]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - arch/arm/kernel/perf_event.c
Add a document describing the padata interface
[mirror_ubuntu-hirsute-kernel.git] / arch / arm / kernel / perf_event.c
1 #undef DEBUG
2
3 /*
4 * ARM performance counter support.
5 *
6 * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
7 *
8 * ARMv7 support: Jean Pihet <jpihet@mvista.com>
9 * 2010 (c) MontaVista Software, LLC.
10 *
11 * This code is based on the sparc64 perf event code, which is in turn based
12 * on the x86 code. Callchain code is based on the ARM OProfile backtrace
13 * code.
14 */
15 #define pr_fmt(fmt) "hw perfevents: " fmt
16
17 #include <linux/interrupt.h>
18 #include <linux/kernel.h>
19 #include <linux/perf_event.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
29 static const struct pmu_irqs *pmu_irqs;
30
31 /*
32 * Hardware lock to serialize accesses to PMU registers. Needed for the
33 * read/modify/write sequences.
34 */
35 DEFINE_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. */
49 struct 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 };
68 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
69
70 struct arm_pmu {
71 char *name;
72 irqreturn_t (*handle_irq)(int irq_num, void *dev);
73 void (*enable)(struct hw_perf_event *evt, int idx);
74 void (*disable)(struct hw_perf_event *evt, int idx);
75 int (*event_map)(int evt);
76 u64 (*raw_event)(u64);
77 int (*get_event_idx)(struct cpu_hw_events *cpuc,
78 struct hw_perf_event *hwc);
79 u32 (*read_counter)(int idx);
80 void (*write_counter)(int idx, u32 val);
81 void (*start)(void);
82 void (*stop)(void);
83 int num_events;
84 u64 max_period;
85 };
86
87 /* Set at runtime when we know what CPU type we are. */
88 static const struct arm_pmu *armpmu;
89
90 #define HW_OP_UNSUPPORTED 0xFFFF
91
92 #define C(_x) \
93 PERF_COUNT_HW_CACHE_##_x
94
95 #define CACHE_OP_UNSUPPORTED 0xFFFF
96
97 static unsigned armpmu_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
98 [PERF_COUNT_HW_CACHE_OP_MAX]
99 [PERF_COUNT_HW_CACHE_RESULT_MAX];
100
101 static int
102 armpmu_map_cache_event(u64 config)
103 {
104 unsigned int cache_type, cache_op, cache_result, ret;
105
106 cache_type = (config >> 0) & 0xff;
107 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
108 return -EINVAL;
109
110 cache_op = (config >> 8) & 0xff;
111 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
112 return -EINVAL;
113
114 cache_result = (config >> 16) & 0xff;
115 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
116 return -EINVAL;
117
118 ret = (int)armpmu_perf_cache_map[cache_type][cache_op][cache_result];
119
120 if (ret == CACHE_OP_UNSUPPORTED)
121 return -ENOENT;
122
123 return ret;
124 }
125
126 static int
127 armpmu_event_set_period(struct perf_event *event,
128 struct hw_perf_event *hwc,
129 int idx)
130 {
131 s64 left = atomic64_read(&hwc->period_left);
132 s64 period = hwc->sample_period;
133 int ret = 0;
134
135 if (unlikely(left <= -period)) {
136 left = period;
137 atomic64_set(&hwc->period_left, left);
138 hwc->last_period = period;
139 ret = 1;
140 }
141
142 if (unlikely(left <= 0)) {
143 left += period;
144 atomic64_set(&hwc->period_left, left);
145 hwc->last_period = period;
146 ret = 1;
147 }
148
149 if (left > (s64)armpmu->max_period)
150 left = armpmu->max_period;
151
152 atomic64_set(&hwc->prev_count, (u64)-left);
153
154 armpmu->write_counter(idx, (u64)(-left) & 0xffffffff);
155
156 perf_event_update_userpage(event);
157
158 return ret;
159 }
160
161 static u64
162 armpmu_event_update(struct perf_event *event,
163 struct hw_perf_event *hwc,
164 int idx)
165 {
166 int shift = 64 - 32;
167 s64 prev_raw_count, new_raw_count;
168 s64 delta;
169
170 again:
171 prev_raw_count = atomic64_read(&hwc->prev_count);
172 new_raw_count = armpmu->read_counter(idx);
173
174 if (atomic64_cmpxchg(&hwc->prev_count, prev_raw_count,
175 new_raw_count) != prev_raw_count)
176 goto again;
177
178 delta = (new_raw_count << shift) - (prev_raw_count << shift);
179 delta >>= shift;
180
181 atomic64_add(delta, &event->count);
182 atomic64_sub(delta, &hwc->period_left);
183
184 return new_raw_count;
185 }
186
187 static void
188 armpmu_disable(struct perf_event *event)
189 {
190 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
191 struct hw_perf_event *hwc = &event->hw;
192 int idx = hwc->idx;
193
194 WARN_ON(idx < 0);
195
196 clear_bit(idx, cpuc->active_mask);
197 armpmu->disable(hwc, idx);
198
199 barrier();
200
201 armpmu_event_update(event, hwc, idx);
202 cpuc->events[idx] = NULL;
203 clear_bit(idx, cpuc->used_mask);
204
205 perf_event_update_userpage(event);
206 }
207
208 static void
209 armpmu_read(struct perf_event *event)
210 {
211 struct hw_perf_event *hwc = &event->hw;
212
213 /* Don't read disabled counters! */
214 if (hwc->idx < 0)
215 return;
216
217 armpmu_event_update(event, hwc, hwc->idx);
218 }
219
220 static void
221 armpmu_unthrottle(struct perf_event *event)
222 {
223 struct hw_perf_event *hwc = &event->hw;
224
225 /*
226 * Set the period again. Some counters can't be stopped, so when we
227 * were throttled we simply disabled the IRQ source and the counter
228 * may have been left counting. If we don't do this step then we may
229 * get an interrupt too soon or *way* too late if the overflow has
230 * happened since disabling.
231 */
232 armpmu_event_set_period(event, hwc, hwc->idx);
233 armpmu->enable(hwc, hwc->idx);
234 }
235
236 static int
237 armpmu_enable(struct perf_event *event)
238 {
239 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
240 struct hw_perf_event *hwc = &event->hw;
241 int idx;
242 int err = 0;
243
244 /* If we don't have a space for the counter then finish early. */
245 idx = armpmu->get_event_idx(cpuc, hwc);
246 if (idx < 0) {
247 err = idx;
248 goto out;
249 }
250
251 /*
252 * If there is an event in the counter we are going to use then make
253 * sure it is disabled.
254 */
255 event->hw.idx = idx;
256 armpmu->disable(hwc, idx);
257 cpuc->events[idx] = event;
258 set_bit(idx, cpuc->active_mask);
259
260 /* Set the period for the event. */
261 armpmu_event_set_period(event, hwc, idx);
262
263 /* Enable the event. */
264 armpmu->enable(hwc, idx);
265
266 /* Propagate our changes to the userspace mapping. */
267 perf_event_update_userpage(event);
268
269 out:
270 return err;
271 }
272
273 static struct pmu pmu = {
274 .enable = armpmu_enable,
275 .disable = armpmu_disable,
276 .unthrottle = armpmu_unthrottle,
277 .read = armpmu_read,
278 };
279
280 static int
281 validate_event(struct cpu_hw_events *cpuc,
282 struct perf_event *event)
283 {
284 struct hw_perf_event fake_event = event->hw;
285
286 if (event->pmu && event->pmu != &pmu)
287 return 0;
288
289 return armpmu->get_event_idx(cpuc, &fake_event) >= 0;
290 }
291
292 static int
293 validate_group(struct perf_event *event)
294 {
295 struct perf_event *sibling, *leader = event->group_leader;
296 struct cpu_hw_events fake_pmu;
297
298 memset(&fake_pmu, 0, sizeof(fake_pmu));
299
300 if (!validate_event(&fake_pmu, leader))
301 return -ENOSPC;
302
303 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
304 if (!validate_event(&fake_pmu, sibling))
305 return -ENOSPC;
306 }
307
308 if (!validate_event(&fake_pmu, event))
309 return -ENOSPC;
310
311 return 0;
312 }
313
314 static int
315 armpmu_reserve_hardware(void)
316 {
317 int i;
318 int err;
319
320 pmu_irqs = reserve_pmu();
321 if (IS_ERR(pmu_irqs)) {
322 pr_warning("unable to reserve pmu\n");
323 return PTR_ERR(pmu_irqs);
324 }
325
326 init_pmu();
327
328 if (pmu_irqs->num_irqs < 1) {
329 pr_err("no irqs for PMUs defined\n");
330 return -ENODEV;
331 }
332
333 for (i = 0; i < pmu_irqs->num_irqs; ++i) {
334 err = request_irq(pmu_irqs->irqs[i], armpmu->handle_irq,
335 IRQF_DISABLED | IRQF_NOBALANCING,
336 "armpmu", NULL);
337 if (err) {
338 pr_warning("unable to request IRQ%d for ARM "
339 "perf counters\n", pmu_irqs->irqs[i]);
340 break;
341 }
342 }
343
344 if (err) {
345 for (i = i - 1; i >= 0; --i)
346 free_irq(pmu_irqs->irqs[i], NULL);
347 release_pmu(pmu_irqs);
348 pmu_irqs = NULL;
349 }
350
351 return err;
352 }
353
354 static void
355 armpmu_release_hardware(void)
356 {
357 int i;
358
359 for (i = pmu_irqs->num_irqs - 1; i >= 0; --i)
360 free_irq(pmu_irqs->irqs[i], NULL);
361 armpmu->stop();
362
363 release_pmu(pmu_irqs);
364 pmu_irqs = NULL;
365 }
366
367 static atomic_t active_events = ATOMIC_INIT(0);
368 static DEFINE_MUTEX(pmu_reserve_mutex);
369
370 static void
371 hw_perf_event_destroy(struct perf_event *event)
372 {
373 if (atomic_dec_and_mutex_lock(&active_events, &pmu_reserve_mutex)) {
374 armpmu_release_hardware();
375 mutex_unlock(&pmu_reserve_mutex);
376 }
377 }
378
379 static int
380 __hw_perf_event_init(struct perf_event *event)
381 {
382 struct hw_perf_event *hwc = &event->hw;
383 int mapping, err;
384
385 /* Decode the generic type into an ARM event identifier. */
386 if (PERF_TYPE_HARDWARE == event->attr.type) {
387 mapping = armpmu->event_map(event->attr.config);
388 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
389 mapping = armpmu_map_cache_event(event->attr.config);
390 } else if (PERF_TYPE_RAW == event->attr.type) {
391 mapping = armpmu->raw_event(event->attr.config);
392 } else {
393 pr_debug("event type %x not supported\n", event->attr.type);
394 return -EOPNOTSUPP;
395 }
396
397 if (mapping < 0) {
398 pr_debug("event %x:%llx not supported\n", event->attr.type,
399 event->attr.config);
400 return mapping;
401 }
402
403 /*
404 * Check whether we need to exclude the counter from certain modes.
