Commit | Line | Data |
---|---|---|
03089688 WD |
1 | /* |
2 | * PMU support | |
3 | * | |
4 | * Copyright (C) 2012 ARM Limited | |
5 | * Author: Will Deacon <will.deacon@arm.com> | |
6 | * | |
7 | * This code is based heavily on the ARMv7 perf event code. | |
8 | * | |
9 | * This program is free software; you can redistribute it and/or modify | |
10 | * it under the terms of the GNU General Public License version 2 as | |
11 | * published by the Free Software Foundation. | |
12 | * | |
13 | * This program is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | * GNU General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU General Public License | |
19 | * along with this program. If not, see <http://www.gnu.org/licenses/>. | |
20 | */ | |
21 | #define pr_fmt(fmt) "hw perfevents: " fmt | |
22 | ||
23 | #include <linux/bitmap.h> | |
24 | #include <linux/interrupt.h> | |
25 | #include <linux/kernel.h> | |
26 | #include <linux/export.h> | |
27 | #include <linux/perf_event.h> | |
28 | #include <linux/platform_device.h> | |
29 | #include <linux/spinlock.h> | |
30 | #include <linux/uaccess.h> | |
31 | ||
32 | #include <asm/cputype.h> | |
33 | #include <asm/irq.h> | |
34 | #include <asm/irq_regs.h> | |
35 | #include <asm/pmu.h> | |
36 | #include <asm/stacktrace.h> | |
37 | ||
38 | /* | |
39 | * ARMv8 supports a maximum of 32 events. | |
40 | * The cycle counter is included in this total. | |
41 | */ | |
42 | #define ARMPMU_MAX_HWEVENTS 32 | |
43 | ||
44 | static DEFINE_PER_CPU(struct perf_event * [ARMPMU_MAX_HWEVENTS], hw_events); | |
45 | static DEFINE_PER_CPU(unsigned long [BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)], used_mask); | |
46 | static DEFINE_PER_CPU(struct pmu_hw_events, cpu_hw_events); | |
47 | ||
48 | #define to_arm_pmu(p) (container_of(p, struct arm_pmu, pmu)) | |
49 | ||
50 | /* Set at runtime when we know what CPU type we are. */ | |
51 | static struct arm_pmu *cpu_pmu; | |
52 | ||
53 | int | |
54 | armpmu_get_max_events(void) | |
55 | { | |
56 | int max_events = 0; | |
57 | ||
58 | if (cpu_pmu != NULL) | |
59 | max_events = cpu_pmu->num_events; | |
60 | ||
61 | return max_events; | |
62 | } | |
63 | EXPORT_SYMBOL_GPL(armpmu_get_max_events); | |
64 | ||
65 | int perf_num_counters(void) | |
66 | { | |
67 | return armpmu_get_max_events(); | |
68 | } | |
69 | EXPORT_SYMBOL_GPL(perf_num_counters); | |
70 | ||
71 | #define HW_OP_UNSUPPORTED 0xFFFF | |
72 | ||
73 | #define C(_x) \ | |
74 | PERF_COUNT_HW_CACHE_##_x | |
75 | ||
76 | #define CACHE_OP_UNSUPPORTED 0xFFFF | |
77 | ||
78 | static int | |
79 | armpmu_map_cache_event(const unsigned (*cache_map) | |
80 | [PERF_COUNT_HW_CACHE_MAX] | |
81 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
82 | [PERF_COUNT_HW_CACHE_RESULT_MAX], | |
83 | u64 config) | |
84 | { | |
85 | unsigned int cache_type, cache_op, cache_result, ret; | |
86 | ||
87 | cache_type = (config >> 0) & 0xff; | |
88 | if (cache_type >= PERF_COUNT_HW_CACHE_MAX) | |
89 | return -EINVAL; | |
90 | ||
91 | cache_op = (config >> 8) & 0xff; | |
92 | if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) | |
93 | return -EINVAL; | |
94 | ||
95 | cache_result = (config >> 16) & 0xff; | |
96 | if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) | |
97 | return -EINVAL; | |
98 | ||
99 | ret = (int)(*cache_map)[cache_type][cache_op][cache_result]; | |
100 | ||
101 | if (ret == CACHE_OP_UNSUPPORTED) | |
102 | return -ENOENT; | |
103 | ||
104 | return ret; | |
105 | } | |
106 | ||
107 | static int | |
108 | armpmu_map_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config) | |
109 | { | |
110 | int mapping = (*event_map)[config]; | |
111 | return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping; | |
112 | } | |
113 | ||
114 | static int | |
115 | armpmu_map_raw_event(u32 raw_event_mask, u64 config) | |
116 | { | |
117 | return (int)(config & raw_event_mask); | |
118 | } | |
119 | ||
120 | static int map_cpu_event(struct perf_event *event, | |
121 | const unsigned (*event_map)[PERF_COUNT_HW_MAX], | |
122 | const unsigned (*cache_map) | |
123 | [PERF_COUNT_HW_CACHE_MAX] | |
124 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
125 | [PERF_COUNT_HW_CACHE_RESULT_MAX], | |
126 | u32 raw_event_mask) | |
127 | { | |
128 | u64 config = event->attr.config; | |
129 | ||
130 | switch (event->attr.type) { | |
131 | case PERF_TYPE_HARDWARE: | |
132 | return armpmu_map_event(event_map, config); | |
133 | case PERF_TYPE_HW_CACHE: | |
134 | return armpmu_map_cache_event(cache_map, config); | |
135 | case PERF_TYPE_RAW: | |
136 | return armpmu_map_raw_event(raw_event_mask, config); | |
137 | } | |
138 | ||
139 | return -ENOENT; | |
140 | } | |
141 | ||
142 | int | |
143 | armpmu_event_set_period(struct perf_event *event, | |
144 | struct hw_perf_event *hwc, | |
145 | int idx) | |
146 | { | |
147 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
148 | s64 left = local64_read(&hwc->period_left); | |
149 | s64 period = hwc->sample_period; | |
150 | int ret = 0; | |
151 | ||
152 | if (unlikely(left <= -period)) { | |
153 | left = period; | |
154 | local64_set(&hwc->period_left, left); | |
155 | hwc->last_period = period; | |
156 | ret = 1; | |
157 | } | |
158 | ||
159 | if (unlikely(left <= 0)) { | |
160 | left += period; | |
161 | local64_set(&hwc->period_left, left); | |
162 | hwc->last_period = period; | |
163 | ret = 1; | |
164 | } | |
165 | ||
166 | if (left > (s64)armpmu->max_period) | |
167 | left = armpmu->max_period; | |
168 | ||
169 | local64_set(&hwc->prev_count, (u64)-left); | |
170 | ||
171 | armpmu->write_counter(idx, (u64)(-left) & 0xffffffff); | |
172 | ||
173 | perf_event_update_userpage(event); | |
174 | ||
175 | return ret; | |
176 | } | |
177 | ||
178 | u64 | |
179 | armpmu_event_update(struct perf_event *event, | |
