xref: /openbmc/linux/arch/s390/kernel/perf_cpum_cf.c (revision 275876e2)
1 /*
2  * Performance event support for s390x - CPU-measurement Counter Facility
3  *
4  *  Copyright IBM Corp. 2012
5  *  Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License (version 2 only)
9  * as published by the Free Software Foundation.
10  */
11 #define KMSG_COMPONENT	"cpum_cf"
12 #define pr_fmt(fmt)	KMSG_COMPONENT ": " fmt
13 
14 #include <linux/kernel.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/perf_event.h>
17 #include <linux/percpu.h>
18 #include <linux/notifier.h>
19 #include <linux/init.h>
20 #include <linux/export.h>
21 #include <asm/ctl_reg.h>
22 #include <asm/irq.h>
23 #include <asm/cpu_mf.h>
24 
25 /* CPU-measurement counter facility supports these CPU counter sets:
26  * For CPU counter sets:
27  *    Basic counter set:	     0-31
28  *    Problem-state counter set:    32-63
29  *    Crypto-activity counter set:  64-127
30  *    Extented counter set:	   128-159
31  */
32 enum cpumf_ctr_set {
33 	/* CPU counter sets */
34 	CPUMF_CTR_SET_BASIC   = 0,
35 	CPUMF_CTR_SET_USER    = 1,
36 	CPUMF_CTR_SET_CRYPTO  = 2,
37 	CPUMF_CTR_SET_EXT     = 3,
38 
39 	/* Maximum number of counter sets */
40 	CPUMF_CTR_SET_MAX,
41 };
42 
43 #define CPUMF_LCCTL_ENABLE_SHIFT    16
44 #define CPUMF_LCCTL_ACTCTL_SHIFT     0
45 static const u64 cpumf_state_ctl[CPUMF_CTR_SET_MAX] = {
46 	[CPUMF_CTR_SET_BASIC]	= 0x02,
47 	[CPUMF_CTR_SET_USER]	= 0x04,
48 	[CPUMF_CTR_SET_CRYPTO]	= 0x08,
49 	[CPUMF_CTR_SET_EXT]	= 0x01,
50 };
51 
52 static void ctr_set_enable(u64 *state, int ctr_set)
53 {
54 	*state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT;
55 }
56 static void ctr_set_disable(u64 *state, int ctr_set)
57 {
58 	*state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ENABLE_SHIFT);
59 }
60 static void ctr_set_start(u64 *state, int ctr_set)
61 {
62 	*state |= cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT;
63 }
64 static void ctr_set_stop(u64 *state, int ctr_set)
65 {
66 	*state &= ~(cpumf_state_ctl[ctr_set] << CPUMF_LCCTL_ACTCTL_SHIFT);
67 }
68 
69 /* Local CPUMF event structure */
70 struct cpu_hw_events {
71 	struct cpumf_ctr_info	info;
72 	atomic_t		ctr_set[CPUMF_CTR_SET_MAX];
73 	u64			state, tx_state;
74 	unsigned int		flags;
75 };
76 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
77 	.ctr_set = {
78 		[CPUMF_CTR_SET_BASIC]  = ATOMIC_INIT(0),
79 		[CPUMF_CTR_SET_USER]   = ATOMIC_INIT(0),
80 		[CPUMF_CTR_SET_CRYPTO] = ATOMIC_INIT(0),
81 		[CPUMF_CTR_SET_EXT]    = ATOMIC_INIT(0),
82 	},
83 	.state = 0,
84 	.flags = 0,
85 };
86 
87 static int get_counter_set(u64 event)
88 {
89 	int set = -1;
90 
91 	if (event < 32)
92 		set = CPUMF_CTR_SET_BASIC;
