treewide: remove redundant IS_ERR() before error code check
[linux/fpc-iii.git] / arch / mips / kernel / perf_event_mipsxx.c
blob128fc9999c564f4c1b31ef63d4ef702f146193ba
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Linux performance counter support for MIPS.
5 * Copyright (C) 2010 MIPS Technologies, Inc.
6 * Copyright (C) 2011 Cavium Networks, Inc.
7 * Author: Deng-Cheng Zhu
9 * This code is based on the implementation for ARM, which is in turn
10 * based on the sparc64 perf event code and the x86 code. Performance
11 * counter access is based on the MIPS Oprofile code. And the callchain
12 * support references the code of MIPS stacktrace.c.
15 #include <linux/cpumask.h>
16 #include <linux/interrupt.h>
17 #include <linux/smp.h>
18 #include <linux/kernel.h>
19 #include <linux/perf_event.h>
20 #include <linux/uaccess.h>
22 #include <asm/irq.h>
23 #include <asm/irq_regs.h>
24 #include <asm/stacktrace.h>
25 #include <asm/time.h> /* For perf_irq */
27 #define MIPS_MAX_HWEVENTS 4
28 #define MIPS_TCS_PER_COUNTER 2
29 #define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1)
31 struct cpu_hw_events {
32 /* Array of events on this cpu. */
33 struct perf_event *events[MIPS_MAX_HWEVENTS];
36 * Set the bit (indexed by the counter number) when the counter
37 * is used for an event.
39 unsigned long used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
42 * Software copy of the control register for each performance counter.
43 * MIPS CPUs vary in performance counters. They use this differently,
44 * and even may not use it.
46 unsigned int saved_ctrl[MIPS_MAX_HWEVENTS];
48 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
49 .saved_ctrl = {0},
52 /* The description of MIPS performance events. */
53 struct mips_perf_event {
54 unsigned int event_id;
56 * MIPS performance counters are indexed starting from 0.
57 * CNTR_EVEN indicates the indexes of the counters to be used are
58 * even numbers.
60 unsigned int cntr_mask;
61 #define CNTR_EVEN 0x55555555
62 #define CNTR_ODD 0xaaaaaaaa
63 #define CNTR_ALL 0xffffffff
64 enum {
65 T = 0,
66 V = 1,
67 P = 2,
68 } range;
71 static struct mips_perf_event raw_event;
72 static DEFINE_MUTEX(raw_event_mutex);
74 #define C(x) PERF_COUNT_HW_CACHE_##x
76 struct mips_pmu {
77 u64 max_period;
78 u64 valid_count;
79 u64 overflow;
80 const char *name;
81 int irq;
82 u64 (*read_counter)(unsigned int idx);
83 void (*write_counter)(unsigned int idx, u64 val);
84 const struct mips_perf_event *(*map_raw_event)(u64 config);
85 const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
86 const struct mips_perf_event (*cache_event_map)
87 [PERF_COUNT_HW_CACHE_MAX]
88 [PERF_COUNT_HW_CACHE_OP_MAX]
89 [PERF_COUNT_HW_CACHE_RESULT_MAX];
90 unsigned int num_counters;
93 static struct mips_pmu mipspmu;
95 #define M_PERFCTL_EVENT(event) (((event) << MIPS_PERFCTRL_EVENT_S) & \
96 MIPS_PERFCTRL_EVENT)
97 #define M_PERFCTL_VPEID(vpe) ((vpe) << MIPS_PERFCTRL_VPEID_S)
99 #ifdef CONFIG_CPU_BMIPS5000
100 #define M_PERFCTL_MT_EN(filter) 0
101 #else /* !CONFIG_CPU_BMIPS5000 */
102 #define M_PERFCTL_MT_EN(filter) (filter)
103 #endif /* CONFIG_CPU_BMIPS5000 */
105 #define M_TC_EN_ALL M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_ALL)
106 #define M_TC_EN_VPE M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_VPE)
107 #define M_TC_EN_TC M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_TC)
109 #define M_PERFCTL_COUNT_EVENT_WHENEVER (MIPS_PERFCTRL_EXL | \
110 MIPS_PERFCTRL_K | \
111 MIPS_PERFCTRL_U | \
112 MIPS_PERFCTRL_S | \
113 MIPS_PERFCTRL_IE)
115 #ifdef CONFIG_MIPS_MT_SMP
116 #define M_PERFCTL_CONFIG_MASK 0x3fff801f
117 #else
118 #define M_PERFCTL_CONFIG_MASK 0x1f
119 #endif
122 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
123 static DEFINE_RWLOCK(pmuint_rwlock);
125 #if defined(CONFIG_CPU_BMIPS5000)
126 #define vpe_id() (cpu_has_mipsmt_pertccounters ? \
127 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK))
128 #else
129 #define vpe_id() (cpu_has_mipsmt_pertccounters ? \
130 0 : cpu_vpe_id(&current_cpu_data))
131 #endif
133 /* Copied from op_model_mipsxx.c */
134 static unsigned int vpe_shift(void)
136 if (num_possible_cpus() > 1)
137 return 1;
139 return 0;
142 static unsigned int counters_total_to_per_cpu(unsigned int counters)
144 return counters >> vpe_shift();
147 #else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
148 #define vpe_id() 0
150 #endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
152 static void resume_local_counters(void);
153 static void pause_local_counters(void);
154 static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
155 static int mipsxx_pmu_handle_shared_irq(void);
157 static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
159 if (vpe_id() == 1)
160 idx = (idx + 2) & 3;
161 return idx;
164 static u64 mipsxx_pmu_read_counter(unsigned int idx)
166 idx = mipsxx_pmu_swizzle_perf_idx(idx);
168 switch (idx) {
169 case 0:
171 * The counters are unsigned, we must cast to truncate
172 * off the high bits.
174 return (u32)read_c0_perfcntr0();
175 case 1:
176 return (u32)read_c0_perfcntr1();
177 case 2:
178 return (u32)read_c0_perfcntr2();
179 case 3:
180 return (u32)read_c0_perfcntr3();
181 default:
182 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
183 return 0;
187 static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
189 idx = mipsxx_pmu_swizzle_perf_idx(idx);
191 switch (idx) {
192 case 0:
193 return read_c0_perfcntr0_64();
194 case 1:
195 return read_c0_perfcntr1_64();
196 case 2:
197 return read_c0_perfcntr2_64();
198 case 3:
199 return read_c0_perfcntr3_64();
200 default:
201 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
202 return 0;
206 static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
208 idx = mipsxx_pmu_swizzle_perf_idx(idx);
210 switch (idx) {
211 case 0:
212 write_c0_perfcntr0(val);
213 return;
214 case 1:
215 write_c0_perfcntr1(val);
216 return;
217 case 2:
218 write_c0_perfcntr2(val);
219 return;
220 case 3:
221 write_c0_perfcntr3(val);
222 return;
226 static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
228 idx = mipsxx_pmu_swizzle_perf_idx(idx);
230 switch (idx) {
231 case 0:
232 write_c0_perfcntr0_64(val);
233 return;
234 case 1:
235 write_c0_perfcntr1_64(val);
236 return;
237 case 2:
238 write_c0_perfcntr2_64(val);
239 return;
240 case 3:
241 write_c0_perfcntr3_64(val);
242 return;
246 static unsigned int mipsxx_pmu_read_control(unsigned int idx)
248 idx = mipsxx_pmu_swizzle_perf_idx(idx);
250 switch (idx) {
251 case 0:
252 return read_c0_perfctrl0();
253 case 1:
254 return read_c0_perfctrl1();
255 case 2:
256 return read_c0_perfctrl2();
257 case 3:
258 return read_c0_perfctrl3();
259 default:
260 WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
261 return 0;
265 static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
267 idx = mipsxx_pmu_swizzle_perf_idx(idx);
269 switch (idx) {
270 case 0:
271 write_c0_perfctrl0(val);
272 return;
273 case 1:
274 write_c0_perfctrl1(val);
275 return;
276 case 2:
277 write_c0_perfctrl2(val);
278 return;
279 case 3:
280 write_c0_perfctrl3(val);
281 return;
285 static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
286 struct hw_perf_event *hwc)
288 int i;
291 * We only need to care the counter mask. The range has been
292 * checked definitely.
