xref: /linux/arch/x86/events/intel/cstate.c (revision d195c39052d1da278a00a6744ce59c383b67b191)
1 /*
2  * 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,GLM,CNL,TNT
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,GLM,
48  *						CNL,KBL,CML,TNT
49  *			       Scope: Core
50  *	MSR_CORE_C6_RESIDENCY: CORE C6 Residency Counter
51  *			       perf code: 0x02
52  *			       Available model: SLM,AMT,NHM,WSM,SNB,IVB,HSW,BDW,
53  *						SKL,KNL,GLM,CNL,KBL,CML,ICL,TGL,
54  *						TNT
55  *			       Scope: Core
56  *	MSR_CORE_C7_RESIDENCY: CORE C7 Residency Counter
57  *			       perf code: 0x03
58  *			       Available model: SNB,IVB,HSW,BDW,SKL,CNL,KBL,CML,
59  *						ICL,TGL
60  *			       Scope: Core
61  *	MSR_PKG_C2_RESIDENCY:  Package C2 Residency Counter.
62  *			       perf code: 0x00
63  *			       Available model: SNB,IVB,HSW,BDW,SKL,KNL,GLM,CNL,
64  *						KBL,CML,ICL,TGL,TNT
65  *			       Scope: Package (physical package)
66  *	MSR_PKG_C3_RESIDENCY:  Package C3 Residency Counter.
67  *			       perf code: 0x01
68  *			       Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL,KNL,
69  *						GLM,CNL,KBL,CML,ICL,TGL,TNT
70  *			       Scope: Package (physical package)
71  *	MSR_PKG_C6_RESIDENCY:  Package C6 Residency Counter.
72  *			       perf code: 0x02
73  *			       Available model: SLM,AMT,NHM,WSM,SNB,IVB,HSW,BDW,
74  *						SKL,KNL,GLM,CNL,KBL,CML,ICL,TGL,
75  *						TNT
76  *			       Scope: Package (physical package)
77  *	MSR_PKG_C7_RESIDENCY:  Package C7 Residency Counter.
78  *			       perf code: 0x03
79  *			       Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL,CNL,
80  *						KBL,CML,ICL,TGL
81  *			       Scope: Package (physical package)
82  *	MSR_PKG_C8_RESIDENCY:  Package C8 Residency Counter.
83  *			       perf code: 0x04
84  *			       Available model: HSW ULT,KBL,CNL,CML,ICL,TGL
85  *			       Scope: Package (physical package)
86  *	MSR_PKG_C9_RESIDENCY:  Package C9 Residency Counter.
87  *			       perf code: 0x05
88  *			       Available model: HSW ULT,KBL,CNL,CML,ICL,TGL
89  *			       Scope: Package (physical package)
90  *	MSR_PKG_C10_RESIDENCY: Package C10 Residency Counter.
91  *			       perf code: 0x06
92  *			       Available model: HSW ULT,KBL,GLM,CNL,CML,ICL,TGL,
93  *						TNT
94  *			       Scope: Package (physical package)
95  *
96  */
97 
98 #include <linux/module.h>
99 #include <linux/slab.h>
100 #include <linux/perf_event.h>
101 #include <linux/nospec.h>
102 #include <asm/cpu_device_id.h>
103 #include <asm/intel-family.h>
104 #include "../perf_event.h"
105 #include "../probe.h"
106 
