4 * ARM performance counter support.
6 * Copyright (C) 2009 picoChip Designs, Ltd., Jamie Iles
7 * Copyright (C) 2010 ARM Ltd., Will Deacon <will.deacon@arm.com>
9 * This code is based on the sparc64 perf event code, which is in turn based
12 #define pr_fmt(fmt) "hw perfevents: " fmt
14 #include <linux/kernel.h>
15 #include <linux/platform_device.h>
16 #include <linux/pm_runtime.h>
17 #include <linux/irq.h>
18 #include <linux/irqdesc.h>
20 #include <asm/irq_regs.h>
24 armpmu_map_cache_event(const unsigned (*cache_map)
25 [PERF_COUNT_HW_CACHE_MAX]
26 [PERF_COUNT_HW_CACHE_OP_MAX]
27 [PERF_COUNT_HW_CACHE_RESULT_MAX],
30 unsigned int cache_type, cache_op, cache_result, ret;
32 cache_type = (config >> 0) & 0xff;
33 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
36 cache_op = (config >> 8) & 0xff;
37 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
40 cache_result = (config >> 16) & 0xff;
41 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
44 ret = (int)(*cache_map)[cache_type][cache_op][cache_result];
46 if (ret == CACHE_OP_UNSUPPORTED)
53 armpmu_map_hw_event(const unsigned (*event_map)[PERF_COUNT_HW_MAX], u64 config)
57 if (config >= PERF_COUNT_HW_MAX)
60 mapping = (*event_map)[config];
61 return mapping == HW_OP_UNSUPPORTED ? -ENOENT : mapping;
65 armpmu_map_raw_event(u32 raw_event_mask, u64 config)
67 return (int)(config & raw_event_mask);
71 armpmu_map_event(struct perf_event *event,
72 const unsigned (*event_map)[PERF_COUNT_HW_MAX],
73 const unsigned (*cache_map)
74 [PERF_COUNT_HW_CACHE_MAX]
75 [PERF_COUNT_HW_CACHE_OP_MAX]
76 [PERF_COUNT_HW_CACHE_RESULT_MAX],
79 u64 config = event->attr.config;
80 int type = event->attr.type;
82 if (type == event->pmu->type)
83 return armpmu_map_raw_event(raw_event_mask, config);
86 case PERF_TYPE_HARDWARE:
87 return armpmu_map_hw_event(event_map, config);
88 case PERF_TYPE_HW_CACHE:
89 return armpmu_map_cache_event(cache_map, config);
91 return armpmu_map_raw_event(raw_event_mask, config);
97 int armpmu_event_set_period(struct perf_event *event)
99 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
100 struct hw_perf_event *hwc = &event->hw;
101 s64 left = local64_read(&hwc->period_left);
102 s64 period = hwc->sample_period;
105 if (unlikely(left <= -period)) {
107 local64_set(&hwc->period_left, left);
108 hwc->last_period = period;
112 if (unlikely(left <= 0)) {
114 local64_set(&hwc->period_left, left);
115 hwc->last_period = period;
120 * Limit the maximum period to prevent the counter value
121 * from overtaking the one we are about to program. In
122 * effect we are reducing max_period to account for
123 * interrupt latency (and we are being very conservative).
125 if (left > (armpmu->max_period >> 1))
126 left = armpmu->max_period >> 1;
128 local64_set(&hwc->prev_count, (u64)-left);
130 armpmu->write_counter(event, (u64)(-left) & 0xffffffff);
132 perf_event_update_userpage(event);
137 u64 armpmu_event_update(struct perf_event *event)
139 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
140 struct hw_perf_event *hwc = &event->hw;
141 u64 delta, prev_raw_count, new_raw_count;
144 prev_raw_count = local64_read(&hwc->prev_count);
145 new_raw_count = armpmu->read_counter(event);
147 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
148 new_raw_count) != prev_raw_count)
151 delta = (new_raw_count - prev_raw_count) & armpmu->max_period;
153 local64_add(delta, &event->count);
154 local64_sub(delta, &hwc->period_left);
156 return new_raw_count;
160 armpmu_read(struct perf_event *event)
162 armpmu_event_update(event);
166 armpmu_stop(struct perf_event *event, int flags)
168 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
169 struct hw_perf_event *hwc = &event->hw;
172 * ARM pmu always has to update the counter, so ignore
173 * PERF_EF_UPDATE, see comments in armpmu_start().
