xref: /openbmc/linux/arch/x86/events/intel/cstate.c (revision a44e4f3a)
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
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
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  *			       Scope: Core
55  *	MSR_CORE_C7_RESIDENCY: CORE C7 Residency Counter
56  *			       perf code: 0x03
57  *			       Available model: SNB,IVB,HSW,BDW,SKL,CNL,KBL,CML,
58  *						ICL,TGL
59  *			       Scope: Core
60  *	MSR_PKG_C2_RESIDENCY:  Package C2 Residency Counter.
61  *			       perf code: 0x00
62  *			       Available model: SNB,IVB,HSW,BDW,SKL,KNL,GLM,CNL,
63  *						KBL,CML,ICL,TGL
64  *			       Scope: Package (physical package)
65  *	MSR_PKG_C3_RESIDENCY:  Package C3 Residency Counter.
66  *			       perf code: 0x01
67  *			       Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL,KNL,
68  *						GLM,CNL,KBL,CML,ICL,TGL
69  *			       Scope: Package (physical package)
70  *	MSR_PKG_C6_RESIDENCY:  Package C6 Residency Counter.
71  *			       perf code: 0x02
72  *			       Available model: SLM,AMT,NHM,WSM,SNB,IVB,HSW,BDW
73  *						SKL,KNL,GLM,CNL,KBL,CML,ICL,TGL
74  *			       Scope: Package (physical package)
75  *	MSR_PKG_C7_RESIDENCY:  Package C7 Residency Counter.
76  *			       perf code: 0x03
77  *			       Available model: NHM,WSM,SNB,IVB,HSW,BDW,SKL,CNL,
78  *						KBL,CML,ICL,TGL
79  *			       Scope: Package (physical package)
80  *	MSR_PKG_C8_RESIDENCY:  Package C8 Residency Counter.
81  *			       perf code: 0x04
82  *			       Available model: HSW ULT,KBL,CNL,CML,ICL,TGL
83  *			       Scope: Package (physical package)
84  *	MSR_PKG_C9_RESIDENCY:  Package C9 Residency Counter.
85  *			       perf code: 0x05
86  *			       Available model: HSW ULT,KBL,CNL,CML,ICL,TGL
87  *			       Scope: Package (physical package)
88  *	MSR_PKG_C10_RESIDENCY: Package C10 Residency Counter.
89  *			       perf code: 0x06
90  *			       Available model: HSW ULT,KBL,GLM,CNL,CML,ICL,TGL
91  *			       Scope: Package (physical package)
92  *
93  */
94 
95 #include <linux/module.h>
96 #include <linux/slab.h>
97 #include <linux/perf_event.h>
98 #include <linux/nospec.h>
99 #include <asm/cpu_device_id.h>
100 #include <asm/intel-family.h>
101 #include "../perf_event.h"
102 #include "../probe.h"
103 
104 MODULE_LICENSE("GPL");
105 
106 #define DEFINE_CSTATE_FORMAT_ATTR(_var, _name, _format)		\
107 static ssize_t __cstate_##_var##_show(struct kobject *kobj,	\
108 				struct kobj_attribute *attr,	\
109 				char *page)			\
110 {								\
111 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);		\
112 	return sprintf(page, _format "\n");			\
113 }								\
114 static struct kobj_attribute format_attr_##_var =		\
115 	__ATTR(_name, 0444, __cstate_##_var##_show, NULL)
116 
117 static ssize_t cstate_get_attr_cpumask(struct device *dev,
118 				       struct device_attribute *attr,
119 				       char *buf);
120 
121 /* Model -> events mapping */
122 struct cstate_model {
123 	unsigned long		core_events;
124 	unsigned long		pkg_events;
125 	unsigned long		quirks;
126 };
127 
128 /* Quirk flags */
129 #define SLM_PKG_C6_USE_C7_MSR	(1UL << 0)
130 #define KNL_CORE_C6_MSR		(1UL << 1)
131 
132 struct perf_cstate_msr {
