xref: /linux/arch/mips/kernel/perf_event_mipsxx.c (revision 4f2c0a4acffbec01079c28f839422e64ddeff004)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Linux performance counter support for MIPS.
4  *
5  * Copyright (C) 2010 MIPS Technologies, Inc.
6  * Copyright (C) 2011 Cavium Networks, Inc.
7  * Author: Deng-Cheng Zhu
8  *
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.
13  */
14 
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>
21 
22 #include <asm/irq.h>
23 #include <asm/irq_regs.h>
24 #include <asm/stacktrace.h>
25 #include <asm/time.h> /* For perf_irq */
26 
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)
30 
31 struct cpu_hw_events {
32 	/* Array of events on this cpu. */
33 	struct perf_event	*events[MIPS_MAX_HWEVENTS];
34 
35 	/*
36 	 * Set the bit (indexed by the counter number) when the counter
37 	 * is used for an event.
38 	 */
39 	unsigned long		used_mask[BITS_TO_LONGS(MIPS_MAX_HWEVENTS)];
40 
41 	/*
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.
45 	 */
46 	unsigned int		saved_ctrl[MIPS_MAX_HWEVENTS];
47 };
48 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
49 	.saved_ctrl = {0},
50 };
51 
52 /* The description of MIPS performance events. */
53 struct mips_perf_event {
54 	unsigned int event_id;
55 	/*
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.
59 	 */
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;
69 };
70 
71 static struct mips_perf_event raw_event;
72 static DEFINE_MUTEX(raw_event_mutex);
73 
74 #define C(x) PERF_COUNT_HW_CACHE_##x
75 
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;
91 };
92 
93 static int counter_bits;
94 static struct mips_pmu mipspmu;
95 
96 #define M_PERFCTL_EVENT(event)		(((event) << MIPS_PERFCTRL_EVENT_S) & \
97 					 MIPS_PERFCTRL_EVENT)
98 #define M_PERFCTL_VPEID(vpe)		((vpe)	  << MIPS_PERFCTRL_VPEID_S)
99 
100 #ifdef CONFIG_CPU_BMIPS5000
101 #define M_PERFCTL_MT_EN(filter)		0
102 #else /* !CONFIG_CPU_BMIPS5000 */
103 #define M_PERFCTL_MT_EN(filter)		(filter)
104 #endif /* CONFIG_CPU_BMIPS5000 */
105 
106 #define	   M_TC_EN_ALL			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_ALL)
107 #define	   M_TC_EN_VPE			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_VPE)
108 #define	   M_TC_EN_TC			M_PERFCTL_MT_EN(MIPS_PERFCTRL_MT_EN_TC)
109 
110 #define M_PERFCTL_COUNT_EVENT_WHENEVER	(MIPS_PERFCTRL_EXL |		\
111 					 MIPS_PERFCTRL_K |		\
112 					 MIPS_PERFCTRL_U |		\
113 					 MIPS_PERFCTRL_S |		\
114 					 MIPS_PERFCTRL_IE)
115 
116 #ifdef CONFIG_MIPS_MT_SMP
117 #define M_PERFCTL_CONFIG_MASK		0x3fff801f
118 #else
119 #define M_PERFCTL_CONFIG_MASK		0x1f
120 #endif
121 
122 #define CNTR_BIT_MASK(n)	(((n) == 64) ? ~0ULL : ((1ULL<<(n))-1))
123 
124 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
125 static DEFINE_RWLOCK(pmuint_rwlock);
126 
127 #if defined(CONFIG_CPU_BMIPS5000)
128 #define vpe_id()	(cpu_has_mipsmt_pertccounters ? \
129 			 0 : (smp_processor_id() & MIPS_CPUID_TO_COUNTER_MASK))
130 #else
131 #define vpe_id()	(cpu_has_mipsmt_pertccounters ? \
132 			 0 : cpu_vpe_id(&current_cpu_data))
133 #endif
134 
135 /* Copied from op_model_mipsxx.c */
vpe_shift(void)136 static unsigned int vpe_shift(void)
137 {
138 	if (num_possible_cpus() > 1)
139 		return 1;
140 
141 	return 0;
142 }
143 
counters_total_to_per_cpu(unsigned int counters)144 static unsigned int counters_total_to_per_cpu(unsigned int counters)
145 {
146 	return counters >> vpe_shift();
147 }
148 
149 #else /* !CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
150 #define vpe_id()	0
151 
152 #endif /* CONFIG_MIPS_PERF_SHARED_TC_COUNTERS */
153 
154 static void resume_local_counters(void);
155 static void pause_local_counters(void);
156 static irqreturn_t mipsxx_pmu_handle_irq(int, void *);
157 static int mipsxx_pmu_handle_shared_irq(void);
158 
159 /* 0: Not Loongson-3
160  * 1: Loongson-3A1000/3B1000/3B1500
161  * 2: Loongson-3A2000/3A3000
162  * 3: Loongson-3A4000+
163  */
164 
165 #define LOONGSON_PMU_TYPE0 0
166 #define LOONGSON_PMU_TYPE1 1
167 #define LOONGSON_PMU_TYPE2 2
168 #define LOONGSON_PMU_TYPE3 3
169 
get_loongson3_pmu_type(void)170 static inline int get_loongson3_pmu_type(void)
171 {
172 	if (boot_cpu_type() != CPU_LOONGSON64)
173 		return LOONGSON_PMU_TYPE0;
174 	if ((boot_cpu_data.processor_id & PRID_COMP_MASK) == PRID_COMP_LEGACY)
175 		return LOONGSON_PMU_TYPE1;
176 	if ((boot_cpu_data.processor_id & PRID_IMP_MASK) == PRID_IMP_LOONGSON_64C)
177 		return LOONGSON_PMU_TYPE2;
178 	if ((boot_cpu_data.processor_id & PRID_IMP_MASK) == PRID_IMP_LOONGSON_64G)
179 		return LOONGSON_PMU_TYPE3;
180 
181 	return LOONGSON_PMU_TYPE0;
182 }
183 
mipsxx_pmu_swizzle_perf_idx(unsigned int idx)184 static unsigned int mipsxx_pmu_swizzle_perf_idx(unsigned int idx)
185 {
186 	if (vpe_id() == 1)
187 		idx = (idx + 2) & 3;
188 	return idx;
189 }
190 
mipsxx_pmu_read_counter(unsigned int idx)191 static u64 mipsxx_pmu_read_counter(unsigned int idx)
192 {
193 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
194 
195 	switch (idx) {
196 	case 0:
197 		/*
198 		 * The counters are unsigned, we must cast to truncate
199 		 * off the high bits.
200 		 */
201 		return (u32)read_c0_perfcntr0();
202 	case 1:
203 		return (u32)read_c0_perfcntr1();
204 	case 2:
205 		return (u32)read_c0_perfcntr2();
206 	case 3:
207 		return (u32)read_c0_perfcntr3();
208 	default:
209 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
210 		return 0;
211 	}
212 }
213 
mipsxx_pmu_read_counter_64(unsigned int idx)214 static u64 mipsxx_pmu_read_counter_64(unsigned int idx)
215 {
216 	u64 mask = CNTR_BIT_MASK(counter_bits);
217 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
218 
219 	switch (idx) {
220 	case 0:
221 		return read_c0_perfcntr0_64() & mask;
222 	case 1:
223 		return read_c0_perfcntr1_64() & mask;
224 	case 2:
225 		return read_c0_perfcntr2_64() & mask;
226 	case 3:
227 		return read_c0_perfcntr3_64() & mask;
228 	default:
229 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
230 		return 0;
231 	}
232 }
233 
mipsxx_pmu_write_counter(unsigned int idx,u64 val)234 static void mipsxx_pmu_write_counter(unsigned int idx, u64 val)
235 {
236 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
237 
238 	switch (idx) {
239 	case 0:
240 		write_c0_perfcntr0(val);
241 		return;
242 	case 1:
243 		write_c0_perfcntr1(val);
244 		return;
245 	case 2:
246 		write_c0_perfcntr2(val);
247 		return;
248 	case 3:
249 		write_c0_perfcntr3(val);
250 		return;
251 	}
252 }
253 
mipsxx_pmu_write_counter_64(unsigned int idx,u64 val)254 static void mipsxx_pmu_write_counter_64(unsigned int idx, u64 val)
255 {
256 	val &= CNTR_BIT_MASK(counter_bits);
257 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
258 
259 	switch (idx) {
260 	case 0:
261 		write_c0_perfcntr0_64(val);
262 		return;
263 	case 1:
264 		write_c0_perfcntr1_64(val);
265 		return;
266 	case 2:
267 		write_c0_perfcntr2_64(val);
268 		return;
269 	case 3:
270 		write_c0_perfcntr3_64(val);
271 		return;
272 	}
273 }
274 
mipsxx_pmu_read_control(unsigned int idx)275 static unsigned int mipsxx_pmu_read_control(unsigned int idx)
276 {
277 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
278 
279 	switch (idx) {
280 	case 0:
281 		return read_c0_perfctrl0();
282 	case 1:
283 		return read_c0_perfctrl1();
284 	case 2:
285 		return read_c0_perfctrl2();
286 	case 3:
287 		return read_c0_perfctrl3();
288 	default:
289 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
290 		return 0;
291 	}
292 }
293 
mipsxx_pmu_write_control(unsigned int idx,unsigned int val)294 static void mipsxx_pmu_write_control(unsigned int idx, unsigned int val)
295 {
296 	idx = mipsxx_pmu_swizzle_perf_idx(idx);
297 
298 	switch (idx) {
299 	case 0:
300 		write_c0_perfctrl0(val);
301 		return;
302 	case 1:
303 		write_c0_perfctrl1(val);
304 		return;
305 	case 2:
306 		write_c0_perfctrl2(val);
307 		return;
308 	case 3:
309 		write_c0_perfctrl3(val);
310 		return;
311 	}
312 }
313 
mipsxx_pmu_alloc_counter(struct cpu_hw_events * cpuc,struct hw_perf_event * hwc)314 static int mipsxx_pmu_alloc_counter(struct cpu_hw_events *cpuc,
315 				    struct hw_perf_event *hwc)
316 {
317 	int i;
318 	unsigned long cntr_mask;
319 
320 	/*
321 	 * We only need to care the counter mask. The range has been
322 	 * checked definitely.