405 * The ARM performance counters are on all of the time so if someone
406 * has asked us for some excludes then we have to fail.
407 */
408 if (event->attr.exclude_kernel || event->attr.exclude_user ||
409 event->attr.exclude_hv || event->attr.exclude_idle) {
410 pr_debug("ARM performance counters do not support "
411 "mode exclusion\n");
412 return -EPERM;
413 }
414
415 /*
416 * We don't assign an index until we actually place the event onto
417 * hardware. Use -1 to signify that we haven't decided where to put it
418 * yet. For SMP systems, each core has it's own PMU so we can't do any
419 * clever allocation or constraints checking at this point.
420 */
421 hwc->idx = -1;
422
423 /*
424 * Store the event encoding into the config_base field. config and
425 * event_base are unused as the only 2 things we need to know are
426 * the event mapping and the counter to use. The counter to use is
427 * also the indx and the config_base is the event type.
428 */
429 hwc->config_base = (unsigned long)mapping;
430 hwc->config = 0;
431 hwc->event_base = 0;
432
433 if (!hwc->sample_period) {
434 hwc->sample_period = armpmu->max_period;
435 hwc->last_period = hwc->sample_period;
436 atomic64_set(&hwc->period_left, hwc->sample_period);
437 }
438
439 err = 0;
440 if (event->group_leader != event) {
441 err = validate_group(event);
442 if (err)
443 return -EINVAL;
444 }
445
446 return err;
447 }
448
449 const struct pmu *
450 hw_perf_event_init(struct perf_event *event)
451 {
452 int err = 0;
453
454 if (!armpmu)
455 return ERR_PTR(-ENODEV);
456
457 event->destroy = hw_perf_event_destroy;
458
459 if (!atomic_inc_not_zero(&active_events)) {
460 if (atomic_read(&active_events) > perf_max_events) {
461 atomic_dec(&active_events);
462 return ERR_PTR(-ENOSPC);
463 }
464
465 mutex_lock(&pmu_reserve_mutex);
466 if (atomic_read(&active_events) == 0) {
467 err = armpmu_reserve_hardware();
468 }
469
470 if (!err)
471 atomic_inc(&active_events);
472 mutex_unlock(&pmu_reserve_mutex);
473 }
474
475 if (err)
476 return ERR_PTR(err);
477
478 err = __hw_perf_event_init(event);
479 if (err)
480 hw_perf_event_destroy(event);
481
482 return err ? ERR_PTR(err) : &pmu;
483 }
484
485 void
486 hw_perf_enable(void)
487 {
488 /* Enable all of the perf events on hardware. */
489 int idx;
490 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
491
492 if (!armpmu)
493 return;
494
495 for (idx = 0; idx <= armpmu->num_events; ++idx) {
496 struct perf_event *event = cpuc->events[idx];
497
498 if (!event)
499 continue;
500
501 armpmu->enable(&event->hw, idx);
502 }
503
504 armpmu->start();
505 }
506
507 void
508 hw_perf_disable(void)
509 {
510 if (armpmu)
511 armpmu->stop();
512 }
513
514 /*
515 * ARMv6 Performance counter handling code.
516 *
517 * ARMv6 has 2 configurable performance counters and a single cycle counter.
518 * They all share a single reset bit but can be written to zero so we can use
519 * that for a reset.
520 *
521 * The counters can't be individually enabled or disabled so when we remove
522 * one event and replace it with another we could get spurious counts from the
523 * wrong event. However, we can take advantage of the fact that the
524 * performance counters can export events to the event bus, and the event bus
525 * itself can be monitored. This requires that we *don't* export the events to
526 * the event bus. The procedure for disabling a configurable counter is:
527 * - change the counter to count the ETMEXTOUT[0] signal (0x20). This
528 * effectively stops the counter from counting.
529 * - disable the counter's interrupt generation (each counter has it's
530 * own interrupt enable bit).
531 * Once stopped, the counter value can be written as 0 to reset.
532 *
533 * To enable a counter:
534 * - enable the counter's interrupt generation.
535 * - set the new event type.
536 *
537 * Note: the dedicated cycle counter only counts cycles and can't be
538 * enabled/disabled independently of the others. When we want to disable the
539 * cycle counter, we have to just disable the interrupt reporting and start
540 * ignoring that counter. When re-enabling, we have to reset the value and
541 * enable the interrupt.
542 */
543
544 enum armv6_perf_types {
545 ARMV6_PERFCTR_ICACHE_MISS = 0x0,
546 ARMV6_PERFCTR_IBUF_STALL = 0x1,
547 ARMV6_PERFCTR_DDEP_STALL = 0x2,
548 ARMV6_PERFCTR_ITLB_MISS = 0x3,
549 ARMV6_PERFCTR_DTLB_MISS = 0x4,
550 ARMV6_PERFCTR_BR_EXEC = 0x5,
551 ARMV6_PERFCTR_BR_MISPREDICT = 0x6,
552 ARMV6_PERFCTR_INSTR_EXEC = 0x7,
553 ARMV6_PERFCTR_DCACHE_HIT = 0x9,
554 ARMV6_PERFCTR_DCACHE_ACCESS = 0xA,
555 ARMV6_PERFCTR_DCACHE_MISS = 0xB,
556 ARMV6_PERFCTR_DCACHE_WBACK = 0xC,
557 ARMV6_PERFCTR_SW_PC_CHANGE = 0xD,
558 ARMV6_PERFCTR_MAIN_TLB_MISS = 0xF,
559 ARMV6_PERFCTR_EXPL_D_ACCESS = 0x10,
560 ARMV6_PERFCTR_LSU_FULL_STALL = 0x11,
561 ARMV6_PERFCTR_WBUF_DRAINED = 0x12,
562 ARMV6_PERFCTR_CPU_CYCLES = 0xFF,
563 ARMV6_PERFCTR_NOP = 0x20,
564 };
565
566 enum armv6_counters {
567 ARMV6_CYCLE_COUNTER = 1,
568 ARMV6_COUNTER0,
569 ARMV6_COUNTER1,
570 };
571
572 /*
573 * The hardware events that we support. We do support cache operations but
574 * we have harvard caches and no way to combine instruction and data
575 * accesses/misses in hardware.
576 */
577 static const unsigned armv6_perf_map[PERF_COUNT_HW_MAX] = {
578 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6_PERFCTR_CPU_CYCLES,
579 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6_PERFCTR_INSTR_EXEC,
580 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
581 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
582 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6_PERFCTR_BR_EXEC,
583 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6_PERFCTR_BR_MISPREDICT,
584 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
585 };
586
587 static const unsigned armv6_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
588 [PERF_COUNT_HW_CACHE_OP_MAX]
589 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
590 [C(L1D)] = {
591 /*
592 * The performance counters don't differentiate between read
593 * and write accesses/misses so this isn't strictly correct,
594 * but it's the best we can do. Writes and reads get
595 * combined.
596 */
597 [C(OP_READ)] = {
598 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
599 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
600 },
601 [C(OP_WRITE)] = {
602 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
603 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
604 },
605 [C(OP_PREFETCH)] = {
606 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
607 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
608 },
609 },
610 [C(L1I)] = {
611 [C(OP_READ)] = {
612 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
613 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
614 },
615 [C(OP_WRITE)] = {
616 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
617 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
618 },
619 [C(OP_PREFETCH)] = {
620 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
621 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
622 },
623 },
624 [C(LL)] = {
625 [C(OP_READ)] = {
626 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
627 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
628 },
629 [C(OP_WRITE)] = {
630 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
631 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
632 },
633 [C(OP_PREFETCH)] = {
634 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
635 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
636 },
637 },
638 [C(DTLB)] = {
639 /*
640 * The ARM performance counters can count micro DTLB misses,
641 * micro ITLB misses and main TLB misses. There isn't an event
642 * for TLB misses, so use the micro misses here and if users
643 * want the main TLB misses they can use a raw counter.
