]>
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> | |
66aa8d6a | 25 | #include <linux/irq.h> |
03089688 WD |
26 | #include <linux/kernel.h> |
27 | #include <linux/export.h> | |
28 | #include <linux/perf_event.h> | |
29 | #include <linux/platform_device.h> | |
30 | #include <linux/spinlock.h> | |
31 | #include <linux/uaccess.h> | |
32 | ||
33 | #include <asm/cputype.h> | |
34 | #include <asm/irq.h> | |
35 | #include <asm/irq_regs.h> | |
36 | #include <asm/pmu.h> | |
37 | #include <asm/stacktrace.h> | |
38 | ||
39 | /* | |
40 | * ARMv8 supports a maximum of 32 events. | |
41 | * The cycle counter is included in this total. | |
42 | */ | |
43 | #define ARMPMU_MAX_HWEVENTS 32 | |
44 | ||
45 | static DEFINE_PER_CPU(struct perf_event * [ARMPMU_MAX_HWEVENTS], hw_events); | |
46 | static DEFINE_PER_CPU(unsigned long [BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)], used_mask); | |
47 | static DEFINE_PER_CPU(struct pmu_hw_events, cpu_hw_events); | |
48 | ||
49 | #define to_arm_pmu(p) (container_of(p, struct arm_pmu, pmu)) | |
50 | ||
51 | /* Set at runtime when we know what CPU type we are. */ | |
52 | static struct arm_pmu *cpu_pmu; | |
53 | ||
54 | int | |
55 | armpmu_get_max_events(void) | |
56 | { | |
57 | int max_events = 0; | |
58 | ||
59 | if (cpu_pmu != NULL) | |
60 | max_events = cpu_pmu->num_events; | |
61 | ||
62 | return max_events; | |
63 | } | |
64 | EXPORT_SYMBOL_GPL(armpmu_get_max_events); | |
65 | ||
66 | int perf_num_counters(void) | |
67 | { | |
68 | return armpmu_get_max_events(); | |
69 | } | |
70 | EXPORT_SYMBOL_GPL(perf_num_counters); | |
71 | ||
72 | #define HW_OP_UNSUPPORTED 0xFFFF | |
73 | ||
74 | #define C(_x) \ | |
75 | PERF_COUNT_HW_CACHE_##_x | |
76 | ||
77 | #define CACHE_OP_UNSUPPORTED 0xFFFF | |
78 | ||
79 | static int | |
80 | armpmu_map_cache_event(const unsigned (*cache_map) | |
81 | [PERF_COUNT_HW_CACHE_MAX] | |
82 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
83 | [PERF_COUNT_HW_CACHE_RESULT_MAX], | |
84 | u64 config) | |
85 | { | |
86 | unsigned int cache_type, cache_op, cache_result, ret; | |
87 | ||
88 | cache_type = (config >> 0) & 0xff; | |
89 | if (cache_type >= PERF_COUNT_HW_CACHE_MAX) | |
90 | return -EINVAL; | |
91 | ||
92 | cache_op = (config >> 8) & 0xff; | |
93 | if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX) | |
94 | return -EINVAL; | |
95 | ||
96 | cache_result = (config >> 16) & 0xff; | |
97 | if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX) | |
98 | return -EINVAL; | |
99 | ||
100 | ret = (int)(*cache_map)[cache_type][cache_op][cache_result]; | |
101 | ||
102 | if (ret == CACHE_OP_UNSUPPORTED) | |
103 | return -ENOENT; | |
104 | ||
105 | return ret; | |
106 | } | |
107 | ||
108 | static int | |
109 | armpmu_map_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config) | |
110 | { | |
868f6fea WD |
111 | int mapping; |
112 | ||
113 | if (config >= PERF_COUNT_HW_MAX) | |
114 | return -EINVAL; | |
115 | ||
116 | mapping = (*event_map)[config]; | |
03089688 WD |
117 | return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping; |
118 | } | |
119 | ||
120 | static int | |
121 | armpmu_map_raw_event(u32 raw_event_mask, u64 config) | |
122 | { | |
123 | return (int)(config & raw_event_mask); | |
124 | } | |
125 | ||
126 | static int map_cpu_event(struct perf_event *event, | |
127 | const unsigned (*event_map)[PERF_COUNT_HW_MAX], | |
128 | const unsigned (*cache_map) | |
129 | [PERF_COUNT_HW_CACHE_MAX] | |
130 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
131 | [PERF_COUNT_HW_CACHE_RESULT_MAX], | |
132 | u32 raw_event_mask) | |
133 | { | |
134 | u64 config = event->attr.config; | |
135 | ||
136 | switch (event->attr.type) { | |
137 | case PERF_TYPE_HARDWARE: | |
138 | return armpmu_map_event(event_map, config); | |
139 | case PERF_TYPE_HW_CACHE: | |
140 | return armpmu_map_cache_event(cache_map, config); | |
141 | case PERF_TYPE_RAW: | |
142 | return armpmu_map_raw_event(raw_event_mask, config); | |
143 | } | |
144 | ||
145 | return -ENOENT; | |
146 | } | |
147 | ||
148 | int | |
149 | armpmu_event_set_period(struct perf_event *event, | |
150 | struct hw_perf_event *hwc, | |
151 | int idx) | |
152 | { | |
153 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
154 | s64 left = local64_read(&hwc->period_left); | |
155 | s64 period = hwc->sample_period; | |
156 | int ret = 0; | |
157 | ||
158 | if (unlikely(left <= -period)) { | |
159 | left = period; | |
160 | local64_set(&hwc->period_left, left); | |
161 | hwc->last_period = period; | |
162 | ret = 1; | |
163 | } | |
164 | ||
165 | if (unlikely(left <= 0)) { | |
166 | left += period; | |
167 | local64_set(&hwc->period_left, left); | |
168 | hwc->last_period = period; | |
169 | ret = 1; | |
170 | } | |
171 | ||
172 | if (left > (s64)armpmu->max_period) | |
173 | left = armpmu->max_period; | |
174 | ||
175 | local64_set(&hwc->prev_count, (u64)-left); | |
176 | ||
177 | armpmu->write_counter(idx, (u64)(-left) & 0xffffffff); | |
178 | ||
179 | perf_event_update_userpage(event); | |
180 | ||
181 | return ret; | |
182 | } | |
183 | ||
184 | u64 | |
185 | armpmu_event_update(struct perf_event *event, | |
186 | struct hw_perf_event *hwc, | |
187 | int idx) | |
188 | { | |
189 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
190 | u64 delta, prev_raw_count, new_raw_count; | |
191 | ||
192 | again: | |
193 | prev_raw_count = local64_read(&hwc->prev_count); | |
194 | new_raw_count = armpmu->read_counter(idx); | |
195 | ||
196 | if (local64_cmpxchg(&hwc->prev_count, prev_raw_count, | |
197 | new_raw_count) != prev_raw_count) | |
198 | goto again; | |
199 | ||
200 | delta = (new_raw_count - prev_raw_count) & armpmu->max_period; | |
201 | ||
202 | local64_add(delta, &event->count); | |
203 | local64_sub(delta, &hwc->period_left); | |
204 | ||
205 | return new_raw_count; | |
206 | } | |
207 | ||
208 | static void | |
209 | armpmu_read(struct perf_event *event) | |
210 | { | |
211 | struct hw_perf_event *hwc = &event->hw; | |
212 | ||
213 | /* Don't read disabled counters! */ | |
214 | if (hwc->idx < 0) | |
215 | return; | |
216 | ||
217 | armpmu_event_update(event, hwc, hwc->idx); | |
218 | } | |
219 | ||
220 | static void | |
221 | armpmu_stop(struct perf_event *event, int flags) | |
222 | { | |
223 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
224 | struct hw_perf_event *hwc = &event->hw; | |
225 | ||
226 | /* | |
