xref: /linux/arch/powerpc/perf/core-fsl-emb.c (revision 68550cbc6129159b7a6434796b721e8b66ee12f6)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Performance event support - Freescale Embedded Performance Monitor
4  *
5  * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
6  * Copyright 2010 Freescale Semiconductor, Inc.
7  */
8 #include <linux/kernel.h>
9 #include <linux/sched.h>
10 #include <linux/perf_event.h>
11 #include <linux/percpu.h>
12 #include <linux/hardirq.h>
13 #include <asm/reg_fsl_emb.h>
14 #include <asm/pmc.h>
15 #include <asm/machdep.h>
16 #include <asm/firmware.h>
17 #include <asm/ptrace.h>
18 
19 struct cpu_hw_events {
20 	int n_events;
21 	int disabled;
22 	u8  pmcs_enabled;
23 	struct perf_event *event[MAX_HWEVENTS];
24 };
25 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
26 
27 static struct fsl_emb_pmu *ppmu;
28 
29 /* Number of perf_events counting hardware events */
30 static atomic_t num_events;
31 /* Used to avoid races in calling reserve/release_pmc_hardware */
32 static DEFINE_MUTEX(pmc_reserve_mutex);
33 
34 static void perf_event_interrupt(struct pt_regs *regs);
35 
36 /*
37  * Read one performance monitor counter (PMC).
38  */
39 static unsigned long read_pmc(int idx)
40 {
41 	unsigned long val;
42 
43 	switch (idx) {
44 	case 0:
45 		val = mfpmr(PMRN_PMC0);
46 		break;
47 	case 1:
48 		val = mfpmr(PMRN_PMC1);
49 		break;
50 	case 2:
51 		val = mfpmr(PMRN_PMC2);
52 		break;
53 	case 3:
54 		val = mfpmr(PMRN_PMC3);
55 		break;
56 	case 4:
57 		val = mfpmr(PMRN_PMC4);
58 		break;
59 	case 5:
60 		val = mfpmr(PMRN_PMC5);
61 		break;
62 	default:
63 		printk(KERN_ERR "oops trying to read PMC%d\n", idx);
64 		val = 0;
65 	}
66 	return val;
67 }
68 
69 /*
70  * Write one PMC.
71  */
72 static void write_pmc(int idx, unsigned long val)
73 {
74 	switch (idx) {
75 	case 0:
76 		mtpmr(PMRN_PMC0, val);
77 		break;
78 	case 1:
79 		mtpmr(PMRN_PMC1, val);
80 		break;
81 	case 2:
82 		mtpmr(PMRN_PMC2, val);
83 		break;
84 	case 3:
85 		mtpmr(PMRN_PMC3, val);
86 		break;
87 	case 4:
88 		mtpmr(PMRN_PMC4, val);
89 		break;
90 	case 5:
91 		mtpmr(PMRN_PMC5, val);
92 		break;
93 	default:
94 		printk(KERN_ERR "oops trying to write PMC%d\n", idx);
95 	}
96 
97 	isync();
98 }
99 
100 /*
101  * Write one local control A register
102  */
103 static void write_pmlca(int idx, unsigned long val)
104 {
105 	switch (idx) {
106 	case 0:
107 		mtpmr(PMRN_PMLCA0, val);
108 		break;
109 	case 1:
110 		mtpmr(PMRN_PMLCA1, val);
111 		break;
112 	case 2:
113 		mtpmr(PMRN_PMLCA2, val);
114 		break;
115 	case 3:
116 		mtpmr(PMRN_PMLCA3, val);
117 		break;
118 	case 4:
119 		mtpmr(PMRN_PMLCA4, val);
120 		break;
121 	case 5:
122 		mtpmr(PMRN_PMLCA5, val);
123 		break;
124 	default:
125 		printk(KERN_ERR "oops trying to write PMLCA%d\n", idx);
126 	}
127 
128 	isync();
129 }
130 
131 /*
132  * Write one local control B register
133  */
134 static void write_pmlcb(int idx, unsigned long val)
135 {
136 	switch (idx) {
137 	case 0:
138 		mtpmr(PMRN_PMLCB0, val);
139 		break;
140 	case 1:
141 		mtpmr(PMRN_PMLCB1, val);
142 		break;
143 	case 2:
144 		mtpmr(PMRN_PMLCB2, val);
145 		break;
146 	case 3:
147 		mtpmr(PMRN_PMLCB3, val);
148 		break;
149 	case 4:
150 		mtpmr(PMRN_PMLCB4, val);
151 		break;
152 	case 5:
153 		mtpmr(PMRN_PMLCB5, val);
154 		break;
155 	default:
156 		printk(KERN_ERR "oops trying to write PMLCB%d\n", idx);
157 	}
158 
159 	isync();
160 }
161 
162 static void fsl_emb_pmu_read(struct perf_event *event)
163 {
164 	s64 val, delta, prev;
165 
166 	if (event->hw.state & PERF_HES_STOPPED)
167 		return;
168 
169 	/*
170 	 * Performance monitor interrupts come even when interrupts
171 	 * are soft-disabled, as long as interrupts are hard-enabled.
