xref: /linux/arch/powerpc/perf/core-book3s.c (revision 95e9fd10f06cb5642028b6b851e32b8c8afb4571)
1 /*
2  * Performance event support - powerpc architecture code
3  *
4  * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version
9  * 2 of the License, or (at your option) any later version.
10  */
11 #include <linux/kernel.h>
12 #include <linux/sched.h>
13 #include <linux/perf_event.h>
14 #include <linux/percpu.h>
15 #include <linux/hardirq.h>
16 #include <asm/reg.h>
17 #include <asm/pmc.h>
18 #include <asm/machdep.h>
19 #include <asm/firmware.h>
20 #include <asm/ptrace.h>
21 
22 struct cpu_hw_events {
23 	int n_events;
24 	int n_percpu;
25 	int disabled;
26 	int n_added;
27 	int n_limited;
28 	u8  pmcs_enabled;
29 	struct perf_event *event[MAX_HWEVENTS];
30 	u64 events[MAX_HWEVENTS];
31 	unsigned int flags[MAX_HWEVENTS];
32 	unsigned long mmcr[3];
33 	struct perf_event *limited_counter[MAX_LIMITED_HWCOUNTERS];
34 	u8  limited_hwidx[MAX_LIMITED_HWCOUNTERS];
35 	u64 alternatives[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
36 	unsigned long amasks[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
37 	unsigned long avalues[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
38 
39 	unsigned int group_flag;
40 	int n_txn_start;
41 };
42 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
43 
44 struct power_pmu *ppmu;
45 
46 /*
47  * Normally, to ignore kernel events we set the FCS (freeze counters
48  * in supervisor mode) bit in MMCR0, but if the kernel runs with the
49  * hypervisor bit set in the MSR, or if we are running on a processor
50  * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
51  * then we need to use the FCHV bit to ignore kernel events.
52  */
53 static unsigned int freeze_events_kernel = MMCR0_FCS;
54 
55 /*
56  * 32-bit doesn't have MMCRA but does have an MMCR2,
57  * and a few other names are different.
58  */
59 #ifdef CONFIG_PPC32
60 
61 #define MMCR0_FCHV		0
62 #define MMCR0_PMCjCE		MMCR0_PMCnCE
63 
64 #define SPRN_MMCRA		SPRN_MMCR2
65 #define MMCRA_SAMPLE_ENABLE	0
66 
67 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
68 {
69 	return 0;
70 }
71 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
72 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
73 {
74 	return 0;
75 }
76 static inline void perf_read_regs(struct pt_regs *regs)
77 {
78 	regs->result = 0;
79 }
80 static inline int perf_intr_is_nmi(struct pt_regs *regs)
81 {
82 	return 0;
83 }
84 
85 #endif /* CONFIG_PPC32 */
86 
87 /*
88  * Things that are specific to 64-bit implementations.
89  */
90 #ifdef CONFIG_PPC64
91 
92 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
93 {
94 	unsigned long mmcra = regs->dsisr;
95 
96 	if ((mmcra & MMCRA_SAMPLE_ENABLE) && !(ppmu->flags & PPMU_ALT_SIPR)) {
97 		unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
98 		if (slot > 1)
99 			return 4 * (slot - 1);
100 	}
101 	return 0;
102 }
103 
104 /*
105  * The user wants a data address recorded.
106  * If we're not doing instruction sampling, give them the SDAR
107  * (sampled data address).  If we are doing instruction sampling, then
108  * only give them the SDAR if it corresponds to the instruction
109  * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC
110  * bit in MMCRA.
111  */
112 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
113 {
114 	unsigned long mmcra = regs->dsisr;
115 	unsigned long sdsync = (ppmu->flags & PPMU_ALT_SIPR) ?
116 		POWER6_MMCRA_SDSYNC : MMCRA_SDSYNC;
117 
118 	if (!(mmcra & MMCRA_SAMPLE_ENABLE) || (mmcra & sdsync))
119 		*addrp = mfspr(SPRN_SDAR);
120 }
121 
122 static bool mmcra_sihv(unsigned long mmcra)
123 {
124 	unsigned long sihv = MMCRA_SIHV;
125 
126 	if (ppmu->flags & PPMU_ALT_SIPR)
127 		sihv = POWER6_MMCRA_SIHV;
128 
129 	return !!(mmcra & sihv);
130 }
131 
132 static bool mmcra_sipr(unsigned long mmcra)
133 {
134 	unsigned long sipr = MMCRA_SIPR;
135 
136 	if (ppmu->flags & PPMU_ALT_SIPR)
137 		sipr = POWER6_MMCRA_SIPR;
138 
139 	return !!(mmcra & sipr);
140 }
141 
142 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
143 {
144 	if (regs->msr & MSR_PR)
145 		return PERF_RECORD_MISC_USER;
146 	if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
147 		return PERF_RECORD_MISC_HYPERVISOR;
148 	return PERF_RECORD_MISC_KERNEL;
149 }
150 
151 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
152 {
153 	unsigned long mmcra = regs->dsisr;
154 	unsigned long use_siar = regs->result;
155 
156 	if (!use_siar)
157 		return perf_flags_from_msr(regs);
158 
159 	/*
160 	 * If we don't have flags in MMCRA, rather than using
161 	 * the MSR, we intuit the flags from the address in
162 	 * SIAR which should give slightly more reliable
163 	 * results
164 	 */
165 	if (ppmu->flags & PPMU_NO_SIPR) {
166 		unsigned long siar = mfspr(SPRN_SIAR);
167 		if (siar >= PAGE_OFFSET)
168 			return PERF_RECORD_MISC_KERNEL;
169 		return PERF_RECORD_MISC_USER;
170 	}
171 
172 	/* PR has priority over HV, so order below is important */
173 	if (mmcra_sipr(mmcra))
174 		return PERF_RECORD_MISC_USER;
175 	if (mmcra_sihv(mmcra) && (freeze_events_kernel != MMCR0_FCHV))
176 		return PERF_RECORD_MISC_HYPERVISOR;
177 	return PERF_RECORD_MISC_KERNEL;
178 }
179 
180 /*
181  * Overload regs->dsisr to store MMCRA so we only need to read it once
182  * on each interrupt.
183  * Overload regs->result to specify whether we should use the MSR (result
184  * is zero) or the SIAR (result is non zero).