294 unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
296 for (i = mipspmu.num_counters - 1; i >= 0; i--) {
298 * Note that some MIPS perf events can be counted by both
299 * even and odd counters, wheresas many other are only by
300 * even _or_ odd counters. This introduces an issue that
301 * when the former kind of event takes the counter the
302 * latter kind of event wants to use, then the "counter
303 * allocation" for the latter event will fail. In fact if
304 * they can be dynamically swapped, they both feel happy.
305 * But here we leave this issue alone for now.
307 if (test_bit(i, &cntr_mask) &&
308 !test_and_set_bit(i, cpuc->used_mask))
309 return i;
312 return -EAGAIN;
315 static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
317 struct perf_event *event = container_of(evt, struct perf_event, hw);
318 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
319 unsigned int range = evt->event_base >> 24;
321 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
323 cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
324 (evt->config_base & M_PERFCTL_CONFIG_MASK) |
325 /* Make sure interrupt enabled. */
326 MIPS_PERFCTRL_IE;
328 if (IS_ENABLED(CONFIG_CPU_BMIPS5000)) {
329 /* enable the counter for the calling thread */
330 cpuc->saved_ctrl[idx] |=
331 (1 << (12 + vpe_id())) | BRCM_PERFCTRL_TC;
332 } else if (IS_ENABLED(CONFIG_MIPS_MT_SMP) && range > V) {
333 /* The counter is processor wide. Set it up to count all TCs. */
334 pr_debug("Enabling perf counter for all TCs\n");
335 cpuc->saved_ctrl[idx] |= M_TC_EN_ALL;
336 } else {
337 unsigned int cpu, ctrl;
340 * Set up the counter for a particular CPU when event->cpu is
341 * a valid CPU number. Otherwise set up the counter for the CPU
342 * scheduling this thread.
344 cpu = (event->cpu >= 0) ? event->cpu : smp_processor_id();
346 ctrl = M_PERFCTL_VPEID(cpu_vpe_id(&cpu_data[cpu]));
347 ctrl |= M_TC_EN_VPE;
348 cpuc->saved_ctrl[idx] |= ctrl;
349 pr_debug("Enabling perf counter for CPU%d\n", cpu);
352 * We do not actually let the counter run. Leave it until start().
356 static void mipsxx_pmu_disable_event(int idx)
358 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
359 unsigned long flags;
361 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
363 local_irq_save(flags);
364 cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
365 ~M_PERFCTL_COUNT_EVENT_WHENEVER;
366 mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
367 local_irq_restore(flags);
370 static int mipspmu_event_set_period(struct perf_event *event,
371 struct hw_perf_event *hwc,
372 int idx)
374 u64 left = local64_read(&hwc->period_left);
375 u64 period = hwc->sample_period;
376 int ret = 0;
378 if (unlikely((left + period) & (1ULL << 63))) {
379 /* left underflowed by more than period. */
380 left = period;
381 local64_set(&hwc->period_left, left);
382 hwc->last_period = period;
383 ret = 1;
384 } else if (unlikely((left + period) <= period)) {
385 /* left underflowed by less than period. */
386 left += period;
387 local64_set(&hwc->period_left, left);
388 hwc->last_period = period;
389 ret = 1;
392 if (left > mipspmu.max_period) {
393 left = mipspmu.max_period;
394 local64_set(&hwc->period_left, left);
397 local64_set(&hwc->prev_count, mipspmu.overflow - left);
399 mipspmu.write_counter(idx, mipspmu.overflow - left);
401 perf_event_update_userpage(event);
403 return ret;
406 static void mipspmu_event_update(struct perf_event *event,
407 struct hw_perf_event *hwc,
408 int idx)
410 u64 prev_raw_count, new_raw_count;
411 u64 delta;
413 again:
414 prev_raw_count = local64_read(&hwc->prev_count);
415 new_raw_count = mipspmu.read_counter(idx);
417 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
418 new_raw_count) != prev_raw_count)
419 goto again;
421 delta = new_raw_count - prev_raw_count;
423 local64_add(delta, &event->count);
424 local64_sub(delta, &hwc->period_left);
427 static void mipspmu_start(struct perf_event *event, int flags)
429 struct hw_perf_event *hwc = &event->hw;
431 if (flags & PERF_EF_RELOAD)
432 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
434 hwc->state = 0;
436 /* Set the period for the event. */
437 mipspmu_event_set_period(event, hwc, hwc->idx);
439 /* Enable the event. */
440 mipsxx_pmu_enable_event(hwc, hwc->idx);
443 static void mipspmu_stop(struct perf_event *event, int flags)
445 struct hw_perf_event *hwc = &event->hw;
447 if (!(hwc->state & PERF_HES_STOPPED)) {
448 /* We are working on a local event. */
449 mipsxx_pmu_disable_event(hwc->idx);
450 barrier();
451 mipspmu_event_update(event, hwc, hwc->idx);
452 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
456 static int mipspmu_add(struct perf_event *event, int flags)
458 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
459 struct hw_perf_event *hwc = &event->hw;
460 int idx;
461 int err = 0;
463 perf_pmu_disable(event->pmu);
465 /* To look for a free counter for this event. */
466 idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
467 if (idx < 0) {
468 err = idx;
469 goto out;
473 * If there is an event in the counter we are going to use then
474 * make sure it is disabled.