107 MODULE_LICENSE("GPL");
108 
109 #define DEFINE_CSTATE_FORMAT_ATTR(_var, _name, _format)		\
110 static ssize_t __cstate_##_var##_show(struct kobject *kobj,	\
111 				struct kobj_attribute *attr,	\
112 				char *page)			\
113 {								\
114 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);		\
115 	return sprintf(page, _format "\n");			\
116 }								\
117 static struct kobj_attribute format_attr_##_var =		\
118 	__ATTR(_name, 0444, __cstate_##_var##_show, NULL)
119 
120 static ssize_t cstate_get_attr_cpumask(struct device *dev,
121 				       struct device_attribute *attr,
122 				       char *buf);
123 
124 /* Model -> events mapping */
125 struct cstate_model {
126 	unsigned long		core_events;
127 	unsigned long		pkg_events;
128 	unsigned long		quirks;
129 };
130 
131 /* Quirk flags */
132 #define SLM_PKG_C6_USE_C7_MSR	(1UL << 0)
133 #define KNL_CORE_C6_MSR		(1UL << 1)
134 
135 struct perf_cstate_msr {
136 	u64	msr;
137 	struct	perf_pmu_events_attr *attr;
138 };
139 
140 
141 /* cstate_core PMU */
142 static struct pmu cstate_core_pmu;
143 static bool has_cstate_core;
144 
145 enum perf_cstate_core_events {
146 	PERF_CSTATE_CORE_C1_RES = 0,
147 	PERF_CSTATE_CORE_C3_RES,
148 	PERF_CSTATE_CORE_C6_RES,
149 	PERF_CSTATE_CORE_C7_RES,
150 
151 	PERF_CSTATE_CORE_EVENT_MAX,
152 };
153 
154 PMU_EVENT_ATTR_STRING(c1-residency, attr_cstate_core_c1, "event=0x00");
155 PMU_EVENT_ATTR_STRING(c3-residency, attr_cstate_core_c3, "event=0x01");
156 PMU_EVENT_ATTR_STRING(c6-residency, attr_cstate_core_c6, "event=0x02");
157 PMU_EVENT_ATTR_STRING(c7-residency, attr_cstate_core_c7, "event=0x03");
158 
159 static unsigned long core_msr_mask;
160 
161 PMU_EVENT_GROUP(events, cstate_core_c1);
162 PMU_EVENT_GROUP(events, cstate_core_c3);
163 PMU_EVENT_GROUP(events, cstate_core_c6);
164 PMU_EVENT_GROUP(events, cstate_core_c7);
165 
166 static bool test_msr(int idx, void *data)
167 {
168 	return test_bit(idx, (unsigned long *) data);
169 }
170 
171 static struct perf_msr core_msr[] = {
172 	[PERF_CSTATE_CORE_C1_RES] = { MSR_CORE_C1_RES,		&group_cstate_core_c1,	test_msr },
173 	[PERF_CSTATE_CORE_C3_RES] = { MSR_CORE_C3_RESIDENCY,	&group_cstate_core_c3,	test_msr },
174 	[PERF_CSTATE_CORE_C6_RES] = { MSR_CORE_C6_RESIDENCY,	&group_cstate_core_c6,	test_msr },
175 	[PERF_CSTATE_CORE_C7_RES] = { MSR_CORE_C7_RESIDENCY,	&group_cstate_core_c7,	test_msr },
176 };
177 
178 static struct attribute *attrs_empty[] = {
179 	NULL,
180 };
181 
182 /*
183  * There are no default events, but we need to create
184  * "events" group (with empty attrs) before updating
185  * it with detected events.
