Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0-only
2#undef DEBUG
3
4/*
5 * ARM performance counter support.
6 *
7 * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
8 * Copyright (C) 2010 ARM Ltd., Will Deacon <will.deacon@arm.com>
9 *
10 * This code is based on the sparc64 perf event code, which is in turn based
11 * on the x86 code.
12 */
13#define pr_fmt(fmt) "hw perfevents: " fmt
14
15#include <linux/bitmap.h>
16#include <linux/cpumask.h>
17#include <linux/cpu_pm.h>
18#include <linux/export.h>
19#include <linux/kernel.h>
20#include <linux/perf/arm_pmu.h>
21#include <linux/slab.h>
22#include <linux/sched/clock.h>
23#include <linux/spinlock.h>
24#include <linux/irq.h>
25#include <linux/irqdesc.h>
26
27#include <asm/irq_regs.h>
28
29static DEFINE_PER_CPU(struct arm_pmu *, cpu_armpmu);
30static DEFINE_PER_CPU(int, cpu_irq);
31
32static inline u64 arm_pmu_event_max_period(struct perf_event *event)
33{
34 if (event->hw.flags & ARMPMU_EVT_64BIT)
35 return GENMASK_ULL(63, 0);
36 else
37 return GENMASK_ULL(31, 0);
38}
39
40static int
41armpmu_map_cache_event(const unsigned (*cache_map)
42 [PERF_COUNT_HW_CACHE_MAX]
43 [PERF_COUNT_HW_CACHE_OP_MAX]
44 [PERF_COUNT_HW_CACHE_RESULT_MAX],
45 u64 config)
46{
47 unsigned int cache_type, cache_op, cache_result, ret;
48
49 cache_type = (config >> 0) & 0xff;
50 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
51 return -EINVAL;
52
53 cache_op = (config >> 8) & 0xff;
54 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
55 return -EINVAL;
56
57 cache_result = (config >> 16) & 0xff;
58 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
59 return -EINVAL;
60
61 if (!cache_map)
62 return -ENOENT;
63
64 ret = (int)(*cache_map)[cache_type][cache_op][cache_result];
65
66 if (ret == CACHE_OP_UNSUPPORTED)
67 return -ENOENT;
68
69 return ret;
70}
71
72static int
73armpmu_map_hw_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config)
74{
75 int mapping;
76
77 if (config >= PERF_COUNT_HW_MAX)
78 return -EINVAL;
79
80 if (!event_map)
81 return -ENOENT;
82
83 mapping = (*event_map)[config];
84 return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping;
85}
86
87static int
88armpmu_map_raw_event(u32 raw_event_mask, u64 config)
89{
90 return (int)(config & raw_event_mask);
91}
92
93int
94armpmu_map_event(struct perf_event *event,
95 const unsigned (*event_map)[PERF_COUNT_HW_MAX],
96 const unsigned (*cache_map)
97 [PERF_COUNT_HW_CACHE_MAX]
98 [PERF_COUNT_HW_CACHE_OP_MAX]
99 [PERF_COUNT_HW_CACHE_RESULT_MAX],
100 u32 raw_event_mask)
101{
102 u64 config = event->attr.config;
103 int type = event->attr.type;
104
105 if (type == event->pmu->type)
106 return armpmu_map_raw_event(raw_event_mask, config);
107
108 switch (type) {
109 case PERF_TYPE_HARDWARE:
110 return armpmu_map_hw_event(event_map, config);
111 case PERF_TYPE_HW_CACHE:
112 return armpmu_map_cache_event(cache_map, config);
113 case PERF_TYPE_RAW:
114 return armpmu_map_raw_event(raw_event_mask, config);
115 }
116
117 return -ENOENT;
118}
119
120int armpmu_event_set_period(struct perf_event *event)
121{
122 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
123 struct hw_perf_event *hwc = &event->hw;
124 s64 left = local64_read(&hwc->period_left);
125 s64 period = hwc->sample_period;
126 u64 max_period;
127 int ret = 0;
128
129 max_period = arm_pmu_event_max_period(event);
130 if (unlikely(left <= -period)) {
131 left = period;
132 local64_set(&hwc->period_left, left);
133 hwc->last_period = period;
134 ret = 1;
135 }
136
137 if (unlikely(left <= 0)) {
138 left += period;
139 local64_set(&hwc->period_left, left);
140 hwc->last_period = period;
141 ret = 1;
142 }
143
144 /*
145 * Limit the maximum period to prevent the counter value
146 * from overtaking the one we are about to program. In
147 * effect we are reducing max_period to account for
148 * interrupt latency (and we are being very conservative).
