Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
Note: File does not exist in v3.1.
  1#undef DEBUG
  2
  3/*
  4 * ARM performance counter support.
  5 *
  6 * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
  7 * Copyright (C) 2010 ARM Ltd., Will Deacon <will.deacon@arm.com>
  8 *
  9 * This code is based on the sparc64 perf event code, which is in turn based
 10 * on the x86 code.
 11 */
 12#define pr_fmt(fmt) "hw perfevents: " fmt
 13
 14#include <linux/bitmap.h>
 15#include <linux/cpumask.h>
 16#include <linux/cpu_pm.h>
 17#include <linux/export.h>
 18#include <linux/kernel.h>
 19#include <linux/perf/arm_pmu.h>
 20#include <linux/slab.h>
 21#include <linux/sched/clock.h>
 22#include <linux/spinlock.h>
 23#include <linux/irq.h>
 24#include <linux/irqdesc.h>
 25
 26#include <asm/irq_regs.h>
 27
 28static DEFINE_PER_CPU(struct arm_pmu *, cpu_armpmu);
 29static DEFINE_PER_CPU(int, cpu_irq);
 30
 31static int
 32armpmu_map_cache_event(const unsigned (*cache_map)
 33				      [PERF_COUNT_HW_CACHE_MAX]
 34				      [PERF_COUNT_HW_CACHE_OP_MAX]
 35				      [PERF_COUNT_HW_CACHE_RESULT_MAX],
 36		       u64 config)
 37{
 38	unsigned int cache_type, cache_op, cache_result, ret;
 39
 40	cache_type = (config >>  0) & 0xff;
 41	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
 42		return -EINVAL;
 43
 44	cache_op = (config >>  8) & 0xff;
 45	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
 46		return -EINVAL;
 47
 48	cache_result = (config >> 16) & 0xff;
 49	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 50		return -EINVAL;
 51
 52	if (!cache_map)
 53		return -ENOENT;
 54
 55	ret = (int)(*cache_map)[cache_type][cache_op][cache_result];
 56
 57	if (ret == CACHE_OP_UNSUPPORTED)
 58		return -ENOENT;
 59
 60	return ret;
 61}
 62
 63static int
 64armpmu_map_hw_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config)
 65{
 66	int mapping;
 67
 68	if (config >= PERF_COUNT_HW_MAX)
 69		return -EINVAL;
 70
 71	if (!event_map)
 72		return -ENOENT;
 73
 74	mapping = (*event_map)[config];
 75	return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping;
 76}
 77
 78static int
 79armpmu_map_raw_event(u32 raw_event_mask, u64 config)
 80{
 81	return (int)(config & raw_event_mask);
 82}
 83
 84int
 85armpmu_map_event(struct perf_event *event,
 86		 const unsigned (*event_map)[PERF_COUNT_HW_MAX],
 87		 const unsigned (*cache_map)
 88				[PERF_COUNT_HW_CACHE_MAX]
 89				[PERF_COUNT_HW_CACHE_OP_MAX]
 90				[PERF_COUNT_HW_CACHE_RESULT_MAX],
 91		 u32 raw_event_mask)
 92{
 93	u64 config = event->attr.config;
 94	int type = event->attr.type;
 95
 96	if (type == event->pmu->type)
 97		return armpmu_map_raw_event(raw_event_mask, config);
 98
 99	switch (type) {
100	case PERF_TYPE_HARDWARE:
101		return armpmu_map_hw_event(event_map, config);
102	case PERF_TYPE_HW_CACHE:
103		return armpmu_map_cache_event(cache_map, config);
104	case PERF_TYPE_RAW:
105		return armpmu_map_raw_event(raw_event_mask, config);
106	}
107
108	return -ENOENT;
109}
110
111int armpmu_event_set_period(struct perf_event *event)
112{
113	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
114	struct hw_perf_event *hwc = &event->hw;
115	s64 left = local64_read(&hwc->period_left);
116	s64 period = hwc->sample_period;
117	int ret = 0;
118
119	if (unlikely(left <= -period)) {
120		left = period;
121		local64_set(&hwc->period_left, left);
122		hwc->last_period = period;
123		ret = 1;
124	}
125
126	if (unlikely(left <= 0)) {
127		left += period;
128		local64_set(&hwc->period_left, left);
129		hwc->last_period = period;
130		ret = 1;
131	}
132
133	/*
134	 * Limit the maximum period to prevent the counter value
135	 * from overtaking the one we are about to program. In
136	 * effect we are reducing max_period to account for
137	 * interrupt latency (and we are being very conservative).
