xref: /openbmc/linux/arch/powerpc/perf/core-book3s.c (revision 920b0442)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Performance event support - powerpc architecture code
4  *
5  * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
6  */
7 #include <linux/kernel.h>
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <linux/perf_event.h>
11 #include <linux/percpu.h>
12 #include <linux/hardirq.h>
13 #include <linux/uaccess.h>
14 #include <asm/reg.h>
15 #include <asm/pmc.h>
16 #include <asm/machdep.h>
17 #include <asm/firmware.h>
18 #include <asm/ptrace.h>
19 #include <asm/code-patching.h>
20 #include <asm/hw_irq.h>
21 #include <asm/interrupt.h>
22 
23 #ifdef CONFIG_PPC64
24 #include "internal.h"
25 #endif
26 
27 #define BHRB_MAX_ENTRIES	32
28 #define BHRB_TARGET		0x0000000000000002
29 #define BHRB_PREDICTION		0x0000000000000001
30 #define BHRB_EA			0xFFFFFFFFFFFFFFFCUL
31 
32 struct cpu_hw_events {
33 	int n_events;
34 	int n_percpu;
35 	int disabled;
36 	int n_added;
37 	int n_limited;
38 	u8  pmcs_enabled;
39 	struct perf_event *event[MAX_HWEVENTS];
40 	u64 events[MAX_HWEVENTS];
41 	unsigned int flags[MAX_HWEVENTS];
42 	struct mmcr_regs mmcr;
43 	struct perf_event *limited_counter[MAX_LIMITED_HWCOUNTERS];
44 	u8  limited_hwidx[MAX_LIMITED_HWCOUNTERS];
45 	u64 alternatives[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
46 	unsigned long amasks[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
47 	unsigned long avalues[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
48 
49 	unsigned int txn_flags;
50 	int n_txn_start;
51 
52 	/* BHRB bits */
53 	u64				bhrb_filter;	/* BHRB HW branch filter */
54 	unsigned int			bhrb_users;
55 	void				*bhrb_context;
56 	struct	perf_branch_stack	bhrb_stack;
57 	struct	perf_branch_entry	bhrb_entries[BHRB_MAX_ENTRIES];
58 	u64				ic_init;
59 
60 	/* Store the PMC values */
61 	unsigned long pmcs[MAX_HWEVENTS];
62 };
63 
64 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
65 
66 static struct power_pmu *ppmu;
67 
68 /*
69  * Normally, to ignore kernel events we set the FCS (freeze counters
70  * in supervisor mode) bit in MMCR0, but if the kernel runs with the
71  * hypervisor bit set in the MSR, or if we are running on a processor
72  * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
73  * then we need to use the FCHV bit to ignore kernel events.
74  */
75 static unsigned int freeze_events_kernel = MMCR0_FCS;
76 
77 /*
78  * 32-bit doesn't have MMCRA but does have an MMCR2,
79  * and a few other names are different.
80  * Also 32-bit doesn't have MMCR3, SIER2 and SIER3.
81  * Define them as zero knowing that any code path accessing
82  * these registers (via mtspr/mfspr) are done under ppmu flag
83  * check for PPMU_ARCH_31 and we will not enter that code path
84  * for 32-bit.
