit),
489 	EVENT_PTR(rapl_gpu_scale),
490 	NULL,
491 };
492 
493 static struct attribute_group rapl_events_gpu_group = {
494 	.name  = "events",
495 	.attrs = rapl_events_gpu,
496 };
497 
498 static struct attribute *rapl_events_psys[] = {
499 	EVENT_PTR(rapl_psys),
500 	EVENT_PTR(rapl_psys_unit),
501 	EVENT_PTR(rapl_psys_scale),
502 	NULL,
503 };
504 
505 static struct attribute_group rapl_events_psys_group = {
506 	.name  = "events",
507 	.attrs = rapl_events_psys,
508 };
509 
test_msr(int idx,void * data)510 static bool test_msr(int idx, void *data)
511 {
512 	return test_bit(idx, (unsigned long *) data);
513 }
514 
515 /* Only lower 32bits of the MSR represents the energy counter */
516 #define RAPL_MSR_MASK 0xFFFFFFFF
517 
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 },
524 };
525 
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 },
532 };
533 
534 /*
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
538  */
539 static struct perf_msr amd_rapl_msrs[] = {
540 	[PERF_RAPL_PP0]  = { 0, &rapl_events_cores_group, 0, false, 0 },
541 	[PERF_RAPL_PKG]  = { MSR_AMD_PKG_ENERGY_STATUS,  &rapl_events_pkg_group,   test_msr, false, RAPL_MSR_MASK },
542 	[PERF_RAPL_RAM]  = { 0, &rapl_events_ram_group,   0, false, 0 },
543 	[PERF_RAPL_PP1]  = { 0, &rapl_events_gpu_group,   0, false, 0 },
544 	[PERF_RAPL_PSYS] = { 0, &rapl_events_psys_group,  0, false, 0 },
545 };
546 
rapl_cpu_offline(unsigned int cpu)547 static int rapl_cpu_offline(unsigned int cpu)
548 {
549 	struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
550 	int target;
551 
552 	/* Check if exiting cpu is used for collecting rapl events */
553 	if (!cpumask_test_and_clear_cpu(cpu, &rapl_cpu_mask))
554 		return 0;
555 
556 	pmu->cpu = -1;
557 	/* Find a new cpu to collect rapl events */
558 	target = cpumask_any_but(topology_die_cpumask(cpu), cpu);
559 
560 	/* Migrate rapl events to the new target */
561 	if (target < nr_cpu_ids) {
562 		cpumask_set_cpu(target, &rapl_cpu_mask);
563 		pmu->cpu = target;
564 		perf_pmu_migrate_context(pmu->pmu, cpu, target);
565 	}
566 	return 0;
567 }
568 
rapl_cpu_online(unsigned int cpu)569 static int rapl_cpu_online(unsigned int cpu)
570 {
571 	struct rapl_pmu *pmu = cpu_to_rapl_pmu(cpu);
572 	int target;
573 
574 	if (!pmu) {
575 		pmu = kzalloc_node(sizeof(*pmu), GFP_KERNEL, cpu_to_node(cpu));
576 		if (!pmu)
577 			return -ENOMEM;
578 
579 		raw_spin_lock_init(&pmu->lock);
580 		INIT_LIST_HEAD(&pmu->active_list);
581 		pmu->pmu = &rapl_pmus->pmu;
582 		pmu->timer_interval = ms_to_ktime(rapl_timer_ms);
583 		rapl_hrtimer_init(pmu);
584 
585 		rapl_pmus->pmus[topology_logical_die_id(cpu)] = pmu;
586 	}
587 
588 	/*
589 	 * Check if there is an online cpu in the package which collects rapl
590 	 * events already.
