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
10 * on the x86 code. Callchain code is based on the ARM OProfile backtrace
13 #define pr_fmt(fmt) "hw perfevents: " fmt
15 #include <linux/interrupt.h>
16 #include <linux/kernel.h>
17 #include <linux/module.h>
18 #include <linux/perf_event.h>
19 #include <linux/platform_device.h>
20 #include <linux/spinlock.h>
21 #include <linux/uaccess.h>
23 #include <asm/cputype.h>
25 #include <asm/irq_regs.h>
27 #include <asm/stacktrace.h>
29 static struct platform_device *pmu_device;
32 * Hardware lock to serialize accesses to PMU registers. Needed for the
33 * read/modify/write sequences.
35 static DEFINE_RAW_SPINLOCK(pmu_lock);
38 * ARMv6 supports a maximum of 3 events, starting from index 0. If we add
39 * another platform that supports more, we need to increase this to be the
40 * largest of all platforms.
42 * ARMv7 supports up to 32 events:
43 * cycle counter CCNT + 31 events counters CNT0..30.
44 * Cortex-A8 has 1+4 counters, Cortex-A9 has 1+6 counters.
46 #define ARMPMU_MAX_HWEVENTS 32
48 /* The events for a given CPU. */
49 struct cpu_hw_events {
51 * The events that are active on the CPU for the given index.
53 struct perf_event *events[ARMPMU_MAX_HWEVENTS];
56 * A 1 bit for an index indicates that the counter is being used for
57 * an event. A 0 means that the counter can be used.
59 unsigned long used_mask[BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)];
62 * A 1 bit for an index indicates that the counter is actively being
65 unsigned long active_mask[BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)];
67 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
70 enum arm_perf_pmu_ids id;
71 cpumask_t active_irqs;
73 irqreturn_t (*handle_irq)(int irq_num, void *dev);
74 void (*enable)(struct hw_perf_event *evt, int idx);
75 void (*disable)(struct hw_perf_event *evt, int idx);
76 int (*get_event_idx)(struct cpu_hw_events *cpuc,
77 struct hw_perf_event *hwc);
78 int (*set_event_filter)(struct hw_perf_event *evt,
79 struct perf_event_attr *attr);
80 u32 (*read_counter)(int idx);
81 void (*write_counter)(int idx, u32 val);
84 void (*reset)(void *);
85 const unsigned (*cache_map)[PERF_COUNT_HW_CACHE_MAX]
86 [PERF_COUNT_HW_CACHE_OP_MAX]
87 [PERF_COUNT_HW_CACHE_RESULT_MAX];
88 const unsigned (*event_map)[PERF_COUNT_HW_MAX];
94 /* Set at runtime when we know what CPU type we are. */
95 static struct arm_pmu *armpmu;
98 armpmu_get_pmu_id(void)
107 EXPORT_SYMBOL_GPL(armpmu_get_pmu_id);
110 armpmu_get_max_events(void)
115 max_events = armpmu->num_events;
119 EXPORT_SYMBOL_GPL(armpmu_get_max_events);
121 int perf_num_counters(void)
123 return armpmu_get_max_events();
125 EXPORT_SYMBOL_GPL(perf_num_counters);
127 #define HW_OP_UNSUPPORTED 0xFFFF
130 PERF_COUNT_HW_CACHE_##_x
132 #define CACHE_OP_UNSUPPORTED 0xFFFF
135 armpmu_map_cache_event(u64 config)
137 unsigned int cache_type, cache_op, cache_result, ret;
139 cache_type = (config >> 0) & 0xff;
140 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
143 cache_op = (config >> 8) & 0xff;
144 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
147 cache_result = (config >> 16) & 0xff;
148 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
151 ret = (int)(*armpmu->cache_map)[cache_type][cache_op][cache_result];
153 if (ret == CACHE_OP_UNSUPPORTED)
160 armpmu_map_event(u64 config)
162 int mapping = (*armpmu->event_map)[config];
163 return mapping == HW_OP_UNSUPPORTED ? -EOPNOTSUPP : mapping;
167 armpmu_map_raw_event(u64 config)
169 return (int)(config & armpmu->raw_event_mask);
173 armpmu_event_set_period(struct perf_event *event,
174 struct hw_perf_event *hwc,
177 s64 left = local64_read(&hwc->period_left);
178 s64 period = hwc->sample_period;
181 if (unlikely(left <= -period)) {
183 local64_set(&hwc->period_left, left);
184 hwc->last_period = period;
188 if (unlikely(left <= 0)) {
190 local64_set(&hwc->period_left, left);
191 hwc->last_period = period;
195 if (left > (s64)armpmu->max_period)
196 left = armpmu->max_period;
198 local64_set(&hwc->prev_count, (u64)-left);
200 armpmu->write_counter(idx, (u64)(-left) & 0xffffffff);
202 perf_event_update_userpage(event);
208 armpmu_event_update(struct perf_event *event,
209 struct hw_perf_event *hwc,
210 int idx, int overflow)
212 u64 delta, prev_raw_count, new_raw_count;
215 prev_raw_count = local64_read(&hwc->prev_count);
216 new_raw_count = armpmu->read_counter(idx);
218 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
219 new_raw_count) != prev_raw_count)
222 new_raw_count &= armpmu->max_period;
223 prev_raw_count &= armpmu->max_period;
226 delta = armpmu->max_period - prev_raw_count + new_raw_count + 1;
228 delta = new_raw_count - prev_raw_count;
230 local64_add(delta, &event->count);
231 local64_sub(delta, &hwc->period_left);
233 return new_raw_count;
237 armpmu_read(struct perf_event *event)
239 struct hw_perf_event *hwc = &event->hw;
241 /* Don't read disabled counters! */
245 armpmu_event_update(event, hwc, hwc->idx, 0);
249 armpmu_stop(struct perf_event *event, int flags)
251 struct hw_perf_event *hwc = &event->hw;
254 * ARM pmu always has to update the counter, so ignore
255 * PERF_EF_UPDATE, see comments in armpmu_start().