323 	 */
324 	if (get_loongson3_pmu_type() == LOONGSON_PMU_TYPE2)
325 		cntr_mask = (hwc->event_base >> 10) & 0xffff;
326 	else
327 		cntr_mask = (hwc->event_base >> 8) & 0xffff;
328 
329 	for (i = mipspmu.num_counters - 1; i >= 0; i--) {
330 		/*
331 		 * Note that some MIPS perf events can be counted by both
332 		 * even and odd counters, whereas many other are only by
333 		 * even _or_ odd counters. This introduces an issue that
334 		 * when the former kind of event takes the counter the
335 		 * latter kind of event wants to use, then the "counter
336 		 * allocation" for the latter event will fail. In fact if
337 		 * they can be dynamically swapped, they both feel happy.
338 		 * But here we leave this issue alone for now.
339 		 */
340 		if (test_bit(i, &cntr_mask) &&
341 			!test_and_set_bit(i, cpuc->used_mask))
342 			return i;
343 	}
344 
345 	return -EAGAIN;
346 }
347 
mipsxx_pmu_enable_event(struct hw_perf_event * evt,int idx)348 static void mipsxx_pmu_enable_event(struct hw_perf_event *evt, int idx)
349 {
350 	struct perf_event *event = container_of(evt, struct perf_event, hw);
351 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
352 	unsigned int range = evt->event_base >> 24;
353 
354 	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
355 
356 	if (get_loongson3_pmu_type() == LOONGSON_PMU_TYPE2)
357 		cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0x3ff) |
358 			(evt->config_base & M_PERFCTL_CONFIG_MASK) |
359 			/* Make sure interrupt enabled. */
360 			MIPS_PERFCTRL_IE;
361 	else
362 		cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
363 			(evt->config_base & M_PERFCTL_CONFIG_MASK) |
364 			/* Make sure interrupt enabled. */
365 			MIPS_PERFCTRL_IE;
366 
367 	if (IS_ENABLED(CONFIG_CPU_BMIPS5000)) {
368 		/* enable the counter for the calling thread */
369 		cpuc->saved_ctrl[idx] |=
370 			(1 << (12 + vpe_id())) | BRCM_PERFCTRL_TC;
371 	} else if (IS_ENABLED(CONFIG_MIPS_MT_SMP) && range > V) {
372 		/* The counter is processor wide. Set it up to count all TCs. */
373 		pr_debug("Enabling perf counter for all TCs\n");
374 		cpuc->saved_ctrl[idx] |= M_TC_EN_ALL;
375 	} else {
376 		unsigned int cpu, ctrl;
377 
378 		/*
379 		 * Set up the counter for a particular CPU when event->cpu is
380 		 * a valid CPU number. Otherwise set up the counter for the CPU
381 		 * scheduling this thread.
382 		 */
383 		cpu = (event->cpu >= 0) ? event->cpu : smp_processor_id();
384 
385 		ctrl = M_PERFCTL_VPEID(cpu_vpe_id(&cpu_data[cpu]));
386 		ctrl |= M_TC_EN_VPE;
387 		cpuc->saved_ctrl[idx] |= ctrl;
388 		pr_debug("Enabling perf counter for CPU%d\n", cpu);
389 	}
390 	/*
391 	 * We do not actually let the counter run. Leave it until start().
392 	 */
393 }
394 
mipsxx_pmu_disable_event(int idx)395 static void mipsxx_pmu_disable_event(int idx)
396 {
397 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
398 	unsigned long flags;
399 
400 	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
401 
402 	local_irq_save(flags);
403 	cpuc->saved_ctrl[idx] = mipsxx_pmu_read_control(idx) &
404 		~M_PERFCTL_COUNT_EVENT_WHENEVER;
405 	mipsxx_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
406 	local_irq_restore(flags);
407 }
408 
mipspmu_event_set_period(struct perf_event * event,struct hw_perf_event * hwc,int idx)409 static int mipspmu_event_set_period(struct perf_event *event,
410 				    struct hw_perf_event *hwc,
411 				    int idx)
412 {
413 	u64 left = local64_read(&hwc->period_left);
414 	u64 period = hwc->sample_period;
415 	int ret = 0;
416 
417 	if (unlikely((left + period) & (1ULL << 63))) {
418 		/* left underflowed by more than period. */
419 		left = period;
420 		local64_set(&hwc->period_left, left);
421 		hwc->last_period = period;
422 		ret = 1;
423 	} else	if (unlikely((left + period) <= period)) {
424 		/* left underflowed by less than period. */
425 		left += period;
426 		local64_set(&hwc->period_left, left);
427 		hwc->last_period = period;
428 		ret = 1;
429 	}
430 
431 	if (left > mipspmu.max_period) {
432 		left = mipspmu.max_period;
433 		local64_set(&hwc->period_left, left);
434 	}
435 
436 	local64_set(&hwc->prev_count, mipspmu.overflow - left);
437 
438 	if (get_loongson3_pmu_type() == LOONGSON_PMU_TYPE2)
439 		mipsxx_pmu_write_control(idx,
440 				M_PERFCTL_EVENT(hwc->event_base & 0x3ff));
441 
442 	mipspmu.write_counter(idx, mipspmu.overflow - left);
443 
444 	perf_event_update_userpage(event);
445 
446 	return ret;
447 }
448 
mipspmu_event_update(struct perf_event * event,struct hw_perf_event * hwc,int idx)449 static void mipspmu_event_update(struct perf_event *event,
450 				 struct hw_perf_event *hwc,
451 				 int idx)
452 {
453 	u64 prev_raw_count, new_raw_count;
454 	u64 delta;
455 
456 again:
457 	prev_raw_count = local64_read(&hwc->prev_count);
458 	new_raw_count = mipspmu.read_counter(idx);
459 
460 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
461 				new_raw_count) != prev_raw_count)
462 		goto again;
463 
464 	delta = new_raw_count - prev_raw_count;
465 
466 	local64_add(delta, &event->count);
467 	local64_sub(delta, &hwc->period_left);
468 }
469 
mipspmu_start(struct perf_event * event,int flags)470 static void mipspmu_start(struct perf_event *event, int flags)
471 {
472 	struct hw_perf_event *hwc = &event->hw;
473 
474 	if (flags & PERF_EF_RELOAD)
475 		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
476 
477 	hwc->state = 0;
478 
479 	/* Set the period for the event. */
480 	mipspmu_event_set_period(event, hwc, hwc->idx);
481 
482 	/* Enable the event. */
483 	mipsxx_pmu_enable_event(hwc, hwc->idx);
484 }
485 
mipspmu_stop(struct perf_event * event,int flags)486 static void mipspmu_stop(struct perf_event *event, int flags)
487 {
488 	struct hw_perf_event *hwc = &event->hw;
489 
490 	if (!(hwc->state & PERF_HES_STOPPED)) {
491 		/* We are working on a local event. */
492 		mipsxx_pmu_disable_event(hwc->idx);
493 		barrier();
494 		mipspmu_event_update(event, hwc, hwc->idx);
495 		hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
496 	}
497 }
498 
mipspmu_add(struct perf_event * event,int flags)499 static int mipspmu_add(struct perf_event *event, int flags)
500 {
501 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
502 	struct hw_perf_event *hwc = &event->hw;
503 	int idx;
504 	int err = 0;
505 
506 	perf_pmu_disable(event->pmu);
507 
508 	/* To look for a free counter for this event. */
509 	idx = mipsxx_pmu_alloc_counter(cpuc, hwc);
510 	if (idx < 0) {
511 		err = idx;
512 		goto out;
513 	}
514 
515 	/*
516 	 * If there is an event in the counter we are going to use then
517 	 * make sure it is disabled.
518 	 */
519 	event->hw.idx = idx;
520 	mipsxx_pmu_disable_event(idx);
521 	cpuc->events[idx] = event;
522 
523 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
524 	if (flags & PERF_EF_START)
525 		mipspmu_start(event, PERF_EF_RELOAD);
526 
527 	/* Propagate our changes to the userspace mapping. */
528 	perf_event_update_userpage(event);
529 
530 out:
531 	perf_pmu_enable(event->pmu);
532 	return err;
533 }
534 
mipspmu_del(struct perf_event * event,int flags)535 static void mipspmu_del(struct perf_event *event, int flags)
536 {
537 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
538 	struct hw_perf_event *hwc = &event->hw;
539 	int idx = hwc->idx;
540 
541 	WARN_ON(idx < 0 || idx >= mipspmu.num_counters);
542 
543 	mipspmu_stop(event, PERF_EF_UPDATE);
544 	cpuc->events[idx] = NULL;
545 	clear_bit(idx, cpuc->used_mask);
546 
547 	perf_event_update_userpage(event);
548 }
549 
mipspmu_read(struct perf_event * event)550 static void mipspmu_read(struct perf_event *event)
551 {
552 	struct hw_perf_event *hwc = &event->hw;
553 
554 	/* Don't read disabled counters! */
555 	if (hwc->idx < 0)
556 		return;
557 
558 	mipspmu_event_update(event, hwc, hwc->idx);
559 }
560 
mipspmu_enable(struct pmu * pmu)561 static void mipspmu_enable(struct pmu *pmu)
562 {
563 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
564 	write_unlock(&pmuint_rwlock);
565 #endif
566 	resume_local_counters();
567 }
568 
569 /*
570  * MIPS performance counters can be per-TC. The control registers can
571  * not be directly accessed across CPUs. Hence if we want to do global
572  * control, we need cross CPU calls. on_each_cpu() can help us, but we
573  * can not make sure this function is called with interrupts enabled. So
574  * here we pause local counters and then grab a rwlock and leave the
575  * counters on other CPUs alone. If any counter interrupt raises while
576  * we own the write lock, simply pause local counters on that CPU and
577  * spin in the handler. Also we know we won't be switched to another
578  * CPU after pausing local counters and before grabbing the lock.