85  */
86 #ifdef CONFIG_PPC32
87 
88 #define MMCR0_FCHV		0
89 #define MMCR0_PMCjCE		MMCR0_PMCnCE
90 #define MMCR0_FC56		0
91 #define MMCR0_PMAO		0
92 #define MMCR0_EBE		0
93 #define MMCR0_BHRBA		0
94 #define MMCR0_PMCC		0
95 #define MMCR0_PMCC_U6		0
96 
97 #define SPRN_MMCRA		SPRN_MMCR2
98 #define SPRN_MMCR3		0
99 #define SPRN_SIER2		0
100 #define SPRN_SIER3		0
101 #define MMCRA_SAMPLE_ENABLE	0
102 #define MMCRA_BHRB_DISABLE     0
103 #define MMCR0_PMCCEXT		0
104 
105 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
106 {
107 	return 0;
108 }
109 static inline void perf_get_data_addr(struct perf_event *event, struct pt_regs *regs, u64 *addrp) { }
110 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
111 {
112 	return 0;
113 }
114 static inline void perf_read_regs(struct pt_regs *regs)
115 {
116 	regs->result = 0;
117 }
118 
119 static inline int siar_valid(struct pt_regs *regs)
120 {
121 	return 1;
122 }
123 
124 static bool is_ebb_event(struct perf_event *event) { return false; }
125 static int ebb_event_check(struct perf_event *event) { return 0; }
126 static void ebb_event_add(struct perf_event *event) { }
127 static void ebb_switch_out(unsigned long mmcr0) { }
128 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
129 {
130 	return cpuhw->mmcr.mmcr0;
131 }
132 
133 static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
134 static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
135 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in) {}
136 static inline void power_pmu_bhrb_read(struct perf_event *event, struct cpu_hw_events *cpuhw) {}
137 static void pmao_restore_workaround(bool ebb) { }
138 #endif /* CONFIG_PPC32 */
139 
140 bool is_sier_available(void)
141 {
142 	if (!ppmu)
143 		return false;
144 
145 	if (ppmu->flags & PPMU_HAS_SIER)
146 		return true;
147 
148 	return false;
149 }
150 
151 /*
152  * Return PMC value corresponding to the
153  * index passed.
154  */
155 unsigned long get_pmcs_ext_regs(int idx)
156 {
157 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
158 
159 	return cpuhw->pmcs[idx];
160 }
161 
162 static bool regs_use_siar(struct pt_regs *regs)
163 {
164 	/*
165 	 * When we take a performance monitor exception the regs are setup
166 	 * using perf_read_regs() which overloads some fields, in particular
167 	 * regs->result to tell us whether to use SIAR.
168 	 *
169 	 * However if the regs are from another exception, eg. a syscall, then
170 	 * they have not been setup using perf_read_regs() and so regs->result
171 	 * is something random.
172 	 */
173 	return ((TRAP(regs) == INTERRUPT_PERFMON) && regs->result);
174 }
175 
176 /*
177  * Things that are specific to 64-bit implementations.
178  */
179 #ifdef CONFIG_PPC64
180 
181 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
182 {
183 	unsigned long mmcra = regs->dsisr;
184 
185 	if ((ppmu->flags & PPMU_HAS_SSLOT) && (mmcra & MMCRA_SAMPLE_ENABLE)) {
186 		unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
187 		if (slot > 1)
188 			return 4 * (slot - 1);
189 	}
190 
191 	return 0;
192 }
193 
194 /*
195  * The user wants a data address recorded.
196  * If we're not doing instruction sampling, give them the SDAR
197  * (sampled data address).  If we are doing instruction sampling, then
198  * only give them the SDAR if it corresponds to the instruction
199  * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC, the
200  * [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA, or the SDAR_VALID bit in SIER.
201  */
202 static inline void perf_get_data_addr(struct perf_event *event, struct pt_regs *regs, u64 *addrp)
203 {
204 	unsigned long mmcra = regs->dsisr;
205 	bool sdar_valid;
206 
207 	if (ppmu->flags & PPMU_HAS_SIER)
208 		sdar_valid = regs->dar & SIER_SDAR_VALID;
209 	else {
210 		unsigned long sdsync;
211 
212 		if (ppmu->flags & PPMU_SIAR_VALID)
213 			sdsync = POWER7P_MMCRA_SDAR_VALID;
214 		else if (ppmu->flags & PPMU_ALT_SIPR)
215 			sdsync = POWER6_MMCRA_SDSYNC;
216 		else if (ppmu->flags & PPMU_NO_SIAR)
