]> bbs.cooldavid.org Git - net-next-2.6.git/blame - arch/arm/kernel/perf_event.c
git://bbs.cooldavid.org / net-next-2.6.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ARM: 6469/1: perf-events: squash compiler warning
[net-next-2.6.git] / arch / arm / kernel / perf_event.c
CommitLineData
1b8873a0
JI		 1#undef DEBUG
2
3/*
4 * ARM performance counter support.
5 *
6 * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
7 *
796d1295
JP		 8 * ARMv7 support: Jean Pihet <jpihet@mvista.com>
9 * 2010 (c) MontaVista Software, LLC.
10 *
1b8873a0
JI		 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>
181193f3 19#include <linux/module.h>
1b8873a0 20#include <linux/perf_event.h>
49c006b9 21#include <linux/platform_device.h>
1b8873a0
JI		 22#include <linux/spinlock.h>
23#include <linux/uaccess.h>
24
25#include <asm/cputype.h>
26#include <asm/irq.h>
27#include <asm/irq_regs.h>
28#include <asm/pmu.h>
29#include <asm/stacktrace.h>
30
49c006b9 31static struct platform_device *pmu_device;
1b8873a0
JI		 32
33/*
34 * Hardware lock to serialize accesses to PMU registers. Needed for the
35 * read/modify/write sequences.
36 */
37DEFINE_SPINLOCK(pmu_lock);
38
39/*
40 * ARMv6 supports a maximum of 3 events, starting from index 1. If we add
41 * another platform that supports more, we need to increase this to be the
42 * largest of all platforms.
796d1295
JP		 43 *
44 * ARMv7 supports up to 32 events:
45 * cycle counter CCNT + 31 events counters CNT0..30.
46 * Cortex-A8 has 1+4 counters, Cortex-A9 has 1+6 counters.
1b8873a0 47 */
796d1295 48#define ARMPMU_MAX_HWEVENTS 33
1b8873a0
JI		 49
50/* The events for a given CPU. */
51struct cpu_hw_events {
52 /*
53 * The events that are active on the CPU for the given index. Index 0
54 * is reserved.
55 */
56 struct perf_event *events[ARMPMU_MAX_HWEVENTS];
57
58 /*
59 * A 1 bit for an index indicates that the counter is being used for
60 * an event. A 0 means that the counter can be used.
61 */
62 unsigned long used_mask[BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)];
63
64 /*
65 * A 1 bit for an index indicates that the counter is actively being
66 * used.
67 */
68 unsigned long active_mask[BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)];
69};
70DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
71
181193f3
WD		 72/* PMU names. */
73static const char *arm_pmu_names[] = {
74 [ARM_PERF_PMU_ID_XSCALE1] = "xscale1",
75 [ARM_PERF_PMU_ID_XSCALE2] = "xscale2",
76 [ARM_PERF_PMU_ID_V6] = "v6",
77 [ARM_PERF_PMU_ID_V6MP] = "v6mpcore",
78 [ARM_PERF_PMU_ID_CA8] = "ARMv7 Cortex-A8",
79 [ARM_PERF_PMU_ID_CA9] = "ARMv7 Cortex-A9",
80};
81
1b8873a0 82struct arm_pmu {
181193f3 83 enum arm_perf_pmu_ids id;
1b8873a0
JI		 84 irqreturn_t (*handle_irq)(int irq_num, void *dev);
85 void (*enable)(struct hw_perf_event *evt, int idx);
86 void (*disable)(struct hw_perf_event *evt, int idx);
87 int (*event_map)(int evt);
88 u64 (*raw_event)(u64);
89 int (*get_event_idx)(struct cpu_hw_events *cpuc,
90 struct hw_perf_event *hwc);
91 u32 (*read_counter)(int idx);
92 void (*write_counter)(int idx, u32 val);
93 void (*start)(void);
94 void (*stop)(void);
95 int num_events;
96 u64 max_period;
97};
98
99/* Set at runtime when we know what CPU type we are. */
100static const struct arm_pmu *armpmu;
101
181193f3
WD		 102enum arm_perf_pmu_ids
103armpmu_get_pmu_id(void)
104{
105 int id = -ENODEV;
106
107 if (armpmu != NULL)
108 id = armpmu->id;
109
110 return id;
111}
112EXPORT_SYMBOL_GPL(armpmu_get_pmu_id);
113
929f5199
WD		 114int
115armpmu_get_max_events(void)
116{
117 int max_events = 0;
118
119 if (armpmu != NULL)
120 max_events = armpmu->num_events;
121
122 return max_events;
123}
124EXPORT_SYMBOL_GPL(armpmu_get_max_events);
125
3bf101ba
MF		 126int perf_num_counters(void)
127{
128 return armpmu_get_max_events();
129}
130EXPORT_SYMBOL_GPL(perf_num_counters);
131
1b8873a0
JI		 132#define HW_OP_UNSUPPORTED 0xFFFF
133
134#define C(_x) \
135 PERF_COUNT_HW_CACHE_##_x
136
137#define CACHE_OP_UNSUPPORTED 0xFFFF
138
139static unsigned armpmu_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
140 [PERF_COUNT_HW_CACHE_OP_MAX]
141 [PERF_COUNT_HW_CACHE_RESULT_MAX];
142
143static int
144armpmu_map_cache_event(u64 config)
145{
146 unsigned int cache_type, cache_op, cache_result, ret;
147
148 cache_type = (config >> 0) & 0xff;
149 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
150 return -EINVAL;
151
152 cache_op = (config >> 8) & 0xff;
153 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
154 return -EINVAL;
155
156 cache_result = (config >> 16) & 0xff;
157 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
158 return -EINVAL;
159
160 ret = (int)armpmu_perf_cache_map[cache_type][cache_op][cache_result];
161
162 if (ret == CACHE_OP_UNSUPPORTED)
163 return -ENOENT;
164
165 return ret;
166}
167
168static int
169armpmu_event_set_period(struct perf_event *event,
170 struct hw_perf_event *hwc,
171 int idx)
172{
e7850595 173 s64 left = local64_read(&hwc->period_left);
1b8873a0
JI		 174 s64 period = hwc->sample_period;
175 int ret = 0;
176
177 if (unlikely(left <= -period)) {
178 left = period;
e7850595 179 local64_set(&hwc->period_left, left);
1b8873a0
JI		 180 hwc->last_period = period;
181 ret = 1;
182 }
183
184 if (unlikely(left <= 0)) {
185 left += period;
e7850595 186 local64_set(&hwc->period_left, left);
1b8873a0
JI		 187 hwc->last_period = period;
188 ret = 1;
189 }
190
191 if (left > (s64)armpmu->max_period)
192 left = armpmu->max_period;
193
e7850595 194 local64_set(&hwc->prev_count, (u64)-left);
1b8873a0
JI		 195
196 armpmu->write_counter(idx, (u64)(-left) & 0xffffffff);
197
198 perf_event_update_userpage(event);
199
200 return ret;
201}
202
203static u64
204armpmu_event_update(struct perf_event *event,
205 struct hw_perf_event *hwc,
206 int idx)
207{
208 int shift = 64 - 32;
209 s64 prev_raw_count, new_raw_count;
446a5a8b 210 u64 delta;
1b8873a0
JI		 211
212again:
e7850595 213 prev_raw_count = local64_read(&hwc->prev_count);
1b8873a0
JI		 214 new_raw_count = armpmu->read_counter(idx);
215
e7850595 216 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
1b8873a0
JI		 217 new_raw_count) != prev_raw_count)
218 goto again;
219
220 delta = (new_raw_count << shift) - (prev_raw_count << shift);
221 delta >>= shift;
222
e7850595
PZ		 223 local64_add(delta, &event->count);
224 local64_sub(delta, &hwc->period_left);
1b8873a0
JI		 225
226 return new_raw_count;
227}
228
229static void
a4eaf7f1 230armpmu_read(struct perf_event *event)
1b8873a0 231{
1b8873a0 232 struct hw_perf_event *hwc = &event->hw;
1b8873a0 233
a4eaf7f1
PZ		 234 /* Don't read disabled counters! */
235 if (hwc->idx < 0)
236 return;
1b8873a0 237
a4eaf7f1 238 armpmu_event_update(event, hwc, hwc->idx);
1b8873a0
JI		 239}
240
241static void
a4eaf7f1 242armpmu_stop(struct perf_event *event, int flags)
1b8873a0
JI		 243{
244 struct hw_perf_event *hwc = &event->hw;
245
a4eaf7f1 246 if (!armpmu)
1b8873a0
JI		 247 return;
248
a4eaf7f1
PZ		 249 /*
250 * ARM pmu always has to update the counter, so ignore
251 * PERF_EF_UPDATE, see comments in armpmu_start().
252 */
253 if (!(hwc->state & PERF_HES_STOPPED)) {
254 armpmu->disable(hwc, hwc->idx);
255 barrier(); /* why? */
256 armpmu_event_update(event, hwc, hwc->idx);
257 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
258 }
1b8873a0
JI		 259}
260
261static void
a4eaf7f1 262armpmu_start(struct perf_event *event, int flags)
1b8873a0
JI		 263{
264 struct hw_perf_event *hwc = &event->hw;
265
a4eaf7f1
PZ		 266 if (!armpmu)
267 return;
268
269 /*
270 * ARM pmu always has to reprogram the period, so ignore
271 * PERF_EF_RELOAD, see the comment below.
272 */
273 if (flags & PERF_EF_RELOAD)
274 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
275
276 hwc->state = 0;
1b8873a0
JI		 277 /*
278 * Set the period again. Some counters can't be stopped, so when we
a4eaf7f1 279 * were stopped we simply disabled the IRQ source and the counter
1b8873a0
JI		 280 * may have been left counting. If we don't do this step then we may
281 * get an interrupt too soon or *way* too late if the overflow has
282 * happened since disabling.
283 */
284 armpmu_event_set_period(event, hwc, hwc->idx);
285 armpmu->enable(hwc, hwc->idx);
286}
287
a4eaf7f1
PZ		 288static void
289armpmu_del(struct perf_event *event, int flags)
290{
291 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
292 struct hw_perf_event *hwc = &event->hw;
293 int idx = hwc->idx;
294
295 WARN_ON(idx < 0);
296
297 clear_bit(idx, cpuc->active_mask);
298 armpmu_stop(event, PERF_EF_UPDATE);
299 cpuc->events[idx] = NULL;
300 clear_bit(idx, cpuc->used_mask);
301
302 perf_event_update_userpage(event);
303}
304
1b8873a0 305static int
a4eaf7f1 306armpmu_add(struct perf_event *event, int flags)
1b8873a0
JI		 307{
308 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
309 struct hw_perf_event *hwc = &event->hw;
310 int idx;
311 int err = 0;
312
33696fc0 313 perf_pmu_disable(event->pmu);
24cd7f54 314
1b8873a0
JI		 315 /* If we don't have a space for the counter then finish early. */
316 idx = armpmu->get_event_idx(cpuc, hwc);
317 if (idx < 0) {
318 err = idx;
319 goto out;
320 }
321
322 /*
323 * If there is an event in the counter we are going to use then make
324 * sure it is disabled.
325 */
326 event->hw.idx = idx;
327 armpmu->disable(hwc, idx);
328 cpuc->events[idx] = event;
329 set_bit(idx, cpuc->active_mask);
330
a4eaf7f1
PZ		 331 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
332 if (flags & PERF_EF_START)
333 armpmu_start(event, PERF_EF_RELOAD);
1b8873a0
JI		 334
335 /* Propagate our changes to the userspace mapping. */
336 perf_event_update_userpage(event);
337
338out:
33696fc0 339 perf_pmu_enable(event->pmu);
1b8873a0
JI		 340 return err;
341}
342
b0a873eb 343static struct pmu pmu;
1b8873a0
JI		 344
345static int
346validate_event(struct cpu_hw_events *cpuc,
347 struct perf_event *event)
348{
349 struct hw_perf_event fake_event = event->hw;
350
65b4711f
WD		 351 if (event->pmu != &pmu || event->state <= PERF_EVENT_STATE_OFF)
352 return 1;
1b8873a0
JI		 353
354 return armpmu->get_event_idx(cpuc, &fake_event) >= 0;
355}
356
357static int
358validate_group(struct perf_event *event)
359{
360 struct perf_event *sibling, *leader = event->group_leader;
361 struct cpu_hw_events fake_pmu;
362
363 memset(&fake_pmu, 0, sizeof(fake_pmu));
364
365 if (!validate_event(&fake_pmu, leader))
366 return -ENOSPC;
367
368 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
369 if (!validate_event(&fake_pmu, sibling))
370 return -ENOSPC;
371 }
372
373 if (!validate_event(&fake_pmu, event))
374 return -ENOSPC;
375
376 return 0;
377}
378
379static int
380armpmu_reserve_hardware(void)
381{
49c006b9 382 int i, err = -ENODEV, irq;
1b8873a0 383
49c006b9
WD		 384 pmu_device = reserve_pmu(ARM_PMU_DEVICE_CPU);
385 if (IS_ERR(pmu_device)) {
1b8873a0 386 pr_warning("unable to reserve pmu\n");
49c006b9 387 return PTR_ERR(pmu_device);
1b8873a0
JI		 388 }
389
49c006b9 390 init_pmu(ARM_PMU_DEVICE_CPU);
1b8873a0 391
49c006b9 392 if (pmu_device->num_resources < 1) {
1b8873a0
JI		 393 pr_err("no irqs for PMUs defined\n");
394 return -ENODEV;
395 }
396
49c006b9
WD		 397 for (i = 0; i < pmu_device->num_resources; ++i) {
398 irq = platform_get_irq(pmu_device, i);
399 if (irq < 0)
400 continue;
401
402 err = request_irq(irq, armpmu->handle_irq,
ddee87f2
WD		 403 IRQF_DISABLED | IRQF_NOBALANCING,
404 "armpmu", NULL);
1b8873a0 405 if (err) {
49c006b9
WD		 406 pr_warning("unable to request IRQ%d for ARM perf "
407 "counters\n", irq);
1b8873a0
JI		 408 break;
409 }
410 }
411
412 if (err) {
49c006b9
WD		 413 for (i = i - 1; i >= 0; --i) {
414 irq = platform_get_irq(pmu_device, i);
415 if (irq >= 0)
416 free_irq(irq, NULL);
417 }
418 release_pmu(pmu_device);
419 pmu_device = NULL;
1b8873a0
JI		 420 }
421
422 return err;
423}
424
425static void
426armpmu_release_hardware(void)
427{
49c006b9 428 int i, irq;
1b8873a0 429
49c006b9
WD		 430 for (i = pmu_device->num_resources - 1; i >= 0; --i) {
431 irq = platform_get_irq(pmu_device, i);
432 if (irq >= 0)
433 free_irq(irq, NULL);
434 }
1b8873a0
JI		 435 armpmu->stop();
436
49c006b9
WD		 437 release_pmu(pmu_device);
438 pmu_device = NULL;
1b8873a0
JI		 439}
440
441static atomic_t active_events = ATOMIC_INIT(0);
442static DEFINE_MUTEX(pmu_reserve_mutex);
443
444static void
445hw_perf_event_destroy(struct perf_event *event)
446{
447 if (atomic_dec_and_mutex_lock(&active_events, &pmu_reserve_mutex)) {
448 armpmu_release_hardware();
449 mutex_unlock(&pmu_reserve_mutex);
450 }
451}
452
453static int
454__hw_perf_event_init(struct perf_event *event)
455{
456 struct hw_perf_event *hwc = &event->hw;
457 int mapping, err;
458
459 /* Decode the generic type into an ARM event identifier. */
460 if (PERF_TYPE_HARDWARE == event->attr.type) {
461 mapping = armpmu->event_map(event->attr.config);
462 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
463 mapping = armpmu_map_cache_event(event->attr.config);
464 } else if (PERF_TYPE_RAW == event->attr.type) {
465 mapping = armpmu->raw_event(event->attr.config);
466 } else {
467 pr_debug("event type %x not supported\n", event->attr.type);
468 return -EOPNOTSUPP;
469 }
470
471 if (mapping < 0) {
472 pr_debug("event %x:%llx not supported\n", event->attr.type,
473 event->attr.config);
474 return mapping;
475 }
476
477 /*
478 * Check whether we need to exclude the counter from certain modes.