579  */
mipspmu_disable(struct pmu * pmu)580 static void mipspmu_disable(struct pmu *pmu)
581 {
582 	pause_local_counters();
583 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
584 	write_lock(&pmuint_rwlock);
585 #endif
586 }
587 
588 static atomic_t active_events = ATOMIC_INIT(0);
589 static DEFINE_MUTEX(pmu_reserve_mutex);
590 static int (*save_perf_irq)(void);
591 
mipspmu_get_irq(void)592 static int mipspmu_get_irq(void)
593 {
594 	int err;
595 
596 	if (mipspmu.irq >= 0) {
597 		/* Request my own irq handler. */
598 		err = request_irq(mipspmu.irq, mipsxx_pmu_handle_irq,
599 				  IRQF_PERCPU | IRQF_NOBALANCING |
600 				  IRQF_NO_THREAD | IRQF_NO_SUSPEND |
601 				  IRQF_SHARED,
602 				  "mips_perf_pmu", &mipspmu);
603 		if (err) {
604 			pr_warn("Unable to request IRQ%d for MIPS performance counters!\n",
605 				mipspmu.irq);
606 		}
607 	} else if (cp0_perfcount_irq < 0) {
608 		/*
609 		 * We are sharing the irq number with the timer interrupt.
610 		 */
611 		save_perf_irq = perf_irq;
612 		perf_irq = mipsxx_pmu_handle_shared_irq;
613 		err = 0;
614 	} else {
615 		pr_warn("The platform hasn't properly defined its interrupt controller\n");
616 		err = -ENOENT;
617 	}
618 
619 	return err;
620 }
621 
mipspmu_free_irq(void)622 static void mipspmu_free_irq(void)
623 {
624 	if (mipspmu.irq >= 0)
625 		free_irq(mipspmu.irq, &mipspmu);
626 	else if (cp0_perfcount_irq < 0)
627 		perf_irq = save_perf_irq;
628 }
629 
630 /*
631  * mipsxx/rm9000/loongson2 have different performance counters, they have
632  * specific low-level init routines.
633  */
634 static void reset_counters(void *arg);
635 static int __hw_perf_event_init(struct perf_event *event);
636 
hw_perf_event_destroy(struct perf_event * event)637 static void hw_perf_event_destroy(struct perf_event *event)
638 {
639 	if (atomic_dec_and_mutex_lock(&active_events,
640 				&pmu_reserve_mutex)) {
641 		/*
642 		 * We must not call the destroy function with interrupts
643 		 * disabled.
644 		 */
645 		on_each_cpu(reset_counters,
646 			(void *)(long)mipspmu.num_counters, 1);
647 		mipspmu_free_irq();
648 		mutex_unlock(&pmu_reserve_mutex);
649 	}
650 }
651 
mipspmu_event_init(struct perf_event * event)652 static int mipspmu_event_init(struct perf_event *event)
653 {
654 	int err = 0;
655 
656 	/* does not support taken branch sampling */
657 	if (has_branch_stack(event))
658 		return -EOPNOTSUPP;
659 
660 	switch (event->attr.type) {
661 	case PERF_TYPE_RAW:
662 	case PERF_TYPE_HARDWARE:
663 	case PERF_TYPE_HW_CACHE:
664 		break;
665 
666 	default:
667 		return -ENOENT;
668 	}
669 
670 	if (event->cpu >= 0 && !cpu_online(event->cpu))
671 		return -ENODEV;
672 
673 	if (!atomic_inc_not_zero(&active_events)) {
674 		mutex_lock(&pmu_reserve_mutex);
675 		if (atomic_read(&active_events) == 0)
676 			err = mipspmu_get_irq();
677 
678 		if (!err)
679 			atomic_inc(&active_events);
680 		mutex_unlock(&pmu_reserve_mutex);
681 	}
682 
683 	if (err)
684 		return err;
685 
686 	return __hw_perf_event_init(event);
687 }
688 
689 static struct pmu pmu = {
690 	.pmu_enable	= mipspmu_enable,
691 	.pmu_disable	= mipspmu_disable,
692 	.event_init	= mipspmu_event_init,
693 	.add		= mipspmu_add,
694 	.del		= mipspmu_del,
695 	.start		= mipspmu_start,
696 	.stop		= mipspmu_stop,
697 	.read		= mipspmu_read,
698 };
699 
mipspmu_perf_event_encode(const struct mips_perf_event * pev)700 static unsigned int mipspmu_perf_event_encode(const struct mips_perf_event *pev)
701 {
702 /*
703  * Top 8 bits for range, next 16 bits for cntr_mask, lowest 8 bits for
704  * event_id.
705  */
706 #ifdef CONFIG_MIPS_MT_SMP
707 	if (num_possible_cpus() > 1)
708 		return ((unsigned int)pev->range << 24) |
709 			(pev->cntr_mask & 0xffff00) |
710 			(pev->event_id & 0xff);
711 	else
712 #endif /* CONFIG_MIPS_MT_SMP */
713 	{
714 		if (get_loongson3_pmu_type() == LOONGSON_PMU_TYPE2)
715 			return (pev->cntr_mask & 0xfffc00) |
716 				(pev->event_id & 0x3ff);
717 		else
718 			return (pev->cntr_mask & 0xffff00) |
719 				(pev->event_id & 0xff);
720 	}
721 }
722 
mipspmu_map_general_event(int idx)723 static const struct mips_perf_event *mipspmu_map_general_event(int idx)
724 {
725 
726 	if ((*mipspmu.general_event_map)[idx].cntr_mask == 0)
727 		return ERR_PTR(-EOPNOTSUPP);
728 	return &(*mipspmu.general_event_map)[idx];
729 }
730 
mipspmu_map_cache_event(u64 config)731 static const struct mips_perf_event *mipspmu_map_cache_event(u64 config)
732 {
733 	unsigned int cache_type, cache_op, cache_result;
734 	const struct mips_perf_event *pev;
735 
736 	cache_type = (config >> 0) & 0xff;
737 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
738 		return ERR_PTR(-EINVAL);
739 
740 	cache_op = (config >> 8) & 0xff;
741 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
742 		return ERR_PTR(-EINVAL);
743 
744 	cache_result = (config >> 16) & 0xff;
745 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
746 		return ERR_PTR(-EINVAL);
747 
748 	pev = &((*mipspmu.cache_event_map)
749 					[cache_type]
750 					[cache_op]
751 					[cache_result]);
752 
753 	if (pev->cntr_mask == 0)
754 		return ERR_PTR(-EOPNOTSUPP);
755 
756 	return pev;
757 
758 }
759 
validate_group(struct perf_event * event)760 static int validate_group(struct perf_event *event)
761 {
762 	struct perf_event *sibling, *leader = event->group_leader;
763 	struct cpu_hw_events fake_cpuc;
764 
765 	memset(&fake_cpuc, 0, sizeof(fake_cpuc));
766 
767 	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
768 		return -EINVAL;
769 
770 	for_each_sibling_event(sibling, leader) {
771 		if (mipsxx_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
772 			return -EINVAL;
773 	}
774 
775 	if (mipsxx_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
776 		return -EINVAL;
777 
778 	return 0;
779 }
780 
781 /* This is needed by specific irq handlers in perf_event_*.c */
handle_associated_event(struct cpu_hw_events * cpuc,int idx,struct perf_sample_data * data,struct pt_regs * regs)782 static void handle_associated_event(struct cpu_hw_events *cpuc,
783 				    int idx, struct perf_sample_data *data,
784 				    struct pt_regs *regs)
785 {
786 	struct perf_event *event = cpuc->events[idx];
787 	struct hw_perf_event *hwc = &event->hw;
788 
789 	mipspmu_event_update(event, hwc, idx);
790 	data->period = event->hw.last_period;
791 	if (!mipspmu_event_set_period(event, hwc, idx))
792 		return;
793 
794 	if (perf_event_overflow(event, data, regs))
795 		mipsxx_pmu_disable_event(idx);
796 }
797 
798 
__n_counters(void)799 static int __n_counters(void)
800 {
801 	if (!cpu_has_perf)
802 		return 0;
803 	if (!(read_c0_perfctrl0() & MIPS_PERFCTRL_M))
804 		return 1;
805 	if (!(read_c0_perfctrl1() & MIPS_PERFCTRL_M))
806 		return 2;
807 	if (!(read_c0_perfctrl2() & MIPS_PERFCTRL_M))
808 		return 3;
809 
810 	return 4;
811 }
812 
n_counters(void)813 static int n_counters(void)
814 {
815 	int counters;
816 
817 	switch (current_cpu_type()) {
818 	case CPU_R10000:
819 		counters = 2;
820 		break;
821 
822 	case CPU_R12000:
823 	case CPU_R14000:
824 	case CPU_R16000:
825 		counters = 4;
826 		break;
827 
828 	default:
829 		counters = __n_counters();
830 	}
831 
832 	return counters;
833 }
834 
loongson3_reset_counters(void * arg)835 static void loongson3_reset_counters(void *arg)
836 {
837 	int counters = (int)(long)arg;
838 
839 	switch (counters) {
840 	case 4:
841 		mipsxx_pmu_write_control(3, 0);
842 		mipspmu.write_counter(3, 0);
843 		mipsxx_pmu_write_control(3, 127<<5);
844 		mipspmu.write_counter(3, 0);
845 		mipsxx_pmu_write_control(3, 191<<5);
846 		mipspmu.write_counter(3, 0);
847 		mipsxx_pmu_write_control(3, 255<<5);
848 		mipspmu.write_counter(3, 0);
849 		mipsxx_pmu_write_control(3, 319<<5);
850 		mipspmu.write_counter(3, 0);
851 		mipsxx_pmu_write_control(3, 383<<5);
852 		mipspmu.write_counter(3, 0);
853 		mipsxx_pmu_write_control(3, 575<<5);
854 		mipspmu.write_counter(3, 0);
855 		fallthrough;
856 	case 3:
857 		mipsxx_pmu_write_control(2, 0);
858 		mipspmu.write_counter(2, 0);
859 		mipsxx_pmu_write_control(2, 127<<5);
860 		mipspmu.write_counter(2, 0);
861 		mipsxx_pmu_write_control(2, 191<<5);
862 		mipspmu.write_counter(2, 0);
863 		mipsxx_pmu_write_control(2, 255<<5);
864 		mipspmu.write_counter(2, 0);
865 		mipsxx_pmu_write_control(2, 319<<5);
866 		mipspmu.write_counter(2, 0);
867 		mipsxx_pmu_write_control(2, 383<<5);
868 		mipspmu.write_counter(2, 0);
869 		mipsxx_pmu_write_control(2, 575<<5);
870 		mipspmu.write_counter(2, 0);
871 		fallthrough;
872 	case 2:
873 		mipsxx_pmu_write_control(1, 0);
874 		mipspmu.write_counter(1, 0);
875 		mipsxx_pmu_write_control(1, 127<<5);
876 		mipspmu.write_counter(1, 0);
877 		mipsxx_pmu_write_control(1, 191<<5);
878 		mipspmu.write_counter(1, 0);
879 		mipsxx_pmu_write_control(1, 255<<5);
880 		mipspmu.write_counter(1, 0);
881 		mipsxx_pmu_write_control(1, 319<<5);
882 		mipspmu.write_counter(1, 0);
883 		mipsxx_pmu_write_control(1, 383<<5);
884 		mipspmu.write_counter(1, 0);
885 		mipsxx_pmu_write_control(1, 575<<5);
886 		mipspmu.write_counter(1, 0);
887 		fallthrough;
888 	case 1:
889 		mipsxx_pmu_write_control(0, 0);
