Loading...
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. Callchain code is based on the ARM OProfile backtrace
11 * code.
12 */
13#define pr_fmt(fmt) "hw perfevents: " fmt
14
15#include <linux/interrupt.h>
16#include <linux/kernel.h>
17#include <linux/module.h>
18#include <linux/perf_event.h>
19#include <linux/platform_device.h>
20#include <linux/spinlock.h>
21#include <linux/uaccess.h>
22
23#include <asm/cputype.h>
24#include <asm/irq.h>
25#include <asm/irq_regs.h>
26#include <asm/pmu.h>
27#include <asm/stacktrace.h>
28
29static struct platform_device *pmu_device;
30
31/*
32 * Hardware lock to serialize accesses to PMU registers. Needed for the
33 * read/modify/write sequences.
34 */
35static DEFINE_RAW_SPINLOCK(pmu_lock);
36
37/*
38 * ARMv6 supports a maximum of 3 events, starting from index 1. If we add
39 * another platform that supports more, we need to increase this to be the
40 * largest of all platforms.
41 *
42 * ARMv7 supports up to 32 events:
43 * cycle counter CCNT + 31 events counters CNT0..30.
44 * Cortex-A8 has 1+4 counters, Cortex-A9 has 1+6 counters.
45 */
46#define ARMPMU_MAX_HWEVENTS 33
47
48/* The events for a given CPU. */
49struct cpu_hw_events {
50 /*
51 * The events that are active on the CPU for the given index. Index 0
52 * is reserved.
53 */
54 struct perf_event *events[ARMPMU_MAX_HWEVENTS];
55
56 /*
57 * A 1 bit for an index indicates that the counter is being used for
58 * an event. A 0 means that the counter can be used.
59 */
60 unsigned long used_mask[BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)];
61
62 /*
63 * A 1 bit for an index indicates that the counter is actively being
64 * used.
65 */
66 unsigned long active_mask[BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)];
67};
68static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
69
70struct arm_pmu {
71 enum arm_perf_pmu_ids id;
72 const char *name;
73 irqreturn_t (*handle_irq)(int irq_num, void *dev);
74 void (*enable)(struct hw_perf_event *evt, int idx);
75 void (*disable)(struct hw_perf_event *evt, int idx);
76 int (*get_event_idx)(struct cpu_hw_events *cpuc,
77 struct hw_perf_event *hwc);
78 u32 (*read_counter)(int idx);
79 void (*write_counter)(int idx, u32 val);
80 void (*start)(void);
81 void (*stop)(void);
82 void (*reset)(void *);
83 const unsigned (*cache_map)[PERF_COUNT_HW_CACHE_MAX]
84 [PERF_COUNT_HW_CACHE_OP_MAX]
85 [PERF_COUNT_HW_CACHE_RESULT_MAX];
86 const unsigned (*event_map)[PERF_COUNT_HW_MAX];
87 u32 raw_event_mask;
88 int num_events;
89 u64 max_period;
90};
91
92/* Set at runtime when we know what CPU type we are. */
93static const struct arm_pmu *armpmu;
94
95enum arm_perf_pmu_ids
96armpmu_get_pmu_id(void)
97{
98 int id = -ENODEV;
99
100 if (armpmu != NULL)
101 id = armpmu->id;
102
103 return id;
104}
105EXPORT_SYMBOL_GPL(armpmu_get_pmu_id);
106
107int
108armpmu_get_max_events(void)
109{
110 int max_events = 0;
111
112 if (armpmu != NULL)
113 max_events = armpmu->num_events;
114
115 return max_events;
116}
117EXPORT_SYMBOL_GPL(armpmu_get_max_events);
118
119int perf_num_counters(void)
120{
121 return armpmu_get_max_events();
122}
123EXPORT_SYMBOL_GPL(perf_num_counters);
124
125#define HW_OP_UNSUPPORTED 0xFFFF
126
127#define C(_x) \
128 PERF_COUNT_HW_CACHE_##_x
129
130#define CACHE_OP_UNSUPPORTED 0xFFFF
131
132static int
133armpmu_map_cache_event(u64 config)
134{
135 unsigned int cache_type, cache_op, cache_result, ret;
136
137 cache_type = (config >> 0) & 0xff;
138 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
139 return -EINVAL;
140
141 cache_op = (config >> 8) & 0xff;
142 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
143 return -EINVAL;
144
145 cache_result = (config >> 16) & 0xff;
146 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
147 return -EINVAL;
148
149 ret = (int)(*armpmu->cache_map)[cache_type][cache_op][cache_result];
150
151 if (ret == CACHE_OP_UNSUPPORTED)
152 return -ENOENT;
153
154 return ret;
155}
156
157static int
158armpmu_map_event(u64 config)
159{
160 int mapping = (*armpmu->event_map)[config];
161 return mapping == HW_OP_UNSUPPORTED ? -EOPNOTSUPP : mapping;
162}
163
164static int