217 			sdsync = MMCRA_SAMPLE_ENABLE;
218 		else
219 			sdsync = MMCRA_SDSYNC;
220 
221 		sdar_valid = mmcra & sdsync;
222 	}
223 
224 	if (!(mmcra & MMCRA_SAMPLE_ENABLE) || sdar_valid)
225 		*addrp = mfspr(SPRN_SDAR);
226 
227 	if (is_kernel_addr(mfspr(SPRN_SDAR)) && event->attr.exclude_kernel)
228 		*addrp = 0;
229 }
230 
231 static bool regs_sihv(struct pt_regs *regs)
232 {
233 	unsigned long sihv = MMCRA_SIHV;
234 
235 	if (ppmu->flags & PPMU_HAS_SIER)
236 		return !!(regs->dar & SIER_SIHV);
237 
238 	if (ppmu->flags & PPMU_ALT_SIPR)
239 		sihv = POWER6_MMCRA_SIHV;
240 
241 	return !!(regs->dsisr & sihv);
242 }
243 
244 static bool regs_sipr(struct pt_regs *regs)
245 {
246 	unsigned long sipr = MMCRA_SIPR;
247 
248 	if (ppmu->flags & PPMU_HAS_SIER)
249 		return !!(regs->dar & SIER_SIPR);
250 
251 	if (ppmu->flags & PPMU_ALT_SIPR)
252 		sipr = POWER6_MMCRA_SIPR;
253 
254 	return !!(regs->dsisr & sipr);
255 }
256 
257 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
258 {
259 	if (regs->msr & MSR_PR)
260 		return PERF_RECORD_MISC_USER;
261 	if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
262 		return PERF_RECORD_MISC_HYPERVISOR;
263 	return PERF_RECORD_MISC_KERNEL;
264 }
265 
266 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
267 {
268 	bool use_siar = regs_use_siar(regs);
269 	unsigned long mmcra = regs->dsisr;
270 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
271 
272 	if (!use_siar)
273 		return perf_flags_from_msr(regs);
274 
275 	/*
276 	 * Check the address in SIAR to identify the
277 	 * privilege levels since the SIER[MSR_HV, MSR_PR]
278 	 * bits are not set for marked events in power10
279 	 * DD1.
280 	 */
281 	if (marked && (ppmu->flags & PPMU_P10_DD1)) {
282 		unsigned long siar = mfspr(SPRN_SIAR);
283 		if (siar) {
284 			if (is_kernel_addr(siar))
285 				return PERF_RECORD_MISC_KERNEL;
286 			return PERF_RECORD_MISC_USER;
287 		} else {
288 			if (is_kernel_addr(regs->nip))
289 				return PERF_RECORD_MISC_KERNEL;
290 			return PERF_RECORD_MISC_USER;
291 		}
292 	}
293 
294 	/*
295 	 * If we don't have flags in MMCRA, rather than using
296 	 * the MSR, we intuit the flags from the address in
297 	 * SIAR which should give slightly more reliable
298 	 * results
299 	 */
300 	if (ppmu->flags & PPMU_NO_SIPR) {
301 		unsigned long siar = mfspr(SPRN_SIAR);
302 		if (is_kernel_addr(siar))
303 			return PERF_RECORD_MISC_KERNEL;
304 		return PERF_RECORD_MISC_USER;
305 	}
306 
307 	/* PR has priority over HV, so order below is important */
308 	if (regs_sipr(regs))
309 		return PERF_RECORD_MISC_USER;
310 
311 	if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
312 		return PERF_RECORD_MISC_HYPERVISOR;
313 
314 	return PERF_RECORD_MISC_KERNEL;
315 }
316 
317 /*
318  * Overload regs->dsisr to store MMCRA so we only need to read it once
319  * on each interrupt.
320  * Overload regs->dar to store SIER if we have it.
321  * Overload regs->result to specify whether we should use the MSR (result
322  * is zero) or the SIAR (result is non zero).
323  */
324 static inline void perf_read_regs(struct pt_regs *regs)
325 {
326 	unsigned long mmcra = mfspr(SPRN_MMCRA);
327 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
328 	int use_siar;
329 
330 	regs->dsisr = mmcra;
331 
332 	if (ppmu->flags & PPMU_HAS_SIER)
333 		regs->dar = mfspr(SPRN_SIER);
334 
335 	/*
336 	 * If this isn't a PMU exception (eg a software event) the SIAR is
337 	 * not valid. Use pt_regs.
338 	 *
339 	 * If it is a marked event use the SIAR.
340 	 *
341 	 * If the PMU doesn't update the SIAR for non marked events use
342 	 * pt_regs.
343 	 *
344 	 * If regs is a kernel interrupt, always use SIAR. Some PMUs have an
345 	 * issue with regs_sipr not being in synch with SIAR in interrupt entry
346 	 * and return sequences, which can result in regs_sipr being true for
347 	 * kernel interrupts and SIAR, which has the effect of causing samples
348 	 * to pile up at mtmsrd MSR[EE] 0->1 or pending irq replay around
349 	 * interrupt entry/exit.