138	 */
139	if (left > (armpmu->max_period >> 1))
140		left = armpmu->max_period >> 1;
141
142	local64_set(&hwc->prev_count, (u64)-left);
143
144	armpmu->write_counter(event, (u64)(-left) & 0xffffffff);
145
146	perf_event_update_userpage(event);
147
148	return ret;
149}
150
151u64 armpmu_event_update(struct perf_event *event)
152{
153	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
154	struct hw_perf_event *hwc = &event->hw;
155	u64 delta, prev_raw_count, new_raw_count;
156
157again:
158	prev_raw_count = local64_read(&hwc->prev_count);
159	new_raw_count = armpmu->read_counter(event);
160
161	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
162			     new_raw_count) != prev_raw_count)
163		goto again;
164
165	delta = (new_raw_count - prev_raw_count) & armpmu->max_period;
166
167	local64_add(delta, &event->count);
168	local64_sub(delta, &hwc->period_left);
169
170	return new_raw_count;
171}
172
173static void
174armpmu_read(struct perf_event *event)
175{
176	armpmu_event_update(event);
177}
178
179static void
180armpmu_stop(struct perf_event *event, int flags)
181{
182	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
183	struct hw_perf_event *hwc = &event->hw;
184
185	/*
186	 * ARM pmu always has to update the counter, so ignore
187	 * PERF_EF_UPDATE, see comments in armpmu_start().
188	 */
189	if (!(hwc->state & PERF_HES_STOPPED)) {
190		armpmu->disable(event);
191		armpmu_event_update(event);
192		hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
193	}
194}
195
196static void armpmu_start(struct perf_event *event, int flags)
197{
198	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
199	struct hw_perf_event *hwc = &event->hw;
200
201	/*
202	 * ARM pmu always has to reprogram the period, so ignore
203	 * PERF_EF_RELOAD, see the comment below.
204	 */
205	if (flags & PERF_EF_RELOAD)
206		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
207
208	hwc->state = 0;
209	/*
210	 * Set the period again. Some counters can't be stopped, so when we
211	 * were stopped we simply disabled the IRQ source and the counter
212	 * may have been left counting. If we don't do this step then we may
213	 * get an interrupt too soon or *way* too late if the overflow has
214	 * happened since disabling.
215	 */
216	armpmu_event_set_period(event);
217	armpmu->enable(event);
218}
219
220static void
221armpmu_del(struct perf_event *event, int flags)
222{
223	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
224	struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
225	struct hw_perf_event *hwc = &event->hw;
226	int idx = hwc->idx;
227
228	armpmu_stop(event, PERF_EF_UPDATE);
229	hw_events->events[idx] = NULL;
230	clear_bit(idx, hw_events->used_mask);
231	if (armpmu->clear_event_idx)
232		armpmu->clear_event_idx(hw_events, event);
233
234	perf_event_update_userpage(event);
235}
236
237static int
238armpmu_add(struct perf_event *event, int flags)
239{
240	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
241	struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
242	struct hw_perf_event *hwc = &event->hw;
243	int idx;
244
245	/* An event following a process won't be stopped earlier */
246	if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
247		return -ENOENT;
248
249	/* If we don't have a space for the counter then finish early. */
250	idx = armpmu->get_event_idx(hw_events, event);
251	if (idx < 0)
252		return idx;
253
254	/*
255	 * If there is an event in the counter we are going to use then make
256	 * sure it is disabled.