257 if (!(hwc->state & PERF_HES_STOPPED)) {
258 armpmu->disable(hwc, hwc->idx);
259 barrier(); /* why? */
260 armpmu_event_update(event, hwc, hwc->idx, 0);
261 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
266 armpmu_start(struct perf_event *event, int flags)
268 struct hw_perf_event *hwc = &event->hw;
271 * ARM pmu always has to reprogram the period, so ignore
272 * PERF_EF_RELOAD, see the comment below.
274 if (flags & PERF_EF_RELOAD)
275 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
279 * Set the period again. Some counters can't be stopped, so when we
280 * were stopped we simply disabled the IRQ source and the counter
281 * may have been left counting. If we don't do this step then we may
282 * get an interrupt too soon or *way* too late if the overflow has
283 * happened since disabling.
285 armpmu_event_set_period(event, hwc, hwc->idx);
286 armpmu->enable(hwc, hwc->idx);
290 armpmu_del(struct perf_event *event, int flags)
292 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
293 struct hw_perf_event *hwc = &event->hw;
298 clear_bit(idx, cpuc->active_mask);
299 armpmu_stop(event, PERF_EF_UPDATE);
300 cpuc->events[idx] = NULL;
301 clear_bit(idx, cpuc->used_mask);
303 perf_event_update_userpage(event);
307 armpmu_add(struct perf_event *event, int flags)
309 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
310 struct hw_perf_event *hwc = &event->hw;
314 perf_pmu_disable(event->pmu);
316 /* If we don't have a space for the counter then finish early. */
317 idx = armpmu->get_event_idx(cpuc, hwc);
324 * If there is an event in the counter we are going to use then make
325 * sure it is disabled.
328 armpmu->disable(hwc, idx);
329 cpuc->events[idx] = event;
330 set_bit(idx, cpuc->active_mask);
332 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
333 if (flags & PERF_EF_START)
334 armpmu_start(event, PERF_EF_RELOAD);
336 /* Propagate our changes to the userspace mapping. */
337 perf_event_update_userpage(event);
340 perf_pmu_enable(event->pmu);
344 static struct pmu pmu;
347 validate_event(struct cpu_hw_events *cpuc,
348 struct perf_event *event)
350 struct hw_perf_event fake_event = event->hw;
351 struct pmu *leader_pmu = event->group_leader->pmu;
353 if (event->pmu != leader_pmu || event->state <= PERF_EVENT_STATE_OFF)
356 return armpmu->get_event_idx(cpuc, &fake_event) >= 0;
360 validate_group(struct perf_event *event)
362 struct perf_event *sibling, *leader = event->group_leader;
363 struct cpu_hw_events fake_pmu;
365 memset(&fake_pmu, 0, sizeof(fake_pmu));
367 if (!validate_event(&fake_pmu, leader))
370 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
371 if (!validate_event(&fake_pmu, sibling))
375 if (!validate_event(&fake_pmu, event))
381 static irqreturn_t armpmu_platform_irq(int irq, void *dev)
383 struct arm_pmu_platdata *plat = dev_get_platdata(&pmu_device->dev);
385 return plat->handle_irq(irq, dev, armpmu->handle_irq);
389 armpmu_release_hardware(void)
393 irqs = min(pmu_device->num_resources, num_possible_cpus());
395 for (i = 0; i < irqs; ++i) {
396 if (!cpumask_test_and_clear_cpu(i, &armpmu->active_irqs))
398 irq = platform_get_irq(pmu_device, i);
404 release_pmu(ARM_PMU_DEVICE_CPU);
408 armpmu_reserve_hardware(void)
410 struct arm_pmu_platdata *plat;
411 irq_handler_t handle_irq;
412 int i, err, irq, irqs;
414 err = reserve_pmu(ARM_PMU_DEVICE_CPU);
416 pr_warning("unable to reserve pmu\n");
420 plat = dev_get_platdata(&pmu_device->dev);
421 if (plat && plat->handle_irq)
422 handle_irq = armpmu_platform_irq;
424 handle_irq = armpmu->handle_irq;
426 irqs = min(pmu_device->num_resources, num_possible_cpus());
428 pr_err("no irqs for PMUs defined\n");
432 for (i = 0; i < irqs; ++i) {
434 irq = platform_get_irq(pmu_device, i);
439 * If we have a single PMU interrupt that we can't shift,
440 * assume that we're running on a uniprocessor machine and
441 * continue. Otherwise, continue without this interrupt.