149 */
150 if (left > (max_period >> 1))
151 left = (max_period >> 1);
152
153 local64_set(&hwc->prev_count, (u64)-left);
154
155 armpmu->write_counter(event, (u64)(-left) & max_period);
156
157 perf_event_update_userpage(event);
158
159 return ret;
160}
161
162u64 armpmu_event_update(struct perf_event *event)
163{
164 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
165 struct hw_perf_event *hwc = &event->hw;
166 u64 delta, prev_raw_count, new_raw_count;
167 u64 max_period = arm_pmu_event_max_period(event);
168
169again:
170 prev_raw_count = local64_read(&hwc->prev_count);
171 new_raw_count = armpmu->read_counter(event);
172
173 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
174 new_raw_count) != prev_raw_count)
175 goto again;
176
177 delta = (new_raw_count - prev_raw_count) & max_period;
178
179 local64_add(delta, &event->count);
180 local64_sub(delta, &hwc->period_left);
181
182 return new_raw_count;
183}
184
185static void
186armpmu_read(struct perf_event *event)
187{
188 armpmu_event_update(event);
189}
190
191static void
192armpmu_stop(struct perf_event *event, int flags)
193{
194 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
195 struct hw_perf_event *hwc = &event->hw;
196
197 /*
198 * ARM pmu always has to update the counter, so ignore
199 * PERF_EF_UPDATE, see comments in armpmu_start().
200 */
201 if (!(hwc->state & PERF_HES_STOPPED)) {
202 armpmu->disable(event);
203 armpmu_event_update(event);
204 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
205 }
206}
207
208static void armpmu_start(struct perf_event *event, int flags)
209{
210 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
211 struct hw_perf_event *hwc = &event->hw;
212
213 /*
214 * ARM pmu always has to reprogram the period, so ignore
215 * PERF_EF_RELOAD, see the comment below.
216 */
217 if (flags & PERF_EF_RELOAD)
218 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
219
220 hwc->state = 0;
221 /*
222 * Set the period again. Some counters can't be stopped, so when we
223 * were stopped we simply disabled the IRQ source and the counter
224 * may have been left counting. If we don't do this step then we may
225 * get an interrupt too soon or *way* too late if the overflow has
226 * happened since disabling.
227 */
228 armpmu_event_set_period(event);
229 armpmu->enable(event);
230}
231
232static void
233armpmu_del(struct perf_event *event, int flags)
234{
235 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
236 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
237 struct hw_perf_event *hwc = &event->hw;
238 int idx = hwc->idx;
239
240 armpmu_stop(event, PERF_EF_UPDATE);
241 hw_events->events[idx] = NULL;
242 armpmu->clear_event_idx(hw_events, event);
243 perf_event_update_userpage(event);
244 /* Clear the allocated counter */
245 hwc->idx = -1;
246}
247
248static int
249armpmu_add(struct perf_event *event, int flags)
250{
251 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
252 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
253 struct hw_perf_event *hwc = &event->hw;
254 int idx;
255
256 /* An event following a process won't be stopped earlier */
257 if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
258 return -ENOENT;
259
260 /* If we don't have a space for the counter then finish early. */
261 idx = armpmu->get_event_idx(hw_events, event);
262 if (idx < 0)
263 return idx;
264
265 /*
266 * If there is an event in the counter we are going to use then make
267 * sure it is disabled.