257	 */
258	event->hw.idx = idx;
259	armpmu->disable(event);
260	hw_events->events[idx] = event;
261
262	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
263	if (flags & PERF_EF_START)
264		armpmu_start(event, PERF_EF_RELOAD);
265
266	/* Propagate our changes to the userspace mapping. */
267	perf_event_update_userpage(event);
268
269	return 0;
270}
271
272static int
273validate_event(struct pmu *pmu, struct pmu_hw_events *hw_events,
274			       struct perf_event *event)
275{
276	struct arm_pmu *armpmu;
277
278	if (is_software_event(event))
279		return 1;
280
281	/*
282	 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
283	 * core perf code won't check that the pmu->ctx == leader->ctx
284	 * until after pmu->event_init(event).
285	 */
286	if (event->pmu != pmu)
287		return 0;
288
289	if (event->state < PERF_EVENT_STATE_OFF)
290		return 1;
291
292	if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
293		return 1;
294
295	armpmu = to_arm_pmu(event->pmu);
296	return armpmu->get_event_idx(hw_events, event) >= 0;
297}
298
299static int
300validate_group(struct perf_event *event)
301{
302	struct perf_event *sibling, *leader = event->group_leader;
303	struct pmu_hw_events fake_pmu;
304
305	/*
306	 * Initialise the fake PMU. We only need to populate the
307	 * used_mask for the purposes of validation.
308	 */
309	memset(&fake_pmu.used_mask, 0, sizeof(fake_pmu.used_mask));
310
311	if (!validate_event(event->pmu, &fake_pmu, leader))
312		return -EINVAL;
313
314	for_each_sibling_event(sibling, leader) {
315		if (!validate_event(event->pmu, &fake_pmu, sibling))
316			return -EINVAL;
317	}
318
319	if (!validate_event(event->pmu, &fake_pmu, event))
320		return -EINVAL;
321
322	return 0;
323}
324
325static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
326{
327	struct arm_pmu *armpmu;
328	int ret;
329	u64 start_clock, finish_clock;
330
331	/*
332	 * we request the IRQ with a (possibly percpu) struct arm_pmu**, but
333	 * the handlers expect a struct arm_pmu*. The percpu_irq framework will
334	 * do any necessary shifting, we just need to perform the first
335	 * dereference.
336	 */
337	armpmu = *(void **)dev;
338	if (WARN_ON_ONCE(!armpmu))
339		return IRQ_NONE;
340
341	start_clock = sched_clock();
342	ret = armpmu->handle_irq(irq, armpmu);
343	finish_clock = sched_clock();
344
345	perf_sample_event_took(finish_clock - start_clock);
346	return ret;
347}
348
349static int
350event_requires_mode_exclusion(struct perf_event_attr *attr)
351{
352	return attr->exclude_idle || attr->exclude_user ||
353	       attr->exclude_kernel || attr->exclude_hv;
354}
355
356static int
357__hw_perf_event_init(struct perf_event *event)
358{
359	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
360	struct hw_perf_event *hwc = &event->hw;
361	int mapping;
362
363	mapping = armpmu->map_event(event);
364
365	if (mapping < 0) {
366		pr_debug("event %x:%llx not supported\n", event->attr.type,
367			 event->attr.config);
368		return mapping;
369	}
370
371	/*
372	 * We don't assign an index until we actually place the event onto
373	 * hardware. Use -1 to signify that we haven't decided where to put it
374	 * yet. For SMP systems, each core has it's own PMU so we can't do any
375	 * clever allocation or constraints checking at this point.
376	 */
377	hwc->idx		= -1;
378	hwc->config_base	= 0;
379	hwc->config		= 0;
380	hwc->event_base		= 0;
381
382	/*
383	 * Check whether we need to exclude the counter from certain modes.