443 if (irq_set_affinity(irq, cpumask_of(i)) && irqs > 1) {
444 pr_warning("unable to set irq affinity (irq=%d, cpu=%u)\n",
449 err = request_irq(irq, handle_irq,
450 IRQF_DISABLED | IRQF_NOBALANCING,
453 pr_err("unable to request IRQ%d for ARM PMU counters\n",
455 armpmu_release_hardware();
459 cpumask_set_cpu(i, &armpmu->active_irqs);
465 static atomic_t active_events = ATOMIC_INIT(0);
466 static DEFINE_MUTEX(pmu_reserve_mutex);
469 hw_perf_event_destroy(struct perf_event *event)
471 if (atomic_dec_and_mutex_lock(&active_events, &pmu_reserve_mutex)) {
472 armpmu_release_hardware();
473 mutex_unlock(&pmu_reserve_mutex);
478 event_requires_mode_exclusion(struct perf_event_attr *attr)
480 return attr->exclude_idle || attr->exclude_user ||
481 attr->exclude_kernel || attr->exclude_hv;
485 __hw_perf_event_init(struct perf_event *event)
487 struct hw_perf_event *hwc = &event->hw;
490 /* Decode the generic type into an ARM event identifier. */
491 if (PERF_TYPE_HARDWARE == event->attr.type) {
492 mapping = armpmu_map_event(event->attr.config);
493 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
494 mapping = armpmu_map_cache_event(event->attr.config);
495 } else if (PERF_TYPE_RAW == event->attr.type) {
496 mapping = armpmu_map_raw_event(event->attr.config);
498 pr_debug("event type %x not supported\n", event->attr.type);
503 pr_debug("event %x:%llx not supported\n", event->attr.type,
509 * We don't assign an index until we actually place the event onto
510 * hardware. Use -1 to signify that we haven't decided where to put it
511 * yet. For SMP systems, each core has it's own PMU so we can't do any
512 * clever allocation or constraints checking at this point.
515 hwc->config_base = 0;
520 * Check whether we need to exclude the counter from certain modes.
522 if ((!armpmu->set_event_filter ||
523 armpmu->set_event_filter(hwc, &event->attr)) &&
524 event_requires_mode_exclusion(&event->attr)) {
525 pr_debug("ARM performance counters do not support "
531 * Store the event encoding into the config_base field.
533 hwc->config_base |= (unsigned long)mapping;
535 if (!hwc->sample_period) {
536 hwc->sample_period = armpmu->max_period;
537 hwc->last_period = hwc->sample_period;
538 local64_set(&hwc->period_left, hwc->sample_period);
542 if (event->group_leader != event) {
543 err = validate_group(event);
551 static int armpmu_event_init(struct perf_event *event)
555 switch (event->attr.type) {
557 case PERF_TYPE_HARDWARE:
558 case PERF_TYPE_HW_CACHE:
565 event->destroy = hw_perf_event_destroy;
567 if (!atomic_inc_not_zero(&active_events)) {
568 mutex_lock(&pmu_reserve_mutex);
569 if (atomic_read(&active_events) == 0) {
570 err = armpmu_reserve_hardware();
574 atomic_inc(&active_events);
575 mutex_unlock(&pmu_reserve_mutex);
581 err = __hw_perf_event_init(event);
583 hw_perf_event_destroy(event);
588 static void armpmu_enable(struct pmu *pmu)
590 /* Enable all of the perf events on hardware. */
591 int idx, enabled = 0;
592 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
594 for (idx = 0; idx < armpmu->num_events; ++idx) {
595 struct perf_event *event = cpuc->events[idx];
600 armpmu->enable(&event->hw, idx);
608 static void armpmu_disable(struct pmu *pmu)
613 static struct pmu pmu = {
614 .pmu_enable = armpmu_enable,
615 .pmu_disable = armpmu_disable,
616 .event_init = armpmu_event_init,
619 .start = armpmu_start,
624 /* Include the PMU-specific implementations. */
625 #include "perf_event_xscale.c"
626 #include "perf_event_v6.c"
627 #include "perf_event_v7.c"
630 * Ensure the PMU has sane values out of reset.