268 */
269 event->hw.idx = idx;
270 armpmu->disable(event);
271 hw_events->events[idx] = event;
272
273 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
274 if (flags & PERF_EF_START)
275 armpmu_start(event, PERF_EF_RELOAD);
276
277 /* Propagate our changes to the userspace mapping. */
278 perf_event_update_userpage(event);
279
280 return 0;
281}
282
283static int
284validate_event(struct pmu *pmu, struct pmu_hw_events *hw_events,
285 struct perf_event *event)
286{
287 struct arm_pmu *armpmu;
288
289 if (is_software_event(event))
290 return 1;
291
292 /*
293 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
294 * core perf code won't check that the pmu->ctx == leader->ctx
295 * until after pmu->event_init(event).
296 */
297 if (event->pmu != pmu)
298 return 0;
299
300 if (event->state < PERF_EVENT_STATE_OFF)
301 return 1;
302
303 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
304 return 1;
305
306 armpmu = to_arm_pmu(event->pmu);
307 return armpmu->get_event_idx(hw_events, event) >= 0;
308}
309
310static int
311validate_group(struct perf_event *event)
312{
313 struct perf_event *sibling, *leader = event->group_leader;
314 struct pmu_hw_events fake_pmu;
315
316 /*
317 * Initialise the fake PMU. We only need to populate the
318 * used_mask for the purposes of validation.
319 */
320 memset(&fake_pmu.used_mask, 0, sizeof(fake_pmu.used_mask));
321
322 if (!validate_event(event->pmu, &fake_pmu, leader))
323 return -EINVAL;
324
325 for_each_sibling_event(sibling, leader) {
326 if (!validate_event(event->pmu, &fake_pmu, sibling))
327 return -EINVAL;
328 }
329
330 if (!validate_event(event->pmu, &fake_pmu, event))
331 return -EINVAL;
332
333 return 0;
334}
335
336static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
337{
338 struct arm_pmu *armpmu;
339 int ret;
340 u64 start_clock, finish_clock;
341
342 /*
343 * we request the IRQ with a (possibly percpu) struct arm_pmu**, but
344 * the handlers expect a struct arm_pmu*. The percpu_irq framework will
345 * do any necessary shifting, we just need to perform the first
346 * dereference.
347 */
348 armpmu = *(void **)dev;
349 if (WARN_ON_ONCE(!armpmu))
350 return IRQ_NONE;
351
352 start_clock = sched_clock();
353 ret = armpmu->handle_irq(armpmu);
354 finish_clock = sched_clock();
355
356 perf_sample_event_took(finish_clock - start_clock);
357 return ret;
358}
359
360static int
361__hw_perf_event_init(struct perf_event *event)
362{
363 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
364 struct hw_perf_event *hwc = &event->hw;
365 int mapping;
366
367 hwc->flags = 0;
368 mapping = armpmu->map_event(event);
369
370 if (mapping < 0) {
371 pr_debug("event %x:%llx not supported\n", event->attr.type,
372 event->attr.config);
373 return mapping;
374 }
375
376 /*
377 * We don't assign an index until we actually place the event onto
378 * hardware. Use -1 to signify that we haven't decided where to put it
379 * yet. For SMP systems, each core has it's own PMU so we can't do any
380 * clever allocation or constraints checking at this point.
381 */
382 hwc->idx = -1;
383 hwc->config_base = 0;
384 hwc->config = 0;
385 hwc->event_base = 0;
386
387 /*
388 * Check whether we need to exclude the counter from certain modes.
389 */
390 if (armpmu->set_event_filter &&
391 armpmu->set_event_filter(hwc, &event->attr)) {
392 pr_debug("ARM performance counters do not support "
393 "mode exclusion\n");
394 return -EOPNOTSUPP;
395 }
396
397 /*
398 * Store the event encoding into the config_base field.