384	 */
385	if ((!armpmu->set_event_filter ||
386	     armpmu->set_event_filter(hwc, &event->attr)) &&
387	     event_requires_mode_exclusion(&event->attr)) {
388		pr_debug("ARM performance counters do not support "
389			 "mode exclusion\n");
390		return -EOPNOTSUPP;
391	}
392
393	/*
394	 * Store the event encoding into the config_base field.
395	 */
396	hwc->config_base	    |= (unsigned long)mapping;
397
398	if (!is_sampling_event(event)) {
399		/*
400		 * For non-sampling runs, limit the sample_period to half
401		 * of the counter width. That way, the new counter value
402		 * is far less likely to overtake the previous one unless
403		 * you have some serious IRQ latency issues.
404		 */
405		hwc->sample_period  = armpmu->max_period >> 1;
406		hwc->last_period    = hwc->sample_period;
407		local64_set(&hwc->period_left, hwc->sample_period);
408	}
409
410	if (event->group_leader != event) {
411		if (validate_group(event) != 0)
412			return -EINVAL;
413	}
414
415	return 0;
416}
417
418static int armpmu_event_init(struct perf_event *event)
419{
420	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
421
422	/*
423	 * Reject CPU-affine events for CPUs that are of a different class to
424	 * that which this PMU handles. Process-following events (where
425	 * event->cpu == -1) can be migrated between CPUs, and thus we have to
426	 * reject them later (in armpmu_add) if they're scheduled on a
427	 * different class of CPU.
428	 */
429	if (event->cpu != -1 &&
430		!cpumask_test_cpu(event->cpu, &armpmu->supported_cpus))
431		return -ENOENT;
432
433	/* does not support taken branch sampling */
434	if (has_branch_stack(event))
435		return -EOPNOTSUPP;
436
437	if (armpmu->map_event(event) == -ENOENT)
438		return -ENOENT;
439
440	return __hw_perf_event_init(event);
441}
442
443static void armpmu_enable(struct pmu *pmu)
444{
445	struct arm_pmu *armpmu = to_arm_pmu(pmu);
446	struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
447	int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
448
449	/* For task-bound events we may be called on other CPUs */
450	if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
451		return;
452
453	if (enabled)
454		armpmu->start(armpmu);
455}
456
457static void armpmu_disable(struct pmu *pmu)
458{
459	struct arm_pmu *armpmu = to_arm_pmu(pmu);
460
461	/* For task-bound events we may be called on other CPUs */
462	if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
463		return;
464
465	armpmu->stop(armpmu);
466}
467
468/*
469 * In heterogeneous systems, events are specific to a particular
470 * microarchitecture, and aren't suitable for another. Thus, only match CPUs of
471 * the same microarchitecture.
472 */
473static int armpmu_filter_match(struct perf_event *event)
474{
475	struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
476	unsigned int cpu = smp_processor_id();
477	return cpumask_test_cpu(cpu, &armpmu->supported_cpus);
478}
479
480static ssize_t armpmu_cpumask_show(struct device *dev,
481				   struct device_attribute *attr, char *buf)
482{
483	struct arm_pmu *armpmu = to_arm_pmu(dev_get_drvdata(dev));
484	return cpumap_print_to_pagebuf(true, buf, &armpmu->supported_cpus);
485}
486
487static DEVICE_ATTR(cpus, S_IRUGO, armpmu_cpumask_show, NULL);
488
489static struct attribute *armpmu_common_attrs[] = {
490	&dev_attr_cpus.attr,
491	NULL,
492};
493
494static struct attribute_group armpmu_common_attr_group = {
495	.attrs = armpmu_common_attrs,
496};
497
498/* Set at runtime when we know what CPU type we are. */
499static struct arm_pmu *__oprofile_cpu_pmu;
500
501/*
502 * Despite the names, these two functions are CPU-specific and are used
503 * by the OProfile/perf code.