479 * The ARM performance counters are on all of the time so if someone
480 * has asked us for some excludes then we have to fail.
481 */
482 if (event->attr.exclude_kernel || event->attr.exclude_user ||
483 event->attr.exclude_hv || event->attr.exclude_idle) {
484 pr_debug("ARM performance counters do not support "
485 "mode exclusion\n");
486 return -EPERM;
487 }
488
489 /*
490 * We don't assign an index until we actually place the event onto
491 * hardware. Use -1 to signify that we haven't decided where to put it
492 * yet. For SMP systems, each core has it's own PMU so we can't do any
493 * clever allocation or constraints checking at this point.
494 */
495 hwc->idx = -1;
496
497 /*
498 * Store the event encoding into the config_base field. config and
499 * event_base are unused as the only 2 things we need to know are
500 * the event mapping and the counter to use. The counter to use is
501 * also the indx and the config_base is the event type.
502 */
503 hwc->config_base = (unsigned long)mapping;
504 hwc->config = 0;
505 hwc->event_base = 0;
506
507 if (!hwc->sample_period) {
508 hwc->sample_period = armpmu->max_period;
509 hwc->last_period = hwc->sample_period;
e7850595 510 local64_set(&hwc->period_left, hwc->sample_period);
1b8873a0
JI		 511 }
512
513 err = 0;
514 if (event->group_leader != event) {
515 err = validate_group(event);
516 if (err)
517 return -EINVAL;
518 }
519
520 return err;
521}
522
b0a873eb 523static int armpmu_event_init(struct perf_event *event)
1b8873a0
JI		 524{
525 int err = 0;
526
b0a873eb
PZ		 527 switch (event->attr.type) {
528 case PERF_TYPE_RAW:
529 case PERF_TYPE_HARDWARE:
530 case PERF_TYPE_HW_CACHE:
531 break;
532
533 default:
534 return -ENOENT;
535 }
536
1b8873a0 537 if (!armpmu)
b0a873eb 538 return -ENODEV;
1b8873a0
JI		 539
540 event->destroy = hw_perf_event_destroy;
541
542 if (!atomic_inc_not_zero(&active_events)) {
1efeb08d 543 if (atomic_read(&active_events) > armpmu->num_events) {
1b8873a0 544 atomic_dec(&active_events);
b0a873eb 545 return -ENOSPC;
1b8873a0
JI		 546 }
547
548 mutex_lock(&pmu_reserve_mutex);
549 if (atomic_read(&active_events) == 0) {
550 err = armpmu_reserve_hardware();
551 }
552
553 if (!err)
554 atomic_inc(&active_events);
555 mutex_unlock(&pmu_reserve_mutex);
556 }
557
558 if (err)
b0a873eb 559 return err;
1b8873a0
JI		 560
561 err = __hw_perf_event_init(event);
562 if (err)
563 hw_perf_event_destroy(event);
564
b0a873eb 565 return err;
1b8873a0
JI		 566}
567
a4eaf7f1 568static void armpmu_enable(struct pmu *pmu)
1b8873a0
JI		 569{
570 /* Enable all of the perf events on hardware. */
571 int idx;
572 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
573
574 if (!armpmu)
575 return;
576
577 for (idx = 0; idx <= armpmu->num_events; ++idx) {
578 struct perf_event *event = cpuc->events[idx];
579
580 if (!event)
581 continue;
582
583 armpmu->enable(&event->hw, idx);
584 }
585
586 armpmu->start();
587}
588
a4eaf7f1 589static void armpmu_disable(struct pmu *pmu)
1b8873a0
JI		 590{
591 if (armpmu)
592 armpmu->stop();
593}
594
33696fc0 595static struct pmu pmu = {
a4eaf7f1
PZ		 596 .pmu_enable = armpmu_enable,
597 .pmu_disable = armpmu_disable,
598 .event_init = armpmu_event_init,
599 .add = armpmu_add,
600 .del = armpmu_del,
601 .start = armpmu_start,
602 .stop = armpmu_stop,
603 .read = armpmu_read,
33696fc0
PZ		 604};
605
1b8873a0
JI		 606/*
607 * ARMv6 Performance counter handling code.
608 *
609 * ARMv6 has 2 configurable performance counters and a single cycle counter.
610 * They all share a single reset bit but can be written to zero so we can use
611 * that for a reset.
612 *
613 * The counters can't be individually enabled or disabled so when we remove
614 * one event and replace it with another we could get spurious counts from the
615 * wrong event. However, we can take advantage of the fact that the
616 * performance counters can export events to the event bus, and the event bus
617 * itself can be monitored. This requires that we *don't* export the events to
618 * the event bus. The procedure for disabling a configurable counter is:
619 * - change the counter to count the ETMEXTOUT[0] signal (0x20). This
620 * effectively stops the counter from counting.
621 * - disable the counter's interrupt generation (each counter has it's
622 * own interrupt enable bit).
623 * Once stopped, the counter value can be written as 0 to reset.
624 *
625 * To enable a counter:
626 * - enable the counter's interrupt generation.
627 * - set the new event type.
628 *
629 * Note: the dedicated cycle counter only counts cycles and can't be
630 * enabled/disabled independently of the others. When we want to disable the
631 * cycle counter, we have to just disable the interrupt reporting and start
632 * ignoring that counter. When re-enabling, we have to reset the value and
633 * enable the interrupt.
634 */
635
636enum armv6_perf_types {
637 ARMV6_PERFCTR_ICACHE_MISS = 0x0,
638 ARMV6_PERFCTR_IBUF_STALL = 0x1,
639 ARMV6_PERFCTR_DDEP_STALL = 0x2,
640 ARMV6_PERFCTR_ITLB_MISS = 0x3,
641 ARMV6_PERFCTR_DTLB_MISS = 0x4,
642 ARMV6_PERFCTR_BR_EXEC = 0x5,
643 ARMV6_PERFCTR_BR_MISPREDICT = 0x6,
644 ARMV6_PERFCTR_INSTR_EXEC = 0x7,
645 ARMV6_PERFCTR_DCACHE_HIT = 0x9,
646 ARMV6_PERFCTR_DCACHE_ACCESS = 0xA,
647 ARMV6_PERFCTR_DCACHE_MISS = 0xB,
648 ARMV6_PERFCTR_DCACHE_WBACK = 0xC,
649 ARMV6_PERFCTR_SW_PC_CHANGE = 0xD,
650 ARMV6_PERFCTR_MAIN_TLB_MISS = 0xF,
651 ARMV6_PERFCTR_EXPL_D_ACCESS = 0x10,
652 ARMV6_PERFCTR_LSU_FULL_STALL = 0x11,
653 ARMV6_PERFCTR_WBUF_DRAINED = 0x12,
654 ARMV6_PERFCTR_CPU_CYCLES = 0xFF,
655 ARMV6_PERFCTR_NOP = 0x20,
656};
657
658enum armv6_counters {
659 ARMV6_CYCLE_COUNTER = 1,
660 ARMV6_COUNTER0,
661 ARMV6_COUNTER1,
662};
663
664/*
665 * The hardware events that we support. We do support cache operations but
666 * we have harvard caches and no way to combine instruction and data
667 * accesses/misses in hardware.
668 */
669static const unsigned armv6_perf_map[PERF_COUNT_HW_MAX] = {
670 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6_PERFCTR_CPU_CYCLES,
671 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6_PERFCTR_INSTR_EXEC,
672 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
673 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
674 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6_PERFCTR_BR_EXEC,
675 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6_PERFCTR_BR_MISPREDICT,
676 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
677};
678
679static const unsigned armv6_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
680 [PERF_COUNT_HW_CACHE_OP_MAX]
681 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
682 [C(L1D)] = {
683 /*
684 * The performance counters don't differentiate between read
685 * and write accesses/misses so this isn't strictly correct,
686 * but it's the best we can do. Writes and reads get
687 * combined.
688 */
689 [C(OP_READ)] = {
690 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
691 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
692 },
693 [C(OP_WRITE)] = {
694 [C(RESULT_ACCESS)] = ARMV6_PERFCTR_DCACHE_ACCESS,
695 [C(RESULT_MISS)] = ARMV6_PERFCTR_DCACHE_MISS,
696 },
697 [C(OP_PREFETCH)] = {
698 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
699 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
700 },
701 },
702 [C(L1I)] = {
703 [C(OP_READ)] = {
704 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
705 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
706 },
707 [C(OP_WRITE)] = {
708 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
709 [C(RESULT_MISS)] = ARMV6_PERFCTR_ICACHE_MISS,
710 },
711 [C(OP_PREFETCH)] = {
712 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
713 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
714 },
715 },
716 [C(LL)] = {
717 [C(OP_READ)] = {
718 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
719 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
720 },
721 [C(OP_WRITE)] = {
722 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
723 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
724 },
725 [C(OP_PREFETCH)] = {
726 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
727 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
728 },
729 },
730 [C(DTLB)] = {
731 /*
732 * The ARM performance counters can count micro DTLB misses,
733 * micro ITLB misses and main TLB misses. There isn't an event
734 * for TLB misses, so use the micro misses here and if users
735 * want the main TLB misses they can use a raw counter.
736 */
737 [C(OP_READ)] = {
738 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
739 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
740 },
741 [C(OP_WRITE)] = {
742 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
743 [C(RESULT_MISS)] = ARMV6_PERFCTR_DTLB_MISS,
744 },
745 [C(OP_PREFETCH)] = {
746 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
747 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
748 },
749 },
750 [C(ITLB)] = {
751 [C(OP_READ)] = {
752 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
753 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
754 },
755 [C(OP_WRITE)] = {
756 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
757 [C(RESULT_MISS)] = ARMV6_PERFCTR_ITLB_MISS,
758 },
759 [C(OP_PREFETCH)] = {
760 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
761 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
762 },
763 },
764 [C(BPU)] = {
765 [C(OP_READ)] = {
766 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
767 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
768 },
769 [C(OP_WRITE)] = {
770 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
771 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
772 },
773 [C(OP_PREFETCH)] = {
774 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
775 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
776 },
777 },
778};
779
780enum armv6mpcore_perf_types {
781 ARMV6MPCORE_PERFCTR_ICACHE_MISS = 0x0,
782 ARMV6MPCORE_PERFCTR_IBUF_STALL = 0x1,
783 ARMV6MPCORE_PERFCTR_DDEP_STALL = 0x2,
784 ARMV6MPCORE_PERFCTR_ITLB_MISS = 0x3,
785 ARMV6MPCORE_PERFCTR_DTLB_MISS = 0x4,
786 ARMV6MPCORE_PERFCTR_BR_EXEC = 0x5,
787 ARMV6MPCORE_PERFCTR_BR_NOTPREDICT = 0x6,
788 ARMV6MPCORE_PERFCTR_BR_MISPREDICT = 0x7,
789 ARMV6MPCORE_PERFCTR_INSTR_EXEC = 0x8,
790 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS = 0xA,
791 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS = 0xB,
792 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS = 0xC,
793 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS = 0xD,
794 ARMV6MPCORE_PERFCTR_DCACHE_EVICTION = 0xE,
795 ARMV6MPCORE_PERFCTR_SW_PC_CHANGE = 0xF,
796 ARMV6MPCORE_PERFCTR_MAIN_TLB_MISS = 0x10,
797 ARMV6MPCORE_PERFCTR_EXPL_MEM_ACCESS = 0x11,
798 ARMV6MPCORE_PERFCTR_LSU_FULL_STALL = 0x12,
799 ARMV6MPCORE_PERFCTR_WBUF_DRAINED = 0x13,
800 ARMV6MPCORE_PERFCTR_CPU_CYCLES = 0xFF,
801};
802
803/*
804 * The hardware events that we support. We do support cache operations but
805 * we have harvard caches and no way to combine instruction and data
806 * accesses/misses in hardware.
807 */
808static const unsigned armv6mpcore_perf_map[PERF_COUNT_HW_MAX] = {
809 [PERF_COUNT_HW_CPU_CYCLES] = ARMV6MPCORE_PERFCTR_CPU_CYCLES,
810 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_INSTR_EXEC,
811 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
812 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
813 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV6MPCORE_PERFCTR_BR_EXEC,
814 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV6MPCORE_PERFCTR_BR_MISPREDICT,
815 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
816};
817
818static const unsigned armv6mpcore_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
819 [PERF_COUNT_HW_CACHE_OP_MAX]
820 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
821 [C(L1D)] = {
822 [C(OP_READ)] = {
823 [C(RESULT_ACCESS)] =
824 ARMV6MPCORE_PERFCTR_DCACHE_RDACCESS,
825 [C(RESULT_MISS)] =
826 ARMV6MPCORE_PERFCTR_DCACHE_RDMISS,
827 },
828 [C(OP_WRITE)] = {
829 [C(RESULT_ACCESS)] =
830 ARMV6MPCORE_PERFCTR_DCACHE_WRACCESS,
831 [C(RESULT_MISS)] =
832 ARMV6MPCORE_PERFCTR_DCACHE_WRMISS,
833 },
834 [C(OP_PREFETCH)] = {
835 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
836 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
837 },
838 },
839 [C(L1I)] = {
840 [C(OP_READ)] = {
841 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
842 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
843 },
844 [C(OP_WRITE)] = {
845 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
846 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ICACHE_MISS,
847 },
848 [C(OP_PREFETCH)] = {
849 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
850 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
851 },
852 },
853 [C(LL)] = {
854 [C(OP_READ)] = {
855 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
856 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
857 },
858 [C(OP_WRITE)] = {
859 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
860 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
861 },
862 [C(OP_PREFETCH)] = {
863 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
864 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
865 },
866 },
867 [C(DTLB)] = {
868 /*
869 * The ARM performance counters can count micro DTLB misses,
870 * micro ITLB misses and main TLB misses. There isn't an event
871 * for TLB misses, so use the micro misses here and if users
872 * want the main TLB misses they can use a raw counter.