165armpmu_map_raw_event(u64 config)
166{
167 return (int)(config & armpmu->raw_event_mask);
168}
169
170static int
171armpmu_event_set_period(struct perf_event *event,
172 struct hw_perf_event *hwc,
173 int idx)
174{
175 s64 left = local64_read(&hwc->period_left);
176 s64 period = hwc->sample_period;
177 int ret = 0;
178
179 if (unlikely(left <= -period)) {
180 left = period;
181 local64_set(&hwc->period_left, left);
182 hwc->last_period = period;
183 ret = 1;
184 }
185
186 if (unlikely(left <= 0)) {
187 left += period;
188 local64_set(&hwc->period_left, left);
189 hwc->last_period = period;
190 ret = 1;
191 }
192
193 if (left > (s64)armpmu->max_period)
194 left = armpmu->max_period;
195
196 local64_set(&hwc->prev_count, (u64)-left);
197
198 armpmu->write_counter(idx, (u64)(-left) & 0xffffffff);
199
200 perf_event_update_userpage(event);
201
202 return ret;
203}
204
205static u64
206armpmu_event_update(struct perf_event *event,
207 struct hw_perf_event *hwc,
208 int idx, int overflow)
209{
210 u64 delta, prev_raw_count, new_raw_count;
211
212again:
213 prev_raw_count = local64_read(&hwc->prev_count);
214 new_raw_count = armpmu->read_counter(idx);
215
216 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
217 new_raw_count) != prev_raw_count)
218 goto again;
219
220 new_raw_count &= armpmu->max_period;
221 prev_raw_count &= armpmu->max_period;
222
223 if (overflow)
224 delta = armpmu->max_period - prev_raw_count + new_raw_count + 1;
225 else
226 delta = new_raw_count - prev_raw_count;
227
228 local64_add(delta, &event->count);
229 local64_sub(delta, &hwc->period_left);
230
231 return new_raw_count;
232}
233
234static void
235armpmu_read(struct perf_event *event)
236{
237 struct hw_perf_event *hwc = &event->hw;
238
239 /* Don't read disabled counters! */
240 if (hwc->idx < 0)
241 return;
242
243 armpmu_event_update(event, hwc, hwc->idx, 0);
244}
245
246static void
247armpmu_stop(struct perf_event *event, int flags)
248{
249 struct hw_perf_event *hwc = &event->hw;
250
251 if (!armpmu)
252 return;
253
254 /*
255 * ARM pmu always has to update the counter, so ignore
256 * PERF_EF_UPDATE, see comments in armpmu_start().
257 */
258 if (!(hwc->state & PERF_HES_STOPPED)) {
259 armpmu->disable(hwc, hwc->idx);
260 barrier(); /* why? */
261 armpmu_event_update(event, hwc, hwc->idx, 0);
262 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
263 }
264}
265
266static void
267armpmu_start(struct perf_event *event, int flags)
268{
269 struct hw_perf_event *hwc = &event->hw;
270
271 if (!armpmu)
272 return;
273
274 /*
275 * ARM pmu always has to reprogram the period, so ignore
276 * PERF_EF_RELOAD, see the comment below.
277 */
278 if (flags & PERF_EF_RELOAD)
279 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
280
281 hwc->state = 0;
282 /*
283 * Set the period again. Some counters can't be stopped, so when we
284 * were stopped we simply disabled the IRQ source and the counter
285 * may have been left counting. If we don't do this step then we may
286 * get an interrupt too soon or *way* too late if the overflow has
287 * happened since disabling.
288 */
289 armpmu_event_set_period(event, hwc, hwc->idx);
290 armpmu->enable(hwc, hwc->idx);
291}
292
293static void
294armpmu_del(struct perf_event *event, int flags)
295{
296 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
297 struct hw_perf_event *hwc = &event->hw;
298 int idx = hwc->idx;
299
300 WARN_ON(idx < 0);
301
302 clear_bit(idx, cpuc->active_mask);
303 armpmu_stop(event, PERF_EF_UPDATE);
304 cpuc->events[idx] = NULL;
305 clear_bit(idx, cpuc->used_mask);
306
307 perf_event_update_userpage(event);
308}
309
310static int
311armpmu_add(struct perf_event *event, int flags)
312{
313 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
314 struct hw_perf_event *hwc = &event->hw;
315 int idx;
316 int err = 0;
317
318 perf_pmu_disable(event->pmu);
319
320 /* If we don't have a space for the counter then finish early. */
321 idx = armpmu->get_event_idx(cpuc, hwc);
322 if (idx < 0) {
323 err = idx;
324 goto out;
325 }
326
327 /*
328 * If there is an event in the counter we are going to use then make
329 * sure it is disabled.
