1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Linux performance counter support for LoongArch.
4  *
5  * Copyright (C) 2022 Loongson Technology Corporation Limited
6  *
7  * Derived from MIPS:
8  * Copyright (C) 2010 MIPS Technologies, Inc.
9  * Copyright (C) 2011 Cavium Networks, Inc.
10  * Author: Deng-Cheng Zhu
11  */
12 
13 #include <linux/cpumask.h>
14 #include <linux/interrupt.h>
15 #include <linux/smp.h>
16 #include <linux/kernel.h>
17 #include <linux/perf_event.h>
18 #include <linux/uaccess.h>
19 #include <linux/sched/task_stack.h>
20 
21 #include <asm/irq.h>
22 #include <asm/irq_regs.h>
23 #include <asm/stacktrace.h>
24 #include <asm/unwind.h>
25 
26 /*
27  * Get the return address for a single stackframe and return a pointer to the
28  * next frame tail.
29  */
30 static unsigned long
31 user_backtrace(struct perf_callchain_entry_ctx *entry, unsigned long fp)
32 {
33 	unsigned long err;
34 	unsigned long __user *user_frame_tail;
35 	struct stack_frame buftail;
36 
37 	user_frame_tail = (unsigned long __user *)(fp - sizeof(struct stack_frame));
38 
39 	/* Also check accessibility of one struct frame_tail beyond */
40 	if (!access_ok(user_frame_tail, sizeof(buftail)))
41 		return 0;
42 
43 	pagefault_disable();
44 	err = __copy_from_user_inatomic(&buftail, user_frame_tail, sizeof(buftail));
45 	pagefault_enable();
46 
47 	if (err || (unsigned long)user_frame_tail >= buftail.fp)
48 		return 0;
49 
50 	perf_callchain_store(entry, buftail.ra);
51 
52 	return buftail.fp;
53 }
54 
55 void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
56 			 struct pt_regs *regs)
57 {
58 	unsigned long fp;
59 
60 	if (perf_guest_state()) {
61 		/* We don't support guest os callchain now */
62 		return;
63 	}
64 
65 	perf_callchain_store(entry, regs->csr_era);
66 
67 	fp = regs->regs[22];
68 
69 	while (entry->nr < entry->max_stack && fp && !((unsigned long)fp & 0xf))
70 		fp = user_backtrace(entry, fp);
71 }
72 
73 void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
74 			   struct pt_regs *regs)
75 {
76 	struct unwind_state state;
77 	unsigned long addr;
78 
79 	for (unwind_start(&state, current, regs);
80 	      !unwind_done(&state); unwind_next_frame(&state)) {
81 		addr = unwind_get_return_address(&state);
82 		if (!addr || perf_callchain_store(entry, addr))
83 			return;
84 	}
85 }
86 
87 #define LOONGARCH_MAX_HWEVENTS 32
88 
89 struct cpu_hw_events {
90 	/* Array of events on this cpu. */
91 	struct perf_event	*events[LOONGARCH_MAX_HWEVENTS];
92 
93 	/*
94 	 * Set the bit (indexed by the counter number) when the counter
95 	 * is used for an event.
96 	 */
97 	unsigned long		used_mask[BITS_TO_LONGS(LOONGARCH_MAX_HWEVENTS)];
98 
99 	/*
100 	 * Software copy of the control register for each performance counter.