644 */
645 [C(OP_READ)] = {
646 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
647 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
648 },
649 [C(OP_WRITE)] = {
650 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
651 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
652 },
653 [C(OP_PREFETCH)] = {
654 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
655 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
656 },
657 },
658 [C(ITLB)] = {
659 [C(OP_READ)] = {
660 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
661 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
662 },
663 [C(OP_WRITE)] = {
664 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
665 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
666 },
667 [C(OP_PREFETCH)] = {
668 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
669 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
670 },
671 },
672 [C(BPU)] = {
673 [C(OP_READ)] = {
674 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
675 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
676 },
677 [C(OP_WRITE)] = {
678 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
679 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
680 },
681 [C(OP_PREFETCH)] = {
682 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
683 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
684 },
685 },
686 };
687
688 enum armv6mpcore_perf_types {
689 ARMV6MPCORE_PERFCTR_ICACHE_MISS = 0x0,
690 ARMV6MPCORE_PERFCTR_IBUF_STALL = 0x1,
691 ARMV6MPCORE_PERFCTR_DDEP_STALL = 0x2,
692 ARMV6MPCORE_PERFCTR_ITLB_MISS = 0x3,
693 ARMV6MPCORE_PERFCTR_DTLB_MISS = 0x4,
694 ARMV6MPCORE_PERFCTR_BR_EXEC = 0x5,
695 ARMV6MPCORE_PERFCTR_BR_NOTPREDICT = 0x6,
696 ARMV6MPCORE_PERFCTR_BR_MISPREDICT = 0x7,
697 ARMV6MPCORE_PERFCTR_INSTR_EXEC = 0x8,
698 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS = 0xA,
699 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS = 0xB,
700 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS = 0xC,
701 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS = 0xD,
702 ARMV6MPCORE_PERFCTR_DCACHE_EVICTION = 0xE,
703 ARMV6MPCORE_PERFCTR_SW_PC_CHANGE = 0xF,
704 ARMV6MPCORE_PERFCTR_MAIN_TLB_MISS = 0x10,
705 ARMV6MPCORE_PERFCTR_EXPL_MEM_ACCESS = 0x11,
706 ARMV6MPCORE_PERFCTR_LSU_FULL_STALL = 0x12,
707 ARMV6MPCORE_PERFCTR_WBUF_DRAINED = 0x13,
708 ARMV6MPCORE_PERFCTR_CPU_CYCLES = 0xFF,
709 };
710
711 /*
712 * The hardware events that we support. We do support cache operations but
713 * we have harvard caches and no way to combine instruction and data
714 * accesses/misses in hardware.
715 */
716 static const unsigned armv6mpcore_perf_map[PERF_COUNT_HW_MAX] = {
717 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6MPCORE_PERFCTR_CPU_CYCLES,
718 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_INSTR_EXEC,
719 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
720 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
721 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_BR_EXEC,
722 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6MPCORE_PERFCTR_BR_MISPREDICT,
723 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
724 };
725
726 static const unsigned armv6mpcore_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
727 [PERF_COUNT_HW_CACHE_OP_MAX]
728 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
729 [C(L1D)] = {
730 [C(OP_READ)] = {
731 [C(RESULT_ACCESS)] =
732 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS,
733 [C(RESULT_MISS)] =
734 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS,
735 },
736 [C(OP_WRITE)] = {
737 [C(RESULT_ACCESS)] =
738 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS,
739 [C(RESULT_MISS)] =
740 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS,
741 },
742 [C(OP_PREFETCH)] = {
743 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
744 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
745 },
746 },
747 [C(L1I)] = {
748 [C(OP_READ)] = {
749 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
750 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
751 },
752 [C(OP_WRITE)] = {
753 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
754 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
755 },
756 [C(OP_PREFETCH)] = {
757 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
758 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
759 },
760 },
761 [C(LL)] = {
762 [C(OP_READ)] = {
763 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
764 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
765 },
766 [C(OP_WRITE)] = {
767 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
768 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
769 },
770 [C(OP_PREFETCH)] = {
771 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
772 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
773 },
774 },
775 [C(DTLB)] = {
776 /*
777 * The ARM performance counters can count micro DTLB misses,
778 * micro ITLB misses and main TLB misses. There isn't an event
779 * for TLB misses, so use the micro misses here and if users
780 * want the main TLB misses they can use a raw counter.
781 */
782 [C(OP_READ)] = {
783 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
784 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
785 },
786 [C(OP_WRITE)] = {
787 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
788 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
789 },
790 [C(OP_PREFETCH)] = {
791 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
792 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
793 },
794 },
795 [C(ITLB)] = {
796 [C(OP_READ)] = {
797 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
798 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
799 },
800 [C(OP_WRITE)] = {
801 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
802 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
803 },
804 [C(OP_PREFETCH)] = {
805 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
806 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
807 },
808 },
809 [C(BPU)] = {
810 [C(OP_READ)] = {
811 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
812 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
813 },
814 [C(OP_WRITE)] = {
815 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
816 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
817 },
818 [C(OP_PREFETCH)] = {
819 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
820 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
821 },
822 },
823 };
824
825 static inline unsigned long
826 armv6_pmcr_read(void)
827 {
828 u32 val;
829 asm volatile("mrc p15, 0, %0, c15, c12, 0" : "=r"(val));
830 return val;
831 }
832
833 static inline void
834 armv6_pmcr_write(unsigned long val)
835 {
836 asm volatile("mcr p15, 0, %0, c15, c12, 0" : : "r"(val));
837 }
838
839 #define ARMV6_PMCR_ENABLE (1 << 0)
840 #define ARMV6_PMCR_CTR01_RESET (1 << 1)
841 #define ARMV6_PMCR_CCOUNT_RESET (1 << 2)
842 #define ARMV6_PMCR_CCOUNT_DIV (1 << 3)
843 #define ARMV6_PMCR_COUNT0_IEN (1 << 4)
844 #define ARMV6_PMCR_COUNT1_IEN (1 << 5)
845 #define ARMV6_PMCR_CCOUNT_IEN (1 << 6)
846 #define ARMV6_PMCR_COUNT0_OVERFLOW (1 << 8)
847 #define ARMV6_PMCR_COUNT1_OVERFLOW (1 << 9)
848 #define ARMV6_PMCR_CCOUNT_OVERFLOW (1 << 10)
849 #define ARMV6_PMCR_EVT_COUNT0_SHIFT 20
850 #define ARMV6_PMCR_EVT_COUNT0_MASK (0xFF << ARMV6_PMCR_EVT_COUNT0_SHIFT)
851 #define ARMV6_PMCR_EVT_COUNT1_SHIFT 12
852 #define ARMV6_PMCR_EVT_COUNT1_MASK (0xFF << ARMV6_PMCR_EVT_COUNT1_SHIFT)
853
854 #define ARMV6_PMCR_OVERFLOWED_MASK \
855 (ARMV6_PMCR_COUNT0_OVERFLOW | ARMV6_PMCR_COUNT1_OVERFLOW | \
856 ARMV6_PMCR_CCOUNT_OVERFLOW)
857
858 static inline int
859 armv6_pmcr_has_overflowed(unsigned long pmcr)
860 {
861 return (pmcr & ARMV6_PMCR_OVERFLOWED_MASK);
862 }
863
864 static inline int
865 armv6_pmcr_counter_has_overflowed(unsigned long pmcr,
866 enum armv6_counters counter)
867 {
868 int ret = 0;
869
870 if (ARMV6_CYCLE_COUNTER == counter)
871 ret = pmcr & ARMV6_PMCR_CCOUNT_OVERFLOW;
872 else if (ARMV6_COUNTER0 == counter)
873 ret = pmcr & ARMV6_PMCR_COUNT0_OVERFLOW;
874 else if (ARMV6_COUNTER1 == counter)
875 ret = pmcr & ARMV6_PMCR_COUNT1_OVERFLOW;
876 else
877 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
878
879 return ret;
880 }
881
882 static inline u32
883 armv6pmu_read_counter(int counter)
884 {
885 unsigned long value = 0;
886
887 if (ARMV6_CYCLE_COUNTER == counter)
888 asm volatile("mrc p15, 0, %0, c15, c12, 1" : "=r"(value));
889 else if (ARMV6_COUNTER0 == counter)
890 asm volatile("mrc p15, 0, %0, c15, c12, 2" : "=r"(value));
891 else if (ARMV6_COUNTER1 == counter)
892 asm volatile("mrc p15, 0, %0, c15, c12, 3" : "=r"(value));
893 else
894 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
895
896 return value;
897 }
898
899 static inline void
900 armv6pmu_write_counter(int counter,
901 u32 value)
902 {
903 if (ARMV6_CYCLE_COUNTER == counter)
904 asm volatile("mcr p15, 0, %0, c15, c12, 1" : : "r"(value));
905 else if (ARMV6_COUNTER0 == counter)
906 asm volatile("mcr p15, 0, %0, c15, c12, 2" : : "r"(value));
907 else if (ARMV6_COUNTER1 == counter)
908 asm volatile("mcr p15, 0, %0, c15, c12, 3" : : "r"(value));
909 else
910 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
911 }
912
913 void
914 armv6pmu_enable_event(struct hw_perf_event *hwc,
915 int idx)
916 {
917 unsigned long val, mask, evt, flags;
918
919 if (ARMV6_CYCLE_COUNTER == idx) {
920 mask = 0;
921 evt = ARMV6_PMCR_CCOUNT_IEN;
922 } else if (ARMV6_COUNTER0 == idx) {
923 mask = ARMV6_PMCR_EVT_COUNT0_MASK;
924 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT0_SHIFT) |
925 ARMV6_PMCR_COUNT0_IEN;
926 } else if (ARMV6_COUNTER1 == idx) {
927 mask = ARMV6_PMCR_EVT_COUNT1_MASK;
928 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT1_SHIFT) |
929 ARMV6_PMCR_COUNT1_IEN;
930 } else {
931 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
932 return;
933 }
934
935 /*
936 * Mask out the current event and set the counter to count the event
937 * that we're interested in.
938 */
939 spin_lock_irqsave(&pmu_lock, flags);
940 val = armv6_pmcr_read();
941 val &= ~mask;
942 val |= evt;
943 armv6_pmcr_write(val);
944 spin_unlock_irqrestore(&pmu_lock, flags);
945 }
946
947 static irqreturn_t
948 armv6pmu_handle_irq(int irq_num,
949 void *dev)
950 {
951 unsigned long pmcr = armv6_pmcr_read();
952 struct perf_sample_data data;
953 struct cpu_hw_events *cpuc;
954 struct pt_regs *regs;
955 int idx;
956
957 if (!armv6_pmcr_has_overflowed(pmcr))
958 return IRQ_NONE;
959
960 regs = get_irq_regs();
961
962 /*
963 * The interrupts are cleared by writing the overflow flags back to
964 * the control register. All of the other bits don't have any effect
965 * if they are rewritten, so write the whole value back.