330 */
331 event->hw.idx = idx;
332 armpmu->disable(hwc, idx);
333 cpuc->events[idx] = event;
334 set_bit(idx, cpuc->active_mask);
335
336 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
337 if (flags & PERF_EF_START)
338 armpmu_start(event, PERF_EF_RELOAD);
339
340 /* Propagate our changes to the userspace mapping. */
341 perf_event_update_userpage(event);
342
343out:
344 perf_pmu_enable(event->pmu);
345 return err;
346}
347
348static struct pmu pmu;
349
350static int
351validate_event(struct cpu_hw_events *cpuc,
352 struct perf_event *event)
353{
354 struct hw_perf_event fake_event = event->hw;
355
356 if (event->pmu != &pmu || event->state <= PERF_EVENT_STATE_OFF)
357 return 1;
358
359 return armpmu->get_event_idx(cpuc, &fake_event) >= 0;
360}
361
362static int
363validate_group(struct perf_event *event)
364{
365 struct perf_event *sibling, *leader = event->group_leader;
366 struct cpu_hw_events fake_pmu;
367
368 memset(&fake_pmu, 0, sizeof(fake_pmu));
369
370 if (!validate_event(&fake_pmu, leader))
371 return -ENOSPC;
372
373 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
374 if (!validate_event(&fake_pmu, sibling))
375 return -ENOSPC;
376 }
377
378 if (!validate_event(&fake_pmu, event))
379 return -ENOSPC;
380
381 return 0;
382}
383
384static irqreturn_t armpmu_platform_irq(int irq, void *dev)
385{
386 struct arm_pmu_platdata *plat = dev_get_platdata(&pmu_device->dev);
387
388 return plat->handle_irq(irq, dev, armpmu->handle_irq);
389}
390
391static int
392armpmu_reserve_hardware(void)
393{
394 struct arm_pmu_platdata *plat;
395 irq_handler_t handle_irq;
396 int i, err = -ENODEV, irq;
397
398 pmu_device = reserve_pmu(ARM_PMU_DEVICE_CPU);
399 if (IS_ERR(pmu_device)) {
400 pr_warning("unable to reserve pmu\n");
401 return PTR_ERR(pmu_device);
402 }
403
404 init_pmu(ARM_PMU_DEVICE_CPU);
405
406 plat = dev_get_platdata(&pmu_device->dev);
407 if (plat && plat->handle_irq)
408 handle_irq = armpmu_platform_irq;
409 else
410 handle_irq = armpmu->handle_irq;
411
412 if (pmu_device->num_resources < 1) {
413 pr_err("no irqs for PMUs defined\n");
414 return -ENODEV;
415 }
416
417 for (i = 0; i < pmu_device->num_resources; ++i) {
418 irq = platform_get_irq(pmu_device, i);
419 if (irq < 0)
420 continue;
421
422 err = request_irq(irq, handle_irq,
423 IRQF_DISABLED | IRQF_NOBALANCING,
424 "armpmu", NULL);
425 if (err) {
426 pr_warning("unable to request IRQ%d for ARM perf "
427 "counters\n", irq);
428 break;
429 }
430 }
431
432 if (err) {
433 for (i = i - 1; i >= 0; --i) {
434 irq = platform_get_irq(pmu_device, i);
435 if (irq >= 0)
436 free_irq(irq, NULL);
437 }
438 release_pmu(ARM_PMU_DEVICE_CPU);
439 pmu_device = NULL;
440 }
441
442 return err;
443}
444
445static void
446armpmu_release_hardware(void)
447{
448 int i, irq;
449
450 for (i = pmu_device->num_resources - 1; i >= 0; --i) {
451 irq = platform_get_irq(pmu_device, i);
452 if (irq >= 0)
453 free_irq(irq, NULL);
454 }
455 armpmu->stop();
456
457 release_pmu(ARM_PMU_DEVICE_CPU);
458 pmu_device = NULL;
459}
460
461static atomic_t active_events = ATOMIC_INIT(0);
462static DEFINE_MUTEX(pmu_reserve_mutex);
463
464static void
465hw_perf_event_destroy(struct perf_event *event)
466{
467 if (atomic_dec_and_mutex_lock(&active_events, &pmu_reserve_mutex)) {
468 armpmu_release_hardware();
469 mutex_unlock(&pmu_reserve_mutex);
470 }
471}
472
473static int
474__hw_perf_event_init(struct perf_event *event)
475{
476 struct hw_perf_event *hwc = &event->hw;
477 int mapping, err;
478
479 /* Decode the generic type into an ARM event identifier. */
480 if (PERF_TYPE_HARDWARE == event->attr.type) {
481 mapping = armpmu_map_event(event->attr.config);
482 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
483 mapping = armpmu_map_cache_event(event->attr.config);
484 } else if (PERF_TYPE_RAW == event->attr.type) {
485 mapping = armpmu_map_raw_event(event->attr.config);
486 } else {
487 pr_debug("event type %x not supported\n", event->attr.type);
488 return -EOPNOTSUPP;
489 }
490
491 if (mapping < 0) {
492 pr_debug("event %x:%llx not supported\n", event->attr.type,
493 event->attr.config);
494 return mapping;
495 }
496
497 /*
498 * Check whether we need to exclude the counter from certain modes.
499 * The ARM performance counters are on all of the time so if someone
500 * has asked us for some excludes then we have to fail.
501 */
502 if (event->attr.exclude_kernel || event->attr.exclude_user ||
503 event->attr.exclude_hv || event->attr.exclude_idle) {
504 pr_debug("ARM performance counters do not support "
505 "mode exclusion\n");
506 return -EPERM;
507 }
508
509 /*
510 * We don't assign an index until we actually place the event onto
511 * hardware. Use -1 to signify that we haven't decided where to put it
512 * yet. For SMP systems, each core has it's own PMU so we can't do any
513 * clever allocation or constraints checking at this point.
514 */
515 hwc->idx = -1;
516
517 /*
518 * Store the event encoding into the config_base field. config and
519 * event_base are unused as the only 2 things we need to know are
520 * the event mapping and the counter to use. The counter to use is
521 * also the indx and the config_base is the event type.