350 	 *
351 	 * If the PMU has HV/PR flags then check to see if they
352 	 * place the exception in userspace. If so, use pt_regs. In
353 	 * continuous sampling mode the SIAR and the PMU exception are
354 	 * not synchronised, so they may be many instructions apart.
355 	 * This can result in confusing backtraces. We still want
356 	 * hypervisor samples as well as samples in the kernel with
357 	 * interrupts off hence the userspace check.
358 	 */
359 	if (TRAP(regs) != INTERRUPT_PERFMON)
360 		use_siar = 0;
361 	else if ((ppmu->flags & PPMU_NO_SIAR))
362 		use_siar = 0;
363 	else if (marked)
364 		use_siar = 1;
365 	else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
366 		use_siar = 0;
367 	else if (!user_mode(regs))
368 		use_siar = 1;
369 	else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
370 		use_siar = 0;
371 	else
372 		use_siar = 1;
373 
374 	regs->result = use_siar;
375 }
376 
377 /*
378  * On processors like P7+ that have the SIAR-Valid bit, marked instructions
379  * must be sampled only if the SIAR-valid bit is set.
380  *
381  * For unmarked instructions and for processors that don't have the SIAR-Valid
382  * bit, assume that SIAR is valid.
383  */
384 static inline int siar_valid(struct pt_regs *regs)
385 {
386 	unsigned long mmcra = regs->dsisr;
387 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
388 
389 	if (marked) {
390 		/*
391 		 * SIER[SIAR_VALID] is not set for some
392 		 * marked events on power10 DD1, so drop
393 		 * the check for SIER[SIAR_VALID] and return true.
394 		 */
395 		if (ppmu->flags & PPMU_P10_DD1)
396 			return 0x1;
397 		else if (ppmu->flags & PPMU_HAS_SIER)
398 			return regs->dar & SIER_SIAR_VALID;
399 
400 		if (ppmu->flags & PPMU_SIAR_VALID)
401 			return mmcra & POWER7P_MMCRA_SIAR_VALID;
402 	}
403 
404 	return 1;
405 }
406 
407 
408 /* Reset all possible BHRB entries */
409 static void power_pmu_bhrb_reset(void)
410 {
411 	asm volatile(PPC_CLRBHRB);
412 }
413 
414 static void power_pmu_bhrb_enable(struct perf_event *event)
415 {
416 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
417 
418 	if (!ppmu->bhrb_nr)
419 		return;
420 
421 	/* Clear BHRB if we changed task context to avoid data leaks */
422 	if (event->ctx->task && cpuhw->bhrb_context != event->ctx) {
423 		power_pmu_bhrb_reset();
424 		cpuhw->bhrb_context = event->ctx;
425 	}
426 	cpuhw->bhrb_users++;
427 	perf_sched_cb_inc(event->ctx->pmu);
428 }
429 
430 static void power_pmu_bhrb_disable(struct perf_event *event)
431 {
432 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
433 
434 	if (!ppmu->bhrb_nr)
435 		return;
436 
437 	WARN_ON_ONCE(!cpuhw->bhrb_users);
438 	cpuhw->bhrb_users--;
439 	perf_sched_cb_dec(event->ctx->pmu);
440 
441 	if (!cpuhw->disabled && !cpuhw->bhrb_users) {
442 		/* BHRB cannot be turned off when other
443 		 * events are active on the PMU.
444 		 */
445 
446 		/* avoid stale pointer */
447 		cpuhw->bhrb_context = NULL;
448 	}
449 }
450 
451 /* Called from ctxsw to prevent one process's branch entries to
452  * mingle with the other process's entries during context switch.