185  */
186 static inline void perf_read_regs(struct pt_regs *regs)
187 {
188 	unsigned long mmcra = mfspr(SPRN_MMCRA);
189 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
190 	int use_siar;
191 
192 	/*
193 	 * If this isn't a PMU exception (eg a software event) the SIAR is
194 	 * not valid. Use pt_regs.
195 	 *
196 	 * If it is a marked event use the SIAR.
197 	 *
198 	 * If the PMU doesn't update the SIAR for non marked events use
199 	 * pt_regs.
200 	 *
201 	 * If the PMU has HV/PR flags then check to see if they
202 	 * place the exception in userspace. If so, use pt_regs. In
203 	 * continuous sampling mode the SIAR and the PMU exception are
204 	 * not synchronised, so they may be many instructions apart.
205 	 * This can result in confusing backtraces. We still want
206 	 * hypervisor samples as well as samples in the kernel with
207 	 * interrupts off hence the userspace check.
208 	 */
209 	if (TRAP(regs) != 0xf00)
210 		use_siar = 0;
211 	else if (marked)
212 		use_siar = 1;
213 	else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
214 		use_siar = 0;
215 	else if (!(ppmu->flags & PPMU_NO_SIPR) && mmcra_sipr(mmcra))
216 		use_siar = 0;
217 	else
218 		use_siar = 1;
219 
220 	regs->dsisr = mmcra;
221 	regs->result = use_siar;
222 }
223 
224 /*
225  * If interrupts were soft-disabled when a PMU interrupt occurs, treat
226  * it as an NMI.
227  */
228 static inline int perf_intr_is_nmi(struct pt_regs *regs)
229 {
230 	return !regs->softe;
231 }
232 
233 #endif /* CONFIG_PPC64 */
234 
235 static void perf_event_interrupt(struct pt_regs *regs);
236 
237 void perf_event_print_debug(void)
238 {
239 }
240 
241 /*
242  * Read one performance monitor counter (PMC).
243  */
244 static unsigned long read_pmc(int idx)
245 {
246 	unsigned long val;
247 
248 	switch (idx) {
249 	case 1:
250 		val = mfspr(SPRN_PMC1);
251 		break;
252 	case 2:
253 		val = mfspr(SPRN_PMC2);
254 		break;
255 	case 3:
256 		val = mfspr(SPRN_PMC3);
257 		break;
258 	case 4:
259 		val = mfspr(SPRN_PMC4);
260 		break;
261 	case 5:
262 		val = mfspr(SPRN_PMC5);
263 		break;
264 	case 6:
265 		val = mfspr(SPRN_PMC6);
266 		break;
267 #ifdef CONFIG_PPC64
268 	case 7:
269 		val = mfspr(SPRN_PMC7);
270 		break;
271 	case 8:
272 		val = mfspr(SPRN_PMC8);
273 		break;
274 #endif /* CONFIG_PPC64 */
275 	default:
276 		printk(KERN_ERR "oops trying to read PMC%d\n", idx);
277 		val = 0;
278 	}
279 	return val;
280 }
281 
282 /*
283  * Write one PMC.
284  */
285 static void write_pmc(int idx, unsigned long val)
286 {
287 	switch (idx) {
288 	case 1:
289 		mtspr(SPRN_PMC1, val);
290 		break;
291 	case 2:
292 		mtspr(SPRN_PMC2, val);
293 		break;
294 	case 3:
295 		mtspr(SPRN_PMC3, val);
296 		break;
297 	case 4:
298 		mtspr(SPRN_PMC4, val);
299 		break;
300 	case 5:
301 		mtspr(SPRN_PMC5, val);
302 		break;
303 	case 6:
304 		mtspr(SPRN_PMC6, val);
305 		break;
306 #ifdef CONFIG_PPC64
307 	case 7:
308 		mtspr(SPRN_PMC7, val);
309 		break;
310 	case 8:
311 		mtspr(SPRN_PMC8, val);
312 		break;
313 #endif /* CONFIG_PPC64 */
314 	default:
315 		printk(KERN_ERR "oops trying to write PMC%d\n", idx);
316 	}
317 }
318 
319 /*
320  * Check if a set of events can all go on the PMU at once.
321  * If they can't, this will look at alternative codes for the events
322  * and see if any combination of alternative codes is feasible.
323  * The feasible set is returned in event_id[].
324  */
325 static int power_check_constraints(struct cpu_hw_events *cpuhw,
326 				   u64 event_id[], unsigned int cflags[],
327 				   int n_ev)
328 {
329 	unsigned long mask, value, nv;
330 	unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
331 	int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
332 	int i, j;
333 	unsigned long addf = ppmu->add_fields;
334 	unsigned long tadd = ppmu->test_adder;
335 
336 	if (n_ev > ppmu->n_counter)
337 		return -1;
338 
339 	/* First see if the events will go on as-is */
340 	for (i = 0; i < n_ev; ++i) {
341 		if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
342 		    && !ppmu->limited_pmc_event(event_id[i])) {
343 			ppmu->get_alternatives(event_id[i], cflags[i],
344 					       cpuhw->alternatives[i]);
345 			event_id[i] = cpuhw->alternatives[i][0];
346 		}
347 		if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
348 					 &cpuhw->avalues[i][0]))
349 			return -1;
350 	}
351 	value = mask = 0;
352 	for (i = 0; i < n_ev; ++i) {
353 		nv = (value | cpuhw->avalues[i][0]) +
354 			(value & cpuhw->avalues[i][0] & addf);
355 		if ((((nv + tadd) ^ value) & mask) != 0 ||
356 		    (((nv + tadd) ^ cpuhw->avalues[i][0]) &
357 		     cpuhw->amasks[i][0]) != 0)
358 			break;
359 		value = nv;
360 		mask |= cpuhw->amasks[i][0];
361 	}
362 	if (i == n_ev)
363 		return 0;	/* all OK */
364 
365 	/* doesn't work, gather alternatives... */
366 	if (!ppmu->get_alternatives)
367 		return -1;
368 	for (i = 0; i < n_ev; ++i) {
369 		choice[i] = 0;
370 		n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
371 						  cpuhw->alternatives[i]);
372 		for (j = 1; j < n_alt[i]; ++j)
373 			ppmu->get_constraint(cpuhw->alternatives[i][j],
374 					     &cpuhw->amasks[i][j],
375 					     &cpuhw->avalues[i][j]);
376 	}
377 
378 	/* enumerate all possibilities and see if any will work */
379 	i = 0;
380 	j = -1;
381 	value = mask = nv = 0;
382 	while (i < n_ev) {
383 		if (j >= 0) {
384 			/* we're backtracking, restore context */
385 			value = svalues[i];
386 			mask = smasks[i];
387 			j = choice[i];
388 		}
389 		/*
390 		 * See if any alternative k for event_id i,
391 		 * where k > j, will satisfy the constraints.