522 */
523 hwc->config_base = (unsigned long)mapping;
524 hwc->config = 0;
525 hwc->event_base = 0;
526
527 if (!hwc->sample_period) {
528 hwc->sample_period = armpmu->max_period;
529 hwc->last_period = hwc->sample_period;
530 local64_set(&hwc->period_left, hwc->sample_period);
531 }
532
533 err = 0;
534 if (event->group_leader != event) {
535 err = validate_group(event);
536 if (err)
537 return -EINVAL;
538 }
539
540 return err;
541}
542
543static int armpmu_event_init(struct perf_event *event)
544{
545 int err = 0;
546
547 switch (event->attr.type) {
548 case PERF_TYPE_RAW:
549 case PERF_TYPE_HARDWARE:
550 case PERF_TYPE_HW_CACHE:
551 break;
552
553 default:
554 return -ENOENT;
555 }
556
557 if (!armpmu)
558 return -ENODEV;
559
560 event->destroy = hw_perf_event_destroy;
561
562 if (!atomic_inc_not_zero(&active_events)) {
563 mutex_lock(&pmu_reserve_mutex);
564 if (atomic_read(&active_events) == 0) {
565 err = armpmu_reserve_hardware();
566 }
567
568 if (!err)
569 atomic_inc(&active_events);
570 mutex_unlock(&pmu_reserve_mutex);
571 }
572
573 if (err)
574 return err;
575
576 err = __hw_perf_event_init(event);
577 if (err)
578 hw_perf_event_destroy(event);
579
580 return err;
581}
582
583static void armpmu_enable(struct pmu *pmu)
584{
585 /* Enable all of the perf events on hardware. */
586 int idx, enabled = 0;
587 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
588
589 if (!armpmu)
590 return;
591
592 for (idx = 0; idx <= armpmu->num_events; ++idx) {
593 struct perf_event *event = cpuc->events[idx];
594
595 if (!event)
596 continue;
597
598 armpmu->enable(&event->hw, idx);
599 enabled = 1;
600 }
601
602 if (enabled)
603 armpmu->start();
604}
605
606static void armpmu_disable(struct pmu *pmu)
607{
608 if (armpmu)
609 armpmu->stop();
610}
611
612static struct pmu pmu = {
613 .pmu_enable = armpmu_enable,
614 .pmu_disable = armpmu_disable,
615 .event_init = armpmu_event_init,
616 .add = armpmu_add,
617 .del = armpmu_del,
618 .start = armpmu_start,
619 .stop = armpmu_stop,
620 .read = armpmu_read,
621};
622
623/* Include the PMU-specific implementations. */
624#include "perf_event_xscale.c"
625#include "perf_event_v6.c"
626#include "perf_event_v7.c"
627
628/*
629 * Ensure the PMU has sane values out of reset.
630 * This requires SMP to be available, so exists as a separate initcall.
631 */
632static int __init
633armpmu_reset(void)
634{
635 if (armpmu && armpmu->reset)
636 return on_each_cpu(armpmu->reset, NULL, 1);
637 return 0;
638}
639arch_initcall(armpmu_reset);
640
641static int __init
642init_hw_perf_events(void)
643{
644 unsigned long cpuid = read_cpuid_id();
645 unsigned long implementor = (cpuid & 0xFF000000) >> 24;
646 unsigned long part_number = (cpuid & 0xFFF0);
647
648 /* ARM Ltd CPUs. */
649 if (0x41 == implementor) {
650 switch (part_number) {
651 case 0xB360: /* ARM1136 */
652 case 0xB560: /* ARM1156 */
653 case 0xB760: /* ARM1176 */
654 armpmu = armv6pmu_init();
655 break;
656 case 0xB020: /* ARM11mpcore */
657 armpmu = armv6mpcore_pmu_init();
658 break;
659 case 0xC080: /* Cortex-A8 */
660 armpmu = armv7_a8_pmu_init();
661 break;
662 case 0xC090: /* Cortex-A9 */
663 armpmu = armv7_a9_pmu_init();
664 break;
665 case 0xC050: /* Cortex-A5 */
666 armpmu = armv7_a5_pmu_init();
667 break;
668 case 0xC0F0: /* Cortex-A15 */
669 armpmu = armv7_a15_pmu_init();
670 break;
671 }
672 /* Intel CPUs [xscale]. */
673 } else if (0x69 == implementor) {
674 part_number = (cpuid >> 13) & 0x7;
675 switch (part_number) {
676 case 1:
677 armpmu = xscale1pmu_init();
678 break;
679 case 2:
680 armpmu = xscale2pmu_init();
681 break;
682 }
683 }
684
685 if (armpmu) {
686 pr_info("enabled with %s PMU driver, %d counters available\n",
687 armpmu->name, armpmu->num_events);
688 } else {
689 pr_info("no hardware support available\n");
690 }
691
692 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
693
694 return 0;
695}
696early_initcall(init_hw_perf_events);
697
698/*
699 * Callchain handling code.
700 */
701
702/*
703 * The registers we're interested in are at the end of the variable
704 * length saved register structure. The fp points at the end of this
705 * structure so the address of this struct is:
706 * (struct frame_tail *)(xxx->fp)-1
707 *
708 * This code has been adapted from the ARM OProfile support.