453  */
454 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
455 {
456 	if (!ppmu->bhrb_nr)
457 		return;
458 
459 	if (sched_in)
460 		power_pmu_bhrb_reset();
461 }
462 /* Calculate the to address for a branch */
463 static __u64 power_pmu_bhrb_to(u64 addr)
464 {
465 	unsigned int instr;
466 	__u64 target;
467 
468 	if (is_kernel_addr(addr)) {
469 		if (copy_from_kernel_nofault(&instr, (void *)addr,
470 				sizeof(instr)))
471 			return 0;
472 
473 		return branch_target(&instr);
474 	}
475 
476 	/* Userspace: need copy instruction here then translate it */
477 	if (copy_from_user_nofault(&instr, (unsigned int __user *)addr,
478 			sizeof(instr)))
479 		return 0;
480 
481 	target = branch_target(&instr);
482 	if ((!target) || (instr & BRANCH_ABSOLUTE))
483 		return target;
484 
485 	/* Translate relative branch target from kernel to user address */
486 	return target - (unsigned long)&instr + addr;
487 }
488 
489 /* Processing BHRB entries */
490 static void power_pmu_bhrb_read(struct perf_event *event, struct cpu_hw_events *cpuhw)
491 {
492 	u64 val;
493 	u64 addr;
494 	int r_index, u_index, pred;
495 
496 	r_index = 0;
497 	u_index = 0;
498 	while (r_index < ppmu->bhrb_nr) {
499 		/* Assembly read function */
500 		val = read_bhrb(r_index++);
501 		if (!val)
502 			/* Terminal marker: End of valid BHRB entries */
503 			break;
504 		else {
505 			addr = val & BHRB_EA;
506 			pred = val & BHRB_PREDICTION;
507 
508 			if (!addr)
509 				/* invalid entry */
510 				continue;
511 
512 			/*
513 			 * BHRB rolling buffer could very much contain the kernel
514 			 * addresses at this point. Check the privileges before
515 			 * exporting it to userspace (avoid exposure of regions
516 			 * where we could have speculative execution)
517 			 * Incase of ISA v3.1, BHRB will capture only user-space
518 			 * addresses, hence include a check before filtering code
519 			 */
520 			if (!(ppmu->flags & PPMU_ARCH_31) &&
521 			    is_kernel_addr(addr) && event->attr.exclude_kernel)
522 				continue;
523 
524 			/* Branches are read most recent first (ie. mfbhrb 0 is
525 			 * the most recent branch).
526 			 * There are two types of valid entries:
527 			 * 1) a target entry which is the to address of a
528 			 *    computed goto like a blr,bctr,btar.  The next
529 			 *    entry read from the bhrb will be branch
530 			 *    corresponding to this target (ie. the actual
531 			 *    blr/bctr/btar instruction).
532 			 * 2) a from address which is an actual branch.  If a
533 			 *    target entry proceeds this, then this is the
534 			 *    matching branch for that target.  If this is not
535 			 *    following a target entry, then this is a branch
536 			 *    where the target is given as an immediate field
537 			 *    in the instruction (ie. an i or b form branch).
538 			 *    In this case we need to read the instruction from
539 			 *    memory to determine the target/to address.
540 			 */
541 
542 			if (val & BHRB_TARGET) {
543 				/* Target branches use two entries
544 				 * (ie. computed gotos/XL form)
545 				 */
546 				cpuhw->bhrb_entries[u_index].to = addr;
547 				cpuhw->bhrb_entries[u_index].mispred = pred;
548 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
549 
550 				/* Get from address in next entry */
551 				val = read_bhrb(r_index++);
552 				addr = val & BHRB_EA;
553 				if (val & BHRB_TARGET) {
554 					/* Shouldn't have two targets in a
555 					   row.. Reset index and try again */
556 					r_index--;
557 					addr = 0;
558 				}
559 				cpuhw->bhrb_entries[u_index].from = addr;
560 			} else {
561 				/* Branches to immediate field
562 				   (ie I or B form) */
563 				cpuhw->bhrb_entries[u_index].from = addr;
564 				cpuhw->bhrb_entries[u_index].to =
565 					power_pmu_bhrb_to(addr);
566 				cpuhw->bhrb_entries[u_index].mispred = pred;
567 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
568 			}
569 			u_index++;
570 
571 		}
572 	}
573 	cpuhw->bhrb_stack.nr = u_index;
574 	cpuhw->bhrb_stack.hw_idx = -1ULL;
575 	return;
576 }
577 
578 static bool is_ebb_event(struct perf_event *event)
579 {
580 	/*
581 	 * This could be a per-PMU callback, but we'd rather avoid the cost. We
582 	 * check that the PMU supports EBB, meaning those that don't can still
583 	 * use bit 63 of the event code for something else if they wish.