504 */
505const char *perf_pmu_name(void)
506{
507	if (!__oprofile_cpu_pmu)
508		return NULL;
509
510	return __oprofile_cpu_pmu->name;
511}
512EXPORT_SYMBOL_GPL(perf_pmu_name);
513
514int perf_num_counters(void)
515{
516	int max_events = 0;
517
518	if (__oprofile_cpu_pmu != NULL)
519		max_events = __oprofile_cpu_pmu->num_events;
520
521	return max_events;
522}
523EXPORT_SYMBOL_GPL(perf_num_counters);
524
525static int armpmu_count_irq_users(const int irq)
526{
527	int cpu, count = 0;
528
529	for_each_possible_cpu(cpu) {
530		if (per_cpu(cpu_irq, cpu) == irq)
531			count++;
532	}
533
534	return count;
535}
536
537void armpmu_free_irq(int irq, int cpu)
538{
539	if (per_cpu(cpu_irq, cpu) == 0)
540		return;
541	if (WARN_ON(irq != per_cpu(cpu_irq, cpu)))
542		return;
543
544	if (!irq_is_percpu_devid(irq))
545		free_irq(irq, per_cpu_ptr(&cpu_armpmu, cpu));
546	else if (armpmu_count_irq_users(irq) == 1)
547		free_percpu_irq(irq, &cpu_armpmu);
548
549	per_cpu(cpu_irq, cpu) = 0;
550}
551
552int armpmu_request_irq(int irq, int cpu)
553{
554	int err = 0;
555	const irq_handler_t handler = armpmu_dispatch_irq;
556	if (!irq)
557		return 0;
558
559	if (!irq_is_percpu_devid(irq)) {
560		unsigned long irq_flags;
561
562		err = irq_force_affinity(irq, cpumask_of(cpu));
563
564		if (err && num_possible_cpus() > 1) {
565			pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
566				irq, cpu);
567			goto err_out;
568		}
569
570		irq_flags = IRQF_PERCPU |
571			    IRQF_NOBALANCING |
572			    IRQF_NO_THREAD;
573
574		irq_set_status_flags(irq, IRQ_NOAUTOEN);
575		err = request_irq(irq, handler, irq_flags, "arm-pmu",
576				  per_cpu_ptr(&cpu_armpmu, cpu));
577	} else if (armpmu_count_irq_users(irq) == 0) {
578		err = request_percpu_irq(irq, handler, "arm-pmu",
579					 &cpu_armpmu);
580	}
581
582	if (err)
583		goto err_out;
584
585	per_cpu(cpu_irq, cpu) = irq;
586	return 0;
587
588err_out:
589	pr_err("unable to request IRQ%d for ARM PMU counters\n", irq);
590	return err;
591}
592
593static int armpmu_get_cpu_irq(struct arm_pmu *pmu, int cpu)
594{
595	struct pmu_hw_events __percpu *hw_events = pmu->hw_events;
596	return per_cpu(hw_events->irq, cpu);
597}
598
599/*
600 * PMU hardware loses all context when a CPU goes offline.
601 * When a CPU is hotplugged back in, since some hardware registers are
602 * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
603 * junk values out of them.