966 */
967 armv6_pmcr_write(pmcr);
968
969 perf_sample_data_init(&data, 0);
970
971 cpuc = &__get_cpu_var(cpu_hw_events);
972 for (idx = 0; idx <= armpmu->num_events; ++idx) {
973 struct perf_event *event = cpuc->events[idx];
974 struct hw_perf_event *hwc;
975
976 if (!test_bit(idx, cpuc->active_mask))
977 continue;
978
979 /*
980 * We have a single interrupt for all counters. Check that
981 * each counter has overflowed before we process it.
982 */
983 if (!armv6_pmcr_counter_has_overflowed(pmcr, idx))
984 continue;
985
986 hwc = &event->hw;
987 armpmu_event_update(event, hwc, idx);
988 data.period = event->hw.last_period;
989 if (!armpmu_event_set_period(event, hwc, idx))
990 continue;
991
992 if (perf_event_overflow(event, 0, &data, regs))
993 armpmu->disable(hwc, idx);
994 }
995
996 /*
997 * Handle the pending perf events.
998 *
999 * Note: this call *must* be run with interrupts enabled. For
1000 * platforms that can have the PMU interrupts raised as a PMI, this
1001 * will not work.
1002 */
1003 perf_event_do_pending();
1004
1005 return IRQ_HANDLED;
1006 }
1007
1008 static void
1009 armv6pmu_start(void)
1010 {
1011 unsigned long flags, val;
1012
1013 spin_lock_irqsave(&pmu_lock, flags);
1014 val = armv6_pmcr_read();
1015 val |= ARMV6_PMCR_ENABLE;
1016 armv6_pmcr_write(val);
1017 spin_unlock_irqrestore(&pmu_lock, flags);
1018 }
1019
1020 void
1021 armv6pmu_stop(void)
1022 {
1023 unsigned long flags, val;
1024
1025 spin_lock_irqsave(&pmu_lock, flags);
1026 val = armv6_pmcr_read();
1027 val &= ~ARMV6_PMCR_ENABLE;
1028 armv6_pmcr_write(val);
1029 spin_unlock_irqrestore(&pmu_lock, flags);
1030 }
1031
1032 static inline int
1033 armv6pmu_event_map(int config)
1034 {
1035 int mapping = armv6_perf_map[config];
1036 if (HW_OP_UNSUPPORTED == mapping)
1037 mapping = -EOPNOTSUPP;
1038 return mapping;
1039 }
1040
1041 static inline int
1042 armv6mpcore_pmu_event_map(int config)
1043 {
1044 int mapping = armv6mpcore_perf_map[config];
1045 if (HW_OP_UNSUPPORTED == mapping)
1046 mapping = -EOPNOTSUPP;
1047 return mapping;
1048 }
1049
1050 static u64
1051 armv6pmu_raw_event(u64 config)
1052 {
1053 return config & 0xff;
1054 }
1055
1056 static int
1057 armv6pmu_get_event_idx(struct cpu_hw_events *cpuc,
1058 struct hw_perf_event *event)
1059 {
1060 /* Always place a cycle counter into the cycle counter. */
1061 if (ARMV6_PERFCTR_CPU_CYCLES == event->config_base) {
1062 if (test_and_set_bit(ARMV6_CYCLE_COUNTER, cpuc->used_mask))
1063 return -EAGAIN;
1064
1065 return ARMV6_CYCLE_COUNTER;
1066 } else {
1067 /*
1068 * For anything other than a cycle counter, try and use
1069 * counter0 and counter1.
1070 */
1071 if (!test_and_set_bit(ARMV6_COUNTER1, cpuc->used_mask)) {
1072 return ARMV6_COUNTER1;
1073 }
1074
1075 if (!test_and_set_bit(ARMV6_COUNTER0, cpuc->used_mask)) {
1076 return ARMV6_COUNTER0;
1077 }
1078
1079 /* The counters are all in use. */
1080 return -EAGAIN;
1081 }
1082 }
1083
1084 static void
1085 armv6pmu_disable_event(struct hw_perf_event *hwc,
1086 int idx)
1087 {
1088 unsigned long val, mask, evt, flags;
1089
1090 if (ARMV6_CYCLE_COUNTER == idx) {
1091 mask = ARMV6_PMCR_CCOUNT_IEN;
1092 evt = 0;
1093 } else if (ARMV6_COUNTER0 == idx) {
1094 mask = ARMV6_PMCR_COUNT0_IEN | ARMV6_PMCR_EVT_COUNT0_MASK;
1095 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT0_SHIFT;
1096 } else if (ARMV6_COUNTER1 == idx) {
1097 mask = ARMV6_PMCR_COUNT1_IEN | ARMV6_PMCR_EVT_COUNT1_MASK;
1098 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT1_SHIFT;
1099 } else {
1100 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
1101 return;
1102 }
1103
1104 /*
1105 * Mask out the current event and set the counter to count the number
1106 * of ETM bus signal assertion cycles. The external reporting should
1107 * be disabled and so this should never increment.
1108 */
1109 spin_lock_irqsave(&pmu_lock, flags);
1110 val = armv6_pmcr_read();
1111 val &= ~mask;
1112 val |= evt;
1113 armv6_pmcr_write(val);
1114 spin_unlock_irqrestore(&pmu_lock, flags);
1115 }
1116
1117 static void
1118 armv6mpcore_pmu_disable_event(struct hw_perf_event *hwc,
1119 int idx)
1120 {
1121 unsigned long val, mask, flags, evt = 0;
1122
1123 if (ARMV6_CYCLE_COUNTER == idx) {
1124 mask = ARMV6_PMCR_CCOUNT_IEN;
1125 } else if (ARMV6_COUNTER0 == idx) {
1126 mask = ARMV6_PMCR_COUNT0_IEN;
1127 } else if (ARMV6_COUNTER1 == idx) {
1128 mask = ARMV6_PMCR_COUNT1_IEN;
1129 } else {
1130 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
1131 return;
1132 }
1133
1134 /*
1135 * Unlike UP ARMv6, we don't have a way of stopping the counters. We
1136 * simply disable the interrupt reporting.
1137 */
1138 spin_lock_irqsave(&pmu_lock, flags);
1139 val = armv6_pmcr_read();
1140 val &= ~mask;
1141 val |= evt;
1142 armv6_pmcr_write(val);
1143 spin_unlock_irqrestore(&pmu_lock, flags);
1144 }
1145
1146 static const struct arm_pmu armv6pmu = {
1147 .name = "v6",
1148 .handle_irq = armv6pmu_handle_irq,
1149 .enable = armv6pmu_enable_event,
1150 .disable = armv6pmu_disable_event,
1151 .event_map = armv6pmu_event_map,
1152 .raw_event = armv6pmu_raw_event,
1153 .read_counter = armv6pmu_read_counter,
1154 .write_counter = armv6pmu_write_counter,
1155 .get_event_idx = armv6pmu_get_event_idx,
1156 .start = armv6pmu_start,
1157 .stop = armv6pmu_stop,
1158 .num_events = 3,
1159 .max_period = (1LLU << 32) - 1,
1160 };
1161
1162 /*
1163 * ARMv6mpcore is almost identical to single core ARMv6 with the exception
1164 * that some of the events have different enumerations and that there is no
1165 * *hack* to stop the programmable counters. To stop the counters we simply
1166 * disable the interrupt reporting and update the event. When unthrottling we
1167 * reset the period and enable the interrupt reporting.
1168 */
1169 static const struct arm_pmu armv6mpcore_pmu = {
1170 .name = "v6mpcore",
1171 .handle_irq = armv6pmu_handle_irq,
1172 .enable = armv6pmu_enable_event,
1173 .disable = armv6mpcore_pmu_disable_event,
1174 .event_map = armv6mpcore_pmu_event_map,
1175 .raw_event = armv6pmu_raw_event,
1176 .read_counter = armv6pmu_read_counter,
1177 .write_counter = armv6pmu_write_counter,
1178 .get_event_idx = armv6pmu_get_event_idx,
1179 .start = armv6pmu_start,
1180 .stop = armv6pmu_stop,
1181 .num_events = 3,
1182 .max_period = (1LLU << 32) - 1,
1183 };
1184
1185 /*
1186 * ARMv7 Cortex-A8 and Cortex-A9 Performance Events handling code.
1187 *
1188 * Copied from ARMv6 code, with the low level code inspired
1189 * by the ARMv7 Oprofile code.
1190 *
1191 * Cortex-A8 has up to 4 configurable performance counters and
1192 * a single cycle counter.
1193 * Cortex-A9 has up to 31 configurable performance counters and
1194 * a single cycle counter.
1195 *
1196 * All counters can be enabled/disabled and IRQ masked separately. The cycle
1197 * counter and all 4 performance counters together can be reset separately.
1198 */
1199
1200 #define ARMV7_PMU_CORTEX_A8_NAME "ARMv7 Cortex-A8"
1201
1202 #define ARMV7_PMU_CORTEX_A9_NAME "ARMv7 Cortex-A9"
1203
1204 /* Common ARMv7 event types */
1205 enum armv7_perf_types {
1206 ARMV7_PERFCTR_PMNC_SW_INCR = 0x00,
1207 ARMV7_PERFCTR_IFETCH_MISS = 0x01,
1208 ARMV7_PERFCTR_ITLB_MISS = 0x02,
1209 ARMV7_PERFCTR_DCACHE_REFILL = 0x03,
1210 ARMV7_PERFCTR_DCACHE_ACCESS = 0x04,
1211 ARMV7_PERFCTR_DTLB_REFILL = 0x05,
1212 ARMV7_PERFCTR_DREAD = 0x06,
1213 ARMV7_PERFCTR_DWRITE = 0x07,
1214
1215 ARMV7_PERFCTR_EXC_TAKEN = 0x09,
1216 ARMV7_PERFCTR_EXC_EXECUTED = 0x0A,
1217 ARMV7_PERFCTR_CID_WRITE = 0x0B,
1218 /* ARMV7_PERFCTR_PC_WRITE is equivalent to HW_BRANCH_INSTRUCTIONS.