180 | struct hw_perf_event *hwc, | |
181 | int idx) | |
182 | { | |
183 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
184 | u64 delta, prev_raw_count, new_raw_count; | |
185 | ||
186 | again: | |
187 | prev_raw_count = local64_read(&hwc->prev_count); | |
188 | new_raw_count = armpmu->read_counter(idx); | |
189 | ||
190 | if (local64_cmpxchg(&hwc->prev_count, prev_raw_count, | |
191 | new_raw_count) != prev_raw_count) | |
192 | goto again; | |
193 | ||
194 | delta = (new_raw_count - prev_raw_count) & armpmu->max_period; | |
195 | ||
196 | local64_add(delta, &event->count); | |
197 | local64_sub(delta, &hwc->period_left); | |
198 | ||
199 | return new_raw_count; | |
200 | } | |
201 | ||
202 | static void | |
203 | armpmu_read(struct perf_event *event) | |
204 | { | |
205 | struct hw_perf_event *hwc = &event->hw; | |
206 | ||
207 | /* Don't read disabled counters! */ | |
208 | if (hwc->idx < 0) | |
209 | return; | |
210 | ||
211 | armpmu_event_update(event, hwc, hwc->idx); | |
212 | } | |
213 | ||
214 | static void | |
215 | armpmu_stop(struct perf_event *event, int flags) | |
216 | { | |
217 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
218 | struct hw_perf_event *hwc = &event->hw; | |
219 | ||
220 | /* | |
221 | * ARM pmu always has to update the counter, so ignore | |
222 | * PERF_EF_UPDATE, see comments in armpmu_start(). | |
223 | */ | |
224 | if (!(hwc->state & PERF_HES_STOPPED)) { | |
225 | armpmu->disable(hwc, hwc->idx); | |
226 | barrier(); /* why? */ | |
227 | armpmu_event_update(event, hwc, hwc->idx); | |
228 | hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE; | |
229 | } | |
230 | } | |
231 | ||
232 | static void | |
233 | armpmu_start(struct perf_event *event, int flags) | |
234 | { | |
235 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
236 | struct hw_perf_event *hwc = &event->hw; | |
237 | ||
238 | /* | |
239 | * ARM pmu always has to reprogram the period, so ignore | |
240 | * PERF_EF_RELOAD, see the comment below. | |
241 | */ | |
242 | if (flags & PERF_EF_RELOAD) | |
243 | WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE)); | |
244 | ||
245 | hwc->state = 0; | |
246 | /* | |
247 | * Set the period again. Some counters can't be stopped, so when we | |
248 | * were stopped we simply disabled the IRQ source and the counter | |
249 | * may have been left counting. If we don't do this step then we may | |
250 | * get an interrupt too soon or *way* too late if the overflow has | |
251 | * happened since disabling. | |
252 | */ | |
253 | armpmu_event_set_period(event, hwc, hwc->idx); | |
254 | armpmu->enable(hwc, hwc->idx); | |
255 | } | |
256 | ||
257 | static void | |
258 | armpmu_del(struct perf_event *event, int flags) | |
259 | { | |
260 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
261 | struct pmu_hw_events *hw_events = armpmu->get_hw_events(); | |
262 | struct hw_perf_event *hwc = &event->hw; | |
263 | int idx = hwc->idx; | |
264 | ||
265 | WARN_ON(idx < 0); | |
266 | ||
267 | armpmu_stop(event, PERF_EF_UPDATE); | |
268 | hw_events->events[idx] = NULL; | |
269 | clear_bit(idx, hw_events->used_mask); | |
270 | ||
271 | perf_event_update_userpage(event); | |
272 | } | |
273 | ||
274 | static int | |
275 | armpmu_add(struct perf_event *event, int flags) | |
276 | { | |
277 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
278 | struct pmu_hw_events *hw_events = armpmu->get_hw_events(); | |
279 | struct hw_perf_event *hwc = &event->hw; | |
280 | int idx; | |
281 | int err = 0; | |
282 | ||
283 | perf_pmu_disable(event->pmu); | |
284 | ||
285 | /* If we don't have a space for the counter then finish early. */ | |
286 | idx = armpmu->get_event_idx(hw_events, hwc); | |
287 | if (idx < 0) { | |
288 | err = idx; | |
289 | goto out; | |
290 | } | |
291 | ||
292 | /* | |
293 | * If there is an event in the counter we are going to use then make | |
294 | * sure it is disabled. | |
295 | */ | |
296 | event->hw.idx = idx; | |
297 | armpmu->disable(hwc, idx); | |
298 | hw_events->events[idx] = event; | |
299 | ||
300 | hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE; | |
301 | if (flags & PERF_EF_START) | |
302 | armpmu_start(event, PERF_EF_RELOAD); | |
303 | ||
304 | /* Propagate our changes to the userspace mapping. */ | |
305 | perf_event_update_userpage(event); | |
306 | ||
307 | out: | |
308 | perf_pmu_enable(event->pmu); | |
309 | return err; | |
310 | } | |
311 | ||
312 | static int | |
313 | validate_event(struct pmu_hw_events *hw_events, | |
314 | struct perf_event *event) | |
315 | { | |
316 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
317 | struct hw_perf_event fake_event = event->hw; | |
318 | struct pmu *leader_pmu = event->group_leader->pmu; | |
319 | ||
320 | if (event->pmu != leader_pmu || event->state <= PERF_EVENT_STATE_OFF) | |
321 | return 1; | |
322 | ||
323 | return armpmu->get_event_idx(hw_events, &fake_event) >= 0; | |
324 | } | |
325 | ||
326 | static int | |
327 | validate_group(struct perf_event *event) | |
328 | { | |
329 | struct perf_event *sibling, *leader = event->group_leader; | |
330 | struct pmu_hw_events fake_pmu; | |
331 | DECLARE_BITMAP(fake_used_mask, ARMPMU_MAX_HWEVENTS); | |
332 | ||
333 | /* | |
334 | * Initialise the fake PMU. We only need to populate the | |
335 | * used_mask for the purposes of validation. | |
336 | */ | |
337 | memset(fake_used_mask, 0, sizeof(fake_used_mask)); | |
338 | fake_pmu.used_mask = fake_used_mask; | |
339 | ||
340 | if (!validate_event(&fake_pmu, leader)) | |
341 | return -EINVAL; | |
342 | ||
343 | list_for_each_entry(sibling, &leader->sibling_list, group_entry) { | |
344 | if (!validate_event(&fake_pmu, sibling)) | |
345 | return -EINVAL; | |
346 | } | |
347 | ||
348 | if (!validate_event(&fake_pmu, event)) | |
349 | return -EINVAL; | |
350 | ||
351 | return 0; | |
