1 /*
2  * Linux performance counter support for MIPS.
3  *
4  * Copyright (C) 2010 MIPS Technologies, Inc.
5  * Copyright (C) 2011 Cavium Networks, Inc.
6  * Author: Deng-Cheng Zhu
7  *
8  * This code is based on the implementation for ARM, which is in turn
9  * based on the sparc64 perf event code and the x86 code. Performance
10  * counter access is based on the MIPS Oprofile code. And the callchain
11  * support references the code of MIPS stacktrace.c.
12  *
13  * This program is free software; you can redistribute it and/or modify
14  * it under the terms of the GNU General Public License version 2 as
15  * published by the Free Software Foundation.
16  */
17 
18 #include <linux/cpumask.h>
19 #include <linux/interrupt.h>
20 #include <linux/smp.h>
21 #include <linux/kernel.h>
22 #include <linux/perf_event.h>
23 #include <linux/uaccess.h>
24 
25 #include <asm/irq.h>
26 #include <asm/irq_regs.h>
27 #include <asm/stacktrace.h>
28 #include <asm/time.h> /* For perf_irq */
29 
30 #define MIPS_MAX_HWEVENTS 4
31 #define MIPS_TCS_PER_COUNTER 2
32 #define MIPS_CPUID_TO_COUNTER_MASK (MIPS_TCS_PER_COUNTER - 1)
33 
34 struct cpu_hw_events {
35 	/* Array of events on this cpu. */
36 	struct perf_event	*events[MIPS_MAX_HWEVENTS];
37 
38 	/*
39 	 * Set the bit (indexed by the counter number) when the counter
40 	 * is used for an event.
41 	 */
42 	unsigned long		used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
43 
44 	/*
45 	 * Software copy of the control register for each performance counter.
46 	 * MIPS CPUs vary in performance counters. They use this differently,
47 	 * and even may not use it.
48 	 */
49 	unsigned int		saved_ctrl[MIPS_MAX_HWEVENTS];
50 };
51 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
52 	.saved_ctrl = {0},
53 };
54 
55 /* The description of MIPS performance events. */
56 struct mips_perf_event {
57 	unsigned int event_id;
58 	/*
59 	 * MIPS performance counters are indexed starting from 0.
60 	 * CNTR_EVEN indicates the indexes of the counters to be used are
61 	 * even numbers.
62 	 */
63 	unsigned int cntr_mask;
64 	#define CNTR_EVEN	0x55555555
65 	#define CNTR_ODD	0xaaaaaaaa
66 	#define CNTR_ALL	0xffffffff
67 #ifdef CONFIG_MIPS_MT_SMP
68 	enum {
69 		T  = 0,
70 		V  = 1,
71 		P  = 2,
72 	} range;
73 #else
74 	#define T
75 	#define V
76 	#define P
77 #endif
78 };
79 
80 static struct mips_perf_event raw_event;
81 static DEFINE_MUTEX(raw_event_mutex);
82 
83 #define C(x) PERF_COUNT_HW_CACHE_##x
84 
85 struct mips_pmu {
86 	u64		max_period;
87 	u64		valid_count;
88 	u64		overflow;
89 	const char	*name;
90 	int		irq;
91 	u64		(*read_counter)(unsigned int idx);
92 	void		(*write_counter)(unsigned int idx, u64 val);
93 	const struct mips_perf_event *(*map_raw_event)(u64 config);
94 	const struct mips_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
95 	const struct mips_perf_event (*cache_event_map)
96 				[PERF_COUNT_HW_CACHE_MAX]
97 				[PERF_COUNT_HW_CACHE_OP_MAX]
98 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
99 	unsigned int	num_counters;
100 };
101 
102 static struct mips_pmu mipspmu;
103 
104 #define M_PERFCTL_EVENT(event)		(((event) << MIPS_PERFCTRL_EVENT_S) & \
105 					 MIPS_PERFCTRL_EVENT)
106 #define M_PERFCTL_VPEID(vpe)		((vpe)	  << MIPS_PERFCTRL_VPEID_S)
107 
108 #ifdef CONFIG_CPU_BMIPS5000
109 #define M_PERFCTL_MT_EN(filter)		0
110 #else /* !CONFIG_CPU_BMIPS5000 */
111 #define M_PERFCTL_MT_EN(filter)		(filter)
112 #endif /* CONFIG_CPU_BMIPS5000 */
113 
114 #define	   M_TC_EN_ALL			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_ALL)
115 #define	   M_TC_EN_VPE			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_VPE)
116 #define	   M_TC_EN_TC			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_TC)
117 
118 #define M_PERFCTL_COUNT_EVENT_WHENEVER	(MIPS_PERFCTRL_EXL |		\
119 					 MIPS_PERFCTRL_K |		\
120 					 MIPS_PERFCTRL_U |		\
121 					 MIPS_PERFCTRL_S |		\
122 					 MIPS_PERFCTRL_IE)
123 
124 #ifdef CONFIG_MIPS_MT_SMP
125 #define M_PERFCTL_CONFIG_MASK		0x3fff801f
126 #else
127 #define M_PERFCTL_CONFIG_MASK		0x1f
128 #endif
129 
130 
131 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
132 static int cpu_has_mipsmt_pertccounters;
133 
134 static DEFINE_RWLOCK(pmuint_rwlock);
135 
136 #if defined(CONFIG_CPU_BMIPS5000)
137 #define vpe_id()	(cpu_has_mipsmt_pertccounters ? \
138 			 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK))
139 #else
140 /*
141  * FIXME: For VSMP, vpe_id() is redefined for Perf-events, because
142  * cpu_data[cpuid].vpe_id reports 0 for _both_ CPUs.
143  */
144 #define vpe_id()	(cpu_has_mipsmt_pertccounters ? \
145 			 0 : smp_processor_id())
146 #endif
147 
148 /* Copied from op_model_mipsxx.c */
149 static unsigned int vpe_shift(void)
150 {
151 	if (num_possible_cpus() > 1)
152 		return 1;
153 
154 	return 0;
155 }
156 
157 static unsigned int counters_total_to_per_cpu(unsigned int counters)
158 {
159 	return counters >> vpe_shift();
160 }
161 
162 #else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
163 #define vpe_id()	0
164 
165 #endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
166 
167 static void resume_local_counters(void);
168 static void pause_local_counters(void);
169 static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
170 static int mipsxx_pmu_handle_shared_irq(void);
171 
172 static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
173 {
174 	if (vpe_id() == 1)
175 		idx = (idx + 2) & 3;
176 	return idx;
177 }
178 
179 static u64 mipsxx_pmu_read_counter(unsigned int idx)
180 {
181 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
182 
183 	switch (idx) {
184 	case 0:
185 		/*
186 		 * The counters are unsigned, we must cast to truncate
187 		 * off the high bits.