709 */
710struct frame_tail {
711 struct frame_tail __user *fp;
712 unsigned long sp;
713 unsigned long lr;
714} __attribute__((packed));
715
716/*
717 * Get the return address for a single stackframe and return a pointer to the
718 * next frame tail.
719 */
720static struct frame_tail __user *
721user_backtrace(struct frame_tail __user *tail,
722 struct perf_callchain_entry *entry)
723{
724 struct frame_tail buftail;
725
726 /* Also check accessibility of one struct frame_tail beyond */
727 if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
728 return NULL;
729 if (__copy_from_user_inatomic(&buftail, tail, sizeof(buftail)))
730 return NULL;
731
732 perf_callchain_store(entry, buftail.lr);
733
734 /*
735 * Frame pointers should strictly progress back up the stack
736 * (towards higher addresses).
737 */
738 if (tail + 1 >= buftail.fp)
739 return NULL;
740
741 return buftail.fp - 1;
742}
743
744void
745perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
746{
747 struct frame_tail __user *tail;
748
749
750 tail = (struct frame_tail __user *)regs->ARM_fp - 1;
751
752 while ((entry->nr < PERF_MAX_STACK_DEPTH) &&
753 tail && !((unsigned long)tail & 0x3))
754 tail = user_backtrace(tail, entry);
755}
756
757/*
758 * Gets called by walk_stackframe() for every stackframe. This will be called
759 * whist unwinding the stackframe and is like a subroutine return so we use
760 * the PC.
761 */
762static int
763callchain_trace(struct stackframe *fr,
764 void *data)
765{
766 struct perf_callchain_entry *entry = data;
767 perf_callchain_store(entry, fr->pc);
768 return 0;
769}
770
771void
772perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
773{
774 struct stackframe fr;
775
776 fr.fp = regs->ARM_fp;
777 fr.sp = regs->ARM_sp;
778 fr.lr = regs->ARM_lr;
779 fr.pc = regs->ARM_pc;
780 walk_stackframe(&fr, callchain_trace, entry);
781}
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. Callchain code is based on the ARM OProfile backtrace
11 * code.
12 */
13#define pr_fmt(fmt) "hw perfevents: " fmt
14
15#include <linux/kernel.h>
16#include <linux/platform_device.h>
17#include <linux/pm_runtime.h>
18#include <linux/uaccess.h>
19#include <linux/irq.h>
20#include <linux/irqdesc.h>
21
22#include <asm/irq_regs.h>
23#include <asm/pmu.h>
24#include <asm/stacktrace.h>
25
26static int
27armpmu_map_cache_event(const unsigned (*cache_map)
28 [PERF_COUNT_HW_CACHE_MAX]
29 [PERF_COUNT_HW_CACHE_OP_MAX]
30 [PERF_COUNT_HW_CACHE_RESULT_MAX],
31 u64 config)
32{
33 unsigned int cache_type, cache_op, cache_result, ret;
34
35 cache_type = (config >> 0) & 0xff;
36 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
37 return -EINVAL;
38
39 cache_op = (config >> 8) & 0xff;
40 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
41 return -EINVAL;
42
43 cache_result = (config >> 16) & 0xff;
44 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
45 return -EINVAL;
46
47 ret = (int)(*cache_map)[cache_type][cache_op][cache_result];
48
49 if (ret == CACHE_OP_UNSUPPORTED)
50 return -ENOENT;
51
52 return ret;
53}
54
55static int
56armpmu_map_hw_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config)
57{
58 int mapping;
59
60 if (config >= PERF_COUNT_HW_MAX)
61 return -EINVAL;
62
63 mapping = (*event_map)[config];
64 return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping;
65}
66
67static int
68armpmu_map_raw_event(u32 raw_event_mask, u64 config)
69{
70 return (int)(config & raw_event_mask);
71}
72
73int
74armpmu_map_event(struct perf_event *event,
75 const unsigned (*event_map)[PERF_COUNT_HW_MAX],
76 const unsigned (*cache_map)
77 [PERF_COUNT_HW_CACHE_MAX]
78 [PERF_COUNT_HW_CACHE_OP_MAX]
79 [PERF_COUNT_HW_CACHE_RESULT_MAX],
80 u32 raw_event_mask)
81{
82 u64 config = event->attr.config;
83
84 switch (event->attr.type) {
85 case PERF_TYPE_HARDWARE:
86 return armpmu_map_hw_event(event_map, config);
87 case PERF_TYPE_HW_CACHE:
88 return armpmu_map_cache_event(cache_map, config);
89 case PERF_TYPE_RAW:
90 return armpmu_map_raw_event(raw_event_mask, config);
91 }
92
93 return -ENOENT;
94}
95
96int armpmu_event_set_period(struct perf_event *event)
97{
98 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
99 struct hw_perf_event *hwc = &event->hw;
100 s64 left = local64_read(&hwc->period_left);
101 s64 period = hwc->sample_period;
102 int ret = 0;
103
104 if (unlikely(left <= -period)) {
105 left = period;
106 local64_set(&hwc->period_left, left);
107 hwc->last_period = period;
108 ret = 1;
109 }
110
111 if (unlikely(left <= 0)) {
112 left += period;
113 local64_set(&hwc->period_left, left);
114 hwc->last_period = period;
115 ret = 1;
116 }
117
118 if (left > (s64)armpmu->max_period)
119 left = armpmu->max_period;
120
121 local64_set(&hwc->prev_count, (u64)-left);
122
123 armpmu->write_counter(event, (u64)(-left) & 0xffffffff);
124
125 perf_event_update_userpage(event);
126
127 return ret;
128}
129
130u64 armpmu_event_update(struct perf_event *event)
131{
132 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
133 struct hw_perf_event *hwc = &event->hw;
134 u64 delta, prev_raw_count, new_raw_count;
135
136again:
137 prev_raw_count = local64_read(&hwc->prev_count);
138 new_raw_count = armpmu->read_counter(event);
139
140 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
141 new_raw_count) != prev_raw_count)
142 goto again;
143
144 delta = (new_raw_count - prev_raw_count) & armpmu->max_period;
145
146 local64_add(delta, &event->count);
147 local64_sub(delta, &hwc->period_left);
148
149 return new_raw_count;
150}
151
152static void
153armpmu_read(struct perf_event *event)
154{
155 armpmu_event_update(event);
156}
157
158static void
159armpmu_stop(struct perf_event *event, int flags)
160{
161 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
162 struct hw_perf_event *hwc = &event->hw;
163
164 /*
165 * ARM pmu always has to update the counter, so ignore
166 * PERF_EF_UPDATE, see comments in armpmu_start().