392 		 */
393 		while (++j < n_alt[i]) {
394 			nv = (value | cpuhw->avalues[i][j]) +
395 				(value & cpuhw->avalues[i][j] & addf);
396 			if ((((nv + tadd) ^ value) & mask) == 0 &&
397 			    (((nv + tadd) ^ cpuhw->avalues[i][j])
398 			     & cpuhw->amasks[i][j]) == 0)
399 				break;
400 		}
401 		if (j >= n_alt[i]) {
402 			/*
403 			 * No feasible alternative, backtrack
404 			 * to event_id i-1 and continue enumerating its
405 			 * alternatives from where we got up to.
406 			 */
407 			if (--i < 0)
408 				return -1;
409 		} else {
410 			/*
411 			 * Found a feasible alternative for event_id i,
412 			 * remember where we got up to with this event_id,
413 			 * go on to the next event_id, and start with
414 			 * the first alternative for it.
415 			 */
416 			choice[i] = j;
417 			svalues[i] = value;
418 			smasks[i] = mask;
419 			value = nv;
420 			mask |= cpuhw->amasks[i][j];
421 			++i;
422 			j = -1;
423 		}
424 	}
425 
426 	/* OK, we have a feasible combination, tell the caller the solution */
427 	for (i = 0; i < n_ev; ++i)
428 		event_id[i] = cpuhw->alternatives[i][choice[i]];
429 	return 0;
430 }
431 
432 /*
433  * Check if newly-added events have consistent settings for
434  * exclude_{user,kernel,hv} with each other and any previously
435  * added events.
436  */
437 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
438 			  int n_prev, int n_new)
439 {
440 	int eu = 0, ek = 0, eh = 0;
441 	int i, n, first;
442 	struct perf_event *event;
443 
444 	n = n_prev + n_new;
445 	if (n <= 1)
446 		return 0;
447 
448 	first = 1;
449 	for (i = 0; i < n; ++i) {
450 		if (cflags[i] & PPMU_LIMITED_PMC_OK) {
451 			cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
452 			continue;
453 		}
454 		event = ctrs[i];
455 		if (first) {
456 			eu = event->attr.exclude_user;
457 			ek = event->attr.exclude_kernel;
458 			eh = event->attr.exclude_hv;
459 			first = 0;
460 		} else if (event->attr.exclude_user != eu ||
461 			   event->attr.exclude_kernel != ek ||
462 			   event->attr.exclude_hv != eh) {
463 			return -EAGAIN;
464 		}
465 	}
466 
467 	if (eu || ek || eh)
468 		for (i = 0; i < n; ++i)
469 			if (cflags[i] & PPMU_LIMITED_PMC_OK)
470 				cflags[i] |= PPMU_LIMITED_PMC_REQD;
471 
472 	return 0;
473 }
474 
475 static u64 check_and_compute_delta(u64 prev, u64 val)
476 {
477 	u64 delta = (val - prev) & 0xfffffffful;
478 
479 	/*
480 	 * POWER7 can roll back counter values, if the new value is smaller
481 	 * than the previous value it will cause the delta and the counter to
482 	 * have bogus values unless we rolled a counter over.  If a coutner is
483 	 * rolled back, it will be smaller, but within 256, which is the maximum
484 	 * number of events to rollback at once.  If we dectect a rollback
485 	 * return 0.  This can lead to a small lack of precision in the
486 	 * counters.
487 	 */
488 	if (prev > val && (prev - val) < 256)
489 		delta = 0;
490 
491 	return delta;
492 }
493 
494 static void power_pmu_read(struct perf_event *event)
495 {
496 	s64 val, delta, prev;
497 
498 	if (event->hw.state & PERF_HES_STOPPED)
499 		return;
500 
501 	if (!event->hw.idx)
502 		return;
503 	/*
504 	 * Performance monitor interrupts come even when interrupts
505 	 * are soft-disabled, as long as interrupts are hard-enabled.
506 	 * Therefore we treat them like NMIs.
507 	 */
508 	do {
509 		prev = local64_read(&event->hw.prev_count);
510 		barrier();
511 		val = read_pmc(event->hw.idx);
512 		delta = check_and_compute_delta(prev, val);
513 		if (!delta)
514 			return;
515 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
516 
517 	local64_add(delta, &event->count);
518 	local64_sub(delta, &event->hw.period_left);
519 }
520 
521 /*
522  * On some machines, PMC5 and PMC6 can't be written, don't respect
523  * the freeze conditions, and don't generate interrupts.  This tells
524  * us if `event' is using such a PMC.
525  */
526 static int is_limited_pmc(int pmcnum)
527 {
528 	return (ppmu->flags & PPMU_LIMITED_PMC5_6)
529 		&& (pmcnum == 5 || pmcnum == 6);
530 }
531 
532 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
533 				    unsigned long pmc5, unsigned long pmc6)
534 {
535 	struct perf_event *event;
536 	u64 val, prev, delta;
537 	int i;
538 
539 	for (i = 0; i < cpuhw->n_limited; ++i) {
540 		event = cpuhw->limited_counter[i];
541 		if (!event->hw.idx)
542 			continue;
543 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
544 		prev = local64_read(&event->hw.prev_count);
545 		event->hw.idx = 0;
546 		delta = check_and_compute_delta(prev, val);
547 		if (delta)
548 			local64_add(delta, &event->count);
549 	}
550 }
551 
552 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
553 				  unsigned long pmc5, unsigned long pmc6)
554 {
555 	struct perf_event *event;
556 	u64 val, prev;
557 	int i;
558 
559 	for (i = 0; i < cpuhw->n_limited; ++i) {
560 		event = cpuhw->limited_counter[i];
561 		event->hw.idx = cpuhw->limited_hwidx[i];
562 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
563 		prev = local64_read(&event->hw.prev_count);
564 		if (check_and_compute_delta(prev, val))
565 			local64_set(&event->hw.prev_count, val);
566 		perf_event_update_userpage(event);
567 	}
568 }
569 
570 /*
571  * Since limited events don't respect the freeze conditions, we
572  * have to read them immediately after freezing or unfreezing the
573  * other events.  We try to keep the values from the limited
574  * events as consistent as possible by keeping the delay (in
575  * cycles and instructions) between freezing/unfreezing and reading
576  * the limited events as small and consistent as possible.