175 if (!(hwc->state & PERF_HES_STOPPED)) {
176 armpmu->disable(event);
177 armpmu_event_update(event);
178 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
182 static void armpmu_start(struct perf_event *event, int flags)
184 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
185 struct hw_perf_event *hwc = &event->hw;
188 * ARM pmu always has to reprogram the period, so ignore
189 * PERF_EF_RELOAD, see the comment below.
191 if (flags & PERF_EF_RELOAD)
192 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
196 * Set the period again. Some counters can't be stopped, so when we
197 * were stopped we simply disabled the IRQ source and the counter
198 * may have been left counting. If we don't do this step then we may
199 * get an interrupt too soon or *way* too late if the overflow has
200 * happened since disabling.
202 armpmu_event_set_period(event);
203 armpmu->enable(event);
207 armpmu_del(struct perf_event *event, int flags)
209 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
210 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
211 struct hw_perf_event *hwc = &event->hw;
214 armpmu_stop(event, PERF_EF_UPDATE);
215 hw_events->events[idx] = NULL;
216 clear_bit(idx, hw_events->used_mask);
217 if (armpmu->clear_event_idx)
218 armpmu->clear_event_idx(hw_events, event);
220 perf_event_update_userpage(event);
224 armpmu_add(struct perf_event *event, int flags)
226 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
227 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
228 struct hw_perf_event *hwc = &event->hw;
232 perf_pmu_disable(event->pmu);
234 /* If we don't have a space for the counter then finish early. */
235 idx = armpmu->get_event_idx(hw_events, event);
242 * If there is an event in the counter we are going to use then make
243 * sure it is disabled.
246 armpmu->disable(event);
247 hw_events->events[idx] = event;
249 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
250 if (flags & PERF_EF_START)
251 armpmu_start(event, PERF_EF_RELOAD);
253 /* Propagate our changes to the userspace mapping. */
254 perf_event_update_userpage(event);
257 perf_pmu_enable(event->pmu);
262 validate_event(struct pmu *pmu, struct pmu_hw_events *hw_events,
263 struct perf_event *event)
265 struct arm_pmu *armpmu;
267 if (is_software_event(event))
271 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
272 * core perf code won't check that the pmu->ctx == leader->ctx
273 * until after pmu->event_init(event).
275 if (event->pmu != pmu)
278 if (event->state < PERF_EVENT_STATE_OFF)
281 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
284 armpmu = to_arm_pmu(event->pmu);
285 return armpmu->get_event_idx(hw_events, event) >= 0;
289 validate_group(struct perf_event *event)
291 struct perf_event *sibling, *leader = event->group_leader;
292 struct pmu_hw_events fake_pmu;
295 * Initialise the fake PMU. We only need to populate the
296 * used_mask for the purposes of validation.
298 memset(&fake_pmu.used_mask, 0, sizeof(fake_pmu.used_mask));
300 if (!validate_event(event->pmu, &fake_pmu, leader))
303 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
304 if (!validate_event(event->pmu, &fake_pmu, sibling))
308 if (!validate_event(event->pmu, &fake_pmu, event))
314 static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
316 struct arm_pmu *armpmu;
317 struct platform_device *plat_device;
318 struct arm_pmu_platdata *plat;
320 u64 start_clock, finish_clock;
323 * we request the IRQ with a (possibly percpu) struct arm_pmu**, but
324 * the handlers expect a struct arm_pmu*. The percpu_irq framework will
325 * do any necessary shifting, we just need to perform the first
328 armpmu = *(void **)dev;
329 plat_device = armpmu->plat_device;
330 plat = dev_get_platdata(&plat_device->dev);
332 start_clock = sched_clock();
333 if (plat && plat->handle_irq)
334 ret = plat->handle_irq(irq, armpmu, armpmu->handle_irq);
336 ret = armpmu->handle_irq(irq, armpmu);
337 finish_clock = sched_clock();
339 perf_sample_event_took(finish_clock - start_clock);
344 armpmu_release_hardware(struct arm_pmu *armpmu)
346 armpmu->free_irq(armpmu);
347 pm_runtime_put_sync(&armpmu->plat_device->dev);
351 armpmu_reserve_hardware(struct arm_pmu *armpmu)
354 struct platform_device *pmu_device = armpmu->plat_device;
359 pm_runtime_get_sync(&pmu_device->dev);
360 err = armpmu->request_irq(armpmu, armpmu_dispatch_irq);
362 armpmu_release_hardware(armpmu);
370 hw_perf_event_destroy(struct perf_event *event)
372 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
373 atomic_t *active_events = &armpmu->active_events;
374 struct mutex *pmu_reserve_mutex = &armpmu->reserve_mutex;
376 if (atomic_dec_and_mutex_lock(active_events, pmu_reserve_mutex)) {
377 armpmu_release_hardware(armpmu);
378 mutex_unlock(pmu_reserve_mutex);
383 event_requires_mode_exclusion(struct perf_event_attr *attr)
385 return attr->exclude_idle || attr->exclude_user ||
386 attr->exclude_kernel || attr->exclude_hv;
390 __hw_perf_event_init(struct perf_event *event)
392 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
393 struct hw_perf_event *hwc = &event->hw;
396 mapping = armpmu->map_event(event);
399 pr_debug("event %x:%llx not supported\n", event->attr.type,
405 * We don't assign an index until we actually place the event onto
406 * hardware. Use -1 to signify that we haven't decided where to put it
407 * yet. For SMP systems, each core has it's own PMU so we can't do any
408 * clever allocation or constraints checking at this point.