188 		 */
189 		return (u32)read_c0_perfcntr0();
190 	case 1:
191 		return (u32)read_c0_perfcntr1();
192 	case 2:
193 		return (u32)read_c0_perfcntr2();
194 	case 3:
195 		return (u32)read_c0_perfcntr3();
196 	default:
197 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
198 		return 0;
199 	}
200 }
201 
202 static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
203 {
204 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
205 
206 	switch (idx) {
207 	case 0:
208 		return read_c0_perfcntr0_64();
209 	case 1:
210 		return read_c0_perfcntr1_64();
211 	case 2:
212 		return read_c0_perfcntr2_64();
213 	case 3:
214 		return read_c0_perfcntr3_64();
215 	default:
216 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
217 		return 0;
218 	}
219 }
220 
221 static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
222 {
223 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
224 
225 	switch (idx) {
226 	case 0:
227 		write_c0_perfcntr0(val);
228 		return;
229 	case 1:
230 		write_c0_perfcntr1(val);
231 		return;
232 	case 2:
233 		write_c0_perfcntr2(val);
234 		return;
235 	case 3:
236 		write_c0_perfcntr3(val);
237 		return;
238 	}
239 }
240 
241 static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
242 {
243 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
244 
245 	switch (idx) {
246 	case 0:
247 		write_c0_perfcntr0_64(val);
248 		return;
249 	case 1:
250 		write_c0_perfcntr1_64(val);
251 		return;
252 	case 2:
253 		write_c0_perfcntr2_64(val);
254 		return;
255 	case 3:
256 		write_c0_perfcntr3_64(val);
257 		return;
258 	}
259 }
260 
261 static unsigned int mipsxx_pmu_read_control(unsigned int idx)
262 {
263 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
264 
265 	switch (idx) {
266 	case 0:
267 		return read_c0_perfctrl0();
268 	case 1:
269 		return read_c0_perfctrl1();
270 	case 2:
271 		return read_c0_perfctrl2();
272 	case 3:
273 		return read_c0_perfctrl3();
274 	default:
275 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
276 		return 0;
277 	}
278 }
279 
280 static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
281 {
282 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
283 
284 	switch (idx) {
285 	case 0:
286 		write_c0_perfctrl0(val);
287 		return;
288 	case 1:
289 		write_c0_perfctrl1(val);
290 		return;
291 	case 2:
292 		write_c0_perfctrl2(val);
293 		return;
294 	case 3:
295 		write_c0_perfctrl3(val);
296 		return;
297 	}
298 }
299 
300 static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
301 				    struct hw_perf_event *hwc)
302 {
303 	int i;
304 
305 	/*
306 	 * We only need to care the counter mask. The range has been
307 	 * checked definitely.
308 	 */
309 	unsigned long cntr_mask = (hwc->event_base >> 8) & 0xffff;
310 
311 	for (i = mipspmu.num_counters - 1; i >= 0; i--) {
312 		/*
313 		 * Note that some MIPS perf events can be counted by both
314 		 * even and odd counters, wheresas many other are only by
315 		 * even _or_ odd counters. This introduces an issue that
316 		 * when the former kind of event takes the counter the
317 		 * latter kind of event wants to use, then the "counter
318 		 * allocation" for the latter event will fail. In fact if
319 		 * they can be dynamically swapped, they both feel happy.
320 		 * But here we leave this issue alone for now.
321 		 */
322 		if (test_bit(i, &cntr_mask) &&
323 			!test_and_set_bit(i, cpuc->used_mask))
324 			return i;
325 	}
326 
327 	return -EAGAIN;
328 }
329 
330 static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
331 {
332 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
333 
334 	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
335 
336 	cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
337 		(evt->config_base & M_PERFCTL_CONFIG_MASK) |
338 		/* Make sure interrupt enabled. */
339 		MIPS_PERFCTRL_IE;
340 	if (IS_ENABLED(CONFIG_CPU_BMIPS5000))
341 		/* enable the counter for the calling thread */
342 		cpuc->saved_ctrl[idx] |=
343 			(1 << (12 + vpe_id())) | BRCM_PERFCTRL_TC;
344 
345 	/*
346 	 * We do not actually let the counter run. Leave it until start().
347 	 */
348 }
349 
350 static void mipsxx_pmu_disable_event(int idx)
351 {
352 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
353 	unsigned long flags;
354 
355 	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
356 
357 	local_irq_save(flags);
358 	cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
359 		~M_PERFCTL_COUNT_EVENT_WHENEVER;
360 	mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
361 	local_irq_restore(flags);
362 }
363 
364 static int mipspmu_event_set_period(struct perf_event *event,
365 				    struct hw_perf_event *hwc,
366 				    int idx)
367 {
368 	u64 left = local64_read(&hwc->period_left);
369 	u64 period = hwc->sample_period;
370 	int ret = 0;
371 
372 	if (unlikely((left + period) & (1ULL << 63))) {
373 		/* left underflowed by more than period. */
374 		left = period;
375 		local64_set(&hwc->period_left, left);
376 		hwc->last_period = period;
377 		ret = 1;
378 	} else	if (unlikely((left + period) <= period)) {
379 		/* left underflowed by less than period. */
380 		left += period;
381 		local64_set(&hwc->period_left, left);
382 		hwc->last_period = period;
383 		ret = 1;
384 	}
385 
386 	if (left > mipspmu.max_period) {
387 		left = mipspmu.max_period;
388 		local64_set(&hwc->period_left, left);
389 	}
390 
391 	local64_set(&hwc->prev_count, mipspmu.overflow - left);
392 
393 	mipspmu.write_counter(idx, mipspmu.overflow - left);
394 
395 	perf_event_update_userpage(event);
396 
397 	return ret;
398 }
399 
400 static void mipspmu_event_update(struct perf_event *event,
401 				 struct hw_perf_event *hwc,
402 				 int idx)
403 {
404 	u64 prev_raw_count, new_raw_count;
405 	u64 delta;
406 
407 again:
408 	prev_raw_count = local64_read(&hwc->prev_count);
409 	new_raw_count = mipspmu.read_counter(idx);
410 
411 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
412 				new_raw_count) != prev_raw_count)
413 		goto again;
414 
415 	delta = new_raw_count - prev_raw_count;
416 
417 	local64_add(delta, &event->count);
418 	local64_sub(delta, &hwc->period_left);
419 }
420 
421 static void mipspmu_start(struct perf_event *event, int flags)
422 {
423 	struct hw_perf_event *hwc = &event->hw;
424 
425 	if (flags & PERF_EF_RELOAD)
426 		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
427 
428 	hwc->state = 0;
429 
430 	/* Set the period for the event. */
431 	mipspmu_event_set_period(event, hwc, hwc->idx);
432 
433 	/* Enable the event. */
434 	mipsxx_pmu_enable_event(hwc, hwc->idx);
435 }
436 
437 static void mipspmu_stop(struct perf_event *event, int flags)
438 {
439 	struct hw_perf_event *hwc = &event->hw;
440 
441 	if (!(hwc->state & PERF_HES_STOPPED)) {
442 		/* We are working on a local event. */
443 		mipsxx_pmu_disable_event(hwc->idx);
444 		barrier();
445 		mipspmu_event_update(event, hwc, hwc->idx);
446 		hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
447 	}
448 }
449 
450 static int mipspmu_add(struct perf_event *event, int flags)
451 {
452 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
453 	struct hw_perf_event *hwc = &event->hw;
454 	int idx;
455 	int err = 0;
456 
457 	perf_pmu_disable(event->pmu);
458 
459 	/* To look for a free counter for this event. */
460 	idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
461 	if (idx < 0) {
462 		err = idx;
463 		goto out;
464 	}
465 
466 	/*
467 	 * If there is an event in the counter we are going to use then
468 	 * make sure it is disabled.
469 	 */
470 	event->hw.idx = idx;
471 	mipsxx_pmu_disable_event(idx);
472 	cpuc->events[idx] = event;
473 
474 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
475 	if (flags & PERF_EF_START)
476 		mipspmu_start(event, PERF_EF_RELOAD);
477 
478 	/* Propagate our changes to the userspace mapping. */
479 	perf_event_update_userpage(event);
480 
481 out:
482 	perf_pmu_enable(event->pmu);
483 	return err;
484 }
485 
486 static void mipspmu_del(struct perf_event *event, int flags)
487 {
488 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
489 	struct hw_perf_event *hwc = &event->hw;
490 	int idx = hwc->idx;
491 
492 	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
493 
494 	mipspmu_stop(event, PERF_EF_UPDATE);
495 	cpuc->events[idx] = NULL;
496 	clear_bit(idx, cpuc->used_mask);
497 
498 	perf_event_update_userpage(event);
499 }
500 
501 static void mipspmu_read(struct perf_event *event)
502 {
503 	struct hw_perf_event *hwc = &event->hw;
504 
505 	/* Don't read disabled counters! */
506 	if (hwc->idx < 0)
507 		return;
508 
509 	mipspmu_event_update(event, hwc, hwc->idx);
510 }
511 
512 static void mipspmu_enable(struct pmu *pmu)
513 {
514 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
515 	write_unlock(&pmuint_rwlock);
516 #endif
517 	resume_local_counters();
518 }
519 
520 /*
521  * MIPS performance counters can be per-TC. The control registers can
522  * not be directly accessed across CPUs. Hence if we want to do global
523  * control, we need cross CPU calls. on_each_cpu() can help us, but we
524  * can not make sure this function is called with interrupts enabled. So
525  * here we pause local counters and then grab a rwlock and leave the
526  * counters on other CPUs alone. If any counter interrupt raises while
527  * we own the write lock, simply pause local counters on that CPU and
528  * spin in the handler. Also we know we won't be switched to another
529  * CPU after pausing local counters and before grabbing the lock.