577  * Therefore, if any limited events are in use, we read them
578  * both, and always in the same order, to minimize variability,
579  * and do it inside the same asm that writes MMCR0.
580  */
581 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
582 {
583 	unsigned long pmc5, pmc6;
584 
585 	if (!cpuhw->n_limited) {
586 		mtspr(SPRN_MMCR0, mmcr0);
587 		return;
588 	}
589 
590 	/*
591 	 * Write MMCR0, then read PMC5 and PMC6 immediately.
592 	 * To ensure we don't get a performance monitor interrupt
593 	 * between writing MMCR0 and freezing/thawing the limited
594 	 * events, we first write MMCR0 with the event overflow
595 	 * interrupt enable bits turned off.
596 	 */
597 	asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
598 		     : "=&r" (pmc5), "=&r" (pmc6)
599 		     : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
600 		       "i" (SPRN_MMCR0),
601 		       "i" (SPRN_PMC5), "i" (SPRN_PMC6));
602 
603 	if (mmcr0 & MMCR0_FC)
604 		freeze_limited_counters(cpuhw, pmc5, pmc6);
605 	else
606 		thaw_limited_counters(cpuhw, pmc5, pmc6);
607 
608 	/*
609 	 * Write the full MMCR0 including the event overflow interrupt
610 	 * enable bits, if necessary.
611 	 */
612 	if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
613 		mtspr(SPRN_MMCR0, mmcr0);
614 }
615 
616 /*
617  * Disable all events to prevent PMU interrupts and to allow
618  * events to be added or removed.
619  */
620 static void power_pmu_disable(struct pmu *pmu)
621 {
622 	struct cpu_hw_events *cpuhw;
623 	unsigned long flags;
624 
625 	if (!ppmu)
626 		return;
627 	local_irq_save(flags);
628 	cpuhw = &__get_cpu_var(cpu_hw_events);
629 
630 	if (!cpuhw->disabled) {
631 		cpuhw->disabled = 1;
632 		cpuhw->n_added = 0;
633 
634 		/*
635 		 * Check if we ever enabled the PMU on this cpu.
636 		 */
637 		if (!cpuhw->pmcs_enabled) {
638 			ppc_enable_pmcs();
639 			cpuhw->pmcs_enabled = 1;
640 		}
641 
642 		/*
643 		 * Disable instruction sampling if it was enabled
644 		 */
645 		if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
646 			mtspr(SPRN_MMCRA,
647 			      cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
648 			mb();
649 		}
650 
651 		/*
652 		 * Set the 'freeze counters' bit.
653 		 * The barrier is to make sure the mtspr has been
654 		 * executed and the PMU has frozen the events
655 		 * before we return.
656 		 */
657 		write_mmcr0(cpuhw, mfspr(SPRN_MMCR0) | MMCR0_FC);
658 		mb();
659 	}
660 	local_irq_restore(flags);
661 }
662 
663 /*
664  * Re-enable all events if disable == 0.
665  * If we were previously disabled and events were added, then
666  * put the new config on the PMU.
667  */
668 static void power_pmu_enable(struct pmu *pmu)
669 {
670 	struct perf_event *event;
671 	struct cpu_hw_events *cpuhw;
672 	unsigned long flags;
673 	long i;
674 	unsigned long val;
675 	s64 left;
676 	unsigned int hwc_index[MAX_HWEVENTS];
677 	int n_lim;
678 	int idx;
679 
680 	if (!ppmu)
681 		return;
682 	local_irq_save(flags);
683 	cpuhw = &__get_cpu_var(cpu_hw_events);
684 	if (!cpuhw->disabled) {
685 		local_irq_restore(flags);
686 		return;
687 	}
688 	cpuhw->disabled = 0;
689 
690 	/*
691 	 * If we didn't change anything, or only removed events,
692 	 * no need to recalculate MMCR* settings and reset the PMCs.
693 	 * Just reenable the PMU with the current MMCR* settings
694 	 * (possibly updated for removal of events).
695 	 */
696 	if (!cpuhw->n_added) {
697 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
698 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
699 		if (cpuhw->n_events == 0)
700 			ppc_set_pmu_inuse(0);
701 		goto out_enable;
702 	}
703 
704 	/*
705 	 * Compute MMCR* values for the new set of events
706 	 */
707 	if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
708 			       cpuhw->mmcr)) {
709 		/* shouldn't ever get here */
710 		printk(KERN_ERR "oops compute_mmcr failed\n");
711 		goto out;
712 	}
713 
714 	/*
715 	 * Add in MMCR0 freeze bits corresponding to the
716 	 * attr.exclude_* bits for the first event.
717 	 * We have already checked that all events have the
718 	 * same values for these bits as the first event.
719 	 */
720 	event = cpuhw->event[0];
721 	if (event->attr.exclude_user)
722 		cpuhw->mmcr[0] |= MMCR0_FCP;
723 	if (event->attr.exclude_kernel)
724 		cpuhw->mmcr[0] |= freeze_events_kernel;
725 	if (event->attr.exclude_hv)
726 		cpuhw->mmcr[0] |= MMCR0_FCHV;
727 
728 	/*
729 	 * Write the new configuration to MMCR* with the freeze
730 	 * bit set and set the hardware events to their initial values.
731 	 * Then unfreeze the events.
732 	 */
733 	ppc_set_pmu_inuse(1);
734 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
735 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
736 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
737 				| MMCR0_FC);
738 
739 	/*
740 	 * Read off any pre-existing events that need to move
741 	 * to another PMC.
742 	 */
743 	for (i = 0; i < cpuhw->n_events; ++i) {
744 		event = cpuhw->event[i];
745 		if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
746 			power_pmu_read(event);
747 			write_pmc(event->hw.idx, 0);
748 			event->hw.idx = 0;
749 		}
750 	}
751 
752 	/*
753 	 * Initialize the PMCs for all the new and moved events.