167 */
168 if (!(hwc->state & PERF_HES_STOPPED)) {
169 armpmu->disable(event);
170 armpmu_event_update(event);
171 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
172 }
173}
174
175static void armpmu_start(struct perf_event *event, int flags)
176{
177 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
178 struct hw_perf_event *hwc = &event->hw;
179
180 /*
181 * ARM pmu always has to reprogram the period, so ignore
182 * PERF_EF_RELOAD, see the comment below.
183 */
184 if (flags & PERF_EF_RELOAD)
185 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
186
187 hwc->state = 0;
188 /*
189 * Set the period again. Some counters can't be stopped, so when we
190 * were stopped we simply disabled the IRQ source and the counter
191 * may have been left counting. If we don't do this step then we may
192 * get an interrupt too soon or *way* too late if the overflow has
193 * happened since disabling.
194 */
195 armpmu_event_set_period(event);
196 armpmu->enable(event);
197}
198
199static void
200armpmu_del(struct perf_event *event, int flags)
201{
202 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
203 struct pmu_hw_events *hw_events = armpmu->get_hw_events();
204 struct hw_perf_event *hwc = &event->hw;
205 int idx = hwc->idx;
206
207 armpmu_stop(event, PERF_EF_UPDATE);
208 hw_events->events[idx] = NULL;
209 clear_bit(idx, hw_events->used_mask);
210 if (armpmu->clear_event_idx)
211 armpmu->clear_event_idx(hw_events, event);
212
213 perf_event_update_userpage(event);
214}
215
216static int
217armpmu_add(struct perf_event *event, int flags)
218{
219 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
220 struct pmu_hw_events *hw_events = armpmu->get_hw_events();
221 struct hw_perf_event *hwc = &event->hw;
222 int idx;
223 int err = 0;
224
225 perf_pmu_disable(event->pmu);
226
227 /* If we don't have a space for the counter then finish early. */
228 idx = armpmu->get_event_idx(hw_events, event);
229 if (idx < 0) {
230 err = idx;
231 goto out;
232 }
233
234 /*
235 * If there is an event in the counter we are going to use then make
236 * sure it is disabled.
237 */
238 event->hw.idx = idx;
239 armpmu->disable(event);
240 hw_events->events[idx] = event;
241
242 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
243 if (flags & PERF_EF_START)
244 armpmu_start(event, PERF_EF_RELOAD);
245
246 /* Propagate our changes to the userspace mapping. */
247 perf_event_update_userpage(event);
248
249out:
250 perf_pmu_enable(event->pmu);
251 return err;
252}
253
254static int
255validate_event(struct pmu_hw_events *hw_events,
256 struct perf_event *event)
257{
258 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
259
260 if (is_software_event(event))
261 return 1;
262
263 if (event->state < PERF_EVENT_STATE_OFF)
264 return 1;
265
266 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
267 return 1;
268
269 return armpmu->get_event_idx(hw_events, event) >= 0;
270}
271
272static int
273validate_group(struct perf_event *event)
274{
275 struct perf_event *sibling, *leader = event->group_leader;
276 struct pmu_hw_events fake_pmu;
277 DECLARE_BITMAP(fake_used_mask, ARMPMU_MAX_HWEVENTS);
278
279 /*
280 * Initialise the fake PMU. We only need to populate the
281 * used_mask for the purposes of validation.