93 	else if (event < 64)
94 		set = CPUMF_CTR_SET_USER;
95 	else if (event < 128)
96 		set = CPUMF_CTR_SET_CRYPTO;
97 	else if (event < 256)
98 		set = CPUMF_CTR_SET_EXT;
99 
100 	return set;
101 }
102 
103 static int validate_event(const struct hw_perf_event *hwc)
104 {
105 	switch (hwc->config_base) {
106 	case CPUMF_CTR_SET_BASIC:
107 	case CPUMF_CTR_SET_USER:
108 	case CPUMF_CTR_SET_CRYPTO:
109 	case CPUMF_CTR_SET_EXT:
110 		/* check for reserved counters */
111 		if ((hwc->config >=  6 && hwc->config <=  31) ||
112 		    (hwc->config >= 38 && hwc->config <=  63) ||
113 		    (hwc->config >= 80 && hwc->config <= 127))
114 			return -EOPNOTSUPP;
115 		break;
116 	default:
117 		return -EINVAL;
118 	}
119 
120 	return 0;
121 }
122 
123 static int validate_ctr_version(const struct hw_perf_event *hwc)
124 {
125 	struct cpu_hw_events *cpuhw;
126 	int err = 0;
127 
128 	cpuhw = &get_cpu_var(cpu_hw_events);
129 
130 	/* check required version for counter sets */
131 	switch (hwc->config_base) {
132 	case CPUMF_CTR_SET_BASIC:
133 	case CPUMF_CTR_SET_USER:
134 		if (cpuhw->info.cfvn < 1)
135 			err = -EOPNOTSUPP;
136 		break;
137 	case CPUMF_CTR_SET_CRYPTO:
138 	case CPUMF_CTR_SET_EXT:
139 		if (cpuhw->info.csvn < 1)
140 			err = -EOPNOTSUPP;
141 		if ((cpuhw->info.csvn == 1 && hwc->config > 159) ||
142 		    (cpuhw->info.csvn == 2 && hwc->config > 175) ||
143 		    (cpuhw->info.csvn  > 2 && hwc->config > 255))
144 			err = -EOPNOTSUPP;
145 		break;
146 	}
147 
148 	put_cpu_var(cpu_hw_events);
149 	return err;
150 }
151 
152 static int validate_ctr_auth(const struct hw_perf_event *hwc)
153 {
154 	struct cpu_hw_events *cpuhw;
155 	u64 ctrs_state;
156 	int err = 0;
157 
158 	cpuhw = &get_cpu_var(cpu_hw_events);
159 
160 	/* check authorization for cpu counter sets */
161 	ctrs_state = cpumf_state_ctl[hwc->config_base];
162 	if (!(ctrs_state & cpuhw->info.auth_ctl))
163 		err = -EPERM;
164 
165 	put_cpu_var(cpu_hw_events);
166 	return err;
167 }
168 
169 /*
170  * Change the CPUMF state to active.
171  * Enable and activate the CPU-counter sets according
172  * to the per-cpu control state.
173  */
174 static void cpumf_pmu_enable(struct pmu *pmu)
175 {
176 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
177 	int err;
178 
179 	if (cpuhw->flags & PMU_F_ENABLED)
180 		return;
181 
182 	err = lcctl(cpuhw->state);
183 	if (err) {
184 		pr_err("Enabling the performance measuring unit "
185 		       "failed with rc=%x\n", err);
186 		return;
187 	}
188 
189 	cpuhw->flags |= PMU_F_ENABLED;
190 }
191 
192 /*
193  * Change the CPUMF state to inactive.
194  * Disable and enable (inactive) the CPU-counter sets according
195  * to the per-cpu control state.