873 */
874 [C(OP_READ)] = {
875 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
876 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
877 },
878 [C(OP_WRITE)] = {
879 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
880 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_DTLB_MISS,
881 },
882 [C(OP_PREFETCH)] = {
883 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
884 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
885 },
886 },
887 [C(ITLB)] = {
888 [C(OP_READ)] = {
889 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
890 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
891 },
892 [C(OP_WRITE)] = {
893 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
894 [C(RESULT_MISS)] = ARMV6MPCORE_PERFCTR_ITLB_MISS,
895 },
896 [C(OP_PREFETCH)] = {
897 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
898 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
899 },
900 },
901 [C(BPU)] = {
902 [C(OP_READ)] = {
903 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
904 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
905 },
906 [C(OP_WRITE)] = {
907 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
908 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
909 },
910 [C(OP_PREFETCH)] = {
911 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
912 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
913 },
914 },
915};
916
917static inline unsigned long
918armv6_pmcr_read(void)
919{
920 u32 val;
921 asm volatile("mrc p15, 0, %0, c15, c12, 0" : "=r"(val));
922 return val;
923}
924
925static inline void
926armv6_pmcr_write(unsigned long val)
927{
928 asm volatile("mcr p15, 0, %0, c15, c12, 0" : : "r"(val));
929}
930
931#define ARMV6_PMCR_ENABLE (1 << 0)
932#define ARMV6_PMCR_CTR01_RESET (1 << 1)
933#define ARMV6_PMCR_CCOUNT_RESET (1 << 2)
934#define ARMV6_PMCR_CCOUNT_DIV (1 << 3)
935#define ARMV6_PMCR_COUNT0_IEN (1 << 4)
936#define ARMV6_PMCR_COUNT1_IEN (1 << 5)
937#define ARMV6_PMCR_CCOUNT_IEN (1 << 6)
938#define ARMV6_PMCR_COUNT0_OVERFLOW (1 << 8)
939#define ARMV6_PMCR_COUNT1_OVERFLOW (1 << 9)
940#define ARMV6_PMCR_CCOUNT_OVERFLOW (1 << 10)
941#define ARMV6_PMCR_EVT_COUNT0_SHIFT 20
942#define ARMV6_PMCR_EVT_COUNT0_MASK (0xFF << ARMV6_PMCR_EVT_COUNT0_SHIFT)
943#define ARMV6_PMCR_EVT_COUNT1_SHIFT 12
944#define ARMV6_PMCR_EVT_COUNT1_MASK (0xFF << ARMV6_PMCR_EVT_COUNT1_SHIFT)
945
946#define ARMV6_PMCR_OVERFLOWED_MASK \
947 (ARMV6_PMCR_COUNT0_OVERFLOW | ARMV6_PMCR_COUNT1_OVERFLOW | \
948 ARMV6_PMCR_CCOUNT_OVERFLOW)
949
950static inline int
951armv6_pmcr_has_overflowed(unsigned long pmcr)
952{
953 return (pmcr & ARMV6_PMCR_OVERFLOWED_MASK);
954}
955
956static inline int
957armv6_pmcr_counter_has_overflowed(unsigned long pmcr,
958 enum armv6_counters counter)
959{
960 int ret = 0;
961
962 if (ARMV6_CYCLE_COUNTER == counter)
963 ret = pmcr & ARMV6_PMCR_CCOUNT_OVERFLOW;
964 else if (ARMV6_COUNTER0 == counter)
965 ret = pmcr & ARMV6_PMCR_COUNT0_OVERFLOW;
966 else if (ARMV6_COUNTER1 == counter)
967 ret = pmcr & ARMV6_PMCR_COUNT1_OVERFLOW;
968 else
969 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
970
971 return ret;
972}
973
974static inline u32
975armv6pmu_read_counter(int counter)
976{
977 unsigned long value = 0;
978
979 if (ARMV6_CYCLE_COUNTER == counter)
980 asm volatile("mrc p15, 0, %0, c15, c12, 1" : "=r"(value));
981 else if (ARMV6_COUNTER0 == counter)
982 asm volatile("mrc p15, 0, %0, c15, c12, 2" : "=r"(value));
983 else if (ARMV6_COUNTER1 == counter)
984 asm volatile("mrc p15, 0, %0, c15, c12, 3" : "=r"(value));
985 else
986 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
987
988 return value;
989}
990
991static inline void
992armv6pmu_write_counter(int counter,
993 u32 value)
994{
995 if (ARMV6_CYCLE_COUNTER == counter)
996 asm volatile("mcr p15, 0, %0, c15, c12, 1" : : "r"(value));
997 else if (ARMV6_COUNTER0 == counter)
998 asm volatile("mcr p15, 0, %0, c15, c12, 2" : : "r"(value));
999 else if (ARMV6_COUNTER1 == counter)
1000 asm volatile("mcr p15, 0, %0, c15, c12, 3" : : "r"(value));
1001 else
1002 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
1003}
1004
1005void
1006armv6pmu_enable_event(struct hw_perf_event *hwc,
1007 int idx)
1008{
1009 unsigned long val, mask, evt, flags;
1010
1011 if (ARMV6_CYCLE_COUNTER == idx) {
1012 mask = 0;
1013 evt = ARMV6_PMCR_CCOUNT_IEN;
1014 } else if (ARMV6_COUNTER0 == idx) {
1015 mask = ARMV6_PMCR_EVT_COUNT0_MASK;
1016 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT0_SHIFT) |
1017 ARMV6_PMCR_COUNT0_IEN;
1018 } else if (ARMV6_COUNTER1 == idx) {
1019 mask = ARMV6_PMCR_EVT_COUNT1_MASK;
1020 evt = (hwc->config_base << ARMV6_PMCR_EVT_COUNT1_SHIFT) |
1021 ARMV6_PMCR_COUNT1_IEN;
1022 } else {
1023 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
1024 return;
1025 }
1026
1027 /*
1028 * Mask out the current event and set the counter to count the event
1029 * that we're interested in.
1030 */
1031 spin_lock_irqsave(&pmu_lock, flags);
1032 val = armv6_pmcr_read();
1033 val &= ~mask;
1034 val |= evt;
1035 armv6_pmcr_write(val);
1036 spin_unlock_irqrestore(&pmu_lock, flags);
1037}
1038
1039static irqreturn_t
1040armv6pmu_handle_irq(int irq_num,
1041 void *dev)
1042{
1043 unsigned long pmcr = armv6_pmcr_read();
1044 struct perf_sample_data data;
1045 struct cpu_hw_events *cpuc;
1046 struct pt_regs *regs;
1047 int idx;
1048
1049 if (!armv6_pmcr_has_overflowed(pmcr))
1050 return IRQ_NONE;
1051
1052 regs = get_irq_regs();
1053
1054 /*
1055 * The interrupts are cleared by writing the overflow flags back to
1056 * the control register. All of the other bits don't have any effect
1057 * if they are rewritten, so write the whole value back.
1058 */
1059 armv6_pmcr_write(pmcr);
1060
dc1d628a 1061 perf_sample_data_init(&data, 0);
1b8873a0
JI		 1062
1063 cpuc = &__get_cpu_var(cpu_hw_events);
1064 for (idx = 0; idx <= armpmu->num_events; ++idx) {
1065 struct perf_event *event = cpuc->events[idx];
1066 struct hw_perf_event *hwc;
1067
1068 if (!test_bit(idx, cpuc->active_mask))
1069 continue;
1070
1071 /*
1072 * We have a single interrupt for all counters. Check that
1073 * each counter has overflowed before we process it.
1074 */
1075 if (!armv6_pmcr_counter_has_overflowed(pmcr, idx))
1076 continue;
1077
1078 hwc = &event->hw;
1079 armpmu_event_update(event, hwc, idx);
1080 data.period = event->hw.last_period;
1081 if (!armpmu_event_set_period(event, hwc, idx))
1082 continue;
1083
1084 if (perf_event_overflow(event, 0, &data, regs))
1085 armpmu->disable(hwc, idx);
1086 }
1087
1088 /*
1089 * Handle the pending perf events.
1090 *
25d35847
WD		 1091 * Note: this call *must* be run with interrupts disabled. For
1092 * platforms that can have the PMU interrupts raised as an NMI, this
1b8873a0
JI		 1093 * will not work.
1094 */
e360adbe 1095 irq_work_run();
1b8873a0
JI		 1096
1097 return IRQ_HANDLED;
1098}
1099
1100static void
1101armv6pmu_start(void)
1102{
1103 unsigned long flags, val;
1104
1105 spin_lock_irqsave(&pmu_lock, flags);
1106 val = armv6_pmcr_read();
1107 val |= ARMV6_PMCR_ENABLE;
1108 armv6_pmcr_write(val);
1109 spin_unlock_irqrestore(&pmu_lock, flags);
1110}
1111
1112void
1113armv6pmu_stop(void)
1114{
1115 unsigned long flags, val;
1116
1117 spin_lock_irqsave(&pmu_lock, flags);
1118 val = armv6_pmcr_read();
1119 val &= ~ARMV6_PMCR_ENABLE;
1120 armv6_pmcr_write(val);
1121 spin_unlock_irqrestore(&pmu_lock, flags);
1122}
1123
1124static inline int
1125armv6pmu_event_map(int config)
1126{
1127 int mapping = armv6_perf_map[config];
1128 if (HW_OP_UNSUPPORTED == mapping)
1129 mapping = -EOPNOTSUPP;
1130 return mapping;
1131}
1132
1133static inline int
1134armv6mpcore_pmu_event_map(int config)
1135{
1136 int mapping = armv6mpcore_perf_map[config];
1137 if (HW_OP_UNSUPPORTED == mapping)
1138 mapping = -EOPNOTSUPP;
1139 return mapping;
1140}
1141
1142static u64
1143armv6pmu_raw_event(u64 config)
1144{
1145 return config & 0xff;
1146}
1147
1148static int
1149armv6pmu_get_event_idx(struct cpu_hw_events *cpuc,
1150 struct hw_perf_event *event)
1151{
1152 /* Always place a cycle counter into the cycle counter. */
1153 if (ARMV6_PERFCTR_CPU_CYCLES == event->config_base) {
1154 if (test_and_set_bit(ARMV6_CYCLE_COUNTER, cpuc->used_mask))
1155 return -EAGAIN;
1156
1157 return ARMV6_CYCLE_COUNTER;
1158 } else {
1159 /*
1160 * For anything other than a cycle counter, try and use
1161 * counter0 and counter1.
1162 */
1163 if (!test_and_set_bit(ARMV6_COUNTER1, cpuc->used_mask)) {
1164 return ARMV6_COUNTER1;
1165 }
1166
1167 if (!test_and_set_bit(ARMV6_COUNTER0, cpuc->used_mask)) {
1168 return ARMV6_COUNTER0;
1169 }
1170
1171 /* The counters are all in use. */
1172 return -EAGAIN;
1173 }
1174}
1175
1176static void
1177armv6pmu_disable_event(struct hw_perf_event *hwc,
1178 int idx)
1179{
1180 unsigned long val, mask, evt, flags;
1181
1182 if (ARMV6_CYCLE_COUNTER == idx) {
1183 mask = ARMV6_PMCR_CCOUNT_IEN;
1184 evt = 0;
1185 } else if (ARMV6_COUNTER0 == idx) {
1186 mask = ARMV6_PMCR_COUNT0_IEN | ARMV6_PMCR_EVT_COUNT0_MASK;
1187 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT0_SHIFT;
1188 } else if (ARMV6_COUNTER1 == idx) {
1189 mask = ARMV6_PMCR_COUNT1_IEN | ARMV6_PMCR_EVT_COUNT1_MASK;
1190 evt = ARMV6_PERFCTR_NOP << ARMV6_PMCR_EVT_COUNT1_SHIFT;
1191 } else {
1192 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
1193 return;
1194 }
1195
1196 /*
1197 * Mask out the current event and set the counter to count the number
1198 * of ETM bus signal assertion cycles. The external reporting should
1199 * be disabled and so this should never increment.
1200 */
1201 spin_lock_irqsave(&pmu_lock, flags);
1202 val = armv6_pmcr_read();
1203 val &= ~mask;
1204 val |= evt;
1205 armv6_pmcr_write(val);
1206 spin_unlock_irqrestore(&pmu_lock, flags);
1207}
1208
1209static void
1210armv6mpcore_pmu_disable_event(struct hw_perf_event *hwc,
1211 int idx)
1212{
1213 unsigned long val, mask, flags, evt = 0;
1214
1215 if (ARMV6_CYCLE_COUNTER == idx) {
1216 mask = ARMV6_PMCR_CCOUNT_IEN;
1217 } else if (ARMV6_COUNTER0 == idx) {
1218 mask = ARMV6_PMCR_COUNT0_IEN;
1219 } else if (ARMV6_COUNTER1 == idx) {
1220 mask = ARMV6_PMCR_COUNT1_IEN;
1221 } else {
1222 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
1223 return;
1224 }
1225
1226 /*
1227 * Unlike UP ARMv6, we don't have a way of stopping the counters. We
1228 * simply disable the interrupt reporting.
1229 */
1230 spin_lock_irqsave(&pmu_lock, flags);
1231 val = armv6_pmcr_read();
1232 val &= ~mask;
1233 val |= evt;
1234 armv6_pmcr_write(val);
1235 spin_unlock_irqrestore(&pmu_lock, flags);
1236}
1237
1238static const struct arm_pmu armv6pmu = {
181193f3 1239 .id = ARM_PERF_PMU_ID_V6,
1b8873a0
JI		 1240 .handle_irq = armv6pmu_handle_irq,
1241 .enable = armv6pmu_enable_event,
1242 .disable = armv6pmu_disable_event,
1243 .event_map = armv6pmu_event_map,
1244 .raw_event = armv6pmu_raw_event,
1245 .read_counter = armv6pmu_read_counter,
1246 .write_counter = armv6pmu_write_counter,
1247 .get_event_idx = armv6pmu_get_event_idx,
1248 .start = armv6pmu_start,
1249 .stop = armv6pmu_stop,
1250 .num_events = 3,
1251 .max_period = (1LLU << 32) - 1,
1252};
1253
1254/*
1255 * ARMv6mpcore is almost identical to single core ARMv6 with the exception
1256 * that some of the events have different enumerations and that there is no
1257 * *hack* to stop the programmable counters. To stop the counters we simply
1258 * disable the interrupt reporting and update the event. When unthrottling we
1259 * reset the period and enable the interrupt reporting.