890 		mipspmu.write_counter(0, 0);
891 		mipsxx_pmu_write_control(0, 127<<5);
892 		mipspmu.write_counter(0, 0);
893 		mipsxx_pmu_write_control(0, 191<<5);
894 		mipspmu.write_counter(0, 0);
895 		mipsxx_pmu_write_control(0, 255<<5);
896 		mipspmu.write_counter(0, 0);
897 		mipsxx_pmu_write_control(0, 319<<5);
898 		mipspmu.write_counter(0, 0);
899 		mipsxx_pmu_write_control(0, 383<<5);
900 		mipspmu.write_counter(0, 0);
901 		mipsxx_pmu_write_control(0, 575<<5);
902 		mipspmu.write_counter(0, 0);
903 		break;
904 	}
905 }
906 
reset_counters(void * arg)907 static void reset_counters(void *arg)
908 {
909 	int counters = (int)(long)arg;
910 
911 	if (get_loongson3_pmu_type() == LOONGSON_PMU_TYPE2) {
912 		loongson3_reset_counters(arg);
913 		return;
914 	}
915 
916 	switch (counters) {
917 	case 4:
918 		mipsxx_pmu_write_control(3, 0);
919 		mipspmu.write_counter(3, 0);
920 		fallthrough;
921 	case 3:
922 		mipsxx_pmu_write_control(2, 0);
923 		mipspmu.write_counter(2, 0);
924 		fallthrough;
925 	case 2:
926 		mipsxx_pmu_write_control(1, 0);
927 		mipspmu.write_counter(1, 0);
928 		fallthrough;
929 	case 1:
930 		mipsxx_pmu_write_control(0, 0);
931 		mipspmu.write_counter(0, 0);
932 		break;
933 	}
934 }
935 
936 /* 24K/34K/1004K/interAptiv/loongson1 cores share the same event map. */
937 static const struct mips_perf_event mipsxxcore_event_map
938 				[PERF_COUNT_HW_MAX] = {
939 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
940 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
941 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_EVEN, T },
942 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
943 };
944 
945 /* 74K/proAptiv core has different branch event code. */
946 static const struct mips_perf_event mipsxxcore_event_map2
947 				[PERF_COUNT_HW_MAX] = {
948 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, P },
949 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
950 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x27, CNTR_EVEN, T },
951 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x27, CNTR_ODD, T },
952 };
953 
954 static const struct mips_perf_event i6x00_event_map[PERF_COUNT_HW_MAX] = {
955 	[PERF_COUNT_HW_CPU_CYCLES]          = { 0x00, CNTR_EVEN | CNTR_ODD },
956 	[PERF_COUNT_HW_INSTRUCTIONS]        = { 0x01, CNTR_EVEN | CNTR_ODD },
957 	/* These only count dcache, not icache */
958 	[PERF_COUNT_HW_CACHE_REFERENCES]    = { 0x45, CNTR_EVEN | CNTR_ODD },
959 	[PERF_COUNT_HW_CACHE_MISSES]        = { 0x48, CNTR_EVEN | CNTR_ODD },
960 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x15, CNTR_EVEN | CNTR_ODD },
961 	[PERF_COUNT_HW_BRANCH_MISSES]       = { 0x16, CNTR_EVEN | CNTR_ODD },
962 };
963 
964 static const struct mips_perf_event loongson3_event_map1[PERF_COUNT_HW_MAX] = {
965 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN },
966 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x00, CNTR_ODD },
967 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x01, CNTR_EVEN },
968 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x01, CNTR_ODD },
969 };
970 
971 static const struct mips_perf_event loongson3_event_map2[PERF_COUNT_HW_MAX] = {
972 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x80, CNTR_ALL },
973 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x81, CNTR_ALL },
974 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x18, CNTR_ALL },
975 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x94, CNTR_ALL },
976 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x9c, CNTR_ALL },
977 };
978 
979 static const struct mips_perf_event loongson3_event_map3[PERF_COUNT_HW_MAX] = {
980 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_ALL },
981 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_ALL },
982 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x1c, CNTR_ALL },
983 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x1d, CNTR_ALL },
984 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02, CNTR_ALL },
985 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x08, CNTR_ALL },
986 };
987 
988 static const struct mips_perf_event octeon_event_map[PERF_COUNT_HW_MAX] = {
989 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x01, CNTR_ALL },
990 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x03, CNTR_ALL },
991 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x2b, CNTR_ALL },
992 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x2e, CNTR_ALL	 },
993 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x08, CNTR_ALL },
994 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x09, CNTR_ALL },
995 	[PERF_COUNT_HW_BUS_CYCLES] = { 0x25, CNTR_ALL },
996 };
997 
998 static const struct mips_perf_event bmips5000_event_map
999 				[PERF_COUNT_HW_MAX] = {
1000 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00, CNTR_EVEN | CNTR_ODD, T },
1001 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01, CNTR_EVEN | CNTR_ODD, T },
1002 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x02, CNTR_ODD, T },
1003 };
1004 
1005 /* 24K/34K/1004K/interAptiv/loongson1 cores share the same cache event map. */
1006 static const struct mips_perf_event mipsxxcore_cache_map
1007 				[PERF_COUNT_HW_CACHE_MAX]
1008 				[PERF_COUNT_HW_CACHE_OP_MAX]
1009 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1010 [C(L1D)] = {
1011 	/*
1012 	 * Like some other architectures (e.g. ARM), the performance
1013 	 * counters don't differentiate between read and write
1014 	 * accesses/misses, so this isn't strictly correct, but it's the
1015 	 * best we can do. Writes and reads get combined.
1016 	 */
1017 	[C(OP_READ)] = {
1018 		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
1019 		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
1020 	},
1021 	[C(OP_WRITE)] = {
1022 		[C(RESULT_ACCESS)]	= { 0x0a, CNTR_EVEN, T },
1023 		[C(RESULT_MISS)]	= { 0x0b, CNTR_EVEN | CNTR_ODD, T },
1024 	},
1025 },
1026 [C(L1I)] = {
1027 	[C(OP_READ)] = {
1028 		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
1029 		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
1030 	},
1031 	[C(OP_WRITE)] = {
1032 		[C(RESULT_ACCESS)]	= { 0x09, CNTR_EVEN, T },
1033 		[C(RESULT_MISS)]	= { 0x09, CNTR_ODD, T },
1034 	},
1035 	[C(OP_PREFETCH)] = {
1036 		[C(RESULT_ACCESS)]	= { 0x14, CNTR_EVEN, T },
1037 		/*
1038 		 * Note that MIPS has only "hit" events countable for
1039 		 * the prefetch operation.
1040 		 */
1041 	},
1042 },
1043 [C(LL)] = {
1044 	[C(OP_READ)] = {
1045 		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
1046 		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
1047 	},
1048 	[C(OP_WRITE)] = {
1049 		[C(RESULT_ACCESS)]	= { 0x15, CNTR_ODD, P },
1050 		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN, P },
1051 	},
1052 },
1053 [C(DTLB)] = {
1054 	[C(OP_READ)] = {
1055 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
1056 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
1057 	},
1058 	[C(OP_WRITE)] = {
1059 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
1060 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
1061 	},
1062 },
1063 [C(ITLB)] = {
1064 	[C(OP_READ)] = {
1065 		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
1066 		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
1067 	},
1068 	[C(OP_WRITE)] = {
1069 		[C(RESULT_ACCESS)]	= { 0x05, CNTR_EVEN, T },
1070 		[C(RESULT_MISS)]	= { 0x05, CNTR_ODD, T },
1071 	},
1072 },
1073 [C(BPU)] = {
1074 	/* Using the same code for *HW_BRANCH* */
1075 	[C(OP_READ)] = {
1076 		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
1077 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1078 	},
1079 	[C(OP_WRITE)] = {
1080 		[C(RESULT_ACCESS)]	= { 0x02, CNTR_EVEN, T },
1081 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1082 	},
1083 },
1084 };
1085 
1086 /* 74K/proAptiv core has completely different cache event map. */
1087 static const struct mips_perf_event mipsxxcore_cache_map2
1088 				[PERF_COUNT_HW_CACHE_MAX]
1089 				[PERF_COUNT_HW_CACHE_OP_MAX]
1090 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1091 [C(L1D)] = {
1092 	/*
1093 	 * Like some other architectures (e.g. ARM), the performance
1094 	 * counters don't differentiate between read and write
1095 	 * accesses/misses, so this isn't strictly correct, but it's the
1096 	 * best we can do. Writes and reads get combined.
1097 	 */
1098 	[C(OP_READ)] = {
1099 		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
1100 		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
1101 	},
1102 	[C(OP_WRITE)] = {
1103 		[C(RESULT_ACCESS)]	= { 0x17, CNTR_ODD, T },
1104 		[C(RESULT_MISS)]	= { 0x18, CNTR_ODD, T },
1105 	},
1106 },
1107 [C(L1I)] = {
1108 	[C(OP_READ)] = {
1109 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
1110 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
1111 	},
1112 	[C(OP_WRITE)] = {
1113 		[C(RESULT_ACCESS)]	= { 0x06, CNTR_EVEN, T },
1114 		[C(RESULT_MISS)]	= { 0x06, CNTR_ODD, T },
1115 	},
1116 	[C(OP_PREFETCH)] = {
1117 		[C(RESULT_ACCESS)]	= { 0x34, CNTR_EVEN, T },
1118 		/*
1119 		 * Note that MIPS has only "hit" events countable for
1120 		 * the prefetch operation.