352 | } | |
353 | ||
354 | static void | |
355 | armpmu_release_hardware(struct arm_pmu *armpmu) | |
356 | { | |
357 | int i, irq, irqs; | |
358 | struct platform_device *pmu_device = armpmu->plat_device; | |
359 | ||
360 | irqs = min(pmu_device->num_resources, num_possible_cpus()); | |
361 | ||
362 | for (i = 0; i < irqs; ++i) { | |
363 | if (!cpumask_test_and_clear_cpu(i, &armpmu->active_irqs)) | |
364 | continue; | |
365 | irq = platform_get_irq(pmu_device, i); | |
366 | if (irq >= 0) | |
367 | free_irq(irq, armpmu); | |
368 | } | |
369 | } | |
370 | ||
371 | static int | |
372 | armpmu_reserve_hardware(struct arm_pmu *armpmu) | |
373 | { | |
374 | int i, err, irq, irqs; | |
375 | struct platform_device *pmu_device = armpmu->plat_device; | |
376 | ||
377 | if (!pmu_device) { | |
378 | pr_err("no PMU device registered\n"); | |
379 | return -ENODEV; | |
380 | } | |
381 | ||
382 | irqs = min(pmu_device->num_resources, num_possible_cpus()); | |
383 | if (irqs < 1) { | |
384 | pr_err("no irqs for PMUs defined\n"); | |
385 | return -ENODEV; | |
386 | } | |
387 | ||
388 | for (i = 0; i < irqs; ++i) { | |
389 | err = 0; | |
390 | irq = platform_get_irq(pmu_device, i); | |
391 | if (irq < 0) | |
392 | continue; | |
393 | ||
394 | /* | |
395 | * If we have a single PMU interrupt that we can't shift, | |
396 | * assume that we're running on a uniprocessor machine and | |
397 | * continue. Otherwise, continue without this interrupt. | |
398 | */ | |
399 | if (irq_set_affinity(irq, cpumask_of(i)) && irqs > 1) { | |
400 | pr_warning("unable to set irq affinity (irq=%d, cpu=%u)\n", | |
401 | irq, i); | |
402 | continue; | |
403 | } | |
404 | ||
405 | err = request_irq(irq, armpmu->handle_irq, | |
406 | IRQF_NOBALANCING, | |
407 | "arm-pmu", armpmu); | |
408 | if (err) { | |
409 | pr_err("unable to request IRQ%d for ARM PMU counters\n", | |
410 | irq); | |
411 | armpmu_release_hardware(armpmu); | |
412 | return err; | |
413 | } | |
414 | ||
415 | cpumask_set_cpu(i, &armpmu->active_irqs); | |
416 | } | |
417 | ||
418 | return 0; | |
419 | } | |
420 | ||
421 | static void | |
422 | hw_perf_event_destroy(struct perf_event *event) | |
423 | { | |
424 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
425 | atomic_t *active_events = &armpmu->active_events; | |
426 | struct mutex *pmu_reserve_mutex = &armpmu->reserve_mutex; | |
427 | ||
428 | if (atomic_dec_and_mutex_lock(active_events, pmu_reserve_mutex)) { | |
429 | armpmu_release_hardware(armpmu); | |
430 | mutex_unlock(pmu_reserve_mutex); | |
431 | } | |
432 | } | |
433 | ||
434 | static int | |
435 | event_requires_mode_exclusion(struct perf_event_attr *attr) | |
436 | { | |
437 | return attr->exclude_idle || attr->exclude_user || | |
438 | attr->exclude_kernel || attr->exclude_hv; | |
439 | } | |
440 | ||
441 | static int | |
442 | __hw_perf_event_init(struct perf_event *event) | |
443 | { | |
444 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
445 | struct hw_perf_event *hwc = &event->hw; | |
446 | int mapping, err; | |
447 | ||
448 | mapping = armpmu->map_event(event); | |
449 | ||
450 | if (mapping < 0) { | |
451 | pr_debug("event %x:%llx not supported\n", event->attr.type, | |
452 | event->attr.config); | |
453 | return mapping; | |
454 | } | |
455 | ||
456 | /* | |
457 | * We don't assign an index until we actually place the event onto | |
458 | * hardware. Use -1 to signify that we haven't decided where to put it | |
459 | * yet. For SMP systems, each core has it's own PMU so we can't do any | |
460 | * clever allocation or constraints checking at this point. | |
461 | */ | |
462 | hwc->idx = -1; | |
463 | hwc->config_base = 0; | |
464 | hwc->config = 0; | |
465 | hwc->event_base = 0; | |
466 | ||
467 | /* | |
468 | * Check whether we need to exclude the counter from certain modes. | |
469 | */ | |
470 | if ((!armpmu->set_event_filter || | |
471 | armpmu->set_event_filter(hwc, &event->attr)) && | |
472 | event_requires_mode_exclusion(&event->attr)) { | |
473 | pr_debug("ARM performance counters do not support mode exclusion\n"); | |
474 | return -EPERM; | |
475 | } | |
476 | ||
477 | /* | |
478 | * Store the event encoding into the config_base field. | |
479 | */ | |
480 | hwc->config_base |= (unsigned long)mapping; | |
481 | ||
482 | if (!hwc->sample_period) { | |
483 | /* | |
484 | * For non-sampling runs, limit the sample_period to half | |
485 | * of the counter width. That way, the new counter value | |
486 | * is far less likely to overtake the previous one unless | |
487 | * you have some serious IRQ latency issues. | |
488 | */ | |
489 | hwc->sample_period = armpmu->max_period >> 1; | |
490 | hwc->last_period = hwc->sample_period; | |
491 | local64_set(&hwc->period_left, hwc->sample_period); | |
492 | } | |
493 | ||
494 | err = 0; | |
495 | if (event->group_leader != event) { | |
496 | err = validate_group(event); | |
497 | if (err) | |
498 | return -EINVAL; | |
499 | } | |
500 | ||
501 | return err; | |
502 | } | |
503 | ||
504 | static int armpmu_event_init(struct perf_event *event) | |
505 | { | |
506 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
507 | int err = 0; | |
508 | atomic_t *active_events = &armpmu->active_events; | |
509 | ||
510 | if (armpmu->map_event(event) == -ENOENT) | |
511 | return -ENOENT; | |
512 | ||
513 | event->destroy = hw_perf_event_destroy; | |
514 | ||
515 | if (!atomic_inc_not_zero(active_events)) { | |
516 | mutex_lock(&armpmu->reserve_mutex); | |
517 | if (atomic_read(active_events) == 0) | |
518 | err = armpmu_reserve_hardware(armpmu); | |
519 | ||
520 | if (!err) | |
521 | atomic_inc(active_events); | |
522 | mutex_unlock(&armpmu->reserve_mutex); | |
523 | } | |
524 | ||
525 | if (err) | |
526 | return err; | |
527 | ||
528 | err = __hw_perf_event_init(event); | |
529 | if (err) | |