399 */
400 hwc->config_base |= (unsigned long)mapping;
401
402 if (!is_sampling_event(event)) {
403 /*
404 * For non-sampling runs, limit the sample_period to half
405 * of the counter width. That way, the new counter value
406 * is far less likely to overtake the previous one unless
407 * you have some serious IRQ latency issues.
408 */
409 hwc->sample_period = arm_pmu_event_max_period(event) >> 1;
410 hwc->last_period = hwc->sample_period;
411 local64_set(&hwc->period_left, hwc->sample_period);
412 }
413
414 if (event->group_leader != event) {
415 if (validate_group(event) != 0)
416 return -EINVAL;
417 }
418
419 return 0;
420}
421
422static int armpmu_event_init(struct perf_event *event)
423{
424 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
425
426 /*
427 * Reject CPU-affine events for CPUs that are of a different class to
428 * that which this PMU handles. Process-following events (where
429 * event->cpu == -1) can be migrated between CPUs, and thus we have to
430 * reject them later (in armpmu_add) if they're scheduled on a
431 * different class of CPU.
432 */
433 if (event->cpu != -1 &&
434 !cpumask_test_cpu(event->cpu, &armpmu->supported_cpus))
435 return -ENOENT;
436
437 /* does not support taken branch sampling */
438 if (has_branch_stack(event))
439 return -EOPNOTSUPP;
440
441 if (armpmu->map_event(event) == -ENOENT)
442 return -ENOENT;
443
444 return __hw_perf_event_init(event);
445}
446
447static void armpmu_enable(struct pmu *pmu)
448{
449 struct arm_pmu *armpmu = to_arm_pmu(pmu);
450 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
451 int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
452
453 /* For task-bound events we may be called on other CPUs */
454 if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
455 return;
456
457 if (enabled)
458 armpmu->start(armpmu);
459}
460
461static void armpmu_disable(struct pmu *pmu)
462{
463 struct arm_pmu *armpmu = to_arm_pmu(pmu);
464
465 /* For task-bound events we may be called on other CPUs */
466 if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
467 return;
468
469 armpmu->stop(armpmu);
470}
471
472/*
473 * In heterogeneous systems, events are specific to a particular
474 * microarchitecture, and aren't suitable for another. Thus, only match CPUs of
475 * the same microarchitecture.
476 */
477static int armpmu_filter_match(struct perf_event *event)
478{
479 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
480 unsigned int cpu = smp_processor_id();
481 int ret;
482
483 ret = cpumask_test_cpu(cpu, &armpmu->supported_cpus);
484 if (ret && armpmu->filter_match)
485 return armpmu->filter_match(event);
486
487 return ret;
488}
489
490static ssize_t armpmu_cpumask_show(struct device *dev,
491 struct device_attribute *attr, char *buf)
492{
493 struct arm_pmu *armpmu = to_arm_pmu(dev_get_drvdata(dev));
494 return cpumap_print_to_pagebuf(true, buf, &armpmu->supported_cpus);
495}
496
497static DEVICE_ATTR(cpus, S_IRUGO, armpmu_cpumask_show, NULL);
498
499static struct attribute *armpmu_common_attrs[] = {
500 &dev_attr_cpus.attr,
501 NULL,
502};
503
504static struct attribute_group armpmu_common_attr_group = {
505 .attrs = armpmu_common_attrs,
506};
507
508/* Set at runtime when we know what CPU type we are. */
509static struct arm_pmu *__oprofile_cpu_pmu;
510
511/*
512 * Despite the names, these two functions are CPU-specific and are used
513 * by the OProfile/perf code.