530  */
531 static void mipspmu_disable(struct pmu *pmu)
532 {
533 	pause_local_counters();
534 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
535 	write_lock(&pmuint_rwlock);
536 #endif
537 }
538 
539 static atomic_t active_events = ATOMIC_INIT(0);
540 static DEFINE_MUTEX(pmu_reserve_mutex);
541 static int (*save_perf_irq)(void);
542 
543 static int mipspmu_get_irq(void)
544 {
545 	int err;
546 
547 	if (mipspmu.irq >= 0) {
548 		/* Request my own irq handler. */
549 		err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
550 				  IRQF_PERCPU | IRQF_NOBALANCING |
551 				  IRQF_NO_THREAD | IRQF_NO_SUSPEND |
552 				  IRQF_SHARED,
553 				  "mips_perf_pmu", &mipspmu);
554 		if (err) {
555 			pr_warn("Unable to request IRQ%d for MIPS performance counters!\n",
556 				mipspmu.irq);
557 		}
558 	} else if (cp0_perfcount_irq < 0) {
559 		/*
560 		 * We are sharing the irq number with the timer interrupt.
561 		 */
562 		save_perf_irq = perf_irq;
563 		perf_irq = mipsxx_pmu_handle_shared_irq;
564 		err = 0;
565 	} else {
566 		pr_warn("The platform hasn't properly defined its interrupt controller\n");
567 		err = -ENOENT;
568 	}
569 
570 	return err;
571 }
572 
573 static void mipspmu_free_irq(void)
574 {
575 	if (mipspmu.irq >= 0)
576 		free_irq(mipspmu.irq, &mipspmu);
577 	else if (cp0_perfcount_irq < 0)
578 		perf_irq = save_perf_irq;
579 }
580 
581 /*
582  * mipsxx/rm9000/loongson2 have different performance counters, they have
583  * specific low-level init routines.
584  */
585 static void reset_counters(void *arg);
586 static int __hw_perf_event_init(struct perf_event *event);
587 
588 static void hw_perf_event_destroy(struct perf_event *event)
589 {
590 	if (atomic_dec_and_mutex_lock(&active_events,
591 				&pmu_reserve_mutex)) {
592 		/*
593 		 * We must not call the destroy function with interrupts
594 		 * disabled.
595 		 */
596 		on_each_cpu(reset_counters,
597 			(void *)(long)mipspmu.num_counters, 1);
598 		mipspmu_free_irq();
599 		mutex_unlock(&pmu_reserve_mutex);
600 	}
601 }
602 
603 static int mipspmu_event_init(struct perf_event *event)
604 {
605 	int err = 0;
606 
607 	/* does not support taken branch sampling */
608 	if (has_branch_stack(event))
609 		return -EOPNOTSUPP;
610 
611 	switch (event->attr.type) {
612 	case PERF_TYPE_RAW:
613 	case PERF_TYPE_HARDWARE:
614 	case PERF_TYPE_HW_CACHE:
615 		break;
616 
617 	default:
618 		return -ENOENT;
619 	}
620 
621 	if (event->cpu >= nr_cpumask_bits ||
622 	    (event->cpu >= 0 && !cpu_online(event->cpu)))
623 		return -ENODEV;
624 
625 	if (!atomic_inc_not_zero(&active_events)) {
626 		mutex_lock(&pmu_reserve_mutex);
627 		if (atomic_read(&active_events) == 0)
628 			err = mipspmu_get_irq();
629 
630 		if (!err)
631 			atomic_inc(&active_events);
632 		mutex_unlock(&pmu_reserve_mutex);
633 	}
634 
635 	if (err)
636 		return err;
637 
638 	return __hw_perf_event_init(event);
639 }
640 
641 static struct pmu pmu = {
642 	.pmu_enable	= mipspmu_enable,
643 	.pmu_disable	= mipspmu_disable,
644 	.event_init	= mipspmu_event_init,
645 	.add		= mipspmu_add,
646 	.del		= mipspmu_del,
647 	.start		= mipspmu_start,
648 	.stop		= mipspmu_stop,
649 	.read		= mipspmu_read,
650 };
651 
652 static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
653 {
654 /*
655  * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
656  * event_id.
657  */
658 #ifdef CONFIG_MIPS_MT_SMP
659 	return ((unsigned int)pev->range << 24) |
660 		(pev->cntr_mask & 0xffff00) |
661 		(pev->event_id & 0xff);
662 #else
663 	return (pev->cntr_mask & 0xffff00) |
664 		(pev->event_id & 0xff);
665 #endif
666 }
667 
668 static const struct mips_perf_event *mipspmu_map_general_event(int idx)
669 {
670 
671 	if ((*mipspmu.general_event_map)[idx].cntr_mask == 0)
672 		return ERR_PTR(-EOPNOTSUPP);
673 	return &(*mipspmu.general_event_map)[idx];
674 }
675 
676 static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
677 {
678 	unsigned int cache_type, cache_op, cache_result;
679 	const struct mips_perf_event *pev;
680 
681 	cache_type = (config >> 0) & 0xff;
682 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
683 		return ERR_PTR(-EINVAL);
684 
685 	cache_op = (config >> 8) & 0xff;
686 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
687 		return ERR_PTR(-EINVAL);
688 
689 	cache_result = (config >> 16) & 0xff;
690 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
691 		return ERR_PTR(-EINVAL);
692 
693 	pev = &((*mipspmu.cache_event_map)
694 					[cache_type]
695 					[cache_op]
696 					[cache_result]);
697 
698 	if (pev->cntr_mask == 0)
699 		return ERR_PTR(-EOPNOTSUPP);
700 
701 	return pev;
702 
703 }
704 
705 static int validate_group(struct perf_event *event)
706 {
707 	struct perf_event *sibling, *leader = event->group_leader;
708 	struct cpu_hw_events fake_cpuc;
709 
710 	memset(&fake_cpuc, 0, sizeof(fake_cpuc));
711 
712 	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
713 		return -EINVAL;
714 
715 	list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
716 		if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
717 			return -EINVAL;
718 	}
719 
720 	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
721 		return -EINVAL;
722 
723 	return 0;
724 }
725 
726 /* This is needed by specific irq handlers in perf_event_*.c */
727 static void handle_associated_event(struct cpu_hw_events *cpuc,
728 				    int idx, struct perf_sample_data *data,
729 				    struct pt_regs *regs)
730 {
731 	struct perf_event *event = cpuc->events[idx];
732 	struct hw_perf_event *hwc = &event->hw;
733 
734 	mipspmu_event_update(event, hwc, idx);
735 	data->period = event->hw.last_period;
736 	if (!mipspmu_event_set_period(event, hwc, idx))
737 		return;
738 
739 	if (perf_event_overflow(event, data, regs))
740 		mipsxx_pmu_disable_event(idx);
741 }
742 
743 
744 static int __n_counters(void)
745 {
746 	if (!cpu_has_perf)
747 		return 0;
748 	if (!(read_c0_perfctrl0() & MIPS_PERFCTRL_M))
749 		return 1;
750 	if (!(read_c0_perfctrl1() & MIPS_PERFCTRL_M))
751 		return 2;
752 	if (!(read_c0_perfctrl2() & MIPS_PERFCTRL_M))
753 		return 3;
754 
755 	return 4;
756 }
757 
758 static int n_counters(void)
759 {
760 	int counters;
761 
762 	switch (current_cpu_type()) {
763 	case CPU_R10000:
764 		counters = 2;
765 		break;
766 
767 	case CPU_R12000:
768 	case CPU_R14000:
769 	case CPU_R16000:
770 		counters = 4;
771 		break;
772 
773 	default:
774 		counters = __n_counters();
775 	}
776 
777 	return counters;
778 }
779 
780 static void reset_counters(void *arg)
781 {
782 	int counters = (int)(long)arg;
783 	switch (counters) {
784 	case 4:
785 		mipsxx_pmu_write_control(3, 0);
786 		mipspmu.write_counter(3, 0);
787 	case 3:
788 		mipsxx_pmu_write_control(2, 0);
789 		mipspmu.write_counter(2, 0);
790 	case 2:
791 		mipsxx_pmu_write_control(1, 0);
792 		mipspmu.write_counter(1, 0);
793 	case 1:
794 		mipsxx_pmu_write_control(0, 0);
795 		mipspmu.write_counter(0, 0);
796 	}
797 }
798 
799 /* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */
800 static const struct mips_perf_event mipsxxcore_event_map
801 				[PERF_COUNT_HW_MAX] = {
802 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
803 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
804 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
805 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