476 event->hw.idx = idx;
477 mipsxx_pmu_disable_event(idx);
478 cpuc->events[idx] = event;
480 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
481 if (flags & PERF_EF_START)
482 mipspmu_start(event, PERF_EF_RELOAD);
484 /* Propagate our changes to the userspace mapping. */
485 perf_event_update_userpage(event);
487 out:
488 perf_pmu_enable(event->pmu);
489 return err;
492 static void mipspmu_del(struct perf_event *event, int flags)
494 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
495 struct hw_perf_event *hwc = &event->hw;
496 int idx = hwc->idx;
498 WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
500 mipspmu_stop(event, PERF_EF_UPDATE);
501 cpuc->events[idx] = NULL;
502 clear_bit(idx, cpuc->used_mask);
504 perf_event_update_userpage(event);
507 static void mipspmu_read(struct perf_event *event)
509 struct hw_perf_event *hwc = &event->hw;
511 /* Don't read disabled counters! */
512 if (hwc->idx < 0)
513 return;
515 mipspmu_event_update(event, hwc, hwc->idx);
518 static void mipspmu_enable(struct pmu *pmu)
520 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
521 write_unlock(&pmuint_rwlock);
522 #endif
523 resume_local_counters();
527 * MIPS performance counters can be per-TC. The control registers can
528 * not be directly accessed across CPUs. Hence if we want to do global
529 * control, we need cross CPU calls. on_each_cpu() can help us, but we
530 * can not make sure this function is called with interrupts enabled. So
531 * here we pause local counters and then grab a rwlock and leave the
532 * counters on other CPUs alone. If any counter interrupt raises while
533 * we own the write lock, simply pause local counters on that CPU and
534 * spin in the handler. Also we know we won't be switched to another
535 * CPU after pausing local counters and before grabbing the lock.
537 static void mipspmu_disable(struct pmu *pmu)
539 pause_local_counters();
540 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
541 write_lock(&pmuint_rwlock);
542 #endif
545 static atomic_t active_events = ATOMIC_INIT(0);
546 static DEFINE_MUTEX(pmu_reserve_mutex);
547 static int (*save_perf_irq)(void);
549 static int mipspmu_get_irq(void)
551 int err;
553 if (mipspmu.irq >= 0) {
554 /* Request my own irq handler. */
555 err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
556 IRQF_PERCPU | IRQF_NOBALANCING |
557 IRQF_NO_THREAD | IRQF_NO_SUSPEND |
558 IRQF_SHARED,
559 "mips_perf_pmu", &mipspmu);
560 if (err) {
561 pr_warn("Unable to request IRQ%d for MIPS performance counters!\n",
562 mipspmu.irq);
564 } else if (cp0_perfcount_irq < 0) {
566 * We are sharing the irq number with the timer interrupt.
568 save_perf_irq = perf_irq;
569 perf_irq = mipsxx_pmu_handle_shared_irq;
570 err = 0;
571 } else {
572 pr_warn("The platform hasn't properly defined its interrupt controller\n");
573 err = -ENOENT;
576 return err;
579 static void mipspmu_free_irq(void)
581 if (mipspmu.irq >= 0)
582 free_irq(mipspmu.irq, &mipspmu);
583 else if (cp0_perfcount_irq < 0)
584 perf_irq = save_perf_irq;
588 * mipsxx/rm9000/loongson2 have different performance counters, they have
589 * specific low-level init routines.
591 static void reset_counters(void *arg);
592 static int __hw_perf_event_init(struct perf_event *event);
594 static void hw_perf_event_destroy(struct perf_event *event)
596 if (atomic_dec_and_mutex_lock(&active_events,
597 &pmu_reserve_mutex)) {
599 * We must not call the destroy function with interrupts
600 * disabled.
602 on_each_cpu(reset_counters,
603 (void *)(long)mipspmu.num_counters, 1);
604 mipspmu_free_irq();
605 mutex_unlock(&pmu_reserve_mutex);
609 static int mipspmu_event_init(struct perf_event *event)
611 int err = 0;
613 /* does not support taken branch sampling */
614 if (has_branch_stack(event))
615 return -EOPNOTSUPP;
617 switch (event->attr.type) {
618 case PERF_TYPE_RAW:
619 case PERF_TYPE_HARDWARE:
620 case PERF_TYPE_HW_CACHE:
621 break;
623 default:
624 return -ENOENT;
627 if (event->cpu >= 0 && !cpu_online(event->cpu))
628 return -ENODEV;
630 if (!atomic_inc_not_zero(&active_events)) {
631 mutex_lock(&pmu_reserve_mutex);
632 if (atomic_read(&active_events) == 0)
633 err = mipspmu_get_irq();
635 if (!err)
636 atomic_inc(&active_events);
637 mutex_unlock(&pmu_reserve_mutex);
640 if (err)
641 return err;
643 return __hw_perf_event_init(event);
646 static struct pmu pmu = {
647 .pmu_enable = mipspmu_enable,
648 .pmu_disable = mipspmu_disable,
649 .event_init = mipspmu_event_init,
650 .add = mipspmu_add,
651 .del = mipspmu_del,
652 .start = mipspmu_start,
653 .stop = mipspmu_stop,
654 .read = mipspmu_read,
657 static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
660 * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
661 * event_id.
663 #ifdef CONFIG_MIPS_MT_SMP
664 if (num_possible_cpus() > 1)
665 return ((unsigned int)pev->range << 24) |
666 (pev->cntr_mask & 0xffff00) |
667 (pev->event_id & 0xff);
668 else
669 #endif /* CONFIG_MIPS_MT_SMP */
670 return ((pev->cntr_mask & 0xffff00) |
671 (pev->event_id & 0xff));
674 static const struct mips_perf_event *mipspmu_map_general_event(int idx)
677 if ((*mipspmu.general_event_map)[idx].cntr_mask == 0)
678 return ERR_PTR(-EOPNOTSUPP);
679 return &(*mipspmu.general_event_map)[idx];
682 static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
684 unsigned int cache_type, cache_op, cache_result;
685 const struct mips_perf_event *pev;
687 cache_type = (config >> 0) & 0xff;
688 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
689 return ERR_PTR(-EINVAL);
691 cache_op = (config >> 8) & 0xff;
692 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
693 return ERR_PTR(-EINVAL);
695 cache_result = (config >> 16) & 0xff;
696 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
697 return ERR_PTR(-EINVAL);
699 pev = &((*mipspmu.cache_event_map)
700 [cache_type]
701 [cache_op]
702 [cache_result]);
704 if (pev->cntr_mask == 0)
705 return ERR_PTR(-EOPNOTSUPP);
707 return pev;
711 static int validate_group(struct perf_event *event)
713 struct perf_event *sibling, *leader = event->group_leader;
714 struct cpu_hw_events fake_cpuc;
716 memset(&fake_cpuc, 0, sizeof(fake_cpuc));
718 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
719 return -EINVAL;
721 for_each_sibling_event(sibling, leader) {
722 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
723 return -EINVAL;
726 if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
727 return -EINVAL;
729 return 0;
732 /* This is needed by specific irq handlers in perf_event_*.c */
733 static void handle_associated_event(struct cpu_hw_events *cpuc,
734 int idx, struct perf_sample_data *data,
735 struct pt_regs *regs)
737 struct perf_event *event = cpuc->events[idx];
738 struct hw_perf_event *hwc = &event->hw;
740 mipspmu_event_update(event, hwc, idx);
741 data->period = event->hw.last_period;
742 if (!mipspmu_event_set_period(event, hwc, idx))
743 return;
745 if (perf_event_overflow(event, data, regs))
746 mipsxx_pmu_disable_event(idx);
750 static int __n_counters(void)
752 if (!cpu_has_perf)
753 return 0;
754 if (!(read_c0_perfctrl0() & MIPS_PERFCTRL_M))
755 return 1;
756 if (!(read_c0_perfctrl1() & MIPS_PERFCTRL_M))
757 return 2;
758 if (!(read_c0_perfctrl2() & MIPS_PERFCTRL_M))
759 return 3;
761 return 4;
764 static int n_counters(void)
766 int counters;
768 switch (current_cpu_type()) {