227 | * ARM pmu always has to update the counter, so ignore | |
228 | * PERF_EF_UPDATE, see comments in armpmu_start(). | |
229 | */ | |
230 | if (!(hwc->state & PERF_HES_STOPPED)) { | |
231 | armpmu->disable(hwc, hwc->idx); | |
232 | barrier(); /* why? */ | |
233 | armpmu_event_update(event, hwc, hwc->idx); | |
234 | hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE; | |
235 | } | |
236 | } | |
237 | ||
238 | static void | |
239 | armpmu_start(struct perf_event *event, int flags) | |
240 | { | |
241 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
242 | struct hw_perf_event *hwc = &event->hw; | |
243 | ||
244 | /* | |
245 | * ARM pmu always has to reprogram the period, so ignore | |
246 | * PERF_EF_RELOAD, see the comment below. | |
247 | */ | |
248 | if (flags & PERF_EF_RELOAD) | |
249 | WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE)); | |
250 | ||
251 | hwc->state = 0; | |
252 | /* | |
253 | * Set the period again. Some counters can't be stopped, so when we | |
254 | * were stopped we simply disabled the IRQ source and the counter | |
255 | * may have been left counting. If we don't do this step then we may | |
256 | * get an interrupt too soon or *way* too late if the overflow has | |
257 | * happened since disabling. | |
258 | */ | |
259 | armpmu_event_set_period(event, hwc, hwc->idx); | |
260 | armpmu->enable(hwc, hwc->idx); | |
261 | } | |
262 | ||
263 | static void | |
264 | armpmu_del(struct perf_event *event, int flags) | |
265 | { | |
266 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
267 | struct pmu_hw_events *hw_events = armpmu->get_hw_events(); | |
268 | struct hw_perf_event *hwc = &event->hw; | |
269 | int idx = hwc->idx; | |
270 | ||
271 | WARN_ON(idx < 0); | |
272 | ||
273 | armpmu_stop(event, PERF_EF_UPDATE); | |
274 | hw_events->events[idx] = NULL; | |
275 | clear_bit(idx, hw_events->used_mask); | |
276 | ||
277 | perf_event_update_userpage(event); | |
278 | } | |
279 | ||
280 | static int | |
281 | armpmu_add(struct perf_event *event, int flags) | |
282 | { | |
283 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
284 | struct pmu_hw_events *hw_events = armpmu->get_hw_events(); | |
285 | struct hw_perf_event *hwc = &event->hw; | |
286 | int idx; | |
287 | int err = 0; | |
288 | ||
289 | perf_pmu_disable(event->pmu); | |
290 | ||
291 | /* If we don't have a space for the counter then finish early. */ | |
292 | idx = armpmu->get_event_idx(hw_events, hwc); | |
293 | if (idx < 0) { | |
294 | err = idx; | |
295 | goto out; | |
296 | } | |
297 | ||
298 | /* | |
299 | * If there is an event in the counter we are going to use then make | |
300 | * sure it is disabled. | |
301 | */ | |
302 | event->hw.idx = idx; | |
303 | armpmu->disable(hwc, idx); | |
304 | hw_events->events[idx] = event; | |
305 | ||
306 | hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE; | |
307 | if (flags & PERF_EF_START) | |
308 | armpmu_start(event, PERF_EF_RELOAD); | |
309 | ||
310 | /* Propagate our changes to the userspace mapping. */ | |
311 | perf_event_update_userpage(event); | |
312 | ||
313 | out: | |
314 | perf_pmu_enable(event->pmu); | |
315 | return err; | |
316 | } | |
317 | ||
318 | static int | |
319 | validate_event(struct pmu_hw_events *hw_events, | |
320 | struct perf_event *event) | |
321 | { | |
322 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
323 | struct hw_perf_event fake_event = event->hw; | |
324 | struct pmu *leader_pmu = event->group_leader->pmu; | |
325 | ||
ee7538a0 WD |
326 | if (is_software_event(event)) |
327 | return 1; | |
328 | ||
8455e6ec WD |
329 | if (event->pmu != leader_pmu || event->state < PERF_EVENT_STATE_OFF) |
330 | return 1; | |
331 | ||
332 | if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec) | |
03089688 WD |
333 | return 1; |
334 | ||
335 | return armpmu->get_event_idx(hw_events, &fake_event) >= 0; | |
336 | } | |
337 | ||
338 | static int | |
339 | validate_group(struct perf_event *event) | |
340 | { | |
341 | struct perf_event *sibling, *leader = event->group_leader; | |
342 | struct pmu_hw_events fake_pmu; | |
343 | DECLARE_BITMAP(fake_used_mask, ARMPMU_MAX_HWEVENTS); | |
344 | ||
345 | /* | |
346 | * Initialise the fake PMU. We only need to populate the | |
347 | * used_mask for the purposes of validation. | |
348 | */ | |
349 | memset(fake_used_mask, 0, sizeof(fake_used_mask)); | |
350 | fake_pmu.used_mask = fake_used_mask; | |
351 | ||
352 | if (!validate_event(&fake_pmu, leader)) | |
353 | return -EINVAL; | |
354 | ||
355 | list_for_each_entry(sibling, &leader->sibling_list, group_entry) { | |
356 | if (!validate_event(&fake_pmu, sibling)) | |
357 | return -EINVAL; | |
358 | } | |
359 | ||
360 | if (!validate_event(&fake_pmu, event)) | |
361 | return -EINVAL; | |
362 | ||
363 | return 0; | |
364 | } | |
365 | ||
66aa8d6a VK |
366 | static void |
367 | armpmu_disable_percpu_irq(void *data) | |
368 | { | |
369 | unsigned int irq = *(unsigned int *)data; | |
370 | disable_percpu_irq(irq); | |
371 | } | |
372 | ||
03089688 WD |
373 | static void |
374 | armpmu_release_hardware(struct arm_pmu *armpmu) | |
375 | { | |
66aa8d6a VK |
376 | int irq; |
377 | unsigned int i, irqs; | |
03089688 WD |
378 | struct platform_device *pmu_device = armpmu->plat_device; |
379 | ||
380 | irqs = min(pmu_device->num_resources, num_possible_cpus()); | |
66aa8d6a VK |
381 | if (!irqs) |
382 | return; | |
03089688 | 383 | |
66aa8d6a VK |
384 | irq = platform_get_irq(pmu_device, 0); |
385 | if (irq <= 0) | |
386 | return; | |
387 | ||
388 | if (irq_is_percpu(irq)) { | |
389 | on_each_cpu(armpmu_disable_percpu_irq, &irq, 1); | |
390 | free_percpu_irq(irq, &cpu_hw_events); | |
391 | } else { | |
392 | for (i = 0; i < irqs; ++i) { | |
393 | if (!cpumask_test_and_clear_cpu(i, &armpmu->active_irqs)) | |
394 | continue; | |
395 | irq = platform_get_irq(pmu_device, i); | |
396 | if (irq > 0) | |
397 | free_irq(irq, armpmu); | |
398 | } | |
03089688 WD |
399 | } |
400 | } | |
401 | ||
66aa8d6a VK |
402 | static void |
403 | armpmu_enable_percpu_irq(void *data) | |
404 | { | |
405 | unsigned int irq = *(unsigned int *)data; | |
406 | enable_percpu_irq(irq, IRQ_TYPE_NONE); | |
407 | } | |
408 | ||
03089688 WD |
409 | static int |
410 | armpmu_reserve_hardware(struct arm_pmu *armpmu) | |
411 | { | |
66aa8d6a VK |
412 | int err, irq; |
413 | unsigned int i, irqs; | |
03089688 WD |
414 | struct platform_device *pmu_device = armpmu->plat_device; |
415 | ||