101 	 */
102 	unsigned int		saved_ctrl[LOONGARCH_MAX_HWEVENTS];
103 };
104 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
105 	.saved_ctrl = {0},
106 };
107 
108 /* The description of LoongArch performance events. */
109 struct loongarch_perf_event {
110 	unsigned int event_id;
111 };
112 
113 static struct loongarch_perf_event raw_event;
114 static DEFINE_MUTEX(raw_event_mutex);
115 
116 #define C(x) PERF_COUNT_HW_CACHE_##x
117 #define HW_OP_UNSUPPORTED		0xffffffff
118 #define CACHE_OP_UNSUPPORTED		0xffffffff
119 
120 #define PERF_MAP_ALL_UNSUPPORTED					\
121 	[0 ... PERF_COUNT_HW_MAX - 1] = {HW_OP_UNSUPPORTED}
122 
123 #define PERF_CACHE_MAP_ALL_UNSUPPORTED					\
124 [0 ... C(MAX) - 1] = {							\
125 	[0 ... C(OP_MAX) - 1] = {					\
126 		[0 ... C(RESULT_MAX) - 1] = {CACHE_OP_UNSUPPORTED},	\
127 	},								\
128 }
129 
130 struct loongarch_pmu {
131 	u64		max_period;
132 	u64		valid_count;
133 	u64		overflow;
134 	const char	*name;
135 	unsigned int	num_counters;
136 	u64		(*read_counter)(unsigned int idx);
137 	void		(*write_counter)(unsigned int idx, u64 val);
138 	const struct loongarch_perf_event *(*map_raw_event)(u64 config);
139 	const struct loongarch_perf_event (*general_event_map)[PERF_COUNT_HW_MAX];
140 	const struct loongarch_perf_event (*cache_event_map)
141 				[PERF_COUNT_HW_CACHE_MAX]
142 				[PERF_COUNT_HW_CACHE_OP_MAX]
143 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
144 };
145 
146 static struct loongarch_pmu loongarch_pmu;
147 
148 #define M_PERFCTL_EVENT(event)	(event & CSR_PERFCTRL_EVENT)
149 
150 #define M_PERFCTL_COUNT_EVENT_WHENEVER	(CSR_PERFCTRL_PLV0 |	\
151 					CSR_PERFCTRL_PLV1 |	\
152 					CSR_PERFCTRL_PLV2 |	\
153 					CSR_PERFCTRL_PLV3 |	\
154 					CSR_PERFCTRL_IE)
155 
156 #define M_PERFCTL_CONFIG_MASK		0x1f0000
157 
158 static void pause_local_counters(void);
159 static void resume_local_counters(void);
160 
161 static u64 loongarch_pmu_read_counter(unsigned int idx)
162 {
163 	u64 val = -1;
164 
165 	switch (idx) {
166 	case 0:
167 		val = read_csr_perfcntr0();
168 		break;
169 	case 1:
170 		val = read_csr_perfcntr1();
171 		break;
172 	case 2:
173 		val = read_csr_perfcntr2();
174 		break;
175 	case 3:
176 		val = read_csr_perfcntr3();
177 		break;
178 	default:
179 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
180 		return 0;
181 	}
182 
183 	return val;
184 }
185 
186 static void loongarch_pmu_write_counter(unsigned int idx, u64 val)
187 {
188 	switch (idx) {
189 	case 0:
190 		write_csr_perfcntr0(val);
191 		return;
192 	case 1:
193 		write_csr_perfcntr1(val);
194 		return;
195 	case 2:
196 		write_csr_perfcntr2(val);
197 		return;
198 	case 3:
199 		write_csr_perfcntr3(val);
200 		return;
201 	default:
202 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
203 		return;
204 	}
205 }
206 
207 static unsigned int loongarch_pmu_read_control(unsigned int idx)
208 {
209 	unsigned int val = -1;
210 
211 	switch (idx) {
212 	case 0:
213 		val = read_csr_perfctrl0();
214 		break;
215 	case 1:
216 		val = read_csr_perfctrl1();
217 		break;
218 	case 2:
219 		val = read_csr_perfctrl2();
220 		break;
221 	case 3:
222 		val = read_csr_perfctrl3();
223 		break;
224 	default:
225 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
226 		return 0;
227 	}
228 
229 	return val;
230 }
231 