769 case CPU_R10000:
770 counters = 2;
771 break;
773 case CPU_R12000:
774 case CPU_R14000:
775 case CPU_R16000:
776 counters = 4;
777 break;
779 default:
780 counters = __n_counters();
783 return counters;
786 static void reset_counters(void *arg)
788 int counters = (int)(long)arg;
789 switch (counters) {
790 case 4:
791 mipsxx_pmu_write_control(3, 0);
792 mipspmu.write_counter(3, 0);
793 /* fall through */
794 case 3:
795 mipsxx_pmu_write_control(2, 0);
796 mipspmu.write_counter(2, 0);
797 /* fall through */
798 case 2:
799 mipsxx_pmu_write_control(1, 0);
800 mipspmu.write_counter(1, 0);
801 /* fall through */
802 case 1:
803 mipsxx_pmu_write_control(0, 0);
804 mipspmu.write_counter(0, 0);
805 /* fall through */
809 /* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */
810 static const struct mips_perf_event mipsxxcore_event_map
811 [PERF_COUNT_HW_MAX] = {
812 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
813 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
814 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
815 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
818 /* 74K/proAptiv core has different branch event code. */
819 static const struct mips_perf_event mipsxxcore_event_map2
820 [PERF_COUNT_HW_MAX] = {
821 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
822 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
823 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
824 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
827 static const struct mips_perf_event i6x00_event_map[PERF_COUNT_HW_MAX] = {
828 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD },
829 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD },
830 /* These only count dcache, not icache */
831 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x45, CNTR_EVEN | CNTR_ODD },
832 [PERF_COUNT_HW_CACHE_MISSES] = { 0x48, CNTR_EVEN | CNTR_ODD },
833 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x15, CNTR_EVEN | CNTR_ODD },
834 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x16, CNTR_EVEN | CNTR_ODD },
837 static const struct mips_perf_event loongson3_event_map[PERF_COUNT_HW_MAX] = {
838 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN },
839 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD },
840 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN },
841 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD },
844 static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
845 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
846 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
847 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
848 [PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL },
849 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
850 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
851 [PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
854 static const struct mips_perf_event bmips5000_event_map
855 [PERF_COUNT_HW_MAX] = {
856 [PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T },
857 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
858 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
861 static const struct mips_perf_event xlp_event_map[PERF_COUNT_HW_MAX] = {
862 [PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
863 [PERF_COUNT_HW_INSTRUCTIONS] = { 0x18, CNTR_ALL }, /* PAPI_TOT_INS */
864 [PERF_COUNT_HW_CACHE_REFERENCES] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
865 [PERF_COUNT_HW_CACHE_MISSES] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
866 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x1b, CNTR_ALL }, /* PAPI_BR_CN */
867 [PERF_COUNT_HW_BRANCH_MISSES] = { 0x1c, CNTR_ALL }, /* PAPI_BR_MSP */
870 /* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */
871 static const struct mips_perf_event mipsxxcore_cache_map
872 [PERF_COUNT_HW_CACHE_MAX]
873 [PERF_COUNT_HW_CACHE_OP_MAX]
874 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
875 [C(L1D)] = {
877 * Like some other architectures (e.g. ARM), the performance
878 * counters don't differentiate between read and write
879 * accesses/misses, so this isn't strictly correct, but it's the
880 * best we can do. Writes and reads get combined.
882 [C(OP_READ)] = {
883 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T },
884 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
886 [C(OP_WRITE)] = {
887 [C(RESULT_ACCESS)] = { 0x0a, CNTR_EVEN, T },
888 [C(RESULT_MISS)] = { 0x0b, CNTR_EVEN | CNTR_ODD, T },
891 [C(L1I)] = {
892 [C(OP_READ)] = {
893 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T },
894 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T },
896 [C(OP_WRITE)] = {
897 [C(RESULT_ACCESS)] = { 0x09, CNTR_EVEN, T },
898 [C(RESULT_MISS)] = { 0x09, CNTR_ODD, T },
900 [C(OP_PREFETCH)] = {
901 [C(RESULT_ACCESS)] = { 0x14, CNTR_EVEN, T },
903 * Note that MIPS has only "hit" events countable for
904 * the prefetch operation.
908 [C(LL)] = {
909 [C(OP_READ)] = {
910 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P },
911 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P },
913 [C(OP_WRITE)] = {
914 [C(RESULT_ACCESS)] = { 0x15, CNTR_ODD, P },
915 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN, P },
918 [C(DTLB)] = {
919 [C(OP_READ)] = {
920 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
921 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
923 [C(OP_WRITE)] = {
924 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
925 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
928 [C(ITLB)] = {
929 [C(OP_READ)] = {
930 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T },
931 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T },
933 [C(OP_WRITE)] = {
934 [C(RESULT_ACCESS)] = { 0x05, CNTR_EVEN, T },
935 [C(RESULT_MISS)] = { 0x05, CNTR_ODD, T },
938 [C(BPU)] = {
939 /* Using the same code for *HW_BRANCH* */
940 [C(OP_READ)] = {
941 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T },
942 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
944 [C(OP_WRITE)] = {
945 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN, T },
946 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
951 /* 74K/proAptiv core has completely different cache event map. */
952 static const struct mips_perf_event mipsxxcore_cache_map2
953 [PERF_COUNT_HW_CACHE_MAX]
954 [PERF_COUNT_HW_CACHE_OP_MAX]
955 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
956 [C(L1D)] = {
958 * Like some other architectures (e.g. ARM), the performance
959 * counters don't differentiate between read and write
960 * accesses/misses, so this isn't strictly correct, but it's the
961 * best we can do. Writes and reads get combined.
963 [C(OP_READ)] = {
964 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T },
965 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T },
967 [C(OP_WRITE)] = {
968 [C(RESULT_ACCESS)] = { 0x17, CNTR_ODD, T },
969 [C(RESULT_MISS)] = { 0x18, CNTR_ODD, T },
972 [C(L1I)] = {
973 [C(OP_READ)] = {
974 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
975 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
977 [C(OP_WRITE)] = {
978 [C(RESULT_ACCESS)] = { 0x06, CNTR_EVEN, T },
979 [C(RESULT_MISS)] = { 0x06, CNTR_ODD, T },
981 [C(OP_PREFETCH)] = {
982 [C(RESULT_ACCESS)] = { 0x34, CNTR_EVEN, T },
984 * Note that MIPS has only "hit" events countable for
985 * the prefetch operation.