186  */
187 static struct attribute_group core_events_attr_group = {
188 	.name = "events",
189 	.attrs = attrs_empty,
190 };
191 
192 DEFINE_CSTATE_FORMAT_ATTR(core_event, event, "config:0-63");
193 static struct attribute *core_format_attrs[] = {
194 	&format_attr_core_event.attr,
195 	NULL,
196 };
197 
198 static struct attribute_group core_format_attr_group = {
199 	.name = "format",
200 	.attrs = core_format_attrs,
201 };
202 
203 static cpumask_t cstate_core_cpu_mask;
204 static DEVICE_ATTR(cpumask, S_IRUGO, cstate_get_attr_cpumask, NULL);
205 
206 static struct attribute *cstate_cpumask_attrs[] = {
207 	&dev_attr_cpumask.attr,
208 	NULL,
209 };
210 
211 static struct attribute_group cpumask_attr_group = {
212 	.attrs = cstate_cpumask_attrs,
213 };
214 
215 static const struct attribute_group *core_attr_groups[] = {
216 	&core_events_attr_group,
217 	&core_format_attr_group,
218 	&cpumask_attr_group,
219 	NULL,
220 };
221 
222 /* cstate_pkg PMU */
223 static struct pmu cstate_pkg_pmu;
224 static bool has_cstate_pkg;
225 
226 enum perf_cstate_pkg_events {
227 	PERF_CSTATE_PKG_C2_RES = 0,
228 	PERF_CSTATE_PKG_C3_RES,
229 	PERF_CSTATE_PKG_C6_RES,
230 	PERF_CSTATE_PKG_C7_RES,
231 	PERF_CSTATE_PKG_C8_RES,
232 	PERF_CSTATE_PKG_C9_RES,
233 	PERF_CSTATE_PKG_C10_RES,
234 
235 	PERF_CSTATE_PKG_EVENT_MAX,
236 };
237 
238 PMU_EVENT_ATTR_STRING(c2-residency,  attr_cstate_pkg_c2,  "event=0x00");
239 PMU_EVENT_ATTR_STRING(c3-residency,  attr_cstate_pkg_c3,  "event=0x01");
240 PMU_EVENT_ATTR_STRING(c6-residency,  attr_cstate_pkg_c6,  "event=0x02");
241 PMU_EVENT_ATTR_STRING(c7-residency,  attr_cstate_pkg_c7,  "event=0x03");
242 PMU_EVENT_ATTR_STRING(c8-residency,  attr_cstate_pkg_c8,  "event=0x04");
243 PMU_EVENT_ATTR_STRING(c9-residency,  attr_cstate_pkg_c9,  "event=0x05");
244 PMU_EVENT_ATTR_STRING(c10-residency, attr_cstate_pkg_c10, "event=0x06");
245 
246 static unsigned long pkg_msr_mask;
247 
248 PMU_EVENT_GROUP(events, cstate_pkg_c2);
249 PMU_EVENT_GROUP(events, cstate_pkg_c3);
250 PMU_EVENT_GROUP(events, cstate_pkg_c6);
251 PMU_EVENT_GROUP(events, cstate_pkg_c7);
252 PMU_EVENT_GROUP(events, cstate_pkg_c8);
253 PMU_EVENT_GROUP(events, cstate_pkg_c9);
254 PMU_EVENT_GROUP(events, cstate_pkg_c10);
255 
256 static struct perf_msr pkg_msr[] = {
257 	[PERF_CSTATE_PKG_C2_RES]  = { MSR_PKG_C2_RESIDENCY,	&group_cstate_pkg_c2,	test_msr },
258 	[PERF_CSTATE_PKG_C3_RES]  = { MSR_PKG_C3_RESIDENCY,	&group_cstate_pkg_c3,	test_msr },
259 	[PERF_CSTATE_PKG_C6_RES]  = { MSR_PKG_C6_RESIDENCY,	&group_cstate_pkg_c6,	test_msr },
260 	[PERF_CSTATE_PKG_C7_RES]  = { MSR_PKG_C7_RESIDENCY,	&group_cstate_pkg_c7,	test_msr },
261 	[PERF_CSTATE_PKG_C8_RES]  = { MSR_PKG_C8_RESIDENCY,	&group_cstate_pkg_c8,	test_msr },
262 	[PERF_CSTATE_PKG_C9_RES]  = { MSR_PKG_C9_RESIDENCY,	&group_cstate_pkg_c9,	test_msr },
263 	[PERF_CSTATE_PKG_C10_RES] = { MSR_PKG_C10_RESIDENCY,	&group_cstate_pkg_c10,	test_msr },
264 };
265 
266 static struct attribute_group pkg_events_attr_group = {
267 	.name = "events",
268 	.attrs = attrs_empty,
269 };
270 
271 DEFINE_CSTATE_FORMAT_ATTR(pkg_event, event, "config:0-63");
272 static struct attribute *pkg_format_attrs[] = {