232 static void loongarch_pmu_write_control(unsigned int idx, unsigned int val)
233 {
234 	switch (idx) {
235 	case 0:
236 		write_csr_perfctrl0(val);
237 		return;
238 	case 1:
239 		write_csr_perfctrl1(val);
240 		return;
241 	case 2:
242 		write_csr_perfctrl2(val);
243 		return;
244 	case 3:
245 		write_csr_perfctrl3(val);
246 		return;
247 	default:
248 		WARN_ONCE(1, "Invalid performance counter number (%d)\n", idx);
249 		return;
250 	}
251 }
252 
253 static int loongarch_pmu_alloc_counter(struct cpu_hw_events *cpuc, struct hw_perf_event *hwc)
254 {
255 	int i;
256 
257 	for (i = 0; i < loongarch_pmu.num_counters; i++) {
258 		if (!test_and_set_bit(i, cpuc->used_mask))
259 			return i;
260 	}
261 
262 	return -EAGAIN;
263 }
264 
265 static void loongarch_pmu_enable_event(struct hw_perf_event *evt, int idx)
266 {
267 	unsigned int cpu;
268 	struct perf_event *event = container_of(evt, struct perf_event, hw);
269 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
270 
271 	WARN_ON(idx < 0 || idx >= loongarch_pmu.num_counters);
272 
273 	/* Make sure interrupt enabled. */
274 	cpuc->saved_ctrl[idx] = M_PERFCTL_EVENT(evt->event_base & 0xff) |
275 		(evt->config_base & M_PERFCTL_CONFIG_MASK) | CSR_PERFCTRL_IE;
276 
277 	cpu = (event->cpu >= 0) ? event->cpu : smp_processor_id();
278 
279 	/*
280 	 * We do not actually let the counter run. Leave it until start().
281 	 */
282 	pr_debug("Enabling perf counter for CPU%d\n", cpu);
283 }
284 
285 static void loongarch_pmu_disable_event(int idx)
286 {
287 	unsigned long flags;
288 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
289 
290 	WARN_ON(idx < 0 || idx >= loongarch_pmu.num_counters);
291 
292 	local_irq_save(flags);
293 	cpuc->saved_ctrl[idx] = loongarch_pmu_read_control(idx) &
294 		~M_PERFCTL_COUNT_EVENT_WHENEVER;
295 	loongarch_pmu_write_control(idx, cpuc->saved_ctrl[idx]);
296 	local_irq_restore(flags);
297 }
298 
299 static int loongarch_pmu_event_set_period(struct perf_event *event,
300 				    struct hw_perf_event *hwc,
301 				    int idx)
302 {
303 	int ret = 0;
304 	u64 left = local64_read(&hwc->period_left);
305 	u64 period = hwc->sample_period;
306 
307 	if (unlikely((left + period) & (1ULL << 63))) {
308 		/* left underflowed by more than period. */
309 		left = period;
310 		local64_set(&hwc->period_left, left);
311 		hwc->last_period = period;
312 		ret = 1;
313 	} else	if (unlikely((left + period) <= period)) {
314 		/* left underflowed by less than period. */
315 		left += period;
316 		local64_set(&hwc->period_left, left);
317 		hwc->last_period = period;
318 		ret = 1;
319 	}
320 
321 	if (left > loongarch_pmu.max_period) {
322 		left = loongarch_pmu.max_period;
323 		local64_set(&hwc->period_left, left);
324 	}
325 
326 	local64_set(&hwc->prev_count, loongarch_pmu.overflow - left);
327 
328 	loongarch_pmu.write_counter(idx, loongarch_pmu.overflow - left);
329 
330 	perf_event_update_userpage(event);
331 
332 	return ret;
333 }
334 
335 static void loongarch_pmu_event_update(struct perf_event *event,
336 				 struct hw_perf_event *hwc,
337 				 int idx)
338 {
339 	u64 delta;
340 	u64 prev_raw_count, new_raw_count;
341 
342 again:
343 	prev_raw_count = local64_read(&hwc->prev_count);
344 	new_raw_count = loongarch_pmu.read_counter(idx);
345 
346 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
347 				new_raw_count) != prev_raw_count)