989 [C(LL)] = {
990 [C(OP_READ)] = {
991 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P },
992 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P },
994 [C(OP_WRITE)] = {
995 [C(RESULT_ACCESS)] = { 0x1c, CNTR_ODD, P },
996 [C(RESULT_MISS)] = { 0x1d, CNTR_EVEN, P },
1000 * 74K core does not have specific DTLB events. proAptiv core has
1001 * "speculative" DTLB events which are numbered 0x63 (even/odd) and
1002 * not included here. One can use raw events if really needed.
1004 [C(ITLB)] = {
1005 [C(OP_READ)] = {
1006 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T },
1007 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T },
1009 [C(OP_WRITE)] = {
1010 [C(RESULT_ACCESS)] = { 0x04, CNTR_EVEN, T },
1011 [C(RESULT_MISS)] = { 0x04, CNTR_ODD, T },
1014 [C(BPU)] = {
1015 /* Using the same code for *HW_BRANCH* */
1016 [C(OP_READ)] = {
1017 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T },
1018 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T },
1020 [C(OP_WRITE)] = {
1021 [C(RESULT_ACCESS)] = { 0x27, CNTR_EVEN, T },
1022 [C(RESULT_MISS)] = { 0x27, CNTR_ODD, T },
1027 static const struct mips_perf_event i6x00_cache_map
1028 [PERF_COUNT_HW_CACHE_MAX]
1029 [PERF_COUNT_HW_CACHE_OP_MAX]
1030 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1031 [C(L1D)] = {
1032 [C(OP_READ)] = {
1033 [C(RESULT_ACCESS)] = { 0x46, CNTR_EVEN | CNTR_ODD },
1034 [C(RESULT_MISS)] = { 0x49, CNTR_EVEN | CNTR_ODD },
1036 [C(OP_WRITE)] = {
1037 [C(RESULT_ACCESS)] = { 0x47, CNTR_EVEN | CNTR_ODD },
1038 [C(RESULT_MISS)] = { 0x4a, CNTR_EVEN | CNTR_ODD },
1041 [C(L1I)] = {
1042 [C(OP_READ)] = {
1043 [C(RESULT_ACCESS)] = { 0x84, CNTR_EVEN | CNTR_ODD },
1044 [C(RESULT_MISS)] = { 0x85, CNTR_EVEN | CNTR_ODD },
1047 [C(DTLB)] = {
1048 /* Can't distinguish read & write */
1049 [C(OP_READ)] = {
1050 [C(RESULT_ACCESS)] = { 0x40, CNTR_EVEN | CNTR_ODD },
1051 [C(RESULT_MISS)] = { 0x41, CNTR_EVEN | CNTR_ODD },
1053 [C(OP_WRITE)] = {
1054 [C(RESULT_ACCESS)] = { 0x40, CNTR_EVEN | CNTR_ODD },
1055 [C(RESULT_MISS)] = { 0x41, CNTR_EVEN | CNTR_ODD },
1058 [C(BPU)] = {
1059 /* Conditional branches / mispredicted */
1060 [C(OP_READ)] = {
1061 [C(RESULT_ACCESS)] = { 0x15, CNTR_EVEN | CNTR_ODD },
1062 [C(RESULT_MISS)] = { 0x16, CNTR_EVEN | CNTR_ODD },
1067 static const struct mips_perf_event loongson3_cache_map
1068 [PERF_COUNT_HW_CACHE_MAX]
1069 [PERF_COUNT_HW_CACHE_OP_MAX]
1070 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1071 [C(L1D)] = {
1073 * Like some other architectures (e.g. ARM), the performance
1074 * counters don't differentiate between read and write
1075 * accesses/misses, so this isn't strictly correct, but it's the
1076 * best we can do. Writes and reads get combined.
1078 [C(OP_READ)] = {
1079 [C(RESULT_MISS)] = { 0x04, CNTR_ODD },
1081 [C(OP_WRITE)] = {
1082 [C(RESULT_MISS)] = { 0x04, CNTR_ODD },
1085 [C(L1I)] = {
1086 [C(OP_READ)] = {
1087 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN },
1089 [C(OP_WRITE)] = {
1090 [C(RESULT_MISS)] = { 0x04, CNTR_EVEN },
1093 [C(DTLB)] = {
1094 [C(OP_READ)] = {
1095 [C(RESULT_MISS)] = { 0x09, CNTR_ODD },
1097 [C(OP_WRITE)] = {
1098 [C(RESULT_MISS)] = { 0x09, CNTR_ODD },
1101 [C(ITLB)] = {
1102 [C(OP_READ)] = {
1103 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD },
1105 [C(OP_WRITE)] = {
1106 [C(RESULT_MISS)] = { 0x0c, CNTR_ODD },
1109 [C(BPU)] = {
1110 /* Using the same code for *HW_BRANCH* */
1111 [C(OP_READ)] = {
1112 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN },
1113 [C(RESULT_MISS)] = { 0x02, CNTR_ODD },
1115 [C(OP_WRITE)] = {
1116 [C(RESULT_ACCESS)] = { 0x02, CNTR_EVEN },
1117 [C(RESULT_MISS)] = { 0x02, CNTR_ODD },
1122 /* BMIPS5000 */
1123 static const struct mips_perf_event bmips5000_cache_map
1124 [PERF_COUNT_HW_CACHE_MAX]
1125 [PERF_COUNT_HW_CACHE_OP_MAX]
1126 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1127 [C(L1D)] = {
1129 * Like some other architectures (e.g. ARM), the performance
1130 * counters don't differentiate between read and write
1131 * accesses/misses, so this isn't strictly correct, but it's the
1132 * best we can do. Writes and reads get combined.
1134 [C(OP_READ)] = {
1135 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T },
1136 [C(RESULT_MISS)] = { 12, CNTR_ODD, T },
1138 [C(OP_WRITE)] = {
1139 [C(RESULT_ACCESS)] = { 12, CNTR_EVEN, T },
1140 [C(RESULT_MISS)] = { 12, CNTR_ODD, T },
1143 [C(L1I)] = {
1144 [C(OP_READ)] = {
1145 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T },
1146 [C(RESULT_MISS)] = { 10, CNTR_ODD, T },
1148 [C(OP_WRITE)] = {
1149 [C(RESULT_ACCESS)] = { 10, CNTR_EVEN, T },
1150 [C(RESULT_MISS)] = { 10, CNTR_ODD, T },
1152 [C(OP_PREFETCH)] = {
1153 [C(RESULT_ACCESS)] = { 23, CNTR_EVEN, T },
1155 * Note that MIPS has only "hit" events countable for
1156 * the prefetch operation.