584 	 */
585 	return (ppmu->flags & PPMU_ARCH_207S) &&
586 	       ((event->attr.config >> PERF_EVENT_CONFIG_EBB_SHIFT) & 1);
587 }
588 
589 static int ebb_event_check(struct perf_event *event)
590 {
591 	struct perf_event *leader = event->group_leader;
592 
593 	/* Event and group leader must agree on EBB */
594 	if (is_ebb_event(leader) != is_ebb_event(event))
595 		return -EINVAL;
596 
597 	if (is_ebb_event(event)) {
598 		if (!(event->attach_state & PERF_ATTACH_TASK))
599 			return -EINVAL;
600 
601 		if (!leader->attr.pinned || !leader->attr.exclusive)
602 			return -EINVAL;
603 
604 		if (event->attr.freq ||
605 		    event->attr.inherit ||
606 		    event->attr.sample_type ||
607 		    event->attr.sample_period ||
608 		    event->attr.enable_on_exec)
609 			return -EINVAL;
610 	}
611 
612 	return 0;
613 }
614 
615 static void ebb_event_add(struct perf_event *event)
616 {
617 	if (!is_ebb_event(event) || current->thread.used_ebb)
618 		return;
619 
620 	/*
621 	 * IFF this is the first time we've added an EBB event, set
622 	 * PMXE in the user MMCR0 so we can detect when it's cleared by
623 	 * userspace. We need this so that we can context switch while
624 	 * userspace is in the EBB handler (where PMXE is 0).
625 	 */
626 	current->thread.used_ebb = 1;
627 	current->thread.mmcr0 |= MMCR0_PMXE;
628 }
629 
630 static void ebb_switch_out(unsigned long mmcr0)
631 {
632 	if (!(mmcr0 & MMCR0_EBE))
633 		return;
634 
635 	current->thread.siar  = mfspr(SPRN_SIAR);
636 	current->thread.sier  = mfspr(SPRN_SIER);
637 	current->thread.sdar  = mfspr(SPRN_SDAR);
638 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
639 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
640 	if (ppmu->flags & PPMU_ARCH_31) {
641 		current->thread.mmcr3 = mfspr(SPRN_MMCR3);
642 		current->thread.sier2 = mfspr(SPRN_SIER2);
643 		current->thread.sier3 = mfspr(SPRN_SIER3);
644 	}
645 }
646 
647 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
648 {
649 	unsigned long mmcr0 = cpuhw->mmcr.mmcr0;
650 
651 	if (!ebb)
652 		goto out;
653 
654 	/* Enable EBB and read/write to all 6 PMCs and BHRB for userspace */
655 	mmcr0 |= MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC_U6;
656 
657 	/*
658 	 * Add any bits from the user MMCR0, FC or PMAO. This is compatible
659 	 * with pmao_restore_workaround() because we may add PMAO but we never
660 	 * clear it here.
661 	 */
662 	mmcr0 |= current->thread.mmcr0;
663 
664 	/*
665 	 * Be careful not to set PMXE if userspace had it cleared. This is also
666 	 * compatible with pmao_restore_workaround() because it has already
667 	 * cleared PMXE and we leave PMAO alone.
668 	 */
669 	if (!(current->thread.mmcr0 & MMCR0_PMXE))
670 		mmcr0 &= ~MMCR0_PMXE;
671 
672 	mtspr(SPRN_SIAR, current->thread.siar);
673 	mtspr(SPRN_SIER, current->thread.sier);
674 	mtspr(SPRN_SDAR, current->thread.sdar);
675 
676 	/*
677 	 * Merge the kernel & user values of MMCR2. The semantics we implement
678 	 * are that the user MMCR2 can set bits, ie. cause counters to freeze,
679 	 * but not clear bits. If a task wants to be able to clear bits, ie.
680 	 * unfreeze counters, it should not set exclude_xxx in its events and
681 	 * instead manage the MMCR2 entirely by itself.
682 	 */
683 	mtspr(SPRN_MMCR2, cpuhw->mmcr.mmcr2 | current->thread.mmcr2);
684 
685 	if (ppmu->flags & PPMU_ARCH_31) {
686 		mtspr(SPRN_MMCR3, current->thread.mmcr3);
687 		mtspr(SPRN_SIER2, current->thread.sier2);
688 		mtspr(SPRN_SIER3, current->thread.sier3);
689 	}
690 out:
691 	return mmcr0;
692 }
693 
694 static void pmao_restore_workaround(bool ebb)
695 {
696 	unsigned pmcs[6];
697 
698 	if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
699 		return;
700 
701 	/*
702 	 * On POWER8E there is a hardware defect which affects the PMU context
703 	 * switch logic, ie. power_pmu_disable/enable().