196  */
197 static void cpumf_pmu_disable(struct pmu *pmu)
198 {
199 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
200 	int err;
201 	u64 inactive;
202 
203 	if (!(cpuhw->flags & PMU_F_ENABLED))
204 		return;
205 
206 	inactive = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
207 	err = lcctl(inactive);
208 	if (err) {
209 		pr_err("Disabling the performance measuring unit "
210 		       "failed with rc=%x\n", err);
211 		return;
212 	}
213 
214 	cpuhw->flags &= ~PMU_F_ENABLED;
215 }
216 
217 
218 /* Number of perf events counting hardware events */
219 static atomic_t num_events = ATOMIC_INIT(0);
220 /* Used to avoid races in calling reserve/release_cpumf_hardware */
221 static DEFINE_MUTEX(pmc_reserve_mutex);
222 
223 /* CPU-measurement alerts for the counter facility */
224 static void cpumf_measurement_alert(struct ext_code ext_code,
225 				    unsigned int alert, unsigned long unused)
226 {
227 	struct cpu_hw_events *cpuhw;
228 
229 	if (!(alert & CPU_MF_INT_CF_MASK))
230 		return;
231 
232 	inc_irq_stat(IRQEXT_CMC);
233 	cpuhw = &__get_cpu_var(cpu_hw_events);
234 
235 	/* Measurement alerts are shared and might happen when the PMU
236 	 * is not reserved.  Ignore these alerts in this case. */
237 	if (!(cpuhw->flags & PMU_F_RESERVED))
238 		return;
239 
240 	/* counter authorization change alert */
241 	if (alert & CPU_MF_INT_CF_CACA)
242 		qctri(&cpuhw->info);
243 
244 	/* loss of counter data alert */
245 	if (alert & CPU_MF_INT_CF_LCDA)
246 		pr_err("CPU[%i] Counter data was lost\n", smp_processor_id());
247 }
248 
249 #define PMC_INIT      0
250 #define PMC_RELEASE   1
251 static void setup_pmc_cpu(void *flags)
252 {
253 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
254 
255 	switch (*((int *) flags)) {
256 	case PMC_INIT:
257 		memset(&cpuhw->info, 0, sizeof(cpuhw->info));
258 		qctri(&cpuhw->info);
259 		cpuhw->flags |= PMU_F_RESERVED;
260 		break;
261 
262 	case PMC_RELEASE:
263 		cpuhw->flags &= ~PMU_F_RESERVED;
264 		break;
265 	}
266 
267 	/* Disable CPU counter sets */
268 	lcctl(0);
269 }
270 
271 /* Initialize the CPU-measurement facility */
272 static int reserve_pmc_hardware(void)
273 {
274 	int flags = PMC_INIT;
275 
276 	on_each_cpu(setup_pmc_cpu, &flags, 1);
277 	irq_subclass_register(IRQ_SUBCLASS_MEASUREMENT_ALERT);
278 
279 	return 0;
280 }
281 
282 /* Release the CPU-measurement facility */
283 static void release_pmc_hardware(void)
284 {
285 	int flags = PMC_RELEASE;
286 
287 	on_each_cpu(setup_pmc_cpu, &flags, 1);
288 	irq_subclass_unregister(IRQ_SUBCLASS_MEASUREMENT_ALERT);
289 }
290 
291 /* Release the PMU if event is the last perf event */
292 static void hw_perf_event_destroy(struct perf_event *event)
293 {
294 	if (!atomic_add_unless(&num_events, -1, 1)) {
295 		mutex_lock(&pmc_reserve_mutex);
296 		if (atomic_dec_return(&num_events) == 0)
297 			release_pmc_hardware();
298 		mutex_unlock(&pmc_reserve_mutex);
299 	}
300 }
301 
302 /* CPUMF <-> perf event mappings for kernel+userspace (basic set) */
303 static const int cpumf_generic_events_basic[] = {
304 	[PERF_COUNT_HW_CPU_CYCLES]	    = 0,
305 	[PERF_COUNT_HW_INSTRUCTIONS]	    = 1,
306 	[PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
307 	[PERF_COUNT_HW_CACHE_MISSES]	    = -1,
308 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
309 	[PERF_COUNT_HW_BRANCH_MISSES]	    = -1,
310 	[PERF_COUNT_HW_BUS_CYCLES]	    = -1,
311 };
312 /* CPUMF <-> perf event mappings for userspace (problem-state set) */
313 static const int cpumf_generic_events_user[] = {
314 	[PERF_COUNT_HW_CPU_CYCLES]	    = 32,
315 	[PERF_COUNT_HW_INSTRUCTIONS]	    = 33,
316 	[PERF_COUNT_HW_CACHE_REFERENCES]    = -1,