530 | hw_perf_event_destroy(event); | |
531 | ||
532 | return err; | |
533 | } | |
534 | ||
535 | static void armpmu_enable(struct pmu *pmu) | |
536 | { | |
537 | struct arm_pmu *armpmu = to_arm_pmu(pmu); | |
538 | struct pmu_hw_events *hw_events = armpmu->get_hw_events(); | |
539 | int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events); | |
540 | ||
541 | if (enabled) | |
542 | armpmu->start(); | |
543 | } | |
544 | ||
545 | static void armpmu_disable(struct pmu *pmu) | |
546 | { | |
547 | struct arm_pmu *armpmu = to_arm_pmu(pmu); | |
548 | armpmu->stop(); | |
549 | } | |
550 | ||
551 | static void __init armpmu_init(struct arm_pmu *armpmu) | |
552 | { | |
553 | atomic_set(&armpmu->active_events, 0); | |
554 | mutex_init(&armpmu->reserve_mutex); | |
555 | ||
556 | armpmu->pmu = (struct pmu) { | |
557 | .pmu_enable = armpmu_enable, | |
558 | .pmu_disable = armpmu_disable, | |
559 | .event_init = armpmu_event_init, | |
560 | .add = armpmu_add, | |
561 | .del = armpmu_del, | |
562 | .start = armpmu_start, | |
563 | .stop = armpmu_stop, | |
564 | .read = armpmu_read, | |
565 | }; | |
566 | } | |
567 | ||
568 | int __init armpmu_register(struct arm_pmu *armpmu, char *name, int type) | |
569 | { | |
570 | armpmu_init(armpmu); | |
571 | return perf_pmu_register(&armpmu->pmu, name, type); | |
572 | } | |
573 | ||
574 | /* | |
575 | * ARMv8 PMUv3 Performance Events handling code. | |
576 | * Common event types. | |
577 | */ | |
578 | enum armv8_pmuv3_perf_types { | |
579 | /* Required events. */ | |
580 | ARMV8_PMUV3_PERFCTR_PMNC_SW_INCR = 0x00, | |
581 | ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL = 0x03, | |
582 | ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS = 0x04, | |
583 | ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED = 0x10, | |
584 | ARMV8_PMUV3_PERFCTR_CLOCK_CYCLES = 0x11, | |
585 | ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED = 0x12, | |
586 | ||
587 | /* At least one of the following is required. */ | |
588 | ARMV8_PMUV3_PERFCTR_INSTR_EXECUTED = 0x08, | |
589 | ARMV8_PMUV3_PERFCTR_OP_SPEC = 0x1B, | |
590 | ||
591 | /* Common architectural events. */ | |
592 | ARMV8_PMUV3_PERFCTR_MEM_READ = 0x06, | |
593 | ARMV8_PMUV3_PERFCTR_MEM_WRITE = 0x07, | |
594 | ARMV8_PMUV3_PERFCTR_EXC_TAKEN = 0x09, | |
595 | ARMV8_PMUV3_PERFCTR_EXC_EXECUTED = 0x0A, | |
596 | ARMV8_PMUV3_PERFCTR_CID_WRITE = 0x0B, | |
597 | ARMV8_PMUV3_PERFCTR_PC_WRITE = 0x0C, | |
598 | ARMV8_PMUV3_PERFCTR_PC_IMM_BRANCH = 0x0D, | |
599 | ARMV8_PMUV3_PERFCTR_PC_PROC_RETURN = 0x0E, | |
600 | ARMV8_PMUV3_PERFCTR_MEM_UNALIGNED_ACCESS = 0x0F, | |
601 | ARMV8_PMUV3_PERFCTR_TTBR_WRITE = 0x1C, | |
602 | ||
603 | /* Common microarchitectural events. */ | |
604 | ARMV8_PMUV3_PERFCTR_L1_ICACHE_REFILL = 0x01, | |
605 | ARMV8_PMUV3_PERFCTR_ITLB_REFILL = 0x02, | |
606 | ARMV8_PMUV3_PERFCTR_DTLB_REFILL = 0x05, | |
607 | ARMV8_PMUV3_PERFCTR_MEM_ACCESS = 0x13, | |
608 | ARMV8_PMUV3_PERFCTR_L1_ICACHE_ACCESS = 0x14, | |
609 | ARMV8_PMUV3_PERFCTR_L1_DCACHE_WB = 0x15, | |
610 | ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS = 0x16, | |
611 | ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL = 0x17, | |
612 | ARMV8_PMUV3_PERFCTR_L2_CACHE_WB = 0x18, | |
613 | ARMV8_PMUV3_PERFCTR_BUS_ACCESS = 0x19, | |
614 | ARMV8_PMUV3_PERFCTR_MEM_ERROR = 0x1A, | |
615 | ARMV8_PMUV3_PERFCTR_BUS_CYCLES = 0x1D, | |
616 | ||
617 | /* | |
618 | * This isn't an architected event. | |
619 | * We detect this event number and use the cycle counter instead. | |
620 | */ | |
621 | ARMV8_PMUV3_PERFCTR_CPU_CYCLES = 0xFF, | |
622 | }; | |
623 | ||
624 | /* PMUv3 HW events mapping. */ | |
625 | static const unsigned armv8_pmuv3_perf_map[PERF_COUNT_HW_MAX] = { | |
626 | [PERF_COUNT_HW_CPU_CYCLES] = ARMV8_PMUV3_PERFCTR_CPU_CYCLES, | |
627 | [PERF_COUNT_HW_INSTRUCTIONS] = ARMV8_PMUV3_PERFCTR_INSTR_EXECUTED, | |
628 | [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, | |
629 | [PERF_COUNT_HW_CACHE_MISSES] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, | |
630 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = HW_OP_UNSUPPORTED, | |
631 | [PERF_COUNT_HW_BRANCH_MISSES] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, | |
632 | [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED, | |
633 | [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = HW_OP_UNSUPPORTED, | |
634 | [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = HW_OP_UNSUPPORTED, | |
635 | }; | |
636 | ||
637 | static const unsigned armv8_pmuv3_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] | |
638 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
639 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
640 | [C(L1D)] = { | |
641 | [C(OP_READ)] = { | |
642 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, | |
643 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, | |
644 | }, | |
645 | [C(OP_WRITE)] = { | |
646 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, | |
647 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, | |
648 | }, | |
649 | [C(OP_PREFETCH)] = { | |
650 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
651 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
652 | }, | |
653 | }, | |
654 | [C(L1I)] = { | |
655 | [C(OP_READ)] = { | |
656 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
657 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
658 | }, | |
659 | [C(OP_WRITE)] = { | |
660 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
661 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
662 | }, | |
663 | [C(OP_PREFETCH)] = { | |
664 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
665 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
666 | }, | |
667 | }, | |
668 | [C(LL)] = { | |
669 | [C(OP_READ)] = { | |
670 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
