xref: /linux/arch/x86/events/intel/cstate.c (revision 7d7d1bf1d1dabe435ef50efb051724b8664749cb)
1 /*
2  * perf_event_intel_cstate.c: support cstate residency counters
3  *
4  * Copyright (C) 2015, Intel Corp.
5  * Author: Kan Liang (kan.liang@intel.com)
6  *
7  * This library is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Library General Public
9  * License as published by the Free Software Foundation; either
10  * version 2 of the License, or (at your option) any later version.
11  *
12  * This library is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Library General Public License for more details.
16  *
17  */
18 
19 /*
20  * This file export cstate related free running (read-only) counters
21  * for perf. These counters may be use simultaneously by other tools,
22  * such as turbostat. However, it still make sense to implement them
23  * in perf. Because we can conveniently collect them together with
24  * other events, and allow to use them from tools without special MSR
25  * access code.
26  *
27  * The events only support system-wide mode counting. There is no
28  * sampling support because it is not supported by the hardware.
29  *
30  * According to counters' scope and category, two PMUs are registered
31  * with the perf_event core subsystem.
32  *  - 'cstate_core': The counter is available for each physical core.
33  *    The counters include CORE_C*_RESIDENCY.
34  *  - 'cstate_pkg': The counter is available for each physical package.
35  *    The counters include PKG_C*_RESIDENCY.
36  *
37  * All of these counters are specified in the Intel® 64 and IA-32
38  * Architectures Software Developer.s Manual Vol3b.
39  *
40  * Model specific counters:
41  *	MSR_CORE_C1_RES: CORE C1 Residency Counter
42  *			 perf code: 0x00
43  *			 Available model: SLM,AMT
44  *			 Scope: Core (each processor core has a MSR)
45  *	MSR_CORE_C3_RESIDENCY: CORE C3 Residency Counter
46  *			       perf code: 0x01
47  *			       Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL
48  *			       Scope: Core
49  *	MSR_CORE_C6_RESIDENCY: CORE C6 Residency Counter
50  *			       perf code: 0x02
51  *			       Available model: SLM,AMT,NHM,WSM,SNB,IVB,HSW,BDW,SKL
52  *			       Scope: Core
53  *	MSR_CORE_C7_RESIDENCY: CORE C7 Residency Counter
54  *			       perf code: 0x03
55  *			       Available model: SNB,IVB,HSW,BDW,SKL
56  *			       Scope: Core
57  *	MSR_PKG_C2_RESIDENCY:  Package C2 Residency Counter.
58  *			       perf code: 0x00
59  *			       Available model: SNB,IVB,HSW,BDW,SKL
60  *			       Scope: Package (physical package)
61  *	MSR_PKG_C3_RESIDENCY:  Package C3 Residency Counter.
62  *			       perf code: 0x01
63  *			       Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL
64  *			       Scope: Package (physical package)
65  *	MSR_PKG_C6_RESIDENCY:  Package C6 Residency Counter.
66  *			       perf code: 0x02
67  *			       Available model: SLM,AMT,NHM,WSM,SNB,IVB,HSW,BDW,SKL
68  *			       Scope: Package (physical package)
69  *	MSR_PKG_C7_RESIDENCY:  Package C7 Residency Counter.
70  *			       perf code: 0x03
71  *			       Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL
72  *			       Scope: Package (physical package)
73  *	MSR_PKG_C8_RESIDENCY:  Package C8 Residency Counter.
74  *			       perf code: 0x04
75  *			       Available model: HSW ULT only
76  *			       Scope: Package (physical package)
77  *	MSR_PKG_C9_RESIDENCY:  Package C9 Residency Counter.
78  *			       perf code: 0x05
79  *			       Available model: HSW ULT only
80  *			       Scope: Package (physical package)
81  *	MSR_PKG_C10_RESIDENCY: Package C10 Residency Counter.