604 */
605static int arm_perf_starting_cpu(unsigned int cpu, struct hlist_node *node)
606{
607	struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
608	int irq;
609
610	if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
611		return 0;
612	if (pmu->reset)
613		pmu->reset(pmu);
614
615	per_cpu(cpu_armpmu, cpu) = pmu;
616
617	irq = armpmu_get_cpu_irq(pmu, cpu);
618	if (irq) {
619		if (irq_is_percpu_devid(irq))
620			enable_percpu_irq(irq, IRQ_TYPE_NONE);
621		else
622			enable_irq(irq);
623	}
624
625	return 0;
626}
627
628static int arm_perf_teardown_cpu(unsigned int cpu, struct hlist_node *node)
629{
630	struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
631	int irq;
632
633	if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
634		return 0;
635
636	irq = armpmu_get_cpu_irq(pmu, cpu);
637	if (irq) {
638		if (irq_is_percpu_devid(irq))
639			disable_percpu_irq(irq);
640		else
641			disable_irq_nosync(irq);
642	}
643
644	per_cpu(cpu_armpmu, cpu) = NULL;
645
646	return 0;
647}
648
649#ifdef CONFIG_CPU_PM
650static void cpu_pm_pmu_setup(struct arm_pmu *armpmu, unsigned long cmd)
651{
652	struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
653	struct perf_event *event;
654	int idx;
655
656	for (idx = 0; idx < armpmu->num_events; idx++) {
657		/*
658		 * If the counter is not used skip it, there is no
659		 * need of stopping/restarting it.
660		 */
661		if (!test_bit(idx, hw_events->used_mask))
662			continue;
663
664		event = hw_events->events[idx];
665
666		switch (cmd) {
667		case CPU_PM_ENTER:
668			/*
669			 * Stop and update the counter
670			 */
671			armpmu_stop(event, PERF_EF_UPDATE);
672			break;
673		case CPU_PM_EXIT:
674		case CPU_PM_ENTER_FAILED:
675			 /*
676			  * Restore and enable the counter.
677			  * armpmu_start() indirectly calls
678			  *
679			  * perf_event_update_userpage()
680			  *
681			  * that requires RCU read locking to be functional,
682			  * wrap the call within RCU_NONIDLE to make the
683			  * RCU subsystem aware this cpu is not idle from
684			  * an RCU perspective for the armpmu_start() call
685			  * duration.
686			  */
687			RCU_NONIDLE(armpmu_start(event, PERF_EF_RELOAD));
688			break;
689		default:
690			break;
691		}
692	}
693}
694
695static int cpu_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
696			     void *v)
697{
698	struct arm_pmu *armpmu = container_of(b, struct arm_pmu, cpu_pm_nb);
699	struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
700	int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
701
702	if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
703		return NOTIFY_DONE;
704
705	/*
706	 * Always reset the PMU registers on power-up even if
707	 * there are no events running.
708	 */
709	if (cmd == CPU_PM_EXIT && armpmu->reset)
710		armpmu->reset(armpmu);
711
712	if (!enabled)
713		return NOTIFY_OK;
714
715	switch (cmd) {
716	case CPU_PM_ENTER:
717		armpmu->stop(armpmu);
718		cpu_pm_pmu_setup(armpmu, cmd);
719		break;
720	case CPU_PM_EXIT:
721		cpu_pm_pmu_setup(armpmu, cmd);
722	case CPU_PM_ENTER_FAILED:
723		armpmu->start(armpmu);
724		break;
725	default:
726		return NOTIFY_DONE;
727	}
728
729	return NOTIFY_OK;
730}
731
732static int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu)
733{
734	cpu_pmu->cpu_pm_nb.notifier_call = cpu_pm_pmu_notify;
735	return cpu_pm_register_notifier(&cpu_pmu->cpu_pm_nb);
736}
737
738static void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu)
739{
740	cpu_pm_unregister_notifier(&cpu_pmu->cpu_pm_nb);
741}