754 	 */
755 	cpuhw->n_limited = n_lim = 0;
756 	for (i = 0; i < cpuhw->n_events; ++i) {
757 		event = cpuhw->event[i];
758 		if (event->hw.idx)
759 			continue;
760 		idx = hwc_index[i] + 1;
761 		if (is_limited_pmc(idx)) {
762 			cpuhw->limited_counter[n_lim] = event;
763 			cpuhw->limited_hwidx[n_lim] = idx;
764 			++n_lim;
765 			continue;
766 		}
767 		val = 0;
768 		if (event->hw.sample_period) {
769 			left = local64_read(&event->hw.period_left);
770 			if (left < 0x80000000L)
771 				val = 0x80000000L - left;
772 		}
773 		local64_set(&event->hw.prev_count, val);
774 		event->hw.idx = idx;
775 		if (event->hw.state & PERF_HES_STOPPED)
776 			val = 0;
777 		write_pmc(idx, val);
778 		perf_event_update_userpage(event);
779 	}
780 	cpuhw->n_limited = n_lim;
781 	cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
782 
783  out_enable:
784 	mb();
785 	write_mmcr0(cpuhw, cpuhw->mmcr[0]);
786 
787 	/*
788 	 * Enable instruction sampling if necessary
789 	 */
790 	if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
791 		mb();
792 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
793 	}
794 
795  out:
796 	local_irq_restore(flags);
797 }
798 
799 static int collect_events(struct perf_event *group, int max_count,
800 			  struct perf_event *ctrs[], u64 *events,
801 			  unsigned int *flags)
802 {
803 	int n = 0;
804 	struct perf_event *event;
805 
806 	if (!is_software_event(group)) {
807 		if (n >= max_count)
808 			return -1;
809 		ctrs[n] = group;
810 		flags[n] = group->hw.event_base;
811 		events[n++] = group->hw.config;
812 	}
813 	list_for_each_entry(event, &group->sibling_list, group_entry) {
814 		if (!is_software_event(event) &&
815 		    event->state != PERF_EVENT_STATE_OFF) {
816 			if (n >= max_count)
817 				return -1;
818 			ctrs[n] = event;
819 			flags[n] = event->hw.event_base;
820 			events[n++] = event->hw.config;
821 		}
822 	}
823 	return n;
824 }
825 
826 /*
827  * Add a event to the PMU.
828  * If all events are not already frozen, then we disable and
829  * re-enable the PMU in order to get hw_perf_enable to do the
830  * actual work of reconfiguring the PMU.
831  */
832 static int power_pmu_add(struct perf_event *event, int ef_flags)
833 {
834 	struct cpu_hw_events *cpuhw;
835 	unsigned long flags;
836 	int n0;
837 	int ret = -EAGAIN;
838 
839 	local_irq_save(flags);
840 	perf_pmu_disable(event->pmu);
841 
842 	/*
843 	 * Add the event to the list (if there is room)
844 	 * and check whether the total set is still feasible.
845 	 */
846 	cpuhw = &__get_cpu_var(cpu_hw_events);
847 	n0 = cpuhw->n_events;
848 	if (n0 >= ppmu->n_counter)
849 		goto out;
850 	cpuhw->event[n0] = event;
851 	cpuhw->events[n0] = event->hw.config;
852 	cpuhw->flags[n0] = event->hw.event_base;
853 
854 	if (!(ef_flags & PERF_EF_START))
855 		event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
856 
857 	/*
858 	 * If group events scheduling transaction was started,
859 	 * skip the schedulability test here, it will be performed
860 	 * at commit time(->commit_txn) as a whole
861 	 */
862 	if (cpuhw->group_flag & PERF_EVENT_TXN)
863 		goto nocheck;
864 
865 	if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
866 		goto out;
867 	if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
868 		goto out;
869 	event->hw.config = cpuhw->events[n0];
870 
871 nocheck:
872 	++cpuhw->n_events;
873 	++cpuhw->n_added;
874 
875 	ret = 0;
876  out:
877 	perf_pmu_enable(event->pmu);
878 	local_irq_restore(flags);
879 	return ret;
880 }
881 
882 /*
883  * Remove a event from the PMU.
884  */
885 static void power_pmu_del(struct perf_event *event, int ef_flags)
886 {
887 	struct cpu_hw_events *cpuhw;
888 	long i;
889 	unsigned long flags;
890 
891 	local_irq_save(flags);
892 	perf_pmu_disable(event->pmu);
893 
894 	power_pmu_read(event);
895 
896 	cpuhw = &__get_cpu_var(cpu_hw_events);
897 	for (i = 0; i < cpuhw->n_events; ++i) {
898 		if (event == cpuhw->event[i]) {
899 			while (++i < cpuhw->n_events) {
900 				cpuhw->event[i-1] = cpuhw->event[i];
901 				cpuhw->events[i-1] = cpuhw->events[i];
902 				cpuhw->flags[i-1] = cpuhw->flags[i];
903 			}
904 			--cpuhw->n_events;
905 			ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
906 			if (event->hw.idx) {
907 				write_pmc(event->hw.idx, 0);
908 				event->hw.idx = 0;
909 			}
910 			perf_event_update_userpage(event);
911 			break;
912 		}
913 	}
914 	for (i = 0; i < cpuhw->n_limited; ++i)
915 		if (event == cpuhw->limited_counter[i])
916 			break;
917 	if (i < cpuhw->n_limited) {
918 		while (++i < cpuhw->n_limited) {
919 			cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
920 			cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
921 		}
922 		--cpuhw->n_limited;
923 	}
924 	if (cpuhw->n_events == 0) {
925 		/* disable exceptions if no events are running */
926 		cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
927 	}
928 
929 	perf_pmu_enable(event->pmu);
930 	local_irq_restore(flags);
931 }
932 
933 /*
934  * POWER-PMU does not support disabling individual counters, hence
935  * program their cycle counter to their max value and ignore the interrupts.