671 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
672 | }, | |
673 | [C(OP_WRITE)] = { | |
674 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
675 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
676 | }, | |
677 | [C(OP_PREFETCH)] = { | |
678 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
679 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
680 | }, | |
681 | }, | |
682 | [C(DTLB)] = { | |
683 | [C(OP_READ)] = { | |
684 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
685 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
686 | }, | |
687 | [C(OP_WRITE)] = { | |
688 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
689 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
690 | }, | |
691 | [C(OP_PREFETCH)] = { | |
692 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
693 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
694 | }, | |
695 | }, | |
696 | [C(ITLB)] = { | |
697 | [C(OP_READ)] = { | |
698 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
699 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
700 | }, | |
701 | [C(OP_WRITE)] = { | |
702 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
703 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
704 | }, | |
705 | [C(OP_PREFETCH)] = { | |
706 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
707 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
708 | }, | |
709 | }, | |
710 | [C(BPU)] = { | |
711 | [C(OP_READ)] = { | |
712 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, | |
713 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, | |
714 | }, | |
715 | [C(OP_WRITE)] = { | |
716 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, | |
717 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, | |
718 | }, | |
719 | [C(OP_PREFETCH)] = { | |
720 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
721 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
722 | }, | |
723 | }, | |
724 | [C(NODE)] = { | |
725 | [C(OP_READ)] = { | |
726 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
727 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
728 | }, | |
729 | [C(OP_WRITE)] = { | |
730 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
731 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
732 | }, | |
733 | [C(OP_PREFETCH)] = { | |
734 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
735 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
736 | }, | |
737 | }, | |
738 | }; | |
739 | ||
740 | /* | |
741 | * Perf Events' indices | |
742 | */ | |
743 | #define ARMV8_IDX_CYCLE_COUNTER 0 | |
744 | #define ARMV8_IDX_COUNTER0 1 | |
745 | #define ARMV8_IDX_COUNTER_LAST (ARMV8_IDX_CYCLE_COUNTER + cpu_pmu->num_events - 1) | |
746 | ||
747 | #define ARMV8_MAX_COUNTERS 32 | |
748 | #define ARMV8_COUNTER_MASK (ARMV8_MAX_COUNTERS - 1) | |
749 | ||
750 | /* | |
751 | * ARMv8 low level PMU access | |
752 | */ | |
753 | ||
754 | /* | |
755 | * Perf Event to low level counters mapping | |
756 | */ | |
757 | #define ARMV8_IDX_TO_COUNTER(x) \ | |
758 | (((x) - ARMV8_IDX_COUNTER0) & ARMV8_COUNTER_MASK) | |
759 | ||
760 | /* | |
761 | * Per-CPU PMCR: config reg | |
762 | */ | |
763 | #define ARMV8_PMCR_E (1 << 0) /* Enable all counters */ | |
764 | #define ARMV8_PMCR_P (1 << 1) /* Reset all counters */ | |
765 | #define ARMV8_PMCR_C (1 << 2) /* Cycle counter reset */ | |
766 | #define ARMV8_PMCR_D (1 << 3) /* CCNT counts every 64th cpu cycle */ | |
767 | #define ARMV8_PMCR_X (1 << 4) /* Export to ETM */ | |
768 | #define ARMV8_PMCR_DP (1 << 5) /* Disable CCNT if non-invasive debug*/ | |
769 | #define ARMV8_PMCR_N_SHIFT 11 /* Number of counters supported */ | |
770 | #define ARMV8_PMCR_N_MASK 0x1f | |
771 | #define ARMV8_PMCR_MASK 0x3f /* Mask for writable bits */ | |
772 | ||
773 | /* | |
774 | * PMOVSR: counters overflow flag status reg | |
775 | */ | |
776 | #define ARMV8_OVSR_MASK 0xffffffff /* Mask for writable bits */ | |
777 | #define ARMV8_OVERFLOWED_MASK ARMV8_OVSR_MASK | |
778 | ||
779 | /* | |
780 | * PMXEVTYPER: Event selection reg | |
781 | */ | |
782 | #define ARMV8_EVTYPE_MASK 0xc00000ff /* Mask for writable bits */ | |
783 | #define ARMV8_EVTYPE_EVENT 0xff /* Mask for EVENT bits */ | |
784 | ||
785 | /* | |
786 | * Event filters for PMUv3 | |
787 | */ | |
788 | #define ARMV8_EXCLUDE_EL1 (1 << 31) | |
789 | #define ARMV8_EXCLUDE_EL0 (1 << 30) | |
790 | #define ARMV8_INCLUDE_EL2 (1 << 27) | |
791 | ||
792 | static inline u32 armv8pmu_pmcr_read(void) | |
793 | { | |
794 | u32 val; | |
795 | asm volatile("mrs %0, pmcr_el0" : "=r" (val)); | |
796 | return val; | |
797 | } | |
798 | ||
799 | static inline void armv8pmu_pmcr_write(u32 val) | |
800 | { | |
801 | val &= ARMV8_PMCR_MASK; | |
802 | isb(); | |
803 | asm volatile("msr pmcr_el0, %0" :: "r" (val)); | |
804 | } | |
805 | ||
806 | static inline int armv8pmu_has_overflowed(u32 pmovsr) | |
807 | { | |
808 | return pmovsr & ARMV8_OVERFLOWED_MASK; | |
809 | } | |
810 | ||
811 | static inline int armv8pmu_counter_valid(int idx) | |
812 | { | |
813 | return idx >= ARMV8_IDX_CYCLE_COUNTER && idx <= ARMV8_IDX_COUNTER_LAST; | |
814 | } | |
815 | ||
816 | static inline int armv8pmu_counter_has_overflowed(u32 pmnc, int idx) | |
817 | { | |
818 | int ret = 0; | |
819 | u32 counter; | |
820 | ||
821 | if (!armv8pmu_counter_valid(idx)) { | |
822 | pr_err("CPU%u checking wrong counter %d overflow status\n", | |
823 | smp_processor_id(), idx); | |
824 | } else { | |
825 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
826 | ret = pmnc & BIT(counter); | |
827 | } | |
828 | ||
829 | return ret; | |
830 | } | |
831 | ||
832 | static inline int armv8pmu_select_counter(int idx) | |
833 | { | |
834 | u32 counter; | |
835 | ||
836 | if (!armv8pmu_counter_valid(idx)) { | |
837 | pr_err("CPU%u selecting wrong PMNC counter %d\n", | |