416 | if (!pmu_device) { | |
417 | pr_err("no PMU device registered\n"); | |
418 | return -ENODEV; | |
419 | } | |
420 | ||
421 | irqs = min(pmu_device->num_resources, num_possible_cpus()); | |
66aa8d6a | 422 | if (!irqs) { |
03089688 WD |
423 | pr_err("no irqs for PMUs defined\n"); |
424 | return -ENODEV; | |
425 | } | |
426 | ||
66aa8d6a VK |
427 | irq = platform_get_irq(pmu_device, 0); |
428 | if (irq <= 0) { | |
429 | pr_err("failed to get valid irq for PMU device\n"); | |
430 | return -ENODEV; | |
431 | } | |
432 | ||
433 | if (irq_is_percpu(irq)) { | |
434 | err = request_percpu_irq(irq, armpmu->handle_irq, | |
435 | "arm-pmu", &cpu_hw_events); | |
03089688 | 436 | |
03089688 | 437 | if (err) { |
66aa8d6a VK |
438 | pr_err("unable to request percpu IRQ%d for ARM PMU counters\n", |
439 | irq); | |
03089688 WD |
440 | armpmu_release_hardware(armpmu); |
441 | return err; | |
442 | } | |
443 | ||
66aa8d6a VK |
444 | on_each_cpu(armpmu_enable_percpu_irq, &irq, 1); |
445 | } else { | |
446 | for (i = 0; i < irqs; ++i) { | |
447 | err = 0; | |
448 | irq = platform_get_irq(pmu_device, i); | |
449 | if (irq <= 0) | |
450 | continue; | |
451 | ||
452 | /* | |
453 | * If we have a single PMU interrupt that we can't shift, | |
454 | * assume that we're running on a uniprocessor machine and | |
455 | * continue. Otherwise, continue without this interrupt. | |
456 | */ | |
457 | if (irq_set_affinity(irq, cpumask_of(i)) && irqs > 1) { | |
458 | pr_warning("unable to set irq affinity (irq=%d, cpu=%u)\n", | |
459 | irq, i); | |
460 | continue; | |
461 | } | |
462 | ||
463 | err = request_irq(irq, armpmu->handle_irq, | |
464 | IRQF_NOBALANCING, | |
465 | "arm-pmu", armpmu); | |
466 | if (err) { | |
467 | pr_err("unable to request IRQ%d for ARM PMU counters\n", | |
468 | irq); | |
469 | armpmu_release_hardware(armpmu); | |
470 | return err; | |
471 | } | |
472 | ||
473 | cpumask_set_cpu(i, &armpmu->active_irqs); | |
474 | } | |
03089688 WD |
475 | } |
476 | ||
477 | return 0; | |
478 | } | |
479 | ||
480 | static void | |
481 | hw_perf_event_destroy(struct perf_event *event) | |
482 | { | |
483 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
484 | atomic_t *active_events = &armpmu->active_events; | |
485 | struct mutex *pmu_reserve_mutex = &armpmu->reserve_mutex; | |
486 | ||
487 | if (atomic_dec_and_mutex_lock(active_events, pmu_reserve_mutex)) { | |
488 | armpmu_release_hardware(armpmu); | |
489 | mutex_unlock(pmu_reserve_mutex); | |
490 | } | |
491 | } | |
492 | ||
493 | static int | |
494 | event_requires_mode_exclusion(struct perf_event_attr *attr) | |
495 | { | |
496 | return attr->exclude_idle || attr->exclude_user || | |
497 | attr->exclude_kernel || attr->exclude_hv; | |
498 | } | |
499 | ||
500 | static int | |
501 | __hw_perf_event_init(struct perf_event *event) | |
502 | { | |
503 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
504 | struct hw_perf_event *hwc = &event->hw; | |
505 | int mapping, err; | |
506 | ||
507 | mapping = armpmu->map_event(event); | |
508 | ||
509 | if (mapping < 0) { | |
510 | pr_debug("event %x:%llx not supported\n", event->attr.type, | |
511 | event->attr.config); | |
512 | return mapping; | |
513 | } | |
514 | ||
515 | /* | |
516 | * We don't assign an index until we actually place the event onto | |
517 | * hardware. Use -1 to signify that we haven't decided where to put it | |
518 | * yet. For SMP systems, each core has it's own PMU so we can't do any | |
519 | * clever allocation or constraints checking at this point. | |
520 | */ | |
521 | hwc->idx = -1; | |
522 | hwc->config_base = 0; | |
523 | hwc->config = 0; | |
524 | hwc->event_base = 0; | |
525 | ||
526 | /* | |
527 | * Check whether we need to exclude the counter from certain modes. | |
528 | */ | |
529 | if ((!armpmu->set_event_filter || | |
530 | armpmu->set_event_filter(hwc, &event->attr)) && | |
531 | event_requires_mode_exclusion(&event->attr)) { | |
532 | pr_debug("ARM performance counters do not support mode exclusion\n"); | |
533 | return -EPERM; | |
534 | } | |
535 | ||
536 | /* | |
537 | * Store the event encoding into the config_base field. | |
538 | */ | |
539 | hwc->config_base |= (unsigned long)mapping; | |
540 | ||
541 | if (!hwc->sample_period) { | |
542 | /* | |
543 | * For non-sampling runs, limit the sample_period to half | |
544 | * of the counter width. That way, the new counter value | |
545 | * is far less likely to overtake the previous one unless | |
546 | * you have some serious IRQ latency issues. | |
547 | */ | |
548 | hwc->sample_period = armpmu->max_period >> 1; | |
549 | hwc->last_period = hwc->sample_period; | |
550 | local64_set(&hwc->period_left, hwc->sample_period); | |
551 | } | |
552 | ||
553 | err = 0; | |
554 | if (event->group_leader != event) { | |
555 | err = validate_group(event); | |
556 | if (err) | |
557 | return -EINVAL; | |
558 | } | |
559 | ||
560 | return err; | |
561 | } | |
562 | ||
563 | static int armpmu_event_init(struct perf_event *event) | |
564 | { | |
565 | struct arm_pmu *armpmu = to_arm_pmu(event->pmu); | |
566 | int err = 0; | |
567 | atomic_t *active_events = &armpmu->active_events; | |
568 | ||
569 | if (armpmu->map_event(event) == -ENOENT) | |
570 | return -ENOENT; | |
571 | ||
572 | event->destroy = hw_perf_event_destroy; | |
573 | ||
574 | if (!atomic_inc_not_zero(active_events)) { | |
575 | mutex_lock(&armpmu->reserve_mutex); | |
576 | if (atomic_read(active_events) == 0) | |
577 | err = armpmu_reserve_hardware(armpmu); | |
578 | ||
579 | if (!err) | |
580 | atomic_inc(active_events); | |
581 | mutex_unlock(&armpmu->reserve_mutex); | |
582 | } | |
583 | ||
584 | if (err) | |
585 | return err; | |
586 | ||
587 | err = __hw_perf_event_init(event); | |
588 | if (err) | |
589 | hw_perf_event_destroy(event); | |
590 | ||
591 | return err; | |
592 | } | |
593 | ||
594 | static void armpmu_enable(struct pmu *pmu) | |
595 | { | |
596 | struct arm_pmu *armpmu = to_arm_pmu(pmu); | |
597 | struct pmu_hw_events *hw_events = armpmu->get_hw_events(); | |
598 | int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events); | |
599 | ||
600 | if (enabled) | |
601 | armpmu->start(); | |
602 | } | |
603 | ||
604 | static void armpmu_disable(struct pmu *pmu) | |
605 | { | |
606 | struct arm_pmu *armpmu = to_arm_pmu(pmu); | |
607 | armpmu->stop(); | |
608 | } | |
609 | ||
610 | static void __init armpmu_init(struct arm_pmu *armpmu) | |
611 | { | |