1260 */
1261static const struct arm_pmu armv6mpcore_pmu = {
181193f3 1262 .id = ARM_PERF_PMU_ID_V6MP,
1b8873a0
JI		 1263 .handle_irq = armv6pmu_handle_irq,
1264 .enable = armv6pmu_enable_event,
1265 .disable = armv6mpcore_pmu_disable_event,
1266 .event_map = armv6mpcore_pmu_event_map,
1267 .raw_event = armv6pmu_raw_event,
1268 .read_counter = armv6pmu_read_counter,
1269 .write_counter = armv6pmu_write_counter,
1270 .get_event_idx = armv6pmu_get_event_idx,
1271 .start = armv6pmu_start,
1272 .stop = armv6pmu_stop,
1273 .num_events = 3,
1274 .max_period = (1LLU << 32) - 1,
1275};
1276
796d1295
JP		 1277/*
1278 * ARMv7 Cortex-A8 and Cortex-A9 Performance Events handling code.
1279 *
1280 * Copied from ARMv6 code, with the low level code inspired
1281 * by the ARMv7 Oprofile code.
1282 *
1283 * Cortex-A8 has up to 4 configurable performance counters and
1284 * a single cycle counter.
1285 * Cortex-A9 has up to 31 configurable performance counters and
1286 * a single cycle counter.
1287 *
1288 * All counters can be enabled/disabled and IRQ masked separately. The cycle
1289 * counter and all 4 performance counters together can be reset separately.
1290 */
1291
796d1295
JP		 1292/* Common ARMv7 event types */
1293enum armv7_perf_types {
1294 ARMV7_PERFCTR_PMNC_SW_INCR = 0x00,
1295 ARMV7_PERFCTR_IFETCH_MISS = 0x01,
1296 ARMV7_PERFCTR_ITLB_MISS = 0x02,
1297 ARMV7_PERFCTR_DCACHE_REFILL = 0x03,
1298 ARMV7_PERFCTR_DCACHE_ACCESS = 0x04,
1299 ARMV7_PERFCTR_DTLB_REFILL = 0x05,
1300 ARMV7_PERFCTR_DREAD = 0x06,
1301 ARMV7_PERFCTR_DWRITE = 0x07,
1302
1303 ARMV7_PERFCTR_EXC_TAKEN = 0x09,
1304 ARMV7_PERFCTR_EXC_EXECUTED = 0x0A,
1305 ARMV7_PERFCTR_CID_WRITE = 0x0B,
1306 /* ARMV7_PERFCTR_PC_WRITE is equivalent to HW_BRANCH_INSTRUCTIONS.
1307 * It counts:
1308 * - all branch instructions,
1309 * - instructions that explicitly write the PC,
1310 * - exception generating instructions.
1311 */
1312 ARMV7_PERFCTR_PC_WRITE = 0x0C,
1313 ARMV7_PERFCTR_PC_IMM_BRANCH = 0x0D,
1314 ARMV7_PERFCTR_UNALIGNED_ACCESS = 0x0F,
1315 ARMV7_PERFCTR_PC_BRANCH_MIS_PRED = 0x10,
1316 ARMV7_PERFCTR_CLOCK_CYCLES = 0x11,
1317
1318 ARMV7_PERFCTR_PC_BRANCH_MIS_USED = 0x12,
1319
1320 ARMV7_PERFCTR_CPU_CYCLES = 0xFF
1321};
1322
1323/* ARMv7 Cortex-A8 specific event types */
1324enum armv7_a8_perf_types {
1325 ARMV7_PERFCTR_INSTR_EXECUTED = 0x08,
1326
1327 ARMV7_PERFCTR_PC_PROC_RETURN = 0x0E,
1328
1329 ARMV7_PERFCTR_WRITE_BUFFER_FULL = 0x40,
1330 ARMV7_PERFCTR_L2_STORE_MERGED = 0x41,
1331 ARMV7_PERFCTR_L2_STORE_BUFF = 0x42,
1332 ARMV7_PERFCTR_L2_ACCESS = 0x43,
1333 ARMV7_PERFCTR_L2_CACH_MISS = 0x44,
1334 ARMV7_PERFCTR_AXI_READ_CYCLES = 0x45,
1335 ARMV7_PERFCTR_AXI_WRITE_CYCLES = 0x46,
1336 ARMV7_PERFCTR_MEMORY_REPLAY = 0x47,
1337 ARMV7_PERFCTR_UNALIGNED_ACCESS_REPLAY = 0x48,
1338 ARMV7_PERFCTR_L1_DATA_MISS = 0x49,
1339 ARMV7_PERFCTR_L1_INST_MISS = 0x4A,
1340 ARMV7_PERFCTR_L1_DATA_COLORING = 0x4B,
1341 ARMV7_PERFCTR_L1_NEON_DATA = 0x4C,
1342 ARMV7_PERFCTR_L1_NEON_CACH_DATA = 0x4D,
1343 ARMV7_PERFCTR_L2_NEON = 0x4E,
1344 ARMV7_PERFCTR_L2_NEON_HIT = 0x4F,
1345 ARMV7_PERFCTR_L1_INST = 0x50,
1346 ARMV7_PERFCTR_PC_RETURN_MIS_PRED = 0x51,
1347 ARMV7_PERFCTR_PC_BRANCH_FAILED = 0x52,
1348 ARMV7_PERFCTR_PC_BRANCH_TAKEN = 0x53,
1349 ARMV7_PERFCTR_PC_BRANCH_EXECUTED = 0x54,
1350 ARMV7_PERFCTR_OP_EXECUTED = 0x55,
1351 ARMV7_PERFCTR_CYCLES_INST_STALL = 0x56,
1352 ARMV7_PERFCTR_CYCLES_INST = 0x57,
1353 ARMV7_PERFCTR_CYCLES_NEON_DATA_STALL = 0x58,
1354 ARMV7_PERFCTR_CYCLES_NEON_INST_STALL = 0x59,
1355 ARMV7_PERFCTR_NEON_CYCLES = 0x5A,
1356
1357 ARMV7_PERFCTR_PMU0_EVENTS = 0x70,
1358 ARMV7_PERFCTR_PMU1_EVENTS = 0x71,
1359 ARMV7_PERFCTR_PMU_EVENTS = 0x72,
1360};
1361
1362/* ARMv7 Cortex-A9 specific event types */
1363enum armv7_a9_perf_types {
1364 ARMV7_PERFCTR_JAVA_HW_BYTECODE_EXEC = 0x40,
1365 ARMV7_PERFCTR_JAVA_SW_BYTECODE_EXEC = 0x41,
1366 ARMV7_PERFCTR_JAZELLE_BRANCH_EXEC = 0x42,
1367
1368 ARMV7_PERFCTR_COHERENT_LINE_MISS = 0x50,
1369 ARMV7_PERFCTR_COHERENT_LINE_HIT = 0x51,
1370
1371 ARMV7_PERFCTR_ICACHE_DEP_STALL_CYCLES = 0x60,
1372 ARMV7_PERFCTR_DCACHE_DEP_STALL_CYCLES = 0x61,
1373 ARMV7_PERFCTR_TLB_MISS_DEP_STALL_CYCLES = 0x62,
1374 ARMV7_PERFCTR_STREX_EXECUTED_PASSED = 0x63,
1375 ARMV7_PERFCTR_STREX_EXECUTED_FAILED = 0x64,
1376 ARMV7_PERFCTR_DATA_EVICTION = 0x65,
1377 ARMV7_PERFCTR_ISSUE_STAGE_NO_INST = 0x66,
1378 ARMV7_PERFCTR_ISSUE_STAGE_EMPTY = 0x67,
1379 ARMV7_PERFCTR_INST_OUT_OF_RENAME_STAGE = 0x68,
1380
1381 ARMV7_PERFCTR_PREDICTABLE_FUNCT_RETURNS = 0x6E,
1382
1383 ARMV7_PERFCTR_MAIN_UNIT_EXECUTED_INST = 0x70,
1384 ARMV7_PERFCTR_SECOND_UNIT_EXECUTED_INST = 0x71,
1385 ARMV7_PERFCTR_LD_ST_UNIT_EXECUTED_INST = 0x72,
1386 ARMV7_PERFCTR_FP_EXECUTED_INST = 0x73,
1387 ARMV7_PERFCTR_NEON_EXECUTED_INST = 0x74,
1388
1389 ARMV7_PERFCTR_PLD_FULL_DEP_STALL_CYCLES = 0x80,
1390 ARMV7_PERFCTR_DATA_WR_DEP_STALL_CYCLES = 0x81,
1391 ARMV7_PERFCTR_ITLB_MISS_DEP_STALL_CYCLES = 0x82,
1392 ARMV7_PERFCTR_DTLB_MISS_DEP_STALL_CYCLES = 0x83,
1393 ARMV7_PERFCTR_MICRO_ITLB_MISS_DEP_STALL_CYCLES = 0x84,
1394 ARMV7_PERFCTR_MICRO_DTLB_MISS_DEP_STALL_CYCLES = 0x85,
1395 ARMV7_PERFCTR_DMB_DEP_STALL_CYCLES = 0x86,
1396
1397 ARMV7_PERFCTR_INTGR_CLK_ENABLED_CYCLES = 0x8A,
1398 ARMV7_PERFCTR_DATA_ENGINE_CLK_EN_CYCLES = 0x8B,
1399
1400 ARMV7_PERFCTR_ISB_INST = 0x90,
1401 ARMV7_PERFCTR_DSB_INST = 0x91,
1402 ARMV7_PERFCTR_DMB_INST = 0x92,
1403 ARMV7_PERFCTR_EXT_INTERRUPTS = 0x93,
1404
1405 ARMV7_PERFCTR_PLE_CACHE_LINE_RQST_COMPLETED = 0xA0,
1406 ARMV7_PERFCTR_PLE_CACHE_LINE_RQST_SKIPPED = 0xA1,
1407 ARMV7_PERFCTR_PLE_FIFO_FLUSH = 0xA2,
1408 ARMV7_PERFCTR_PLE_RQST_COMPLETED = 0xA3,
1409 ARMV7_PERFCTR_PLE_FIFO_OVERFLOW = 0xA4,
1410 ARMV7_PERFCTR_PLE_RQST_PROG = 0xA5
1411};
1412
1413/*
1414 * Cortex-A8 HW events mapping
1415 *
1416 * The hardware events that we support. We do support cache operations but
1417 * we have harvard caches and no way to combine instruction and data
1418 * accesses/misses in hardware.
1419 */
1420static const unsigned armv7_a8_perf_map[PERF_COUNT_HW_MAX] = {
1421 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
1422 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED,
1423 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
1424 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
1425 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
1426 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1427 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
1428};
1429
1430static const unsigned armv7_a8_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
1431 [PERF_COUNT_HW_CACHE_OP_MAX]
1432 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1433 [C(L1D)] = {
1434 /*
1435 * The performance counters don't differentiate between read
1436 * and write accesses/misses so this isn't strictly correct,
1437 * but it's the best we can do. Writes and reads get
1438 * combined.
1439 */
1440 [C(OP_READ)] = {
1441 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1442 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1443 },
1444 [C(OP_WRITE)] = {
1445 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1446 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1447 },
1448 [C(OP_PREFETCH)] = {
1449 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1450 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1451 },
1452 },
1453 [C(L1I)] = {
1454 [C(OP_READ)] = {
1455 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_INST,
1456 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_INST_MISS,
1457 },
1458 [C(OP_WRITE)] = {
1459 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_INST,
1460 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_INST_MISS,
1461 },
1462 [C(OP_PREFETCH)] = {
1463 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1464 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1465 },
1466 },
1467 [C(LL)] = {
1468 [C(OP_READ)] = {
1469 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_ACCESS,
1470 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACH_MISS,
1471 },
1472 [C(OP_WRITE)] = {
1473 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_ACCESS,
1474 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACH_MISS,
1475 },
1476 [C(OP_PREFETCH)] = {
1477 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1478 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1479 },
1480 },
1481 [C(DTLB)] = {
1482 /*
1483 * Only ITLB misses and DTLB refills are supported.
1484 * If users want the DTLB refills misses a raw counter
1485 * must be used.
1486 */
1487 [C(OP_READ)] = {
1488 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1489 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1490 },
1491 [C(OP_WRITE)] = {
1492 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1493 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1494 },
1495 [C(OP_PREFETCH)] = {
1496 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1497 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1498 },
1499 },
1500 [C(ITLB)] = {
1501 [C(OP_READ)] = {
1502 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1503 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1504 },
1505 [C(OP_WRITE)] = {
1506 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1507 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1508 },
1509 [C(OP_PREFETCH)] = {
1510 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1511 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1512 },
1513 },
1514 [C(BPU)] = {
1515 [C(OP_READ)] = {
1516 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1517 [C(RESULT_MISS)]
1518 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1519 },
1520 [C(OP_WRITE)] = {
1521 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1522 [C(RESULT_MISS)]
1523 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1524 },
1525 [C(OP_PREFETCH)] = {
1526 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1527 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1528 },
1529 },
1530};
1531
1532/*
1533 * Cortex-A9 HW events mapping
1534 */
1535static const unsigned armv7_a9_perf_map[PERF_COUNT_HW_MAX] = {
1536 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES,
1537 [PERF_COUNT_HW_INSTRUCTIONS] =
1538 ARMV7_PERFCTR_INST_OUT_OF_RENAME_STAGE,
1539 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_COHERENT_LINE_HIT,
1540 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_COHERENT_LINE_MISS,
1541 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE,
1542 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1543 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_CLOCK_CYCLES,
1544};
1545
1546static const unsigned armv7_a9_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
1547 [PERF_COUNT_HW_CACHE_OP_MAX]
1548 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1549 [C(L1D)] = {
1550 /*
1551 * The performance counters don't differentiate between read
1552 * and write accesses/misses so this isn't strictly correct,
1553 * but it's the best we can do. Writes and reads get
1554 * combined.