348 		goto again;
349 
350 	delta = new_raw_count - prev_raw_count;
351 
352 	local64_add(delta, &event->count);
353 	local64_sub(delta, &hwc->period_left);
354 }
355 
356 static void loongarch_pmu_start(struct perf_event *event, int flags)
357 {
358 	struct hw_perf_event *hwc = &event->hw;
359 
360 	if (flags & PERF_EF_RELOAD)
361 		WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
362 
363 	hwc->state = 0;
364 
365 	/* Set the period for the event. */
366 	loongarch_pmu_event_set_period(event, hwc, hwc->idx);
367 
368 	/* Enable the event. */
369 	loongarch_pmu_enable_event(hwc, hwc->idx);
370 }
371 
372 static void loongarch_pmu_stop(struct perf_event *event, int flags)
373 {
374 	struct hw_perf_event *hwc = &event->hw;
375 
376 	if (!(hwc->state & PERF_HES_STOPPED)) {
377 		/* We are working on a local event. */
378 		loongarch_pmu_disable_event(hwc->idx);
379 		barrier();
380 		loongarch_pmu_event_update(event, hwc, hwc->idx);
381 		hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
382 	}
383 }
384 
385 static int loongarch_pmu_add(struct perf_event *event, int flags)
386 {
387 	int idx, err = 0;
388 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
389 	struct hw_perf_event *hwc = &event->hw;
390 
391 	perf_pmu_disable(event->pmu);
392 
393 	/* To look for a free counter for this event. */
394 	idx = loongarch_pmu_alloc_counter(cpuc, hwc);
395 	if (idx < 0) {
396 		err = idx;
397 		goto out;
398 	}
399 
400 	/*
401 	 * If there is an event in the counter we are going to use then
402 	 * make sure it is disabled.
403 	 */
404 	event->hw.idx = idx;
405 	loongarch_pmu_disable_event(idx);
406 	cpuc->events[idx] = event;
407 
408 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
409 	if (flags & PERF_EF_START)
410 		loongarch_pmu_start(event, PERF_EF_RELOAD);
411 
412 	/* Propagate our changes to the userspace mapping. */
413 	perf_event_update_userpage(event);
414 
415 out:
416 	perf_pmu_enable(event->pmu);
417 	return err;
418 }
419 
420 static void loongarch_pmu_del(struct perf_event *event, int flags)
421 {
422 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
423 	struct hw_perf_event *hwc = &event->hw;
424 	int idx = hwc->idx;
425 
426 	WARN_ON(idx < 0 || idx >= loongarch_pmu.num_counters);
427 
428 	loongarch_pmu_stop(event, PERF_EF_UPDATE);
429 	cpuc->events[idx] = NULL;
430 	clear_bit(idx, cpuc->used_mask);
431 
432 	perf_event_update_userpage(event);
433 }
434 
435 static void loongarch_pmu_read(struct perf_event *event)
436 {
437 	struct hw_perf_event *hwc = &event->hw;
438 
439 	/* Don't read disabled counters! */
440 	if (hwc->idx < 0)
441 		return;
442 
443 	loongarch_pmu_event_update(event, hwc, hwc->idx);
444 }
445 
446 static void loongarch_pmu_enable(struct pmu *pmu)
447 {
448 	resume_local_counters();
449 }
450 
451 static void loongarch_pmu_disable(struct pmu *pmu)
452 {
453 	pause_local_counters();
454 }
455 
456 static DEFINE_MUTEX(pmu_reserve_mutex);
457 static atomic_t active_events = ATOMIC_INIT(0);
458 
459 static int get_pmc_irq(void)
460 {
461 	struct irq_domain *d = irq_find_matching_fwnode(cpuintc_handle, DOMAIN_BUS_ANY);
462 
463 	if (d)
464 		return irq_create_mapping(d, INT_PCOV);
465 
466 	return -EINVAL;
467 }
468 
469 static void reset_counters(void *arg);
470 static int __hw_perf_event_init(struct perf_event *event);