282 */
283 memset(fake_used_mask, 0, sizeof(fake_used_mask));
284 fake_pmu.used_mask = fake_used_mask;
285
286 if (!validate_event(&fake_pmu, leader))
287 return -EINVAL;
288
289 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
290 if (!validate_event(&fake_pmu, sibling))
291 return -EINVAL;
292 }
293
294 if (!validate_event(&fake_pmu, event))
295 return -EINVAL;
296
297 return 0;
298}
299
300static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
301{
302 struct arm_pmu *armpmu;
303 struct platform_device *plat_device;
304 struct arm_pmu_platdata *plat;
305 int ret;
306 u64 start_clock, finish_clock;
307
308 if (irq_is_percpu(irq))
309 dev = *(void **)dev;
310 armpmu = dev;
311 plat_device = armpmu->plat_device;
312 plat = dev_get_platdata(&plat_device->dev);
313
314 start_clock = sched_clock();
315 if (plat && plat->handle_irq)
316 ret = plat->handle_irq(irq, dev, armpmu->handle_irq);
317 else
318 ret = armpmu->handle_irq(irq, dev);
319 finish_clock = sched_clock();
320
321 perf_sample_event_took(finish_clock - start_clock);
322 return ret;
323}
324
325static void
326armpmu_release_hardware(struct arm_pmu *armpmu)
327{
328 armpmu->free_irq(armpmu);
329 pm_runtime_put_sync(&armpmu->plat_device->dev);
330}
331
332static int
333armpmu_reserve_hardware(struct arm_pmu *armpmu)
334{
335 int err;
336 struct platform_device *pmu_device = armpmu->plat_device;
337
338 if (!pmu_device)
339 return -ENODEV;
340
341 pm_runtime_get_sync(&pmu_device->dev);
342 err = armpmu->request_irq(armpmu, armpmu_dispatch_irq);
343 if (err) {
344 armpmu_release_hardware(armpmu);
345 return err;
346 }
347
348 return 0;
349}
350
351static void
352hw_perf_event_destroy(struct perf_event *event)
353{
354 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
355 atomic_t *active_events = &armpmu->active_events;
356 struct mutex *pmu_reserve_mutex = &armpmu->reserve_mutex;
357
358 if (atomic_dec_and_mutex_lock(active_events, pmu_reserve_mutex)) {
359 armpmu_release_hardware(armpmu);
360 mutex_unlock(pmu_reserve_mutex);
361 }
362}
363
364static int
365event_requires_mode_exclusion(struct perf_event_attr *attr)
366{
367 return attr->exclude_idle || attr->exclude_user ||
368 attr->exclude_kernel || attr->exclude_hv;
369}
370
371static int
372__hw_perf_event_init(struct perf_event *event)
373{
374 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
375 struct hw_perf_event *hwc = &event->hw;
376 int mapping;
377
378 mapping = armpmu->map_event(event);
379
380 if (mapping < 0) {
381 pr_debug("event %x:%llx not supported\n", event->attr.type,
382 event->attr.config);
383 return mapping;
384 }
385
386 /*
387 * We don't assign an index until we actually place the event onto
388 * hardware. Use -1 to signify that we haven't decided where to put it
389 * yet. For SMP systems, each core has it's own PMU so we can't do any
390 * clever allocation or constraints checking at this point.
391 */
392 hwc->idx = -1;
393 hwc->config_base = 0;
394 hwc->config = 0;
395 hwc->event_base = 0;
396
397 /*
398 * Check whether we need to exclude the counter from certain modes.
399 */
400 if ((!armpmu->set_event_filter ||
401 armpmu->set_event_filter(hwc, &event->attr)) &&
402 event_requires_mode_exclusion(&event->attr)) {
403 pr_debug("ARM performance counters do not support "
404 "mode exclusion\n");
405 return -EOPNOTSUPP;
406 }
407
408 /*
409 * Store the event encoding into the config_base field.
410 */
411 hwc->config_base |= (unsigned long)mapping;
412
413 if (!hwc->sample_period) {
414 /*
415 * For non-sampling runs, limit the sample_period to half
416 * of the counter width. That way, the new counter value
417 * is far less likely to overtake the previous one unless
418 * you have some serious IRQ latency issues.
419 */
420 hwc->sample_period = armpmu->max_period >> 1;
421 hwc->last_period = hwc->sample_period;
422 local64_set(&hwc->period_left, hwc->sample_period);
423 }
424
425 if (event->group_leader != event) {
426 if (validate_group(event) != 0)
427 return -EINVAL;
428 }
429
430 return 0;
431}
432
433static int armpmu_event_init(struct perf_event *event)
434{
435 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
436 int err = 0;
437 atomic_t *active_events = &armpmu->active_events;
438
439 /* does not support taken branch sampling */
440 if (has_branch_stack(event))
441 return -EOPNOTSUPP;
442
443 if (armpmu->map_event(event) == -ENOENT)
444 return -ENOENT;
445
446 event->destroy = hw_perf_event_destroy;
447
448 if (!atomic_inc_not_zero(active_events)) {
449 mutex_lock(&armpmu->reserve_mutex);
450 if (atomic_read(active_events) == 0)
451 err = armpmu_reserve_hardware(armpmu);
452
453 if (!err)
454 atomic_inc(active_events);
455 mutex_unlock(&armpmu->reserve_mutex);
456 }
457
458 if (err)
459 return err;
460
461 err = __hw_perf_event_init(event);
462 if (err)
463 hw_perf_event_destroy(event);
464
465 return err;
466}
467
468static void armpmu_enable(struct pmu *pmu)