273 	&format_attr_pkg_event.attr,
274 	NULL,
275 };
276 static struct attribute_group pkg_format_attr_group = {
277 	.name = "format",
278 	.attrs = pkg_format_attrs,
279 };
280 
281 static cpumask_t cstate_pkg_cpu_mask;
282 
283 static const struct attribute_group *pkg_attr_groups[] = {
284 	&pkg_events_attr_group,
285 	&pkg_format_attr_group,
286 	&cpumask_attr_group,
287 	NULL,
288 };
289 
290 static ssize_t cstate_get_attr_cpumask(struct device *dev,
291 				       struct device_attribute *attr,
292 				       char *buf)
293 {
294 	struct pmu *pmu = dev_get_drvdata(dev);
295 
296 	if (pmu == &cstate_core_pmu)
297 		return cpumap_print_to_pagebuf(true, buf, &cstate_core_cpu_mask);
298 	else if (pmu == &cstate_pkg_pmu)
299 		return cpumap_print_to_pagebuf(true, buf, &cstate_pkg_cpu_mask);
300 	else
301 		return 0;
302 }
303 
304 static int cstate_pmu_event_init(struct perf_event *event)
305 {
306 	u64 cfg = event->attr.config;
307 	int cpu;
308 
309 	if (event->attr.type != event->pmu->type)
310 		return -ENOENT;
311 
312 	/* unsupported modes and filters */
313 	if (event->attr.sample_period) /* no sampling */
314 		return -EINVAL;
315 
316 	if (event->cpu < 0)
317 		return -EINVAL;
318 
319 	if (event->pmu == &cstate_core_pmu) {
320 		if (cfg >= PERF_CSTATE_CORE_EVENT_MAX)
321 			return -EINVAL;
322 		cfg = array_index_nospec((unsigned long)cfg, PERF_CSTATE_CORE_EVENT_MAX);
323 		if (!(core_msr_mask & (1 << cfg)))
324 			return -EINVAL;
325 		event->hw.event_base = core_msr[cfg].msr;
326 		cpu = cpumask_any_and(&cstate_core_cpu_mask,
327 				      topology_sibling_cpumask(event->cpu));
328 	} else if (event->pmu == &cstate_pkg_pmu) {
329 		if (cfg >= PERF_CSTATE_PKG_EVENT_MAX)
330 			return -EINVAL;
331 		cfg = array_index_nospec((unsigned long)cfg, PERF_CSTATE_PKG_EVENT_MAX);
332 		if (!(pkg_msr_mask & (1 << cfg)))
333 			return -EINVAL;
334 		event->hw.event_base = pkg_msr[cfg].msr;
335 		cpu = cpumask_any_and(&cstate_pkg_cpu_mask,
336 				      topology_die_cpumask(event->cpu));
337 	} else {
338 		return -ENOENT;
339 	}
340 
341 	if (cpu >= nr_cpu_ids)
342 		return -ENODEV;
343 
344 	event->cpu = cpu;
345 	event->hw.config = cfg;
346 	event->hw.idx = -1;
347 	return 0;
348 }
349 
350 static inline u64 cstate_pmu_read_counter(struct perf_event *event)
351 {
352 	u64 val;
353 
354 	rdmsrl(event->hw.event_base, val);
355 	return val;
356 }
357 
358 static void cstate_pmu_event_update(struct perf_event *event)
359 {
360 	struct hw_perf_event *hwc = &event->hw;
361 	u64 prev_raw_count, new_raw_count;
362 
363 again:
364 	prev_raw_count = local64_read(&hwc->prev_count);
365 	new_raw_count = cstate_pmu_read_counter(event);
366 
367 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
368 			    new_raw_count) != prev_raw_count)
369 		goto again;
370 
371 	local64_add(new_raw_count - prev_raw_count, &event->count);
372 }
373 
374 static void cstate_pmu_event_start(struct perf_event *event, int mode)
375 {
376 	local64_set(&event->hw.prev_count, cstate_pmu_read_counter(event));
377 }
378 
379 static void cstate_pmu_event_stop(struct perf_event *event, int mode)
380 {
381 	cstate_pmu_event_update(event);
382 }
383 