82  *			       perf code: 0x06
83  *			       Available model: HSW ULT only
84  *			       Scope: Package (physical package)
85  *
86  */
87 
88 #include <linux/module.h>
89 #include <linux/slab.h>
90 #include <linux/perf_event.h>
91 #include <asm/cpu_device_id.h>
92 #include <asm/intel-family.h>
93 #include "../perf_event.h"
94 
95 MODULE_LICENSE("GPL");
96 
97 #define DEFINE_CSTATE_FORMAT_ATTR(_var, _name, _format)		\
98 static ssize_t __cstate_##_var##_show(struct kobject *kobj,	\
99 				struct kobj_attribute *attr,	\
100 				char *page)			\
101 {								\
102 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);		\
103 	return sprintf(page, _format "\n");			\
104 }								\
105 static struct kobj_attribute format_attr_##_var =		\
106 	__ATTR(_name, 0444, __cstate_##_var##_show, NULL)
107 
108 static ssize_t cstate_get_attr_cpumask(struct device *dev,
109 				       struct device_attribute *attr,
110 				       char *buf);
111 
112 /* Model -> events mapping */
113 struct cstate_model {
114 	unsigned long		core_events;
115 	unsigned long		pkg_events;
116 	unsigned long		quirks;
117 };
118 
119 /* Quirk flags */
120 #define SLM_PKG_C6_USE_C7_MSR	(1UL << 0)
121 
122 struct perf_cstate_msr {
123 	u64	msr;
124 	struct	perf_pmu_events_attr *attr;
125 };
126 
127 
128 /* cstate_core PMU */
129 static struct pmu cstate_core_pmu;
130 static bool has_cstate_core;
131 
132 enum perf_cstate_core_events {
133 	PERF_CSTATE_CORE_C1_RES = 0,
134 	PERF_CSTATE_CORE_C3_RES,
135 	PERF_CSTATE_CORE_C6_RES,
136 	PERF_CSTATE_CORE_C7_RES,
137 
138 	PERF_CSTATE_CORE_EVENT_MAX,
139 };
140 
141 PMU_EVENT_ATTR_STRING(c1-residency, evattr_cstate_core_c1, "event=0x00");
142 PMU_EVENT_ATTR_STRING(c3-residency, evattr_cstate_core_c3, "event=0x01");
143 PMU_EVENT_ATTR_STRING(c6-residency, evattr_cstate_core_c6, "event=0x02");
144 PMU_EVENT_ATTR_STRING(c7-residency, evattr_cstate_core_c7, "event=0x03");
145 
146 static struct perf_cstate_msr core_msr[] = {
147 	[PERF_CSTATE_CORE_C1_RES] = { MSR_CORE_C1_RES,		&evattr_cstate_core_c1 },
148 	[PERF_CSTATE_CORE_C3_RES] = { MSR_CORE_C3_RESIDENCY,	&evattr_cstate_core_c3 },
149 	[PERF_CSTATE_CORE_C6_RES] = { MSR_CORE_C6_RESIDENCY,	&evattr_cstate_core_c6 },
150 	[PERF_CSTATE_CORE_C7_RES] = { MSR_CORE_C7_RESIDENCY,	&evattr_cstate_core_c7 },
151 };
152 
153 static struct attribute *core_events_attrs[PERF_CSTATE_CORE_EVENT_MAX + 1] = {
154 	NULL,
155 };
156 
157 static struct attribute_group core_events_attr_group = {
158 	.name = "events",
159 	.attrs = core_events_attrs,
160 };
161 
162 DEFINE_CSTATE_FORMAT_ATTR(core_event, event, "config:0-63");
163 static struct attribute *core_format_attrs[] = {
164 	&format_attr_core_event.attr,
165 	NULL,
166 };
167 
168 static struct attribute_group core_format_attr_group = {
169 	.name = "format",
170 	.attrs = core_format_attrs,
171 };
172 
173 static cpumask_t cstate_core_cpu_mask;
174 static DEVICE_ATTR(cpumask, S_IRUGO, cstate_get_attr_cpumask, NULL);
175 
176 static struct attribute *cstate_cpumask_attrs[] = {
177 	&dev_attr_cpumask.attr,
178 	NULL,
179 };
180 
181 static struct attribute_group cpumask_attr_group = {