1160 [C(LL)] = {
1161 [C(OP_READ)] = {
1162 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P },
1163 [C(RESULT_MISS)] = { 28, CNTR_ODD, P },
1165 [C(OP_WRITE)] = {
1166 [C(RESULT_ACCESS)] = { 28, CNTR_EVEN, P },
1167 [C(RESULT_MISS)] = { 28, CNTR_ODD, P },
1170 [C(BPU)] = {
1171 /* Using the same code for *HW_BRANCH* */
1172 [C(OP_READ)] = {
1173 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
1175 [C(OP_WRITE)] = {
1176 [C(RESULT_MISS)] = { 0x02, CNTR_ODD, T },
1182 static const struct mips_perf_event octeon_cache_map
1183 [PERF_COUNT_HW_CACHE_MAX]
1184 [PERF_COUNT_HW_CACHE_OP_MAX]
1185 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1186 [C(L1D)] = {
1187 [C(OP_READ)] = {
1188 [C(RESULT_ACCESS)] = { 0x2b, CNTR_ALL },
1189 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL },
1191 [C(OP_WRITE)] = {
1192 [C(RESULT_ACCESS)] = { 0x30, CNTR_ALL },
1195 [C(L1I)] = {
1196 [C(OP_READ)] = {
1197 [C(RESULT_ACCESS)] = { 0x18, CNTR_ALL },
1199 [C(OP_PREFETCH)] = {
1200 [C(RESULT_ACCESS)] = { 0x19, CNTR_ALL },
1203 [C(DTLB)] = {
1205 * Only general DTLB misses are counted use the same event for
1206 * read and write.
1208 [C(OP_READ)] = {
1209 [C(RESULT_MISS)] = { 0x35, CNTR_ALL },
1211 [C(OP_WRITE)] = {
1212 [C(RESULT_MISS)] = { 0x35, CNTR_ALL },
1215 [C(ITLB)] = {
1216 [C(OP_READ)] = {
1217 [C(RESULT_MISS)] = { 0x37, CNTR_ALL },
1222 static const struct mips_perf_event xlp_cache_map
1223 [PERF_COUNT_HW_CACHE_MAX]
1224 [PERF_COUNT_HW_CACHE_OP_MAX]
1225 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1226 [C(L1D)] = {
1227 [C(OP_READ)] = {
1228 [C(RESULT_ACCESS)] = { 0x31, CNTR_ALL }, /* PAPI_L1_DCR */
1229 [C(RESULT_MISS)] = { 0x30, CNTR_ALL }, /* PAPI_L1_LDM */
1231 [C(OP_WRITE)] = {
1232 [C(RESULT_ACCESS)] = { 0x2f, CNTR_ALL }, /* PAPI_L1_DCW */
1233 [C(RESULT_MISS)] = { 0x2e, CNTR_ALL }, /* PAPI_L1_STM */
1236 [C(L1I)] = {
1237 [C(OP_READ)] = {
1238 [C(RESULT_ACCESS)] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
1239 [C(RESULT_MISS)] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
1242 [C(LL)] = {
1243 [C(OP_READ)] = {
1244 [C(RESULT_ACCESS)] = { 0x35, CNTR_ALL }, /* PAPI_L2_DCR */
1245 [C(RESULT_MISS)] = { 0x37, CNTR_ALL }, /* PAPI_L2_LDM */
1247 [C(OP_WRITE)] = {
1248 [C(RESULT_ACCESS)] = { 0x34, CNTR_ALL }, /* PAPI_L2_DCA */
1249 [C(RESULT_MISS)] = { 0x36, CNTR_ALL }, /* PAPI_L2_DCM */
1252 [C(DTLB)] = {
1254 * Only general DTLB misses are counted use the same event for
1255 * read and write.
1257 [C(OP_READ)] = {
1258 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1260 [C(OP_WRITE)] = {
1261 [C(RESULT_MISS)] = { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1264 [C(ITLB)] = {
1265 [C(OP_READ)] = {
1266 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1268 [C(OP_WRITE)] = {
1269 [C(RESULT_MISS)] = { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1272 [C(BPU)] = {
1273 [C(OP_READ)] = {
1274 [C(RESULT_MISS)] = { 0x25, CNTR_ALL },
1279 static int __hw_perf_event_init(struct perf_event *event)
1281 struct perf_event_attr *attr = &event->attr;
1282 struct hw_perf_event *hwc = &event->hw;
1283 const struct mips_perf_event *pev;
1284 int err;
1286 /* Returning MIPS event descriptor for generic perf event. */
1287 if (PERF_TYPE_HARDWARE == event->attr.type) {
1288 if (event->attr.config >= PERF_COUNT_HW_MAX)
1289 return -EINVAL;
1290 pev = mipspmu_map_general_event(event->attr.config);
1291 } else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1292 pev = mipspmu_map_cache_event(event->attr.config);
1293 } else if (PERF_TYPE_RAW == event->attr.type) {
1294 /* We are working on the global raw event. */
1295 mutex_lock(&raw_event_mutex);
1296 pev = mipspmu.map_raw_event(event->attr.config);
1297 } else {
1298 /* The event type is not (yet) supported. */
1299 return -EOPNOTSUPP;
1302 if (IS_ERR(pev)) {
1303 if (PERF_TYPE_RAW == event->attr.type)
1304 mutex_unlock(&raw_event_mutex);
1305 return PTR_ERR(pev);
1309 * We allow max flexibility on how each individual counter shared
1310 * by the single CPU operates (the mode exclusion and the range).
1312 hwc->config_base = MIPS_PERFCTRL_IE;
1314 hwc->event_base = mipspmu_perf_event_encode(pev);
1315 if (PERF_TYPE_RAW == event->attr.type)
1316 mutex_unlock(&raw_event_mutex);
1318 if (!attr->exclude_user)
1319 hwc->config_base |= MIPS_PERFCTRL_U;
1320 if (!attr->exclude_kernel) {
1321 hwc->config_base |= MIPS_PERFCTRL_K;
1322 /* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1323 hwc->config_base |= MIPS_PERFCTRL_EXL;
1325 if (!attr->exclude_hv)
1326 hwc->config_base |= MIPS_PERFCTRL_S;
1328 hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1330 * The event can belong to another cpu. We do not assign a local
1331 * counter for it for now.
1333 hwc->idx = -1;
1334 hwc->config = 0;
1336 if (!hwc->sample_period) {
1337 hwc->sample_period = mipspmu.max_period;
1338 hwc->last_period = hwc->sample_period;
1339 local64_set(&hwc->period_left, hwc->sample_period);
1342 err = 0;
1343 if (event->group_leader != event)
1344 err = validate_group(event);
1346 event->destroy = hw_perf_event_destroy;
1348 if (err)
1349 event->destroy(event);
1351 return err;
1354 static void pause_local_counters(void)
1356 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1357 int ctr = mipspmu.num_counters;
1358 unsigned long flags;
1360 local_irq_save(flags);
1361 do {
1362 ctr--;
1363 cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1364 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1365 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1366 } while (ctr > 0);
1367 local_irq_restore(flags);
1370 static void resume_local_counters(void)
1372 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1373 int ctr = mipspmu.num_counters;
1375 do {
1376 ctr--;
1377 mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1378 } while (ctr > 0);
1381 static int mipsxx_pmu_handle_shared_irq(void)
1383 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1384 struct perf_sample_data data;
1385 unsigned int counters = mipspmu.num_counters;
1386 u64 counter;
1387 int n, handled = IRQ_NONE;
1388 struct pt_regs *regs;
1390 if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI))
1391 return handled;
1393 * First we pause the local counters, so that when we are locked
1394 * here, the counters are all paused. When it gets locked due to
1395 * perf_disable(), the timer interrupt handler will be delayed.
1397 * See also mipsxx_pmu_start().