806 };
807 
808 /* 74K/proAptiv core has different branch event code. */
809 static const struct mips_perf_event mipsxxcore_event_map2
810 				[PERF_COUNT_HW_MAX] = {
811 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
812 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
813 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
814 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
815 };
816 
817 static const struct mips_perf_event i6400_event_map[PERF_COUNT_HW_MAX] = {
818 	[PERF_COUNT_HW_CPU_CYCLES]          = { 0x00, CNTR_EVEN | CNTR_ODD },
819 	[PERF_COUNT_HW_INSTRUCTIONS]        = { 0x01, CNTR_EVEN | CNTR_ODD },
820 	/* These only count dcache, not icache */
821 	[PERF_COUNT_HW_CACHE_REFERENCES]    = { 0x45, CNTR_EVEN | CNTR_ODD },
822 	[PERF_COUNT_HW_CACHE_MISSES]        = { 0x48, CNTR_EVEN | CNTR_ODD },
823 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x15, CNTR_EVEN | CNTR_ODD },
824 	[PERF_COUNT_HW_BRANCH_MISSES]       = { 0x16, CNTR_EVEN | CNTR_ODD },
825 };
826 
827 static const struct mips_perf_event loongson3_event_map[PERF_COUNT_HW_MAX] = {
828 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN },
829 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD },
830 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN },
831 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD },
832 };
833 
834 static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
835 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
836 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
837 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
838 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL	 },
839 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
840 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
841 	[PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
842 };
843 
844 static const struct mips_perf_event bmips5000_event_map
845 				[PERF_COUNT_HW_MAX] = {
846 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T },
847 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
848 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
849 };
850 
851 static const struct mips_perf_event xlp_event_map[PERF_COUNT_HW_MAX] = {
852 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
853 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x18, CNTR_ALL }, /* PAPI_TOT_INS */
854 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
855 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
856 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x1b, CNTR_ALL }, /* PAPI_BR_CN */
857 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x1c, CNTR_ALL }, /* PAPI_BR_MSP */
858 };
859 
860 /* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */
861 static const struct mips_perf_event mipsxxcore_cache_map
862 				[PERF_COUNT_HW_CACHE_MAX]
863 				[PERF_COUNT_HW_CACHE_OP_MAX]
864 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
865 [C(L1D)] = {
866 	/*
867 	 * Like some other architectures (e.g. ARM), the performance
868 	 * counters don't differentiate between read and write
869 	 * accesses/misses, so this isn't strictly correct, but it's the
870 	 * best we can do. Writes and reads get combined.
871 	 */
872 	[C(OP_READ)] = {
873 		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
874 		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
875 	},
876 	[C(OP_WRITE)] = {
877 		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
878 		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
879 	},
880 },
881 [C(L1I)] = {
882 	[C(OP_READ)] = {
883 		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
884 		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
885 	},
886 	[C(OP_WRITE)] = {
887 		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
888 		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
889 	},
890 	[C(OP_PREFETCH)] = {
891 		[C(RESULT_ACCESS)]	= { 0x14, CNTR_EVEN, T },
892 		/*
893 		 * Note that MIPS has only "hit" events countable for
894 		 * the prefetch operation.
895 		 */
896 	},
897 },
898 [C(LL)] = {
899 	[C(OP_READ)] = {
900 		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
901 		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
902 	},
903 	[C(OP_WRITE)] = {
904 		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
905 		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
906 	},
907 },
908 [C(DTLB)] = {
909 	[C(OP_READ)] = {
910 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
911 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
912 	},
913 	[C(OP_WRITE)] = {
914 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
915 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
916 	},
917 },
918 [C(ITLB)] = {
919 	[C(OP_READ)] = {
920 		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
921 		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
922 	},
923 	[C(OP_WRITE)] = {
924 		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
925 		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
926 	},
927 },
928 [C(BPU)] = {
929 	/* Using the same code for *HW_BRANCH* */
930 	[C(OP_READ)] = {
931 		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
932 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
933 	},
934 	[C(OP_WRITE)] = {
935 		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
936 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
937 	},
938 },
939 };
940 
941 /* 74K/proAptiv core has completely different cache event map. */
942 static const struct mips_perf_event mipsxxcore_cache_map2
943 				[PERF_COUNT_HW_CACHE_MAX]
944 				[PERF_COUNT_HW_CACHE_OP_MAX]
945 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
946 [C(L1D)] = {
947 	/*
948 	 * Like some other architectures (e.g. ARM), the performance
949 	 * counters don't differentiate between read and write
950 	 * accesses/misses, so this isn't strictly correct, but it's the
951 	 * best we can do. Writes and reads get combined.
952 	 */
953 	[C(OP_READ)] = {
954 		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
955 		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
956 	},
957 	[C(OP_WRITE)] = {
958 		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
959 		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
960 	},
961 },
962 [C(L1I)] = {
963 	[C(OP_READ)] = {
964 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
965 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
966 	},
967 	[C(OP_WRITE)] = {
968 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
969 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
970 	},
971 	[C(OP_PREFETCH)] = {
972 		[C(RESULT_ACCESS)]	= { 0x34, CNTR_EVEN, T },
973 		/*
974 		 * Note that MIPS has only "hit" events countable for
975 		 * the prefetch operation.
976 		 */
977 	},
978 },
979 [C(LL)] = {
980 	[C(OP_READ)] = {
981 		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
982 		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN, P },
983 	},
984 	[C(OP_WRITE)] = {
985 		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
986 		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN, P },
987 	},
988 },
989 /*
990  * 74K core does not have specific DTLB events. proAptiv core has
991  * "speculative" DTLB events which are numbered 0x63 (even/odd) and
992  * not included here. One can use raw events if really needed.