182 	.attrs = cstate_cpumask_attrs,
183 };
184 
185 static const struct attribute_group *core_attr_groups[] = {
186 	&core_events_attr_group,
187 	&core_format_attr_group,
188 	&cpumask_attr_group,
189 	NULL,
190 };
191 
192 /* cstate_pkg PMU */
193 static struct pmu cstate_pkg_pmu;
194 static bool has_cstate_pkg;
195 
196 enum perf_cstate_pkg_events {
197 	PERF_CSTATE_PKG_C2_RES = 0,
198 	PERF_CSTATE_PKG_C3_RES,
199 	PERF_CSTATE_PKG_C6_RES,
200 	PERF_CSTATE_PKG_C7_RES,
201 	PERF_CSTATE_PKG_C8_RES,
202 	PERF_CSTATE_PKG_C9_RES,
203 	PERF_CSTATE_PKG_C10_RES,
204 
205 	PERF_CSTATE_PKG_EVENT_MAX,
206 };
207 
208 PMU_EVENT_ATTR_STRING(c2-residency, evattr_cstate_pkg_c2, "event=0x00");
209 PMU_EVENT_ATTR_STRING(c3-residency, evattr_cstate_pkg_c3, "event=0x01");
210 PMU_EVENT_ATTR_STRING(c6-residency, evattr_cstate_pkg_c6, "event=0x02");
211 PMU_EVENT_ATTR_STRING(c7-residency, evattr_cstate_pkg_c7, "event=0x03");
212 PMU_EVENT_ATTR_STRING(c8-residency, evattr_cstate_pkg_c8, "event=0x04");
213 PMU_EVENT_ATTR_STRING(c9-residency, evattr_cstate_pkg_c9, "event=0x05");
214 PMU_EVENT_ATTR_STRING(c10-residency, evattr_cstate_pkg_c10, "event=0x06");
215 
216 static struct perf_cstate_msr pkg_msr[] = {
217 	[PERF_CSTATE_PKG_C2_RES] = { MSR_PKG_C2_RESIDENCY,	&evattr_cstate_pkg_c2 },
218 	[PERF_CSTATE_PKG_C3_RES] = { MSR_PKG_C3_RESIDENCY,	&evattr_cstate_pkg_c3 },
219 	[PERF_CSTATE_PKG_C6_RES] = { MSR_PKG_C6_RESIDENCY,	&evattr_cstate_pkg_c6 },
220 	[PERF_CSTATE_PKG_C7_RES] = { MSR_PKG_C7_RESIDENCY,	&evattr_cstate_pkg_c7 },
221 	[PERF_CSTATE_PKG_C8_RES] = { MSR_PKG_C8_RESIDENCY,	&evattr_cstate_pkg_c8 },
222 	[PERF_CSTATE_PKG_C9_RES] = { MSR_PKG_C9_RESIDENCY,	&evattr_cstate_pkg_c9 },
223 	[PERF_CSTATE_PKG_C10_RES] = { MSR_PKG_C10_RESIDENCY,	&evattr_cstate_pkg_c10 },
224 };
225 
226 static struct attribute *pkg_events_attrs[PERF_CSTATE_PKG_EVENT_MAX + 1] = {
227 	NULL,
228 };
229 
230 static struct attribute_group pkg_events_attr_group = {
231 	.name = "events",
232 	.attrs = pkg_events_attrs,
233 };
234 
235 DEFINE_CSTATE_FORMAT_ATTR(pkg_event, event, "config:0-63");
236 static struct attribute *pkg_format_attrs[] = {
237 	&format_attr_pkg_event.attr,
238 	NULL,
239 };
240 static struct attribute_group pkg_format_attr_group = {
241 	.name = "format",
242 	.attrs = pkg_format_attrs,
243 };
244 
245 static cpumask_t cstate_pkg_cpu_mask;
246 
247 static const struct attribute_group *pkg_attr_groups[] = {
248 	&pkg_events_attr_group,
249 	&pkg_format_attr_group,
250 	&cpumask_attr_group,
251 	NULL,
252 };
253 
254 static ssize_t cstate_get_attr_cpumask(struct device *dev,
255 				       struct device_attribute *attr,
256 				       char *buf)
257 {
258 	struct pmu *pmu = dev_get_drvdata(dev);
259 
260 	if (pmu == &cstate_core_pmu)
261 		return cpumap_print_to_pagebuf(true, buf, &cstate_core_cpu_mask);
262 	else if (pmu == &cstate_pkg_pmu)
263 		return cpumap_print_to_pagebuf(true, buf, &cstate_pkg_cpu_mask);
264 	else
265 		return 0;
266 }
267 
268 static int cstate_pmu_event_init(struct perf_event *event)
269 {
270 	u64 cfg = event->attr.config;
271 	int cpu;
272 
273 	if (event->attr.type != event->pmu->type)