612 | atomic_set(&armpmu->active_events, 0); | |
613 | mutex_init(&armpmu->reserve_mutex); | |
614 | ||
615 | armpmu->pmu = (struct pmu) { | |
616 | .pmu_enable = armpmu_enable, | |
617 | .pmu_disable = armpmu_disable, | |
618 | .event_init = armpmu_event_init, | |
619 | .add = armpmu_add, | |
620 | .del = armpmu_del, | |
621 | .start = armpmu_start, | |
622 | .stop = armpmu_stop, | |
623 | .read = armpmu_read, | |
624 | }; | |
625 | } | |
626 | ||
627 | int __init armpmu_register(struct arm_pmu *armpmu, char *name, int type) | |
628 | { | |
629 | armpmu_init(armpmu); | |
630 | return perf_pmu_register(&armpmu->pmu, name, type); | |
631 | } | |
632 | ||
633 | /* | |
634 | * ARMv8 PMUv3 Performance Events handling code. | |
635 | * Common event types. | |
636 | */ | |
637 | enum armv8_pmuv3_perf_types { | |
638 | /* Required events. */ | |
639 | ARMV8_PMUV3_PERFCTR_PMNC_SW_INCR = 0x00, | |
640 | ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL = 0x03, | |
641 | ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS = 0x04, | |
642 | ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED = 0x10, | |
643 | ARMV8_PMUV3_PERFCTR_CLOCK_CYCLES = 0x11, | |
644 | ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED = 0x12, | |
645 | ||
646 | /* At least one of the following is required. */ | |
647 | ARMV8_PMUV3_PERFCTR_INSTR_EXECUTED = 0x08, | |
648 | ARMV8_PMUV3_PERFCTR_OP_SPEC = 0x1B, | |
649 | ||
650 | /* Common architectural events. */ | |
651 | ARMV8_PMUV3_PERFCTR_MEM_READ = 0x06, | |
652 | ARMV8_PMUV3_PERFCTR_MEM_WRITE = 0x07, | |
653 | ARMV8_PMUV3_PERFCTR_EXC_TAKEN = 0x09, | |
654 | ARMV8_PMUV3_PERFCTR_EXC_EXECUTED = 0x0A, | |
655 | ARMV8_PMUV3_PERFCTR_CID_WRITE = 0x0B, | |
656 | ARMV8_PMUV3_PERFCTR_PC_WRITE = 0x0C, | |
657 | ARMV8_PMUV3_PERFCTR_PC_IMM_BRANCH = 0x0D, | |
658 | ARMV8_PMUV3_PERFCTR_PC_PROC_RETURN = 0x0E, | |
659 | ARMV8_PMUV3_PERFCTR_MEM_UNALIGNED_ACCESS = 0x0F, | |
660 | ARMV8_PMUV3_PERFCTR_TTBR_WRITE = 0x1C, | |
661 | ||
662 | /* Common microarchitectural events. */ | |
663 | ARMV8_PMUV3_PERFCTR_L1_ICACHE_REFILL = 0x01, | |
664 | ARMV8_PMUV3_PERFCTR_ITLB_REFILL = 0x02, | |
665 | ARMV8_PMUV3_PERFCTR_DTLB_REFILL = 0x05, | |
666 | ARMV8_PMUV3_PERFCTR_MEM_ACCESS = 0x13, | |
667 | ARMV8_PMUV3_PERFCTR_L1_ICACHE_ACCESS = 0x14, | |
668 | ARMV8_PMUV3_PERFCTR_L1_DCACHE_WB = 0x15, | |
669 | ARMV8_PMUV3_PERFCTR_L2_CACHE_ACCESS = 0x16, | |
670 | ARMV8_PMUV3_PERFCTR_L2_CACHE_REFILL = 0x17, | |
671 | ARMV8_PMUV3_PERFCTR_L2_CACHE_WB = 0x18, | |
672 | ARMV8_PMUV3_PERFCTR_BUS_ACCESS = 0x19, | |
673 | ARMV8_PMUV3_PERFCTR_MEM_ERROR = 0x1A, | |
674 | ARMV8_PMUV3_PERFCTR_BUS_CYCLES = 0x1D, | |
03089688 WD |
675 | }; |
676 | ||
677 | /* PMUv3 HW events mapping. */ | |
678 | static const unsigned armv8_pmuv3_perf_map[PERF_COUNT_HW_MAX] = { | |
f46f979f | 679 | [PERF_COUNT_HW_CPU_CYCLES] = ARMV8_PMUV3_PERFCTR_CLOCK_CYCLES, |
03089688 WD |
680 | [PERF_COUNT_HW_INSTRUCTIONS] = ARMV8_PMUV3_PERFCTR_INSTR_EXECUTED, |
681 | [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, | |
682 | [PERF_COUNT_HW_CACHE_MISSES] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, | |
683 | [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = HW_OP_UNSUPPORTED, | |
684 | [PERF_COUNT_HW_BRANCH_MISSES] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, | |
685 | [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED, | |
686 | [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = HW_OP_UNSUPPORTED, | |
687 | [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = HW_OP_UNSUPPORTED, | |
688 | }; | |
689 | ||
690 | static const unsigned armv8_pmuv3_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] | |
691 | [PERF_COUNT_HW_CACHE_OP_MAX] | |
692 | [PERF_COUNT_HW_CACHE_RESULT_MAX] = { | |
693 | [C(L1D)] = { | |
694 | [C(OP_READ)] = { | |
695 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, | |
696 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, | |
697 | }, | |
698 | [C(OP_WRITE)] = { | |
699 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_ACCESS, | |
700 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_L1_DCACHE_REFILL, | |
701 | }, | |
702 | [C(OP_PREFETCH)] = { | |
703 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
704 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
705 | }, | |
706 | }, | |
707 | [C(L1I)] = { | |
708 | [C(OP_READ)] = { | |
709 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
710 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
711 | }, | |
712 | [C(OP_WRITE)] = { | |
713 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
714 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
715 | }, | |
716 | [C(OP_PREFETCH)] = { | |
717 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
718 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
719 | }, | |
720 | }, | |
721 | [C(LL)] = { | |
722 | [C(OP_READ)] = { | |
723 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
724 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
725 | }, | |
726 | [C(OP_WRITE)] = { | |
727 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
728 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
729 | }, | |
730 | [C(OP_PREFETCH)] = { | |
731 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
732 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
733 | }, | |
734 | }, | |
735 | [C(DTLB)] = { | |
736 | [C(OP_READ)] = { | |
737 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
738 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
739 | }, | |
740 | [C(OP_WRITE)] = { | |
741 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
742 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
743 | }, | |
744 | [C(OP_PREFETCH)] = { | |
745 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
746 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
747 | }, | |
748 | }, | |
749 | [C(ITLB)] = { | |
750 | [C(OP_READ)] = { | |
751 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
752 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
753 | }, | |
754 | [C(OP_WRITE)] = { | |
755 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
756 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
757 | }, | |
758 | [C(OP_PREFETCH)] = { | |
759 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
760 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
761 | }, | |
762 | }, | |
763 | [C(BPU)] = { | |
764 | [C(OP_READ)] = { | |