936  */
937 
938 static void power_pmu_start(struct perf_event *event, int ef_flags)
939 {
940 	unsigned long flags;
941 	s64 left;
942 	unsigned long val;
943 
944 	if (!event->hw.idx || !event->hw.sample_period)
945 		return;
946 
947 	if (!(event->hw.state & PERF_HES_STOPPED))
948 		return;
949 
950 	if (ef_flags & PERF_EF_RELOAD)
951 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
952 
953 	local_irq_save(flags);
954 	perf_pmu_disable(event->pmu);
955 
956 	event->hw.state = 0;
957 	left = local64_read(&event->hw.period_left);
958 
959 	val = 0;
960 	if (left < 0x80000000L)
961 		val = 0x80000000L - left;
962 
963 	write_pmc(event->hw.idx, val);
964 
965 	perf_event_update_userpage(event);
966 	perf_pmu_enable(event->pmu);
967 	local_irq_restore(flags);
968 }
969 
970 static void power_pmu_stop(struct perf_event *event, int ef_flags)
971 {
972 	unsigned long flags;
973 
974 	if (!event->hw.idx || !event->hw.sample_period)
975 		return;
976 
977 	if (event->hw.state & PERF_HES_STOPPED)
978 		return;
979 
980 	local_irq_save(flags);
981 	perf_pmu_disable(event->pmu);
982 
983 	power_pmu_read(event);
984 	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
985 	write_pmc(event->hw.idx, 0);
986 
987 	perf_event_update_userpage(event);
988 	perf_pmu_enable(event->pmu);
989 	local_irq_restore(flags);
990 }
991 
992 /*
993  * Start group events scheduling transaction
994  * Set the flag to make pmu::enable() not perform the
995  * schedulability test, it will be performed at commit time
996  */
997 void power_pmu_start_txn(struct pmu *pmu)
998 {
999 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1000 
1001 	perf_pmu_disable(pmu);
1002 	cpuhw->group_flag |= PERF_EVENT_TXN;
1003 	cpuhw->n_txn_start = cpuhw->n_events;
1004 }
1005 
1006 /*
1007  * Stop group events scheduling transaction
1008  * Clear the flag and pmu::enable() will perform the
1009  * schedulability test.
1010  */
1011 void power_pmu_cancel_txn(struct pmu *pmu)
1012 {
1013 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1014 
1015 	cpuhw->group_flag &= ~PERF_EVENT_TXN;
1016 	perf_pmu_enable(pmu);
1017 }
1018 
1019 /*
1020  * Commit group events scheduling transaction
1021  * Perform the group schedulability test as a whole
1022  * Return 0 if success
1023  */
1024 int power_pmu_commit_txn(struct pmu *pmu)
1025 {
1026 	struct cpu_hw_events *cpuhw;
1027 	long i, n;
1028 
1029 	if (!ppmu)
1030 		return -EAGAIN;
1031 	cpuhw = &__get_cpu_var(cpu_hw_events);
1032 	n = cpuhw->n_events;
1033 	if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1034 		return -EAGAIN;
1035 	i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1036 	if (i < 0)
1037 		return -EAGAIN;
1038 
1039 	for (i = cpuhw->n_txn_start; i < n; ++i)
1040 		cpuhw->event[i]->hw.config = cpuhw->events[i];
1041 
1042 	cpuhw->group_flag &= ~PERF_EVENT_TXN;
1043 	perf_pmu_enable(pmu);
1044 	return 0;
1045 }
1046 
1047 /*
1048  * Return 1 if we might be able to put event on a limited PMC,
1049  * or 0 if not.
1050  * A event can only go on a limited PMC if it counts something
1051  * that a limited PMC can count, doesn't require interrupts, and
1052  * doesn't exclude any processor mode.
1053  */
1054 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1055 				 unsigned int flags)
1056 {
1057 	int n;
1058 	u64 alt[MAX_EVENT_ALTERNATIVES];
1059 
1060 	if (event->attr.exclude_user
1061 	    || event->attr.exclude_kernel
1062 	    || event->attr.exclude_hv
1063 	    || event->attr.sample_period)
1064 		return 0;
1065 
1066 	if (ppmu->limited_pmc_event(ev))
1067 		return 1;
1068 
1069 	/*
1070 	 * The requested event_id isn't on a limited PMC already;
1071 	 * see if any alternative code goes on a limited PMC.
1072 	 */
1073 	if (!ppmu->get_alternatives)
1074 		return 0;
1075 
1076 	flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1077 	n = ppmu->get_alternatives(ev, flags, alt);
1078 
1079 	return n > 0;
1080 }
1081 
1082 /*
1083  * Find an alternative event_id that goes on a normal PMC, if possible,
1084  * and return the event_id code, or 0 if there is no such alternative.
1085  * (Note: event_id code 0 is "don't count" on all machines.)
1086  */
1087 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1088 {
1089 	u64 alt[MAX_EVENT_ALTERNATIVES];
1090 	int n;
1091 
1092 	flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1093 	n = ppmu->get_alternatives(ev, flags, alt);
1094 	if (!n)
1095 		return 0;
1096 	return alt[0];
1097 }
1098 
1099 /* Number of perf_events counting hardware events */
1100 static atomic_t num_events;
1101 /* Used to avoid races in calling reserve/release_pmc_hardware */
1102 static DEFINE_MUTEX(pmc_reserve_mutex);
1103 
1104 /*
1105  * Release the PMU if this is the last perf_event.
1106  */
1107 static void hw_perf_event_destroy(struct perf_event *event)
1108 {
1109 	if (!atomic_add_unless(&num_events, -1, 1)) {
1110 		mutex_lock(&pmc_reserve_mutex);
1111 		if (atomic_dec_return(&num_events) == 0)
1112 			release_pmc_hardware();
1113 		mutex_unlock(&pmc_reserve_mutex);
1114 	}
1115 }
1116 
1117 /*
1118  * Translate a generic cache event_id config to a raw event_id code.
1119  */
1120 static int hw_perf_cache_event(u64 config, u64 *eventp)
1121 {
1122 	unsigned long type, op, result;
1123 	int ev;
1124 
1125 	if (!ppmu->cache_events)
1126 		return -EINVAL;
1127 
1128 	/* unpack config */
1129 	type = config & 0xff;
1130 	op = (config >> 8) & 0xff;
1131 	result = (config >> 16) & 0xff;
1132 
1133 	if (type >= PERF_COUNT_HW_CACHE_MAX ||
1134 	    op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1135 	    result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1136 		return -EINVAL;
1137 
1138 	ev = (*ppmu->cache_events)[type][op][result];
1139 	if (ev == 0)
1140 		return -EOPNOTSUPP;
1141 	if (ev == -1)
1142 		return -EINVAL;
1143 	*eventp = ev;
1144 	return 0;
1145 }
1146 
1147 static int power_pmu_event_init(struct perf_event *event)
1148 {
1149 	u64 ev;
1150 	unsigned long flags;
1151 	struct perf_event *ctrs[MAX_HWEVENTS];
1152 	u64 events[MAX_HWEVENTS];
1153 	unsigned int cflags[MAX_HWEVENTS];
1154 	int n;
1155 	int err;
1156 	struct cpu_hw_events *cpuhw;
1157 
1158 	if (!ppmu)
1159 		return -ENOENT;
1160 
1161 	/* does not support taken branch sampling */
1162 	if (has_branch_stack(event))
1163 		return -EOPNOTSUPP;
1164 
1165 	switch (event->attr.type) {
1166 	case PERF_TYPE_HARDWARE:
1167 		ev = event->attr.config;
1168 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1169 			return -EOPNOTSUPP;
1170 		ev = ppmu->generic_events[ev];
1171 		break;
1172 	case PERF_TYPE_HW_CACHE:
1173 		err = hw_perf_cache_event(event->attr.config, &ev);
1174 		if (err)
1175 			return err;
1176 		break;
1177 	case PERF_TYPE_RAW:
1178 		ev = event->attr.config;
1179 		break;
1180 	default:
1181 		return -ENOENT;
1182 	}
1183 
1184 	event->hw.config_base = ev;
1185 	event->hw.idx = 0;
1186 
1187 	/*
1188 	 * If we are not running on a hypervisor, force the
1189 	 * exclude_hv bit to 0 so that we don't care what
1190 	 * the user set it to.