274 		return -ENOENT;
275 
276 	/* unsupported modes and filters */
277 	if (event->attr.exclude_user   ||
278 	    event->attr.exclude_kernel ||
279 	    event->attr.exclude_hv     ||
280 	    event->attr.exclude_idle   ||
281 	    event->attr.exclude_host   ||
282 	    event->attr.exclude_guest  ||
283 	    event->attr.sample_period) /* no sampling */
284 		return -EINVAL;
285 
286 	if (event->cpu < 0)
287 		return -EINVAL;
288 
289 	if (event->pmu == &cstate_core_pmu) {
290 		if (cfg >= PERF_CSTATE_CORE_EVENT_MAX)
291 			return -EINVAL;
292 		if (!core_msr[cfg].attr)
293 			return -EINVAL;
294 		event->hw.event_base = core_msr[cfg].msr;
295 		cpu = cpumask_any_and(&cstate_core_cpu_mask,
296 				      topology_sibling_cpumask(event->cpu));
297 	} else if (event->pmu == &cstate_pkg_pmu) {
298 		if (cfg >= PERF_CSTATE_PKG_EVENT_MAX)
299 			return -EINVAL;
300 		if (!pkg_msr[cfg].attr)
301 			return -EINVAL;
302 		event->hw.event_base = pkg_msr[cfg].msr;
303 		cpu = cpumask_any_and(&cstate_pkg_cpu_mask,
304 				      topology_core_cpumask(event->cpu));
305 	} else {
306 		return -ENOENT;
307 	}
308 
309 	if (cpu >= nr_cpu_ids)
310 		return -ENODEV;
311 
312 	event->cpu = cpu;
313 	event->hw.config = cfg;
314 	event->hw.idx = -1;
315 	return 0;
316 }
317 
318 static inline u64 cstate_pmu_read_counter(struct perf_event *event)
319 {
320 	u64 val;
321 
322 	rdmsrl(event->hw.event_base, val);
323 	return val;
324 }
325 
326 static void cstate_pmu_event_update(struct perf_event *event)
327 {
328 	struct hw_perf_event *hwc = &event->hw;
329 	u64 prev_raw_count, new_raw_count;
330 
331 again:
332 	prev_raw_count = local64_read(&hwc->prev_count);
333 	new_raw_count = cstate_pmu_read_counter(event);
334 
335 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
336 			    new_raw_count) != prev_raw_count)
337 		goto again;
338 
339 	local64_add(new_raw_count - prev_raw_count, &event->count);
340 }
341 
342 static void cstate_pmu_event_start(struct perf_event *event, int mode)
343 {
344 	local64_set(&event->hw.prev_count, cstate_pmu_read_counter(event));
345 }
346 
347 static void cstate_pmu_event_stop(struct perf_event *event, int mode)
348 {
349 	cstate_pmu_event_update(event);
350 }
351 
352 static void cstate_pmu_event_del(struct perf_event *event, int mode)
353 {
354 	cstate_pmu_event_stop(event, PERF_EF_UPDATE);
355 }
356 
357 static int cstate_pmu_event_add(struct perf_event *event, int mode)
358 {
359 	if (mode & PERF_EF_START)
360 		cstate_pmu_event_start(event, mode);
361 
362 	return 0;
363 }
364 
365 /*
366  * Check if exiting cpu is the designated reader. If so migrate the
367  * events when there is a valid target available
368  */
369 static void cstate_cpu_exit(int cpu)
370 {
371 	unsigned int target;
372 
373 	if (has_cstate_core &&
374 	    cpumask_test_and_clear_cpu(cpu, &cstate_core_cpu_mask)) {
375 
376 		target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
377 		/* Migrate events if there is a valid target */
378 		if (target < nr_cpu_ids) {
379 			cpumask_set_cpu(target, &cstate_core_cpu_mask);
380 			perf_pmu_migrate_context(&cstate_core_pmu, cpu, target);
