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>
30 * Hardware lock to serialize accesses to PMU registers. Needed for the
31 * read/modify/write sequences.
33 static DEFINE_RAW_SPINLOCK(pmu_lock);
36 * ARMv6 supports a maximum of 3 events, starting from index 0. If we add
37 * another platform that supports more, we need to increase this to be the
38 * largest of all platforms.
40 * ARMv7 supports up to 32 events:
41 * cycle counter CCNT + 31 events counters CNT0..30.
42 * Cortex-A8 has 1+4 counters, Cortex-A9 has 1+6 counters.
44 #define ARMPMU_MAX_HWEVENTS 32
46 /* The events for a given CPU. */
47 struct cpu_hw_events {
49 * The events that are active on the CPU for the given index.
51 struct perf_event *events[ARMPMU_MAX_HWEVENTS];
54 * A 1 bit for an index indicates that the counter is being used for
55 * an event. A 0 means that the counter can be used.
57 unsigned long used_mask[BITS_TO_LONGS(ARMPMU_MAX_HWEVENTS)];
59 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
62 enum arm_perf_pmu_ids id;
63 cpumask_t active_irqs;
65 irqreturn_t (*handle_irq)(int irq_num, void *dev);
66 void (*enable)(struct hw_perf_event *evt, int idx);
67 void (*disable)(struct hw_perf_event *evt, int idx);
68 int (*get_event_idx)(struct cpu_hw_events *cpuc,
69 struct hw_perf_event *hwc);
70 int (*set_event_filter)(struct hw_perf_event *evt,
71 struct perf_event_attr *attr);
72 u32 (*read_counter)(int idx);
73 void (*write_counter)(int idx, u32 val);
76 void (*reset)(void *);
77 const unsigned (*cache_map)[PERF_COUNT_HW_CACHE_MAX]
78 [PERF_COUNT_HW_CACHE_OP_MAX]
79 [PERF_COUNT_HW_CACHE_RESULT_MAX];
80 const unsigned (*event_map)[PERF_COUNT_HW_MAX];
83 atomic_t active_events;
84 struct mutex reserve_mutex;
86 struct platform_device *plat_device;
87 struct cpu_hw_events *(*get_hw_events)(void);
90 /* Set at runtime when we know what CPU type we are. */
91 static struct arm_pmu *armpmu;
94 armpmu_get_pmu_id(void)
103 EXPORT_SYMBOL_GPL(armpmu_get_pmu_id);
106 armpmu_get_max_events(void)
111 max_events = armpmu->num_events;
115 EXPORT_SYMBOL_GPL(armpmu_get_max_events);
117 int perf_num_counters(void)
119 return armpmu_get_max_events();
121 EXPORT_SYMBOL_GPL(perf_num_counters);
123 #define HW_OP_UNSUPPORTED 0xFFFF
126 PERF_COUNT_HW_CACHE_##_x
128 #define CACHE_OP_UNSUPPORTED 0xFFFF
131 armpmu_map_cache_event(u64 config)
133 unsigned int cache_type, cache_op, cache_result, ret;
135 cache_type = (config >> 0) & 0xff;
136 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
139 cache_op = (config >> 8) & 0xff;
140 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
143 cache_result = (config >> 16) & 0xff;
144 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
147 ret = (int)(*armpmu->cache_map)[cache_type][cache_op][cache_result];
149 if (ret == CACHE_OP_UNSUPPORTED)
156 armpmu_map_event(u64 config)
158 int mapping = (*armpmu->event_map)[config];
159 return mapping == HW_OP_UNSUPPORTED ? -EOPNOTSUPP : mapping;
163 armpmu_map_raw_event(u64 config)
165 return (int)(config & armpmu->raw_event_mask);
169 armpmu_event_set_period(struct perf_event *event,
170 struct hw_perf_event *hwc,