471 
472 static void hw_perf_event_destroy(struct perf_event *event)
473 {
474 	if (atomic_dec_and_mutex_lock(&active_events, &pmu_reserve_mutex)) {
475 		on_each_cpu(reset_counters, NULL, 1);
476 		free_irq(get_pmc_irq(), &loongarch_pmu);
477 		mutex_unlock(&pmu_reserve_mutex);
478 	}
479 }
480 
481 static void handle_associated_event(struct cpu_hw_events *cpuc, int idx,
482 			struct perf_sample_data *data, struct pt_regs *regs)
483 {
484 	struct perf_event *event = cpuc->events[idx];
485 	struct hw_perf_event *hwc = &event->hw;
486 
487 	loongarch_pmu_event_update(event, hwc, idx);
488 	data->period = event->hw.last_period;
489 	if (!loongarch_pmu_event_set_period(event, hwc, idx))
490 		return;
491 
492 	if (perf_event_overflow(event, data, regs))
493 		loongarch_pmu_disable_event(idx);
494 }
495 
496 static irqreturn_t pmu_handle_irq(int irq, void *dev)
497 {
498 	int n;
499 	int handled = IRQ_NONE;
500 	uint64_t counter;
501 	struct pt_regs *regs;
502 	struct perf_sample_data data;
503 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
504 
505 	/*
506 	 * First we pause the local counters, so that when we are locked
507 	 * here, the counters are all paused. When it gets locked due to
508 	 * perf_disable(), the timer interrupt handler will be delayed.
509 	 *
510 	 * See also loongarch_pmu_start().
511 	 */
512 	pause_local_counters();
513 
514 	regs = get_irq_regs();
515 
516 	perf_sample_data_init(&data, 0, 0);
517 
518 	for (n = 0; n < loongarch_pmu.num_counters; n++) {
519 		if (test_bit(n, cpuc->used_mask)) {
520 			counter = loongarch_pmu.read_counter(n);
521 			if (counter & loongarch_pmu.overflow) {
522 				handle_associated_event(cpuc, n, &data, regs);
523 				handled = IRQ_HANDLED;
524 			}
525 		}
526 	}
527 
528 	resume_local_counters();
529 
530 	/*
531 	 * Do all the work for the pending perf events. We can do this
532 	 * in here because the performance counter interrupt is a regular
533 	 * interrupt, not NMI.
534 	 */
535 	if (handled == IRQ_HANDLED)
536 		irq_work_run();
537 
538 	return handled;
539 }
540 
541 static int loongarch_pmu_event_init(struct perf_event *event)
542 {
543 	int r, irq;
544 	unsigned long flags;
545 
546 	/* does not support taken branch sampling */
547 	if (has_branch_stack(event))
548 		return -EOPNOTSUPP;
549 
550 	switch (event->attr.type) {
551 	case PERF_TYPE_RAW:
552 	case PERF_TYPE_HARDWARE:
553 	case PERF_TYPE_HW_CACHE:
554 		break;
555 
556 	default:
557 		/* Init it to avoid false validate_group */
558 		event->hw.event_base = 0xffffffff;
559 		return -ENOENT;
560 	}
561 
562 	if (event->cpu >= 0 && !cpu_online(event->cpu))
563 		return -ENODEV;
564 
565 	irq = get_pmc_irq();
566 	flags = IRQF_PERCPU | IRQF_NOBALANCING | IRQF_NO_THREAD | IRQF_NO_SUSPEND | IRQF_SHARED;
567 	if (!atomic_inc_not_zero(&active_events)) {
568 		mutex_lock(&pmu_reserve_mutex);
569 		if (atomic_read(&active_events) == 0) {
570 			r = request_irq(irq, pmu_handle_irq, flags, "Perf_PMU", &loongarch_pmu);
571 			if (r < 0) {
572 				mutex_unlock(&pmu_reserve_mutex);
573 				pr_warn("PMU IRQ request failed\n");
574 				return -ENODEV;
575 			}
576 		}
577 		atomic_inc(&active_events);
578 		mutex_unlock(&pmu_reserve_mutex);
579 	}
580 
581 	return __hw_perf_event_init(event);
582 }
583 
584 static struct pmu pmu = {
585 	.pmu_enable	= loongarch_pmu_enable,