514 */
515const char *perf_pmu_name(void)
516{
517 if (!__oprofile_cpu_pmu)
518 return NULL;
519
520 return __oprofile_cpu_pmu->name;
521}
522EXPORT_SYMBOL_GPL(perf_pmu_name);
523
524int perf_num_counters(void)
525{
526 int max_events = 0;
527
528 if (__oprofile_cpu_pmu != NULL)
529 max_events = __oprofile_cpu_pmu->num_events;
530
531 return max_events;
532}
533EXPORT_SYMBOL_GPL(perf_num_counters);
534
535static int armpmu_count_irq_users(const int irq)
536{
537 int cpu, count = 0;
538
539 for_each_possible_cpu(cpu) {
540 if (per_cpu(cpu_irq, cpu) == irq)
541 count++;
542 }
543
544 return count;
545}
546
547void armpmu_free_irq(int irq, int cpu)
548{
549 if (per_cpu(cpu_irq, cpu) == 0)
550 return;
551 if (WARN_ON(irq != per_cpu(cpu_irq, cpu)))
552 return;
553
554 if (!irq_is_percpu_devid(irq))
555 free_irq(irq, per_cpu_ptr(&cpu_armpmu, cpu));
556 else if (armpmu_count_irq_users(irq) == 1)
557 free_percpu_irq(irq, &cpu_armpmu);
558
559 per_cpu(cpu_irq, cpu) = 0;
560}
561
562int armpmu_request_irq(int irq, int cpu)
563{
564 int err = 0;
565 const irq_handler_t handler = armpmu_dispatch_irq;
566 if (!irq)
567 return 0;
568
569 if (!irq_is_percpu_devid(irq)) {
570 unsigned long irq_flags;
571
572 err = irq_force_affinity(irq, cpumask_of(cpu));
573
574 if (err && num_possible_cpus() > 1) {
575 pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
576 irq, cpu);
577 goto err_out;
578 }
579
580 irq_flags = IRQF_PERCPU |
581 IRQF_NOBALANCING |
582 IRQF_NO_THREAD;
583
584 irq_set_status_flags(irq, IRQ_NOAUTOEN);
585 err = request_irq(irq, handler, irq_flags, "arm-pmu",
586 per_cpu_ptr(&cpu_armpmu, cpu));
587 } else if (armpmu_count_irq_users(irq) == 0) {
588 err = request_percpu_irq(irq, handler, "arm-pmu",
589 &cpu_armpmu);
590 }
591
592 if (err)
593 goto err_out;
594
595 per_cpu(cpu_irq, cpu) = irq;
596 return 0;
597
598err_out:
599 pr_err("unable to request IRQ%d for ARM PMU counters\n", irq);
600 return err;
601}
602
603static int armpmu_get_cpu_irq(struct arm_pmu *pmu, int cpu)
604{
605 struct pmu_hw_events __percpu *hw_events = pmu->hw_events;
606 return per_cpu(hw_events->irq, cpu);
607}
608
609/*
610 * PMU hardware loses all context when a CPU goes offline.
611 * When a CPU is hotplugged back in, since some hardware registers are
612 * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
613 * junk values out of them.
614 */
615static int arm_perf_starting_cpu(unsigned int cpu, struct hlist_node *node)
616{
617 struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
618 int irq;
619
620 if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
621 return 0;
622 if (pmu->reset)
623 pmu->reset(pmu);
624
625 per_cpu(cpu_armpmu, cpu) = pmu;
626
627 irq = armpmu_get_cpu_irq(pmu, cpu);
628 if (irq) {
629 if (irq_is_percpu_devid(irq))
630 enable_percpu_irq(irq, IRQ_TYPE_NONE);
631 else
632 enable_irq(irq);
633 }
634
635 return 0;
636}
637
638static int arm_perf_teardown_cpu(unsigned int cpu, struct hlist_node *node)
639{
640 struct arm_pmu *pmu = hlist_entry_safe(node, struct arm_pmu, node);
641 int irq;
642
643 if (!cpumask_test_cpu(cpu, &pmu->supported_cpus))
644 return 0;
645
646 irq = armpmu_get_cpu_irq(pmu, cpu);
647 if (irq) {
648 if (irq_is_percpu_devid(irq))
649 disable_percpu_irq(irq);
650 else
651 disable_irq_nosync(irq);
652 }
653
654 per_cpu(cpu_armpmu, cpu) = NULL;
655
656 return 0;
657}
658
659#ifdef CONFIG_CPU_PM
660static void cpu_pm_pmu_setup(struct arm_pmu *armpmu, unsigned long cmd)
661{
662 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
663 struct perf_event *event;
664 int idx;
665
666 for (idx = 0; idx < armpmu->num_events; idx++) {
667 event = hw_events->events[idx];
668 if (!event)
669 continue;
670
671 switch (cmd) {
672 case CPU_PM_ENTER:
673 /*
674 * Stop and update the counter
675 */
676 armpmu_stop(event, PERF_EF_UPDATE);
677 break;
678 case CPU_PM_EXIT:
679 case CPU_PM_ENTER_FAILED:
680 /*
681 * Restore and enable the counter.