411 hwc->config_base = 0;
416 * Check whether we need to exclude the counter from certain modes.
418 if ((!armpmu->set_event_filter ||
419 armpmu->set_event_filter(hwc, &event->attr)) &&
420 event_requires_mode_exclusion(&event->attr)) {
421 pr_debug("ARM performance counters do not support "
427 * Store the event encoding into the config_base field.
429 hwc->config_base |= (unsigned long)mapping;
431 if (!is_sampling_event(event)) {
433 * For non-sampling runs, limit the sample_period to half
434 * of the counter width. That way, the new counter value
435 * is far less likely to overtake the previous one unless
436 * you have some serious IRQ latency issues.
438 hwc->sample_period = armpmu->max_period >> 1;
439 hwc->last_period = hwc->sample_period;
440 local64_set(&hwc->period_left, hwc->sample_period);
443 if (event->group_leader != event) {
444 if (validate_group(event) != 0)
451 static int armpmu_event_init(struct perf_event *event)
453 struct arm_pmu *armpmu = to_arm_pmu(event->pmu);
455 atomic_t *active_events = &armpmu->active_events;
457 /* does not support taken branch sampling */
458 if (has_branch_stack(event))
461 if (armpmu->map_event(event) == -ENOENT)
464 event->destroy = hw_perf_event_destroy;
466 if (!atomic_inc_not_zero(active_events)) {
467 mutex_lock(&armpmu->reserve_mutex);
468 if (atomic_read(active_events) == 0)
469 err = armpmu_reserve_hardware(armpmu);
472 atomic_inc(active_events);
473 mutex_unlock(&armpmu->reserve_mutex);
479 err = __hw_perf_event_init(event);
481 hw_perf_event_destroy(event);
486 static void armpmu_enable(struct pmu *pmu)
488 struct arm_pmu *armpmu = to_arm_pmu(pmu);
489 struct pmu_hw_events *hw_events = this_cpu_ptr(armpmu->hw_events);
490 int enabled = bitmap_weight(hw_events->used_mask, armpmu->num_events);
493 armpmu->start(armpmu);
496 static void armpmu_disable(struct pmu *pmu)
498 struct arm_pmu *armpmu = to_arm_pmu(pmu);
499 armpmu->stop(armpmu);
503 static int armpmu_runtime_resume(struct device *dev)
505 struct arm_pmu_platdata *plat = dev_get_platdata(dev);
507 if (plat && plat->runtime_resume)
508 return plat->runtime_resume(dev);
513 static int armpmu_runtime_suspend(struct device *dev)
515 struct arm_pmu_platdata *plat = dev_get_platdata(dev);
517 if (plat && plat->runtime_suspend)
518 return plat->runtime_suspend(dev);
524 const struct dev_pm_ops armpmu_dev_pm_ops = {
525 SET_RUNTIME_PM_OPS(armpmu_runtime_suspend, armpmu_runtime_resume, NULL)
528 static void armpmu_init(struct arm_pmu *armpmu)
530 atomic_set(&armpmu->active_events, 0);
531 mutex_init(&armpmu->reserve_mutex);
533 armpmu->pmu = (struct pmu) {
534 .pmu_enable = armpmu_enable,
535 .pmu_disable = armpmu_disable,
536 .event_init = armpmu_event_init,
539 .start = armpmu_start,
545 int armpmu_register(struct arm_pmu *armpmu, int type)
548 pm_runtime_enable(&armpmu->plat_device->dev);
549 pr_info("enabled with %s PMU driver, %d counters available\n",
550 armpmu->name, armpmu->num_events);
551 return perf_pmu_register(&armpmu->pmu, armpmu->name, type);