1399 pause_local_counters();
1400 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1401 read_lock(&pmuint_rwlock);
1402 #endif
1404 regs = get_irq_regs();
1406 perf_sample_data_init(&data, 0, 0);
1408 for (n = counters - 1; n >= 0; n--) {
1409 if (!test_bit(n, cpuc->used_mask))
1410 continue;
1412 counter = mipspmu.read_counter(n);
1413 if (!(counter & mipspmu.overflow))
1414 continue;
1416 handle_associated_event(cpuc, n, &data, regs);
1417 handled = IRQ_HANDLED;
1420 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1421 read_unlock(&pmuint_rwlock);
1422 #endif
1423 resume_local_counters();
1426 * Do all the work for the pending perf events. We can do this
1427 * in here because the performance counter interrupt is a regular
1428 * interrupt, not NMI.
1430 if (handled == IRQ_HANDLED)
1431 irq_work_run();
1433 return handled;
1436 static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1438 return mipsxx_pmu_handle_shared_irq();
1441 /* 24K */
1442 #define IS_BOTH_COUNTERS_24K_EVENT(b) \
1443 ((b) == 0 || (b) == 1 || (b) == 11)
1445 /* 34K */
1446 #define IS_BOTH_COUNTERS_34K_EVENT(b) \
1447 ((b) == 0 || (b) == 1 || (b) == 11)
1448 #ifdef CONFIG_MIPS_MT_SMP
1449 #define IS_RANGE_P_34K_EVENT(r, b) \
1450 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1451 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 || \
1452 (r) == 176 || ((b) >= 50 && (b) <= 55) || \
1453 ((b) >= 64 && (b) <= 67))
1454 #define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
1455 #endif
1457 /* 74K */
1458 #define IS_BOTH_COUNTERS_74K_EVENT(b) \
1459 ((b) == 0 || (b) == 1)
1461 /* proAptiv */
1462 #define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b) \
1463 ((b) == 0 || (b) == 1)
1464 /* P5600 */
1465 #define IS_BOTH_COUNTERS_P5600_EVENT(b) \
1466 ((b) == 0 || (b) == 1)
1468 /* 1004K */
1469 #define IS_BOTH_COUNTERS_1004K_EVENT(b) \
1470 ((b) == 0 || (b) == 1 || (b) == 11)
1471 #ifdef CONFIG_MIPS_MT_SMP
1472 #define IS_RANGE_P_1004K_EVENT(r, b) \
1473 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1474 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 || \
1475 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) || \
1476 (r) == 188 || (b) == 61 || (b) == 62 || \
1477 ((b) >= 64 && (b) <= 67))
1478 #define IS_RANGE_V_1004K_EVENT(r) ((r) == 47)
1479 #endif
1481 /* interAptiv */
1482 #define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b) \
1483 ((b) == 0 || (b) == 1 || (b) == 11)
1484 #ifdef CONFIG_MIPS_MT_SMP
1485 /* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */
1486 #define IS_RANGE_P_INTERAPTIV_EVENT(r, b) \
1487 ((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 || \
1488 (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 || \
1489 (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 && \
1490 (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 || \
1491 ((b) >= 64 && (b) <= 67))
1492 #define IS_RANGE_V_INTERAPTIV_EVENT(r) ((r) == 47 || (r) == 175)
1493 #endif
1495 /* BMIPS5000 */
1496 #define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b) \
1497 ((b) == 0 || (b) == 1)
1501 * For most cores the user can use 0-255 raw events, where 0-127 for the events
1502 * of even counters, and 128-255 for odd counters. Note that bit 7 is used to
1503 * indicate the even/odd bank selector. So, for example, when user wants to take
1504 * the Event Num of 15 for odd counters (by referring to the user manual), then
1505 * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F)
1506 * to be used.
1508 * Some newer cores have even more events, in which case the user can use raw
1509 * events 0-511, where 0-255 are for the events of even counters, and 256-511
1510 * are for odd counters, so bit 8 is used to indicate the even/odd bank selector.
1512 static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1514 /* currently most cores have 7-bit event numbers */
1515 unsigned int raw_id = config & 0xff;
1516 unsigned int base_id = raw_id & 0x7f;
1518 switch (current_cpu_type()) {
1519 case CPU_24K:
1520 if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1521 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1522 else
1523 raw_event.cntr_mask =
1524 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1525 #ifdef CONFIG_MIPS_MT_SMP
1527 * This is actually doing nothing. Non-multithreading
1528 * CPUs will not check and calculate the range.
1530 raw_event.range = P;
1531 #endif
1532 break;
1533 case CPU_34K:
1534 if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1535 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1536 else
1537 raw_event.cntr_mask =
1538 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1539 #ifdef CONFIG_MIPS_MT_SMP
1540 if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1541 raw_event.range = P;
1542 else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1543 raw_event.range = V;
1544 else
1545 raw_event.range = T;
1546 #endif
1547 break;
1548 case CPU_74K:
1549 case CPU_1074K:
1550 if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1551 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1552 else
1553 raw_event.cntr_mask =
1554 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1555 #ifdef CONFIG_MIPS_MT_SMP
1556 raw_event.range = P;
1557 #endif
1558 break;
1559 case CPU_PROAPTIV:
1560 if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id))
1561 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1562 else
1563 raw_event.cntr_mask =
1564 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1565 #ifdef CONFIG_MIPS_MT_SMP
1566 raw_event.range = P;
1567 #endif
1568 break;
1569 case CPU_P5600:
1570 case CPU_P6600:
1571 /* 8-bit event numbers */
1572 raw_id = config & 0x1ff;
1573 base_id = raw_id & 0xff;
1574 if (IS_BOTH_COUNTERS_P5600_EVENT(base_id))
1575 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1576 else
1577 raw_event.cntr_mask =
1578 raw_id > 255 ? CNTR_ODD : CNTR_EVEN;
1579 #ifdef CONFIG_MIPS_MT_SMP
1580 raw_event.range = P;
1581 #endif
1582 break;
1583 case CPU_I6400:
1584 case CPU_I6500:
1585 /* 8-bit event numbers */
1586 base_id = config & 0xff;
1587 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1588 break;
1589 case CPU_1004K:
1590 if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1591 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1592 else
1593 raw_event.cntr_mask =
1594 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1595 #ifdef CONFIG_MIPS_MT_SMP
1596 if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1597 raw_event.range = P;
1598 else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1599 raw_event.range = V;
1600 else
1601 raw_event.range = T;
1602 #endif
1603 break;
1604 case CPU_INTERAPTIV:
1605 if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id))