1219 * It counts:
1220 * - all branch instructions,
1221 * - instructions that explicitly write the PC,
1222 * - exception generating instructions.
1223 */
1224 ARMV7_PERFCTR_PC_WRITE = 0x0C,
1225 ARMV7_PERFCTR_PC_IMM_BRANCH = 0x0D,
1226 ARMV7_PERFCTR_UNALIGNED_ACCESS = 0x0F,
1227 ARMV7_PERFCTR_PC_BRANCH_MIS_PRED = 0x10,
1228 ARMV7_PERFCTR_CLOCK_CYCLES = 0x11,
1229
1230 ARMV7_PERFCTR_PC_BRANCH_MIS_USED = 0x12,
1231
1232 ARMV7_PERFCTR_CPU_CYCLES = 0xFF
1233 };
1234
1235 /* ARMv7 Cortex-A8 specific event types */
1236 enum armv7_a8_perf_types {
1237 ARMV7_PERFCTR_INSTR_EXECUTED = 0x08,
1238
1239 ARMV7_PERFCTR_PC_PROC_RETURN = 0x0E,
1240
1241 ARMV7_PERFCTR_WRITE_BUFFER_FULL = 0x40,
1242 ARMV7_PERFCTR_L2_STORE_MERGED = 0x41,
1243 ARMV7_PERFCTR_L2_STORE_BUFF = 0x42,
1244 ARMV7_PERFCTR_L2_ACCESS = 0x43,
1245 ARMV7_PERFCTR_L2_CACH_MISS = 0x44,
1246 ARMV7_PERFCTR_AXI_READ_CYCLES = 0x45,
1247 ARMV7_PERFCTR_AXI_WRITE_CYCLES = 0x46,
1248 ARMV7_PERFCTR_MEMORY_REPLAY = 0x47,
1249 ARMV7_PERFCTR_UNALIGNED_ACCESS_REPLAY = 0x48,
1250 ARMV7_PERFCTR_L1_DATA_MISS = 0x49,
1251 ARMV7_PERFCTR_L1_INST_MISS = 0x4A,
1252 ARMV7_PERFCTR_L1_DATA_COLORING = 0x4B,
1253 ARMV7_PERFCTR_L1_NEON_DATA = 0x4C,
1254 ARMV7_PERFCTR_L1_NEON_CACH_DATA = 0x4D,
1255 ARMV7_PERFCTR_L2_NEON = 0x4E,
1256 ARMV7_PERFCTR_L2_NEON_HIT = 0x4F,
1257 ARMV7_PERFCTR_L1_INST = 0x50,
1258 ARMV7_PERFCTR_PC_RETURN_MIS_PRED = 0x51,
1259 ARMV7_PERFCTR_PC_BRANCH_FAILED = 0x52,
1260 ARMV7_PERFCTR_PC_BRANCH_TAKEN = 0x53,
1261 ARMV7_PERFCTR_PC_BRANCH_EXECUTED = 0x54,
1262 ARMV7_PERFCTR_OP_EXECUTED = 0x55,
1263 ARMV7_PERFCTR_CYCLES_INST_STALL = 0x56,
1264 ARMV7_PERFCTR_CYCLES_INST = 0x57,
1265 ARMV7_PERFCTR_CYCLES_NEON_DATA_STALL = 0x58,
1266 ARMV7_PERFCTR_CYCLES_NEON_INST_STALL = 0x59,
1267 ARMV7_PERFCTR_NEON_CYCLES = 0x5A,
1268
1269 ARMV7_PERFCTR_PMU0_EVENTS = 0x70,
1270 ARMV7_PERFCTR_PMU1_EVENTS = 0x71,
1271 ARMV7_PERFCTR_PMU_EVENTS = 0x72,
1272 };
1273
1274 /* ARMv7 Cortex-A9 specific event types */
1275 enum armv7_a9_perf_types {
1276 ARMV7_PERFCTR_JAVA_HW_BYTECODE_EXEC = 0x40,
1277 ARMV7_PERFCTR_JAVA_SW_BYTECODE_EXEC = 0x41,
1278 ARMV7_PERFCTR_JAZELLE_BRANCH_EXEC = 0x42,
1279
1280 ARMV7_PERFCTR_COHERENT_LINE_MISS = 0x50,
1281 ARMV7_PERFCTR_COHERENT_LINE_HIT = 0x51,
1282
1283 ARMV7_PERFCTR_ICACHE_DEP_STALL_CYCLES = 0x60,
1284 ARMV7_PERFCTR_DCACHE_DEP_STALL_CYCLES = 0x61,
1285 ARMV7_PERFCTR_TLB_MISS_DEP_STALL_CYCLES = 0x62,
1286 ARMV7_PERFCTR_STREX_EXECUTED_PASSED = 0x63,
1287 ARMV7_PERFCTR_STREX_EXECUTED_FAILED = 0x64,
1288 ARMV7_PERFCTR_DATA_EVICTION = 0x65,
1289 ARMV7_PERFCTR_ISSUE_STAGE_NO_INST = 0x66,
1290 ARMV7_PERFCTR_ISSUE_STAGE_EMPTY = 0x67,
1291 ARMV7_PERFCTR_INST_OUT_OF_RENAME_STAGE = 0x68,
1292
1293 ARMV7_PERFCTR_PREDICTABLE_FUNCT_RETURNS = 0x6E,
1294
1295 ARMV7_PERFCTR_MAIN_UNIT_EXECUTED_INST = 0x70,
1296 ARMV7_PERFCTR_SECOND_UNIT_EXECUTED_INST = 0x71,
1297 ARMV7_PERFCTR_LD_ST_UNIT_EXECUTED_INST = 0x72,
1298 ARMV7_PERFCTR_FP_EXECUTED_INST = 0x73,
1299 ARMV7_PERFCTR_NEON_EXECUTED_INST = 0x74,
1300
1301 ARMV7_PERFCTR_PLD_FULL_DEP_STALL_CYCLES = 0x80,
1302 ARMV7_PERFCTR_DATA_WR_DEP_STALL_CYCLES = 0x81,
1303 ARMV7_PERFCTR_ITLB_MISS_DEP_STALL_CYCLES = 0x82,
1304 ARMV7_PERFCTR_DTLB_MISS_DEP_STALL_CYCLES = 0x83,
1305 ARMV7_PERFCTR_MICRO_ITLB_MISS_DEP_STALL_CYCLES = 0x84,
1306 ARMV7_PERFCTR_MICRO_DTLB_MISS_DEP_STALL_CYCLES = 0x85,
1307 ARMV7_PERFCTR_DMB_DEP_STALL_CYCLES = 0x86,
1308
1309 ARMV7_PERFCTR_INTGR_CLK_ENABLED_CYCLES = 0x8A,
1310 ARMV7_PERFCTR_DATA_ENGINE_CLK_EN_CYCLES = 0x8B,
1311
1312 ARMV7_PERFCTR_ISB_INST = 0x90,
1313 ARMV7_PERFCTR_DSB_INST = 0x91,
1314 ARMV7_PERFCTR_DMB_INST = 0x92,
1315 ARMV7_PERFCTR_EXT_INTERRUPTS = 0x93,
1316
1317 ARMV7_PERFCTR_PLE_CACHE_LINE_RQST_COMPLETED = 0xA0,
1318 ARMV7_PERFCTR_PLE_CACHE_LINE_RQST_SKIPPED = 0xA1,
1319 ARMV7_PERFCTR_PLE_FIFO_FLUSH = 0xA2,
1320 ARMV7_PERFCTR_PLE_RQST_COMPLETED = 0xA3,
1321 ARMV7_PERFCTR_PLE_FIFO_OVERFLOW = 0xA4,
1322 ARMV7_PERFCTR_PLE_RQST_PROG = 0xA5
1323 };
1324
1325 /*
1326 * Cortex-A8 HW events mapping
1327 *
1328 * The hardware events that we support. We do support cache operations but
1329 * we have harvard caches and no way to combine instruction and data
1330 * accesses/misses in hardware.
1331 */
1332 static const unsigned armv7_a8_perf_map[PERF_COUNT_HW_MAX] = {
1333 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
1334 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
1335 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
1336 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
1337 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
1338 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1339 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
1340 };
1341
1342 static const unsigned armv7_a8_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
1343 [PERF_COUNT_HW_CACHE_OP_MAX]
1344 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1345 [C(L1D)] = {
1346 /*
1347 * The performance counters don't differentiate between read
1348 * and write accesses/misses so this isn't strictly correct,
1349 * but it's the best we can do. Writes and reads get
1350 * combined.
1351 */
1352 [C(OP_READ)] = {
1353 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1354 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1355 },
1356 [C(OP_WRITE)] = {
1357 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1358 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1359 },
1360 [C(OP_PREFETCH)] = {
1361 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1362 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1363 },
1364 },
1365 [C(L1I)] = {
1366 [C(OP_READ)] = {
1367 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_INST,
1368 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_INST_MISS,
1369 },
1370 [C(OP_WRITE)] = {
1371 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_INST,
1372 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_INST_MISS,
1373 },
1374 [C(OP_PREFETCH)] = {
1375 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1376 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1377 },
1378 },
1379 [C(LL)] = {
1380 [C(OP_READ)] = {
1381 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_ACCESS,
1382 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACH_MISS,
1383 },
1384 [C(OP_WRITE)] = {
1385 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_ACCESS,
1386 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACH_MISS,
1387 },
1388 [C(OP_PREFETCH)] = {
1389 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1390 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1391 },
1392 },
1393 [C(DTLB)] = {
1394 /*
1395 * Only ITLB misses and DTLB refills are supported.