765 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, | |
766 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, | |
767 | }, | |
768 | [C(OP_WRITE)] = { | |
769 | [C(RESULT_ACCESS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_PRED, | |
770 | [C(RESULT_MISS)] = ARMV8_PMUV3_PERFCTR_PC_BRANCH_MIS_PRED, | |
771 | }, | |
772 | [C(OP_PREFETCH)] = { | |
773 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
774 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
775 | }, | |
776 | }, | |
777 | [C(NODE)] = { | |
778 | [C(OP_READ)] = { | |
779 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
780 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
781 | }, | |
782 | [C(OP_WRITE)] = { | |
783 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
784 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
785 | }, | |
786 | [C(OP_PREFETCH)] = { | |
787 | [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, | |
788 | [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, | |
789 | }, | |
790 | }, | |
791 | }; | |
792 | ||
793 | /* | |
794 | * Perf Events' indices | |
795 | */ | |
796 | #define ARMV8_IDX_CYCLE_COUNTER 0 | |
797 | #define ARMV8_IDX_COUNTER0 1 | |
798 | #define ARMV8_IDX_COUNTER_LAST (ARMV8_IDX_CYCLE_COUNTER + cpu_pmu->num_events - 1) | |
799 | ||
800 | #define ARMV8_MAX_COUNTERS 32 | |
801 | #define ARMV8_COUNTER_MASK (ARMV8_MAX_COUNTERS - 1) | |
802 | ||
803 | /* | |
804 | * ARMv8 low level PMU access | |
805 | */ | |
806 | ||
807 | /* | |
808 | * Perf Event to low level counters mapping | |
809 | */ | |
810 | #define ARMV8_IDX_TO_COUNTER(x) \ | |
811 | (((x) - ARMV8_IDX_COUNTER0) & ARMV8_COUNTER_MASK) | |
812 | ||
813 | /* | |
814 | * Per-CPU PMCR: config reg | |
815 | */ | |
816 | #define ARMV8_PMCR_E (1 << 0) /* Enable all counters */ | |
817 | #define ARMV8_PMCR_P (1 << 1) /* Reset all counters */ | |
818 | #define ARMV8_PMCR_C (1 << 2) /* Cycle counter reset */ | |
819 | #define ARMV8_PMCR_D (1 << 3) /* CCNT counts every 64th cpu cycle */ | |
820 | #define ARMV8_PMCR_X (1 << 4) /* Export to ETM */ | |
821 | #define ARMV8_PMCR_DP (1 << 5) /* Disable CCNT if non-invasive debug*/ | |
822 | #define ARMV8_PMCR_N_SHIFT 11 /* Number of counters supported */ | |
823 | #define ARMV8_PMCR_N_MASK 0x1f | |
824 | #define ARMV8_PMCR_MASK 0x3f /* Mask for writable bits */ | |
825 | ||
826 | /* | |
827 | * PMOVSR: counters overflow flag status reg | |
828 | */ | |
829 | #define ARMV8_OVSR_MASK 0xffffffff /* Mask for writable bits */ | |
830 | #define ARMV8_OVERFLOWED_MASK ARMV8_OVSR_MASK | |
831 | ||
832 | /* | |
833 | * PMXEVTYPER: Event selection reg | |
834 | */ | |
c019de3d VK |
835 | #define ARMV8_EVTYPE_MASK 0xc80003ff /* Mask for writable bits */ |
836 | #define ARMV8_EVTYPE_EVENT 0x3ff /* Mask for EVENT bits */ | |
03089688 WD |
837 | |
838 | /* | |
839 | * Event filters for PMUv3 | |
840 | */ | |
841 | #define ARMV8_EXCLUDE_EL1 (1 << 31) | |
842 | #define ARMV8_EXCLUDE_EL0 (1 << 30) | |
843 | #define ARMV8_INCLUDE_EL2 (1 << 27) | |
844 | ||
845 | static inline u32 armv8pmu_pmcr_read(void) | |
846 | { | |
847 | u32 val; | |
848 | asm volatile("mrs %0, pmcr_el0" : "=r" (val)); | |
849 | return val; | |
850 | } | |
851 | ||
852 | static inline void armv8pmu_pmcr_write(u32 val) | |
853 | { | |
854 | val &= ARMV8_PMCR_MASK; | |
855 | isb(); | |
856 | asm volatile("msr pmcr_el0, %0" :: "r" (val)); | |
857 | } | |
858 | ||
859 | static inline int armv8pmu_has_overflowed(u32 pmovsr) | |
860 | { | |
861 | return pmovsr & ARMV8_OVERFLOWED_MASK; | |
862 | } | |
863 | ||
864 | static inline int armv8pmu_counter_valid(int idx) | |
865 | { | |
866 | return idx >= ARMV8_IDX_CYCLE_COUNTER && idx <= ARMV8_IDX_COUNTER_LAST; | |
867 | } | |
868 | ||
869 | static inline int armv8pmu_counter_has_overflowed(u32 pmnc, int idx) | |
870 | { | |
871 | int ret = 0; | |
872 | u32 counter; | |
873 | ||
874 | if (!armv8pmu_counter_valid(idx)) { | |
875 | pr_err("CPU%u checking wrong counter %d overflow status\n", | |
876 | smp_processor_id(), idx); | |
877 | } else { | |
878 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
879 | ret = pmnc & BIT(counter); | |
880 | } | |
881 | ||
882 | return ret; | |
883 | } | |
884 | ||
885 | static inline int armv8pmu_select_counter(int idx) | |
886 | { | |
887 | u32 counter; | |
888 | ||
889 | if (!armv8pmu_counter_valid(idx)) { | |
890 | pr_err("CPU%u selecting wrong PMNC counter %d\n", | |
891 | smp_processor_id(), idx); | |
892 | return -EINVAL; | |
893 | } | |
894 | ||
895 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
896 | asm volatile("msr pmselr_el0, %0" :: "r" (counter)); | |
897 | isb(); | |
898 | ||
899 | return idx; | |
900 | } | |
901 | ||
902 | static inline u32 armv8pmu_read_counter(int idx) | |
903 | { | |
904 | u32 value = 0; | |
905 | ||
906 | if (!armv8pmu_counter_valid(idx)) | |
907 | pr_err("CPU%u reading wrong counter %d\n", | |
908 | smp_processor_id(), idx); | |
909 | else if (idx == ARMV8_IDX_CYCLE_COUNTER) | |
910 | asm volatile("mrs %0, pmccntr_el0" : "=r" (value)); | |
911 | else if (armv8pmu_select_counter(idx) == idx) | |
912 | asm volatile("mrs %0, pmxevcntr_el0" : "=r" (value)); | |
913 | ||
914 | return value; | |
915 | } | |
916 | ||
917 | static inline void armv8pmu_write_counter(int idx, u32 value) | |
918 | { | |
919 | if (!armv8pmu_counter_valid(idx)) | |
920 | pr_err("CPU%u writing wrong counter %d\n", | |
921 | smp_processor_id(), idx); | |
922 | else if (idx == ARMV8_IDX_CYCLE_COUNTER) | |
923 | asm volatile("msr pmccntr_el0, %0" :: "r" (value)); | |
924 | else if (armv8pmu_select_counter(idx) == idx) | |
925 | asm volatile("msr pmxevcntr_el0, %0" :: "r" (value)); | |
926 | } | |
927 | ||
928 | static inline void armv8pmu_write_evtype(int idx, u32 val) | |
929 | { | |
930 | if (armv8pmu_select_counter(idx) == idx) { | |
931 | val &= ARMV8_EVTYPE_MASK; | |
932 | asm volatile("msr pmxevtyper_el0, %0" :: "r" (val)); | |
933 | } | |
934 | } | |
935 | ||
936 | static inline int armv8pmu_enable_counter(int idx) | |
937 | { | |
938 | u32 counter; | |
939 | ||
940 | if (!armv8pmu_counter_valid(idx)) { | |
941 | pr_err("CPU%u enabling wrong PMNC counter %d\n", | |
942 | smp_processor_id(), idx); | |
943 | return -EINVAL; | |
944 | } | |
945 | ||
946 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
947 | asm volatile("msr pmcntenset_el0, %0" :: "r" (BIT(counter))); | |
948 | return idx; | |
949 | } | |
950 | ||