133 	u64	msr;
134 	struct	perf_pmu_events_attr *attr;
135 };
136 
137 
138 /* cstate_core PMU */
139 static struct pmu cstate_core_pmu;
140 static bool has_cstate_core;
141 
142 enum perf_cstate_core_events {
143 	PERF_CSTATE_CORE_C1_RES = 0,
144 	PERF_CSTATE_CORE_C3_RES,
145 	PERF_CSTATE_CORE_C6_RES,
146 	PERF_CSTATE_CORE_C7_RES,
147 
148 	PERF_CSTATE_CORE_EVENT_MAX,
149 };
150 
151 PMU_EVENT_ATTR_STRING(c1-residency, attr_cstate_core_c1, "event=0x00");
152 PMU_EVENT_ATTR_STRING(c3-residency, attr_cstate_core_c3, "event=0x01");
153 PMU_EVENT_ATTR_STRING(c6-residency, attr_cstate_core_c6, "event=0x02");
154 PMU_EVENT_ATTR_STRING(c7-residency, attr_cstate_core_c7, "event=0x03");
155 
156 static unsigned long core_msr_mask;
157 
158 PMU_EVENT_GROUP(events, cstate_core_c1);
159 PMU_EVENT_GROUP(events, cstate_core_c3);
160 PMU_EVENT_GROUP(events, cstate_core_c6);
161 PMU_EVENT_GROUP(events, cstate_core_c7);
162 
163 static bool test_msr(int idx, void *data)
164 {
165 	return test_bit(idx, (unsigned long *) data);
166 }
167 
168 static struct perf_msr core_msr[] = {
169 	[PERF_CSTATE_CORE_C1_RES] = { MSR_CORE_C1_RES,		&group_cstate_core_c1,	test_msr },
170 	[PERF_CSTATE_CORE_C3_RES] = { MSR_CORE_C3_RESIDENCY,	&group_cstate_core_c3,	test_msr },
171 	[PERF_CSTATE_CORE_C6_RES] = { MSR_CORE_C6_RESIDENCY,	&group_cstate_core_c6,	test_msr },
172 	[PERF_CSTATE_CORE_C7_RES] = { MSR_CORE_C7_RESIDENCY,	&group_cstate_core_c7,	test_msr },
173 };
174 
175 static struct attribute *attrs_empty[] = {
176 	NULL,
177 };
178 
179 /*
180  * There are no default events, but we need to create
181  * "events" group (with empty attrs) before updating
182  * it with detected events.
183  */
184 static struct attribute_group core_events_attr_group = {
185 	.name = "events",
186 	.attrs = attrs_empty,
187 };
188 
189 DEFINE_CSTATE_FORMAT_ATTR(core_event, event, "config:0-63");
190 static struct attribute *core_format_attrs[] = {
191 	&format_attr_core_event.attr,
192 	NULL,
193 };
194 
195 static struct attribute_group core_format_attr_group = {
196 	.name = "format",
197 	.attrs = core_format_attrs,
198 };
199 
200 static cpumask_t cstate_core_cpu_mask;
201 static DEVICE_ATTR(cpumask, S_IRUGO, cstate_get_attr_cpumask, NULL);
202 
203 static struct attribute *cstate_cpumask_attrs[] = {
204 	&dev_attr_cpumask.attr,
205 	NULL,
206 };
207 
208 static struct attribute_group cpumask_attr_group = {
209 	.attrs = cstate_cpumask_attrs,
210 };
211 
212 static const struct attribute_group *core_attr_groups[] = {
213 	&core_events_attr_group,
214 	&core_format_attr_group,
215 	&cpumask_attr_group,
216 	NULL,
217 };
218 
219 /* cstate_pkg PMU */
220 static struct pmu cstate_pkg_pmu;
221 static bool has_cstate_pkg;
222 
223 enum perf_cstate_pkg_events {
224 	PERF_CSTATE_PKG_C2_RES = 0,
225 	PERF_CSTATE_PKG_C3_RES,
226 	PERF_CSTATE_PKG_C6_RES,
227 	PERF_CSTATE_PKG_C7_RES,
228 	PERF_CSTATE_PKG_C8_RES,
229 	PERF_CSTATE_PKG_C9_RES,
230 	PERF_CSTATE_PKG_C10_RES,
231 
232 	PERF_CSTATE_PKG_EVENT_MAX,
233 };
234 
235 PMU_EVENT_ATTR_STRING(c2-residency,  attr_cstate_pkg_c2,  "event=0x00");
236 PMU_EVENT_ATTR_STRING(c3-residency,  attr_cstate_pkg_c3,  "event=0x01");