1121 		 */
1122 	},
1123 },
1124 [C(LL)] = {
1125 	[C(OP_READ)] = {
1126 		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
1127 		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN, P },
1128 	},
1129 	[C(OP_WRITE)] = {
1130 		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ODD, P },
1131 		[C(RESULT_MISS)]	= { 0x1d, CNTR_EVEN, P },
1132 	},
1133 },
1134 /*
1135  * 74K core does not have specific DTLB events. proAptiv core has
1136  * "speculative" DTLB events which are numbered 0x63 (even/odd) and
1137  * not included here. One can use raw events if really needed.
1138  */
1139 [C(ITLB)] = {
1140 	[C(OP_READ)] = {
1141 		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
1142 		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
1143 	},
1144 	[C(OP_WRITE)] = {
1145 		[C(RESULT_ACCESS)]	= { 0x04, CNTR_EVEN, T },
1146 		[C(RESULT_MISS)]	= { 0x04, CNTR_ODD, T },
1147 	},
1148 },
1149 [C(BPU)] = {
1150 	/* Using the same code for *HW_BRANCH* */
1151 	[C(OP_READ)] = {
1152 		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1153 		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1154 	},
1155 	[C(OP_WRITE)] = {
1156 		[C(RESULT_ACCESS)]	= { 0x27, CNTR_EVEN, T },
1157 		[C(RESULT_MISS)]	= { 0x27, CNTR_ODD, T },
1158 	},
1159 },
1160 };
1161 
1162 static const struct mips_perf_event i6x00_cache_map
1163 				[PERF_COUNT_HW_CACHE_MAX]
1164 				[PERF_COUNT_HW_CACHE_OP_MAX]
1165 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1166 [C(L1D)] = {
1167 	[C(OP_READ)] = {
1168 		[C(RESULT_ACCESS)]	= { 0x46, CNTR_EVEN | CNTR_ODD },
1169 		[C(RESULT_MISS)]	= { 0x49, CNTR_EVEN | CNTR_ODD },
1170 	},
1171 	[C(OP_WRITE)] = {
1172 		[C(RESULT_ACCESS)]	= { 0x47, CNTR_EVEN | CNTR_ODD },
1173 		[C(RESULT_MISS)]	= { 0x4a, CNTR_EVEN | CNTR_ODD },
1174 	},
1175 },
1176 [C(L1I)] = {
1177 	[C(OP_READ)] = {
1178 		[C(RESULT_ACCESS)]	= { 0x84, CNTR_EVEN | CNTR_ODD },
1179 		[C(RESULT_MISS)]	= { 0x85, CNTR_EVEN | CNTR_ODD },
1180 	},
1181 },
1182 [C(DTLB)] = {
1183 	/* Can't distinguish read & write */
1184 	[C(OP_READ)] = {
1185 		[C(RESULT_ACCESS)]	= { 0x40, CNTR_EVEN | CNTR_ODD },
1186 		[C(RESULT_MISS)]	= { 0x41, CNTR_EVEN | CNTR_ODD },
1187 	},
1188 	[C(OP_WRITE)] = {
1189 		[C(RESULT_ACCESS)]	= { 0x40, CNTR_EVEN | CNTR_ODD },
1190 		[C(RESULT_MISS)]	= { 0x41, CNTR_EVEN | CNTR_ODD },
1191 	},
1192 },
1193 [C(BPU)] = {
1194 	/* Conditional branches / mispredicted */
1195 	[C(OP_READ)] = {
1196 		[C(RESULT_ACCESS)]	= { 0x15, CNTR_EVEN | CNTR_ODD },
1197 		[C(RESULT_MISS)]	= { 0x16, CNTR_EVEN | CNTR_ODD },
1198 	},
1199 },
1200 };
1201 
1202 static const struct mips_perf_event loongson3_cache_map1
1203 				[PERF_COUNT_HW_CACHE_MAX]
1204 				[PERF_COUNT_HW_CACHE_OP_MAX]
1205 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1206 [C(L1D)] = {
1207 	/*
1208 	 * Like some other architectures (e.g. ARM), the performance
1209 	 * counters don't differentiate between read and write
1210 	 * accesses/misses, so this isn't strictly correct, but it's the
1211 	 * best we can do. Writes and reads get combined.
1212 	 */
1213 	[C(OP_READ)] = {
1214 		[C(RESULT_MISS)]        = { 0x04, CNTR_ODD },
1215 	},
1216 	[C(OP_WRITE)] = {
1217 		[C(RESULT_MISS)]        = { 0x04, CNTR_ODD },
1218 	},
1219 },
1220 [C(L1I)] = {
1221 	[C(OP_READ)] = {
1222 		[C(RESULT_MISS)]        = { 0x04, CNTR_EVEN },
1223 	},
1224 	[C(OP_WRITE)] = {
1225 		[C(RESULT_MISS)]        = { 0x04, CNTR_EVEN },
1226 	},
1227 },
1228 [C(DTLB)] = {
1229 	[C(OP_READ)] = {
1230 		[C(RESULT_MISS)]        = { 0x09, CNTR_ODD },
1231 	},
1232 	[C(OP_WRITE)] = {
1233 		[C(RESULT_MISS)]        = { 0x09, CNTR_ODD },
1234 	},
1235 },
1236 [C(ITLB)] = {
1237 	[C(OP_READ)] = {
1238 		[C(RESULT_MISS)]        = { 0x0c, CNTR_ODD },
1239 	},
1240 	[C(OP_WRITE)] = {
1241 		[C(RESULT_MISS)]        = { 0x0c, CNTR_ODD },
1242 	},
1243 },
1244 [C(BPU)] = {
1245 	/* Using the same code for *HW_BRANCH* */
1246 	[C(OP_READ)] = {
1247 		[C(RESULT_ACCESS)]      = { 0x01, CNTR_EVEN },
1248 		[C(RESULT_MISS)]        = { 0x01, CNTR_ODD },
1249 	},
1250 	[C(OP_WRITE)] = {
1251 		[C(RESULT_ACCESS)]      = { 0x01, CNTR_EVEN },
1252 		[C(RESULT_MISS)]        = { 0x01, CNTR_ODD },
1253 	},
1254 },
1255 };
1256 
1257 static const struct mips_perf_event loongson3_cache_map2
1258 				[PERF_COUNT_HW_CACHE_MAX]
1259 				[PERF_COUNT_HW_CACHE_OP_MAX]
1260 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1261 [C(L1D)] = {
1262 	/*
1263 	 * Like some other architectures (e.g. ARM), the performance
1264 	 * counters don't differentiate between read and write
1265 	 * accesses/misses, so this isn't strictly correct, but it's the
1266 	 * best we can do. Writes and reads get combined.
1267 	 */
1268 	[C(OP_READ)] = {
1269 		[C(RESULT_ACCESS)]	= { 0x156, CNTR_ALL },
1270 	},
1271 	[C(OP_WRITE)] = {
1272 		[C(RESULT_ACCESS)]	= { 0x155, CNTR_ALL },
1273 		[C(RESULT_MISS)]        = { 0x153, CNTR_ALL },
1274 	},
1275 },
1276 [C(L1I)] = {
1277 	[C(OP_READ)] = {
1278 		[C(RESULT_MISS)]	= { 0x18, CNTR_ALL },
1279 	},
1280 	[C(OP_WRITE)] = {
1281 		[C(RESULT_MISS)]        = { 0x18, CNTR_ALL },
1282 	},
1283 },
1284 [C(LL)] = {
1285 	[C(OP_READ)] = {
1286 		[C(RESULT_ACCESS)]	= { 0x1b6, CNTR_ALL },
1287 	},
1288 	[C(OP_WRITE)] = {
1289 		[C(RESULT_ACCESS)]	= { 0x1b7, CNTR_ALL },
1290 	},
1291 	[C(OP_PREFETCH)] = {
1292 		[C(RESULT_ACCESS)]	= { 0x1bf, CNTR_ALL },
1293 	},
1294 },
1295 [C(DTLB)] = {
1296 	[C(OP_READ)] = {
1297 		[C(RESULT_MISS)]        = { 0x92, CNTR_ALL },
1298 	},
1299 	[C(OP_WRITE)] = {
1300 		[C(RESULT_MISS)]        = { 0x92, CNTR_ALL },
1301 	},
1302 },
1303 [C(ITLB)] = {
1304 	[C(OP_READ)] = {
1305 		[C(RESULT_MISS)]	= { 0x1a, CNTR_ALL },
1306 	},
1307 	[C(OP_WRITE)] = {
1308 		[C(RESULT_MISS)]	= { 0x1a, CNTR_ALL },
1309 	},
1310 },
1311 [C(BPU)] = {
1312 	/* Using the same code for *HW_BRANCH* */
1313 	[C(OP_READ)] = {
1314 		[C(RESULT_ACCESS)]      = { 0x94, CNTR_ALL },
1315 		[C(RESULT_MISS)]        = { 0x9c, CNTR_ALL },
1316 	},
1317 },
1318 };
1319 
1320 static const struct mips_perf_event loongson3_cache_map3
1321 				[PERF_COUNT_HW_CACHE_MAX]
1322 				[PERF_COUNT_HW_CACHE_OP_MAX]
1323 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1324 [C(L1D)] = {
1325 	/*
1326 	 * Like some other architectures (e.g. ARM), the performance
1327 	 * counters don't differentiate between read and write
1328 	 * accesses/misses, so this isn't strictly correct, but it's the
1329 	 * best we can do. Writes and reads get combined.