951 | static inline int armv8pmu_disable_counter(int idx) | |
952 | { | |
953 | u32 counter; | |
954 | ||
955 | if (!armv8pmu_counter_valid(idx)) { | |
956 | pr_err("CPU%u disabling wrong PMNC counter %d\n", | |
957 | smp_processor_id(), idx); | |
958 | return -EINVAL; | |
959 | } | |
960 | ||
961 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
962 | asm volatile("msr pmcntenclr_el0, %0" :: "r" (BIT(counter))); | |
963 | return idx; | |
964 | } | |
965 | ||
966 | static inline int armv8pmu_enable_intens(int idx) | |
967 | { | |
968 | u32 counter; | |
969 | ||
970 | if (!armv8pmu_counter_valid(idx)) { | |
971 | pr_err("CPU%u enabling wrong PMNC counter IRQ enable %d\n", | |
972 | smp_processor_id(), idx); | |
973 | return -EINVAL; | |
974 | } | |
975 | ||
976 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
977 | asm volatile("msr pmintenset_el1, %0" :: "r" (BIT(counter))); | |
978 | return idx; | |
979 | } | |
980 | ||
981 | static inline int armv8pmu_disable_intens(int idx) | |
982 | { | |
983 | u32 counter; | |
984 | ||
985 | if (!armv8pmu_counter_valid(idx)) { | |
986 | pr_err("CPU%u disabling wrong PMNC counter IRQ enable %d\n", | |
987 | smp_processor_id(), idx); | |
988 | return -EINVAL; | |
989 | } | |
990 | ||
991 | counter = ARMV8_IDX_TO_COUNTER(idx); | |
992 | asm volatile("msr pmintenclr_el1, %0" :: "r" (BIT(counter))); | |
993 | isb(); | |
994 | /* Clear the overflow flag in case an interrupt is pending. */ | |
995 | asm volatile("msr pmovsclr_el0, %0" :: "r" (BIT(counter))); | |
996 | isb(); | |
997 | return idx; | |
998 | } | |
999 | ||
1000 | static inline u32 armv8pmu_getreset_flags(void) | |
1001 | { | |
1002 | u32 value; | |
1003 | ||
1004 | /* Read */ | |
1005 | asm volatile("mrs %0, pmovsclr_el0" : "=r" (value)); | |
1006 | ||
1007 | /* Write to clear flags */ | |
1008 | value &= ARMV8_OVSR_MASK; | |
1009 | asm volatile("msr pmovsclr_el0, %0" :: "r" (value)); | |
1010 | ||
1011 | return value; | |
1012 | } | |
1013 | ||
1014 | static void armv8pmu_enable_event(struct hw_perf_event *hwc, int idx) | |
1015 | { | |
1016 | unsigned long flags; | |
1017 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
1018 | ||
1019 | /* | |
1020 | * Enable counter and interrupt, and set the counter to count | |
1021 | * the event that we're interested in. | |
1022 | */ | |
1023 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
1024 | ||
1025 | /* | |
1026 | * Disable counter | |
1027 | */ | |
1028 | armv8pmu_disable_counter(idx); | |
1029 | ||
1030 | /* | |
1031 | * Set event (if destined for PMNx counters). | |
1032 | */ | |
1033 | armv8pmu_write_evtype(idx, hwc->config_base); | |
1034 | ||
1035 | /* | |
1036 | * Enable interrupt for this counter | |
1037 | */ | |
1038 | armv8pmu_enable_intens(idx); | |
1039 | ||
1040 | /* | |
1041 | * Enable counter | |
1042 | */ | |
1043 | armv8pmu_enable_counter(idx); | |
1044 | ||
1045 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
1046 | } | |
1047 | ||
1048 | static void armv8pmu_disable_event(struct hw_perf_event *hwc, int idx) | |
1049 | { | |
1050 | unsigned long flags; | |
1051 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
1052 | ||
1053 | /* | |
1054 | * Disable counter and interrupt | |
1055 | */ | |
1056 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
1057 | ||
1058 | /* | |
1059 | * Disable counter | |
1060 | */ | |
1061 | armv8pmu_disable_counter(idx); | |
1062 | ||
1063 | /* | |
1064 | * Disable interrupt for this counter | |
1065 | */ | |
1066 | armv8pmu_disable_intens(idx); | |
1067 | ||
1068 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
1069 | } | |
1070 | ||
1071 | static irqreturn_t armv8pmu_handle_irq(int irq_num, void *dev) | |
1072 | { | |
1073 | u32 pmovsr; | |
1074 | struct perf_sample_data data; | |
1075 | struct pmu_hw_events *cpuc; | |
1076 | struct pt_regs *regs; | |
1077 | int idx; | |
1078 | ||
1079 | /* | |
1080 | * Get and reset the IRQ flags | |
1081 | */ | |
1082 | pmovsr = armv8pmu_getreset_flags(); | |
1083 | ||
1084 | /* | |
1085 | * Did an overflow occur? | |
1086 | */ | |
1087 | if (!armv8pmu_has_overflowed(pmovsr)) | |
1088 | return IRQ_NONE; | |
1089 | ||
1090 | /* | |
1091 | * Handle the counter(s) overflow(s) | |
1092 | */ | |
1093 | regs = get_irq_regs(); | |
1094 | ||
1436c1aa | 1095 | cpuc = this_cpu_ptr(&cpu_hw_events); |
03089688 WD |
1096 | for (idx = 0; idx < cpu_pmu->num_events; ++idx) { |
1097 | struct perf_event *event = cpuc->events[idx]; | |
1098 | struct hw_perf_event *hwc; | |
1099 | ||
1100 | /* Ignore if we don't have an event. */ | |
1101 | if (!event) | |
1102 | continue; | |
1103 | ||
1104 | /* | |
1105 | * We have a single interrupt for all counters. Check that | |
1106 | * each counter has overflowed before we process it. | |
1107 | */ | |
1108 | if (!armv8pmu_counter_has_overflowed(pmovsr, idx)) | |
1109 | continue; | |
1110 | ||
1111 | hwc = &event->hw; | |
1112 | armpmu_event_update(event, hwc, idx); | |
1113 | perf_sample_data_init(&data, 0, hwc->last_period); | |
1114 | if (!armpmu_event_set_period(event, hwc, idx)) | |
1115 | continue; | |
1116 | ||
1117 | if (perf_event_overflow(event, &data, regs)) | |
1118 | cpu_pmu->disable(hwc, idx); | |
1119 | } | |
1120 | ||
1121 | /* | |
1122 | * Handle the pending perf events. | |
1123 | * | |
1124 | * Note: this call *must* be run with interrupts disabled. For | |
1125 | * platforms that can have the PMU interrupts raised as an NMI, this | |
1126 | * will not work. | |
1127 | */ | |
1128 | irq_work_run(); | |
1129 | ||
1130 | return IRQ_HANDLED; | |
1131 | } | |
1132 | ||
1133 | static void armv8pmu_start(void) | |
1134 | { | |
1135 | unsigned long flags; | |
1136 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
1137 | ||
1138 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
1139 | /* Enable all counters */ | |
1140 | armv8pmu_pmcr_write(armv8pmu_pmcr_read() | ARMV8_PMCR_E); | |
1141 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
1142 | } | |
1143 | ||
1144 | static void armv8pmu_stop(void) | |
1145 | { | |
1146 | unsigned long flags; | |
1147 | struct pmu_hw_events *events = cpu_pmu->get_hw_events(); | |
1148 | ||
1149 | raw_spin_lock_irqsave(&events->pmu_lock, flags); | |
1150 | /* Disable all counters */ | |
1151 | armv8pmu_pmcr_write(armv8pmu_pmcr_read() & ~ARMV8_PMCR_E); | |
1152 | raw_spin_unlock_irqrestore(&events->pmu_lock, flags); | |
1153 | } | |
1154 | ||