704 	 *
705 	 * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
706 	 * by the hardware. Sometime later the actual PMU exception is
707 	 * delivered.
708 	 *
709 	 * If we context switch, or simply disable/enable, the PMU prior to the
710 	 * exception arriving, the exception will be lost when we clear PMAO.
711 	 *
712 	 * When we reenable the PMU, we will write the saved MMCR0 with PMAO
713 	 * set, and this _should_ generate an exception. However because of the
714 	 * defect no exception is generated when we write PMAO, and we get
715 	 * stuck with no counters counting but no exception delivered.
716 	 *
717 	 * The workaround is to detect this case and tweak the hardware to
718 	 * create another pending PMU exception.
719 	 *
720 	 * We do that by setting up PMC6 (cycles) for an imminent overflow and
721 	 * enabling the PMU. That causes a new exception to be generated in the
722 	 * chip, but we don't take it yet because we have interrupts hard
723 	 * disabled. We then write back the PMU state as we want it to be seen
724 	 * by the exception handler. When we reenable interrupts the exception
725 	 * handler will be called and see the correct state.
726 	 *
727 	 * The logic is the same for EBB, except that the exception is gated by
728 	 * us having interrupts hard disabled as well as the fact that we are
729 	 * not in userspace. The exception is finally delivered when we return
730 	 * to userspace.
731 	 */
732 
733 	/* Only if PMAO is set and PMAO_SYNC is clear */
734 	if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
735 		return;
736 
737 	/* If we're doing EBB, only if BESCR[GE] is set */
738 	if (ebb && !(current->thread.bescr & BESCR_GE))
739 		return;
740 
741 	/*
742 	 * We are already soft-disabled in power_pmu_enable(). We need to hard
743 	 * disable to actually prevent the PMU exception from firing.
744 	 */
745 	hard_irq_disable();
746 
747 	/*
748 	 * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
749 	 * Using read/write_pmc() in a for loop adds 12 function calls and
750 	 * almost doubles our code size.
751 	 */
752 	pmcs[0] = mfspr(SPRN_PMC1);
753 	pmcs[1] = mfspr(SPRN_PMC2);
754 	pmcs[2] = mfspr(SPRN_PMC3);
755 	pmcs[3] = mfspr(SPRN_PMC4);
756 	pmcs[4] = mfspr(SPRN_PMC5);
757 	pmcs[5] = mfspr(SPRN_PMC6);
758 
759 	/* Ensure all freeze bits are unset */
760 	mtspr(SPRN_MMCR2, 0);
761 
762 	/* Set up PMC6 to overflow in one cycle */
763 	mtspr(SPRN_PMC6, 0x7FFFFFFE);
764 
765 	/* Enable exceptions and unfreeze PMC6 */
766 	mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
767 
768 	/* Now we need to refreeze and restore the PMCs */
769 	mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
770 
771 	mtspr(SPRN_PMC1, pmcs[0]);
772 	mtspr(SPRN_PMC2, pmcs[1]);
773 	mtspr(SPRN_PMC3, pmcs[2]);
774 	mtspr(SPRN_PMC4, pmcs[3]);
775 	mtspr(SPRN_PMC5, pmcs[4]);
776 	mtspr(SPRN_PMC6, pmcs[5]);
777 }
778 
779 /*
780  * If the perf subsystem wants performance monitor interrupts as soon as
781  * possible (e.g., to sample the instruction address and stack chain),
782  * this should return true. The IRQ masking code can then enable MSR[EE]
783  * in some places (e.g., interrupt handlers) that allows PMI interrupts
784  * through to improve accuracy of profiles, at the cost of some performance.
785  *
786  * The PMU counters can be enabled by other means (e.g., sysfs raw SPR
787  * access), but in that case there is no need for prompt PMI handling.
788  *
789  * This currently returns true if any perf counter is being used. It
790  * could possibly return false if only events are being counted rather than
791  * samples being taken, but for now this is good enough.