586 	.pmu_disable	= loongarch_pmu_disable,
587 	.event_init	= loongarch_pmu_event_init,
588 	.add		= loongarch_pmu_add,
589 	.del		= loongarch_pmu_del,
590 	.start		= loongarch_pmu_start,
591 	.stop		= loongarch_pmu_stop,
592 	.read		= loongarch_pmu_read,
593 };
594 
595 static unsigned int loongarch_pmu_perf_event_encode(const struct loongarch_perf_event *pev)
596 {
597 	return (pev->event_id & 0xff);
598 }
599 
600 static const struct loongarch_perf_event *loongarch_pmu_map_general_event(int idx)
601 {
602 	const struct loongarch_perf_event *pev;
603 
604 	pev = &(*loongarch_pmu.general_event_map)[idx];
605 
606 	if (pev->event_id == HW_OP_UNSUPPORTED)
607 		return ERR_PTR(-ENOENT);
608 
609 	return pev;
610 }
611 
612 static const struct loongarch_perf_event *loongarch_pmu_map_cache_event(u64 config)
613 {
614 	unsigned int cache_type, cache_op, cache_result;
615 	const struct loongarch_perf_event *pev;
616 
617 	cache_type = (config >> 0) & 0xff;
618 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
619 		return ERR_PTR(-EINVAL);
620 
621 	cache_op = (config >> 8) & 0xff;
622 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
623 		return ERR_PTR(-EINVAL);
624 
625 	cache_result = (config >> 16) & 0xff;
626 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
627 		return ERR_PTR(-EINVAL);
628 
629 	pev = &((*loongarch_pmu.cache_event_map)
630 					[cache_type]
631 					[cache_op]
632 					[cache_result]);
633 
634 	if (pev->event_id == CACHE_OP_UNSUPPORTED)
635 		return ERR_PTR(-ENOENT);
636 
637 	return pev;
638 }
639 
640 static int validate_group(struct perf_event *event)
641 {
642 	struct cpu_hw_events fake_cpuc;
643 	struct perf_event *sibling, *leader = event->group_leader;
644 
645 	memset(&fake_cpuc, 0, sizeof(fake_cpuc));
646 
647 	if (loongarch_pmu_alloc_counter(&fake_cpuc, &leader->hw) < 0)
648 		return -EINVAL;
649 
650 	for_each_sibling_event(sibling, leader) {
651 		if (loongarch_pmu_alloc_counter(&fake_cpuc, &sibling->hw) < 0)
652 			return -EINVAL;
653 	}
654 
655 	if (loongarch_pmu_alloc_counter(&fake_cpuc, &event->hw) < 0)
656 		return -EINVAL;
657 
658 	return 0;
659 }
660 
661 static void reset_counters(void *arg)
662 {
663 	int n;
664 	int counters = loongarch_pmu.num_counters;
665 
666 	for (n = 0; n < counters; n++) {
667 		loongarch_pmu_write_control(n, 0);
668 		loongarch_pmu.write_counter(n, 0);
669 	}
670 }
671 
672 static const struct loongarch_perf_event loongson_event_map[PERF_COUNT_HW_MAX] = {
673 	PERF_MAP_ALL_UNSUPPORTED,
674 	[PERF_COUNT_HW_CPU_CYCLES] = { 0x00 },
675 	[PERF_COUNT_HW_INSTRUCTIONS] = { 0x01 },
676 	[PERF_COUNT_HW_CACHE_REFERENCES] = { 0x08 },
677 	[PERF_COUNT_HW_CACHE_MISSES] = { 0x09 },
678 	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = { 0x02 },
679 	[PERF_COUNT_HW_BRANCH_MISSES] = { 0x03 },
680 };
681 
682 static const struct loongarch_perf_event loongson_cache_map
683 				[PERF_COUNT_HW_CACHE_MAX]
684 				[PERF_COUNT_HW_CACHE_OP_MAX]
685 				[PERF_COUNT_HW_CACHE_RESULT_MAX] = {
686 PERF_CACHE_MAP_ALL_UNSUPPORTED,
687 [C(L1D)] = {
688 	/*
689 	 * Like some other architectures (e.g. ARM), the performance
690 	 * counters don't differentiate between read and write
691 	 * accesses/misses, so this isn't strictly correct, but it's the
692 	 * best we can do. Writes and reads get combined.