1555 */
1556 [C(OP_READ)] = {
1557 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1558 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1559 },
1560 [C(OP_WRITE)] = {
1561 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_DCACHE_ACCESS,
1562 [C(RESULT_MISS)] = ARMV7_PERFCTR_DCACHE_REFILL,
1563 },
1564 [C(OP_PREFETCH)] = {
1565 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1566 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1567 },
1568 },
1569 [C(L1I)] = {
1570 [C(OP_READ)] = {
1571 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1572 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
1573 },
1574 [C(OP_WRITE)] = {
1575 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1576 [C(RESULT_MISS)] = ARMV7_PERFCTR_IFETCH_MISS,
1577 },
1578 [C(OP_PREFETCH)] = {
1579 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1580 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1581 },
1582 },
1583 [C(LL)] = {
1584 [C(OP_READ)] = {
1585 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1586 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1587 },
1588 [C(OP_WRITE)] = {
1589 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1590 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1591 },
1592 [C(OP_PREFETCH)] = {
1593 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1594 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1595 },
1596 },
1597 [C(DTLB)] = {
1598 /*
1599 * Only ITLB misses and DTLB refills are supported.
1600 * If users want the DTLB refills misses a raw counter
1601 * must be used.
1602 */
1603 [C(OP_READ)] = {
1604 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1605 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1606 },
1607 [C(OP_WRITE)] = {
1608 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1609 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL,
1610 },
1611 [C(OP_PREFETCH)] = {
1612 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1613 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1614 },
1615 },
1616 [C(ITLB)] = {
1617 [C(OP_READ)] = {
1618 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1619 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1620 },
1621 [C(OP_WRITE)] = {
1622 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1623 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_MISS,
1624 },
1625 [C(OP_PREFETCH)] = {
1626 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1627 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1628 },
1629 },
1630 [C(BPU)] = {
1631 [C(OP_READ)] = {
1632 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1633 [C(RESULT_MISS)]
1634 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1635 },
1636 [C(OP_WRITE)] = {
1637 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_WRITE,
1638 [C(RESULT_MISS)]
1639 = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED,
1640 },
1641 [C(OP_PREFETCH)] = {
1642 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
1643 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
1644 },
1645 },
1646};
1647
1648/*
1649 * Perf Events counters
1650 */
1651enum armv7_counters {
1652 ARMV7_CYCLE_COUNTER = 1, /* Cycle counter */
1653 ARMV7_COUNTER0 = 2, /* First event counter */
1654};
1655
1656/*
1657 * The cycle counter is ARMV7_CYCLE_COUNTER.
1658 * The first event counter is ARMV7_COUNTER0.
1659 * The last event counter is (ARMV7_COUNTER0 + armpmu->num_events - 1).
1660 */
1661#define ARMV7_COUNTER_LAST (ARMV7_COUNTER0 + armpmu->num_events - 1)
1662
1663/*
1664 * ARMv7 low level PMNC access
1665 */
1666
1667/*
1668 * Per-CPU PMNC: config reg
1669 */
1670#define ARMV7_PMNC_E (1 << 0) /* Enable all counters */
1671#define ARMV7_PMNC_P (1 << 1) /* Reset all counters */
1672#define ARMV7_PMNC_C (1 << 2) /* Cycle counter reset */
1673#define ARMV7_PMNC_D (1 << 3) /* CCNT counts every 64th cpu cycle */
1674#define ARMV7_PMNC_X (1 << 4) /* Export to ETM */
1675#define ARMV7_PMNC_DP (1 << 5) /* Disable CCNT if non-invasive debug*/
1676#define ARMV7_PMNC_N_SHIFT 11 /* Number of counters supported */
1677#define ARMV7_PMNC_N_MASK 0x1f
1678#define ARMV7_PMNC_MASK 0x3f /* Mask for writable bits */
1679
1680/*
1681 * Available counters
1682 */
1683#define ARMV7_CNT0 0 /* First event counter */
1684#define ARMV7_CCNT 31 /* Cycle counter */
1685
1686/* Perf Event to low level counters mapping */
1687#define ARMV7_EVENT_CNT_TO_CNTx (ARMV7_COUNTER0 - ARMV7_CNT0)
1688
1689/*
1690 * CNTENS: counters enable reg
1691 */
1692#define ARMV7_CNTENS_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1693#define ARMV7_CNTENS_C (1 << ARMV7_CCNT)
1694
1695/*
1696 * CNTENC: counters disable reg
1697 */
1698#define ARMV7_CNTENC_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1699#define ARMV7_CNTENC_C (1 << ARMV7_CCNT)
1700
1701/*
1702 * INTENS: counters overflow interrupt enable reg
1703 */
1704#define ARMV7_INTENS_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1705#define ARMV7_INTENS_C (1 << ARMV7_CCNT)
1706
1707/*
1708 * INTENC: counters overflow interrupt disable reg
1709 */
1710#define ARMV7_INTENC_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1711#define ARMV7_INTENC_C (1 << ARMV7_CCNT)
1712
1713/*
1714 * EVTSEL: Event selection reg
1715 */
d10fca9f 1716#define ARMV7_EVTSEL_MASK 0xff /* Mask for writable bits */
796d1295
JP		 1717
1718/*
1719 * SELECT: Counter selection reg
1720 */
1721#define ARMV7_SELECT_MASK 0x1f /* Mask for writable bits */
1722
1723/*
1724 * FLAG: counters overflow flag status reg
1725 */
1726#define ARMV7_FLAG_P(idx) (1 << (idx - ARMV7_EVENT_CNT_TO_CNTx))
1727#define ARMV7_FLAG_C (1 << ARMV7_CCNT)
1728#define ARMV7_FLAG_MASK 0xffffffff /* Mask for writable bits */
1729#define ARMV7_OVERFLOWED_MASK ARMV7_FLAG_MASK
1730
1731static inline unsigned long armv7_pmnc_read(void)
1732{
1733 u32 val;
1734 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(val));
1735 return val;
1736}
1737
1738static inline void armv7_pmnc_write(unsigned long val)
1739{
1740 val &= ARMV7_PMNC_MASK;
1741 asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r"(val));
1742}
1743
1744static inline int armv7_pmnc_has_overflowed(unsigned long pmnc)
1745{
1746 return pmnc & ARMV7_OVERFLOWED_MASK;
1747}
1748
1749static inline int armv7_pmnc_counter_has_overflowed(unsigned long pmnc,
1750 enum armv7_counters counter)
1751{
c3b291d9 1752 int ret = 0;
796d1295
JP		 1753
1754 if (counter == ARMV7_CYCLE_COUNTER)
1755 ret = pmnc & ARMV7_FLAG_C;
1756 else if ((counter >= ARMV7_COUNTER0) && (counter <= ARMV7_COUNTER_LAST))
1757 ret = pmnc & ARMV7_FLAG_P(counter);
1758 else
1759 pr_err("CPU%u checking wrong counter %d overflow status\n",
1760 smp_processor_id(), counter);
1761
1762 return ret;
1763}
1764
1765static inline int armv7_pmnc_select_counter(unsigned int idx)
1766{
1767 u32 val;
1768
1769 if ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST)) {
1770 pr_err("CPU%u selecting wrong PMNC counter"
1771 " %d\n", smp_processor_id(), idx);
1772 return -1;
1773 }
1774
1775 val = (idx - ARMV7_EVENT_CNT_TO_CNTx) & ARMV7_SELECT_MASK;
1776 asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (val));
1777
1778 return idx;
1779}
1780
1781static inline u32 armv7pmu_read_counter(int idx)
1782{
1783 unsigned long value = 0;
1784
1785 if (idx == ARMV7_CYCLE_COUNTER)
1786 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (value));
1787 else if ((idx >= ARMV7_COUNTER0) && (idx <= ARMV7_COUNTER_LAST)) {
1788 if (armv7_pmnc_select_counter(idx) == idx)
1789 asm volatile("mrc p15, 0, %0, c9, c13, 2"
1790 : "=r" (value));
1791 } else
1792 pr_err("CPU%u reading wrong counter %d\n",
1793 smp_processor_id(), idx);
1794
1795 return value;
1796}
1797
1798static inline void armv7pmu_write_counter(int idx, u32 value)
1799{
1800 if (idx == ARMV7_CYCLE_COUNTER)
1801 asm volatile("mcr p15, 0, %0, c9, c13, 0" : : "r" (value));
1802 else if ((idx >= ARMV7_COUNTER0) && (idx <= ARMV7_COUNTER_LAST)) {
1803 if (armv7_pmnc_select_counter(idx) == idx)
1804 asm volatile("mcr p15, 0, %0, c9, c13, 2"
1805 : : "r" (value));
1806 } else
1807 pr_err("CPU%u writing wrong counter %d\n",
1808 smp_processor_id(), idx);
1809}
1810
1811static inline void armv7_pmnc_write_evtsel(unsigned int idx, u32 val)
1812{
1813 if (armv7_pmnc_select_counter(idx) == idx) {
1814 val &= ARMV7_EVTSEL_MASK;
1815 asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (val));
1816 }
1817}
1818
1819static inline u32 armv7_pmnc_enable_counter(unsigned int idx)
1820{
1821 u32 val;
1822
1823 if ((idx != ARMV7_CYCLE_COUNTER) &&
1824 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1825 pr_err("CPU%u enabling wrong PMNC counter"
1826 " %d\n", smp_processor_id(), idx);
1827 return -1;
1828 }
1829
1830 if (idx == ARMV7_CYCLE_COUNTER)
1831 val = ARMV7_CNTENS_C;
1832 else
1833 val = ARMV7_CNTENS_P(idx);
1834
1835 asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (val));
1836
1837 return idx;
1838}
1839
1840static inline u32 armv7_pmnc_disable_counter(unsigned int idx)
1841{
1842 u32 val;
1843
1844
1845 if ((idx != ARMV7_CYCLE_COUNTER) &&
1846 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1847 pr_err("CPU%u disabling wrong PMNC counter"
1848 " %d\n", smp_processor_id(), idx);
1849 return -1;
1850 }
1851
1852 if (idx == ARMV7_CYCLE_COUNTER)
1853 val = ARMV7_CNTENC_C;
1854 else
1855 val = ARMV7_CNTENC_P(idx);
1856
1857 asm volatile("mcr p15, 0, %0, c9, c12, 2" : : "r" (val));
1858
1859 return idx;
1860}
1861
1862static inline u32 armv7_pmnc_enable_intens(unsigned int idx)
1863{
1864 u32 val;
1865
1866 if ((idx != ARMV7_CYCLE_COUNTER) &&
1867 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1868 pr_err("CPU%u enabling wrong PMNC counter"
1869 " interrupt enable %d\n", smp_processor_id(), idx);
1870 return -1;
1871 }
1872
1873 if (idx == ARMV7_CYCLE_COUNTER)
1874 val = ARMV7_INTENS_C;
1875 else
1876 val = ARMV7_INTENS_P(idx);
1877
1878 asm volatile("mcr p15, 0, %0, c9, c14, 1" : : "r" (val));
1879
1880 return idx;
1881}
1882
1883static inline u32 armv7_pmnc_disable_intens(unsigned int idx)
1884{
1885 u32 val;
1886
1887 if ((idx != ARMV7_CYCLE_COUNTER) &&
1888 ((idx < ARMV7_COUNTER0) || (idx > ARMV7_COUNTER_LAST))) {
1889 pr_err("CPU%u disabling wrong PMNC counter"
1890 " interrupt enable %d\n", smp_processor_id(), idx);
1891 return -1;
1892 }
1893
1894 if (idx == ARMV7_CYCLE_COUNTER)
1895 val = ARMV7_INTENC_C;
1896 else
1897 val = ARMV7_INTENC_P(idx);
1898
1899 asm volatile("mcr p15, 0, %0, c9, c14, 2" : : "r" (val));
1900
1901 return idx;
1902}
1903
1904static inline u32 armv7_pmnc_getreset_flags(void)
1905{
1906 u32 val;
1907
1908 /* Read */
1909 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
1910
1911 /* Write to clear flags */
1912 val &= ARMV7_FLAG_MASK;
1913 asm volatile("mcr p15, 0, %0, c9, c12, 3" : : "r" (val));
1914
1915 return val;
1916}
1917
1918#ifdef DEBUG
1919static void armv7_pmnc_dump_regs(void)
1920{
1921 u32 val;
1922 unsigned int cnt;
1923
1924 printk(KERN_INFO "PMNC registers dump:\n");
1925
1926 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (val));
1927 printk(KERN_INFO "PMNC =0x%08x\n", val);
1928
1929 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r" (val));
1930 printk(KERN_INFO "CNTENS=0x%08x\n", val);
1931
1932 asm volatile("mrc p15, 0, %0, c9, c14, 1" : "=r" (val));
1933 printk(KERN_INFO "INTENS=0x%08x\n", val);
1934
1935 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val));
1936 printk(KERN_INFO "FLAGS =0x%08x\n", val);
1937
1938 asm volatile("mrc p15, 0, %0, c9, c12, 5" : "=r" (val));
1939 printk(KERN_INFO "SELECT=0x%08x\n", val);
1940
1941 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (val));
1942 printk(KERN_INFO "CCNT =0x%08x\n", val);
1943
1944 for (cnt = ARMV7_COUNTER0; cnt < ARMV7_COUNTER_LAST; cnt++) {
1945 armv7_pmnc_select_counter(cnt);
1946 asm volatile("mrc p15, 0, %0, c9, c13, 2" : "=r" (val));
1947 printk(KERN_INFO "CNT[%d] count =0x%08x\n",
1948 cnt-ARMV7_EVENT_CNT_TO_CNTx, val);
1949 asm volatile("mrc p15, 0, %0, c9, c13, 1" : "=r" (val));
1950 printk(KERN_INFO "CNT[%d] evtsel=0x%08x\n",
1951 cnt-ARMV7_EVENT_CNT_TO_CNTx, val);
1952 }
1953}
1954#endif
1955
1956void armv7pmu_enable_event(struct hw_perf_event *hwc, int idx)
1957{
1958 unsigned long flags;
1959
1960 /*
1961 * Enable counter and interrupt, and set the counter to count
1962 * the event that we're interested in.