172 	 * Therefore we treat them like NMIs.
173 	 */
174 	do {
175 		prev = local64_read(&event->hw.prev_count);
176 		barrier();
177 		val = read_pmc(event->hw.idx);
178 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
179 
180 	/* The counters are only 32 bits wide */
181 	delta = (val - prev) & 0xfffffffful;
182 	local64_add(delta, &event->count);
183 	local64_sub(delta, &event->hw.period_left);
184 }
185 
186 /*
187  * Disable all events to prevent PMU interrupts and to allow
188  * events to be added or removed.
189  */
190 static void fsl_emb_pmu_disable(struct pmu *pmu)
191 {
192 	struct cpu_hw_events *cpuhw;
193 	unsigned long flags;
194 
195 	local_irq_save(flags);
196 	cpuhw = this_cpu_ptr(&cpu_hw_events);
197 
198 	if (!cpuhw->disabled) {
199 		cpuhw->disabled = 1;
200 
201 		/*
202 		 * Check if we ever enabled the PMU on this cpu.
203 		 */
204 		if (!cpuhw->pmcs_enabled) {
205 			ppc_enable_pmcs();
206 			cpuhw->pmcs_enabled = 1;
207 		}
208 
209 		if (atomic_read(&num_events)) {
210 			/*
211 			 * Set the 'freeze all counters' bit, and disable
212 			 * interrupts.  The barrier is to make sure the
213 			 * mtpmr has been executed and the PMU has frozen
214 			 * the events before we return.
215 			 */
216 
217 			mtpmr(PMRN_PMGC0, PMGC0_FAC);
218 			isync();
219 		}
220 	}
221 	local_irq_restore(flags);
222 }
223 
224 /*
225  * Re-enable all events if disable == 0.
226  * If we were previously disabled and events were added, then
227  * put the new config on the PMU.
228  */
229 static void fsl_emb_pmu_enable(struct pmu *pmu)
230 {
231 	struct cpu_hw_events *cpuhw;
232 	unsigned long flags;
233 
234 	local_irq_save(flags);
235 	cpuhw = this_cpu_ptr(&cpu_hw_events);
236 	if (!cpuhw->disabled)
237 		goto out;
238 
239 	cpuhw->disabled = 0;
240 	ppc_set_pmu_inuse(cpuhw->n_events != 0);
241 
242 	if (cpuhw->n_events > 0) {
243 		mtpmr(PMRN_PMGC0, PMGC0_PMIE | PMGC0_FCECE);
244 		isync();
245 	}
246 
247  out:
248 	local_irq_restore(flags);
249 }
250 
251 static int collect_events(struct perf_event *group, int max_count,
252 			  struct perf_event *ctrs[])
253 {
254 	int n = 0;
255 	struct perf_event *event;
256 
257 	if (!is_software_event(group)) {
258 		if (n >= max_count)
259 			return -1;
260 		ctrs[n] = group;
261 		n++;
262 	}
263 	for_each_sibling_event(event, group) {
264 		if (!is_software_event(event) &&
265 		    event->state != PERF_EVENT_STATE_OFF) {
266 			if (n >= max_count)
267 				return -1;
268 			ctrs[n] = event;
269 			n++;
270 		}
271 	}
272 	return n;
273 }
274 
275 /* context locked on entry */
276 static int fsl_emb_pmu_add(struct perf_event *event, int flags)
277 {
278 	struct cpu_hw_events *cpuhw;
279 	int ret = -EAGAIN;
280 	int num_counters = ppmu->n_counter;
281 	u64 val;
282 	int i;
283 
284 	perf_pmu_disable(event->pmu);
285 	cpuhw = &get_cpu_var(cpu_hw_events);
286 
287 	if (event->hw.config & FSL_EMB_EVENT_RESTRICTED)
288 		num_counters = ppmu->n_restricted;
289 
290 	/*
291 	 * Allocate counters from top-down, so that restricted-capable
292 	 * counters are kept free as long as possible.