1330 	 */
1331 	[C(OP_READ)] = {
1332 		[C(RESULT_ACCESS)]      = { 0x1e, CNTR_ALL },
1333 		[C(RESULT_MISS)]        = { 0x1f, CNTR_ALL },
1334 	},
1335 	[C(OP_PREFETCH)] = {
1336 		[C(RESULT_ACCESS)]	= { 0xaa, CNTR_ALL },
1337 		[C(RESULT_MISS)]	= { 0xa9, CNTR_ALL },
1338 	},
1339 },
1340 [C(L1I)] = {
1341 	[C(OP_READ)] = {
1342 		[C(RESULT_ACCESS)]	= { 0x1c, CNTR_ALL },
1343 		[C(RESULT_MISS)]	= { 0x1d, CNTR_ALL },
1344 	},
1345 },
1346 [C(LL)] = {
1347 	[C(OP_READ)] = {
1348 		[C(RESULT_ACCESS)]	= { 0x2e, CNTR_ALL },
1349 		[C(RESULT_MISS)]	= { 0x2f, CNTR_ALL },
1350 	},
1351 },
1352 [C(DTLB)] = {
1353 	[C(OP_READ)] = {
1354 		[C(RESULT_ACCESS)]      = { 0x14, CNTR_ALL },
1355 		[C(RESULT_MISS)]	= { 0x1b, CNTR_ALL },
1356 	},
1357 },
1358 [C(ITLB)] = {
1359 	[C(OP_READ)] = {
1360 		[C(RESULT_MISS)]	= { 0x1a, CNTR_ALL },
1361 	},
1362 },
1363 [C(BPU)] = {
1364 	/* Using the same code for *HW_BRANCH* */
1365 	[C(OP_READ)] = {
1366 		[C(RESULT_ACCESS)]      = { 0x02, CNTR_ALL },
1367 		[C(RESULT_MISS)]        = { 0x08, CNTR_ALL },
1368 	},
1369 },
1370 };
1371 
1372 /* BMIPS5000 */
1373 static const struct mips_perf_event bmips5000_cache_map
1374 				[PERF_COUNT_HW_CACHE_MAX]
1375 				[PERF_COUNT_HW_CACHE_OP_MAX]
1376 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1377 [C(L1D)] = {
1378 	/*
1379 	 * Like some other architectures (e.g. ARM), the performance
1380 	 * counters don't differentiate between read and write
1381 	 * accesses/misses, so this isn't strictly correct, but it's the
1382 	 * best we can do. Writes and reads get combined.
1383 	 */
1384 	[C(OP_READ)] = {
1385 		[C(RESULT_ACCESS)]	= { 12, CNTR_EVEN, T },
1386 		[C(RESULT_MISS)]	= { 12, CNTR_ODD, T },
1387 	},
1388 	[C(OP_WRITE)] = {
1389 		[C(RESULT_ACCESS)]	= { 12, CNTR_EVEN, T },
1390 		[C(RESULT_MISS)]	= { 12, CNTR_ODD, T },
1391 	},
1392 },
1393 [C(L1I)] = {
1394 	[C(OP_READ)] = {
1395 		[C(RESULT_ACCESS)]	= { 10, CNTR_EVEN, T },
1396 		[C(RESULT_MISS)]	= { 10, CNTR_ODD, T },
1397 	},
1398 	[C(OP_WRITE)] = {
1399 		[C(RESULT_ACCESS)]	= { 10, CNTR_EVEN, T },
1400 		[C(RESULT_MISS)]	= { 10, CNTR_ODD, T },
1401 	},
1402 	[C(OP_PREFETCH)] = {
1403 		[C(RESULT_ACCESS)]	= { 23, CNTR_EVEN, T },
1404 		/*
1405 		 * Note that MIPS has only "hit" events countable for
1406 		 * the prefetch operation.
1407 		 */
1408 	},
1409 },
1410 [C(LL)] = {
1411 	[C(OP_READ)] = {
1412 		[C(RESULT_ACCESS)]	= { 28, CNTR_EVEN, P },
1413 		[C(RESULT_MISS)]	= { 28, CNTR_ODD, P },
1414 	},
1415 	[C(OP_WRITE)] = {
1416 		[C(RESULT_ACCESS)]	= { 28, CNTR_EVEN, P },
1417 		[C(RESULT_MISS)]	= { 28, CNTR_ODD, P },
1418 	},
1419 },
1420 [C(BPU)] = {
1421 	/* Using the same code for *HW_BRANCH* */
1422 	[C(OP_READ)] = {
1423 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1424 	},
1425 	[C(OP_WRITE)] = {
1426 		[C(RESULT_MISS)]	= { 0x02, CNTR_ODD, T },
1427 	},
1428 },
1429 };
1430 
1431 static const struct mips_perf_event octeon_cache_map
1432 				[PERF_COUNT_HW_CACHE_MAX]
1433 				[PERF_COUNT_HW_CACHE_OP_MAX]
1434 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
1435 [C(L1D)] = {
1436 	[C(OP_READ)] = {
1437 		[C(RESULT_ACCESS)]	= { 0x2b, CNTR_ALL },
1438 		[C(RESULT_MISS)]	= { 0x2e, CNTR_ALL },
1439 	},
1440 	[C(OP_WRITE)] = {
1441 		[C(RESULT_ACCESS)]	= { 0x30, CNTR_ALL },
1442 	},
1443 },
1444 [C(L1I)] = {
1445 	[C(OP_READ)] = {
1446 		[C(RESULT_ACCESS)]	= { 0x18, CNTR_ALL },
1447 	},
1448 	[C(OP_PREFETCH)] = {
1449 		[C(RESULT_ACCESS)]	= { 0x19, CNTR_ALL },
1450 	},
1451 },
1452 [C(DTLB)] = {
1453 	/*
1454 	 * Only general DTLB misses are counted use the same event for
1455 	 * read and write.
1456 	 */
1457 	[C(OP_READ)] = {
1458 		[C(RESULT_MISS)]	= { 0x35, CNTR_ALL },
1459 	},
1460 	[C(OP_WRITE)] = {
1461 		[C(RESULT_MISS)]	= { 0x35, CNTR_ALL },
1462 	},
1463 },
1464 [C(ITLB)] = {
1465 	[C(OP_READ)] = {
1466 		[C(RESULT_MISS)]	= { 0x37, CNTR_ALL },
1467 	},
1468 },
1469 };
1470 
__hw_perf_event_init(struct perf_event * event)1471 static int __hw_perf_event_init(struct perf_event *event)
1472 {
1473 	struct perf_event_attr *attr = &event->attr;
1474 	struct hw_perf_event *hwc = &event->hw;
1475 	const struct mips_perf_event *pev;
1476 	int err;
1477 
1478 	/* Returning MIPS event descriptor for generic perf event. */
1479 	if (PERF_TYPE_HARDWARE == event->attr.type) {
1480 		if (event->attr.config >= PERF_COUNT_HW_MAX)
1481 			return -EINVAL;
1482 		pev = mipspmu_map_general_event(event->attr.config);
1483 	} else if (PERF_TYPE_HW_CACHE == event->attr.type) {
1484 		pev = mipspmu_map_cache_event(event->attr.config);
1485 	} else if (PERF_TYPE_RAW == event->attr.type) {
1486 		/* We are working on the global raw event. */
1487 		mutex_lock(&raw_event_mutex);
1488 		pev = mipspmu.map_raw_event(event->attr.config);
1489 	} else {
1490 		/* The event type is not (yet) supported. */
1491 		return -EOPNOTSUPP;
1492 	}
1493 
1494 	if (IS_ERR(pev)) {
1495 		if (PERF_TYPE_RAW == event->attr.type)
1496 			mutex_unlock(&raw_event_mutex);
1497 		return PTR_ERR(pev);
1498 	}
1499 
1500 	/*
1501 	 * We allow max flexibility on how each individual counter shared
1502 	 * by the single CPU operates (the mode exclusion and the range).
1503 	 */
1504 	hwc->config_base = MIPS_PERFCTRL_IE;
1505 
1506 	hwc->event_base = mipspmu_perf_event_encode(pev);
1507 	if (PERF_TYPE_RAW == event->attr.type)
1508 		mutex_unlock(&raw_event_mutex);
1509 
1510 	if (!attr->exclude_user)
1511 		hwc->config_base |= MIPS_PERFCTRL_U;
1512 	if (!attr->exclude_kernel) {
1513 		hwc->config_base |= MIPS_PERFCTRL_K;
1514 		/* MIPS kernel mode: KSU == 00b || EXL == 1 || ERL == 1 */
1515 		hwc->config_base |= MIPS_PERFCTRL_EXL;
1516 	}
1517 	if (!attr->exclude_hv)
1518 		hwc->config_base |= MIPS_PERFCTRL_S;
1519 
1520 	hwc->config_base &= M_PERFCTL_CONFIG_MASK;
1521 	/*
1522 	 * The event can belong to another cpu. We do not assign a local
1523 	 * counter for it for now.
1524 	 */
1525 	hwc->idx = -1;
1526 	hwc->config = 0;
1527 
1528 	if (!hwc->sample_period) {
1529 		hwc->sample_period  = mipspmu.max_period;
1530 		hwc->last_period    = hwc->sample_period;
1531 		local64_set(&hwc->period_left, hwc->sample_period);
1532 	}
1533 
1534 	err = 0;
1535 	if (event->group_leader != event)
1536 		err = validate_group(event);
1537 
1538 	event->destroy = hw_perf_event_destroy;
1539 
1540 	if (err)
1541 		event->destroy(event);
1542 
1543 	return err;
1544 }
1545 
pause_local_counters(void)1546 static void pause_local_counters(void)
1547 {
1548 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1549 	int ctr = mipspmu.num_counters;
1550 	unsigned long flags;
1551 
1552 	local_irq_save(flags);
1553 	do {
1554 		ctr--;
1555 		cpuc->saved_ctrl[ctr] = mipsxx_pmu_read_control(ctr);
1556 		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
1557 					 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
1558 	} while (ctr > 0);
1559 	local_irq_restore(flags);
1560 }
1561 
resume_local_counters(void)1562 static void resume_local_counters(void)
1563 {
1564 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1565 	int ctr = mipspmu.num_counters;
1566 
1567 	do {
1568 		ctr--;
1569 		mipsxx_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
1570 	} while (ctr > 0);
1571 }
1572 
mipsxx_pmu_handle_shared_irq(void)1573 static int mipsxx_pmu_handle_shared_irq(void)
1574 {
1575 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1576 	struct perf_sample_data data;
1577 	unsigned int counters = mipspmu.num_counters;
1578 	u64 counter;
1579 	int n, handled = IRQ_NONE;
1580 	struct pt_regs *regs;
1581 
1582 	if (cpu_has_perf_cntr_intr_bit && !(read_c0_cause() & CAUSEF_PCI))
1583 		return handled;
1584 	/*
1585 	 * First we pause the local counters, so that when we are locked
1586 	 * here, the counters are all paused. When it gets locked due to
1587 	 * perf_disable(), the timer interrupt handler will be delayed.