1155 | static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc, | |
1156 | struct hw_perf_event *event) | |
1157 | { | |
1158 | int idx; | |
1159 | unsigned long evtype = event->config_base & ARMV8_EVTYPE_EVENT; | |
1160 | ||
1161 | /* Always place a cycle counter into the cycle counter. */ | |
f46f979f | 1162 | if (evtype == ARMV8_PMUV3_PERFCTR_CLOCK_CYCLES) { |
03089688 WD |
1163 | if (test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask)) |
1164 | return -EAGAIN; | |
1165 | ||
1166 | return ARMV8_IDX_CYCLE_COUNTER; | |
1167 | } | |
1168 | ||
1169 | /* | |
1170 | * For anything other than a cycle counter, try and use | |
1171 | * the events counters | |
1172 | */ | |
1173 | for (idx = ARMV8_IDX_COUNTER0; idx < cpu_pmu->num_events; ++idx) { | |
1174 | if (!test_and_set_bit(idx, cpuc->used_mask)) | |
1175 | return idx; | |
1176 | } | |
1177 | ||
1178 | /* The counters are all in use. */ | |
1179 | return -EAGAIN; | |
1180 | } | |
1181 | ||
1182 | /* | |
1183 | * Add an event filter to a given event. This will only work for PMUv2 PMUs. | |
1184 | */ | |
1185 | static int armv8pmu_set_event_filter(struct hw_perf_event *event, | |
1186 | struct perf_event_attr *attr) | |
1187 | { | |
1188 | unsigned long config_base = 0; | |
1189 | ||
1190 | if (attr->exclude_idle) | |
1191 | return -EPERM; | |
1192 | if (attr->exclude_user) | |
1193 | config_base |= ARMV8_EXCLUDE_EL0; | |
1194 | if (attr->exclude_kernel) | |
1195 | config_base |= ARMV8_EXCLUDE_EL1; | |
1196 | if (!attr->exclude_hv) | |
1197 | config_base |= ARMV8_INCLUDE_EL2; | |
1198 | ||
1199 | /* | |
1200 | * Install the filter into config_base as this is used to | |
1201 | * construct the event type. | |
1202 | */ | |
1203 | event->config_base = config_base; | |
1204 | ||
1205 | return 0; | |
1206 | } | |
1207 | ||
1208 | static void armv8pmu_reset(void *info) | |
1209 | { | |
1210 | u32 idx, nb_cnt = cpu_pmu->num_events; | |
1211 | ||
1212 | /* The counter and interrupt enable registers are unknown at reset. */ | |
1213 | for (idx = ARMV8_IDX_CYCLE_COUNTER; idx < nb_cnt; ++idx) | |
1214 | armv8pmu_disable_event(NULL, idx); | |
1215 | ||
1216 | /* Initialize & Reset PMNC: C and P bits. */ | |
1217 | armv8pmu_pmcr_write(ARMV8_PMCR_P | ARMV8_PMCR_C); | |
1218 | ||
1219 | /* Disable access from userspace. */ | |
1220 | asm volatile("msr pmuserenr_el0, %0" :: "r" (0)); | |
1221 | } | |
1222 | ||
1223 | static int armv8_pmuv3_map_event(struct perf_event *event) | |
1224 | { | |
1225 | return map_cpu_event(event, &armv8_pmuv3_perf_map, | |
c019de3d VK |
1226 | &armv8_pmuv3_perf_cache_map, |
1227 | ARMV8_EVTYPE_EVENT); | |
03089688 WD |
1228 | } |
1229 | ||
1230 | static struct arm_pmu armv8pmu = { | |
1231 | .handle_irq = armv8pmu_handle_irq, | |
1232 | .enable = armv8pmu_enable_event, | |
1233 | .disable = armv8pmu_disable_event, | |
1234 | .read_counter = armv8pmu_read_counter, | |
1235 | .write_counter = armv8pmu_write_counter, | |
1236 | .get_event_idx = armv8pmu_get_event_idx, | |
1237 | .start = armv8pmu_start, | |
1238 | .stop = armv8pmu_stop, | |
1239 | .reset = armv8pmu_reset, | |
1240 | .max_period = (1LLU << 32) - 1, | |
1241 | }; | |
1242 | ||
1243 | static u32 __init armv8pmu_read_num_pmnc_events(void) | |
1244 | { | |
1245 | u32 nb_cnt; | |
1246 | ||
1247 | /* Read the nb of CNTx counters supported from PMNC */ | |
1248 | nb_cnt = (armv8pmu_pmcr_read() >> ARMV8_PMCR_N_SHIFT) & ARMV8_PMCR_N_MASK; | |
1249 | ||
1250 | /* Add the CPU cycles counter and return */ | |
1251 | return nb_cnt + 1; | |
1252 | } | |
1253 | ||
1254 | static struct arm_pmu *__init armv8_pmuv3_pmu_init(void) | |
1255 | { | |
1256 | armv8pmu.name = "arm/armv8-pmuv3"; | |
1257 | armv8pmu.map_event = armv8_pmuv3_map_event; | |
1258 | armv8pmu.num_events = armv8pmu_read_num_pmnc_events(); | |
1259 | armv8pmu.set_event_filter = armv8pmu_set_event_filter; | |
1260 | return &armv8pmu; | |
1261 | } | |
1262 | ||
1263 | /* | |
1264 | * Ensure the PMU has sane values out of reset. | |
1265 | * This requires SMP to be available, so exists as a separate initcall. | |
1266 | */ | |
1267 | static int __init | |
1268 | cpu_pmu_reset(void) | |
1269 | { | |
1270 | if (cpu_pmu && cpu_pmu->reset) | |
1271 | return on_each_cpu(cpu_pmu->reset, NULL, 1); | |
1272 | return 0; | |
1273 | } | |
1274 | arch_initcall(cpu_pmu_reset); | |
1275 | ||
1276 | /* | |
1277 | * PMU platform driver and devicetree bindings. | |
1278 | */ | |
1279 | static struct of_device_id armpmu_of_device_ids[] = { | |
1280 | {.compatible = "arm,armv8-pmuv3"}, | |
1281 | {}, | |
1282 | }; | |
1283 | ||
b881bc46 | 1284 | static int armpmu_device_probe(struct platform_device *pdev) |
03089688 WD |
1285 | { |
1286 | if (!cpu_pmu) | |
1287 | return -ENODEV; | |
1288 | ||
1289 | cpu_pmu->plat_device = pdev; | |
1290 | return 0; | |
1291 | } | |
1292 | ||
1293 | static struct platform_driver armpmu_driver = { | |
1294 | .driver = { | |
1295 | .name = "arm-pmu", | |
1296 | .of_match_table = armpmu_of_device_ids, | |
1297 | }, | |
1298 | .probe = armpmu_device_probe, | |
1299 | }; | |
1300 | ||
1301 | static int __init register_pmu_driver(void) | |
1302 | { | |
1303 | return platform_driver_register(&armpmu_driver); | |
1304 | } | |
1305 | device_initcall(register_pmu_driver); | |
1306 | ||
1307 | static struct pmu_hw_events *armpmu_get_cpu_events(void) | |
1308 | { | |
1436c1aa | 1309 | return this_cpu_ptr(&cpu_hw_events); |
03089688 WD |
1310 | } |
1311 | ||
1312 | static void __init cpu_pmu_init(struct arm_pmu *armpmu) | |
1313 | { | |
1314 | int cpu; | |
1315 | for_each_possible_cpu(cpu) { | |
1316 | struct pmu_hw_events *events = &per_cpu(cpu_hw_events, cpu); | |
1317 | events->events = per_cpu(hw_events, cpu); | |
1318 | events->used_mask = per_cpu(used_mask, cpu); | |
1319 | raw_spin_lock_init(&events->pmu_lock); | |
1320 | } | |
1321 | armpmu->get_hw_events = armpmu_get_cpu_events; | |
1322 | } | |
1323 | ||
1324 | static int __init init_hw_perf_events(void) | |
1325 | { | |
1326 | u64 dfr = read_cpuid(ID_AA64DFR0_EL1); | |
1327 | ||
1328 | switch ((dfr >> 8) & 0xf) { | |
1329 | case 0x1: /* PMUv3 */ | |
1330 | cpu_pmu = armv8_pmuv3_pmu_init(); | |
1331 | break; | |
1332 | } | |
1333 | ||
1334 | if (cpu_pmu) { | |
1335 | pr_info("enabled with %s PMU driver, %d counters available\n", | |
1336 | cpu_pmu->name, cpu_pmu->num_events); | |
1337 | cpu_pmu_init(cpu_pmu); | |
1338 | armpmu_register(cpu_pmu, "cpu", PERF_TYPE_RAW); | |