293 	 */
294 	for (i = num_counters - 1; i >= 0; i--) {
295 		if (cpuhw->event[i])
296 			continue;
297 
298 		break;
299 	}
300 
301 	if (i < 0)
302 		goto out;
303 
304 	event->hw.idx = i;
305 	cpuhw->event[i] = event;
306 	++cpuhw->n_events;
307 
308 	val = 0;
309 	if (event->hw.sample_period) {
310 		s64 left = local64_read(&event->hw.period_left);
311 		if (left < 0x80000000L)
312 			val = 0x80000000L - left;
313 	}
314 	local64_set(&event->hw.prev_count, val);
315 
316 	if (unlikely(!(flags & PERF_EF_START))) {
317 		event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
318 		val = 0;
319 	} else {
320 		event->hw.state &= ~(PERF_HES_STOPPED | PERF_HES_UPTODATE);
321 	}
322 
323 	write_pmc(i, val);
324 	perf_event_update_userpage(event);
325 
326 	write_pmlcb(i, event->hw.config >> 32);
327 	write_pmlca(i, event->hw.config_base);
328 
329 	ret = 0;
330  out:
331 	put_cpu_var(cpu_hw_events);
332 	perf_pmu_enable(event->pmu);
333 	return ret;
334 }
335 
336 /* context locked on entry */
337 static void fsl_emb_pmu_del(struct perf_event *event, int flags)
338 {
339 	struct cpu_hw_events *cpuhw;
340 	int i = event->hw.idx;
341 
342 	perf_pmu_disable(event->pmu);
343 	if (i < 0)
344 		goto out;
345 
346 	fsl_emb_pmu_read(event);
347 
348 	cpuhw = &get_cpu_var(cpu_hw_events);
349 
350 	WARN_ON(event != cpuhw->event[event->hw.idx]);
351 
352 	write_pmlca(i, 0);
353 	write_pmlcb(i, 0);
354 	write_pmc(i, 0);
355 
356 	cpuhw->event[i] = NULL;
357 	event->hw.idx = -1;
358 
359 	/*
360 	 * TODO: if at least one restricted event exists, and we
361 	 * just freed up a non-restricted-capable counter, and
362 	 * there is a restricted-capable counter occupied by
363 	 * a non-restricted event, migrate that event to the
364 	 * vacated counter.
365 	 */
366 
367 	cpuhw->n_events--;
368 
369  out:
370 	perf_pmu_enable(event->pmu);
371 	put_cpu_var(cpu_hw_events);
372 }
373 
374 static void fsl_emb_pmu_start(struct perf_event *event, int ef_flags)
375 {
376 	unsigned long flags;
377 	unsigned long val;
378 	s64 left;
379 
380 	if (event->hw.idx < 0 || !event->hw.sample_period)
381 		return;
382 
383 	if (!(event->hw.state & PERF_HES_STOPPED))
384 		return;
385 
386 	if (ef_flags & PERF_EF_RELOAD)
387 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
388 
389 	local_irq_save(flags);
390 	perf_pmu_disable(event->pmu);
391 
392 	event->hw.state = 0;
393 	left = local64_read(&event->hw.period_left);
394 	val = 0;
395 	if (left < 0x80000000L)
396 		val = 0x80000000L - left;
397 	write_pmc(event->hw.idx, val);
398 
399 	perf_event_update_userpage(event);
400 	perf_pmu_enable(event->pmu);
401 	local_irq_restore(flags);
402 }
403 
404 static void fsl_emb_pmu_stop(struct perf_event *event, int ef_flags)
405 {
406 	unsigned long flags;
407 
408 	if (event->hw.idx < 0 || !event->hw.sample_period)
409 		return;
410 
411 	if (event->hw.state & PERF_HES_STOPPED)
412 		return;
413 
414 	local_irq_save(flags);
415 	perf_pmu_disable(event->pmu);
416 
417 	fsl_emb_pmu_read(event);
418 	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
419 	write_pmc(event->hw.idx, 0);
420 
421 	perf_event_update_userpage(event);
422 	perf_pmu_enable(event->pmu);
423 	local_irq_restore(flags);
424 }
425 
426 /*
427  * Release the PMU if this is the last perf_event.