1606 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1607 else
1608 raw_event.cntr_mask =
1609 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1610 #ifdef CONFIG_MIPS_MT_SMP
1611 if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id))
1612 raw_event.range = P;
1613 else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id)))
1614 raw_event.range = V;
1615 else
1616 raw_event.range = T;
1617 #endif
1618 break;
1619 case CPU_BMIPS5000:
1620 if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id))
1621 raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1622 else
1623 raw_event.cntr_mask =
1624 raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1625 break;
1626 case CPU_LOONGSON64:
1627 raw_event.cntr_mask = raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1628 break;
1631 raw_event.event_id = base_id;
1633 return &raw_event;
1636 static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1638 unsigned int raw_id = config & 0xff;
1639 unsigned int base_id = raw_id & 0x7f;
1642 raw_event.cntr_mask = CNTR_ALL;
1643 raw_event.event_id = base_id;
1645 if (current_cpu_type() == CPU_CAVIUM_OCTEON2) {
1646 if (base_id > 0x42)
1647 return ERR_PTR(-EOPNOTSUPP);
1648 } else {
1649 if (base_id > 0x3a)
1650 return ERR_PTR(-EOPNOTSUPP);
1653 switch (base_id) {
1654 case 0x00:
1655 case 0x0f:
1656 case 0x1e:
1657 case 0x1f:
1658 case 0x2f:
1659 case 0x34:
1660 case 0x3b ... 0x3f:
1661 return ERR_PTR(-EOPNOTSUPP);
1662 default:
1663 break;
1666 return &raw_event;
1669 static const struct mips_perf_event *xlp_pmu_map_raw_event(u64 config)
1671 unsigned int raw_id = config & 0xff;
1673 /* Only 1-63 are defined */
1674 if ((raw_id < 0x01) || (raw_id > 0x3f))
1675 return ERR_PTR(-EOPNOTSUPP);
1677 raw_event.cntr_mask = CNTR_ALL;
1678 raw_event.event_id = raw_id;
1680 return &raw_event;
1683 static int __init
1684 init_hw_perf_events(void)
1686 int counters, irq;
1687 int counter_bits;
1689 pr_info("Performance counters: ");
1691 counters = n_counters();
1692 if (counters == 0) {
1693 pr_cont("No available PMU.\n");
1694 return -ENODEV;
1697 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1698 if (!cpu_has_mipsmt_pertccounters)
1699 counters = counters_total_to_per_cpu(counters);
1700 #endif
1702 if (get_c0_perfcount_int)
1703 irq = get_c0_perfcount_int();
1704 else if (cp0_perfcount_irq >= 0)
1705 irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1706 else
1707 irq = -1;
1709 mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1711 switch (current_cpu_type()) {
1712 case CPU_24K:
1713 mipspmu.name = "mips/24K";
1714 mipspmu.general_event_map = &mipsxxcore_event_map;
1715 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1716 break;
1717 case CPU_34K:
1718 mipspmu.name = "mips/34K";
1719 mipspmu.general_event_map = &mipsxxcore_event_map;
1720 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1721 break;
1722 case CPU_74K:
1723 mipspmu.name = "mips/74K";
1724 mipspmu.general_event_map = &mipsxxcore_event_map2;
1725 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1726 break;
1727 case CPU_PROAPTIV:
1728 mipspmu.name = "mips/proAptiv";
1729 mipspmu.general_event_map = &mipsxxcore_event_map2;
1730 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1731 break;
1732 case CPU_P5600:
1733 mipspmu.name = "mips/P5600";
1734 mipspmu.general_event_map = &mipsxxcore_event_map2;
1735 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1736 break;
1737 case CPU_P6600:
1738 mipspmu.name = "mips/P6600";
1739 mipspmu.general_event_map = &mipsxxcore_event_map2;
1740 mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1741 break;
1742 case CPU_I6400:
1743 mipspmu.name = "mips/I6400";
1744 mipspmu.general_event_map = &i6x00_event_map;
1745 mipspmu.cache_event_map = &i6x00_cache_map;
1746 break;
1747 case CPU_I6500:
1748 mipspmu.name = "mips/I6500";
1749 mipspmu.general_event_map = &i6x00_event_map;
1750 mipspmu.cache_event_map = &i6x00_cache_map;
1751 break;
1752 case CPU_1004K:
1753 mipspmu.name = "mips/1004K";
1754 mipspmu.general_event_map = &mipsxxcore_event_map;
1755 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1756 break;
1757 case CPU_1074K:
1758 mipspmu.name = "mips/1074K";
1759 mipspmu.general_event_map = &mipsxxcore_event_map;
1760 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1761 break;
1762 case CPU_INTERAPTIV:
1763 mipspmu.name = "mips/interAptiv";
1764 mipspmu.general_event_map = &mipsxxcore_event_map;
1765 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1766 break;
1767 case CPU_LOONGSON32:
1768 mipspmu.name = "mips/loongson1";
1769 mipspmu.general_event_map = &mipsxxcore_event_map;
1770 mipspmu.cache_event_map = &mipsxxcore_cache_map;
1771 break;
1772 case CPU_LOONGSON64:
1773 mipspmu.name = "mips/loongson3";
1774 mipspmu.general_event_map = &loongson3_event_map;
1775 mipspmu.cache_event_map = &loongson3_cache_map;
1776 break;
1777 case CPU_CAVIUM_OCTEON:
1778 case CPU_CAVIUM_OCTEON_PLUS:
1779 case CPU_CAVIUM_OCTEON2:
1780 mipspmu.name = "octeon";
1781 mipspmu.general_event_map = &octeon_event_map;
1782 mipspmu.cache_event_map = &octeon_cache_map;
1783 mipspmu.map_raw_event = octeon_pmu_map_raw_event;
1784 break;
1785 case CPU_BMIPS5000:
1786 mipspmu.name = "BMIPS5000";
1787 mipspmu.general_event_map = &bmips5000_event_map;
1788 mipspmu.cache_event_map = &bmips5000_cache_map;
1789 break;
1790 case CPU_XLP:
1791 mipspmu.name = "xlp";
1792 mipspmu.general_event_map = &xlp_event_map;
1793 mipspmu.cache_event_map = &xlp_cache_map;
1794 mipspmu.map_raw_event = xlp_pmu_map_raw_event;
1795 break;
1796 default:
1797 pr_cont("Either hardware does not support performance "
1798 "counters, or not yet implemented.\n");
1799 return -ENODEV;
1802 mipspmu.num_counters = counters;
1803 mipspmu.irq = irq;
1805 if (read_c0_perfctrl0() & MIPS_PERFCTRL_W) {
1806 mipspmu.max_period = (1ULL << 63) - 1;
1807 mipspmu.valid_count = (1ULL << 63) - 1;
1808 mipspmu.overflow = 1ULL << 63;
1809 mipspmu.read_counter = mipsxx_pmu_read_counter_64;
1810 mipspmu.write_counter = mipsxx_pmu_write_counter_64;
1811 counter_bits = 64;
1812 } else {
1813 mipspmu.max_period = (1ULL << 31) - 1;
1814 mipspmu.valid_count = (1ULL << 31) - 1;
1815 mipspmu.overflow = 1ULL << 31;
1816 mipspmu.read_counter = mipsxx_pmu_read_counter;
1817 mipspmu.write_counter = mipsxx_pmu_write_counter;
1818 counter_bits = 32;
1821 on_each_cpu(reset_counters, (void *)(long)counters, 1);
1823 pr_cont("%s PMU enabled, %d %d-bit counters available to each "
1824 "CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
1825 irq < 0 ? " (share with timer interrupt)" : "");
1827 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1829 return 0;
1831 early_initcall(init_hw_perf_events);