173 s64 left = local64_read(&hwc->period_left);
174 s64 period = hwc->sample_period;
177 if (unlikely(left <= -period)) {
179 local64_set(&hwc->period_left, left);
180 hwc->last_period = period;
184 if (unlikely(left <= 0)) {
186 local64_set(&hwc->period_left, left);
187 hwc->last_period = period;
191 if (left > (s64)armpmu->max_period)
192 left = armpmu->max_period;
194 local64_set(&hwc->prev_count, (u64)-left);
196 armpmu->write_counter(idx, (u64)(-left) & 0xffffffff);
198 perf_event_update_userpage(event);
204 armpmu_event_update(struct perf_event *event,
205 struct hw_perf_event *hwc,
206 int idx, int overflow)
208 u64 delta, prev_raw_count, new_raw_count;
211 prev_raw_count = local64_read(&hwc->prev_count);
212 new_raw_count = armpmu->read_counter(idx);
214 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
215 new_raw_count) != prev_raw_count)
218 new_raw_count &= armpmu->max_period;
219 prev_raw_count &= armpmu->max_period;
222 delta = armpmu->max_period - prev_raw_count + new_raw_count + 1;
224 delta = new_raw_count - prev_raw_count;
226 local64_add(delta, &event->count);
227 local64_sub(delta, &hwc->period_left);
229 return new_raw_count;
233 armpmu_read(struct perf_event *event)
235 struct hw_perf_event *hwc = &event->hw;
237 /* Don't read disabled counters! */
241 armpmu_event_update(event, hwc, hwc->idx, 0);
245 armpmu_stop(struct perf_event *event, int flags)
247 struct hw_perf_event *hwc = &event->hw;
250 * ARM pmu always has to update the counter, so ignore
251 * PERF_EF_UPDATE, see comments in armpmu_start().
253 if (!(hwc->state & PERF_HES_STOPPED)) {
254 armpmu->disable(hwc, hwc->idx);
255 barrier(); /* why? */
256 armpmu_event_update(event, hwc, hwc->idx, 0);
257 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
262 armpmu_start(struct perf_event *event, int flags)
264 struct hw_perf_event *hwc = &event->hw;
267 * ARM pmu always has to reprogram the period, so ignore
268 * PERF_EF_RELOAD, see the comment below.
270 if (flags & PERF_EF_RELOAD)
271 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
275 * Set the period again. Some counters can't be stopped, so when we
276 * were stopped we simply disabled the IRQ source and the counter
277 * may have been left counting. If we don't do this step then we may
278 * get an interrupt too soon or *way* too late if the overflow has
279 * happened since disabling.
281 armpmu_event_set_period(event, hwc, hwc->idx);
282 armpmu->enable(hwc, hwc->idx);
286 armpmu_del(struct perf_event *event, int flags)
288 struct cpu_hw_events *cpuc = armpmu->get_hw_events();
289 struct hw_perf_event *hwc = &event->hw;
294 armpmu_stop(event, PERF_EF_UPDATE);
295 cpuc->events[idx] = NULL;
296 clear_bit(idx, cpuc->used_mask);
298 perf_event_update_userpage(event);
302 armpmu_add(struct perf_event *event, int flags)
304 struct cpu_hw_events *cpuc = armpmu->get_hw_events();
305 struct hw_perf_event *hwc = &event->hw;
309 perf_pmu_disable(event->pmu);
311 /* If we don't have a space for the counter then finish early. */
312 idx = armpmu->get_event_idx(cpuc, hwc);
319 * If there is an event in the counter we are going to use then make
320 * sure it is disabled.