682 * armpmu_start() indirectly calls
683 *
684 * perf_event_update_userpage()
685 *
686 * that requires RCU read locking to be functional,
687 * wrap the call within RCU_NONIDLE to make the
688 * RCU subsystem aware this cpu is not idle from
689 * an RCU perspective for the armpmu_start() call
690 * duration.
691 */
692 RCU_NONIDLE(armpmu_start(event, PERF_EF_RELOAD));
693 break;
694 default:
695 break;
696 }
697 }
698}
699
700static int cpu_pm_pmu_notify(struct notifier_block *b, unsigned long cmd,
701 void *v)
702{
703 struct arm_pmu *armpmu = container_of(b, struct arm_pmu, cpu_pm_nb);
704 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
705 int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
706
707 if (!cpumask_test_cpu(smp_processor_id(), &armpmu->supported_cpus))
708 return NOTIFY_DONE;
709
710 /*
711 * Always reset the PMU registers on power-up even if
712 * there are no events running.
713 */
714 if (cmd == CPU_PM_EXIT && armpmu->reset)
715 armpmu->reset(armpmu);
716
717 if (!enabled)
718 return NOTIFY_OK;
719
720 switch (cmd) {
721 case CPU_PM_ENTER:
722 armpmu->stop(armpmu);
723 cpu_pm_pmu_setup(armpmu, cmd);
724 break;
725 case CPU_PM_EXIT:
726 case CPU_PM_ENTER_FAILED:
727 cpu_pm_pmu_setup(armpmu, cmd);
728 armpmu->start(armpmu);
729 break;
730 default:
731 return NOTIFY_DONE;
732 }
733
734 return NOTIFY_OK;
735}
736
737static int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu)
738{
739 cpu_pmu->cpu_pm_nb.notifier_call = cpu_pm_pmu_notify;
740 return cpu_pm_register_notifier(&cpu_pmu->cpu_pm_nb);
741}
742
743static void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu)
744{
745 cpu_pm_unregister_notifier(&cpu_pmu->cpu_pm_nb);
746}
747#else
748static inline int cpu_pm_pmu_register(struct arm_pmu *cpu_pmu) { return 0; }
749static inline void cpu_pm_pmu_unregister(struct arm_pmu *cpu_pmu) { }
750#endif
751
752static int cpu_pmu_init(struct arm_pmu *cpu_pmu)
753{
754 int err;
755
756 err = cpuhp_state_add_instance(CPUHP_AP_PERF_ARM_STARTING,
757 &cpu_pmu->node);
758 if (err)
759 goto out;
760
761 err = cpu_pm_pmu_register(cpu_pmu);
762 if (err)
763 goto out_unregister;
764
765 return 0;
766
767out_unregister:
768 cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
769 &cpu_pmu->node);
770out:
771 return err;
772}
773
774static void cpu_pmu_destroy(struct arm_pmu *cpu_pmu)
775{
776 cpu_pm_pmu_unregister(cpu_pmu);
777 cpuhp_state_remove_instance_nocalls(CPUHP_AP_PERF_ARM_STARTING,
778 &cpu_pmu->node);
779}
780
781static struct arm_pmu *__armpmu_alloc(gfp_t flags)
782{
783 struct arm_pmu *pmu;
784 int cpu;