1396 * If users want the DTLB refills misses a raw counter
1397 * must be used.
1398 */
1399 [C(OP_READ)] = {
1400 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1401 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1402 },
1403 [C(OP_WRITE)] = {
1404 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1405 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1406 },
1407 [C(OP_PREFETCH)] = {
1408 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1409 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1410 },
1411 },
1412 [C(ITLB)] = {
1413 [C(OP_READ)] = {
1414 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1415 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1416 },
1417 [C(OP_WRITE)] = {
1418 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1419 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1420 },
1421 [C(OP_PREFETCH)] = {
1422 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1423 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1424 },
1425 },
1426 [C(BPU)] = {
1427 [C(OP_READ)] = {
1428 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1429 [C(RESULT_MISS)]
1430 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1431 },
1432 [C(OP_WRITE)] = {
1433 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1434 [C(RESULT_MISS)]
1435 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1436 },
1437 [C(OP_PREFETCH)] = {
1438 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1439 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1440 },
1441 },
1442 };
1443
1444 /*
1445 * Cortex-A9 HW events mapping
1446 */
1447 static const unsigned armv7_a9_perf_map[PERF_COUNT_HW_MAX] = {
1448 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
1449 [PERF_COUNT_HW_INSTRUCTIONS] =
1450 ARMV7_PERFCTR_INST_OUT_OF_RENAME_STAGE,
1451 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_COHERENT_LINE_HIT,
1452 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_COHERENT_LINE_MISS,
1453 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
1454 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1455 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
1456 };
1457
1458 static const unsigned armv7_a9_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
1459 [PERF_COUNT_HW_CACHE_OP_MAX]
1460 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1461 [C(L1D)] = {
1462 /*
1463 * The performance counters don't differentiate between read
1464 * and write accesses/misses so this isn't strictly correct,
1465 * but it's the best we can do. Writes and reads get
1466 * combined.
1467 */
1468 [C(OP_READ)] = {
1469 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1470 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1471 },
1472 [C(OP_WRITE)] = {
1473 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1474 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1475 },
1476 [C(OP_PREFETCH)] = {
1477 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1478 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1479 },
1480 },
1481 [C(L1I)] = {
1482 [C(OP_READ)] = {
1483 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1484 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
1485 },
1486 [C(OP_WRITE)] = {
1487 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1488 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
1489 },
1490 [C(OP_PREFETCH)] = {
1491 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1492 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1493 },
1494 },
1495 [C(LL)] = {
1496 [C(OP_READ)] = {
1497 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1498 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1499 },
1500 [C(OP_WRITE)] = {
1501 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1502 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1503 },
1504 [C(OP_PREFETCH)] = {
1505 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1506 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1507 },
1508 },
1509 [C(DTLB)] = {
1510 /*
1511 * Only ITLB misses and DTLB refills are supported.
1512 * If users want the DTLB refills misses a raw counter
1513 * must be used.
1514 */
1515 [C(OP_READ)] = {
1516 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1517 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1518 },
1519 [C(OP_WRITE)] = {
1520 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1521 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1522 },
1523 [C(OP_PREFETCH)] = {
1524 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1525 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1526 },
1527 },
1528 [C(ITLB)] = {
1529 [C(OP_READ)] = {
1530 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1531 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1532 },
1533 [C(OP_WRITE)] = {
1534 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1535 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1536 },
1537 [C(OP_PREFETCH)] = {
1538 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1539 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1540 },
1541 },
1542 [C(BPU)] = {
1543 [C(OP_READ)] = {
1544 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1545 [C(RESULT_MISS)]
1546 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1547 },
1548 [C(OP_WRITE)] = {
1549 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1550 [C(RESULT_MISS)]
1551 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1552 },
1553 [C(OP_PREFETCH)] = {
1554 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1555 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1556 },
1557 },
1558 };
1559
1560 /*
1561 * Perf Events counters
1562 */
1563 enum armv7_counters {
1564 ARMV7_CYCLE_COUNTER = 1, /* Cycle counter */
1565 ARMV7_COUNTER0 = 2, /* First event counter */
1566 };
1567
1568 /*
1569 * The cycle counter is ARMV7_CYCLE_COUNTER.
1570 * The first event counter is ARMV7_COUNTER0.
1571 * The last event counter is (ARMV7_COUNTER0 + armpmu->num_events - 1).
1572 */
1573 #define ARMV7_COUNTER_LAST (ARMV7_COUNTER0 + armpmu->num_events - 1)
1574
1575 /*
1576 * ARMv7 low level PMNC access
1577 */
1578
1579 /*
1580 * Per-CPU PMNC: config reg
1581 */
1582 #define ARMV7_PMNC_E (1 << 0) /* Enable all counters */
1583 #define ARMV7_PMNC_P (1 << 1) /* Reset all counters */
1584 #define ARMV7_PMNC_C (1 << 2) /* Cycle counter reset */
1585 #define ARMV7_PMNC_D (1 << 3) /* CCNT counts every 64th cpu cycle */
1586 #define ARMV7_PMNC_X (1 << 4) /* Export to ETM */
1587 #define ARMV7_PMNC_DP (1 << 5) /* Disable CCNT if non-invasive debug*/
1588 #define ARMV7_PMNC_N_SHIFT 11 /* Number of counters supported */
1589 #define ARMV7_PMNC_N_MASK 0x1f
1590 #define ARMV7_PMNC_MASK 0x3f /* Mask for writable bits */
1591
1592 /*
1593 * Available counters
1594 */
1595 #define ARMV7_CNT0 0 /* First event counter */
1596 #define ARMV7_CCNT 31 /* Cycle counter */
1597
1598 /* Perf Event to low level counters mapping */
1599 #define ARMV7_EVENT_CNT_TO_CNTx (ARMV7_COUNTER0 - ARMV7_CNT0)
1600
1601 /*
1602 * CNTENS: counters enable reg
1603 */
1604 #define ARMV7_CNTENS_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1605 #define ARMV7_CNTENS_C (1 << ARMV7_CCNT)
1606
1607 /*
1608 * CNTENC: counters disable reg
1609 */
1610 #define ARMV7_CNTENC_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1611 #define ARMV7_CNTENC_C (1 << ARMV7_CCNT)
1612
1613 /*
1614 * INTENS: counters overflow interrupt enable reg
1615 */
1616 #define ARMV7_INTENS_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1617 #define ARMV7_INTENS_C (1 << ARMV7_CCNT)
1618
1619 /*
1620 * INTENC: counters overflow interrupt disable reg
1621 */
1622 #define ARMV7_INTENC_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1623 #define ARMV7_INTENC_C (1 << ARMV7_CCNT)
1624
1625 /*
1626 * EVTSEL: Event selection reg
1627 */
1628 #define ARMV7_EVTSEL_MASK 0xff /* Mask for writable bits */
1629
1630 /*
1631 * SELECT: Counter selection reg
1632 */
1633 #define ARMV7_SELECT_MASK 0x1f /* Mask for writable bits */
1634
1635 /*
1636 * FLAG: counters overflow flag status reg
1637 */
1638 #define ARMV7_FLAG_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1639 #define ARMV7_FLAG_C (1 << ARMV7_CCNT)
1640 #define ARMV7_FLAG_MASK 0xffffffff /* Mask for writable bits */
1641 #define ARMV7_OVERFLOWED_MASK ARMV7_FLAG_MASK
1642
1643 static inline unsigned long armv7_pmnc_read(void)
1644 {
1645 u32 val;
1646 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(val));
1647 return val;
1648 }
1649
1650 static inline void armv7_pmnc_write(unsigned long val)
1651 {
1652 val &= ARMV7_PMNC_MASK;
1653 asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r"(val));
1654 }
1655
1656 static inline int armv7_pmnc_has_overflowed(unsigned long pmnc)
1657 {