323 armpmu->disable(hwc, idx);
324 cpuc->events[idx] = event;
326 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
327 if (flags & PERF_EF_START)
328 armpmu_start(event, PERF_EF_RELOAD);
330 /* Propagate our changes to the userspace mapping. */
331 perf_event_update_userpage(event);
334 perf_pmu_enable(event->pmu);
338 static struct pmu pmu;
341 validate_event(struct cpu_hw_events *cpuc,
342 struct perf_event *event)
344 struct hw_perf_event fake_event = event->hw;
345 struct pmu *leader_pmu = event->group_leader->pmu;
347 if (event->pmu != leader_pmu || event->state <= PERF_EVENT_STATE_OFF)
350 return armpmu->get_event_idx(cpuc, &fake_event) >= 0;
354 validate_group(struct perf_event *event)
356 struct perf_event *sibling, *leader = event->group_leader;
357 struct cpu_hw_events fake_pmu;
359 memset(&fake_pmu, 0, sizeof(fake_pmu));
361 if (!validate_event(&fake_pmu, leader))
364 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
365 if (!validate_event(&fake_pmu, sibling))
369 if (!validate_event(&fake_pmu, event))
375 static irqreturn_t armpmu_platform_irq(int irq, void *dev)
377 struct platform_device *plat_device = armpmu->plat_device;
378 struct arm_pmu_platdata *plat = dev_get_platdata(&plat_device->dev);
380 return plat->handle_irq(irq, dev, armpmu->handle_irq);
384 armpmu_release_hardware(void)
387 struct platform_device *pmu_device = armpmu->plat_device;
389 irqs = min(pmu_device->num_resources, num_possible_cpus());
391 for (i = 0; i < irqs; ++i) {
392 if (!cpumask_test_and_clear_cpu(i, &armpmu->active_irqs))
394 irq = platform_get_irq(pmu_device, i);
400 release_pmu(ARM_PMU_DEVICE_CPU);
404 armpmu_reserve_hardware(void)
406 struct arm_pmu_platdata *plat;
407 irq_handler_t handle_irq;
408 int i, err, irq, irqs;
409 struct platform_device *pmu_device = armpmu->plat_device;
411 err = reserve_pmu(ARM_PMU_DEVICE_CPU);
413 pr_warning("unable to reserve pmu\n");
417 plat = dev_get_platdata(&pmu_device->dev);
418 if (plat && plat->handle_irq)
419 handle_irq = armpmu_platform_irq;
421 handle_irq = armpmu->handle_irq;
423 irqs = min(pmu_device->num_resources, num_possible_cpus());
425 pr_err("no irqs for PMUs defined\n");
429 for (i = 0; i < irqs; ++i) {
431 irq = platform_get_irq(pmu_device, i);
436 * If we have a single PMU interrupt that we can't shift,
437 * assume that we're running on a uniprocessor machine and
438 * continue. Otherwise, continue without this interrupt.
440 if (irq_set_affinity(irq, cpumask_of(i)) && irqs > 1) {
441 pr_warning("unable to set irq affinity (irq=%d, cpu=%u)\n",
446 err = request_irq(irq, handle_irq,
447 IRQF_DISABLED | IRQF_NOBALANCING,
450 pr_err("unable to request IRQ%d for ARM PMU counters\n",
452 armpmu_release_hardware();
456 cpumask_set_cpu(i, &armpmu->active_irqs);
463 hw_perf_event_destroy(struct perf_event *event)
465 atomic_t *active_events = &armpmu->active_events;
466 struct mutex *pmu_reserve_mutex = &armpmu->reserve_mutex;
468 if (atomic_dec_and_mutex_lock(active_events, pmu_reserve_mutex)) {
469 armpmu_release_hardware();
470 mutex_unlock(pmu_reserve_mutex);
475 event_requires_mode_exclusion(struct perf_event_attr *attr)
477 return attr->exclude_idle || attr->exclude_user ||
478 attr->exclude_kernel || attr->exclude_hv;
482 __hw_perf_event_init(struct perf_event *event)
484 struct hw_perf_event *hwc = &event->hw;
487 /* Decode the generic type into an ARM event identifier. */
488 if (PERF_TYPE_HARDWARE == event->attr.type) {
489 mapping = armpmu_map_event(event->attr.config);
490 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
491 mapping = armpmu_map_cache_event(event->attr.config);
492 } else if (PERF_TYPE_RAW == event->attr.type) {
493 mapping = armpmu_map_raw_event(event->attr.config);
495 pr_debug("event type %x not supported\n", event->attr.type);
500 pr_debug("event %x:%llx not supported\n", event->attr.type,
506 * We don't assign an index until we actually place the event onto
507 * hardware. Use -1 to signify that we haven't decided where to put it
508 * yet. For SMP systems, each core has it's own PMU so we can't do any
509 * clever allocation or constraints checking at this point.