785
786 pmu = kzalloc(sizeof(*pmu), flags);
787 if (!pmu) {
788 pr_info("failed to allocate PMU device!\n");
789 goto out;
790 }
791
792 pmu->hw_events = alloc_percpu_gfp(struct pmu_hw_events, flags);
793 if (!pmu->hw_events) {
794 pr_info("failed to allocate per-cpu PMU data.\n");
795 goto out_free_pmu;
796 }
797
798 pmu->pmu = (struct pmu) {
799 .pmu_enable = armpmu_enable,
800 .pmu_disable = armpmu_disable,
801 .event_init = armpmu_event_init,
802 .add = armpmu_add,
803 .del = armpmu_del,
804 .start = armpmu_start,
805 .stop = armpmu_stop,
806 .read = armpmu_read,
807 .filter_match = armpmu_filter_match,
808 .attr_groups = pmu->attr_groups,
809 /*
810 * This is a CPU PMU potentially in a heterogeneous
811 * configuration (e.g. big.LITTLE). This is not an uncore PMU,
812 * and we have taken ctx sharing into account (e.g. with our
813 * pmu::filter_match callback and pmu::event_init group
814 * validation).
815 */
816 .capabilities = PERF_PMU_CAP_HETEROGENEOUS_CPUS,
817 };
818
819 pmu->attr_groups[ARMPMU_ATTR_GROUP_COMMON] =
820 &armpmu_common_attr_group;
821
822 for_each_possible_cpu(cpu) {
823 struct pmu_hw_events *events;
824
825 events = per_cpu_ptr(pmu->hw_events, cpu);
826 raw_spin_lock_init(&events->pmu_lock);
827 events->percpu_pmu = pmu;
828 }
829
830 return pmu;
831
832out_free_pmu:
833 kfree(pmu);
834out:
835 return NULL;
836}
837
838struct arm_pmu *armpmu_alloc(void)
839{
840 return __armpmu_alloc(GFP_KERNEL);
841}
842
843struct arm_pmu *armpmu_alloc_atomic(void)
844{
845 return __armpmu_alloc(GFP_ATOMIC);
846}
847
848
849void armpmu_free(struct arm_pmu *pmu)
850{
851 free_percpu(pmu->hw_events);
852 kfree(pmu);
853}
854
855int armpmu_register(struct arm_pmu *pmu)
856{
857 int ret;
858
859 ret = cpu_pmu_init(pmu);
860 if (ret)
861 return ret;
862
863 if (!pmu->set_event_filter)
864 pmu->pmu.capabilities |= PERF_PMU_CAP_NO_EXCLUDE;
865
866 ret = perf_pmu_register(&pmu->pmu, pmu->name, -1);
867 if (ret)
868 goto out_destroy;
869
870 if (!__oprofile_cpu_pmu)
871 __oprofile_cpu_pmu = pmu;
872
873 pr_info("enabled with %s PMU driver, %d counters available\n",
874 pmu->name, pmu->num_events);
875
876 return 0;
877
878out_destroy:
879 cpu_pmu_destroy(pmu);
880 return ret;
881}
882
883static int arm_pmu_hp_init(void)
884{
885 int ret;
886
887 ret = cpuhp_setup_state_multi(CPUHP_AP_PERF_ARM_STARTING,
888 "perf/arm/pmu:starting",
889 arm_perf_starting_cpu,
890 arm_perf_teardown_cpu);
891 if (ret)
892 pr_err("CPU hotplug notifier for ARM PMU could not be registered: %d\n",
893 ret);
894 return ret;
895}
896subsys_initcall(arm_pmu_hp_init);