693 	 */
694 	[C(OP_READ)] = {
695 		[C(RESULT_ACCESS)]	= { 0x8 },
696 		[C(RESULT_MISS)]	= { 0x9 },
697 	},
698 	[C(OP_WRITE)] = {
699 		[C(RESULT_ACCESS)]	= { 0x8 },
700 		[C(RESULT_MISS)]	= { 0x9 },
701 	},
702 	[C(OP_PREFETCH)] = {
703 		[C(RESULT_ACCESS)]	= { 0xaa },
704 		[C(RESULT_MISS)]	= { 0xa9 },
705 	},
706 },
707 [C(L1I)] = {
708 	[C(OP_READ)] = {
709 		[C(RESULT_ACCESS)]	= { 0x6 },
710 		[C(RESULT_MISS)]	= { 0x7 },
711 	},
712 },
713 [C(LL)] = {
714 	[C(OP_READ)] = {
715 		[C(RESULT_ACCESS)]	= { 0xc },
716 		[C(RESULT_MISS)]	= { 0xd },
717 	},
718 	[C(OP_WRITE)] = {
719 		[C(RESULT_ACCESS)]	= { 0xc },
720 		[C(RESULT_MISS)]	= { 0xd },
721 	},
722 },
723 [C(ITLB)] = {
724 	[C(OP_READ)] = {
725 		[C(RESULT_MISS)]    = { 0x3b },
726 	},
727 },
728 [C(DTLB)] = {
729 	[C(OP_READ)] = {
730 		[C(RESULT_ACCESS)]	= { 0x4 },
731 		[C(RESULT_MISS)]	= { 0x3c },
732 	},
733 	[C(OP_WRITE)] = {
734 		[C(RESULT_ACCESS)]	= { 0x4 },
735 		[C(RESULT_MISS)]	= { 0x3c },
736 	},
737 },
738 [C(BPU)] = {
739 	/* Using the same code for *HW_BRANCH* */
740 	[C(OP_READ)] = {
741 		[C(RESULT_ACCESS)]  = { 0x02 },
742 		[C(RESULT_MISS)]    = { 0x03 },
743 	},
744 },
745 };
746 
747 static int __hw_perf_event_init(struct perf_event *event)
748 {
749 	int err;
750 	struct hw_perf_event *hwc = &event->hw;
751 	struct perf_event_attr *attr = &event->attr;
752 	const struct loongarch_perf_event *pev;
753 
754 	/* Returning LoongArch event descriptor for generic perf event. */
755 	if (PERF_TYPE_HARDWARE == event->attr.type) {
756 		if (event->attr.config >= PERF_COUNT_HW_MAX)
757 			return -EINVAL;
758 		pev = loongarch_pmu_map_general_event(event->attr.config);
759 	} else if (PERF_TYPE_HW_CACHE == event->attr.type) {
760 		pev = loongarch_pmu_map_cache_event(event->attr.config);
761 	} else if (PERF_TYPE_RAW == event->attr.type) {
762 		/* We are working on the global raw event. */
763 		mutex_lock(&raw_event_mutex);
764 		pev = loongarch_pmu.map_raw_event(event->attr.config);
765 	} else {
766 		/* The event type is not (yet) supported. */
767 		return -EOPNOTSUPP;
768 	}
769 
770 	if (IS_ERR(pev)) {
771 		if (PERF_TYPE_RAW == event->attr.type)
772 			mutex_unlock(&raw_event_mutex);
773 		return PTR_ERR(pev);
774 	}
775 
776 	/*
777 	 * We allow max flexibility on how each individual counter shared
778 	 * by the single CPU operates (the mode exclusion and the range).