993  */
994 [C(ITLB)] = {
995 	[C(OP_READ)] = {
996 		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
997 		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
998 	},
999 	[C(OP_WRITE)] = {
1000 		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
1001 		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
1002 	},
1003 },
1004 [C(BPU)] = {
1005 	/* Using the same code for *HW_BRANCH* */
1006 	[C(OP_READ)] = {
1007 		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1008 		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1009 	},
1010 	[C(OP_WRITE)] = {
1011 		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1012 		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1013 	},
1014 },
1015 };
1016 
1017 static const struct mips_perf_event i6400_cache_map
1018 				[PERF_COUNT_HW_CACHE_MAX]
1019 				[PERF_COUNT_HW_CACHE_OP_MAX]
1020 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1021 [C(L1D)] = {
1022 	[C(OP_READ)] = {
1023 		[C(RESULT_ACCESS)]	= { 0x46, CNTR_EVEN | CNTR_ODD },
1024 		[C(RESULT_MISS)]	= { 0x49, CNTR_EVEN | CNTR_ODD },
1025 	},
1026 	[C(OP_WRITE)] = {
1027 		[C(RESULT_ACCESS)]	= { 0x47, CNTR_EVEN | CNTR_ODD },
1028 		[C(RESULT_MISS)]	= { 0x4a, CNTR_EVEN | CNTR_ODD },
1029 	},
1030 },
1031 [C(L1I)] = {
1032 	[C(OP_READ)] = {
1033 		[C(RESULT_ACCESS)]	= { 0x84, CNTR_EVEN | CNTR_ODD },
1034 		[C(RESULT_MISS)]	= { 0x85, CNTR_EVEN | CNTR_ODD },
1035 	},
1036 },
1037 [C(DTLB)] = {
1038 	/* Can't distinguish read & write */
1039 	[C(OP_READ)] = {
1040 		[C(RESULT_ACCESS)]	= { 0x40, CNTR_EVEN | CNTR_ODD },
1041 		[C(RESULT_MISS)]	= { 0x41, CNTR_EVEN | CNTR_ODD },
1042 	},
1043 	[C(OP_WRITE)] = {
1044 		[C(RESULT_ACCESS)]	= { 0x40, CNTR_EVEN | CNTR_ODD },
1045 		[C(RESULT_MISS)]	= { 0x41, CNTR_EVEN | CNTR_ODD },
1046 	},
1047 },
1048 [C(BPU)] = {
1049 	/* Conditional branches / mispredicted */
1050 	[C(OP_READ)] = {
1051 		[C(RESULT_ACCESS)]	= { 0x15, CNTR_EVEN | CNTR_ODD },
1052 		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN | CNTR_ODD },
1053 	},
1054 },
1055 };
1056 
1057 static const struct mips_perf_event loongson3_cache_map
1058 				[PERF_COUNT_HW_CACHE_MAX]
1059 				[PERF_COUNT_HW_CACHE_OP_MAX]
1060 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1061 [C(L1D)] = {
1062 	/*
1063 	 * Like some other architectures (e.g. ARM), the performance
1064 	 * counters don't differentiate between read and write
1065 	 * accesses/misses, so this isn't strictly correct, but it's the
1066 	 * best we can do. Writes and reads get combined.
1067 	 */
1068 	[C(OP_READ)] = {
1069 		[C(RESULT_MISS)]        = { 0x04, CNTR_ODD },
1070 	},
1071 	[C(OP_WRITE)] = {
1072 		[C(RESULT_MISS)]        = { 0x04, CNTR_ODD },
1073 	},
1074 },
1075 [C(L1I)] = {
1076 	[C(OP_READ)] = {
1077 		[C(RESULT_MISS)]        = { 0x04, CNTR_EVEN },
1078 	},
1079 	[C(OP_WRITE)] = {
1080 		[C(RESULT_MISS)]        = { 0x04, CNTR_EVEN },
1081 	},
1082 },
1083 [C(DTLB)] = {
1084 	[C(OP_READ)] = {
1085 		[C(RESULT_MISS)]        = { 0x09, CNTR_ODD },
1086 	},
1087 	[C(OP_WRITE)] = {
1088 		[C(RESULT_MISS)]        = { 0x09, CNTR_ODD },
1089 	},
1090 },
1091 [C(ITLB)] = {
1092 	[C(OP_READ)] = {
1093 		[C(RESULT_MISS)]        = { 0x0c, CNTR_ODD },
1094 	},
1095 	[C(OP_WRITE)] = {
1096 		[C(RESULT_MISS)]        = { 0x0c, CNTR_ODD },
1097 	},
1098 },
1099 [C(BPU)] = {
1100 	/* Using the same code for *HW_BRANCH* */
1101 	[C(OP_READ)] = {
1102 		[C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN },
1103 		[C(RESULT_MISS)]        = { 0x02, CNTR_ODD },
1104 	},
1105 	[C(OP_WRITE)] = {
1106 		[C(RESULT_ACCESS)]      = { 0x02, CNTR_EVEN },
1107 		[C(RESULT_MISS)]        = { 0x02, CNTR_ODD },
1108 	},
1109 },
1110 };
1111 
1112 /* BMIPS5000 */
1113 static const struct mips_perf_event bmips5000_cache_map
1114 				[PERF_COUNT_HW_CACHE_MAX]
1115 				[PERF_COUNT_HW_CACHE_OP_MAX]
1116 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1117 [C(L1D)] = {
1118 	/*
1119 	 * Like some other architectures (e.g. ARM), the performance
1120 	 * counters don't differentiate between read and write
1121 	 * accesses/misses, so this isn't strictly correct, but it's the
1122 	 * best we can do. Writes and reads get combined.
1123 	 */
1124 	[C(OP_READ)] = {
1125 		[C(RESULT_ACCESS)]	= { 12, CNTR_EVEN, T },
1126 		[C(RESULT_MISS)]	= { 12, CNTR_ODD, T },
1127 	},
1128 	[C(OP_WRITE)] = {
1129 		[C(RESULT_ACCESS)]	= { 12, CNTR_EVEN, T },
1130 		[C(RESULT_MISS)]	= { 12, CNTR_ODD, T },
1131 	},
1132 },
1133 [C(L1I)] = {
1134 	[C(OP_READ)] = {
1135 		[C(RESULT_ACCESS)]	= { 10, CNTR_EVEN, T },
1136 		[C(RESULT_MISS)]	= { 10, CNTR_ODD, T },
1137 	},
1138 	[C(OP_WRITE)] = {
1139 		[C(RESULT_ACCESS)]	= { 10, CNTR_EVEN, T },
1140 		[C(RESULT_MISS)]	= { 10, CNTR_ODD, T },
1141 	},
1142 	[C(OP_PREFETCH)] = {
1143 		[C(RESULT_ACCESS)]	= { 23, CNTR_EVEN, T },
1144 		/*
1145 		 * Note that MIPS has only "hit" events countable for
1146 		 * the prefetch operation.
1147 		 */
1148 	},
1149 },
1150 [C(LL)] = {
1151 	[C(OP_READ)] = {
1152 		[C(RESULT_ACCESS)]	= { 28, CNTR_EVEN, P },
1153 		[C(RESULT_MISS)]	= { 28, CNTR_ODD, P },
1154 	},
1155 	[C(OP_WRITE)] = {
1156 		[C(RESULT_ACCESS)]	= { 28, CNTR_EVEN, P },
1157 		[C(RESULT_MISS)]	= { 28, CNTR_ODD, P },
1158 	},
1159 },
1160 [C(BPU)] = {
1161 	/* Using the same code for *HW_BRANCH* */
1162 	[C(OP_READ)] = {
1163 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1164 	},
1165 	[C(OP_WRITE)] = {
1166 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1167 	},
1168 },
1169 };
1170 
1171 
1172 static const struct mips_perf_event octeon_cache_map
1173 				[PERF_COUNT_HW_CACHE_MAX]
1174 				[PERF_COUNT_HW_CACHE_OP_MAX]
1175 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1176 [C(L1D)] = {
1177 	[C(OP_READ)] = {
1178 		[C(RESULT_ACCESS)]	= { 0x2b, CNTR_ALL },
1179 		[C(RESULT_MISS)]	= { 0x2e, CNTR_ALL },
1180 	},
1181 	[C(OP_WRITE)] = {
1182 		[C(RESULT_ACCESS)]	= { 0x30, CNTR_ALL },
1183 	},
1184 },
1185 [C(L1I)] = {
1186 	[C(OP_READ)] = {
1187 		[C(RESULT_ACCESS)]	= { 0x18, CNTR_ALL },
1188 	},
1189 	[C(OP_PREFETCH)] = {
1190 		[C(RESULT_ACCESS)]	= { 0x19, CNTR_ALL },
1191 	},
1192 },
1193 [C(DTLB)] = {
1194 	/*
1195 	 * Only general DTLB misses are counted use the same event for
1196 	 * read and write.