838 | smp_processor_id(), idx); | |
839 | return -EINVAL; | |
840 | } | |
841 | ||
842 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
843 | asm volatile("msr pmselr_el0, %0" :: "r" (counter)); | |
844 | isb(); | |
845 | ||
846 | return idx; | |
847 | } | |
848 | ||
849 | static inline u32 armv8pmu_read_counter(int idx) | |
850 | { | |
851 | u32 value = 0; | |
852 | ||
853 | if (!armv8pmu_counter_valid(idx)) | |
854 | pr_err("CPU%u reading wrong counter %d\n", | |
855 | smp_processor_id(), idx); | |
856 | else if (idx == ARMV8_IDX_CYCLE_COUNTER) | |
857 | asm volatile("mrs %0, pmccntr_el0" : "=r" (value)); | |
858 | else if (armv8pmu_select_counter(idx) == idx) | |
859 | asm volatile("mrs %0, pmxevcntr_el0" : "=r" (value)); | |
860 | ||
861 | return value; | |
862 | } | |
863 | ||
864 | static inline void armv8pmu_write_counter(int idx, u32 value) | |
865 | { | |
866 | if (!armv8pmu_counter_valid(idx)) | |
867 | pr_err("CPU%u writing wrong counter %d\n", | |
868 | smp_processor_id(), idx); | |
869 | else if (idx == ARMV8_IDX_CYCLE_COUNTER) | |
870 | asm volatile("msr pmccntr_el0, %0" :: "r" (value)); | |
871 | else if (armv8pmu_select_counter(idx) == idx) | |
872 | asm volatile("msr pmxevcntr_el0, %0" :: "r" (value)); | |
873 | } | |
874 | ||
875 | static inline void armv8pmu_write_evtype(int idx, u32 val) | |
876 | { | |
877 | if (armv8pmu_select_counter(idx) == idx) { | |
878 | val &= ARMV8_EVTYPE_MASK; | |
879 | asm volatile("msr pmxevtyper_el0, %0" :: "r" (val)); | |
880 | } | |
881 | } | |
882 | ||
883 | static inline int armv8pmu_enable_counter(int idx) | |
884 | { | |
885 | u32 counter; | |
886 | ||
887 | if (!armv8pmu_counter_valid(idx)) { | |
888 | pr_err("CPU%u enabling wrong PMNC counter %d\n", | |
889 | smp_processor_id(), idx); | |
890 | return -EINVAL; | |
891 | } | |
892 | ||
893 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
894 | asm volatile("msr pmcntenset_el0, %0" :: "r" (BIT(counter))); | |
895 | return idx; | |
896 | } | |
897 | ||
898 | static inline int armv8pmu_disable_counter(int idx) | |
899 | { | |
900 | u32 counter; | |
901 | ||
902 | if (!armv8pmu_counter_valid(idx)) { | |
903 | pr_err("CPU%u disabling wrong PMNC counter %d\n", | |
904 | smp_processor_id(), idx); | |
905 | return -EINVAL; | |
906 | } | |
907 | ||
908 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
909 | asm volatile("msr pmcntenclr_el0, %0" :: "r" (BIT(counter))); | |
910 | return idx; | |
911 | } | |
912 | ||
913 | static inline int armv8pmu_enable_intens(int idx) | |
914 | { | |
915 | u32 counter; | |
916 | ||
917 | if (!armv8pmu_counter_valid(idx)) { | |
918 | pr_err("CPU%u enabling wrong PMNC counter IRQ enable %d\n", | |
919 | smp_processor_id(), idx); | |
920 | return -EINVAL; | |
921 | } | |
922 | ||
923 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
924 | asm volatile("msr pmintenset_el1, %0" :: "r" (BIT(counter))); | |
925 | return idx; | |
926 | } | |
927 | ||
928 | static inline int armv8pmu_disable_intens(int idx) | |
929 | { | |
930 | u32 counter; | |
931 | ||
932 | if (!armv8pmu_counter_valid(idx)) { | |
933 | pr_err("CPU%u disabling wrong PMNC counter IRQ enable %d\n", | |
934 | smp_processor_id(), idx); | |
935 | return -EINVAL; | |
936 | } | |
937 | ||
938 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
939 | asm volatile("msr pmintenclr_el1, %0" :: "r" (BIT(counter))); | |
940 | isb(); | |
941 | /* Clear the overflow flag in case an interrupt is pending. */ | |
942 | asm volatile("msr pmovsclr_el0, %0" :: "r" (BIT(counter))); | |
943 | isb(); | |
944 | return idx; | |
945 | } | |
946 | ||
947 | static inline u32 armv8pmu_getreset_flags(void) | |
948 | { | |
949 | u32 value; | |
950 | ||
951 | /* Read */ | |
952 | asm volatile("mrs %0, pmovsclr_el0" : "=r" (value)); | |
953 | ||
954 | /* Write to clear flags */ | |
955 | value &= ARMV8_OVSR_MASK; | |
956 | asm volatile("msr pmovsclr_el0, %0" :: "r" (value)); | |
957 | ||
958 | return value; | |
959 | } | |
960 | ||
961 | static void armv8pmu_enable_event(struct hw_perf_event *hwc, int idx) | |
962 | { | |
963 | unsigned long flags; | |
964 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
965 | ||
966 | /* | |
967 | * Enable counter and interrupt, and set the counter to count | |
968 | * the event that we're interested in. | |
969 | */ | |
970 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
971 | ||
972 | /* | |
973 | * Disable counter | |
974 | */ | |
975 | armv8pmu_disable_counter(idx); | |
976 | ||
977 | /* | |
978 | * Set event (if destined for PMNx counters). | |
979 | */ | |
980 | armv8pmu_write_evtype(idx, hwc->config_base); | |
981 | ||
982 | /* | |
983 | * Enable interrupt for this counter | |
984 | */ | |
985 | armv8pmu_enable_intens(idx); | |
986 | ||
987 | /* | |
988 | * Enable counter | |
989 | */ | |
990 | armv8pmu_enable_counter(idx); | |
991 | ||
992 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
993 | } | |
994 | ||
995 | static void armv8pmu_disable_event(struct hw_perf_event *hwc, int idx) | |
996 | { | |
997 | unsigned long flags; | |
998 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
999 | ||
1000 | /* | |
1001 | * Disable counter and interrupt | |
1002 | */ | |
1003 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
1004 | ||
1005 | /* | |
1006 | * Disable counter | |
1007 | */ | |
1008 | armv8pmu_disable_counter(idx); | |
1009 | ||
1010 | /* | |
1011 | * Disable interrupt for this counter | |
1012 | */ | |
1013 | armv8pmu_disable_intens(idx); | |
1014 | ||
1015 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
1016 | } | |
1017 | ||
1018 | static irqreturn_t armv8pmu_handle_irq(int irq_num, void *dev) | |
1019 | { | |
1020 | u32 pmovsr; | |
1021 | struct perf_sample_data data; | |
1022 | struct pmu_hw_events *cpuc; | |