384 static void cstate_pmu_event_del(struct perf_event *event, int mode)
385 {
386 	cstate_pmu_event_stop(event, PERF_EF_UPDATE);
387 }
388 
389 static int cstate_pmu_event_add(struct perf_event *event, int mode)
390 {
391 	if (mode & PERF_EF_START)
392 		cstate_pmu_event_start(event, mode);
393 
394 	return 0;
395 }
396 
397 /*
398  * Check if exiting cpu is the designated reader. If so migrate the
399  * events when there is a valid target available
400  */
401 static int cstate_cpu_exit(unsigned int cpu)
402 {
403 	unsigned int target;
404 
405 	if (has_cstate_core &&
406 	    cpumask_test_and_clear_cpu(cpu, &cstate_core_cpu_mask)) {
407 
408 		target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
409 		/* Migrate events if there is a valid target */
410 		if (target < nr_cpu_ids) {
411 			cpumask_set_cpu(target, &cstate_core_cpu_mask);
412 			perf_pmu_migrate_context(&cstate_core_pmu, cpu, target);
413 		}
414 	}
415 
416 	if (has_cstate_pkg &&
417 	    cpumask_test_and_clear_cpu(cpu, &cstate_pkg_cpu_mask)) {
418 
419 		target = cpumask_any_but(topology_die_cpumask(cpu), cpu);
420 		/* Migrate events if there is a valid target */
421 		if (target < nr_cpu_ids) {
422 			cpumask_set_cpu(target, &cstate_pkg_cpu_mask);
423 			perf_pmu_migrate_context(&cstate_pkg_pmu, cpu, target);
424 		}
425 	}
426 	return 0;
427 }
428 
429 static int cstate_cpu_init(unsigned int cpu)
430 {
431 	unsigned int target;
432 
433 	/*
434 	 * If this is the first online thread of that core, set it in
435 	 * the core cpu mask as the designated reader.
436 	 */
437 	target = cpumask_any_and(&cstate_core_cpu_mask,
438 				 topology_sibling_cpumask(cpu));
439 
440 	if (has_cstate_core && target >= nr_cpu_ids)
441 		cpumask_set_cpu(cpu, &cstate_core_cpu_mask);
442 
443 	/*
444 	 * If this is the first online thread of that package, set it
445 	 * in the package cpu mask as the designated reader.
446 	 */
447 	target = cpumask_any_and(&cstate_pkg_cpu_mask,
448 				 topology_die_cpumask(cpu));
449 	if (has_cstate_pkg && target >= nr_cpu_ids)
450 		cpumask_set_cpu(cpu, &cstate_pkg_cpu_mask);
451 
452 	return 0;
453 }
454 
455 static const struct attribute_group *core_attr_update[] = {
456 	&group_cstate_core_c1,
457 	&group_cstate_core_c3,
458 	&group_cstate_core_c6,
459 	&group_cstate_core_c7,
460 	NULL,
461 };
462 
463 static const struct attribute_group *pkg_attr_update[] = {
464 	&group_cstate_pkg_c2,
465 	&group_cstate_pkg_c3,
466 	&group_cstate_pkg_c6,
467 	&group_cstate_pkg_c7,
468 	&group_cstate_pkg_c8,
469 	&group_cstate_pkg_c9,
470 	&group_cstate_pkg_c10,
471 	NULL,
472 };
473 
474 static struct pmu cstate_core_pmu = {
475 	.attr_groups	= core_attr_groups,
476 	.attr_update	= core_attr_update,
477 	.name		= "cstate_core",
478 	.task_ctx_nr	= perf_invalid_context,
479 	.event_init	= cstate_pmu_event_init,
480 	.add		= cstate_pmu_event_add,
481 	.del		= cstate_pmu_event_del,
482 	.start		= cstate_pmu_event_start,
483 	.stop		= cstate_pmu_event_stop,
484 	.read		= cstate_pmu_event_update,
485 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT | PERF_PMU_CAP_NO_EXCLUDE,
486 	.module		= THIS_MODULE,
487 };
488 
489 static struct pmu cstate_pkg_pmu = {