381 		}
382 	}
383 
384 	if (has_cstate_pkg &&
385 	    cpumask_test_and_clear_cpu(cpu, &cstate_pkg_cpu_mask)) {
386 
387 		target = cpumask_any_but(topology_core_cpumask(cpu), cpu);
388 		/* Migrate events if there is a valid target */
389 		if (target < nr_cpu_ids) {
390 			cpumask_set_cpu(target, &cstate_pkg_cpu_mask);
391 			perf_pmu_migrate_context(&cstate_pkg_pmu, cpu, target);
392 		}
393 	}
394 }
395 
396 static void cstate_cpu_init(int cpu)
397 {
398 	unsigned int target;
399 
400 	/*
401 	 * If this is the first online thread of that core, set it in
402 	 * the core cpu mask as the designated reader.
403 	 */
404 	target = cpumask_any_and(&cstate_core_cpu_mask,
405 				 topology_sibling_cpumask(cpu));
406 
407 	if (has_cstate_core && target >= nr_cpu_ids)
408 		cpumask_set_cpu(cpu, &cstate_core_cpu_mask);
409 
410 	/*
411 	 * If this is the first online thread of that package, set it
412 	 * in the package cpu mask as the designated reader.
413 	 */
414 	target = cpumask_any_and(&cstate_pkg_cpu_mask,
415 				 topology_core_cpumask(cpu));
416 	if (has_cstate_pkg && target >= nr_cpu_ids)
417 		cpumask_set_cpu(cpu, &cstate_pkg_cpu_mask);
418 }
419 
420 static int cstate_cpu_notifier(struct notifier_block *self,
421 			       unsigned long action, void *hcpu)
422 {
423 	unsigned int cpu = (long)hcpu;
424 
425 	switch (action & ~CPU_TASKS_FROZEN) {
426 	case CPU_STARTING:
427 		cstate_cpu_init(cpu);
428 		break;
429 	case CPU_DOWN_PREPARE:
430 		cstate_cpu_exit(cpu);
431 		break;
432 	default:
433 		break;
434 	}
435 	return NOTIFY_OK;
436 }
437 
438 static struct notifier_block cstate_cpu_nb = {
439 	.notifier_call	= cstate_cpu_notifier,
440 	.priority       = CPU_PRI_PERF + 1,
441 };
442 
443 static struct pmu cstate_core_pmu = {
444 	.attr_groups	= core_attr_groups,
445 	.name		= "cstate_core",
446 	.task_ctx_nr	= perf_invalid_context,
447 	.event_init	= cstate_pmu_event_init,
448 	.add		= cstate_pmu_event_add,
449 	.del		= cstate_pmu_event_del,
450 	.start		= cstate_pmu_event_start,
451 	.stop		= cstate_pmu_event_stop,
452 	.read		= cstate_pmu_event_update,
453 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT,
454 };
455 
456 static struct pmu cstate_pkg_pmu = {
457 	.attr_groups	= pkg_attr_groups,
458 	.name		= "cstate_pkg",
459 	.task_ctx_nr	= perf_invalid_context,
460 	.event_init	= cstate_pmu_event_init,
461 	.add		= cstate_pmu_event_add,
462 	.del		= cstate_pmu_event_del,
463 	.start		= cstate_pmu_event_start,
464 	.stop		= cstate_pmu_event_stop,
465 	.read		= cstate_pmu_event_update,
466 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT,
467 };
468 
469 static const struct cstate_model nhm_cstates __initconst = {
470 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
471 				  BIT(PERF_CSTATE_CORE_C6_RES),
472 
473 	.pkg_events		= BIT(PERF_CSTATE_PKG_C3_RES) |
474 				  BIT(PERF_CSTATE_PKG_C6_RES) |
475 				  BIT(PERF_CSTATE_PKG_C7_RES),
476 };
477 
478 static const struct cstate_model snb_cstates __initconst = {
479 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
480 				  BIT(PERF_CSTATE_CORE_C6_RES) |