591 	 */
592 	target = cpumask_any_and(&rapl_cpu_mask, topology_die_cpumask(cpu));
593 	if (target < nr_cpu_ids)
594 		return 0;
595 
596 	cpumask_set_cpu(cpu, &rapl_cpu_mask);
597 	pmu->cpu = cpu;
598 	return 0;
599 }
600 
rapl_check_hw_unit(struct rapl_model * rm)601 static int rapl_check_hw_unit(struct rapl_model *rm)
602 {
603 	u64 msr_rapl_power_unit_bits;
604 	int i;
605 
606 	/* protect rdmsrl() to handle virtualization */
607 	if (rdmsrl_safe(rm->msr_power_unit, &msr_rapl_power_unit_bits))
608 		return -1;
609 	for (i = 0; i < NR_RAPL_DOMAINS; i++)
610 		rapl_hw_unit[i] = (msr_rapl_power_unit_bits >> 8) & 0x1FULL;
611 
612 	switch (rm->unit_quirk) {
613 	/*
614 	 * DRAM domain on HSW server and KNL has fixed energy unit which can be
615 	 * different than the unit from power unit MSR. See
616 	 * "Intel Xeon Processor E5-1600 and E5-2600 v3 Product Families, V2
617 	 * of 2. Datasheet, September 2014, Reference Number: 330784-001 "
618 	 */
619 	case RAPL_UNIT_QUIRK_INTEL_HSW:
620 		rapl_hw_unit[PERF_RAPL_RAM] = 16;
621 		break;
622 	/* SPR uses a fixed energy unit for Psys domain. */
623 	case RAPL_UNIT_QUIRK_INTEL_SPR:
624 		rapl_hw_unit[PERF_RAPL_PSYS] = 0;
625 		break;
626 	default:
627 		break;
628 	}
629 
630 
631 	/*
632 	 * Calculate the timer rate:
633 	 * Use reference of 200W for scaling the timeout to avoid counter
634 	 * overflows. 200W = 200 Joules/sec
635 	 * Divide interval by 2 to avoid lockstep (2 * 100)
636 	 * if hw unit is 32, then we use 2 ms 1/200/2
637 	 */
638 	rapl_timer_ms = 2;
639 	if (rapl_hw_unit[0] < 32) {
640 		rapl_timer_ms = (1000 / (2 * 100));
641 		rapl_timer_ms *= (1ULL << (32 - rapl_hw_unit[0] - 1));
642 	}
643 	return 0;
644 }
645 
rapl_advertise(void)646 static void __init rapl_advertise(void)
647 {
648 	int i;
649 
650 	pr_info("API unit is 2^-32 Joules, %d fixed counters, %llu ms ovfl timer\n",
651 		hweight32(rapl_cntr_mask), rapl_timer_ms);
652 
653 	for (i = 0; i < NR_RAPL_DOMAINS; i++) {
654 		if (rapl_cntr_mask & (1 << i)) {
655 			pr_info("hw unit of domain %s 2^-%d Joules\n",
656 				rapl_domain_names[i], rapl_hw_unit[i]);
657 		}
658 	}
659 }
660 
cleanup_rapl_pmus(void)661 static void cleanup_rapl_pmus(void)
662 {
663 	int i;
664 
665 	for (i = 0; i < rapl_pmus->maxdie; i++)
666 		kfree(rapl_pmus->pmus[i]);
667 	kfree(rapl_pmus);
668 }
669 
670 static const struct attribute_group *rapl_attr_update[] = {
671 	&rapl_events_cores_group,
672 	&rapl_events_pkg_group,
673 	&rapl_events_ram_group,
674 	&rapl_events_gpu_group,
675 	&rapl_events_psys_group,
676 	NULL,
677 };
678 
init_rapl_pmus(void)679 static int __init init_rapl_pmus(void)
680 {
681 	int maxdie = topology_max_packages() * topology_max_die_per_package();
682 	size_t size;
683 
684 	size = sizeof(*rapl_pmus) + maxdie * sizeof(struct rapl_pmu *);
685 	rapl_pmus = kzalloc(size, GFP_KERNEL);
686 	if (!rapl_pmus)
687 		return -ENOMEM;
688 
689 	rapl_pmus->maxdie		= maxdie;
690 	rapl_pmus->pmu.attr_groups	= rapl_attr_groups;
691 	rapl_pmus->pmu.attr_update	= rapl_attr_update;
692 	rapl_pmus->pmu.task_ctx_nr	= perf_invalid_context;
693 	rapl_pmus->pmu.event_init	= rapl_pmu_event_init;
694 	rapl_pmus->pmu.add		= rapl_pmu_event_add;
695 	rapl_pmus->pmu.del		= rapl_pmu_event_del;
696 	rapl_pmus->pmu.start		= rapl_pmu_event_start;
697 	rapl_pmus->pmu.stop		= rapl_pmu_event_stop;
698 	rapl_pmus->pmu.read		= rapl_pmu_event_read;
699 	rapl_pmus->pmu.module		= THIS_MODULE;
700 	rapl_pmus->pmu.capabilities	= PERF_PMU_CAP_NO_EXCLUDE;
701 	return 0;
702 }
703 
704 static struct rapl_model model_snb = {
705 	.events		= BIT(PERF_RAPL_PP0) |
706 			  BIT(PERF_RAPL_PKG) |
707 			  BIT(PERF_RAPL_PP1),
708 	.msr_power_unit = MSR_RAPL_POWER_UNIT,
709 	.rapl_msrs      = intel_rapl_msrs,
710 };
711 
712 static struct rapl_model model_snbep = {
713 	.events		= BIT(PERF_RAPL_PP0) |
714 			  BIT(PERF_RAPL_PKG) |
715 			  BIT(PERF_RAPL_RAM),
716 	.msr_power_unit = MSR_RAPL_POWER_UNIT,
717 	.rapl_msrs      = intel_rapl_msrs,
718 };
719 
720 static struct rapl_model model_hsw = {