1191 	 */
1192 	if (!firmware_has_feature(FW_FEATURE_LPAR))
1193 		event->attr.exclude_hv = 0;
1194 
1195 	/*
1196 	 * If this is a per-task event, then we can use
1197 	 * PM_RUN_* events interchangeably with their non RUN_*
1198 	 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1199 	 * XXX we should check if the task is an idle task.
1200 	 */
1201 	flags = 0;
1202 	if (event->attach_state & PERF_ATTACH_TASK)
1203 		flags |= PPMU_ONLY_COUNT_RUN;
1204 
1205 	/*
1206 	 * If this machine has limited events, check whether this
1207 	 * event_id could go on a limited event.
1208 	 */
1209 	if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1210 		if (can_go_on_limited_pmc(event, ev, flags)) {
1211 			flags |= PPMU_LIMITED_PMC_OK;
1212 		} else if (ppmu->limited_pmc_event(ev)) {
1213 			/*
1214 			 * The requested event_id is on a limited PMC,
1215 			 * but we can't use a limited PMC; see if any
1216 			 * alternative goes on a normal PMC.
1217 			 */
1218 			ev = normal_pmc_alternative(ev, flags);
1219 			if (!ev)
1220 				return -EINVAL;
1221 		}
1222 	}
1223 
1224 	/*
1225 	 * If this is in a group, check if it can go on with all the
1226 	 * other hardware events in the group.  We assume the event
1227 	 * hasn't been linked into its leader's sibling list at this point.
1228 	 */
1229 	n = 0;
1230 	if (event->group_leader != event) {
1231 		n = collect_events(event->group_leader, ppmu->n_counter - 1,
1232 				   ctrs, events, cflags);
1233 		if (n < 0)
1234 			return -EINVAL;
1235 	}
1236 	events[n] = ev;
1237 	ctrs[n] = event;
1238 	cflags[n] = flags;
1239 	if (check_excludes(ctrs, cflags, n, 1))
1240 		return -EINVAL;
1241 
1242 	cpuhw = &get_cpu_var(cpu_hw_events);
1243 	err = power_check_constraints(cpuhw, events, cflags, n + 1);
1244 	put_cpu_var(cpu_hw_events);
1245 	if (err)
1246 		return -EINVAL;
1247 
1248 	event->hw.config = events[n];
1249 	event->hw.event_base = cflags[n];
1250 	event->hw.last_period = event->hw.sample_period;
1251 	local64_set(&event->hw.period_left, event->hw.last_period);
1252 
1253 	/*
1254 	 * See if we need to reserve the PMU.
1255 	 * If no events are currently in use, then we have to take a
1256 	 * mutex to ensure that we don't race with another task doing
1257 	 * reserve_pmc_hardware or release_pmc_hardware.
1258 	 */
1259 	err = 0;
1260 	if (!atomic_inc_not_zero(&num_events)) {
1261 		mutex_lock(&pmc_reserve_mutex);
1262 		if (atomic_read(&num_events) == 0 &&
1263 		    reserve_pmc_hardware(perf_event_interrupt))
1264 			err = -EBUSY;
1265 		else
1266 			atomic_inc(&num_events);
1267 		mutex_unlock(&pmc_reserve_mutex);
1268 	}
1269 	event->destroy = hw_perf_event_destroy;
1270 
1271 	return err;
1272 }
1273 
1274 static int power_pmu_event_idx(struct perf_event *event)
1275 {
1276 	return event->hw.idx;
1277 }
1278 
1279 struct pmu power_pmu = {
1280 	.pmu_enable	= power_pmu_enable,
1281 	.pmu_disable	= power_pmu_disable,
1282 	.event_init	= power_pmu_event_init,
1283 	.add		= power_pmu_add,
1284 	.del		= power_pmu_del,
1285 	.start		= power_pmu_start,
1286 	.stop		= power_pmu_stop,
1287 	.read		= power_pmu_read,
1288 	.start_txn	= power_pmu_start_txn,
1289 	.cancel_txn	= power_pmu_cancel_txn,
1290 	.commit_txn	= power_pmu_commit_txn,
1291 	.event_idx	= power_pmu_event_idx,
1292 };
1293 
1294 /*
1295  * A counter has overflowed; update its count and record
1296  * things if requested.  Note that interrupts are hard-disabled
1297  * here so there is no possibility of being interrupted.
1298  */
1299 static void record_and_restart(struct perf_event *event, unsigned long val,
1300 			       struct pt_regs *regs)
1301 {
1302 	u64 period = event->hw.sample_period;
1303 	s64 prev, delta, left;
1304 	int record = 0;
1305 
1306 	if (event->hw.state & PERF_HES_STOPPED) {
1307 		write_pmc(event->hw.idx, 0);
1308 		return;
1309 	}
1310 
1311 	/* we don't have to worry about interrupts here */
1312 	prev = local64_read(&event->hw.prev_count);
1313 	delta = check_and_compute_delta(prev, val);
1314 	local64_add(delta, &event->count);
1315 
1316 	/*
1317 	 * See if the total period for this event has expired,
1318 	 * and update for the next period.