631 * This requires SMP to be available, so exists as a separate initcall.
636 if (armpmu && armpmu->reset)
637 return on_each_cpu(armpmu->reset, NULL, 1);
640 arch_initcall(armpmu_reset);
643 * PMU platform driver and devicetree bindings.
645 static struct of_device_id armpmu_of_device_ids[] = {
646 {.compatible = "arm,cortex-a9-pmu"},
647 {.compatible = "arm,cortex-a8-pmu"},
648 {.compatible = "arm,arm1136-pmu"},
649 {.compatible = "arm,arm1176-pmu"},
653 static struct platform_device_id armpmu_plat_device_ids[] = {
658 static int __devinit armpmu_device_probe(struct platform_device *pdev)
664 static struct platform_driver armpmu_driver = {
667 .of_match_table = armpmu_of_device_ids,
669 .probe = armpmu_device_probe,
670 .id_table = armpmu_plat_device_ids,
673 static int __init register_pmu_driver(void)
675 return platform_driver_register(&armpmu_driver);
677 device_initcall(register_pmu_driver);
680 * CPU PMU identification and registration.
683 init_hw_perf_events(void)
685 unsigned long cpuid = read_cpuid_id();
686 unsigned long implementor = (cpuid & 0xFF000000) >> 24;
687 unsigned long part_number = (cpuid & 0xFFF0);
690 if (0x41 == implementor) {
691 switch (part_number) {
692 case 0xB360: /* ARM1136 */
693 case 0xB560: /* ARM1156 */
694 case 0xB760: /* ARM1176 */
695 armpmu = armv6pmu_init();
697 case 0xB020: /* ARM11mpcore */
698 armpmu = armv6mpcore_pmu_init();
700 case 0xC080: /* Cortex-A8 */
701 armpmu = armv7_a8_pmu_init();
703 case 0xC090: /* Cortex-A9 */
704 armpmu = armv7_a9_pmu_init();
706 case 0xC050: /* Cortex-A5 */
707 armpmu = armv7_a5_pmu_init();
709 case 0xC0F0: /* Cortex-A15 */
710 armpmu = armv7_a15_pmu_init();
713 /* Intel CPUs [xscale]. */
714 } else if (0x69 == implementor) {
715 part_number = (cpuid >> 13) & 0x7;
716 switch (part_number) {
718 armpmu = xscale1pmu_init();
721 armpmu = xscale2pmu_init();
727 pr_info("enabled with %s PMU driver, %d counters available\n",
728 armpmu->name, armpmu->num_events);
729 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
731 pr_info("no hardware support available\n");
736 early_initcall(init_hw_perf_events);
739 * Callchain handling code.
743 * The registers we're interested in are at the end of the variable
744 * length saved register structure. The fp points at the end of this
745 * structure so the address of this struct is:
746 * (struct frame_tail *)(xxx->fp)-1
748 * This code has been adapted from the ARM OProfile support.
751 struct frame_tail __user *fp;
754 } __attribute__((packed));
757 * Get the return address for a single stackframe and return a pointer to the
760 static struct frame_tail __user *
761 user_backtrace(struct frame_tail __user *tail,
762 struct perf_callchain_entry *entry)
764 struct frame_tail buftail;
766 /* Also check accessibility of one struct frame_tail beyond */
767 if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
769 if (__copy_from_user_inatomic(&buftail, tail, sizeof(buftail)))
772 perf_callchain_store(entry, buftail.lr);
775 * Frame pointers should strictly progress back up the stack
776 * (towards higher addresses).
778 if (tail + 1 >= buftail.fp)
781 return buftail.fp - 1;
785 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
787 struct frame_tail __user *tail;
790 tail = (struct frame_tail __user *)regs->ARM_fp - 1;
792 while ((entry->nr < PERF_MAX_STACK_DEPTH) &&
793 tail && !((unsigned long)tail & 0x3))
794 tail = user_backtrace(tail, entry);
798 * Gets called by walk_stackframe() for every stackframe. This will be called
799 * whist unwinding the stackframe and is like a subroutine return so we use
803 callchain_trace(struct stackframe *fr,
806 struct perf_callchain_entry *entry = data;
807 perf_callchain_store(entry, fr->pc);
812 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
814 struct stackframe fr;
816 fr.fp = regs->ARM_fp;
817 fr.sp = regs->ARM_sp;
818 fr.lr = regs->ARM_lr;
819 fr.pc = regs->ARM_pc;
820 walk_stackframe(&fr, callchain_trace, entry);