469{
470 struct arm_pmu *armpmu = to_arm_pmu(pmu);
471 struct pmu_hw_events *hw_events = armpmu->get_hw_events();
472 int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
473
474 if (enabled)
475 armpmu->start(armpmu);
476}
477
478static void armpmu_disable(struct pmu *pmu)
479{
480 struct arm_pmu *armpmu = to_arm_pmu(pmu);
481 armpmu->stop(armpmu);
482}
483
484#ifdef CONFIG_PM_RUNTIME
485static int armpmu_runtime_resume(struct device *dev)
486{
487 struct arm_pmu_platdata *plat = dev_get_platdata(dev);
488
489 if (plat && plat->runtime_resume)
490 return plat->runtime_resume(dev);
491
492 return 0;
493}
494
495static int armpmu_runtime_suspend(struct device *dev)
496{
497 struct arm_pmu_platdata *plat = dev_get_platdata(dev);
498
499 if (plat && plat->runtime_suspend)
500 return plat->runtime_suspend(dev);
501
502 return 0;
503}
504#endif
505
506const struct dev_pm_ops armpmu_dev_pm_ops = {
507 SET_RUNTIME_PM_OPS(armpmu_runtime_suspend, armpmu_runtime_resume, NULL)
508};
509
510static void armpmu_init(struct arm_pmu *armpmu)
511{
512 atomic_set(&armpmu->active_events, 0);
513 mutex_init(&armpmu->reserve_mutex);
514
515 armpmu->pmu = (struct pmu) {
516 .pmu_enable = armpmu_enable,
517 .pmu_disable = armpmu_disable,
518 .event_init = armpmu_event_init,
519 .add = armpmu_add,
520 .del = armpmu_del,
521 .start = armpmu_start,
522 .stop = armpmu_stop,
523 .read = armpmu_read,
524 };
525}
526
527int armpmu_register(struct arm_pmu *armpmu, int type)
528{
529 armpmu_init(armpmu);
530 pm_runtime_enable(&armpmu->plat_device->dev);
531 pr_info("enabled with %s PMU driver, %d counters available\n",
532 armpmu->name, armpmu->num_events);
533 return perf_pmu_register(&armpmu->pmu, armpmu->name, type);
534}
535
536/*
537 * Callchain handling code.
538 */
539
540/*
541 * The registers we're interested in are at the end of the variable
542 * length saved register structure. The fp points at the end of this
543 * structure so the address of this struct is:
544 * (struct frame_tail *)(xxx->fp)-1
545 *
546 * This code has been adapted from the ARM OProfile support.
547 */
548struct frame_tail {
549 struct frame_tail __user *fp;
550 unsigned long sp;
551 unsigned long lr;
552} __attribute__((packed));
553
554/*
555 * Get the return address for a single stackframe and return a pointer to the
556 * next frame tail.
557 */
558static struct frame_tail __user *
559user_backtrace(struct frame_tail __user *tail,
560 struct perf_callchain_entry *entry)
561{
562 struct frame_tail buftail;
563
564 /* Also check accessibility of one struct frame_tail beyond */
565 if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
566 return NULL;
567 if (__copy_from_user_inatomic(&buftail, tail, sizeof(buftail)))
568 return NULL;
569
570 perf_callchain_store(entry, buftail.lr);
571
572 /*
573 * Frame pointers should strictly progress back up the stack
574 * (towards higher addresses).
575 */
576 if (tail + 1 >= buftail.fp)
577 return NULL;
578
579 return buftail.fp - 1;
580}
581
582void
583perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
584{
585 struct frame_tail __user *tail;
586
587 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
588 /* We don't support guest os callchain now */
589 return;
590 }
591
592 perf_callchain_store(entry, regs->ARM_pc);
593 tail = (struct frame_tail __user *)regs->ARM_fp - 1;
594
595 while ((entry->nr < PERF_MAX_STACK_DEPTH) &&
596 tail && !((unsigned long)tail & 0x3))
597 tail = user_backtrace(tail, entry);
598}
599
600/*
601 * Gets called by walk_stackframe() for every stackframe. This will be called
602 * whist unwinding the stackframe and is like a subroutine return so we use
603 * the PC.
604 */
605static int
606callchain_trace(struct stackframe *fr,
607 void *data)
608{
609 struct perf_callchain_entry *entry = data;
610 perf_callchain_store(entry, fr->pc);
611 return 0;
612}
613
614void
615perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
616{
617 struct stackframe fr;
618
619 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
620 /* We don't support guest os callchain now */
621 return;
622 }
623
624 fr.fp = regs->ARM_fp;
625 fr.sp = regs->ARM_sp;
626 fr.lr = regs->ARM_lr;
627 fr.pc = regs->ARM_pc;
628 walk_stackframe(&fr, callchain_trace, entry);
629}
630
631unsigned long perf_instruction_pointer(struct pt_regs *regs)
632{
633 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
634 return perf_guest_cbs->get_guest_ip();
635
636 return instruction_pointer(regs);
637}
638
639unsigned long perf_misc_flags(struct pt_regs *regs)
640{
641 int misc = 0;
642
643 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
644 if (perf_guest_cbs->is_user_mode())
645 misc |= PERF_RECORD_MISC_GUEST_USER;
646 else
647 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
648 } else {
649 if (user_mode(regs))
650 misc |= PERF_RECORD_MISC_USER;
651 else
652 misc |= PERF_RECORD_MISC_KERNEL;
653 }
654
655 return misc;
656}