428  */
429 static void hw_perf_event_destroy(struct perf_event *event)
430 {
431 	if (!atomic_add_unless(&num_events, -1, 1)) {
432 		mutex_lock(&pmc_reserve_mutex);
433 		if (atomic_dec_return(&num_events) == 0)
434 			release_pmc_hardware();
435 		mutex_unlock(&pmc_reserve_mutex);
436 	}
437 }
438 
439 /*
440  * Translate a generic cache event_id config to a raw event_id code.
441  */
442 static int hw_perf_cache_event(u64 config, u64 *eventp)
443 {
444 	unsigned long type, op, result;
445 	int ev;
446 
447 	if (!ppmu->cache_events)
448 		return -EINVAL;
449 
450 	/* unpack config */
451 	type = config & 0xff;
452 	op = (config >> 8) & 0xff;
453 	result = (config >> 16) & 0xff;
454 
455 	if (type >= PERF_COUNT_HW_CACHE_MAX ||
456 	    op >= PERF_COUNT_HW_CACHE_OP_MAX ||
457 	    result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
458 		return -EINVAL;
459 
460 	ev = (*ppmu->cache_events)[type][op][result];
461 	if (ev == 0)
462 		return -EOPNOTSUPP;
463 	if (ev == -1)
464 		return -EINVAL;
465 	*eventp = ev;
466 	return 0;
467 }
468 
469 static int fsl_emb_pmu_event_init(struct perf_event *event)
470 {
471 	u64 ev;
472 	struct perf_event *events[MAX_HWEVENTS];
473 	int n;
474 	int err;
475 	int num_restricted;
476 	int i;
477 
478 	if (ppmu->n_counter > MAX_HWEVENTS) {
479 		WARN(1, "No. of perf counters (%d) is higher than max array size(%d)\n",
480 			ppmu->n_counter, MAX_HWEVENTS);
481 		ppmu->n_counter = MAX_HWEVENTS;
482 	}
483 
484 	switch (event->attr.type) {
485 	case PERF_TYPE_HARDWARE:
486 		ev = event->attr.config;
487 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
488 			return -EOPNOTSUPP;
489 		ev = ppmu->generic_events[ev];
490 		break;
491 
492 	case PERF_TYPE_HW_CACHE:
493 		err = hw_perf_cache_event(event->attr.config, &ev);
494 		if (err)
495 			return err;
496 		break;
497 
498 	case PERF_TYPE_RAW:
499 		ev = event->attr.config;
500 		break;
501 
502 	default:
503 		return -ENOENT;
504 	}
505 
506 	event->hw.config = ppmu->xlate_event(ev);
507 	if (!(event->hw.config & FSL_EMB_EVENT_VALID))
508 		return -EINVAL;
509 
510 	/*
511 	 * If this is in a group, check if it can go on with all the
512 	 * other hardware events in the group.  We assume the event
513 	 * hasn't been linked into its leader's sibling list at this point.
514 	 */
515 	n = 0;
516 	if (event->group_leader != event) {
517 		n = collect_events(event->group_leader,
518 		                   ppmu->n_counter - 1, events);
519 		if (n < 0)
520 			return -EINVAL;
521 	}
522 
523 	if (event->hw.config & FSL_EMB_EVENT_RESTRICTED) {
524 		num_restricted = 0;
525 		for (i = 0; i < n; i++) {
526 			if (events[i]->hw.config & FSL_EMB_EVENT_RESTRICTED)
527 				num_restricted++;
528 		}
529 
530 		if (num_restricted >= ppmu->n_restricted)
531 			return -EINVAL;
532 	}
533 
534 	event->hw.idx = -1;
535 
536 	event->hw.config_base = PMLCA_CE | PMLCA_FCM1 |
537 	                        (u32)((ev << 16) & PMLCA_EVENT_MASK);
538 
539 	if (event->attr.exclude_user)
540 		event->hw.config_base |= PMLCA_FCU;
541 	if (event->attr.exclude_kernel)
542 		event->hw.config_base |= PMLCA_FCS;
543 	if (event->attr.exclude_idle)
544 		return -ENOTSUPP;
545 
546 	event->hw.last_period = event->hw.sample_period;
547 	local64_set(&event->hw.period_left, event->hw.last_period);
548 
549 	/*
550 	 * See if we need to reserve the PMU.
551 	 * If no events are currently in use, then we have to take a
552 	 * mutex to ensure that we don't race with another task doing
553 	 * reserve_pmc_hardware or release_pmc_hardware.