1339 | } else { | |
1340 | pr_info("no hardware support available\n"); | |
1341 | } | |
1342 | ||
1343 | return 0; | |
1344 | } | |
1345 | early_initcall(init_hw_perf_events); | |
1346 | ||
1347 | /* | |
1348 | * Callchain handling code. | |
1349 | */ | |
1350 | struct frame_tail { | |
23c7d70d JP |
1351 | struct frame_tail __user *fp; |
1352 | unsigned long lr; | |
03089688 WD |
1353 | } __attribute__((packed)); |
1354 | ||
1355 | /* | |
1356 | * Get the return address for a single stackframe and return a pointer to the | |
1357 | * next frame tail. | |
1358 | */ | |
1359 | static struct frame_tail __user * | |
1360 | user_backtrace(struct frame_tail __user *tail, | |
1361 | struct perf_callchain_entry *entry) | |
1362 | { | |
1363 | struct frame_tail buftail; | |
1364 | unsigned long err; | |
1365 | ||
1366 | /* Also check accessibility of one struct frame_tail beyond */ | |
1367 | if (!access_ok(VERIFY_READ, tail, sizeof(buftail))) | |
1368 | return NULL; | |
1369 | ||
1370 | pagefault_disable(); | |
1371 | err = __copy_from_user_inatomic(&buftail, tail, sizeof(buftail)); | |
1372 | pagefault_enable(); | |
1373 | ||
1374 | if (err) | |
1375 | return NULL; | |
1376 | ||
1377 | perf_callchain_store(entry, buftail.lr); | |
1378 | ||
1379 | /* | |
1380 | * Frame pointers should strictly progress back up the stack | |
1381 | * (towards higher addresses). | |
1382 | */ | |
1383 | if (tail >= buftail.fp) | |
1384 | return NULL; | |
1385 | ||
1386 | return buftail.fp; | |
1387 | } | |
1388 | ||
ff268ff7 | 1389 | #ifdef CONFIG_COMPAT |
23c7d70d JP |
1390 | /* |
1391 | * The registers we're interested in are at the end of the variable | |
1392 | * length saved register structure. The fp points at the end of this | |
1393 | * structure so the address of this struct is: | |
1394 | * (struct compat_frame_tail *)(xxx->fp)-1 | |
1395 | * | |
1396 | * This code has been adapted from the ARM OProfile support. | |
1397 | */ | |
1398 | struct compat_frame_tail { | |
1399 | compat_uptr_t fp; /* a (struct compat_frame_tail *) in compat mode */ | |
1400 | u32 sp; | |
1401 | u32 lr; | |
1402 | } __attribute__((packed)); | |
1403 | ||
1404 | static struct compat_frame_tail __user * | |
1405 | compat_user_backtrace(struct compat_frame_tail __user *tail, | |
1406 | struct perf_callchain_entry *entry) | |
1407 | { | |
1408 | struct compat_frame_tail buftail; | |
1409 | unsigned long err; | |
1410 | ||
1411 | /* Also check accessibility of one struct frame_tail beyond */ | |
1412 | if (!access_ok(VERIFY_READ, tail, sizeof(buftail))) | |
1413 | return NULL; | |
1414 | ||
1415 | pagefault_disable(); | |
1416 | err = __copy_from_user_inatomic(&buftail, tail, sizeof(buftail)); | |
1417 | pagefault_enable(); | |
1418 | ||
1419 | if (err) | |
1420 | return NULL; | |
1421 | ||
1422 | perf_callchain_store(entry, buftail.lr); | |
1423 | ||
1424 | /* | |
1425 | * Frame pointers should strictly progress back up the stack | |
1426 | * (towards higher addresses). | |
1427 | */ | |
1428 | if (tail + 1 >= (struct compat_frame_tail __user *) | |
1429 | compat_ptr(buftail.fp)) | |
1430 | return NULL; | |
1431 | ||
1432 | return (struct compat_frame_tail __user *)compat_ptr(buftail.fp) - 1; | |
1433 | } | |
ff268ff7 | 1434 | #endif /* CONFIG_COMPAT */ |
23c7d70d | 1435 | |
03089688 WD |
1436 | void perf_callchain_user(struct perf_callchain_entry *entry, |
1437 | struct pt_regs *regs) | |
1438 | { | |
75e42462 MZ |
1439 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) { |
1440 | /* We don't support guest os callchain now */ | |
1441 | return; | |
1442 | } | |
1443 | ||
abc41254 | 1444 | perf_callchain_store(entry, regs->pc); |
03089688 | 1445 | |
23c7d70d JP |
1446 | if (!compat_user_mode(regs)) { |
1447 | /* AARCH64 mode */ | |
1448 | struct frame_tail __user *tail; | |
1449 | ||
1450 | tail = (struct frame_tail __user *)regs->regs[29]; | |
1451 | ||
1452 | while (entry->nr < PERF_MAX_STACK_DEPTH && | |
1453 | tail && !((unsigned long)tail & 0xf)) | |
1454 | tail = user_backtrace(tail, entry); | |
1455 | } else { | |
ff268ff7 | 1456 | #ifdef CONFIG_COMPAT |
23c7d70d JP |
1457 | /* AARCH32 compat mode */ |
1458 | struct compat_frame_tail __user *tail; | |
1459 | ||
1460 | tail = (struct compat_frame_tail __user *)regs->compat_fp - 1; | |
1461 | ||
1462 | while ((entry->nr < PERF_MAX_STACK_DEPTH) && | |
1463 | tail && !((unsigned long)tail & 0x3)) | |
1464 | tail = compat_user_backtrace(tail, entry); | |
ff268ff7 | 1465 | #endif |
23c7d70d | 1466 | } |
03089688 WD |
1467 | } |
1468 | ||
1469 | /* | |
1470 | * Gets called by walk_stackframe() for every stackframe. This will be called | |
1471 | * whist unwinding the stackframe and is like a subroutine return so we use | |
1472 | * the PC. | |
1473 | */ | |
1474 | static int callchain_trace(struct stackframe *frame, void *data) | |
1475 | { | |
1476 | struct perf_callchain_entry *entry = data; | |
1477 | perf_callchain_store(entry, frame->pc); | |
1478 | return 0; | |
1479 | } | |
1480 | ||
1481 | void perf_callchain_kernel(struct perf_callchain_entry *entry, | |
1482 | struct pt_regs *regs) | |
1483 | { | |
1484 | struct stackframe frame; | |
1485 | ||
75e42462 MZ |
1486 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) { |
1487 | /* We don't support guest os callchain now */ | |
1488 | return; | |
1489 | } | |
1490 | ||
03089688 WD |
1491 | frame.fp = regs->regs[29]; |
1492 | frame.sp = regs->sp; | |
1493 | frame.pc = regs->pc; | |
23c7d70d | 1494 | |
03089688 WD |
1495 | walk_stackframe(&frame, callchain_trace, entry); |
1496 | } | |
75e42462 MZ |
1497 | |
1498 | unsigned long perf_instruction_pointer(struct pt_regs *regs) | |
1499 | { | |
1500 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) | |
1501 | return perf_guest_cbs->get_guest_ip(); | |
1502 | ||
1503 | return instruction_pointer(regs); | |
1504 | } | |
1505 | ||
1506 | unsigned long perf_misc_flags(struct pt_regs *regs) | |
1507 | { | |
1508 | int misc = 0; | |
1509 | ||
1510 | if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) { | |
1511 | if (perf_guest_cbs->is_user_mode()) | |
1512 | misc |= PERF_RECORD_MISC_GUEST_USER; | |
1513 | else | |
1514 | misc |= PERF_RECORD_MISC_GUEST_KERNEL; | |
1515 | } else { | |
1516 | if (user_mode(regs)) | |
1517 | misc |= PERF_RECORD_MISC_USER; | |
1518 | else | |
1519 | misc |= PERF_RECORD_MISC_KERNEL; | |
1520 | } | |
1521 | ||
1522 | return misc; | |
1523 | } |