1658 return pmnc & ARMV7_OVERFLOWED_MASK;
1659 }
1660
1661 static inline int armv7_pmnc_counter_has_overflowed(unsigned long pmnc,
1662 enum armv7_counters counter)
1663 {
1664 int ret;
1665
1666 if (counter == ARMV7_CYCLE_COUNTER)
1667 ret = pmnc & ARMV7_FLAG_C;
1668 else if ((counter >= ARMV7_COUNTER0) && (counter <= ARMV7_COUNTER_LAST))
1669 ret = pmnc & ARMV7_FLAG_P(counter);
1670 else
1671 pr_err("CPU%u checking wrong counter %d overflow status\n",
1672 smp_processor_id(), counter);
1673
1674 return ret;
1675 }
1676
1677 static inline int armv7_pmnc_select_counter(unsigned int idx)
1678 {
1679 u32 val;
1680
1681 if ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST)) {
1682 pr_err("CPU%u selecting wrong PMNC counter"
1683 " %d\n", smp_processor_id(), idx);
1684 return -1;
1685 }
1686
1687 val = (idx - ARMV7_EVENT_CNT_TO_CNTx) & ARMV7_SELECT_MASK;
1688 asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (val));
1689
1690 return idx;
1691 }
1692
1693 static inline u32 armv7pmu_read_counter(int idx)
1694 {
1695 unsigned long value = 0;
1696
1697 if (idx == ARMV7_CYCLE_COUNTER)
1698 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (value));
1699 else if ((idx >= ARMV7_COUNTER0) && (idx <= ARMV7_COUNTER_LAST)) {
1700 if (armv7_pmnc_select_counter(idx) == idx)
1701 asm volatile("mrc p15, 0, %0, c9, c13, 2"
1702 : "=r" (value));
1703 } else
1704 pr_err("CPU%u reading wrong counter %d\n",
1705 smp_processor_id(), idx);
1706
1707 return value;
1708 }
1709
1710 static inline void armv7pmu_write_counter(int idx, u32 value)
1711 {
1712 if (idx == ARMV7_CYCLE_COUNTER)
1713 asm volatile("mcr p15, 0, %0, c9, c13, 0" : : "r" (value));
1714 else if ((idx >= ARMV7_COUNTER0) && (idx <= ARMV7_COUNTER_LAST)) {
1715 if (armv7_pmnc_select_counter(idx) == idx)
1716 asm volatile("mcr p15, 0, %0, c9, c13, 2"
1717 : : "r" (value));
1718 } else
1719 pr_err("CPU%u writing wrong counter %d\n",
1720 smp_processor_id(), idx);
1721 }
1722
1723 static inline void armv7_pmnc_write_evtsel(unsigned int idx, u32 val)
1724 {
1725 if (armv7_pmnc_select_counter(idx) == idx) {
1726 val &= ARMV7_EVTSEL_MASK;
1727 asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (val));
1728 }
1729 }
1730
1731 static inline u32 armv7_pmnc_enable_counter(unsigned int idx)
1732 {
1733 u32 val;
1734
1735 if ((idx != ARMV7_CYCLE_COUNTER) &&
1736 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1737 pr_err("CPU%u enabling wrong PMNC counter"
1738 " %d\n", smp_processor_id(), idx);
1739 return -1;
1740 }
1741
1742 if (idx == ARMV7_CYCLE_COUNTER)
1743 val = ARMV7_CNTENS_C;
1744 else
1745 val = ARMV7_CNTENS_P(idx);
1746
1747 asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (val));
1748
1749 return idx;
1750 }
1751
1752 static inline u32 armv7_pmnc_disable_counter(unsigned int idx)
1753 {
1754 u32 val;
1755
1756
1757 if ((idx != ARMV7_CYCLE_COUNTER) &&
1758 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1759 pr_err("CPU%u disabling wrong PMNC counter"
1760 " %d\n", smp_processor_id(), idx);
1761 return -1;
1762 }
1763
1764 if (idx == ARMV7_CYCLE_COUNTER)
1765 val = ARMV7_CNTENC_C;
1766 else
1767 val = ARMV7_CNTENC_P(idx);
1768
1769 asm volatile("mcr p15, 0, %0, c9, c12, 2" : : "r" (val));
1770
1771 return idx;
1772 }
1773
1774 static inline u32 armv7_pmnc_enable_intens(unsigned int idx)
1775 {
1776 u32 val;
1777
1778 if ((idx != ARMV7_CYCLE_COUNTER) &&
1779 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1780 pr_err("CPU%u enabling wrong PMNC counter"
1781 " interrupt enable %d\n", smp_processor_id(), idx);
1782 return -1;
1783 }
1784
1785 if (idx == ARMV7_CYCLE_COUNTER)
1786 val = ARMV7_INTENS_C;
1787 else
1788 val = ARMV7_INTENS_P(idx);
1789
1790 asm volatile("mcr p15, 0, %0, c9, c14, 1" : : "r" (val));
1791
1792 return idx;
1793 }
1794
1795 static inline u32 armv7_pmnc_disable_intens(unsigned int idx)
1796 {
1797 u32 val;
1798
1799 if ((idx != ARMV7_CYCLE_COUNTER) &&
1800 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1801 pr_err("CPU%u disabling wrong PMNC counter"
1802 " interrupt enable %d\n", smp_processor_id(), idx);
1803 return -1;
1804 }
1805
1806 if (idx == ARMV7_CYCLE_COUNTER)
1807 val = ARMV7_INTENC_C;
1808 else
1809 val = ARMV7_INTENC_P(idx);
1810
1811 asm volatile("mcr p15, 0, %0, c9, c14, 2" : : "r" (val));
1812
1813 return idx;
1814 }
1815
1816 static inline u32 armv7_pmnc_getreset_flags(void)
1817 {
1818 u32 val;
1819
1820 /* Read */
1821 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
1822
1823 /* Write to clear flags */
1824 val &= ARMV7_FLAG_MASK;
1825 asm volatile("mcr p15, 0, %0, c9, c12, 3" : : "r" (val));
1826
1827 return val;
1828 }
1829
1830 #ifdef DEBUG
1831 static void armv7_pmnc_dump_regs(void)
1832 {
1833 u32 val;
1834 unsigned int cnt;
1835
1836 printk(KERN_INFO "PMNC registers dump:\n");
1837
1838 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (val));
1839 printk(KERN_INFO "PMNC =0x%08x\n", val);
1840
1841 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r" (val));
1842 printk(KERN_INFO "CNTENS=0x%08x\n", val);
1843
1844 asm volatile("mrc p15, 0, %0, c9, c14, 1" : "=r" (val));
1845 printk(KERN_INFO "INTENS=0x%08x\n", val);
1846
1847 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
1848 printk(KERN_INFO "FLAGS =0x%08x\n", val);
1849
1850 asm volatile("mrc p15, 0, %0, c9, c12, 5" : "=r" (val));
1851 printk(KERN_INFO "SELECT=0x%08x\n", val);
1852
1853 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (val));
1854 printk(KERN_INFO "CCNT =0x%08x\n", val);
1855
1856 for (cnt = ARMV7_COUNTER0; cnt < ARMV7_COUNTER_LAST; cnt++) {
1857 armv7_pmnc_select_counter(cnt);
1858 asm volatile("mrc p15, 0, %0, c9, c13, 2" : "=r" (val));
1859 printk(KERN_INFO "CNT[%d] count =0x%08x\n",
1860 cnt-ARMV7_EVENT_CNT_TO_CNTx, val);
1861 asm volatile("mrc p15, 0, %0, c9, c13, 1" : "=r" (val));
1862 printk(KERN_INFO "CNT[%d] evtsel=0x%08x\n",
1863 cnt-ARMV7_EVENT_CNT_TO_CNTx, val);
1864 }
1865 }
1866 #endif
1867
1868 void armv7pmu_enable_event(struct hw_perf_event *hwc, int idx)
1869 {
1870 unsigned long flags;
1871
1872 /*
1873 * Enable counter and interrupt, and set the counter to count
1874 * the event that we're interested in.
1875 */
1876 spin_lock_irqsave(&pmu_lock, flags);
1877
1878 /*
1879 * Disable counter
1880 */
1881 armv7_pmnc_disable_counter(idx);
1882
1883 /*
1884 * Set event (if destined for PMNx counters)
1885 * We don't need to set the event if it's a cycle count
1886 */
1887 if (idx != ARMV7_CYCLE_COUNTER)
1888 armv7_pmnc_write_evtsel(idx, hwc->config_base);
1889
1890 /*
1891 * Enable interrupt for this counter
1892 */
1893 armv7_pmnc_enable_intens(idx);
1894
1895 /*
1896 * Enable counter
1897 */
1898 armv7_pmnc_enable_counter(idx);
1899
1900 spin_unlock_irqrestore(&pmu_lock, flags);
1901 }
1902
1903 static void armv7pmu_disable_event(struct hw_perf_event *hwc, int idx)
1904 {
1905 unsigned long flags;
1906
1907 /*
1908 * Disable counter and interrupt
1909 */
1910 spin_lock_irqsave(&pmu_lock, flags);
1911
1912 /*
1913 * Disable counter
1914 */
1915 armv7_pmnc_disable_counter(idx);
1916
1917 /*
1918 * Disable interrupt for this counter
1919 */
1920 armv7_pmnc_disable_intens(idx);
1921
1922 spin_unlock_irqrestore(&pmu_lock, flags);
1923 }
1924
1925 static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev)
1926 {
1927 unsigned long pmnc;
1928 struct perf_sample_data data;
1929 struct cpu_hw_events *cpuc;
1930 struct pt_regs *regs;
1931 int idx;
1932
1933 /*
1934 * Get and reset the IRQ flags
1935 */
1936 pmnc = armv7_pmnc_getreset_flags();
1937
1938 /*
1939 * Did an overflow occur?
1940 */
1941 if (!armv7_pmnc_has_overflowed(pmnc))
1942 return IRQ_NONE;
1943
1944 /*
1945 * Handle the counter(s) overflow(s)
1946 */
1947 regs = get_irq_regs();
1948
1949 perf_sample_data_init(&data, 0);
1950
1951 cpuc = &__get_cpu_var(cpu_hw_events);
1952 for (idx = 0; idx <= armpmu->num_events; ++idx) {
1953 struct perf_event *event = cpuc->events[idx];
1954 struct hw_perf_event *hwc;
1955
1956 if (!test_bit(idx, cpuc->active_mask))
1957 continue;
1958
1959 /*
1960 * We have a single interrupt for all counters. Check that
1961 * each counter has overflowed before we process it.
1962 */
1963 if (!armv7_pmnc_counter_has_overflowed(pmnc, idx))
1964 continue;
1965
1966 hwc = &event->hw;
1967 armpmu_event_update(event, hwc, idx);
1968 data.period = event->hw.last_period;
1969 if (!armpmu_event_set_period(event, hwc, idx))
1970 continue;
1971
1972 if (perf_event_overflow(event, 0, &data, regs))
1973 armpmu->disable(hwc, idx);
1974 }
1975
1976 /*
1977 * Handle the pending perf events.
1978 *
1979 * Note: this call *must* be run with interrupts enabled. For
1980 * platforms that can have the PMU interrupts raised as a PMI, this
1981 * will not work.