1963 */
1964 spin_lock_irqsave(&pmu_lock, flags);
1965
1966 /*
1967 * Disable counter
1968 */
1969 armv7_pmnc_disable_counter(idx);
1970
1971 /*
1972 * Set event (if destined for PMNx counters)
1973 * We don't need to set the event if it's a cycle count
1974 */
1975 if (idx != ARMV7_CYCLE_COUNTER)
1976 armv7_pmnc_write_evtsel(idx, hwc->config_base);
1977
1978 /*
1979 * Enable interrupt for this counter
1980 */
1981 armv7_pmnc_enable_intens(idx);
1982
1983 /*
1984 * Enable counter
1985 */
1986 armv7_pmnc_enable_counter(idx);
1987
1988 spin_unlock_irqrestore(&pmu_lock, flags);
1989}
1990
1991static void armv7pmu_disable_event(struct hw_perf_event *hwc, int idx)
1992{
1993 unsigned long flags;
1994
1995 /*
1996 * Disable counter and interrupt
1997 */
1998 spin_lock_irqsave(&pmu_lock, flags);
1999
2000 /*
2001 * Disable counter
2002 */
2003 armv7_pmnc_disable_counter(idx);
2004
2005 /*
2006 * Disable interrupt for this counter
2007 */
2008 armv7_pmnc_disable_intens(idx);
2009
2010 spin_unlock_irqrestore(&pmu_lock, flags);
2011}
2012
2013static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev)
2014{
2015 unsigned long pmnc;
2016 struct perf_sample_data data;
2017 struct cpu_hw_events *cpuc;
2018 struct pt_regs *regs;
2019 int idx;
2020
2021 /*
2022 * Get and reset the IRQ flags
2023 */
2024 pmnc = armv7_pmnc_getreset_flags();
2025
2026 /*
2027 * Did an overflow occur?
2028 */
2029 if (!armv7_pmnc_has_overflowed(pmnc))
2030 return IRQ_NONE;
2031
2032 /*
2033 * Handle the counter(s) overflow(s)
2034 */
2035 regs = get_irq_regs();
2036
dc1d628a 2037 perf_sample_data_init(&data, 0);
796d1295
JP		 2038
2039 cpuc = &__get_cpu_var(cpu_hw_events);
2040 for (idx = 0; idx <= armpmu->num_events; ++idx) {
2041 struct perf_event *event = cpuc->events[idx];
2042 struct hw_perf_event *hwc;
2043
2044 if (!test_bit(idx, cpuc->active_mask))
2045 continue;
2046
2047 /*
2048 * We have a single interrupt for all counters. Check that
2049 * each counter has overflowed before we process it.
2050 */
2051 if (!armv7_pmnc_counter_has_overflowed(pmnc, idx))
2052 continue;
2053
2054 hwc = &event->hw;
2055 armpmu_event_update(event, hwc, idx);
2056 data.period = event->hw.last_period;
2057 if (!armpmu_event_set_period(event, hwc, idx))
2058 continue;
2059
2060 if (perf_event_overflow(event, 0, &data, regs))
2061 armpmu->disable(hwc, idx);
2062 }
2063
2064 /*
2065 * Handle the pending perf events.
2066 *
25d35847
WD		 2067 * Note: this call *must* be run with interrupts disabled. For
2068 * platforms that can have the PMU interrupts raised as an NMI, this
796d1295
JP		 2069 * will not work.
2070 */
e360adbe 2071 irq_work_run();
796d1295
JP		 2072
2073 return IRQ_HANDLED;
2074}
2075
2076static void armv7pmu_start(void)
2077{
2078 unsigned long flags;
2079
2080 spin_lock_irqsave(&pmu_lock, flags);
2081 /* Enable all counters */
2082 armv7_pmnc_write(armv7_pmnc_read() | ARMV7_PMNC_E);
2083 spin_unlock_irqrestore(&pmu_lock, flags);
2084}
2085
2086static void armv7pmu_stop(void)
2087{
2088 unsigned long flags;
2089
2090 spin_lock_irqsave(&pmu_lock, flags);
2091 /* Disable all counters */
2092 armv7_pmnc_write(armv7_pmnc_read() & ~ARMV7_PMNC_E);
2093 spin_unlock_irqrestore(&pmu_lock, flags);
2094}
2095
2096static inline int armv7_a8_pmu_event_map(int config)
2097{
2098 int mapping = armv7_a8_perf_map[config];
2099 if (HW_OP_UNSUPPORTED == mapping)
2100 mapping = -EOPNOTSUPP;
2101 return mapping;
2102}
2103
2104static inline int armv7_a9_pmu_event_map(int config)
2105{
2106 int mapping = armv7_a9_perf_map[config];
2107 if (HW_OP_UNSUPPORTED == mapping)
2108 mapping = -EOPNOTSUPP;
2109 return mapping;
2110}
2111
2112static u64 armv7pmu_raw_event(u64 config)
2113{
2114 return config & 0xff;
2115}
2116
2117static int armv7pmu_get_event_idx(struct cpu_hw_events *cpuc,
2118 struct hw_perf_event *event)
2119{
2120 int idx;
2121
2122 /* Always place a cycle counter into the cycle counter. */
2123 if (event->config_base == ARMV7_PERFCTR_CPU_CYCLES) {
2124 if (test_and_set_bit(ARMV7_CYCLE_COUNTER, cpuc->used_mask))
2125 return -EAGAIN;
2126
2127 return ARMV7_CYCLE_COUNTER;
2128 } else {
2129 /*
2130 * For anything other than a cycle counter, try and use
2131 * the events counters
2132 */
2133 for (idx = ARMV7_COUNTER0; idx <= armpmu->num_events; ++idx) {
2134 if (!test_and_set_bit(idx, cpuc->used_mask))
2135 return idx;
2136 }
2137
2138 /* The counters are all in use. */
2139 return -EAGAIN;
2140 }
2141}
2142
2143static struct arm_pmu armv7pmu = {
2144 .handle_irq = armv7pmu_handle_irq,
2145 .enable = armv7pmu_enable_event,
2146 .disable = armv7pmu_disable_event,
2147 .raw_event = armv7pmu_raw_event,
2148 .read_counter = armv7pmu_read_counter,
2149 .write_counter = armv7pmu_write_counter,
2150 .get_event_idx = armv7pmu_get_event_idx,
2151 .start = armv7pmu_start,
2152 .stop = armv7pmu_stop,
2153 .max_period = (1LLU << 32) - 1,
2154};
2155
2156static u32 __init armv7_reset_read_pmnc(void)
2157{
2158 u32 nb_cnt;
2159
2160 /* Initialize & Reset PMNC: C and P bits */
2161 armv7_pmnc_write(ARMV7_PMNC_P | ARMV7_PMNC_C);
2162
2163 /* Read the nb of CNTx counters supported from PMNC */
2164 nb_cnt = (armv7_pmnc_read() >> ARMV7_PMNC_N_SHIFT) & ARMV7_PMNC_N_MASK;
2165
2166 /* Add the CPU cycles counter and return */
2167 return nb_cnt + 1;
2168}
2169
49e6a32f
WD		 2170/*
2171 * ARMv5 [xscale] Performance counter handling code.
2172 *
2173 * Based on xscale OProfile code.
2174 *
2175 * There are two variants of the xscale PMU that we support:
2176 * - xscale1pmu: 2 event counters and a cycle counter
2177 * - xscale2pmu: 4 event counters and a cycle counter
2178 * The two variants share event definitions, but have different
2179 * PMU structures.
2180 */
2181
2182enum xscale_perf_types {
2183 XSCALE_PERFCTR_ICACHE_MISS = 0x00,
2184 XSCALE_PERFCTR_ICACHE_NO_DELIVER = 0x01,
2185 XSCALE_PERFCTR_DATA_STALL = 0x02,
2186 XSCALE_PERFCTR_ITLB_MISS = 0x03,
2187 XSCALE_PERFCTR_DTLB_MISS = 0x04,
2188 XSCALE_PERFCTR_BRANCH = 0x05,
2189 XSCALE_PERFCTR_BRANCH_MISS = 0x06,
2190 XSCALE_PERFCTR_INSTRUCTION = 0x07,
2191 XSCALE_PERFCTR_DCACHE_FULL_STALL = 0x08,
2192 XSCALE_PERFCTR_DCACHE_FULL_STALL_CONTIG = 0x09,
2193 XSCALE_PERFCTR_DCACHE_ACCESS = 0x0A,
2194 XSCALE_PERFCTR_DCACHE_MISS = 0x0B,
2195 XSCALE_PERFCTR_DCACHE_WRITE_BACK = 0x0C,
2196 XSCALE_PERFCTR_PC_CHANGED = 0x0D,
2197 XSCALE_PERFCTR_BCU_REQUEST = 0x10,
2198 XSCALE_PERFCTR_BCU_FULL = 0x11,
2199 XSCALE_PERFCTR_BCU_DRAIN = 0x12,
2200 XSCALE_PERFCTR_BCU_ECC_NO_ELOG = 0x14,
2201 XSCALE_PERFCTR_BCU_1_BIT_ERR = 0x15,
2202 XSCALE_PERFCTR_RMW = 0x16,
2203 /* XSCALE_PERFCTR_CCNT is not hardware defined */
2204 XSCALE_PERFCTR_CCNT = 0xFE,
2205 XSCALE_PERFCTR_UNUSED = 0xFF,
2206};
2207
2208enum xscale_counters {
2209 XSCALE_CYCLE_COUNTER = 1,
2210 XSCALE_COUNTER0,
2211 XSCALE_COUNTER1,
2212 XSCALE_COUNTER2,
2213 XSCALE_COUNTER3,
2214};
2215
2216static const unsigned xscale_perf_map[PERF_COUNT_HW_MAX] = {
2217 [PERF_COUNT_HW_CPU_CYCLES] = XSCALE_PERFCTR_CCNT,
2218 [PERF_COUNT_HW_INSTRUCTIONS] = XSCALE_PERFCTR_INSTRUCTION,
2219 [PERF_COUNT_HW_CACHE_REFERENCES] = HW_OP_UNSUPPORTED,
2220 [PERF_COUNT_HW_CACHE_MISSES] = HW_OP_UNSUPPORTED,
2221 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = XSCALE_PERFCTR_BRANCH,
2222 [PERF_COUNT_HW_BRANCH_MISSES] = XSCALE_PERFCTR_BRANCH_MISS,
2223 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED,
2224};
2225
2226static const unsigned xscale_perf_cache_map[PERF_COUNT_HW_CACHE_MAX]
2227 [PERF_COUNT_HW_CACHE_OP_MAX]
2228 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
2229 [C(L1D)] = {
2230 [C(OP_READ)] = {
2231 [C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
2232 [C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
2233 },
2234 [C(OP_WRITE)] = {
2235 [C(RESULT_ACCESS)] = XSCALE_PERFCTR_DCACHE_ACCESS,
2236 [C(RESULT_MISS)] = XSCALE_PERFCTR_DCACHE_MISS,
2237 },
2238 [C(OP_PREFETCH)] = {
2239 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2240 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2241 },
2242 },
2243 [C(L1I)] = {
2244 [C(OP_READ)] = {
2245 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2246 [C(RESULT_MISS)] = XSCALE_PERFCTR_ICACHE_MISS,
2247 },
2248 [C(OP_WRITE)] = {
2249 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2250 [C(RESULT_MISS)] = XSCALE_PERFCTR_ICACHE_MISS,
2251 },
2252 [C(OP_PREFETCH)] = {
2253 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2254 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2255 },
2256 },
2257 [C(LL)] = {
2258 [C(OP_READ)] = {
2259 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2260 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2261 },
2262 [C(OP_WRITE)] = {
2263 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2264 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2265 },
2266 [C(OP_PREFETCH)] = {
2267 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2268 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2269 },
2270 },
2271 [C(DTLB)] = {
2272 [C(OP_READ)] = {
2273 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2274 [C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
2275 },
2276 [C(OP_WRITE)] = {
2277 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2278 [C(RESULT_MISS)] = XSCALE_PERFCTR_DTLB_MISS,
2279 },
2280 [C(OP_PREFETCH)] = {
2281 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2282 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2283 },
2284 },
2285 [C(ITLB)] = {
2286 [C(OP_READ)] = {
2287 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2288 [C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
2289 },
2290 [C(OP_WRITE)] = {
2291 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2292 [C(RESULT_MISS)] = XSCALE_PERFCTR_ITLB_MISS,
2293 },
2294 [C(OP_PREFETCH)] = {
2295 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2296 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2297 },
2298 },
2299 [C(BPU)] = {
2300 [C(OP_READ)] = {
2301 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2302 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2303 },
2304 [C(OP_WRITE)] = {
2305 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2306 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2307 },
2308 [C(OP_PREFETCH)] = {
2309 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED,
2310 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED,
2311 },
2312 },
2313};
2314
2315#define XSCALE_PMU_ENABLE 0x001
2316#define XSCALE_PMN_RESET 0x002
2317#define XSCALE_CCNT_RESET 0x004
2318#define XSCALE_PMU_RESET (CCNT_RESET | PMN_RESET)
2319#define XSCALE_PMU_CNT64 0x008
2320
2321static inline int
2322xscalepmu_event_map(int config)
2323{
2324 int mapping = xscale_perf_map[config];
2325 if (HW_OP_UNSUPPORTED == mapping)
2326 mapping = -EOPNOTSUPP;
2327 return mapping;
2328}
2329
2330static u64
2331xscalepmu_raw_event(u64 config)
2332{
2333 return config & 0xff;
2334}
2335
2336#define XSCALE1_OVERFLOWED_MASK 0x700
2337#define XSCALE1_CCOUNT_OVERFLOW 0x400
2338#define XSCALE1_COUNT0_OVERFLOW 0x100
2339#define XSCALE1_COUNT1_OVERFLOW 0x200
2340#define XSCALE1_CCOUNT_INT_EN 0x040
2341#define XSCALE1_COUNT0_INT_EN 0x010
2342#define XSCALE1_COUNT1_INT_EN 0x020
2343#define XSCALE1_COUNT0_EVT_SHFT 12
2344#define XSCALE1_COUNT0_EVT_MASK (0xff << XSCALE1_COUNT0_EVT_SHFT)
2345#define XSCALE1_COUNT1_EVT_SHFT 20
2346#define XSCALE1_COUNT1_EVT_MASK (0xff << XSCALE1_COUNT1_EVT_SHFT)
2347
2348static inline u32
2349xscale1pmu_read_pmnc(void)
2350{
2351 u32 val;
2352 asm volatile("mrc p14, 0, %0, c0, c0, 0" : "=r" (val));
2353 return val;
2354}
2355
2356static inline void
2357xscale1pmu_write_pmnc(u32 val)
2358{
2359 /* upper 4bits and 7, 11 are write-as-0 */
2360 val &= 0xffff77f;
2361 asm volatile("mcr p14, 0, %0, c0, c0, 0" : : "r" (val));
2362}
2363
2364static inline int
2365xscale1_pmnc_counter_has_overflowed(unsigned long pmnc,
2366 enum xscale_counters counter)
2367{
2368 int ret = 0;
2369
2370 switch (counter) {
2371 case XSCALE_CYCLE_COUNTER:
2372 ret = pmnc & XSCALE1_CCOUNT_OVERFLOW;
2373 break;
2374 case XSCALE_COUNTER0:
2375 ret = pmnc & XSCALE1_COUNT0_OVERFLOW;
2376 break;
2377 case XSCALE_COUNTER1:
2378 ret = pmnc & XSCALE1_COUNT1_OVERFLOW;
2379 break;
2380 default:
2381 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
2382 }
2383
2384 return ret;
2385}
2386
2387static irqreturn_t
2388xscale1pmu_handle_irq(int irq_num, void *dev)
2389{
2390 unsigned long pmnc;
2391 struct perf_sample_data data;
2392 struct cpu_hw_events *cpuc;
2393 struct pt_regs *regs;
2394 int idx;
2395
2396 /*
2397 * NOTE: there's an A stepping erratum that states if an overflow
2398 * bit already exists and another occurs, the previous
2399 * Overflow bit gets cleared. There's no workaround.