512 hwc->config_base = 0;
517 * Check whether we need to exclude the counter from certain modes.
519 if ((!armpmu->set_event_filter ||
520 armpmu->set_event_filter(hwc, &event->attr)) &&
521 event_requires_mode_exclusion(&event->attr)) {
522 pr_debug("ARM performance counters do not support "
528 * Store the event encoding into the config_base field.
530 hwc->config_base |= (unsigned long)mapping;
532 if (!hwc->sample_period) {
533 hwc->sample_period = armpmu->max_period;
534 hwc->last_period = hwc->sample_period;
535 local64_set(&hwc->period_left, hwc->sample_period);
539 if (event->group_leader != event) {
540 err = validate_group(event);
548 static int armpmu_event_init(struct perf_event *event)
551 atomic_t *active_events = &armpmu->active_events;
553 switch (event->attr.type) {
555 case PERF_TYPE_HARDWARE:
556 case PERF_TYPE_HW_CACHE:
563 event->destroy = hw_perf_event_destroy;
565 if (!atomic_inc_not_zero(active_events)) {
566 mutex_lock(&armpmu->reserve_mutex);
567 if (atomic_read(active_events) == 0)
568 err = armpmu_reserve_hardware();
571 atomic_inc(active_events);
572 mutex_unlock(&armpmu->reserve_mutex);
578 err = __hw_perf_event_init(event);
580 hw_perf_event_destroy(event);
585 static void armpmu_enable(struct pmu *pmu)
587 /* Enable all of the perf events on hardware. */
588 int idx, enabled = 0;
589 struct cpu_hw_events *cpuc = armpmu->get_hw_events();
591 for (idx = 0; idx < armpmu->num_events; ++idx) {
592 struct perf_event *event = cpuc->events[idx];
597 armpmu->enable(&event->hw, idx);
605 static void armpmu_disable(struct pmu *pmu)
610 static struct pmu pmu = {
611 .pmu_enable = armpmu_enable,
612 .pmu_disable = armpmu_disable,
613 .event_init = armpmu_event_init,
616 .start = armpmu_start,
621 static void __init armpmu_init(struct arm_pmu *armpmu)
623 atomic_set(&armpmu->active_events, 0);
624 mutex_init(&armpmu->reserve_mutex);
627 /* Include the PMU-specific implementations. */
628 #include "perf_event_xscale.c"
629 #include "perf_event_v6.c"
630 #include "perf_event_v7.c"
633 * Ensure the PMU has sane values out of reset.
634 * This requires SMP to be available, so exists as a separate initcall.
639 if (armpmu && armpmu->reset)
640 return on_each_cpu(armpmu->reset, NULL, 1);
643 arch_initcall(armpmu_reset);
646 * PMU platform driver and devicetree bindings.
648 static struct of_device_id armpmu_of_device_ids[] = {
649 {.compatible = "arm,cortex-a9-pmu"},
650 {.compatible = "arm,cortex-a8-pmu"},
651 {.compatible = "arm,arm1136-pmu"},
652 {.compatible = "arm,arm1176-pmu"},
656 static struct platform_device_id armpmu_plat_device_ids[] = {
661 static int __devinit armpmu_device_probe(struct platform_device *pdev)
663 armpmu->plat_device = pdev;
667 static struct platform_driver armpmu_driver = {
670 .of_match_table = armpmu_of_device_ids,
672 .probe = armpmu_device_probe,
673 .id_table = armpmu_plat_device_ids,
676 static int __init register_pmu_driver(void)
678 return platform_driver_register(&armpmu_driver);
680 device_initcall(register_pmu_driver);
682 static struct cpu_hw_events *armpmu_get_cpu_events(void)
684 return &__get_cpu_var(cpu_hw_events);
687 static void __init cpu_pmu_init(struct arm_pmu *armpmu)
689 armpmu->get_hw_events = armpmu_get_cpu_events;
693 * CPU PMU identification and registration.