237 PMU_EVENT_ATTR_STRING(c6-residency,  attr_cstate_pkg_c6,  "event=0x02");
238 PMU_EVENT_ATTR_STRING(c7-residency,  attr_cstate_pkg_c7,  "event=0x03");
239 PMU_EVENT_ATTR_STRING(c8-residency,  attr_cstate_pkg_c8,  "event=0x04");
240 PMU_EVENT_ATTR_STRING(c9-residency,  attr_cstate_pkg_c9,  "event=0x05");
241 PMU_EVENT_ATTR_STRING(c10-residency, attr_cstate_pkg_c10, "event=0x06");
242 
243 static unsigned long pkg_msr_mask;
244 
245 PMU_EVENT_GROUP(events, cstate_pkg_c2);
246 PMU_EVENT_GROUP(events, cstate_pkg_c3);
247 PMU_EVENT_GROUP(events, cstate_pkg_c6);
248 PMU_EVENT_GROUP(events, cstate_pkg_c7);
249 PMU_EVENT_GROUP(events, cstate_pkg_c8);
250 PMU_EVENT_GROUP(events, cstate_pkg_c9);
251 PMU_EVENT_GROUP(events, cstate_pkg_c10);
252 
253 static struct perf_msr pkg_msr[] = {
254 	[PERF_CSTATE_PKG_C2_RES]  = { MSR_PKG_C2_RESIDENCY,	&group_cstate_pkg_c2,	test_msr },
255 	[PERF_CSTATE_PKG_C3_RES]  = { MSR_PKG_C3_RESIDENCY,	&group_cstate_pkg_c3,	test_msr },
256 	[PERF_CSTATE_PKG_C6_RES]  = { MSR_PKG_C6_RESIDENCY,	&group_cstate_pkg_c6,	test_msr },
257 	[PERF_CSTATE_PKG_C7_RES]  = { MSR_PKG_C7_RESIDENCY,	&group_cstate_pkg_c7,	test_msr },
258 	[PERF_CSTATE_PKG_C8_RES]  = { MSR_PKG_C8_RESIDENCY,	&group_cstate_pkg_c8,	test_msr },
259 	[PERF_CSTATE_PKG_C9_RES]  = { MSR_PKG_C9_RESIDENCY,	&group_cstate_pkg_c9,	test_msr },
260 	[PERF_CSTATE_PKG_C10_RES] = { MSR_PKG_C10_RESIDENCY,	&group_cstate_pkg_c10,	test_msr },
261 };
262 
263 static struct attribute_group pkg_events_attr_group = {
264 	.name = "events",
265 	.attrs = attrs_empty,
266 };
267 
268 DEFINE_CSTATE_FORMAT_ATTR(pkg_event, event, "config:0-63");
269 static struct attribute *pkg_format_attrs[] = {
270 	&format_attr_pkg_event.attr,
271 	NULL,
272 };
273 static struct attribute_group pkg_format_attr_group = {
274 	.name = "format",
275 	.attrs = pkg_format_attrs,
276 };
277 
278 static cpumask_t cstate_pkg_cpu_mask;
279 
280 static const struct attribute_group *pkg_attr_groups[] = {
281 	&pkg_events_attr_group,
282 	&pkg_format_attr_group,
283 	&cpumask_attr_group,
284 	NULL,
285 };
286 
287 static ssize_t cstate_get_attr_cpumask(struct device *dev,
288 				       struct device_attribute *attr,
289 				       char *buf)
290 {
291 	struct pmu *pmu = dev_get_drvdata(dev);
292 
293 	if (pmu == &cstate_core_pmu)
294 		return cpumap_print_to_pagebuf(true, buf, &cstate_core_cpu_mask);
295 	else if (pmu == &cstate_pkg_pmu)
296 		return cpumap_print_to_pagebuf(true, buf, &cstate_pkg_cpu_mask);
297 	else
298 		return 0;
299 }
300 
301 static int cstate_pmu_event_init(struct perf_event *event)
302 {
303 	u64 cfg = event->attr.config;
304 	int cpu;
305 
306 	if (event->attr.type != event->pmu->type)
307 		return -ENOENT;
308 
309 	/* unsupported modes and filters */
310 	if (event->attr.sample_period) /* no sampling */
311 		return -EINVAL;
312 
313 	if (event->cpu < 0)
314 		return -EINVAL;
315 
316 	if (event->pmu == &cstate_core_pmu) {
317 		if (cfg >= PERF_CSTATE_CORE_EVENT_MAX)
318 			return -EINVAL;
319 		cfg = array_index_nospec((unsigned long)cfg, PERF_CSTATE_CORE_EVENT_MAX);
320 		if (!(core_msr_mask & (1 << cfg)))
321 			return -EINVAL;
322 		event->hw.event_base = core_msr[cfg].msr;
