2 * Linux performance counter support for MIPS.
4 * Copyright (C) 2010 MIPS Technologies, Inc.
5 * Author: Deng-Cheng Zhu
7 * This code is based on the implementation for ARM, which is in turn
8 * based on the sparc64 perf event code and the x86 code. Performance
9 * counter access is based on the MIPS Oprofile code. And the callchain
10 * support references the code of MIPS stacktrace.c.
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License version 2 as
14 * published by the Free Software Foundation.
17 #include <linux/cpumask.h>
18 #include <linux/interrupt.h>
19 #include <linux/smp.h>
20 #include <linux/kernel.h>
21 #include <linux/perf_event.h>
22 #include <linux/uaccess.h>
25 #include <asm/irq_regs.h>
26 #include <asm/stacktrace.h>
27 #include <asm/time.h> /* For perf_irq */
29 /* These are for 32bit counters. For 64bit ones, define them accordingly. */
30 #define MAX_PERIOD ((1ULL << 32) - 1)
31 #define VALID_COUNT 0x7fffffff
33 #define HIGHEST_BIT 31
35 #define MIPS_MAX_HWEVENTS 4
37 struct cpu_hw_events {
38 /* Array of events on this cpu. */
39 struct perf_event *events[MIPS_MAX_HWEVENTS];
42 * Set the bit (indexed by the counter number) when the counter
43 * is used for an event.
45 unsigned long used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
48 * The borrowed MSB for the performance counter. A MIPS performance
49 * counter uses its bit 31 (for 32bit counters) or bit 63 (for 64bit
50 * counters) as a factor of determining whether a counter overflow
51 * should be signaled. So here we use a separate MSB for each
52 * counter to make things easy.
54 unsigned long msbs[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
57 * Software copy of the control register for each performance counter.
58 * MIPS CPUs vary in performance counters. They use this differently,
59 * and even may not use it.
61 unsigned int saved_ctrl[MIPS_MAX_HWEVENTS];
63 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
67 /* The description of MIPS performance events. */
68 struct mips_perf_event {
69 unsigned int event_id;
71 * MIPS performance counters are indexed starting from 0.
72 * CNTR_EVEN indicates the indexes of the counters to be used are
75 unsigned int cntr_mask;
76 #define CNTR_EVEN 0x55555555
77 #define CNTR_ODD 0xaaaaaaaa
78 #ifdef CONFIG_MIPS_MT_SMP
91 #define UNSUPPORTED_PERF_EVENT_ID 0xffffffff
92 #define C(x) PERF_COUNT_HW_CACHE_##x
97 irqreturn_t (*handle_irq)(int irq, void *dev);
98 int (*handle_shared_irq)(void);
101 int (*alloc_counter)(struct cpu_hw_events *cpuc,
102 struct hw_perf_event *hwc);
103 u64 (*read_counter)(unsigned int idx);
104 void (*write_counter)(unsigned int idx, u64 val);
105 void (*enable_event)(struct hw_perf_event *evt, int idx);
106 void (*disable_event)(int idx);
107 const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
108 const struct mips_perf_event (*cache_event_map)
109 [PERF_COUNT_HW_CACHE_MAX]
110 [PERF_COUNT_HW_CACHE_OP_MAX]
111 [PERF_COUNT_HW_CACHE_RESULT_MAX];
112 unsigned int num_counters;
115 static const struct mips_pmu *mipspmu;
118 mipspmu_event_set_period(struct perf_event *event,
119 struct hw_perf_event *hwc,
122 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
123 s64 left = local64_read(&hwc->period_left);
124 s64 period = hwc->sample_period;
129 if (unlikely(left <= -period)) {
131 local64_set(&hwc->period_left, left);
132 hwc->last_period = period;
136 if (unlikely(left <= 0)) {
138 local64_set(&hwc->period_left, left);
139 hwc->last_period = period;
143 if (left > (s64)MAX_PERIOD)
146 local64_set(&hwc->prev_count, (u64)-left);
148 local_irq_save(flags);
149 uleft = (u64)(-left) & MAX_PERIOD;
150 uleft > VALID_COUNT ?
151 set_bit(idx, cpuc->msbs) : clear_bit(idx, cpuc->msbs);
152 mipspmu->write_counter(idx, (u64)(-left) & VALID_COUNT);
153 local_irq_restore(flags);
155 perf_event_update_userpage(event);
160 static int mipspmu_enable(struct perf_event *event)
162 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
163 struct hw_perf_event *hwc = &event->hw;
167 /* To look for a free counter for this event. */
168 idx = mipspmu->alloc_counter(cpuc, hwc);
175 * If there is an event in the counter we are going to use then
176 * make sure it is disabled.