490 	.attr_groups	= pkg_attr_groups,
491 	.attr_update	= pkg_attr_update,
492 	.name		= "cstate_pkg",
493 	.task_ctx_nr	= perf_invalid_context,
494 	.event_init	= cstate_pmu_event_init,
495 	.add		= cstate_pmu_event_add,
496 	.del		= cstate_pmu_event_del,
497 	.start		= cstate_pmu_event_start,
498 	.stop		= cstate_pmu_event_stop,
499 	.read		= cstate_pmu_event_update,
500 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT | PERF_PMU_CAP_NO_EXCLUDE,
501 	.module		= THIS_MODULE,
502 };
503 
504 static const struct cstate_model nhm_cstates __initconst = {
505 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
506 				  BIT(PERF_CSTATE_CORE_C6_RES),
507 
508 	.pkg_events		= BIT(PERF_CSTATE_PKG_C3_RES) |
509 				  BIT(PERF_CSTATE_PKG_C6_RES) |
510 				  BIT(PERF_CSTATE_PKG_C7_RES),
511 };
512 
513 static const struct cstate_model snb_cstates __initconst = {
514 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
515 				  BIT(PERF_CSTATE_CORE_C6_RES) |
516 				  BIT(PERF_CSTATE_CORE_C7_RES),
517 
518 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
519 				  BIT(PERF_CSTATE_PKG_C3_RES) |
520 				  BIT(PERF_CSTATE_PKG_C6_RES) |
521 				  BIT(PERF_CSTATE_PKG_C7_RES),
522 };
523 
524 static const struct cstate_model hswult_cstates __initconst = {
525 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
526 				  BIT(PERF_CSTATE_CORE_C6_RES) |
527 				  BIT(PERF_CSTATE_CORE_C7_RES),
528 
529 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
530 				  BIT(PERF_CSTATE_PKG_C3_RES) |
531 				  BIT(PERF_CSTATE_PKG_C6_RES) |
532 				  BIT(PERF_CSTATE_PKG_C7_RES) |
533 				  BIT(PERF_CSTATE_PKG_C8_RES) |
534 				  BIT(PERF_CSTATE_PKG_C9_RES) |
535 				  BIT(PERF_CSTATE_PKG_C10_RES),
536 };
537 
538 static const struct cstate_model cnl_cstates __initconst = {
539 	.core_events		= BIT(PERF_CSTATE_CORE_C1_RES) |
540 				  BIT(PERF_CSTATE_CORE_C3_RES) |
541 				  BIT(PERF_CSTATE_CORE_C6_RES) |
542 				  BIT(PERF_CSTATE_CORE_C7_RES),
543 
544 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
545 				  BIT(PERF_CSTATE_PKG_C3_RES) |
546 				  BIT(PERF_CSTATE_PKG_C6_RES) |
547 				  BIT(PERF_CSTATE_PKG_C7_RES) |
548 				  BIT(PERF_CSTATE_PKG_C8_RES) |
549 				  BIT(PERF_CSTATE_PKG_C9_RES) |
550 				  BIT(PERF_CSTATE_PKG_C10_RES),
551 };
552 
553 static const struct cstate_model icl_cstates __initconst = {
554 	.core_events		= BIT(PERF_CSTATE_CORE_C6_RES) |
555 				  BIT(PERF_CSTATE_CORE_C7_RES),
556 
557 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
558 				  BIT(PERF_CSTATE_PKG_C3_RES) |
559 				  BIT(PERF_CSTATE_PKG_C6_RES) |
560 				  BIT(PERF_CSTATE_PKG_C7_RES) |
561 				  BIT(PERF_CSTATE_PKG_C8_RES) |
562 				  BIT(PERF_CSTATE_PKG_C9_RES) |
563 				  BIT(PERF_CSTATE_PKG_C10_RES),
564 };
565 
566 static const struct cstate_model slm_cstates __initconst = {
567 	.core_events		= BIT(PERF_CSTATE_CORE_C1_RES) |
568 				  BIT(PERF_CSTATE_CORE_C6_RES),
569 
570 	.pkg_events		= BIT(PERF_CSTATE_PKG_C6_RES),
571 	.quirks			= SLM_PKG_C6_USE_C7_MSR,
572 };
573 
574 
575 static const struct cstate_model knl_cstates __initconst = {
576 	.core_events		= BIT(PERF_CSTATE_CORE_C6_RES),
577 