323 		cpu = cpumask_any_and(&cstate_core_cpu_mask,
324 				      topology_sibling_cpumask(event->cpu));
325 	} else if (event->pmu == &cstate_pkg_pmu) {
326 		if (cfg >= PERF_CSTATE_PKG_EVENT_MAX)
327 			return -EINVAL;
328 		cfg = array_index_nospec((unsigned long)cfg, PERF_CSTATE_PKG_EVENT_MAX);
329 		if (!(pkg_msr_mask & (1 << cfg)))
330 			return -EINVAL;
331 		event->hw.event_base = pkg_msr[cfg].msr;
332 		cpu = cpumask_any_and(&cstate_pkg_cpu_mask,
333 				      topology_die_cpumask(event->cpu));
334 	} else {
335 		return -ENOENT;
336 	}
337 
338 	if (cpu >= nr_cpu_ids)
339 		return -ENODEV;
340 
341 	event->cpu = cpu;
342 	event->hw.config = cfg;
343 	event->hw.idx = -1;
344 	return 0;
345 }
346 
347 static inline u64 cstate_pmu_read_counter(struct perf_event *event)
348 {
349 	u64 val;
350 
351 	rdmsrl(event->hw.event_base, val);
352 	return val;
353 }
354 
355 static void cstate_pmu_event_update(struct perf_event *event)
356 {
357 	struct hw_perf_event *hwc = &event->hw;
358 	u64 prev_raw_count, new_raw_count;
359 
360 again:
361 	prev_raw_count = local64_read(&hwc->prev_count);
362 	new_raw_count = cstate_pmu_read_counter(event);
363 
364 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
365 			    new_raw_count) != prev_raw_count)
366 		goto again;
367 
368 	local64_add(new_raw_count - prev_raw_count, &event->count);
369 }
370 
371 static void cstate_pmu_event_start(struct perf_event *event, int mode)
372 {
373 	local64_set(&event->hw.prev_count, cstate_pmu_read_counter(event));
374 }
375 
376 static void cstate_pmu_event_stop(struct perf_event *event, int mode)
377 {
378 	cstate_pmu_event_update(event);
379 }
380 
381 static void cstate_pmu_event_del(struct perf_event *event, int mode)
382 {
383 	cstate_pmu_event_stop(event, PERF_EF_UPDATE);
384 }
385 
386 static int cstate_pmu_event_add(struct perf_event *event, int mode)
387 {
388 	if (mode & PERF_EF_START)
389 		cstate_pmu_event_start(event, mode);
390 
391 	return 0;
392 }
393 
394 /*
395  * Check if exiting cpu is the designated reader. If so migrate the
396  * events when there is a valid target available
397  */
398 static int cstate_cpu_exit(unsigned int cpu)
399 {
400 	unsigned int target;
401 
402 	if (has_cstate_core &&
403 	    cpumask_test_and_clear_cpu(cpu, &cstate_core_cpu_mask)) {
404 
405 		target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
406 		/* Migrate events if there is a valid target */
407 		if (target < nr_cpu_ids) {
408 			cpumask_set_cpu(target, &cstate_core_cpu_mask);
409 			perf_pmu_migrate_context(&cstate_core_pmu, cpu, target);
410 		}
411 	}
412 
413 	if (has_cstate_pkg &&
414 	    cpumask_test_and_clear_cpu(cpu, &cstate_pkg_cpu_mask)) {
415 
416 		target = cpumask_any_but(topology_die_cpumask(cpu), cpu);
417 		/* Migrate events if there is a valid target */
418 		if (target < nr_cpu_ids) {
419 			cpumask_set_cpu(target, &cstate_pkg_cpu_mask);
420 			perf_pmu_migrate_context(&cstate_pkg_pmu, cpu, target);
421 		}
422 	}
423 	return 0;
424 }
425 
426 static int cstate_cpu_init(unsigned int cpu)
427 {
428 	unsigned int target;
429 
430 	/*
431 	 * If this is the first online thread of that core, set it in
432 	 * the core cpu mask as the designated reader.