1023 | struct pt_regs *regs; | |
1024 | int idx; | |
1025 | ||
1026 | /* | |
1027 | * Get and reset the IRQ flags | |
1028 | */ | |
1029 | pmovsr = armv8pmu_getreset_flags(); | |
1030 | ||
1031 | /* | |
1032 | * Did an overflow occur? | |
1033 | */ | |
1034 | if (!armv8pmu_has_overflowed(pmovsr)) | |
1035 | return IRQ_NONE; | |
1036 | ||
1037 | /* | |
1038 | * Handle the counter(s) overflow(s) | |
1039 | */ | |
1040 | regs = get_irq_regs(); | |
1041 | ||
1042 | cpuc = &__get_cpu_var(cpu_hw_events); | |
1043 | for (idx = 0; idx < cpu_pmu->num_events; ++idx) { | |
1044 | struct perf_event *event = cpuc->events[idx]; | |
1045 | struct hw_perf_event *hwc; | |
1046 | ||
1047 | /* Ignore if we don't have an event. */ | |
1048 | if (!event) | |
1049 | continue; | |
1050 | ||
1051 | /* | |
1052 | * We have a single interrupt for all counters. Check that | |
1053 | * each counter has overflowed before we process it. | |
1054 | */ | |
1055 | if (!armv8pmu_counter_has_overflowed(pmovsr, idx)) | |
1056 | continue; | |
1057 | ||
1058 | hwc = &event->hw; | |
1059 | armpmu_event_update(event, hwc, idx); | |
1060 | perf_sample_data_init(&data, 0, hwc->last_period); | |
1061 | if (!armpmu_event_set_period(event, hwc, idx)) | |
1062 | continue; | |
1063 | ||
1064 | if (perf_event_overflow(event, &data, regs)) | |
1065 | cpu_pmu->disable(hwc, idx); | |
1066 | } | |
1067 | ||
1068 | /* | |
1069 | * Handle the pending perf events. | |
1070 | * | |
1071 | * Note: this call *must* be run with interrupts disabled. For | |
1072 | * platforms that can have the PMU interrupts raised as an NMI, this | |
1073 | * will not work. | |
1074 | */ | |
1075 | irq_work_run(); | |
1076 | ||
1077 | return IRQ_HANDLED; | |
1078 | } | |
1079 | ||
1080 | static void armv8pmu_start(void) | |
1081 | { | |
1082 | unsigned long flags; | |
1083 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
1084 | ||
1085 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
1086 | /* Enable all counters */ | |
1087 | armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMCR_E); | |
1088 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
1089 | } | |
1090 | ||
1091 | static void armv8pmu_stop(void) | |
1092 | { | |
1093 | unsigned long flags; | |
1094 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
1095 | ||
1096 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
1097 | /* Disable all counters */ | |
1098 | armv8pmu_pmcr_write(armv8pmu_pmcr_read() & ~ARMV8_PMCR_E); | |
1099 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
1100 | } | |
1101 | ||
1102 | static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc, | |
1103 | struct hw_perf_event *event) | |
1104 | { | |
1105 | int idx; | |
1106 | unsigned long evtype = event->config_base & ARMV8_EVTYPE_EVENT; | |
1107 | ||
1108 | /* Always place a cycle counter into the cycle counter. */ | |
1109 | if (evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) { | |
1110 | if (test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask)) | |
1111 | return -EAGAIN; | |
1112 | ||
1113 | return ARMV8_IDX_CYCLE_COUNTER; | |
1114 | } | |
1115 | ||
1116 | /* | |
1117 | * For anything other than a cycle counter, try and use | |
1118 | * the events counters | |
1119 | */ | |
1120 | for (idx = ARMV8_IDX_COUNTER0; idx < cpu_pmu->num_events; ++idx) { | |
1121 | if (!test_and_set_bit(idx, cpuc->used_mask)) | |
1122 | return idx; | |
1123 | } | |
1124 | ||
1125 | /* The counters are all in use. */ | |
1126 | return -EAGAIN; | |
1127 | } | |
1128 | ||
1129 | /* | |
1130 | * Add an event filter to a given event. This will only work for PMUv2 PMUs. | |
1131 | */ | |
1132 | static int armv8pmu_set_event_filter(struct hw_perf_event *event, | |
1133 | struct perf_event_attr *attr) | |
1134 | { | |
1135 | unsigned long config_base = 0; | |
1136 | ||
1137 | if (attr->exclude_idle) | |
1138 | return -EPERM; | |
1139 | if (attr->exclude_user) | |
1140 | config_base |= ARMV8_EXCLUDE_EL0; | |
1141 | if (attr->exclude_kernel) | |
1142 | config_base |= ARMV8_EXCLUDE_EL1; | |
1143 | if (!attr->exclude_hv) | |
1144 | config_base |= ARMV8_INCLUDE_EL2; | |
1145 | ||
1146 | /* | |
1147 | * Install the filter into config_base as this is used to | |
1148 | * construct the event type. | |
1149 | */ | |
1150 | event->config_base = config_base; | |
1151 | ||
1152 | return 0; | |
1153 | } | |
1154 | ||
1155 | static void armv8pmu_reset(void *info) | |
1156 | { | |
1157 | u32 idx, nb_cnt = cpu_pmu->num_events; | |
1158 | ||
1159 | /* The counter and interrupt enable registers are unknown at reset. */ | |
1160 | for (idx = ARMV8_IDX_CYCLE_COUNTER; idx < nb_cnt; ++idx) | |
1161 | armv8pmu_disable_event(NULL, idx); | |
1162 | ||
1163 | /* Initialize & Reset PMNC: C and P bits. */ | |
1164 | armv8pmu_pmcr_write(ARMV8_PMCR_P | ARMV8_PMCR_C); | |
1165 | ||
1166 | /* Disable access from userspace. */ | |
1167 | asm volatile("msr pmuserenr_el0, %0" :: "r" (0)); | |
1168 | } | |
1169 | ||
1170 | static int armv8_pmuv3_map_event(struct perf_event *event) | |
1171 | { | |
1172 | return map_cpu_event(event, &armv8_pmuv3_perf_map, | |
1173 | &armv8_pmuv3_perf_cache_map, 0xFF); | |
1174 | } | |
1175 | ||
1176 | static struct arm_pmu armv8pmu = { | |
1177 | .handle_irq = armv8pmu_handle_irq, | |
1178 | .enable = armv8pmu_enable_event, | |
1179 | .disable = armv8pmu_disable_event, | |
1180 | .read_counter = armv8pmu_read_counter, | |
1181 | .write_counter = armv8pmu_write_counter, | |
1182 | .get_event_idx = armv8pmu_get_event_idx, | |
1183 | .start = armv8pmu_start, | |
1184 | .stop = armv8pmu_stop, | |
1185 | .reset = armv8pmu_reset, | |
1186 | .max_period = (1LLU << 32) - 1, | |
1187 | }; | |
1188 | ||
1189 | static u32 __init armv8pmu_read_num_pmnc_events(void) | |
1190 | { | |
1191 | u32 nb_cnt; | |
1192 | ||