2400 * Fixed in B stepping or later.
2401 */
2402 pmnc = xscale1pmu_read_pmnc();
2403
2404 /*
2405 * Write the value back to clear the overflow flags. Overflow
2406 * flags remain in pmnc for use below. We also disable the PMU
2407 * while we process the interrupt.
2408 */
2409 xscale1pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
2410
2411 if (!(pmnc & XSCALE1_OVERFLOWED_MASK))
2412 return IRQ_NONE;
2413
2414 regs = get_irq_regs();
2415
2416 perf_sample_data_init(&data, 0);
2417
2418 cpuc = &__get_cpu_var(cpu_hw_events);
2419 for (idx = 0; idx <= armpmu->num_events; ++idx) {
2420 struct perf_event *event = cpuc->events[idx];
2421 struct hw_perf_event *hwc;
2422
2423 if (!test_bit(idx, cpuc->active_mask))
2424 continue;
2425
2426 if (!xscale1_pmnc_counter_has_overflowed(pmnc, idx))
2427 continue;
2428
2429 hwc = &event->hw;
2430 armpmu_event_update(event, hwc, idx);
2431 data.period = event->hw.last_period;
2432 if (!armpmu_event_set_period(event, hwc, idx))
2433 continue;
2434
2435 if (perf_event_overflow(event, 0, &data, regs))
2436 armpmu->disable(hwc, idx);
2437 }
2438
e360adbe 2439 irq_work_run();
49e6a32f
WD		 2440
2441 /*
2442 * Re-enable the PMU.
2443 */
2444 pmnc = xscale1pmu_read_pmnc() | XSCALE_PMU_ENABLE;
2445 xscale1pmu_write_pmnc(pmnc);
2446
2447 return IRQ_HANDLED;
2448}
2449
2450static void
2451xscale1pmu_enable_event(struct hw_perf_event *hwc, int idx)
2452{
2453 unsigned long val, mask, evt, flags;
2454
2455 switch (idx) {
2456 case XSCALE_CYCLE_COUNTER:
2457 mask = 0;
2458 evt = XSCALE1_CCOUNT_INT_EN;
2459 break;
2460 case XSCALE_COUNTER0:
2461 mask = XSCALE1_COUNT0_EVT_MASK;
2462 evt = (hwc->config_base << XSCALE1_COUNT0_EVT_SHFT) |
2463 XSCALE1_COUNT0_INT_EN;
2464 break;
2465 case XSCALE_COUNTER1:
2466 mask = XSCALE1_COUNT1_EVT_MASK;
2467 evt = (hwc->config_base << XSCALE1_COUNT1_EVT_SHFT) |
2468 XSCALE1_COUNT1_INT_EN;
2469 break;
2470 default:
2471 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
2472 return;
2473 }
2474
2475 spin_lock_irqsave(&pmu_lock, flags);
2476 val = xscale1pmu_read_pmnc();
2477 val &= ~mask;
2478 val |= evt;
2479 xscale1pmu_write_pmnc(val);
2480 spin_unlock_irqrestore(&pmu_lock, flags);
2481}
2482
2483static void
2484xscale1pmu_disable_event(struct hw_perf_event *hwc, int idx)
2485{
2486 unsigned long val, mask, evt, flags;
2487
2488 switch (idx) {
2489 case XSCALE_CYCLE_COUNTER:
2490 mask = XSCALE1_CCOUNT_INT_EN;
2491 evt = 0;
2492 break;
2493 case XSCALE_COUNTER0:
2494 mask = XSCALE1_COUNT0_INT_EN | XSCALE1_COUNT0_EVT_MASK;
2495 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT0_EVT_SHFT;
2496 break;
2497 case XSCALE_COUNTER1:
2498 mask = XSCALE1_COUNT1_INT_EN | XSCALE1_COUNT1_EVT_MASK;
2499 evt = XSCALE_PERFCTR_UNUSED << XSCALE1_COUNT1_EVT_SHFT;
2500 break;
2501 default:
2502 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
2503 return;
2504 }
2505
2506 spin_lock_irqsave(&pmu_lock, flags);
2507 val = xscale1pmu_read_pmnc();
2508 val &= ~mask;
2509 val |= evt;
2510 xscale1pmu_write_pmnc(val);
2511 spin_unlock_irqrestore(&pmu_lock, flags);
2512}
2513
2514static int
2515xscale1pmu_get_event_idx(struct cpu_hw_events *cpuc,
2516 struct hw_perf_event *event)
2517{
2518 if (XSCALE_PERFCTR_CCNT == event->config_base) {
2519 if (test_and_set_bit(XSCALE_CYCLE_COUNTER, cpuc->used_mask))
2520 return -EAGAIN;
2521
2522 return XSCALE_CYCLE_COUNTER;
2523 } else {
2524 if (!test_and_set_bit(XSCALE_COUNTER1, cpuc->used_mask)) {
2525 return XSCALE_COUNTER1;
2526 }
2527
2528 if (!test_and_set_bit(XSCALE_COUNTER0, cpuc->used_mask)) {
2529 return XSCALE_COUNTER0;
2530 }
2531
2532 return -EAGAIN;
2533 }
2534}
2535
2536static void
2537xscale1pmu_start(void)
2538{
2539 unsigned long flags, val;
2540
2541 spin_lock_irqsave(&pmu_lock, flags);
2542 val = xscale1pmu_read_pmnc();
2543 val |= XSCALE_PMU_ENABLE;
2544 xscale1pmu_write_pmnc(val);
2545 spin_unlock_irqrestore(&pmu_lock, flags);
2546}
2547
2548static void
2549xscale1pmu_stop(void)
2550{
2551 unsigned long flags, val;
2552
2553 spin_lock_irqsave(&pmu_lock, flags);
2554 val = xscale1pmu_read_pmnc();
2555 val &= ~XSCALE_PMU_ENABLE;
2556 xscale1pmu_write_pmnc(val);
2557 spin_unlock_irqrestore(&pmu_lock, flags);
2558}
2559
2560static inline u32
2561xscale1pmu_read_counter(int counter)
2562{
2563 u32 val = 0;
2564
2565 switch (counter) {
2566 case XSCALE_CYCLE_COUNTER:
2567 asm volatile("mrc p14, 0, %0, c1, c0, 0" : "=r" (val));
2568 break;
2569 case XSCALE_COUNTER0:
2570 asm volatile("mrc p14, 0, %0, c2, c0, 0" : "=r" (val));
2571 break;
2572 case XSCALE_COUNTER1:
2573 asm volatile("mrc p14, 0, %0, c3, c0, 0" : "=r" (val));
2574 break;
2575 }
2576
2577 return val;
2578}
2579
2580static inline void
2581xscale1pmu_write_counter(int counter, u32 val)
2582{
2583 switch (counter) {
2584 case XSCALE_CYCLE_COUNTER:
2585 asm volatile("mcr p14, 0, %0, c1, c0, 0" : : "r" (val));
2586 break;
2587 case XSCALE_COUNTER0:
2588 asm volatile("mcr p14, 0, %0, c2, c0, 0" : : "r" (val));
2589 break;
2590 case XSCALE_COUNTER1:
2591 asm volatile("mcr p14, 0, %0, c3, c0, 0" : : "r" (val));
2592 break;
2593 }
2594}
2595
2596static const struct arm_pmu xscale1pmu = {
2597 .id = ARM_PERF_PMU_ID_XSCALE1,
2598 .handle_irq = xscale1pmu_handle_irq,
2599 .enable = xscale1pmu_enable_event,
2600 .disable = xscale1pmu_disable_event,
2601 .event_map = xscalepmu_event_map,
2602 .raw_event = xscalepmu_raw_event,
2603 .read_counter = xscale1pmu_read_counter,
2604 .write_counter = xscale1pmu_write_counter,
2605 .get_event_idx = xscale1pmu_get_event_idx,
2606 .start = xscale1pmu_start,
2607 .stop = xscale1pmu_stop,
2608 .num_events = 3,
2609 .max_period = (1LLU << 32) - 1,
2610};
2611
2612#define XSCALE2_OVERFLOWED_MASK 0x01f
2613#define XSCALE2_CCOUNT_OVERFLOW 0x001
2614#define XSCALE2_COUNT0_OVERFLOW 0x002
2615#define XSCALE2_COUNT1_OVERFLOW 0x004
2616#define XSCALE2_COUNT2_OVERFLOW 0x008
2617#define XSCALE2_COUNT3_OVERFLOW 0x010
2618#define XSCALE2_CCOUNT_INT_EN 0x001
2619#define XSCALE2_COUNT0_INT_EN 0x002
2620#define XSCALE2_COUNT1_INT_EN 0x004
2621#define XSCALE2_COUNT2_INT_EN 0x008
2622#define XSCALE2_COUNT3_INT_EN 0x010
2623#define XSCALE2_COUNT0_EVT_SHFT 0
2624#define XSCALE2_COUNT0_EVT_MASK (0xff << XSCALE2_COUNT0_EVT_SHFT)
2625#define XSCALE2_COUNT1_EVT_SHFT 8
2626#define XSCALE2_COUNT1_EVT_MASK (0xff << XSCALE2_COUNT1_EVT_SHFT)
2627#define XSCALE2_COUNT2_EVT_SHFT 16
2628#define XSCALE2_COUNT2_EVT_MASK (0xff << XSCALE2_COUNT2_EVT_SHFT)
2629#define XSCALE2_COUNT3_EVT_SHFT 24
2630#define XSCALE2_COUNT3_EVT_MASK (0xff << XSCALE2_COUNT3_EVT_SHFT)
2631
2632static inline u32
2633xscale2pmu_read_pmnc(void)
2634{
2635 u32 val;
2636 asm volatile("mrc p14, 0, %0, c0, c1, 0" : "=r" (val));
2637 /* bits 1-2 and 4-23 are read-unpredictable */
2638 return val & 0xff000009;
2639}
2640
2641static inline void
2642xscale2pmu_write_pmnc(u32 val)
2643{
2644 /* bits 4-23 are write-as-0, 24-31 are write ignored */
2645 val &= 0xf;
2646 asm volatile("mcr p14, 0, %0, c0, c1, 0" : : "r" (val));
2647}
2648
2649static inline u32
2650xscale2pmu_read_overflow_flags(void)
2651{
2652 u32 val;
2653 asm volatile("mrc p14, 0, %0, c5, c1, 0" : "=r" (val));
2654 return val;
2655}
2656
2657static inline void
2658xscale2pmu_write_overflow_flags(u32 val)
2659{
2660 asm volatile("mcr p14, 0, %0, c5, c1, 0" : : "r" (val));
2661}
2662
2663static inline u32
2664xscale2pmu_read_event_select(void)
2665{
2666 u32 val;
2667 asm volatile("mrc p14, 0, %0, c8, c1, 0" : "=r" (val));
2668 return val;
2669}
2670
2671static inline void
2672xscale2pmu_write_event_select(u32 val)
2673{
2674 asm volatile("mcr p14, 0, %0, c8, c1, 0" : : "r"(val));
2675}
2676
2677static inline u32
2678xscale2pmu_read_int_enable(void)
2679{
2680 u32 val;
2681 asm volatile("mrc p14, 0, %0, c4, c1, 0" : "=r" (val));
2682 return val;
2683}
2684
2685static void
2686xscale2pmu_write_int_enable(u32 val)
2687{
2688 asm volatile("mcr p14, 0, %0, c4, c1, 0" : : "r" (val));
2689}
2690
2691static inline int
2692xscale2_pmnc_counter_has_overflowed(unsigned long of_flags,
2693 enum xscale_counters counter)
2694{
2695 int ret = 0;
2696
2697 switch (counter) {
2698 case XSCALE_CYCLE_COUNTER:
2699 ret = of_flags & XSCALE2_CCOUNT_OVERFLOW;
2700 break;
2701 case XSCALE_COUNTER0:
2702 ret = of_flags & XSCALE2_COUNT0_OVERFLOW;
2703 break;
2704 case XSCALE_COUNTER1:
2705 ret = of_flags & XSCALE2_COUNT1_OVERFLOW;
2706 break;
2707 case XSCALE_COUNTER2:
2708 ret = of_flags & XSCALE2_COUNT2_OVERFLOW;
2709 break;
2710 case XSCALE_COUNTER3:
2711 ret = of_flags & XSCALE2_COUNT3_OVERFLOW;
2712 break;
2713 default:
2714 WARN_ONCE(1, "invalid counter number (%d)\n", counter);
2715 }
2716
2717 return ret;
2718}
2719
2720static irqreturn_t
2721xscale2pmu_handle_irq(int irq_num, void *dev)
2722{
2723 unsigned long pmnc, of_flags;
2724 struct perf_sample_data data;
2725 struct cpu_hw_events *cpuc;
2726 struct pt_regs *regs;
2727 int idx;
2728
2729 /* Disable the PMU. */
2730 pmnc = xscale2pmu_read_pmnc();
2731 xscale2pmu_write_pmnc(pmnc & ~XSCALE_PMU_ENABLE);
2732
2733 /* Check the overflow flag register. */
2734 of_flags = xscale2pmu_read_overflow_flags();
2735 if (!(of_flags & XSCALE2_OVERFLOWED_MASK))
2736 return IRQ_NONE;
2737
2738 /* Clear the overflow bits. */
2739 xscale2pmu_write_overflow_flags(of_flags);
2740
2741 regs = get_irq_regs();
2742
2743 perf_sample_data_init(&data, 0);
2744
2745 cpuc = &__get_cpu_var(cpu_hw_events);
2746 for (idx = 0; idx <= armpmu->num_events; ++idx) {
2747 struct perf_event *event = cpuc->events[idx];
2748 struct hw_perf_event *hwc;
2749
2750 if (!test_bit(idx, cpuc->active_mask))
2751 continue;
2752
2753 if (!xscale2_pmnc_counter_has_overflowed(pmnc, idx))
2754 continue;
2755
2756 hwc = &event->hw;
2757 armpmu_event_update(event, hwc, idx);
2758 data.period = event->hw.last_period;
2759 if (!armpmu_event_set_period(event, hwc, idx))
2760 continue;
2761
2762 if (perf_event_overflow(event, 0, &data, regs))
2763 armpmu->disable(hwc, idx);
2764 }
2765
e360adbe 2766 irq_work_run();
49e6a32f
WD		 2767
2768 /*
2769 * Re-enable the PMU.