433 	 */
434 	target = cpumask_any_and(&cstate_core_cpu_mask,
435 				 topology_sibling_cpumask(cpu));
436 
437 	if (has_cstate_core && target >= nr_cpu_ids)
438 		cpumask_set_cpu(cpu, &cstate_core_cpu_mask);
439 
440 	/*
441 	 * If this is the first online thread of that package, set it
442 	 * in the package cpu mask as the designated reader.
443 	 */
444 	target = cpumask_any_and(&cstate_pkg_cpu_mask,
445 				 topology_die_cpumask(cpu));
446 	if (has_cstate_pkg && target >= nr_cpu_ids)
447 		cpumask_set_cpu(cpu, &cstate_pkg_cpu_mask);
448 
449 	return 0;
450 }
451 
452 static const struct attribute_group *core_attr_update[] = {
453 	&group_cstate_core_c1,
454 	&group_cstate_core_c3,
455 	&group_cstate_core_c6,
456 	&group_cstate_core_c7,
457 	NULL,
458 };
459 
460 static const struct attribute_group *pkg_attr_update[] = {
461 	&group_cstate_pkg_c2,
462 	&group_cstate_pkg_c3,
463 	&group_cstate_pkg_c6,
464 	&group_cstate_pkg_c7,
465 	&group_cstate_pkg_c8,
466 	&group_cstate_pkg_c9,
467 	&group_cstate_pkg_c10,
468 	NULL,
469 };
470 
471 static struct pmu cstate_core_pmu = {
472 	.attr_groups	= core_attr_groups,
473 	.attr_update	= core_attr_update,
474 	.name		= "cstate_core",
475 	.task_ctx_nr	= perf_invalid_context,
476 	.event_init	= cstate_pmu_event_init,
477 	.add		= cstate_pmu_event_add,
478 	.del		= cstate_pmu_event_del,
479 	.start		= cstate_pmu_event_start,
480 	.stop		= cstate_pmu_event_stop,
481 	.read		= cstate_pmu_event_update,
482 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT | PERF_PMU_CAP_NO_EXCLUDE,
483 	.module		= THIS_MODULE,
484 };
485 
486 static struct pmu cstate_pkg_pmu = {
487 	.attr_groups	= pkg_attr_groups,
488 	.attr_update	= pkg_attr_update,
489 	.name		= "cstate_pkg",
490 	.task_ctx_nr	= perf_invalid_context,
491 	.event_init	= cstate_pmu_event_init,
492 	.add		= cstate_pmu_event_add,
493 	.del		= cstate_pmu_event_del,
494 	.start		= cstate_pmu_event_start,
495 	.stop		= cstate_pmu_event_stop,
496 	.read		= cstate_pmu_event_update,
497 	.capabilities	= PERF_PMU_CAP_NO_INTERRUPT | PERF_PMU_CAP_NO_EXCLUDE,
498 	.module		= THIS_MODULE,
499 };
500 
501 static const struct cstate_model nhm_cstates __initconst = {
502 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
503 				  BIT(PERF_CSTATE_CORE_C6_RES),
504 
505 	.pkg_events		= BIT(PERF_CSTATE_PKG_C3_RES) |
506 				  BIT(PERF_CSTATE_PKG_C6_RES) |
507 				  BIT(PERF_CSTATE_PKG_C7_RES),
508 };
509 
510 static const struct cstate_model snb_cstates __initconst = {
511 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
512 				  BIT(PERF_CSTATE_CORE_C6_RES) |
513 				  BIT(PERF_CSTATE_CORE_C7_RES),
514 
515 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
516 				  BIT(PERF_CSTATE_PKG_C3_RES) |