317 	[PERF_COUNT_HW_CACHE_MISSES]	    = -1,
318 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -1,
319 	[PERF_COUNT_HW_BRANCH_MISSES]	    = -1,
320 	[PERF_COUNT_HW_BUS_CYCLES]	    = -1,
321 };
322 
323 static int __hw_perf_event_init(struct perf_event *event)
324 {
325 	struct perf_event_attr *attr = &event->attr;
326 	struct hw_perf_event *hwc = &event->hw;
327 	int err;
328 	u64 ev;
329 
330 	switch (attr->type) {
331 	case PERF_TYPE_RAW:
332 		/* Raw events are used to access counters directly,
333 		 * hence do not permit excludes */
334 		if (attr->exclude_kernel || attr->exclude_user ||
335 		    attr->exclude_hv)
336 			return -EOPNOTSUPP;
337 		ev = attr->config;
338 		break;
339 
340 	case PERF_TYPE_HARDWARE:
341 		ev = attr->config;
342 		/* Count user space (problem-state) only */
343 		if (!attr->exclude_user && attr->exclude_kernel) {
344 			if (ev >= ARRAY_SIZE(cpumf_generic_events_user))
345 				return -EOPNOTSUPP;
346 			ev = cpumf_generic_events_user[ev];
347 
348 		/* No support for kernel space counters only */
349 		} else if (!attr->exclude_kernel && attr->exclude_user) {
350 			return -EOPNOTSUPP;
351 
352 		/* Count user and kernel space */
353 		} else {
354 			if (ev >= ARRAY_SIZE(cpumf_generic_events_basic))
355 				return -EOPNOTSUPP;
356 			ev = cpumf_generic_events_basic[ev];
357 		}
358 		break;
359 
360 	default:
361 		return -ENOENT;
362 	}
363 
364 	if (ev == -1)
365 		return -ENOENT;
366 
367 	if (ev >= PERF_CPUM_CF_MAX_CTR)
368 		return -EINVAL;
369 
370 	/* Use the hardware perf event structure to store the counter number
371 	 * in 'config' member and the counter set to which the counter belongs
372 	 * in the 'config_base'.  The counter set (config_base) is then used
373 	 * to enable/disable the counters.
374 	 */
375 	hwc->config = ev;
376 	hwc->config_base = get_counter_set(ev);
377 
378 	/* Validate the counter that is assigned to this event.
379 	 * Because the counter facility can use numerous counters at the
380 	 * same time without constraints, it is not necessary to explicity
381 	 * validate event groups (event->group_leader != event).
382 	 */
383 	err = validate_event(hwc);
384 	if (err)
385 		return err;
386 
387 	/* Initialize for using the CPU-measurement counter facility */
388 	if (!atomic_inc_not_zero(&num_events)) {
389 		mutex_lock(&pmc_reserve_mutex);
390 		if (atomic_read(&num_events) == 0 && reserve_pmc_hardware())
391 			err = -EBUSY;
392 		else
393 			atomic_inc(&num_events);
394 		mutex_unlock(&pmc_reserve_mutex);
395 	}
396 	event->destroy = hw_perf_event_destroy;
397 
398 	/* Finally, validate version and authorization of the counter set */
399 	err = validate_ctr_auth(hwc);
400 	if (!err)
401 		err = validate_ctr_version(hwc);
402 
403 	return err;
404 }
405 
406 static int cpumf_pmu_event_init(struct perf_event *event)
407 {
408 	int err;
409 
410 	switch (event->attr.type) {
411 	case PERF_TYPE_HARDWARE:
412 	case PERF_TYPE_HW_CACHE:
413 	case PERF_TYPE_RAW:
414 		err = __hw_perf_event_init(event);
415 		break;
416 	default:
417 		return -ENOENT;
418 	}
419 
420 	if (unlikely(err) && event->destroy)
421 		event->destroy(event);
422 
423 	return err;
424 }
425 
426 static int hw_perf_event_reset(struct perf_event *event)
427 {
428 	u64 prev, new;
429 	int err;
430 
431 	do {
432 		prev = local64_read(&event->hw.prev_count);
433 		err = ecctr(event->hw.config, &new);
434 		if (err) {
435 			if (err != 3)
436 				break;
437 			/* The counter is not (yet) available. This
438 			 * might happen if the counter set to which
439 			 * this counter belongs is in the disabled
440 			 * state.