2770 */
2771 pmnc = xscale2pmu_read_pmnc() | XSCALE_PMU_ENABLE;
2772 xscale2pmu_write_pmnc(pmnc);
2773
2774 return IRQ_HANDLED;
2775}
2776
2777static void
2778xscale2pmu_enable_event(struct hw_perf_event *hwc, int idx)
2779{
2780 unsigned long flags, ien, evtsel;
2781
2782 ien = xscale2pmu_read_int_enable();
2783 evtsel = xscale2pmu_read_event_select();
2784
2785 switch (idx) {
2786 case XSCALE_CYCLE_COUNTER:
2787 ien |= XSCALE2_CCOUNT_INT_EN;
2788 break;
2789 case XSCALE_COUNTER0:
2790 ien |= XSCALE2_COUNT0_INT_EN;
2791 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
2792 evtsel |= hwc->config_base << XSCALE2_COUNT0_EVT_SHFT;
2793 break;
2794 case XSCALE_COUNTER1:
2795 ien |= XSCALE2_COUNT1_INT_EN;
2796 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
2797 evtsel |= hwc->config_base << XSCALE2_COUNT1_EVT_SHFT;
2798 break;
2799 case XSCALE_COUNTER2:
2800 ien |= XSCALE2_COUNT2_INT_EN;
2801 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
2802 evtsel |= hwc->config_base << XSCALE2_COUNT2_EVT_SHFT;
2803 break;
2804 case XSCALE_COUNTER3:
2805 ien |= XSCALE2_COUNT3_INT_EN;
2806 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
2807 evtsel |= hwc->config_base << XSCALE2_COUNT3_EVT_SHFT;
2808 break;
2809 default:
2810 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
2811 return;
2812 }
2813
2814 spin_lock_irqsave(&pmu_lock, flags);
2815 xscale2pmu_write_event_select(evtsel);
2816 xscale2pmu_write_int_enable(ien);
2817 spin_unlock_irqrestore(&pmu_lock, flags);
2818}
2819
2820static void
2821xscale2pmu_disable_event(struct hw_perf_event *hwc, int idx)
2822{
2823 unsigned long flags, ien, evtsel;
2824
2825 ien = xscale2pmu_read_int_enable();
2826 evtsel = xscale2pmu_read_event_select();
2827
2828 switch (idx) {
2829 case XSCALE_CYCLE_COUNTER:
2830 ien &= ~XSCALE2_CCOUNT_INT_EN;
2831 break;
2832 case XSCALE_COUNTER0:
2833 ien &= ~XSCALE2_COUNT0_INT_EN;
2834 evtsel &= ~XSCALE2_COUNT0_EVT_MASK;
2835 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT0_EVT_SHFT;
2836 break;
2837 case XSCALE_COUNTER1:
2838 ien &= ~XSCALE2_COUNT1_INT_EN;
2839 evtsel &= ~XSCALE2_COUNT1_EVT_MASK;
2840 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT1_EVT_SHFT;
2841 break;
2842 case XSCALE_COUNTER2:
2843 ien &= ~XSCALE2_COUNT2_INT_EN;
2844 evtsel &= ~XSCALE2_COUNT2_EVT_MASK;
2845 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT2_EVT_SHFT;
2846 break;
2847 case XSCALE_COUNTER3:
2848 ien &= ~XSCALE2_COUNT3_INT_EN;
2849 evtsel &= ~XSCALE2_COUNT3_EVT_MASK;
2850 evtsel |= XSCALE_PERFCTR_UNUSED << XSCALE2_COUNT3_EVT_SHFT;
2851 break;
2852 default:
2853 WARN_ONCE(1, "invalid counter number (%d)\n", idx);
2854 return;
2855 }
2856
2857 spin_lock_irqsave(&pmu_lock, flags);
2858 xscale2pmu_write_event_select(evtsel);
2859 xscale2pmu_write_int_enable(ien);
2860 spin_unlock_irqrestore(&pmu_lock, flags);
2861}
2862
2863static int
2864xscale2pmu_get_event_idx(struct cpu_hw_events *cpuc,
2865 struct hw_perf_event *event)
2866{
2867 int idx = xscale1pmu_get_event_idx(cpuc, event);
2868 if (idx >= 0)
2869 goto out;
2870
2871 if (!test_and_set_bit(XSCALE_COUNTER3, cpuc->used_mask))
2872 idx = XSCALE_COUNTER3;
2873 else if (!test_and_set_bit(XSCALE_COUNTER2, cpuc->used_mask))
2874 idx = XSCALE_COUNTER2;
2875out:
2876 return idx;
2877}
2878
2879static void
2880xscale2pmu_start(void)
2881{
2882 unsigned long flags, val;
2883
2884 spin_lock_irqsave(&pmu_lock, flags);
2885 val = xscale2pmu_read_pmnc() & ~XSCALE_PMU_CNT64;
2886 val |= XSCALE_PMU_ENABLE;
2887 xscale2pmu_write_pmnc(val);
2888 spin_unlock_irqrestore(&pmu_lock, flags);
2889}
2890
2891static void
2892xscale2pmu_stop(void)
2893{
2894 unsigned long flags, val;
2895
2896 spin_lock_irqsave(&pmu_lock, flags);
2897 val = xscale2pmu_read_pmnc();
2898 val &= ~XSCALE_PMU_ENABLE;
2899 xscale2pmu_write_pmnc(val);
2900 spin_unlock_irqrestore(&pmu_lock, flags);
2901}
2902
2903static inline u32
2904xscale2pmu_read_counter(int counter)
2905{
2906 u32 val = 0;
2907
2908 switch (counter) {
2909 case XSCALE_CYCLE_COUNTER:
2910 asm volatile("mrc p14, 0, %0, c1, c1, 0" : "=r" (val));
2911 break;
2912 case XSCALE_COUNTER0:
2913 asm volatile("mrc p14, 0, %0, c0, c2, 0" : "=r" (val));
2914 break;
2915 case XSCALE_COUNTER1:
2916 asm volatile("mrc p14, 0, %0, c1, c2, 0" : "=r" (val));
2917 break;
2918 case XSCALE_COUNTER2:
2919 asm volatile("mrc p14, 0, %0, c2, c2, 0" : "=r" (val));
2920 break;
2921 case XSCALE_COUNTER3:
2922 asm volatile("mrc p14, 0, %0, c3, c2, 0" : "=r" (val));
2923 break;
2924 }
2925
2926 return val;
2927}
2928
2929static inline void
2930xscale2pmu_write_counter(int counter, u32 val)
2931{
2932 switch (counter) {
2933 case XSCALE_CYCLE_COUNTER:
2934 asm volatile("mcr p14, 0, %0, c1, c1, 0" : : "r" (val));
2935 break;
2936 case XSCALE_COUNTER0:
2937 asm volatile("mcr p14, 0, %0, c0, c2, 0" : : "r" (val));
2938 break;
2939 case XSCALE_COUNTER1:
2940 asm volatile("mcr p14, 0, %0, c1, c2, 0" : : "r" (val));
2941 break;
2942 case XSCALE_COUNTER2:
2943 asm volatile("mcr p14, 0, %0, c2, c2, 0" : : "r" (val));
2944 break;
2945 case XSCALE_COUNTER3:
2946 asm volatile("mcr p14, 0, %0, c3, c2, 0" : : "r" (val));
2947 break;
2948 }
2949}
2950
2951static const struct arm_pmu xscale2pmu = {
2952 .id = ARM_PERF_PMU_ID_XSCALE2,
2953 .handle_irq = xscale2pmu_handle_irq,
2954 .enable = xscale2pmu_enable_event,
2955 .disable = xscale2pmu_disable_event,
2956 .event_map = xscalepmu_event_map,
2957 .raw_event = xscalepmu_raw_event,
2958 .read_counter = xscale2pmu_read_counter,
2959 .write_counter = xscale2pmu_write_counter,
2960 .get_event_idx = xscale2pmu_get_event_idx,
2961 .start = xscale2pmu_start,
2962 .stop = xscale2pmu_stop,
2963 .num_events = 5,
2964 .max_period = (1LLU << 32) - 1,
2965};
2966
1b8873a0
JI		 2967static int __init
2968init_hw_perf_events(void)
2969{
2970 unsigned long cpuid = read_cpuid_id();
2971 unsigned long implementor = (cpuid & 0xFF000000) >> 24;
2972 unsigned long part_number = (cpuid & 0xFFF0);
2973
49e6a32f 2974 /* ARM Ltd CPUs. */
1b8873a0
JI		 2975 if (0x41 == implementor) {
2976 switch (part_number) {
2977 case 0xB360: /* ARM1136 */
2978 case 0xB560: /* ARM1156 */
2979 case 0xB760: /* ARM1176 */
2980 armpmu = &armv6pmu;
2981 memcpy(armpmu_perf_cache_map, armv6_perf_cache_map,
2982 sizeof(armv6_perf_cache_map));
1b8873a0
JI		 2983 break;
2984 case 0xB020: /* ARM11mpcore */
2985 armpmu = &armv6mpcore_pmu;
2986 memcpy(armpmu_perf_cache_map,
2987 armv6mpcore_perf_cache_map,
2988 sizeof(armv6mpcore_perf_cache_map));
1b8873a0 2989 break;
796d1295 2990 case 0xC080: /* Cortex-A8 */
181193f3 2991 armv7pmu.id = ARM_PERF_PMU_ID_CA8;
796d1295
JP		 2992 memcpy(armpmu_perf_cache_map, armv7_a8_perf_cache_map,
2993 sizeof(armv7_a8_perf_cache_map));
2994 armv7pmu.event_map = armv7_a8_pmu_event_map;
2995 armpmu = &armv7pmu;
2996
2997 /* Reset PMNC and read the nb of CNTx counters
2998 supported */
2999 armv7pmu.num_events = armv7_reset_read_pmnc();
796d1295
JP		 3000 break;
3001 case 0xC090: /* Cortex-A9 */
181193f3 3002 armv7pmu.id = ARM_PERF_PMU_ID_CA9;
796d1295
JP		 3003 memcpy(armpmu_perf_cache_map, armv7_a9_perf_cache_map,
3004 sizeof(armv7_a9_perf_cache_map));
3005 armv7pmu.event_map = armv7_a9_pmu_event_map;
3006 armpmu = &armv7pmu;
3007
3008 /* Reset PMNC and read the nb of CNTx counters
3009 supported */
3010 armv7pmu.num_events = armv7_reset_read_pmnc();
796d1295 3011 break;
49e6a32f
WD		 3012 }
3013 /* Intel CPUs [xscale]. */
3014 } else if (0x69 == implementor) {
3015 part_number = (cpuid >> 13) & 0x7;
3016 switch (part_number) {
3017 case 1:
3018 armpmu = &xscale1pmu;
3019 memcpy(armpmu_perf_cache_map, xscale_perf_cache_map,
3020 sizeof(xscale_perf_cache_map));
49e6a32f
WD		 3021 break;
3022 case 2:
3023 armpmu = &xscale2pmu;
3024 memcpy(armpmu_perf_cache_map, xscale_perf_cache_map,
3025 sizeof(xscale_perf_cache_map));
49e6a32f 3026 break;
1b8873a0
JI		 3027 }
3028 }
3029
49e6a32f 3030 if (armpmu) {
796d1295 3031 pr_info("enabled with %s PMU driver, %d counters available\n",
49e6a32f
WD		 3032 arm_pmu_names[armpmu->id], armpmu->num_events);
3033 } else {
3034 pr_info("no hardware support available\n");
49e6a32f 3035 }
1b8873a0 3036
b0a873eb
PZ		 3037 perf_pmu_register(&pmu);
3038
1b8873a0
JI		 3039 return 0;
3040}
3041arch_initcall(init_hw_perf_events);
3042
3043/*
3044 * Callchain handling code.
3045 */
1b8873a0
JI		 3046
3047/*
3048 * The registers we're interested in are at the end of the variable
3049 * length saved register structure. The fp points at the end of this
3050 * structure so the address of this struct is:
3051 * (struct frame_tail *)(xxx->fp)-1
3052 *
3053 * This code has been adapted from the ARM OProfile support.
3054 */
3055struct frame_tail {
3056 struct frame_tail *fp;
3057 unsigned long sp;
3058 unsigned long lr;
3059} __attribute__((packed));
3060
3061/*
3062 * Get the return address for a single stackframe and return a pointer to the
3063 * next frame tail.
3064 */
3065static struct frame_tail *
3066user_backtrace(struct frame_tail *tail,
3067 struct perf_callchain_entry *entry)
3068{
3069 struct frame_tail buftail;
3070
3071 /* Also check accessibility of one struct frame_tail beyond */
3072 if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
3073 return NULL;
3074 if (__copy_from_user_inatomic(&buftail, tail, sizeof(buftail)))
3075 return NULL;
3076
70791ce9 3077 perf_callchain_store(entry, buftail.lr);
1b8873a0
JI		 3078
3079 /*
3080 * Frame pointers should strictly progress back up the stack
3081 * (towards higher addresses).
3082 */
3083 if (tail >= buftail.fp)
3084 return NULL;
3085
3086 return buftail.fp - 1;
3087}
3088
56962b44
FW		 3089void
3090perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
1b8873a0
JI		 3091{
3092 struct frame_tail *tail;
3093
1b8873a0
JI		 3094
3095 tail = (struct frame_tail *)regs->ARM_fp - 1;
3096
3097 while (tail && !((unsigned long)tail & 0x3))
3098 tail = user_backtrace(tail, entry);
3099}
3100
3101/*
3102 * Gets called by walk_stackframe() for every stackframe. This will be called
3103 * whist unwinding the stackframe and is like a subroutine return so we use
3104 * the PC.
3105 */
3106static int
3107callchain_trace(struct stackframe *fr,
3108 void *data)
3109{
3110 struct perf_callchain_entry *entry = data;
70791ce9 3111 perf_callchain_store(entry, fr->pc);
1b8873a0
JI		 3112 return 0;
3113}
3114
56962b44
FW		 3115void
3116perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1b8873a0
JI		 3117{
3118 struct stackframe fr;
3119
1b8873a0
JI		 3120 fr.fp = regs->ARM_fp;
3121 fr.sp = regs->ARM_sp;
3122 fr.lr = regs->ARM_lr;
3123 fr.pc = regs->ARM_pc;
3124 walk_stackframe(&fr, callchain_trace, entry);
3125}
Clone of davem's net-next-2.6 tree