517 				  BIT(PERF_CSTATE_PKG_C6_RES) |
518 				  BIT(PERF_CSTATE_PKG_C7_RES),
519 };
520 
521 static const struct cstate_model hswult_cstates __initconst = {
522 	.core_events		= BIT(PERF_CSTATE_CORE_C3_RES) |
523 				  BIT(PERF_CSTATE_CORE_C6_RES) |
524 				  BIT(PERF_CSTATE_CORE_C7_RES),
525 
526 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
527 				  BIT(PERF_CSTATE_PKG_C3_RES) |
528 				  BIT(PERF_CSTATE_PKG_C6_RES) |
529 				  BIT(PERF_CSTATE_PKG_C7_RES) |
530 				  BIT(PERF_CSTATE_PKG_C8_RES) |
531 				  BIT(PERF_CSTATE_PKG_C9_RES) |
532 				  BIT(PERF_CSTATE_PKG_C10_RES),
533 };
534 
535 static const struct cstate_model cnl_cstates __initconst = {
536 	.core_events		= BIT(PERF_CSTATE_CORE_C1_RES) |
537 				  BIT(PERF_CSTATE_CORE_C3_RES) |
538 				  BIT(PERF_CSTATE_CORE_C6_RES) |
539 				  BIT(PERF_CSTATE_CORE_C7_RES),
540 
541 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
542 				  BIT(PERF_CSTATE_PKG_C3_RES) |
543 				  BIT(PERF_CSTATE_PKG_C6_RES) |
544 				  BIT(PERF_CSTATE_PKG_C7_RES) |
545 				  BIT(PERF_CSTATE_PKG_C8_RES) |
546 				  BIT(PERF_CSTATE_PKG_C9_RES) |
547 				  BIT(PERF_CSTATE_PKG_C10_RES),
548 };
549 
550 static const struct cstate_model icl_cstates __initconst = {
551 	.core_events		= BIT(PERF_CSTATE_CORE_C6_RES) |
552 				  BIT(PERF_CSTATE_CORE_C7_RES),
553 
554 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
555 				  BIT(PERF_CSTATE_PKG_C3_RES) |
556 				  BIT(PERF_CSTATE_PKG_C6_RES) |
557 				  BIT(PERF_CSTATE_PKG_C7_RES) |
558 				  BIT(PERF_CSTATE_PKG_C8_RES) |
559 				  BIT(PERF_CSTATE_PKG_C9_RES) |
560 				  BIT(PERF_CSTATE_PKG_C10_RES),
561 };
562 
563 static const struct cstate_model slm_cstates __initconst = {
564 	.core_events		= BIT(PERF_CSTATE_CORE_C1_RES) |
565 				  BIT(PERF_CSTATE_CORE_C6_RES),
566 
567 	.pkg_events		= BIT(PERF_CSTATE_PKG_C6_RES),
568 	.quirks			= SLM_PKG_C6_USE_C7_MSR,
569 };
570 
571 
572 static const struct cstate_model knl_cstates __initconst = {
573 	.core_events		= BIT(PERF_CSTATE_CORE_C6_RES),
574 
575 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
576 				  BIT(PERF_CSTATE_PKG_C3_RES) |
577 				  BIT(PERF_CSTATE_PKG_C6_RES),
578 	.quirks			= KNL_CORE_C6_MSR,
579 };
580 
581 
582 static const struct cstate_model glm_cstates __initconst = {
583 	.core_events		= BIT(PERF_CSTATE_CORE_C1_RES) |
584 				  BIT(PERF_CSTATE_CORE_C3_RES) |
585 				  BIT(PERF_CSTATE_CORE_C6_RES),
586 
587 	.pkg_events		= BIT(PERF_CSTATE_PKG_C2_RES) |
588 				  BIT(PERF_CSTATE_PKG_C3_RES) |
589 				  BIT(PERF_CSTATE_PKG_C6_RES) |
590 				  BIT(PERF_CSTATE_PKG_C10_RES),
591 };
592 
593 
594 #define X86_CSTATES_MODEL(model, states)				\
595 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY, (unsigned long) &(states) }
596 