481 				  BIT(PERF_CSTATE_CORE_C7_RES),
482 
483 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
484 				  BIT(PERF_CSTATE_PKG_C3_RES) |
485 				  BIT(PERF_CSTATE_PKG_C6_RES) |
486 				  BIT(PERF_CSTATE_PKG_C7_RES),
487 };
488 
489 static const struct cstate_model hswult_cstates __initconst = {
490 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
491 				  BIT(PERF_CSTATE_CORE_C6_RES) |
492 				  BIT(PERF_CSTATE_CORE_C7_RES),
493 
494 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
495 				  BIT(PERF_CSTATE_PKG_C3_RES) |
496 				  BIT(PERF_CSTATE_PKG_C6_RES) |
497 				  BIT(PERF_CSTATE_PKG_C7_RES) |
498 				  BIT(PERF_CSTATE_PKG_C8_RES) |
499 				  BIT(PERF_CSTATE_PKG_C9_RES) |
500 				  BIT(PERF_CSTATE_PKG_C10_RES),
501 };
502 
503 static const struct cstate_model slm_cstates __initconst = {
504 	.core_events		= BIT(PERF_CSTATE_CORE_C1_RES) |
505 				  BIT(PERF_CSTATE_CORE_C6_RES),
506 
507 	.pkg_events		= BIT(PERF_CSTATE_PKG_C6_RES),
508 	.quirks			= SLM_PKG_C6_USE_C7_MSR,
509 };
510 
511 #define X86_CSTATES_MODEL(model, states)				\
512 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY, (unsigned long) &(states) }
513 
514 static const struct x86_cpu_id intel_cstates_match[] __initconst = {
515 	X86_CSTATES_MODEL(INTEL_FAM6_NEHALEM,    nhm_cstates),
516 	X86_CSTATES_MODEL(INTEL_FAM6_NEHALEM_EP, nhm_cstates),
517 	X86_CSTATES_MODEL(INTEL_FAM6_NEHALEM_EX, nhm_cstates),
518 
519 	X86_CSTATES_MODEL(INTEL_FAM6_WESTMERE,    nhm_cstates),
520 	X86_CSTATES_MODEL(INTEL_FAM6_WESTMERE_EP, nhm_cstates),
521 	X86_CSTATES_MODEL(INTEL_FAM6_WESTMERE_EX, nhm_cstates),
522 
523 	X86_CSTATES_MODEL(INTEL_FAM6_SANDYBRIDGE,   snb_cstates),
524 	X86_CSTATES_MODEL(INTEL_FAM6_SANDYBRIDGE_X, snb_cstates),
525 
526 	X86_CSTATES_MODEL(INTEL_FAM6_IVYBRIDGE,   snb_cstates),
527 	X86_CSTATES_MODEL(INTEL_FAM6_IVYBRIDGE_X, snb_cstates),
528 
529 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL_CORE, snb_cstates),
530 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL_X,	   snb_cstates),
531 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL_GT3E, snb_cstates),
532 
533 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL_ULT, hswult_cstates),
534 
535 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_SILVERMONT1, slm_cstates),
536 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_SILVERMONT2, slm_cstates),
537 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_AIRMONT,     slm_cstates),
538 
539 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL_CORE,   snb_cstates),
540 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL_XEON_D, snb_cstates),
541 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL_GT3E,   snb_cstates),
542 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL_X,      snb_cstates),
543 
544 	X86_CSTATES_MODEL(INTEL_FAM6_SKYLAKE_MOBILE,  snb_cstates),
545 	X86_CSTATES_MODEL(INTEL_FAM6_SKYLAKE_DESKTOP, snb_cstates),
546 	{ },
547 };
548 MODULE_DEVICE_TABLE(x86cpu, intel_cstates_match);
549 
550 /*
551  * Probe the cstate events and insert the available one into sysfs attrs
552  * Return false if there are no available events.