554 	 */
555 	err = 0;
556 	if (!atomic_inc_not_zero(&num_events)) {
557 		mutex_lock(&pmc_reserve_mutex);
558 		if (atomic_read(&num_events) == 0 &&
559 		    reserve_pmc_hardware(perf_event_interrupt))
560 			err = -EBUSY;
561 		else
562 			atomic_inc(&num_events);
563 		mutex_unlock(&pmc_reserve_mutex);
564 
565 		mtpmr(PMRN_PMGC0, PMGC0_FAC);
566 		isync();
567 	}
568 	event->destroy = hw_perf_event_destroy;
569 
570 	return err;
571 }
572 
573 static struct pmu fsl_emb_pmu = {
574 	.pmu_enable	= fsl_emb_pmu_enable,
575 	.pmu_disable	= fsl_emb_pmu_disable,
576 	.event_init	= fsl_emb_pmu_event_init,
577 	.add		= fsl_emb_pmu_add,
578 	.del		= fsl_emb_pmu_del,
579 	.start		= fsl_emb_pmu_start,
580 	.stop		= fsl_emb_pmu_stop,
581 	.read		= fsl_emb_pmu_read,
582 };
583 
584 /*
585  * A counter has overflowed; update its count and record
586  * things if requested.  Note that interrupts are hard-disabled
587  * here so there is no possibility of being interrupted.
588  */
589 static void record_and_restart(struct perf_event *event, unsigned long val,
590 			       struct pt_regs *regs)
591 {
592 	u64 period = event->hw.sample_period;
593 	s64 prev, delta, left;
594 	int record = 0;
595 
596 	if (event->hw.state & PERF_HES_STOPPED) {
597 		write_pmc(event->hw.idx, 0);
598 		return;
599 	}
600 
601 	/* we don't have to worry about interrupts here */
602 	prev = local64_read(&event->hw.prev_count);
603 	delta = (val - prev) & 0xfffffffful;
604 	local64_add(delta, &event->count);
605 
606 	/*
607 	 * See if the total period for this event has expired,
608 	 * and update for the next period.
609 	 */
610 	val = 0;
611 	left = local64_read(&event->hw.period_left) - delta;
612 	if (period) {
613 		if (left <= 0) {
614 			left += period;
615 			if (left <= 0)
616 				left = period;
617 			record = 1;
618 			event->hw.last_period = event->hw.sample_period;
619 		}
620 		if (left < 0x80000000LL)
621 			val = 0x80000000LL - left;
622 	}
623 
624 	write_pmc(event->hw.idx, val);
625 	local64_set(&event->hw.prev_count, val);
626 	local64_set(&event->hw.period_left, left);
627 	perf_event_update_userpage(event);
628 
629 	/*
630 	 * Finally record data if requested.
631 	 */
632 	if (record) {
633 		struct perf_sample_data data;
634 
635 		perf_sample_data_init(&data, 0, event->hw.last_period);
636 
637 		if (perf_event_overflow(event, &data, regs))
638 			fsl_emb_pmu_stop(event, 0);
639 	}
640 }
641 
642 static void perf_event_interrupt(struct pt_regs *regs)
643 {
644 	int i;
645 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
646 	struct perf_event *event;
647 	unsigned long val;
648 	int found = 0;
649 
650 	for (i = 0; i < ppmu->n_counter; ++i) {
651 		event = cpuhw->event[i];
652 
653 		val = read_pmc(i);
654 		if ((int)val < 0) {
655 			if (event) {
656 				/* event has overflowed */
657 				found = 1;
658 				record_and_restart(event, val, regs);
659 			} else {
660 				/*
661 				 * Disabled counter is negative,
662 				 * reset it just in case.
663 				 */
664 				write_pmc(i, 0);
665 			}
666 		}
667 	}
668 
669 	/* PMM will keep counters frozen until we return from the interrupt. */
670 	mtmsr(mfmsr() | MSR_PMM);
671 	mtpmr(PMRN_PMGC0, PMGC0_PMIE | PMGC0_FCECE);
672 	isync();
673 }
674 
675 void hw_perf_event_setup(int cpu)
676 {
677 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
678 
679 	memset(cpuhw, 0, sizeof(*cpuhw));
680 }
681 
682 int register_fsl_emb_pmu(struct fsl_emb_pmu *pmu)
683 {
684 	if (ppmu)
685 		return -EBUSY;		/* something's already registered */
686 
687 	ppmu = pmu;
688 	pr_info("%s performance monitor hardware support registered\n",
689 		pmu->name);
690 
691 	perf_pmu_register(&fsl_emb_pmu, "cpu", PERF_TYPE_RAW);
692 
693 	return 0;
694 }
695