597 static const struct x86_cpu_id intel_cstates_match[] __initconst = {
598 	X86_CSTATES_MODEL(INTEL_FAM6_NEHALEM,    nhm_cstates),
599 	X86_CSTATES_MODEL(INTEL_FAM6_NEHALEM_EP, nhm_cstates),
600 	X86_CSTATES_MODEL(INTEL_FAM6_NEHALEM_EX, nhm_cstates),
601 
602 	X86_CSTATES_MODEL(INTEL_FAM6_WESTMERE,    nhm_cstates),
603 	X86_CSTATES_MODEL(INTEL_FAM6_WESTMERE_EP, nhm_cstates),
604 	X86_CSTATES_MODEL(INTEL_FAM6_WESTMERE_EX, nhm_cstates),
605 
606 	X86_CSTATES_MODEL(INTEL_FAM6_SANDYBRIDGE,   snb_cstates),
607 	X86_CSTATES_MODEL(INTEL_FAM6_SANDYBRIDGE_X, snb_cstates),
608 
609 	X86_CSTATES_MODEL(INTEL_FAM6_IVYBRIDGE,   snb_cstates),
610 	X86_CSTATES_MODEL(INTEL_FAM6_IVYBRIDGE_X, snb_cstates),
611 
612 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL,   snb_cstates),
613 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL_X, snb_cstates),
614 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL_G, snb_cstates),
615 
616 	X86_CSTATES_MODEL(INTEL_FAM6_HASWELL_L, hswult_cstates),
617 
618 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_SILVERMONT,   slm_cstates),
619 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_SILVERMONT_D, slm_cstates),
620 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_AIRMONT,      slm_cstates),
621 
622 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL,   snb_cstates),
623 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL_D, snb_cstates),
624 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL_G, snb_cstates),
625 	X86_CSTATES_MODEL(INTEL_FAM6_BROADWELL_X, snb_cstates),
626 
627 	X86_CSTATES_MODEL(INTEL_FAM6_SKYLAKE_L, snb_cstates),
628 	X86_CSTATES_MODEL(INTEL_FAM6_SKYLAKE,   snb_cstates),
629 	X86_CSTATES_MODEL(INTEL_FAM6_SKYLAKE_X, snb_cstates),
630 
631 	X86_CSTATES_MODEL(INTEL_FAM6_KABYLAKE_L, hswult_cstates),
632 	X86_CSTATES_MODEL(INTEL_FAM6_KABYLAKE,   hswult_cstates),
633 	X86_CSTATES_MODEL(INTEL_FAM6_COMETLAKE_L, hswult_cstates),
634 	X86_CSTATES_MODEL(INTEL_FAM6_COMETLAKE, hswult_cstates),
635 
636 	X86_CSTATES_MODEL(INTEL_FAM6_CANNONLAKE_L, cnl_cstates),
637 
638 	X86_CSTATES_MODEL(INTEL_FAM6_XEON_PHI_KNL, knl_cstates),
639 	X86_CSTATES_MODEL(INTEL_FAM6_XEON_PHI_KNM, knl_cstates),
640 
641 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_GOLDMONT,   glm_cstates),
642 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_GOLDMONT_D, glm_cstates),
643 
644 	X86_CSTATES_MODEL(INTEL_FAM6_ATOM_GOLDMONT_PLUS, glm_cstates),
645 
646 	X86_CSTATES_MODEL(INTEL_FAM6_ICELAKE_L, icl_cstates),
647 	X86_CSTATES_MODEL(INTEL_FAM6_ICELAKE,   icl_cstates),
648 	X86_CSTATES_MODEL(INTEL_FAM6_TIGERLAKE_L, icl_cstates),
649 	X86_CSTATES_MODEL(INTEL_FAM6_TIGERLAKE, icl_cstates),
650 	{ },
651 };
652 MODULE_DEVICE_TABLE(x86cpu, intel_cstates_match);
653 