1982 */
1983 perf_event_do_pending();
1984
1985 return IRQ_HANDLED;
1986 }
1987
1988 static void armv7pmu_start(void)
1989 {
1990 unsigned long flags;
1991
1992 spin_lock_irqsave(&pmu_lock, flags);
1993 /* Enable all counters */
1994 armv7_pmnc_write(armv7_pmnc_read() | ARMV7_PMNC_E);
1995 spin_unlock_irqrestore(&pmu_lock, flags);
1996 }
1997
1998 static void armv7pmu_stop(void)
1999 {
2000 unsigned long flags;
2001
2002 spin_lock_irqsave(&pmu_lock, flags);
2003 /* Disable all counters */
2004 armv7_pmnc_write(armv7_pmnc_read() & ~ARMV7_PMNC_E);
2005 spin_unlock_irqrestore(&pmu_lock, flags);
2006 }
2007
2008 static inline int armv7_a8_pmu_event_map(int config)
2009 {
2010 int mapping = armv7_a8_perf_map[config];
2011 if (HW_OP_UNSUPPORTED == mapping)
2012 mapping = -EOPNOTSUPP;
2013 return mapping;
2014 }
2015
2016 static inline int armv7_a9_pmu_event_map(int config)
2017 {
2018 int mapping = armv7_a9_perf_map[config];
2019 if (HW_OP_UNSUPPORTED == mapping)
2020 mapping = -EOPNOTSUPP;
2021 return mapping;
2022 }
2023
2024 static u64 armv7pmu_raw_event(u64 config)
2025 {
2026 return config & 0xff;
2027 }
2028
2029 static int armv7pmu_get_event_idx(struct cpu_hw_events *cpuc,
2030 struct hw_perf_event *event)
2031 {
2032 int idx;
2033
2034 /* Always place a cycle counter into the cycle counter. */
2035 if (event->config_base == ARMV7_PERFCTR_CPU_CYCLES) {
2036 if (test_and_set_bit(ARMV7_CYCLE_COUNTER, cpuc->used_mask))
2037 return -EAGAIN;
2038
2039 return ARMV7_CYCLE_COUNTER;
2040 } else {
2041 /*
2042 * For anything other than a cycle counter, try and use
2043 * the events counters
2044 */
2045 for (idx = ARMV7_COUNTER0; idx <= armpmu->num_events; ++idx) {
2046 if (!test_and_set_bit(idx, cpuc->used_mask))
2047 return idx;
2048 }
2049
2050 /* The counters are all in use. */
2051 return -EAGAIN;
2052 }
2053 }
2054
2055 static struct arm_pmu armv7pmu = {
2056 .handle_irq = armv7pmu_handle_irq,
2057 .enable = armv7pmu_enable_event,
2058 .disable = armv7pmu_disable_event,
2059 .raw_event = armv7pmu_raw_event,
2060 .read_counter = armv7pmu_read_counter,
2061 .write_counter = armv7pmu_write_counter,
2062 .get_event_idx = armv7pmu_get_event_idx,
2063 .start = armv7pmu_start,
2064 .stop = armv7pmu_stop,
2065 .max_period = (1LLU << 32) - 1,
2066 };
2067
2068 static u32 __init armv7_reset_read_pmnc(void)
2069 {
2070 u32 nb_cnt;
2071
2072 /* Initialize & Reset PMNC: C and P bits */
2073 armv7_pmnc_write(ARMV7_PMNC_P | ARMV7_PMNC_C);
2074
2075 /* Read the nb of CNTx counters supported from PMNC */
2076 nb_cnt = (armv7_pmnc_read() >> ARMV7_PMNC_N_SHIFT) & ARMV7_PMNC_N_MASK;
2077
2078 /* Add the CPU cycles counter and return */
2079 return nb_cnt + 1;
2080 }
2081
2082 static int __init
2083 init_hw_perf_events(void)
2084 {
2085 unsigned long cpuid = read_cpuid_id();
2086 unsigned long implementor = (cpuid & 0xFF000000) >> 24;
2087 unsigned long part_number = (cpuid & 0xFFF0);
2088
2089 /* We only support ARM CPUs implemented by ARM at the moment. */
2090 if (0x41 == implementor) {
2091 switch (part_number) {
2092 case 0xB360: /* ARM1136 */
2093 case 0xB560: /* ARM1156 */
2094 case 0xB760: /* ARM1176 */
2095 armpmu = &armv6pmu;
2096 memcpy(armpmu_perf_cache_map, armv6_perf_cache_map,
2097 sizeof(armv6_perf_cache_map));
2098 perf_max_events = armv6pmu.num_events;
2099 break;
2100 case 0xB020: /* ARM11mpcore */
2101 armpmu = &armv6mpcore_pmu;
2102 memcpy(armpmu_perf_cache_map,
2103 armv6mpcore_perf_cache_map,
2104 sizeof(armv6mpcore_perf_cache_map));
2105 perf_max_events = armv6mpcore_pmu.num_events;
2106 break;
2107 case 0xC080: /* Cortex-A8 */
2108 armv7pmu.name = ARMV7_PMU_CORTEX_A8_NAME;
2109 memcpy(armpmu_perf_cache_map, armv7_a8_perf_cache_map,
2110 sizeof(armv7_a8_perf_cache_map));
2111 armv7pmu.event_map = armv7_a8_pmu_event_map;
2112 armpmu = &armv7pmu;
2113
2114 /* Reset PMNC and read the nb of CNTx counters
2115 supported */
2116 armv7pmu.num_events = armv7_reset_read_pmnc();
2117 perf_max_events = armv7pmu.num_events;
2118 break;
2119 case 0xC090: /* Cortex-A9 */
2120 armv7pmu.name = ARMV7_PMU_CORTEX_A9_NAME;
2121 memcpy(armpmu_perf_cache_map, armv7_a9_perf_cache_map,
2122 sizeof(armv7_a9_perf_cache_map));
2123 armv7pmu.event_map = armv7_a9_pmu_event_map;
2124 armpmu = &armv7pmu;
2125
2126 /* Reset PMNC and read the nb of CNTx counters
2127 supported */
2128 armv7pmu.num_events = armv7_reset_read_pmnc();
2129 perf_max_events = armv7pmu.num_events;
2130 break;
2131 default:
2132 pr_info("no hardware support available\n");
2133 perf_max_events = -1;
2134 }
2135 }
2136
2137 if (armpmu)
2138 pr_info("enabled with %s PMU driver, %d counters available\n",
2139 armpmu->name, armpmu->num_events);
2140
2141 return 0;
2142 }
2143 arch_initcall(init_hw_perf_events);
2144
2145 /*
2146 * Callchain handling code.
2147 */
2148 static inline void
2149 callchain_store(struct perf_callchain_entry *entry,
2150 u64 ip)
2151 {
2152 if (entry->nr < PERF_MAX_STACK_DEPTH)
2153 entry->ip[entry->nr++] = ip;
2154 }
2155
2156 /*
2157 * The registers we're interested in are at the end of the variable
2158 * length saved register structure. The fp points at the end of this
2159 * structure so the address of this struct is:
2160 * (struct frame_tail *)(xxx->fp)-1
2161 *
2162 * This code has been adapted from the ARM OProfile support.
2163 */
2164 struct frame_tail {
2165 struct frame_tail *fp;
2166 unsigned long sp;
2167 unsigned long lr;
2168 } __attribute__((packed));
2169
2170 /*
2171 * Get the return address for a single stackframe and return a pointer to the
2172 * next frame tail.
2173 */
2174 static struct frame_tail *
2175 user_backtrace(struct frame_tail *tail,
2176 struct perf_callchain_entry *entry)
2177 {
2178 struct frame_tail buftail;
2179
2180 /* Also check accessibility of one struct frame_tail beyond */
2181 if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
2182 return NULL;
2183 if (__copy_from_user_inatomic(&buftail, tail, sizeof(buftail)))
2184 return NULL;
2185
2186 callchain_store(entry, buftail.lr);
2187
2188 /*
2189 * Frame pointers should strictly progress back up the stack
2190 * (towards higher addresses).
2191 */
2192 if (tail >= buftail.fp)
2193 return NULL;
2194
2195 return buftail.fp - 1;
2196 }
2197
2198 static void
2199 perf_callchain_user(struct pt_regs *regs,
2200 struct perf_callchain_entry *entry)
2201 {
2202 struct frame_tail *tail;
2203
2204 callchain_store(entry, PERF_CONTEXT_USER);
2205
2206 if (!user_mode(regs))
2207 regs = task_pt_regs(current);
2208
2209 tail = (struct frame_tail *)regs->ARM_fp - 1;
2210
2211 while (tail && !((unsigned long)tail & 0x3))
2212 tail = user_backtrace(tail, entry);
2213 }
2214
2215 /*
2216 * Gets called by walk_stackframe() for every stackframe. This will be called
2217 * whist unwinding the stackframe and is like a subroutine return so we use
2218 * the PC.
2219 */
2220 static int
2221 callchain_trace(struct stackframe *fr,
2222 void *data)
2223 {
2224 struct perf_callchain_entry *entry = data;
2225 callchain_store(entry, fr->pc);
2226 return 0;
2227 }
2228
2229 static void
2230 perf_callchain_kernel(struct pt_regs *regs,
2231 struct perf_callchain_entry *entry)
2232 {
2233 struct stackframe fr;
2234
2235 callchain_store(entry, PERF_CONTEXT_KERNEL);
2236 fr.fp = regs->ARM_fp;
2237 fr.sp = regs->ARM_sp;
2238 fr.lr = regs->ARM_lr;
2239 fr.pc = regs->ARM_pc;
2240 walk_stackframe(&fr, callchain_trace, entry);
2241 }
2242
2243 static void
2244 perf_do_callchain(struct pt_regs *regs,
2245 struct perf_callchain_entry *entry)
2246 {
2247 int is_user;
2248
2249 if (!regs)
2250 return;
2251
2252 is_user = user_mode(regs);
2253
2254 if (!current || !current->pid)
2255 return;
2256
2257 if (is_user && current->state != TASK_RUNNING)
2258 return;
2259
2260 if (!is_user)
2261 perf_callchain_kernel(regs, entry);
2262
2263 if (current->mm)
2264 perf_callchain_user(regs, entry);
2265 }
2266
2267 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_irq_entry);
2268
2269 struct perf_callchain_entry *
2270 perf_callchain(struct pt_regs *regs)
2271 {
2272 struct perf_callchain_entry *entry = &__get_cpu_var(pmc_irq_entry);
2273
2274 entry->nr = 0;
2275 perf_do_callchain(regs, entry);
2276 return entry;
2277 }