1588 	 *
1589 	 * See also mipsxx_pmu_start().
1590 	 */
1591 	pause_local_counters();
1592 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1593 	read_lock(&pmuint_rwlock);
1594 #endif
1595 
1596 	regs = get_irq_regs();
1597 
1598 	perf_sample_data_init(&data, 0, 0);
1599 
1600 	for (n = counters - 1; n >= 0; n--) {
1601 		if (!test_bit(n, cpuc->used_mask))
1602 			continue;
1603 
1604 		counter = mipspmu.read_counter(n);
1605 		if (!(counter & mipspmu.overflow))
1606 			continue;
1607 
1608 		handle_associated_event(cpuc, n, &data, regs);
1609 		handled = IRQ_HANDLED;
1610 	}
1611 
1612 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1613 	read_unlock(&pmuint_rwlock);
1614 #endif
1615 	resume_local_counters();
1616 
1617 	/*
1618 	 * Do all the work for the pending perf events. We can do this
1619 	 * in here because the performance counter interrupt is a regular
1620 	 * interrupt, not NMI.
1621 	 */
1622 	if (handled == IRQ_HANDLED)
1623 		irq_work_run();
1624 
1625 	return handled;
1626 }
1627 
mipsxx_pmu_handle_irq(int irq,void * dev)1628 static irqreturn_t mipsxx_pmu_handle_irq(int irq, void *dev)
1629 {
1630 	return mipsxx_pmu_handle_shared_irq();
1631 }
1632 
1633 /* 24K */
1634 #define IS_BOTH_COUNTERS_24K_EVENT(b)					\
1635 	((b) == 0 || (b) == 1 || (b) == 11)
1636 
1637 /* 34K */
1638 #define IS_BOTH_COUNTERS_34K_EVENT(b)					\
1639 	((b) == 0 || (b) == 1 || (b) == 11)
1640 #ifdef CONFIG_MIPS_MT_SMP
1641 #define IS_RANGE_P_34K_EVENT(r, b)					\
1642 	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1643 	 (b) == 25 || (b) == 39 || (r) == 44 || (r) == 174 ||		\
1644 	 (r) == 176 || ((b) >= 50 && (b) <= 55) ||			\
1645 	 ((b) >= 64 && (b) <= 67))
1646 #define IS_RANGE_V_34K_EVENT(r) ((r) == 47)
1647 #endif
1648 
1649 /* 74K */
1650 #define IS_BOTH_COUNTERS_74K_EVENT(b)					\
1651 	((b) == 0 || (b) == 1)
1652 
1653 /* proAptiv */
1654 #define IS_BOTH_COUNTERS_PROAPTIV_EVENT(b)				\
1655 	((b) == 0 || (b) == 1)
1656 /* P5600 */
1657 #define IS_BOTH_COUNTERS_P5600_EVENT(b)					\
1658 	((b) == 0 || (b) == 1)
1659 
1660 /* 1004K */
1661 #define IS_BOTH_COUNTERS_1004K_EVENT(b)					\
1662 	((b) == 0 || (b) == 1 || (b) == 11)
1663 #ifdef CONFIG_MIPS_MT_SMP
1664 #define IS_RANGE_P_1004K_EVENT(r, b)					\
1665 	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1666 	 (b) == 25 || (b) == 36 || (b) == 39 || (r) == 44 ||		\
1667 	 (r) == 174 || (r) == 176 || ((b) >= 50 && (b) <= 59) ||	\
1668 	 (r) == 188 || (b) == 61 || (b) == 62 ||			\
1669 	 ((b) >= 64 && (b) <= 67))
1670 #define IS_RANGE_V_1004K_EVENT(r)	((r) == 47)
1671 #endif
1672 
1673 /* interAptiv */
1674 #define IS_BOTH_COUNTERS_INTERAPTIV_EVENT(b)				\
1675 	((b) == 0 || (b) == 1 || (b) == 11)
1676 #ifdef CONFIG_MIPS_MT_SMP
1677 /* The P/V/T info is not provided for "(b) == 38" in SUM, assume P. */
1678 #define IS_RANGE_P_INTERAPTIV_EVENT(r, b)				\
1679 	((b) == 0 || (r) == 18 || (b) == 21 || (b) == 22 ||		\
1680 	 (b) == 25 || (b) == 36 || (b) == 38 || (b) == 39 ||		\
1681 	 (r) == 44 || (r) == 174 || (r) == 176 || ((b) >= 50 &&		\
1682 	 (b) <= 59) || (r) == 188 || (b) == 61 || (b) == 62 ||		\
1683 	 ((b) >= 64 && (b) <= 67))
1684 #define IS_RANGE_V_INTERAPTIV_EVENT(r)	((r) == 47 || (r) == 175)
1685 #endif
1686 
1687 /* BMIPS5000 */
1688 #define IS_BOTH_COUNTERS_BMIPS5000_EVENT(b)				\
1689 	((b) == 0 || (b) == 1)
1690 
1691 
1692 /*
1693  * For most cores the user can use 0-255 raw events, where 0-127 for the events
1694  * of even counters, and 128-255 for odd counters. Note that bit 7 is used to
1695  * indicate the even/odd bank selector. So, for example, when user wants to take
1696  * the Event Num of 15 for odd counters (by referring to the user manual), then
1697  * 128 needs to be added to 15 as the input for the event config, i.e., 143 (0x8F)
1698  * to be used.
1699  *
1700  * Some newer cores have even more events, in which case the user can use raw
1701  * events 0-511, where 0-255 are for the events of even counters, and 256-511
1702  * are for odd counters, so bit 8 is used to indicate the even/odd bank selector.
1703  */
mipsxx_pmu_map_raw_event(u64 config)1704 static const struct mips_perf_event *mipsxx_pmu_map_raw_event(u64 config)
1705 {
1706 	/* currently most cores have 7-bit event numbers */
1707 	int pmu_type;
1708 	unsigned int raw_id = config & 0xff;
1709 	unsigned int base_id = raw_id & 0x7f;
1710 
1711 	switch (current_cpu_type()) {
1712 	case CPU_24K:
1713 		if (IS_BOTH_COUNTERS_24K_EVENT(base_id))
1714 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1715 		else
1716 			raw_event.cntr_mask =
1717 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1718 #ifdef CONFIG_MIPS_MT_SMP
1719 		/*
1720 		 * This is actually doing nothing. Non-multithreading
1721 		 * CPUs will not check and calculate the range.
1722 		 */
1723 		raw_event.range = P;
1724 #endif
1725 		break;
1726 	case CPU_34K:
1727 		if (IS_BOTH_COUNTERS_34K_EVENT(base_id))
1728 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1729 		else
1730 			raw_event.cntr_mask =
1731 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1732 #ifdef CONFIG_MIPS_MT_SMP
1733 		if (IS_RANGE_P_34K_EVENT(raw_id, base_id))
1734 			raw_event.range = P;
1735 		else if (unlikely(IS_RANGE_V_34K_EVENT(raw_id)))
1736 			raw_event.range = V;
1737 		else
1738 			raw_event.range = T;
1739 #endif
1740 		break;
1741 	case CPU_74K:
1742 	case CPU_1074K:
1743 		if (IS_BOTH_COUNTERS_74K_EVENT(base_id))
1744 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1745 		else
1746 			raw_event.cntr_mask =
1747 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1748 #ifdef CONFIG_MIPS_MT_SMP
1749 		raw_event.range = P;
1750 #endif
1751 		break;
1752 	case CPU_PROAPTIV:
1753 		if (IS_BOTH_COUNTERS_PROAPTIV_EVENT(base_id))
1754 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1755 		else
1756 			raw_event.cntr_mask =
1757 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1758 #ifdef CONFIG_MIPS_MT_SMP
1759 		raw_event.range = P;
1760 #endif
1761 		break;
1762 	case CPU_P5600:
1763 	case CPU_P6600:
1764 		/* 8-bit event numbers */
1765 		raw_id = config & 0x1ff;
1766 		base_id = raw_id & 0xff;
1767 		if (IS_BOTH_COUNTERS_P5600_EVENT(base_id))
1768 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1769 		else
1770 			raw_event.cntr_mask =
1771 				raw_id > 255 ? CNTR_ODD : CNTR_EVEN;
1772 #ifdef CONFIG_MIPS_MT_SMP
1773 		raw_event.range = P;
1774 #endif
1775 		break;
1776 	case CPU_I6400:
1777 	case CPU_I6500:
1778 		/* 8-bit event numbers */
1779 		base_id = config & 0xff;
1780 		raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1781 		break;
1782 	case CPU_1004K:
1783 		if (IS_BOTH_COUNTERS_1004K_EVENT(base_id))
1784 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1785 		else
1786 			raw_event.cntr_mask =
1787 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1788 #ifdef CONFIG_MIPS_MT_SMP
1789 		if (IS_RANGE_P_1004K_EVENT(raw_id, base_id))
1790 			raw_event.range = P;
1791 		else if (unlikely(IS_RANGE_V_1004K_EVENT(raw_id)))
1792 			raw_event.range = V;
1793 		else
1794 			raw_event.range = T;
1795 #endif
1796 		break;
1797 	case CPU_INTERAPTIV:
1798 		if (IS_BOTH_COUNTERS_INTERAPTIV_EVENT(base_id))
1799 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1800 		else
1801 			raw_event.cntr_mask =
1802 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1803 #ifdef CONFIG_MIPS_MT_SMP
1804 		if (IS_RANGE_P_INTERAPTIV_EVENT(raw_id, base_id))
1805 			raw_event.range = P;
1806 		else if (unlikely(IS_RANGE_V_INTERAPTIV_EVENT(raw_id)))
1807 			raw_event.range = V;
1808 		else
1809 			raw_event.range = T;
1810 #endif
1811 		break;
1812 	case CPU_BMIPS5000:
1813 		if (IS_BOTH_COUNTERS_BMIPS5000_EVENT(base_id))