578 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
579 				  BIT(PERF_CSTATE_PKG_C3_RES) |
580 				  BIT(PERF_CSTATE_PKG_C6_RES),
581 	.quirks			= KNL_CORE_C6_MSR,
582 };
583 
584 
585 static const struct cstate_model glm_cstates __initconst = {
586 	.core_events		= BIT(PERF_CSTATE_CORE_C1_RES) |
587 				  BIT(PERF_CSTATE_CORE_C3_RES) |
588 				  BIT(PERF_CSTATE_CORE_C6_RES),
589 
590 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
591 				  BIT(PERF_CSTATE_PKG_C3_RES) |
592 				  BIT(PERF_CSTATE_PKG_C6_RES) |
593 				  BIT(PERF_CSTATE_PKG_C10_RES),
594 };
595 
596 
597 static const struct x86_cpu_id intel_cstates_match[] __initconst = {
598 	X86_MATCH_INTEL_FAM6_MODEL(NEHALEM,		&nhm_cstates),
599 	X86_MATCH_INTEL_FAM6_MODEL(NEHALEM_EP,		&nhm_cstates),
600 	X86_MATCH_INTEL_FAM6_MODEL(NEHALEM_EX,		&nhm_cstates),
601 
602 	X86_MATCH_INTEL_FAM6_MODEL(WESTMERE,		&nhm_cstates),
603 	X86_MATCH_INTEL_FAM6_MODEL(WESTMERE_EP,		&nhm_cstates),
604 	X86_MATCH_INTEL_FAM6_MODEL(WESTMERE_EX,		&nhm_cstates),
605 
606 	X86_MATCH_INTEL_FAM6_MODEL(SANDYBRIDGE,		&snb_cstates),
607 	X86_MATCH_INTEL_FAM6_MODEL(SANDYBRIDGE_X,	&snb_cstates),
608 
609 	X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE,		&snb_cstates),
610 	X86_MATCH_INTEL_FAM6_MODEL(IVYBRIDGE_X,		&snb_cstates),
611 
612 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL,		&snb_cstates),
613 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X,		&snb_cstates),
614 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_G,		&snb_cstates),
615 
616 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_L,		&hswult_cstates),
617 
618 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_SILVERMONT,	&slm_cstates),
619 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_SILVERMONT_D,	&slm_cstates),
620 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_AIRMONT,	&slm_cstates),
621 
622 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL,		&snb_cstates),
623 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_D,		&snb_cstates),
624 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_G,		&snb_cstates),
625 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X,		&snb_cstates),
626 
627 	X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_L,		&snb_cstates),
628 	X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE,		&snb_cstates),
629 	X86_MATCH_INTEL_FAM6_MODEL(SKYLAKE_X,		&snb_cstates),
630 
631 	X86_MATCH_INTEL_FAM6_MODEL(KABYLAKE_L,		&hswult_cstates),
632 	X86_MATCH_INTEL_FAM6_MODEL(KABYLAKE,		&hswult_cstates),
633 	X86_MATCH_INTEL_FAM6_MODEL(COMETLAKE_L,		&hswult_cstates),
634 	X86_MATCH_INTEL_FAM6_MODEL(COMETLAKE,		&hswult_cstates),
635 
636 	X86_MATCH_INTEL_FAM6_MODEL(CANNONLAKE_L,	&cnl_cstates),
637 
638 	X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNL,	&knl_cstates),
639 	X86_MATCH_INTEL_FAM6_MODEL(XEON_PHI_KNM,	&knl_cstates),
640 
641 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT,	&glm_cstates),
642 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT_D,	&glm_cstates),
643 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_GOLDMONT_PLUS,	&glm_cstates),
644 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_TREMONT_D,	&glm_cstates),