721 	.events		= BIT(PERF_RAPL_PP0) |
722 			  BIT(PERF_RAPL_PKG) |
723 			  BIT(PERF_RAPL_RAM) |
724 			  BIT(PERF_RAPL_PP1),
725 	.msr_power_unit = MSR_RAPL_POWER_UNIT,
726 	.rapl_msrs      = intel_rapl_msrs,
727 };
728 
729 static struct rapl_model model_hsx = {
730 	.events		= BIT(PERF_RAPL_PP0) |
731 			  BIT(PERF_RAPL_PKG) |
732 			  BIT(PERF_RAPL_RAM),
733 	.unit_quirk	= RAPL_UNIT_QUIRK_INTEL_HSW,
734 	.msr_power_unit = MSR_RAPL_POWER_UNIT,
735 	.rapl_msrs      = intel_rapl_msrs,
736 };
737 
738 static struct rapl_model model_knl = {
739 	.events		= BIT(PERF_RAPL_PKG) |
740 			  BIT(PERF_RAPL_RAM),
741 	.unit_quirk	= RAPL_UNIT_QUIRK_INTEL_HSW,
742 	.msr_power_unit = MSR_RAPL_POWER_UNIT,
743 	.rapl_msrs      = intel_rapl_msrs,
744 };
745 
746 static struct rapl_model model_skl = {
747 	.events		= BIT(PERF_RAPL_PP0) |
748 			  BIT(PERF_RAPL_PKG) |
749 			  BIT(PERF_RAPL_RAM) |
750 			  BIT(PERF_RAPL_PP1) |
751 			  BIT(PERF_RAPL_PSYS),
752 	.msr_power_unit = MSR_RAPL_POWER_UNIT,
753 	.rapl_msrs      = intel_rapl_msrs,
754 };
755 
756 static struct rapl_model model_spr = {
757 	.events		= BIT(PERF_RAPL_PP0) |
758 			  BIT(PERF_RAPL_PKG) |
759 			  BIT(PERF_RAPL_RAM) |
760 			  BIT(PERF_RAPL_PSYS),
761 	.unit_quirk	= RAPL_UNIT_QUIRK_INTEL_SPR,
762 	.msr_power_unit = MSR_RAPL_POWER_UNIT,
763 	.rapl_msrs      = intel_rapl_spr_msrs,
764 };
765 
766 static struct rapl_model model_amd_hygon = {
767 	.events		= BIT(PERF_RAPL_PKG),
768 	.msr_power_unit = MSR_AMD_RAPL_POWER_UNIT,
769 	.rapl_msrs      = amd_rapl_msrs,
770 };
771 
772 static const struct x86_cpu_id rapl_model_match[] __initconst = {
773 	X86_MATCH_FEATURE(X86_FEATURE_RAPL,		&model_amd_hygon),
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(TIGERLAKE_L,		&model_skl),
804 	X86_MATCH_INTEL_FAM6_MODEL(TIGERLAKE,		&model_skl),
805 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE,		&model_skl),
806 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L,		&model_skl),
807 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_N,		&model_skl),
808 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X,	&model_spr),
809 	X86_MATCH_INTEL_FAM6_MODEL(EMERALDRAPIDS_X,	&model_spr),
810 	X86_MATCH_INTEL_FAM6_MODEL(RAPTORLAKE,		&model_skl),
811 	X86_MATCH_INTEL_FAM6_MODEL(RAPTORLAKE_P,	&model_skl),
812 	X86_MATCH_INTEL_FAM6_MODEL(RAPTORLAKE_S,	&model_skl),
813 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE,		&model_skl),
814 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L,	&model_skl),
815 	{},
816 };
817 MODULE_DEVICE_TABLE(x86cpu, rapl_model_match);
818 
rapl_pmu_init(void)819 static int __init rapl_pmu_init(void)
820 {
821 	const struct x86_cpu_id *id;
822 	struct rapl_model *rm;
823 	int ret;
824 
825 	id = x86_match_cpu(rapl_model_match);
826 	if (!id)
827 		return -ENODEV;
828 
829 	rm = (struct rapl_model *) id->driver_data;
830 
831 	rapl_msrs = rm->rapl_msrs;
832 
833 	rapl_cntr_mask = perf_msr_probe(rapl_msrs, PERF_RAPL_MAX,
834 					false, (void *) &rm->events);
835 
836 	ret = rapl_check_hw_unit(rm);
837 	if (ret)
838 		return ret;
839 
840 	ret = init_rapl_pmus();
841 	if (ret)
842 		return ret;
843 
844 	/*
845 	 * Install callbacks. Core will call them for each online cpu.
846 	 */
847 	ret = cpuhp_setup_state(CPUHP_AP_PERF_X86_RAPL_ONLINE,
848 				"perf/x86/rapl:online",
849 				rapl_cpu_online, rapl_cpu_offline);
850 	if (ret)
851 		goto out;
852 
853 	ret = perf_pmu_register(&rapl_pmus->pmu, "power", -1);
854 	if (ret)
855 		goto out1;
856 
857 	rapl_advertise();
858 	return 0;
859 
860 out1:
861 	cpuhp_remove_state(CPUHP_AP_PERF_X86_RAPL_ONLINE);
862 out:
863 	pr_warn("Initialization failed (%d), disabled\n", ret);
864 	cleanup_rapl_pmus();
865 	return ret;
866 }
867 module_init(rapl_pmu_init);
868 
intel_rapl_exit(void)869 static void __exit intel_rapl_exit(void)
870 {
871 	cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_RAPL_ONLINE);
872 	perf_pmu_unregister(&rapl_pmus->pmu);
873 	cleanup_rapl_pmus();
874 }
875 module_exit(intel_rapl_exit);
876