1197 	 */
1198 	[C(OP_READ)] = {
1199 		[C(RESULT_MISS)]	= { 0x35, CNTR_ALL },
1200 	},
1201 	[C(OP_WRITE)] = {
1202 		[C(RESULT_MISS)]	= { 0x35, CNTR_ALL },
1203 	},
1204 },
1205 [C(ITLB)] = {
1206 	[C(OP_READ)] = {
1207 		[C(RESULT_MISS)]	= { 0x37, CNTR_ALL },
1208 	},
1209 },
1210 };
1211 
1212 static const struct mips_perf_event xlp_cache_map
1213 				[PERF_COUNT_HW_CACHE_MAX]
1214 				[PERF_COUNT_HW_CACHE_OP_MAX]
1215 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1216 [C(L1D)] = {
1217 	[C(OP_READ)] = {
1218 		[C(RESULT_ACCESS)]	= { 0x31, CNTR_ALL }, /* PAPI_L1_DCR */
1219 		[C(RESULT_MISS)]	= { 0x30, CNTR_ALL }, /* PAPI_L1_LDM */
1220 	},
1221 	[C(OP_WRITE)] = {
1222 		[C(RESULT_ACCESS)]	= { 0x2f, CNTR_ALL }, /* PAPI_L1_DCW */
1223 		[C(RESULT_MISS)]	= { 0x2e, CNTR_ALL }, /* PAPI_L1_STM */
1224 	},
1225 },
1226 [C(L1I)] = {
1227 	[C(OP_READ)] = {
1228 		[C(RESULT_ACCESS)]	= { 0x04, CNTR_ALL }, /* PAPI_L1_ICA */
1229 		[C(RESULT_MISS)]	= { 0x07, CNTR_ALL }, /* PAPI_L1_ICM */
1230 	},
1231 },
1232 [C(LL)] = {
1233 	[C(OP_READ)] = {
1234 		[C(RESULT_ACCESS)]	= { 0x35, CNTR_ALL }, /* PAPI_L2_DCR */
1235 		[C(RESULT_MISS)]	= { 0x37, CNTR_ALL }, /* PAPI_L2_LDM */
1236 	},
1237 	[C(OP_WRITE)] = {
1238 		[C(RESULT_ACCESS)]	= { 0x34, CNTR_ALL }, /* PAPI_L2_DCA */
1239 		[C(RESULT_MISS)]	= { 0x36, CNTR_ALL }, /* PAPI_L2_DCM */
1240 	},
1241 },
1242 [C(DTLB)] = {
1243 	/*
1244 	 * Only general DTLB misses are counted use the same event for
1245 	 * read and write.
1246 	 */
1247 	[C(OP_READ)] = {
1248 		[C(RESULT_MISS)]	= { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1249 	},
1250 	[C(OP_WRITE)] = {
1251 		[C(RESULT_MISS)]	= { 0x2d, CNTR_ALL }, /* PAPI_TLB_DM */
1252 	},
1253 },
1254 [C(ITLB)] = {
1255 	[C(OP_READ)] = {
1256 		[C(RESULT_MISS)]	= { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1257 	},
1258 	[C(OP_WRITE)] = {
1259 		[C(RESULT_MISS)]	= { 0x08, CNTR_ALL }, /* PAPI_TLB_IM */
1260 	},
1261 },
1262 [C(BPU)] = {
1263 	[C(OP_READ)] = {
1264 		[C(RESULT_MISS)]	= { 0x25, CNTR_ALL },
1265 	},
1266 },
1267 };
1268 
1269 #ifdef CONFIG_MIPS_MT_SMP
1270 static void check_and_calc_range(struct perf_event *event,
1271 				 const struct mips_perf_event *pev)
1272 {
1273 	struct hw_perf_event *hwc = &event->hw;
1274 
1275 	if (event->cpu >= 0) {
1276 		if (pev->range > V) {
1277 			/*
1278 			 * The user selected an event that is processor
1279 			 * wide, while expecting it to be VPE wide.
1280 			 */
1281 			hwc->config_base |= M_TC_EN_ALL;
1282 		} else {
1283 			/*
1284 			 * FIXME: cpu_data[event->cpu].vpe_id reports 0
1285 			 * for both CPUs.
1286 			 */
1287 			hwc->config_base |= M_PERFCTL_VPEID(event->cpu);
1288 			hwc->config_base |= M_TC_EN_VPE;
1289 		}
1290 	} else
1291 		hwc->config_base |= M_TC_EN_ALL;
1292 }
1293 #else
1294 static void check_and_calc_range(struct perf_event *event,
1295 				 const struct mips_perf_event *pev)
1296 {
1297 }
1298 #endif
1299 
1300 static int __hw_perf_event_init(struct perf_event *event)
1301 {
1302 	struct perf_event_attr *attr = &event->attr;
1303 	struct hw_perf_event *hwc = &event->hw;
1304 	const struct mips_perf_event *pev;
1305 	int err;
1306 
1307 	/* Returning MIPS event descriptor for generic perf event. */
1308 	if (PERF_TYPE_HARDWARE == event->attr.type) {
1309 		if (event->attr.config >= PERF_COUNT_HW_MAX)
1310 			return -EINVAL;
1311 		pev = mipspmu_map_general_event(event->attr.config);
1312 	} else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1313 		pev = mipspmu_map_cache_event(event->attr.config);
1314 	} else if (PERF_TYPE_RAW == event->attr.type) {
1315 		/* We are working on the global raw event. */
1316 		mutex_lock(&raw_event_mutex);
1317 		pev = mipspmu.map_raw_event(event->attr.config);
1318 	} else {
1319 		/* The event type is not (yet) supported. */
1320 		return -EOPNOTSUPP;
1321 	}
1322 
1323 	if (IS_ERR(pev)) {
1324 		if (PERF_TYPE_RAW == event->attr.type)
1325 			mutex_unlock(&raw_event_mutex);
1326 		return PTR_ERR(pev);
1327 	}
1328 
1329 	/*
1330 	 * We allow max flexibility on how each individual counter shared
1331 	 * by the single CPU operates (the mode exclusion and the range).
1332 	 */
1333 	hwc->config_base = MIPS_PERFCTRL_IE;
1334 
1335 	/* Calculate range bits and validate it. */
1336 	if (num_possible_cpus() > 1)
1337 		check_and_calc_range(event, pev);
1338 
1339 	hwc->event_base = mipspmu_perf_event_encode(pev);
1340 	if (PERF_TYPE_RAW == event->attr.type)
1341 		mutex_unlock(&raw_event_mutex);
1342 
1343 	if (!attr->exclude_user)
1344 		hwc->config_base |= MIPS_PERFCTRL_U;
1345 	if (!attr->exclude_kernel) {
1346 		hwc->config_base |= MIPS_PERFCTRL_K;
1347 		/* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1348 		hwc->config_base |= MIPS_PERFCTRL_EXL;
1349 	}
1350 	if (!attr->exclude_hv)
1351 		hwc->config_base |= MIPS_PERFCTRL_S;
1352 
1353 	hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1354 	/*
1355 	 * The event can belong to another cpu. We do not assign a local
1356 	 * counter for it for now.
1357 	 */
1358 	hwc->idx = -1;
1359 	hwc->config = 0;
1360 
1361 	if (!hwc->sample_period) {
1362 		hwc->sample_period  = mipspmu.max_period;
1363 		hwc->last_period    = hwc->sample_period;
1364 		local64_set(&hwc->period_left, hwc->sample_period);
1365 	}
1366 
1367 	err = 0;
1368 	if (event->group_leader != event)
1369 		err = validate_group(event);
1370 
1371 	event->destroy = hw_perf_event_destroy;
1372 
1373 	if (err)
1374 		event->destroy(event);
1375 
1376 	return err;
1377 }
1378 
1379 static void pause_local_counters(void)
1380 {
1381 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1382 	int ctr = mipspmu.num_counters;
1383 	unsigned long flags;
1384 
1385 	local_irq_save(flags);
1386 	do {
1387 		ctr--;
1388 		cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1389 		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1390 					 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1391 	} while (ctr > 0);
1392 	local_irq_restore(flags);
1393 }
1394 
1395 static void resume_local_counters(void)
1396 {
1397 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1398 	int ctr = mipspmu.num_counters;
1399 
1400 	do {
1401 		ctr--;
1402 		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1403 	} while (ctr > 0);
1404 }
1405 
1406 static int mipsxx_pmu_handle_shared_irq(void)
1407 {
1408 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1409 	struct perf_sample_data data;
1410 	unsigned int counters = mipspmu.num_counters;
1411 	u64 counter;
1412 	int handled = IRQ_NONE;
1413 	struct pt_regs *regs;
1414 
1415 	if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI))
1416 		return handled;
1417 	/*
1418 	 * First we pause the local counters, so that when we are locked
1419 	 * here, the counters are all paused. When it gets locked due to
1420 	 * perf_disable(), the timer interrupt handler will be delayed.