1193 | /* Read the nb of CNTx counters supported from PMNC */ | |
1194 | nb_cnt = (armv8pmu_pmcr_read() >> ARMV8_PMCR_N_SHIFT) & ARMV8_PMCR_N_MASK; | |
1195 | ||
1196 | /* Add the CPU cycles counter and return */ | |
1197 | return nb_cnt + 1; | |
1198 | } | |
1199 | ||
1200 | static struct arm_pmu *__init armv8_pmuv3_pmu_init(void) | |
1201 | { | |
1202 | armv8pmu.name = "arm/armv8-pmuv3"; | |
1203 | armv8pmu.map_event = armv8_pmuv3_map_event; | |
1204 | armv8pmu.num_events = armv8pmu_read_num_pmnc_events(); | |
1205 | armv8pmu.set_event_filter = armv8pmu_set_event_filter; | |
1206 | return &armv8pmu; | |
1207 | } | |
1208 | ||
1209 | /* | |
1210 | * Ensure the PMU has sane values out of reset. | |
1211 | * This requires SMP to be available, so exists as a separate initcall. | |
1212 | */ | |
1213 | static int __init | |
1214 | cpu_pmu_reset(void) | |
1215 | { | |
1216 | if (cpu_pmu && cpu_pmu->reset) | |
1217 | return on_each_cpu(cpu_pmu->reset, NULL, 1); | |
1218 | return 0; | |
1219 | } | |
1220 | arch_initcall(cpu_pmu_reset); | |
1221 | ||
1222 | /* | |
1223 | * PMU platform driver and devicetree bindings. | |
1224 | */ | |
1225 | static struct of_device_id armpmu_of_device_ids[] = { | |
1226 | {.compatible = "arm,armv8-pmuv3"}, | |
1227 | {}, | |
1228 | }; | |
1229 | ||
1230 | static int __devinit armpmu_device_probe(struct platform_device *pdev) | |
1231 | { | |
1232 | if (!cpu_pmu) | |
1233 | return -ENODEV; | |
1234 | ||
1235 | cpu_pmu->plat_device = pdev; | |
1236 | return 0; | |
1237 | } | |
1238 | ||
1239 | static struct platform_driver armpmu_driver = { | |
1240 | .driver = { | |
1241 | .name = "arm-pmu", | |
1242 | .of_match_table = armpmu_of_device_ids, | |
1243 | }, | |
1244 | .probe = armpmu_device_probe, | |
1245 | }; | |
1246 | ||
1247 | static int __init register_pmu_driver(void) | |
1248 | { | |
1249 | return platform_driver_register(&armpmu_driver); | |
1250 | } | |
1251 | device_initcall(register_pmu_driver); | |
1252 | ||
1253 | static struct pmu_hw_events *armpmu_get_cpu_events(void) | |
1254 | { | |
1255 | return &__get_cpu_var(cpu_hw_events); | |
1256 | } | |
1257 | ||
1258 | static void __init cpu_pmu_init(struct arm_pmu *armpmu) | |
1259 | { | |
1260 | int cpu; | |
1261 | for_each_possible_cpu(cpu) { | |
1262 | struct pmu_hw_events *events = &per_cpu(cpu_hw_events, cpu); | |
1263 | events->events = per_cpu(hw_events, cpu); | |
1264 | events->used_mask = per_cpu(used_mask, cpu); | |
1265 | raw_spin_lock_init(&events->pmu_lock); | |
1266 | } | |
1267 | armpmu->get_hw_events = armpmu_get_cpu_events; | |
1268 | } | |
1269 | ||
1270 | static int __init init_hw_perf_events(void) | |
1271 | { | |
1272 | u64 dfr = read_cpuid(ID_AA64DFR0_EL1); | |
1273 | ||
1274 | switch ((dfr >> 8) & 0xf) { | |
1275 | case 0x1: /* PMUv3 */ | |
1276 | cpu_pmu = armv8_pmuv3_pmu_init(); | |
1277 | break; | |
1278 | } | |
1279 | ||
1280 | if (cpu_pmu) { | |
1281 | pr_info("enabled with %s PMU driver, %d counters available\n", | |
1282 | cpu_pmu->name, cpu_pmu->num_events); | |
1283 | cpu_pmu_init(cpu_pmu); | |
1284 | armpmu_register(cpu_pmu, "cpu", PERF_TYPE_RAW); | |
1285 | } else { | |
1286 | pr_info("no hardware support available\n"); | |
1287 | } | |
1288 | ||
1289 | return 0; | |
1290 | } | |
1291 | early_initcall(init_hw_perf_events); | |
1292 | ||
1293 | /* | |
1294 | * Callchain handling code. | |
1295 | */ | |
1296 | struct frame_tail { | |
1297 | struct frame_tail __user *fp; | |
1298 | unsigned long lr; | |
1299 | } __attribute__((packed)); | |
1300 | ||
1301 | /* | |
1302 | * Get the return address for a single stackframe and return a pointer to the | |
1303 | * next frame tail. | |
1304 | */ | |
1305 | static struct frame_tail __user * | |
1306 | user_backtrace(struct frame_tail __user *tail, | |
1307 | struct perf_callchain_entry *entry) | |
1308 | { | |
1309 | struct frame_tail buftail; | |
1310 | unsigned long err; | |
1311 | ||
1312 | /* Also check accessibility of one struct frame_tail beyond */ | |
1313 | if (!access_ok(VERIFY_READ, tail, sizeof(buftail))) | |
1314 | return NULL; | |
1315 | ||
1316 | pagefault_disable(); | |
1317 | err = __copy_from_user_inatomic(&buftail, tail, sizeof(buftail)); | |
1318 | pagefault_enable(); | |
1319 | ||
1320 | if (err) | |
1321 | return NULL; | |
1322 | ||
1323 | perf_callchain_store(entry, buftail.lr); | |
1324 | ||
1325 | /* | |
1326 | * Frame pointers should strictly progress back up the stack | |
1327 | * (towards higher addresses). | |
1328 | */ | |
1329 | if (tail >= buftail.fp) | |
1330 | return NULL; | |
1331 | ||
1332 | return buftail.fp; | |
1333 | } | |
1334 | ||
1335 | void perf_callchain_user(struct perf_callchain_entry *entry, | |
1336 | struct pt_regs *regs) | |
1337 | { | |
1338 | struct frame_tail __user *tail; | |
1339 | ||
1340 | tail = (struct frame_tail __user *)regs->regs[29]; | |
1341 | ||
1342 | while (entry->nr < PERF_MAX_STACK_DEPTH && | |
1343 | tail && !((unsigned long)tail & 0xf)) | |
1344 | tail = user_backtrace(tail, entry); | |
1345 | } | |
1346 | ||
1347 | /* | |
1348 | * Gets called by walk_stackframe() for every stackframe. This will be called | |
1349 | * whist unwinding the stackframe and is like a subroutine return so we use | |
1350 | * the PC. | |
1351 | */ | |
1352 | static int callchain_trace(struct stackframe *frame, void *data) | |
1353 | { | |
1354 | struct perf_callchain_entry *entry = data; | |
1355 | perf_callchain_store(entry, frame->pc); | |
1356 | return 0; | |
1357 | } | |
1358 | ||
1359 | void perf_callchain_kernel(struct perf_callchain_entry *entry, | |
1360 | struct pt_regs *regs) | |
1361 | { | |
1362 | struct stackframe frame; | |
1363 | ||
1364 | frame.fp = regs->regs[29]; | |
1365 | frame.sp = regs->sp; | |
1366 | frame.pc = regs->pc; | |
1367 | walk_stackframe(&frame, callchain_trace, entry); | |
1368 | } |