742#else
743static inline int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu) { return 0; }
744static inline void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu) { }
745#endif
746
747static int cpu_pmu_init(struct arm_pmu *cpu_pmu)
748{
749	int err;
750
751	err = cpuhp_state_add_instance(CPUHP_AP_PERF_ARM_STARTING,
752				       &cpu_pmu->node);
753	if (err)
754		goto out;
755
756	err = cpu_pm_pmu_register(cpu_pmu);
757	if (err)
758		goto out_unregister;
759
760	return 0;
761
762out_unregister:
763	cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
764					    &cpu_pmu->node);
765out:
766	return err;
767}
768
769static void cpu_pmu_destroy(struct arm_pmu *cpu_pmu)
770{
771	cpu_pm_pmu_unregister(cpu_pmu);
772	cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
773					    &cpu_pmu->node);
774}
775
776static struct arm_pmu *__armpmu_alloc(gfp_t flags)
777{
778	struct arm_pmu *pmu;
779	int cpu;
780
781	pmu = kzalloc(sizeof(*pmu), flags);
782	if (!pmu) {
783		pr_info("failed to allocate PMU device!\n");
784		goto out;
785	}
786
787	pmu->hw_events = alloc_percpu_gfp(struct pmu_hw_events, flags);
788	if (!pmu->hw_events) {
789		pr_info("failed to allocate per-cpu PMU data.\n");
790		goto out_free_pmu;
791	}
792
793	pmu->pmu = (struct pmu) {
794		.pmu_enable	= armpmu_enable,
795		.pmu_disable	= armpmu_disable,
796		.event_init	= armpmu_event_init,
797		.add		= armpmu_add,
798		.del		= armpmu_del,
799		.start		= armpmu_start,
800		.stop		= armpmu_stop,
801		.read		= armpmu_read,
802		.filter_match	= armpmu_filter_match,
803		.attr_groups	= pmu->attr_groups,
804		/*
805		 * This is a CPU PMU potentially in a heterogeneous
806		 * configuration (e.g. big.LITTLE). This is not an uncore PMU,
807		 * and we have taken ctx sharing into account (e.g. with our
808		 * pmu::filter_match callback and pmu::event_init group
809		 * validation).
810		 */
811		.capabilities	= PERF_PMU_CAP_HETEROGENEOUS_CPUS,
812	};
813
814	pmu->attr_groups[ARMPMU_ATTR_GROUP_COMMON] =
815		&armpmu_common_attr_group;
816
817	for_each_possible_cpu(cpu) {
818		struct pmu_hw_events *events;
819
820		events = per_cpu_ptr(pmu->hw_events, cpu);
821		raw_spin_lock_init(&events->pmu_lock);
822		events->percpu_pmu = pmu;
823	}
824
825	return pmu;
826
827out_free_pmu:
828	kfree(pmu);
829out:
830	return NULL;
831}
832
833struct arm_pmu *armpmu_alloc(void)
834{
835	return __armpmu_alloc(GFP_KERNEL);
836}
837
838struct arm_pmu *armpmu_alloc_atomic(void)
839{
840	return __armpmu_alloc(GFP_ATOMIC);
841}
842
843
844void armpmu_free(struct arm_pmu *pmu)
845{
846	free_percpu(pmu->hw_events);
847	kfree(pmu);
848}
849
850int armpmu_register(struct arm_pmu *pmu)
851{
852	int ret;
853
854	ret = cpu_pmu_init(pmu);
855	if (ret)
856		return ret;
857
858	ret = perf_pmu_register(&pmu->pmu, pmu->name, -1);
859	if (ret)
860		goto out_destroy;
861
862	if (!__oprofile_cpu_pmu)
863		__oprofile_cpu_pmu = pmu;
864
865	pr_info("enabled with %s PMU driver, %d counters available\n",
866		pmu->name, pmu->num_events);
867
868	return 0;
869
870out_destroy:
871	cpu_pmu_destroy(pmu);
872	return ret;
873}
874
875static int arm_pmu_hp_init(void)
876{
877	int ret;
878
879	ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_STARTING,
880				      "perf/arm/pmu:starting",
881				      arm_perf_starting_cpu,
882				      arm_perf_teardown_cpu);
883	if (ret)
884		pr_err("CPU hotplug notifier for ARM PMU could not be registered: %d\n",
885		       ret);
886	return ret;
887}
888subsys_initcall(arm_pmu_hp_init);