1319 	 */
1320 	val = 0;
1321 	left = local64_read(&event->hw.period_left) - delta;
1322 	if (period) {
1323 		if (left <= 0) {
1324 			left += period;
1325 			if (left <= 0)
1326 				left = period;
1327 			record = 1;
1328 			event->hw.last_period = event->hw.sample_period;
1329 		}
1330 		if (left < 0x80000000LL)
1331 			val = 0x80000000LL - left;
1332 	}
1333 
1334 	write_pmc(event->hw.idx, val);
1335 	local64_set(&event->hw.prev_count, val);
1336 	local64_set(&event->hw.period_left, left);
1337 	perf_event_update_userpage(event);
1338 
1339 	/*
1340 	 * Finally record data if requested.
1341 	 */
1342 	if (record) {
1343 		struct perf_sample_data data;
1344 
1345 		perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
1346 
1347 		if (event->attr.sample_type & PERF_SAMPLE_ADDR)
1348 			perf_get_data_addr(regs, &data.addr);
1349 
1350 		if (perf_event_overflow(event, &data, regs))
1351 			power_pmu_stop(event, 0);
1352 	}
1353 }
1354 
1355 /*
1356  * Called from generic code to get the misc flags (i.e. processor mode)
1357  * for an event_id.
1358  */
1359 unsigned long perf_misc_flags(struct pt_regs *regs)
1360 {
1361 	u32 flags = perf_get_misc_flags(regs);
1362 
1363 	if (flags)
1364 		return flags;
1365 	return user_mode(regs) ? PERF_RECORD_MISC_USER :
1366 		PERF_RECORD_MISC_KERNEL;
1367 }
1368 
1369 /*
1370  * Called from generic code to get the instruction pointer
1371  * for an event_id.
1372  */
1373 unsigned long perf_instruction_pointer(struct pt_regs *regs)
1374 {
1375 	unsigned long use_siar = regs->result;
1376 
1377 	if (use_siar)
1378 		return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
1379 	else
1380 		return regs->nip;
1381 }
1382 
1383 static bool pmc_overflow(unsigned long val)
1384 {
1385 	if ((int)val < 0)
1386 		return true;
1387 
1388 	/*
1389 	 * Events on POWER7 can roll back if a speculative event doesn't
1390 	 * eventually complete. Unfortunately in some rare cases they will
1391 	 * raise a performance monitor exception. We need to catch this to
1392 	 * ensure we reset the PMC. In all cases the PMC will be 256 or less
1393 	 * cycles from overflow.
1394 	 *
1395 	 * We only do this if the first pass fails to find any overflowing
1396 	 * PMCs because a user might set a period of less than 256 and we
1397 	 * don't want to mistakenly reset them.
1398 	 */
1399 	if (__is_processor(PV_POWER7) && ((0x80000000 - val) <= 256))
1400 		return true;
1401 
1402 	return false;
1403 }
1404 
1405 /*
1406  * Performance monitor interrupt stuff
1407  */
1408 static void perf_event_interrupt(struct pt_regs *regs)
1409 {
1410 	int i;
1411 	struct cpu_hw_events *cpuhw = &__get_cpu_var(cpu_hw_events);
1412 	struct perf_event *event;
1413 	unsigned long val;
1414 	int found = 0;
1415 	int nmi;
1416 
1417 	if (cpuhw->n_limited)
1418 		freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
1419 					mfspr(SPRN_PMC6));
1420 
1421 	perf_read_regs(regs);
1422 
1423 	nmi = perf_intr_is_nmi(regs);
1424 	if (nmi)
1425 		nmi_enter();
1426 	else
1427 		irq_enter();
1428 
1429 	for (i = 0; i < cpuhw->n_events; ++i) {
1430 		event = cpuhw->event[i];
1431 		if (!event->hw.idx || is_limited_pmc(event->hw.idx))
1432 			continue;
1433 		val = read_pmc(event->hw.idx);
1434 		if (pmc_overflow(val)) {
1435 			/* event has overflowed */
1436 			found = 1;
1437 			record_and_restart(event, val, regs);
1438 		}
1439 	}
1440 
1441 	/*
1442 	 * In case we didn't find and reset the event that caused
1443 	 * the interrupt, scan all events and reset any that are
1444 	 * negative, to avoid getting continual interrupts.
1445 	 * Any that we processed in the previous loop will not be negative.
1446 	 */
1447 	if (!found) {
1448 		for (i = 0; i < ppmu->n_counter; ++i) {
1449 			if (is_limited_pmc(i + 1))
1450 				continue;
1451 			val = read_pmc(i + 1);
1452 			if (pmc_overflow(val))
1453 				write_pmc(i + 1, 0);
1454 		}
1455 	}
1456 
1457 	/*
1458 	 * Reset MMCR0 to its normal value.  This will set PMXE and
1459 	 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
1460 	 * and thus allow interrupts to occur again.
1461 	 * XXX might want to use MSR.PM to keep the events frozen until
1462 	 * we get back out of this interrupt.
1463 	 */
1464 	write_mmcr0(cpuhw, cpuhw->mmcr[0]);
1465 
1466 	if (nmi)
1467 		nmi_exit();
1468 	else
1469 		irq_exit();
1470 }
1471 
1472 static void power_pmu_setup(int cpu)
1473 {
1474 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
1475 
1476 	if (!ppmu)
1477 		return;
1478 	memset(cpuhw, 0, sizeof(*cpuhw));
1479 	cpuhw->mmcr[0] = MMCR0_FC;
1480 }
1481 
1482 static int __cpuinit
1483 power_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1484 {
1485 	unsigned int cpu = (long)hcpu;
1486 
1487 	switch (action & ~CPU_TASKS_FROZEN) {
1488 	case CPU_UP_PREPARE:
1489 		power_pmu_setup(cpu);
1490 		break;
1491 
1492 	default:
1493 		break;
1494 	}
1495 
1496 	return NOTIFY_OK;
1497 }
1498 
1499 int __cpuinit register_power_pmu(struct power_pmu *pmu)
1500 {
1501 	if (ppmu)
1502 		return -EBUSY;		/* something's already registered */
1503 
1504 	ppmu = pmu;
1505 	pr_info("%s performance monitor hardware support registered\n",
1506 		pmu->name);
1507 
1508 #ifdef MSR_HV
1509 	/*
1510 	 * Use FCHV to ignore kernel events if MSR.HV is set.
1511 	 */
1512 	if (mfmsr() & MSR_HV)
1513 		freeze_events_kernel = MMCR0_FCHV;
1514 #endif /* CONFIG_PPC64 */
1515 
1516 	perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
1517 	perf_cpu_notifier(power_pmu_notifier);
1518 
1519 	return 0;
1520 }
1521