1421 	 *
1422 	 * See also mipsxx_pmu_start().
1423 	 */
1424 	pause_local_counters();
1425 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1426 	read_lock(&pmuint_rwlock);
1427 #endif
1428 
1429 	regs = get_irq_regs();
1430 
1431 	perf_sample_data_init(&data, 0, 0);
1432 
1433 	switch (counters) {
1434 #define HANDLE_COUNTER(n)						\
1435 	case n + 1:							\
1436 		if (test_bit(n, cpuc->used_mask)) {			\
1437 			counter = mipspmu.read_counter(n);		\
1438 			if (counter & mipspmu.overflow) {		\
1439 				handle_associated_event(cpuc, n, &data, regs); \
1440 				handled = IRQ_HANDLED;			\
1441 			}						\
1442 		}
1443 	HANDLE_COUNTER(3)
1444 	HANDLE_COUNTER(2)
1445 	HANDLE_COUNTER(1)
1446 	HANDLE_COUNTER(0)
1447 	}
1448 
1449 	/*
1450 	 * Do all the work for the pending perf events. We can do this
1451 	 * in here because the performance counter interrupt is a regular
1452 	 * interrupt, not NMI.
1453 	 */
1454 	if (handled == IRQ_HANDLED)
1455 		irq_work_run();
1456 
1457 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1458 	read_unlock(&pmuint_rwlock);
1459 #endif
1460 	resume_local_counters();
1461 	return handled;
1462 }
1463 
1464 static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1465 {
1466 	return mipsxx_pmu_handle_shared_irq();
1467 }
1468 
1469 /* 24K */
1470 #define IS_BOTH_COUNTERS_24K_EVENT(b)					\
1471 	((b) == 0 || (b) == 1 || (b) == 11)
1472 
1473 /* 34K */
1474 #define IS_BOTH_COUNTERS_34K_EVENT(b)					\
1475 	((b) == 0 || (b) == 1 || (b) == 11)
1476 #ifdef CONFIG_MIPS_MT_SMP
1477 #define IS_RANGE_P_34K_EVENT(r, b)					\
1478 	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1479 	 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 ||		\
1480 	 (r) == 176 || ((b) >= 50 && (b) <= 55) ||			\
1481 	 ((b) >= 64 && (b) <= 67))
1482 #define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
1483 #endif
1484 
1485 /* 74K */
1486 #define IS_BOTH_COUNTERS_74K_EVENT(b)					\
1487 	((b) == 0 || (b) == 1)
1488 
1489 /* proAptiv */
1490 #define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b)				\
1491 	((b) == 0 || (b) == 1)
1492 /* P5600 */
1493 #define IS_BOTH_COUNTERS_P5600_EVENT(b)					\
1494 	((b) == 0 || (b) == 1)
1495 
1496 /* 1004K */
1497 #define IS_BOTH_COUNTERS_1004K_EVENT(b)					\
1498 	((b) == 0 || (b) == 1 || (b) == 11)
1499 #ifdef CONFIG_MIPS_MT_SMP
1500 #define IS_RANGE_P_1004K_EVENT(r, b)					\
1501 	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1502 	 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 ||		\
1503 	 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) ||	\
1504 	 (r) == 188 || (b) == 61 || (b) == 62 ||			\
1505 	 ((b) >= 64 && (b) <= 67))
1506 #define IS_RANGE_V_1004K_EVENT(r)	((r) == 47)
1507 #endif
1508 
1509 /* interAptiv */
1510 #define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b)				\
1511 	((b) == 0 || (b) == 1 || (b) == 11)
1512 #ifdef CONFIG_MIPS_MT_SMP
1513 /* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */
1514 #define IS_RANGE_P_INTERAPTIV_EVENT(r, b)				\
1515 	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1516 	 (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 ||		\
1517 	 (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 &&		\
1518 	 (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 ||		\
1519 	 ((b) >= 64 && (b) <= 67))
1520 #define IS_RANGE_V_INTERAPTIV_EVENT(r)	((r) == 47 || (r) == 175)
1521 #endif
1522 
1523 /* BMIPS5000 */
1524 #define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b)				\
1525 	((b) == 0 || (b) == 1)
1526 
1527 
1528 /*
1529  * For most cores the user can use 0-255 raw events, where 0-127 for the events
1530  * of even counters, and 128-255 for odd counters. Note that bit 7 is used to
1531  * indicate the even/odd bank selector. So, for example, when user wants to take
1532  * the Event Num of 15 for odd counters (by referring to the user manual), then
1533  * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F)
1534  * to be used.
1535  *
1536  * Some newer cores have even more events, in which case the user can use raw
1537  * events 0-511, where 0-255 are for the events of even counters, and 256-511
1538  * are for odd counters, so bit 8 is used to indicate the even/odd bank selector.
1539  */
1540 static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1541 {
1542 	/* currently most cores have 7-bit event numbers */
1543 	unsigned int raw_id = config & 0xff;
1544 	unsigned int base_id = raw_id & 0x7f;
1545 
1546 	switch (current_cpu_type()) {
1547 	case CPU_24K:
1548 		if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1549 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1550 		else
1551 			raw_event.cntr_mask =
1552 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1553 #ifdef CONFIG_MIPS_MT_SMP
1554 		/*
1555 		 * This is actually doing nothing. Non-multithreading
1556 		 * CPUs will not check and calculate the range.