1814 			raw_event.cntr_mask = CNTR_EVEN | CNTR_ODD;
1815 		else
1816 			raw_event.cntr_mask =
1817 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1818 		break;
1819 	case CPU_LOONGSON64:
1820 		pmu_type = get_loongson3_pmu_type();
1821 
1822 		switch (pmu_type) {
1823 		case LOONGSON_PMU_TYPE1:
1824 			raw_event.cntr_mask =
1825 				raw_id > 127 ? CNTR_ODD : CNTR_EVEN;
1826 			break;
1827 		case LOONGSON_PMU_TYPE2:
1828 			base_id = config & 0x3ff;
1829 			raw_event.cntr_mask = CNTR_ALL;
1830 
1831 			if ((base_id >= 1 && base_id < 28) ||
1832 				(base_id >= 64 && base_id < 90) ||
1833 				(base_id >= 128 && base_id < 164) ||
1834 				(base_id >= 192 && base_id < 200) ||
1835 				(base_id >= 256 && base_id < 275) ||
1836 				(base_id >= 320 && base_id < 361) ||
1837 				(base_id >= 384 && base_id < 574))
1838 				break;
1839 
1840 			return ERR_PTR(-EOPNOTSUPP);
1841 		case LOONGSON_PMU_TYPE3:
1842 			base_id = raw_id;
1843 			raw_event.cntr_mask = CNTR_ALL;
1844 			break;
1845 		}
1846 		break;
1847 	}
1848 
1849 	raw_event.event_id = base_id;
1850 
1851 	return &raw_event;
1852 }
1853 
octeon_pmu_map_raw_event(u64 config)1854 static const struct mips_perf_event *octeon_pmu_map_raw_event(u64 config)
1855 {
1856 	unsigned int base_id = config & 0x7f;
1857 	unsigned int event_max;
1858 
1859 
1860 	raw_event.cntr_mask = CNTR_ALL;
1861 	raw_event.event_id = base_id;
1862 
1863 	if (current_cpu_type() == CPU_CAVIUM_OCTEON3)
1864 		event_max = 0x5f;
1865 	else if (current_cpu_type() == CPU_CAVIUM_OCTEON2)
1866 		event_max = 0x42;
1867 	else
1868 		event_max = 0x3a;
1869 
1870 	if (base_id > event_max) {
1871 		return ERR_PTR(-EOPNOTSUPP);
1872 	}
1873 
1874 	switch (base_id) {
1875 	case 0x00:
1876 	case 0x0f:
1877 	case 0x1e:
1878 	case 0x1f:
1879 	case 0x2f:
1880 	case 0x34:
1881 	case 0x3e ... 0x3f:
1882 		return ERR_PTR(-EOPNOTSUPP);
1883 	default:
1884 		break;
1885 	}
1886 
1887 	return &raw_event;
1888 }
1889 
1890 static int __init
init_hw_perf_events(void)1891 init_hw_perf_events(void)
1892 {
1893 	int counters, irq, pmu_type;
1894 
1895 	pr_info("Performance counters: ");
1896 
1897 	counters = n_counters();
1898 	if (counters == 0) {
1899 		pr_cont("No available PMU.\n");
1900 		return -ENODEV;
1901 	}
1902 
1903 #ifdef CONFIG_MIPS_PERF_SHARED_TC_COUNTERS
1904 	if (!cpu_has_mipsmt_pertccounters)
1905 		counters = counters_total_to_per_cpu(counters);
1906 #endif
1907 
1908 	if (get_c0_perfcount_int)
1909 		irq = get_c0_perfcount_int();
1910 	else if (cp0_perfcount_irq >= 0)
1911 		irq = MIPS_CPU_IRQ_BASE + cp0_perfcount_irq;
1912 	else
1913 		irq = -1;
1914 
1915 	mipspmu.map_raw_event = mipsxx_pmu_map_raw_event;
1916 
1917 	switch (current_cpu_type()) {
1918 	case CPU_24K:
1919 		mipspmu.name = "mips/24K";
1920 		mipspmu.general_event_map = &mipsxxcore_event_map;
1921 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1922 		break;
1923 	case CPU_34K:
1924 		mipspmu.name = "mips/34K";
1925 		mipspmu.general_event_map = &mipsxxcore_event_map;
1926 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1927 		break;
1928 	case CPU_74K:
1929 		mipspmu.name = "mips/74K";
1930 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1931 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1932 		break;
1933 	case CPU_PROAPTIV:
1934 		mipspmu.name = "mips/proAptiv";
1935 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1936 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1937 		break;
1938 	case CPU_P5600:
1939 		mipspmu.name = "mips/P5600";
1940 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1941 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1942 		break;
1943 	case CPU_P6600:
1944 		mipspmu.name = "mips/P6600";
1945 		mipspmu.general_event_map = &mipsxxcore_event_map2;
1946 		mipspmu.cache_event_map = &mipsxxcore_cache_map2;
1947 		break;
1948 	case CPU_I6400:
1949 		mipspmu.name = "mips/I6400";
1950 		mipspmu.general_event_map = &i6x00_event_map;
1951 		mipspmu.cache_event_map = &i6x00_cache_map;
1952 		break;
1953 	case CPU_I6500:
1954 		mipspmu.name = "mips/I6500";
1955 		mipspmu.general_event_map = &i6x00_event_map;
1956 		mipspmu.cache_event_map = &i6x00_cache_map;
1957 		break;
1958 	case CPU_1004K:
1959 		mipspmu.name = "mips/1004K";
1960 		mipspmu.general_event_map = &mipsxxcore_event_map;
1961 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1962 		break;
1963 	case CPU_1074K:
1964 		mipspmu.name = "mips/1074K";
1965 		mipspmu.general_event_map = &mipsxxcore_event_map;
1966 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1967 		break;
1968 	case CPU_INTERAPTIV:
1969 		mipspmu.name = "mips/interAptiv";
1970 		mipspmu.general_event_map = &mipsxxcore_event_map;
1971 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1972 		break;
1973 	case CPU_LOONGSON32:
1974 		mipspmu.name = "mips/loongson1";
1975 		mipspmu.general_event_map = &mipsxxcore_event_map;
1976 		mipspmu.cache_event_map = &mipsxxcore_cache_map;
1977 		break;
1978 	case CPU_LOONGSON64:
1979 		mipspmu.name = "mips/loongson3";
1980 		pmu_type = get_loongson3_pmu_type();
1981 
1982 		switch (pmu_type) {
1983 		case LOONGSON_PMU_TYPE1:
1984 			counters = 2;
1985 			mipspmu.general_event_map = &loongson3_event_map1;
1986 			mipspmu.cache_event_map = &loongson3_cache_map1;
1987 			break;
1988 		case LOONGSON_PMU_TYPE2:
1989 			counters = 4;
1990 			mipspmu.general_event_map = &loongson3_event_map2;
1991 			mipspmu.cache_event_map = &loongson3_cache_map2;
1992 			break;
1993 		case LOONGSON_PMU_TYPE3:
1994 			counters = 4;
1995 			mipspmu.general_event_map = &loongson3_event_map3;
1996 			mipspmu.cache_event_map = &loongson3_cache_map3;
1997 			break;
1998 		}
1999 		break;
2000 	case CPU_CAVIUM_OCTEON:
2001 	case CPU_CAVIUM_OCTEON_PLUS:
2002 	case CPU_CAVIUM_OCTEON2:
2003 	case CPU_CAVIUM_OCTEON3:
2004 		mipspmu.name = "octeon";
2005 		mipspmu.general_event_map = &octeon_event_map;
2006 		mipspmu.cache_event_map = &octeon_cache_map;
2007 		mipspmu.map_raw_event = octeon_pmu_map_raw_event;
2008 		break;
2009 	case CPU_BMIPS5000:
2010 		mipspmu.name = "BMIPS5000";
2011 		mipspmu.general_event_map = &bmips5000_event_map;
2012 		mipspmu.cache_event_map = &bmips5000_cache_map;
2013 		break;
2014 	default:
2015 		pr_cont("Either hardware does not support performance "
2016 			"counters, or not yet implemented.\n");
2017 		return -ENODEV;
2018 	}
2019 
2020 	mipspmu.num_counters = counters;
2021 	mipspmu.irq = irq;
2022 
2023 	if (read_c0_perfctrl0() & MIPS_PERFCTRL_W) {
2024 		if (get_loongson3_pmu_type() == LOONGSON_PMU_TYPE2) {
2025 			counter_bits = 48;
2026 			mipspmu.max_period = (1ULL << 47) - 1;
2027 			mipspmu.valid_count = (1ULL << 47) - 1;
2028 			mipspmu.overflow = 1ULL << 47;
2029 		} else {
2030 			counter_bits = 64;
2031 			mipspmu.max_period = (1ULL << 63) - 1;
2032 			mipspmu.valid_count = (1ULL << 63) - 1;
2033 			mipspmu.overflow = 1ULL << 63;
2034 		}
2035 		mipspmu.read_counter = mipsxx_pmu_read_counter_64;
2036 		mipspmu.write_counter = mipsxx_pmu_write_counter_64;
2037 	} else {
2038 		counter_bits = 32;
2039 		mipspmu.max_period = (1ULL << 31) - 1;
2040 		mipspmu.valid_count = (1ULL << 31) - 1;
2041 		mipspmu.overflow = 1ULL << 31;
2042 		mipspmu.read_counter = mipsxx_pmu_read_counter;
2043 		mipspmu.write_counter = mipsxx_pmu_write_counter;
2044 	}
2045 
2046 	on_each_cpu(reset_counters, (void *)(long)counters, 1);
2047 
2048 	pr_cont("%s PMU enabled, %d %d-bit counters available to each "
2049 		"CPU, irq %d%s\n", mipspmu.name, counters, counter_bits, irq,
2050 		irq < 0 ? " (share with timer interrupt)" : "");
2051 
2052 	perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
2053 
2054 	return 0;
2055 }
2056 early_initcall(init_hw_perf_events);
2057