654 static int __init cstate_probe(const struct cstate_model *cm)
655 {
656 	/* SLM has different MSR for PKG C6 */
657 	if (cm->quirks & SLM_PKG_C6_USE_C7_MSR)
658 		pkg_msr[PERF_CSTATE_PKG_C6_RES].msr = MSR_PKG_C7_RESIDENCY;
659 
660 	/* KNL has different MSR for CORE C6 */
661 	if (cm->quirks & KNL_CORE_C6_MSR)
662 		pkg_msr[PERF_CSTATE_CORE_C6_RES].msr = MSR_KNL_CORE_C6_RESIDENCY;
663 
664 
665 	core_msr_mask = perf_msr_probe(core_msr, PERF_CSTATE_CORE_EVENT_MAX,
666 				       true, (void *) &cm->core_events);
667 
668 	pkg_msr_mask = perf_msr_probe(pkg_msr, PERF_CSTATE_PKG_EVENT_MAX,
669 				      true, (void *) &cm->pkg_events);
670 
671 	has_cstate_core = !!core_msr_mask;
672 	has_cstate_pkg  = !!pkg_msr_mask;
673 
674 	return (has_cstate_core || has_cstate_pkg) ? 0 : -ENODEV;
675 }
676 
677 static inline void cstate_cleanup(void)
678 {
679 	cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_CSTATE_ONLINE);
680 	cpuhp_remove_state_nocalls(CPUHP_AP_PERF_X86_CSTATE_STARTING);
681 
682 	if (has_cstate_core)
683 		perf_pmu_unregister(&cstate_core_pmu);
684 
685 	if (has_cstate_pkg)
686 		perf_pmu_unregister(&cstate_pkg_pmu);
687 }
688 
689 static int __init cstate_init(void)
690 {
691 	int err;
692 
693 	cpuhp_setup_state(CPUHP_AP_PERF_X86_CSTATE_STARTING,
694 			  "perf/x86/cstate:starting", cstate_cpu_init, NULL);
695 	cpuhp_setup_state(CPUHP_AP_PERF_X86_CSTATE_ONLINE,
696 			  "perf/x86/cstate:online", NULL, cstate_cpu_exit);
697 
698 	if (has_cstate_core) {
699 		err = perf_pmu_register(&cstate_core_pmu, cstate_core_pmu.name, -1);
700 		if (err) {
701 			has_cstate_core = false;
702 			pr_info("Failed to register cstate core pmu\n");
703 			cstate_cleanup();
704 			return err;
705 		}
706 	}
707 
708 	if (has_cstate_pkg) {
709 		if (topology_max_die_per_package() > 1) {
710 			err = perf_pmu_register(&cstate_pkg_pmu,
711 						"cstate_die", -1);
712 		} else {
713 			err = perf_pmu_register(&cstate_pkg_pmu,
714 						cstate_pkg_pmu.name, -1);
715 		}
716 		if (err) {
717 			has_cstate_pkg = false;
718 			pr_info("Failed to register cstate pkg pmu\n");
719 			cstate_cleanup();
720 			return err;
721 		}
722 	}
723 	return 0;
724 }
725 
726 static int __init cstate_pmu_init(void)
727 {
728 	const struct x86_cpu_id *id;
729 	int err;
730 
731 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
732 		return -ENODEV;
733 
734 	id = x86_match_cpu(intel_cstates_match);
735 	if (!id)
736 		return -ENODEV;
737 
738 	err = cstate_probe((const struct cstate_model *) id->driver_data);
739 	if (err)
740 		return err;
741 
742 	return cstate_init();
743 }
744 module_init(cstate_pmu_init);
745 
746 static void __exit cstate_pmu_exit(void)
747 {
748 	cstate_cleanup();
749 }
750 module_exit(cstate_pmu_exit);
751