779 	 */
780 	hwc->config_base = CSR_PERFCTRL_IE;
781 
782 	hwc->event_base = loongarch_pmu_perf_event_encode(pev);
783 	if (PERF_TYPE_RAW == event->attr.type)
784 		mutex_unlock(&raw_event_mutex);
785 
786 	if (!attr->exclude_user) {
787 		hwc->config_base |= CSR_PERFCTRL_PLV3;
788 		hwc->config_base |= CSR_PERFCTRL_PLV2;
789 	}
790 	if (!attr->exclude_kernel) {
791 		hwc->config_base |= CSR_PERFCTRL_PLV0;
792 	}
793 	if (!attr->exclude_hv) {
794 		hwc->config_base |= CSR_PERFCTRL_PLV1;
795 	}
796 
797 	hwc->config_base &= M_PERFCTL_CONFIG_MASK;
798 	/*
799 	 * The event can belong to another cpu. We do not assign a local
800 	 * counter for it for now.
801 	 */
802 	hwc->idx = -1;
803 	hwc->config = 0;
804 
805 	if (!hwc->sample_period) {
806 		hwc->sample_period  = loongarch_pmu.max_period;
807 		hwc->last_period    = hwc->sample_period;
808 		local64_set(&hwc->period_left, hwc->sample_period);
809 	}
810 
811 	err = 0;
812 	if (event->group_leader != event)
813 		err = validate_group(event);
814 
815 	event->destroy = hw_perf_event_destroy;
816 
817 	if (err)
818 		event->destroy(event);
819 
820 	return err;
821 }
822 
823 static void pause_local_counters(void)
824 {
825 	unsigned long flags;
826 	int ctr = loongarch_pmu.num_counters;
827 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
828 
829 	local_irq_save(flags);
830 	do {
831 		ctr--;
832 		cpuc->saved_ctrl[ctr] = loongarch_pmu_read_control(ctr);
833 		loongarch_pmu_write_control(ctr, cpuc->saved_ctrl[ctr] &
834 					 ~M_PERFCTL_COUNT_EVENT_WHENEVER);
835 	} while (ctr > 0);
836 	local_irq_restore(flags);
837 }
838 
839 static void resume_local_counters(void)
840 {
841 	int ctr = loongarch_pmu.num_counters;
842 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
843 
844 	do {
845 		ctr--;
846 		loongarch_pmu_write_control(ctr, cpuc->saved_ctrl[ctr]);
847 	} while (ctr > 0);
848 }
849 
850 static const struct loongarch_perf_event *loongarch_pmu_map_raw_event(u64 config)
851 {
852 	raw_event.event_id = config & 0xff;
853 
854 	return &raw_event;
855 }
856 
857 static int __init init_hw_perf_events(void)
858 {
859 	int counters;
860 
861 	if (!cpu_has_pmp)
862 		return -ENODEV;
863 
864 	pr_info("Performance counters: ");
865 	counters = ((read_cpucfg(LOONGARCH_CPUCFG6) & CPUCFG6_PMNUM) >> 4) + 1;
866 
867 	loongarch_pmu.num_counters = counters;
868 	loongarch_pmu.max_period = (1ULL << 63) - 1;
869 	loongarch_pmu.valid_count = (1ULL << 63) - 1;
870 	loongarch_pmu.overflow = 1ULL << 63;
871 	loongarch_pmu.name = "loongarch/loongson64";
872 	loongarch_pmu.read_counter = loongarch_pmu_read_counter;
873 	loongarch_pmu.write_counter = loongarch_pmu_write_counter;
874 	loongarch_pmu.map_raw_event = loongarch_pmu_map_raw_event;
875 	loongarch_pmu.general_event_map = &loongson_event_map;
876 	loongarch_pmu.cache_event_map = &loongson_cache_map;
877 
878 	on_each_cpu(reset_counters, NULL, 1);
879 
880 	pr_cont("%s PMU enabled, %d %d-bit counters available to each CPU.\n",
881 			loongarch_pmu.name, counters, 64);
882 
883 	perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
884 
885 	return 0;
886 }
887 early_initcall(init_hw_perf_events);
888