553  */
554 static bool __init cstate_probe_msr(const unsigned long evmsk, int max,
555                                    struct perf_cstate_msr *msr,
556                                    struct attribute **attrs)
557 {
558 	bool found = false;
559 	unsigned int bit;
560 	u64 val;
561 
562 	for (bit = 0; bit < max; bit++) {
563 		if (test_bit(bit, &evmsk) && !rdmsrl_safe(msr[bit].msr, &val)) {
564 			*attrs++ = &msr[bit].attr->attr.attr;
565 			found = true;
566 		} else {
567 			msr[bit].attr = NULL;
568 		}
569 	}
570 	*attrs = NULL;
571 
572 	return found;
573 }
574 
575 static int __init cstate_probe(const struct cstate_model *cm)
576 {
577 	/* SLM has different MSR for PKG C6 */
578 	if (cm->quirks & SLM_PKG_C6_USE_C7_MSR)
579 		pkg_msr[PERF_CSTATE_PKG_C6_RES].msr = MSR_PKG_C7_RESIDENCY;
580 
581 	has_cstate_core = cstate_probe_msr(cm->core_events,
582 					   PERF_CSTATE_CORE_EVENT_MAX,
583 					   core_msr, core_events_attrs);
584 
585 	has_cstate_pkg = cstate_probe_msr(cm->pkg_events,
586 					  PERF_CSTATE_PKG_EVENT_MAX,
587 					  pkg_msr, pkg_events_attrs);
588 
589 	return (has_cstate_core || has_cstate_pkg) ? 0 : -ENODEV;
590 }
591 
592 static inline void cstate_cleanup(void)
593 {
594 	if (has_cstate_core)
595 		perf_pmu_unregister(&cstate_core_pmu);
596 
597 	if (has_cstate_pkg)
598 		perf_pmu_unregister(&cstate_pkg_pmu);
599 }
600 
601 static int __init cstate_init(void)
602 {
603 	int cpu, err;
604 
605 	cpu_notifier_register_begin();
606 	for_each_online_cpu(cpu)
607 		cstate_cpu_init(cpu);
608 
609 	if (has_cstate_core) {
610 		err = perf_pmu_register(&cstate_core_pmu, cstate_core_pmu.name, -1);
611 		if (err) {
612 			has_cstate_core = false;
613 			pr_info("Failed to register cstate core pmu\n");
614 			goto out;
615 		}
616 	}
617 
618 	if (has_cstate_pkg) {
619 		err = perf_pmu_register(&cstate_pkg_pmu, cstate_pkg_pmu.name, -1);
620 		if (err) {
621 			has_cstate_pkg = false;
622 			pr_info("Failed to register cstate pkg pmu\n");
623 			cstate_cleanup();
624 			goto out;
625 		}
626 	}
627 	__register_cpu_notifier(&cstate_cpu_nb);
628 out:
629 	cpu_notifier_register_done();
630 	return err;
631 }
632 
633 static int __init cstate_pmu_init(void)
634 {
635 	const struct x86_cpu_id *id;
636 	int err;
637 
638 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
639 		return -ENODEV;
640 
641 	id = x86_match_cpu(intel_cstates_match);
642 	if (!id)
643 		return -ENODEV;
644 
645 	err = cstate_probe((const struct cstate_model *) id->driver_data);
646 	if (err)
647 		return err;
648 
649 	return cstate_init();
650 }
651 module_init(cstate_pmu_init);
652 
653 static void __exit cstate_pmu_exit(void)
654 {
655 	cpu_notifier_register_begin();
656 	__unregister_cpu_notifier(&cstate_cpu_nb);
657 	cstate_cleanup();
658 	cpu_notifier_register_done();
659 }
660 module_exit(cstate_pmu_exit);
661