441 			 */
442 			new = 0;
443 		}
444 	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
445 
446 	return err;
447 }
448 
449 static int hw_perf_event_update(struct perf_event *event)
450 {
451 	u64 prev, new, delta;
452 	int err;
453 
454 	do {
455 		prev = local64_read(&event->hw.prev_count);
456 		err = ecctr(event->hw.config, &new);
457 		if (err)
458 			goto out;
459 	} while (local64_cmpxchg(&event->hw.prev_count, prev, new) != prev);
460 
461 	delta = (prev <= new) ? new - prev
462 			      : (-1ULL - prev) + new + 1;	 /* overflow */
463 	local64_add(delta, &event->count);
464 out:
465 	return err;
466 }
467 
468 static void cpumf_pmu_read(struct perf_event *event)
469 {
470 	if (event->hw.state & PERF_HES_STOPPED)
471 		return;
472 
473 	hw_perf_event_update(event);
474 }
475 
476 static void cpumf_pmu_start(struct perf_event *event, int flags)
477 {
478 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
479 	struct hw_perf_event *hwc = &event->hw;
480 
481 	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
482 		return;
483 
484 	if (WARN_ON_ONCE(hwc->config == -1))
485 		return;
486 
487 	if (flags & PERF_EF_RELOAD)
488 		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
489 
490 	hwc->state = 0;
491 
492 	/* (Re-)enable and activate the counter set */
493 	ctr_set_enable(&cpuhw->state, hwc->config_base);
494 	ctr_set_start(&cpuhw->state, hwc->config_base);
495 
496 	/* The counter set to which this counter belongs can be already active.
497 	 * Because all counters in a set are active, the event->hw.prev_count
498 	 * needs to be synchronized.  At this point, the counter set can be in
499 	 * the inactive or disabled state.
500 	 */
501 	hw_perf_event_reset(event);
502 
503 	/* increment refcount for this counter set */
504 	atomic_inc(&cpuhw->ctr_set[hwc->config_base]);
505 }
506 
507 static void cpumf_pmu_stop(struct perf_event *event, int flags)
508 {
509 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
510 	struct hw_perf_event *hwc = &event->hw;
511 
512 	if (!(hwc->state & PERF_HES_STOPPED)) {
513 		/* Decrement reference count for this counter set and if this
514 		 * is the last used counter in the set, clear activation
515 		 * control and set the counter set state to inactive.
516 		 */
517 		if (!atomic_dec_return(&cpuhw->ctr_set[hwc->config_base]))
518 			ctr_set_stop(&cpuhw->state, hwc->config_base);
519 		event->hw.state |= PERF_HES_STOPPED;
520 	}
521 
522 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
523 		hw_perf_event_update(event);
524 		event->hw.state |= PERF_HES_UPTODATE;
525 	}
526 }
527 
528 static int cpumf_pmu_add(struct perf_event *event, int flags)
529 {
530 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
531 
532 	/* Check authorization for the counter set to which this
533 	 * counter belongs.
534 	 * For group events transaction, the authorization check is
535 	 * done in cpumf_pmu_commit_txn().
536 	 */
537 	if (!(cpuhw->flags & PERF_EVENT_TXN))
538 		if (validate_ctr_auth(&event->hw))
539 			return -EPERM;
540 
541 	ctr_set_enable(&cpuhw->state, event->hw.config_base);
542 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
543 
544 	if (flags & PERF_EF_START)
545 		cpumf_pmu_start(event, PERF_EF_RELOAD);
546 
547 	perf_event_update_userpage(event);
548 
549 	return 0;
550 }
551 
552 static void cpumf_pmu_del(struct perf_event *event, int flags)
553 {
554 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
555 
556 	cpumf_pmu_stop(event, PERF_EF_UPDATE);
557 
558 	/* Check if any counter in the counter set is still used.  If not used,
559 	 * change the counter set to the disabled state.  This also clears the
560 	 * content of all counters in the set.
561 	 *
562 	 * When a new perf event has been added but not yet started, this can
563 	 * clear enable control and resets all counters in a set.  Therefore,
564 	 * cpumf_pmu_start() always has to reenable a counter set.
565 	 */
566 	if (!atomic_read(&cpuhw->ctr_set[event->hw.config_base]))
567 		ctr_set_disable(&cpuhw->state, event->hw.config_base);
568 
569 	perf_event_update_userpage(event);
570 }
571 
572 /*
573  * Start group events scheduling transaction.
574  * Set flags to perform a single test at commit time.
575  */
576 static void cpumf_pmu_start_txn(struct pmu *pmu)
577 {
578 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
579 
580 	perf_pmu_disable(pmu);
581 	cpuhw->flags |= PERF_EVENT_TXN;
582 	cpuhw->tx_state = cpuhw->state;
583 }
584 
585 /*
586  * Stop and cancel a group events scheduling tranctions.
587  * Assumes cpumf_pmu_del() is called for each successful added
588  * cpumf_pmu_add() during the transaction.