179 mipspmu->disable_event(idx);
180 cpuc->events[idx] = event;
182 /* Set the period for the event. */
183 mipspmu_event_set_period(event, hwc, idx);
185 /* Enable the event. */
186 mipspmu->enable_event(hwc, idx);
188 /* Propagate our changes to the userspace mapping. */
189 perf_event_update_userpage(event);
195 static void mipspmu_event_update(struct perf_event *event,
196 struct hw_perf_event *hwc,
199 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
201 int shift = 64 - TOTAL_BITS;
202 s64 prev_raw_count, new_raw_count;
206 prev_raw_count = local64_read(&hwc->prev_count);
207 local_irq_save(flags);
208 /* Make the counter value be a "real" one. */
209 new_raw_count = mipspmu->read_counter(idx);
210 if (new_raw_count & (test_bit(idx, cpuc->msbs) << HIGHEST_BIT)) {
211 new_raw_count &= VALID_COUNT;
212 clear_bit(idx, cpuc->msbs);
214 new_raw_count |= (test_bit(idx, cpuc->msbs) << HIGHEST_BIT);
215 local_irq_restore(flags);
217 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
218 new_raw_count) != prev_raw_count)
221 delta = (new_raw_count << shift) - (prev_raw_count << shift);
224 local64_add(delta, &event->count);
225 local64_sub(delta, &hwc->period_left);
230 static void mipspmu_disable(struct perf_event *event)
232 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
233 struct hw_perf_event *hwc = &event->hw;
237 WARN_ON(idx < 0 || idx >= mipspmu->num_counters);
239 /* We are working on a local event. */
240 mipspmu->disable_event(idx);
244 mipspmu_event_update(event, hwc, idx);
245 cpuc->events[idx] = NULL;
246 clear_bit(idx, cpuc->used_mask);
248 perf_event_update_userpage(event);
251 static void mipspmu_unthrottle(struct perf_event *event)
253 struct hw_perf_event *hwc = &event->hw;
255 mipspmu->enable_event(hwc, hwc->idx);
258 static void mipspmu_read(struct perf_event *event)
260 struct hw_perf_event *hwc = &event->hw;
262 /* Don't read disabled counters! */
266 mipspmu_event_update(event, hwc, hwc->idx);
269 static struct pmu pmu = {
270 .enable = mipspmu_enable,
271 .disable = mipspmu_disable,
272 .unthrottle = mipspmu_unthrottle,
273 .read = mipspmu_read,
276 static atomic_t active_events = ATOMIC_INIT(0);
277 static DEFINE_MUTEX(pmu_reserve_mutex);
278 static int (*save_perf_irq)(void);
280 static int mipspmu_get_irq(void)
284 if (mipspmu->irq >= 0) {
285 /* Request my own irq handler. */
286 err = request_irq(mipspmu->irq, mipspmu->handle_irq,
287 IRQF_DISABLED | IRQF_NOBALANCING,
288 "mips_perf_pmu", NULL);
290 pr_warning("Unable to request IRQ%d for MIPS "
291 "performance counters!\n", mipspmu->irq);
293 } else if (cp0_perfcount_irq < 0) {
295 * We are sharing the irq number with the timer interrupt.
297 save_perf_irq = perf_irq;
298 perf_irq = mipspmu->handle_shared_irq;
301 pr_warning("The platform hasn't properly defined its "
302 "interrupt controller.\n");
309 static void mipspmu_free_irq(void)
311 if (mipspmu->irq >= 0)
312 free_irq(mipspmu->irq, NULL);
313 else if (cp0_perfcount_irq < 0)
314 perf_irq = save_perf_irq;
317 static inline unsigned int
318 mipspmu_perf_event_encode(const struct mips_perf_event *pev)
321 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
324 #ifdef CONFIG_MIPS_MT_SMP
325 return ((unsigned int)pev->range << 24) |
326 (pev->cntr_mask & 0xffff00) |
327 (pev->event_id & 0xff);
329 return (pev->cntr_mask & 0xffff00) |
330 (pev->event_id & 0xff);
334 static const struct mips_perf_event *
335 mipspmu_map_general_event(int idx)
337 const struct mips_perf_event *pev;
339 pev = ((*mipspmu->general_event_map)[idx].event_id ==
340 UNSUPPORTED_PERF_EVENT_ID ? ERR_PTR(-EOPNOTSUPP) :
341 &(*mipspmu->general_event_map)[idx]);
346 static const struct mips_perf_event *
347 mipspmu_map_cache_event(u64 config)
349 unsigned int cache_type, cache_op, cache_result;
350 const struct mips_perf_event *pev;
352 cache_type = (config >> 0) & 0xff;
353 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
354 return ERR_PTR(-EINVAL);
356 cache_op = (config >> 8) & 0xff;
357 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
358 return ERR_PTR(-EINVAL);
360 cache_result = (config >> 16) & 0xff;
361 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
362 return ERR_PTR(-EINVAL);
364 pev = &((*mipspmu->cache_event_map)
369 if (pev->event_id == UNSUPPORTED_PERF_EVENT_ID)
370 return ERR_PTR(-EOPNOTSUPP);
376 static int validate_event(struct cpu_hw_events *cpuc,
377 struct perf_event *event)
379 struct hw_perf_event fake_hwc = event->hw;
381 if (event->pmu && event->pmu != &pmu)
384 return mipspmu->alloc_counter(cpuc, &fake_hwc) >= 0;
387 static int validate_group(struct perf_event *event)
389 struct perf_event *sibling, *leader = event->group_leader;
390 struct cpu_hw_events fake_cpuc;
392 memset(&fake_cpuc, 0, sizeof(fake_cpuc));
394 if (!validate_event(&fake_cpuc, leader))
397 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
398 if (!validate_event(&fake_cpuc, sibling))
402 if (!validate_event(&fake_cpuc, event))
409 * mipsxx/rm9000/loongson2 have different performance counters, they have
410 * specific low-level init routines.