1557 		 */
1558 		raw_event.range = P;
1559 #endif
1560 		break;
1561 	case CPU_34K:
1562 		if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1563 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1564 		else
1565 			raw_event.cntr_mask =
1566 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1567 #ifdef CONFIG_MIPS_MT_SMP
1568 		if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1569 			raw_event.range = P;
1570 		else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1571 			raw_event.range = V;
1572 		else
1573 			raw_event.range = T;
1574 #endif
1575 		break;
1576 	case CPU_74K:
1577 	case CPU_1074K:
1578 		if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1579 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1580 		else
1581 			raw_event.cntr_mask =
1582 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1583 #ifdef CONFIG_MIPS_MT_SMP
1584 		raw_event.range = P;
1585 #endif
1586 		break;
1587 	case CPU_PROAPTIV:
1588 		if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id))
1589 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1590 		else
1591 			raw_event.cntr_mask =
1592 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1593 #ifdef CONFIG_MIPS_MT_SMP
1594 		raw_event.range = P;
1595 #endif
1596 		break;
1597 	case CPU_P5600:
1598 	case CPU_P6600:
1599 	case CPU_I6400:
1600 		/* 8-bit event numbers */
1601 		raw_id = config & 0x1ff;
1602 		base_id = raw_id & 0xff;
1603 		if (IS_BOTH_COUNTERS_P5600_EVENT(base_id))
1604 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1605 		else
1606 			raw_event.cntr_mask =
1607 				raw_id > 255 ? CNTR_ODD : CNTR_EVEN;
1608 #ifdef CONFIG_MIPS_MT_SMP
1609 		raw_event.range = P;
1610 #endif
1611 		break;
1612 	case CPU_1004K:
1613 		if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1614 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1615 		else
1616 			raw_event.cntr_mask =
1617 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1618 #ifdef CONFIG_MIPS_MT_SMP
1619 		if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1620 			raw_event.range = P;
1621 		else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1622 			raw_event.range = V;
1623 		else
1624 			raw_event.range = T;
1625 #endif
1626 		break;
1627 	case CPU_INTERAPTIV:
1628 		if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id))
1629 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1630 		else
1631 			raw_event.cntr_mask =
1632 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1633 #ifdef CONFIG_MIPS_MT_SMP
1634 		if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id))
1635 			raw_event.range = P;
1636 		else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id)))
1637 			raw_event.range = V;
1638 		else
1639 			raw_event.range = T;
1640 #endif
1641 		break;
1642 	case CPU_BMIPS5000:
1643 		if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id))
1644 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1645 		else
1646 			raw_event.cntr_mask =
1647 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1648 		break;
1649 	case CPU_LOONGSON3:
1650 		raw_event.cntr_mask = raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1651 	break;
1652 	}
1653 
1654 	raw_event.event_id = base_id;
1655 
1656 	return &raw_event;
1657 }
1658 
1659 static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1660 {
1661 	unsigned int raw_id = config & 0xff;
1662 	unsigned int base_id = raw_id & 0x7f;
1663 
1664 
1665 	raw_event.cntr_mask = CNTR_ALL;
1666 	raw_event.event_id = base_id;
1667 
1668 	if (current_cpu_type() == CPU_CAVIUM_OCTEON2) {
1669 		if (base_id > 0x42)
1670 			return ERR_PTR(-EOPNOTSUPP);
1671 	} else {
1672 		if (base_id > 0x3a)
1673 			return ERR_PTR(-EOPNOTSUPP);
1674 	}
1675 
1676 	switch (base_id) {
1677 	case 0x00:
1678 	case 0x0f:
1679 	case 0x1e:
1680 	case 0x1f:
1681 	case 0x2f:
1682 	case 0x34:
1683 	case 0x3b ... 0x3f:
1684 		return ERR_PTR(-EOPNOTSUPP);
1685 	default:
1686 		break;
1687 	}
1688 
1689 	return &raw_event;
1690 }
1691 
1692 static const struct mips_perf_event *xlp_pmu_map_raw_event(u64 config)
1693 {
1694 	unsigned int raw_id = config & 0xff;
1695 
1696 	/* Only 1-63 are defined */
1697 	if ((raw_id < 0x01) || (raw_id > 0x3f))
1698 		return ERR_PTR(-EOPNOTSUPP);
1699 
1700 	raw_event.cntr_mask = CNTR_ALL;
1701 	raw_event.event_id = raw_id;
1702 
1703 	return &raw_event;
1704 }
1705 
1706 static int __init
1707 init_hw_perf_events(void)
1708 {
1709 	int counters, irq;
1710 	int counter_bits;
1711 
1712 	pr_info("Performance counters: ");
1713 
1714 	counters = n_counters();
1715 	if (counters == 0) {
1716 		pr_cont("No available PMU.\n");
1717 		return -ENODEV;
1718 	}
1719 
1720 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1721 	cpu_has_mipsmt_pertccounters = read_c0_config7() & (1<<19);
1722 	if (!cpu_has_mipsmt_pertccounters)
1723 		counters = counters_total_to_per_cpu(counters);
1724 #endif
1725 
1726 	if (get_c0_perfcount_int)
1727 		irq = get_c0_perfcount_int();
1728 	else if (cp0_perfcount_irq >= 0)
1729 		irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1730 	else
1731 		irq = -1;
1732 
1733 	mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1734 
1735 	switch (current_cpu_type()) {
1736 	case CPU_24K:
1737 		mipspmu.name = "mips/24K";
1738 		mipspmu.general_event_map = &mipsxxcore_event_map;
1739 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1740 		break;
1741 	case CPU_34K:
1742 		mipspmu.name = "mips/34K";
1743 		mipspmu.general_event_map = &mipsxxcore_event_map;
1744 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1745 		break;
1746 	case CPU_74K:
1747 		mipspmu.name = "mips/74K";
1748 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1749 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1750 		break;
1751 	case CPU_PROAPTIV:
1752 		mipspmu.name = "mips/proAptiv";
1753 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1754 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1755 		break;
1756 	case CPU_P5600:
1757 		mipspmu.name = "mips/P5600";
1758 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1759 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1760 		break;
1761 	case CPU_P6600:
1762 		mipspmu.name = "mips/P6600";
1763 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1764 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1765 		break;
1766 	case CPU_I6400:
1767 		mipspmu.name = "mips/I6400";
1768 		mipspmu.general_event_map = &i6400_event_map;
1769 		mipspmu.cache_event_map = &i6400_cache_map;
1770 		break;
1771 	case CPU_1004K:
1772 		mipspmu.name = "mips/1004K";
1773 		mipspmu.general_event_map = &mipsxxcore_event_map;
1774 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1775 		break;
1776 	case CPU_1074K:
1777 		mipspmu.name = "mips/1074K";
1778 		mipspmu.general_event_map = &mipsxxcore_event_map;
1779 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1780 		break;
1781 	case CPU_INTERAPTIV:
1782 		mipspmu.name = "mips/interAptiv";
1783 		mipspmu.general_event_map = &mipsxxcore_event_map;
1784 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1785 		break;
1786 	case CPU_LOONGSON1:
1787 		mipspmu.name = "mips/loongson1";
1788 		mipspmu.general_event_map = &mipsxxcore_event_map;
1789 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1790 		break;
1791 	case CPU_LOONGSON3:
1792 		mipspmu.name = "mips/loongson3";
1793 		mipspmu.general_event_map = &loongson3_event_map;
1794 		mipspmu.cache_event_map = &loongson3_cache_map;
1795 		break;
1796 	case CPU_CAVIUM_OCTEON:
1797 	case CPU_CAVIUM_OCTEON_PLUS:
1798 	case CPU_CAVIUM_OCTEON2:
1799 		mipspmu.name = "octeon";
1800 		mipspmu.general_event_map = &octeon_event_map;
1801 		mipspmu.cache_event_map = &octeon_cache_map;
1802 		mipspmu.map_raw_event = octeon_pmu_map_raw_event;
1803 		break;
1804 	case CPU_BMIPS5000:
1805 		mipspmu.name = "BMIPS5000";
1806 		mipspmu.general_event_map = &bmips5000_event_map;
1807 		mipspmu.cache_event_map = &bmips5000_cache_map;
1808 		break;
1809 	case CPU_XLP:
1810 		mipspmu.name = "xlp";
1811 		mipspmu.general_event_map = &xlp_event_map;
1812 		mipspmu.cache_event_map = &xlp_cache_map;
1813 		mipspmu.map_raw_event = xlp_pmu_map_raw_event;
1814 		break;
1815 	default:
1816 		pr_cont("Either hardware does not support performance "
1817 			"counters, or not yet implemented.\n");
1818 		return -ENODEV;
1819 	}
1820 
1821 	mipspmu.num_counters = counters;
1822 	mipspmu.irq = irq;
1823 
1824 	if (read_c0_perfctrl0() & MIPS_PERFCTRL_W) {
1825 		mipspmu.max_period = (1ULL << 63) - 1;
1826 		mipspmu.valid_count = (1ULL << 63) - 1;
1827 		mipspmu.overflow = 1ULL << 63;
1828 		mipspmu.read_counter = mipsxx_pmu_read_counter_64;
1829 		mipspmu.write_counter = mipsxx_pmu_write_counter_64;
1830 		counter_bits = 64;
1831 	} else {
1832 		mipspmu.max_period = (1ULL << 31) - 1;
1833 		mipspmu.valid_count = (1ULL << 31) - 1;
1834 		mipspmu.overflow = 1ULL << 31;
1835 		mipspmu.read_counter = mipsxx_pmu_read_counter;
1836 		mipspmu.write_counter = mipsxx_pmu_write_counter;
1837 		counter_bits = 32;
1838 	}
1839 
1840 	on_each_cpu(reset_counters, (void *)(long)counters, 1);
1841 
1842 	pr_cont("%s PMU enabled, %d %d-bit counters available to each "
1843 		"CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
1844 		irq < 0 ? " (share with timer interrupt)" : "");
1845 
1846 	perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1847 
1848 	return 0;
1849 }
1850 early_initcall(init_hw_perf_events);
1851