645 	X86_MATCH_INTEL_FAM6_MODEL(ATOM_TREMONT,	&glm_cstates),
646 
647 	X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_L,		&icl_cstates),
648 	X86_MATCH_INTEL_FAM6_MODEL(ICELAKE,		&icl_cstates),
649 	X86_MATCH_INTEL_FAM6_MODEL(TIGERLAKE_L,		&icl_cstates),
650 	X86_MATCH_INTEL_FAM6_MODEL(TIGERLAKE,		&icl_cstates),
651 	{ },
652 };
653 MODULE_DEVICE_TABLE(x86cpu, intel_cstates_match);
654 
655 static int __init cstate_probe(const struct cstate_model *cm)
656 {
657 	/* SLM has different MSR for PKG C6 */
658 	if (cm->quirks & SLM_PKG_C6_USE_C7_MSR)
659 		pkg_msr[PERF_CSTATE_PKG_C6_RES].msr = MSR_PKG_C7_RESIDENCY;
660 
661 	/* KNL has different MSR for CORE C6 */
662 	if (cm->quirks & KNL_CORE_C6_MSR)
663 		pkg_msr[PERF_CSTATE_CORE_C6_RES].msr = MSR_KNL_CORE_C6_RESIDENCY;
664 
665 
666 	core_msr_mask = perf_msr_probe(core_msr, PERF_CSTATE_CORE_EVENT_MAX,
667 				       true, (void *) &cm->core_events);
668 
669 	pkg_msr_mask = perf_msr_probe(pkg_msr, PERF_CSTATE_PKG_EVENT_MAX,
670 				      true, (void *) &cm->pkg_events);
671 
672 	has_cstate_core = !!core_msr_mask;
673 	has_cstate_pkg  = !!pkg_msr_mask;
674 
675 	return (has_cstate_core || has_cstate_pkg) ? 0 : -ENODEV;
676 }
677 
678 static inline void cstate_cleanup(void)
679 {
680 	cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_CSTATE_ONLINE);
681 	cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_CSTATE_STARTING);
682 
683 	if (has_cstate_core)
684 		perf_pmu_unregister(&cstate_core_pmu);
685 
686 	if (has_cstate_pkg)
687 		perf_pmu_unregister(&cstate_pkg_pmu);
688 }
689 
690 static int __init cstate_init(void)
691 {
692 	int err;
693 
694 	cpuhp_setup_state(CPUHP_AP_PERF_X86_CSTATE_STARTING,
695 			  "perf/x86/cstate:starting", cstate_cpu_init, NULL);
696 	cpuhp_setup_state(CPUHP_AP_PERF_X86_CSTATE_ONLINE,
697 			  "perf/x86/cstate:online", NULL, cstate_cpu_exit);
698 
699 	if (has_cstate_core) {
700 		err = perf_pmu_register(&cstate_core_pmu, cstate_core_pmu.name, -1);
701 		if (err) {
702 			has_cstate_core = false;
703 			pr_info("Failed to register cstate core pmu\n");
704 			cstate_cleanup();
705 			return err;
706 		}
707 	}
708 
709 	if (has_cstate_pkg) {
710 		if (topology_max_die_per_package() > 1) {
711 			err = perf_pmu_register(&cstate_pkg_pmu,
712 						"cstate_die", -1);
713 		} else {
714 			err = perf_pmu_register(&cstate_pkg_pmu,
715 						cstate_pkg_pmu.name, -1);
716 		}
717 		if (err) {
718 			has_cstate_pkg = false;
719 			pr_info("Failed to register cstate pkg pmu\n");
720 			cstate_cleanup();
721 			return err;
722 		}
723 	}
724 	return 0;
725 }
726 
727 static int __init cstate_pmu_init(void)
728 {
729 	const struct x86_cpu_id *id;
730 	int err;
731 
732 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
733 		return -ENODEV;
734 
735 	id = x86_match_cpu(intel_cstates_match);
736 	if (!id)
737 		return -ENODEV;
738 
739 	err = cstate_probe((const struct cstate_model *) id->driver_data);
740 	if (err)
741 		return err;
742 
743 	return cstate_init();
744 }
745 module_init(cstate_pmu_init);
746 
747 static void __exit cstate_pmu_exit(void)
748 {
749 	cstate_cleanup();
750 }
751 module_exit(cstate_pmu_exit);
752