696 init_hw_perf_events(void)
698 unsigned long cpuid = read_cpuid_id();
699 unsigned long implementor = (cpuid & 0xFF000000) >> 24;
700 unsigned long part_number = (cpuid & 0xFFF0);
703 if (0x41 == implementor) {
704 switch (part_number) {
705 case 0xB360: /* ARM1136 */
706 case 0xB560: /* ARM1156 */
707 case 0xB760: /* ARM1176 */
708 armpmu = armv6pmu_init();
710 case 0xB020: /* ARM11mpcore */
711 armpmu = armv6mpcore_pmu_init();
713 case 0xC080: /* Cortex-A8 */
714 armpmu = armv7_a8_pmu_init();
716 case 0xC090: /* Cortex-A9 */
717 armpmu = armv7_a9_pmu_init();
719 case 0xC050: /* Cortex-A5 */
720 armpmu = armv7_a5_pmu_init();
722 case 0xC0F0: /* Cortex-A15 */
723 armpmu = armv7_a15_pmu_init();
726 /* Intel CPUs [xscale]. */
727 } else if (0x69 == implementor) {
728 part_number = (cpuid >> 13) & 0x7;
729 switch (part_number) {
731 armpmu = xscale1pmu_init();
734 armpmu = xscale2pmu_init();
740 pr_info("enabled with %s PMU driver, %d counters available\n",
741 armpmu->name, armpmu->num_events);
742 cpu_pmu_init(armpmu);
744 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
746 pr_info("no hardware support available\n");
751 early_initcall(init_hw_perf_events);
754 * Callchain handling code.
758 * The registers we're interested in are at the end of the variable
759 * length saved register structure. The fp points at the end of this
760 * structure so the address of this struct is:
761 * (struct frame_tail *)(xxx->fp)-1
763 * This code has been adapted from the ARM OProfile support.
766 struct frame_tail __user *fp;
769 } __attribute__((packed));
772 * Get the return address for a single stackframe and return a pointer to the
775 static struct frame_tail __user *
776 user_backtrace(struct frame_tail __user *tail,
777 struct perf_callchain_entry *entry)
779 struct frame_tail buftail;
781 /* Also check accessibility of one struct frame_tail beyond */
782 if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
784 if (__copy_from_user_inatomic(&buftail, tail, sizeof(buftail)))
787 perf_callchain_store(entry, buftail.lr);
790 * Frame pointers should strictly progress back up the stack
791 * (towards higher addresses).
793 if (tail + 1 >= buftail.fp)
796 return buftail.fp - 1;
800 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
802 struct frame_tail __user *tail;
805 tail = (struct frame_tail __user *)regs->ARM_fp - 1;
807 while ((entry->nr < PERF_MAX_STACK_DEPTH) &&
808 tail && !((unsigned long)tail & 0x3))
809 tail = user_backtrace(tail, entry);
813 * Gets called by walk_stackframe() for every stackframe. This will be called
814 * whist unwinding the stackframe and is like a subroutine return so we use
818 callchain_trace(struct stackframe *fr,
821 struct perf_callchain_entry *entry = data;
822 perf_callchain_store(entry, fr->pc);
827 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
829 struct stackframe fr;
831 fr.fp = regs->ARM_fp;
832 fr.sp = regs->ARM_sp;
833 fr.lr = regs->ARM_lr;
834 fr.pc = regs->ARM_pc;
835 walk_stackframe(&fr, callchain_trace, entry);