412 static int __hw_perf_event_init(struct perf_event *event);
414 static void hw_perf_event_destroy(struct perf_event *event)
416 if (atomic_dec_and_mutex_lock(&active_events,
417 &pmu_reserve_mutex)) {
419 * We must not call the destroy function with interrupts
422 on_each_cpu(reset_counters,
423 (void *)(long)mipspmu->num_counters, 1);
425 mutex_unlock(&pmu_reserve_mutex);
429 const struct pmu *hw_perf_event_init(struct perf_event *event)
433 if (!mipspmu || event->cpu >= nr_cpumask_bits ||
434 (event->cpu >= 0 && !cpu_online(event->cpu)))
435 return ERR_PTR(-ENODEV);
437 if (!atomic_inc_not_zero(&active_events)) {
438 if (atomic_read(&active_events) > MIPS_MAX_HWEVENTS) {
439 atomic_dec(&active_events);
440 return ERR_PTR(-ENOSPC);
443 mutex_lock(&pmu_reserve_mutex);
444 if (atomic_read(&active_events) == 0)
445 err = mipspmu_get_irq();
448 atomic_inc(&active_events);
449 mutex_unlock(&pmu_reserve_mutex);
455 err = __hw_perf_event_init(event);
457 hw_perf_event_destroy(event);
459 return err ? ERR_PTR(err) : &pmu;
462 void hw_perf_enable(void)
468 void hw_perf_disable(void)
474 /* This is needed by specific irq handlers in perf_event_*.c */
476 handle_associated_event(struct cpu_hw_events *cpuc,
477 int idx, struct perf_sample_data *data, struct pt_regs *regs)
479 struct perf_event *event = cpuc->events[idx];
480 struct hw_perf_event *hwc = &event->hw;
482 mipspmu_event_update(event, hwc, idx);
483 data->period = event->hw.last_period;
484 if (!mipspmu_event_set_period(event, hwc, idx))
487 if (perf_event_overflow(event, 0, data, regs))
488 mipspmu->disable_event(idx);
491 /* Callchain handling code. */
493 callchain_store(struct perf_callchain_entry *entry,
496 if (entry->nr < PERF_MAX_STACK_DEPTH)
497 entry->ip[entry->nr++] = ip;
501 * Leave userspace callchain empty for now. When we find a way to trace
502 * the user stack callchains, we add here.
505 perf_callchain_user(struct pt_regs *regs,
506 struct perf_callchain_entry *entry)
510 static void save_raw_perf_callchain(struct perf_callchain_entry *entry,
513 unsigned long *sp = (unsigned long *)reg29;
516 while (!kstack_end(sp)) {
518 if (__kernel_text_address(addr)) {
519 callchain_store(entry, addr);
520 if (entry->nr >= PERF_MAX_STACK_DEPTH)
527 perf_callchain_kernel(struct pt_regs *regs,
528 struct perf_callchain_entry *entry)
530 unsigned long sp = regs->regs[29];
531 #ifdef CONFIG_KALLSYMS
532 unsigned long ra = regs->regs[31];
533 unsigned long pc = regs->cp0_epc;
535 callchain_store(entry, PERF_CONTEXT_KERNEL);
536 if (raw_show_trace || !__kernel_text_address(pc)) {
537 unsigned long stack_page =
538 (unsigned long)task_stack_page(current);
539 if (stack_page && sp >= stack_page &&
540 sp <= stack_page + THREAD_SIZE - 32)
541 save_raw_perf_callchain(entry, sp);
545 callchain_store(entry, pc);
546 if (entry->nr >= PERF_MAX_STACK_DEPTH)
548 pc = unwind_stack(current, &sp, pc, &ra);
551 callchain_store(entry, PERF_CONTEXT_KERNEL);
552 save_raw_perf_callchain(entry, sp);
557 perf_do_callchain(struct pt_regs *regs,
558 struct perf_callchain_entry *entry)
565 is_user = user_mode(regs);
567 if (!current || !current->pid)
570 if (is_user && current->state != TASK_RUNNING)
574 perf_callchain_kernel(regs, entry);
576 regs = task_pt_regs(current);
581 perf_callchain_user(regs, entry);
584 static DEFINE_PER_CPU(struct perf_callchain_entry, pmc_irq_entry);
586 struct perf_callchain_entry *
587 perf_callchain(struct pt_regs *regs)
589 struct perf_callchain_entry *entry = &__get_cpu_var(pmc_irq_entry);
592 perf_do_callchain(regs, entry);