589  */
590 static void cpumf_pmu_cancel_txn(struct pmu *pmu)
591 {
592 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
593 
594 	WARN_ON(cpuhw->tx_state != cpuhw->state);
595 
596 	cpuhw->flags &= ~PERF_EVENT_TXN;
597 	perf_pmu_enable(pmu);
598 }
599 
600 /*
601  * Commit the group events scheduling transaction.  On success, the
602  * transaction is closed.   On error, the transaction is kept open
603  * until cpumf_pmu_cancel_txn() is called.
604  */
605 static int cpumf_pmu_commit_txn(struct pmu *pmu)
606 {
607 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
608 	u64 state;
609 
610 	/* check if the updated state can be scheduled */
611 	state = cpuhw->state & ~((1 << CPUMF_LCCTL_ENABLE_SHIFT) - 1);
612 	state >>= CPUMF_LCCTL_ENABLE_SHIFT;
613 	if ((state & cpuhw->info.auth_ctl) != state)
614 		return -EPERM;
615 
616 	cpuhw->flags &= ~PERF_EVENT_TXN;
617 	perf_pmu_enable(pmu);
618 	return 0;
619 }
620 
621 /* Performance monitoring unit for s390x */
622 static struct pmu cpumf_pmu = {
623 	.pmu_enable   = cpumf_pmu_enable,
624 	.pmu_disable  = cpumf_pmu_disable,
625 	.event_init   = cpumf_pmu_event_init,
626 	.add	      = cpumf_pmu_add,
627 	.del	      = cpumf_pmu_del,
628 	.start	      = cpumf_pmu_start,
629 	.stop	      = cpumf_pmu_stop,
630 	.read	      = cpumf_pmu_read,
631 	.start_txn    = cpumf_pmu_start_txn,
632 	.commit_txn   = cpumf_pmu_commit_txn,
633 	.cancel_txn   = cpumf_pmu_cancel_txn,
634 };
635 
636 static int cpumf_pmu_notifier(struct notifier_block *self, unsigned long action,
637 			      void *hcpu)
638 {
639 	unsigned int cpu = (long) hcpu;
640 	int flags;
641 
642 	switch (action & ~CPU_TASKS_FROZEN) {
643 	case CPU_ONLINE:
644 		flags = PMC_INIT;
645 		smp_call_function_single(cpu, setup_pmc_cpu, &flags, 1);
646 		break;
647 	case CPU_DOWN_PREPARE:
648 		flags = PMC_RELEASE;
649 		smp_call_function_single(cpu, setup_pmc_cpu, &flags, 1);
650 		break;
651 	default:
652 		break;
653 	}
654 
655 	return NOTIFY_OK;
656 }
657 
658 static int __init cpumf_pmu_init(void)
659 {
660 	int rc;
661 
662 	if (!cpum_cf_avail())
663 		return -ENODEV;
664 
665 	/* clear bit 15 of cr0 to unauthorize problem-state to
666 	 * extract measurement counters */
667 	ctl_clear_bit(0, 48);
668 
669 	/* register handler for measurement-alert interruptions */
670 	rc = register_external_irq(EXT_IRQ_MEASURE_ALERT,
671 				   cpumf_measurement_alert);
672 	if (rc) {
673 		pr_err("Registering for CPU-measurement alerts "
674 		       "failed with rc=%i\n", rc);
675 		goto out;
676 	}
677 
678 	/* The CPU measurement counter facility does not have overflow
679 	 * interrupts to do sampling.  Sampling must be provided by
680 	 * external means, for example, by timers.
681 	 */
682 	cpumf_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
683 
684 	cpumf_pmu.attr_groups = cpumf_cf_event_group();
685 	rc = perf_pmu_register(&cpumf_pmu, "cpum_cf", PERF_TYPE_RAW);
686 	if (rc) {
687 		pr_err("Registering the cpum_cf PMU failed with rc=%i\n", rc);
688 		unregister_external_irq(EXT_IRQ_MEASURE_ALERT,
689 					cpumf_measurement_alert);
690 		goto out;
691 	}
692 	perf_cpu_notifier(cpumf_pmu_notifier);
693 out:
694 	return rc;
695 }
696 early_initcall(cpumf_pmu_init);
697