792  */
793 bool power_pmu_wants_prompt_pmi(void)
794 {
795 	struct cpu_hw_events *cpuhw;
796 
797 	/*
798 	 * This could simply test local_paca->pmcregs_in_use if that were not
799 	 * under ifdef KVM.
800 	 */
801 	if (!ppmu)
802 		return false;
803 
804 	cpuhw = this_cpu_ptr(&cpu_hw_events);
805 	return cpuhw->n_events;
806 }
807 #endif /* CONFIG_PPC64 */
808 
809 static void perf_event_interrupt(struct pt_regs *regs);
810 
811 /*
812  * Read one performance monitor counter (PMC).
813  */
814 static unsigned long read_pmc(int idx)
815 {
816 	unsigned long val;
817 
818 	switch (idx) {
819 	case 1:
820 		val = mfspr(SPRN_PMC1);
821 		break;
822 	case 2:
823 		val = mfspr(SPRN_PMC2);
824 		break;
825 	case 3:
826 		val = mfspr(SPRN_PMC3);
827 		break;
828 	case 4:
829 		val = mfspr(SPRN_PMC4);
830 		break;
831 	case 5:
832 		val = mfspr(SPRN_PMC5);
833 		break;
834 	case 6:
835 		val = mfspr(SPRN_PMC6);
836 		break;
837 #ifdef CONFIG_PPC64
838 	case 7:
839 		val = mfspr(SPRN_PMC7);
840 		break;
841 	case 8:
842 		val = mfspr(SPRN_PMC8);
843 		break;
844 #endif /* CONFIG_PPC64 */
845 	default:
846 		printk(KERN_ERR "oops trying to read PMC%d\n", idx);
847 		val = 0;
848 	}
849 	return val;
850 }
851 
852 /*
853  * Write one PMC.
854  */
855 static void write_pmc(int idx, unsigned long val)
856 {
857 	switch (idx) {
858 	case 1:
859 		mtspr(SPRN_PMC1, val);
860 		break;
861 	case 2:
862 		mtspr(SPRN_PMC2, val);
863 		break;
864 	case 3:
865 		mtspr(SPRN_PMC3, val);
866 		break;
867 	case 4:
868 		mtspr(SPRN_PMC4, val);
869 		break;
870 	case 5:
871 		mtspr(SPRN_PMC5, val);
872 		break;
873 	case 6:
874 		mtspr(SPRN_PMC6, val);
875 		break;
876 #ifdef CONFIG_PPC64
877 	case 7:
878 		mtspr(SPRN_PMC7, val);
879 		break;
880 	case 8:
881 		mtspr(SPRN_PMC8, val);
882 		break;
883 #endif /* CONFIG_PPC64 */
884 	default:
885 		printk(KERN_ERR "oops trying to write PMC%d\n", idx);
886 	}
887 }
888 
889 static int any_pmc_overflown(struct cpu_hw_events *cpuhw)
890 {
891 	int i, idx;
892 
893 	for (i = 0; i < cpuhw->n_events; i++) {
894 		idx = cpuhw->event[i]->hw.idx;
895 		if ((idx) && ((int)read_pmc(idx) < 0))
896 			return idx;
897 	}
898 
899 	return 0;
900 }
901 
902 /* Called from sysrq_handle_showregs() */
903 void perf_event_print_debug(void)
904 {
905 	unsigned long sdar, sier, flags;
906 	u32 pmcs[MAX_HWEVENTS];
907 	int i;
908 
909 	if (!ppmu) {
910 		pr_info("Performance monitor hardware not registered.\n");
911 		return;
912 	}
913 
914 	if (!ppmu->n_counter)
915 		return;
916 
917 	local_irq_save(flags);
918 
919 	pr_info("CPU: %d PMU registers, ppmu = %s n_counters = %d",
920 		 smp_processor_id(), ppmu->name, ppmu->n_counter);
921 
922 	for (i = 0; i < ppmu->n_counter; i++)
923 		pmcs[i] = read_pmc(i + 1);
924 
925 	for (; i < MAX_HWEVENTS; i++)
926 		pmcs[i] = 0xdeadbeef;
927 
928 	pr_info("PMC1:  %08x PMC2: %08x PMC3: %08x PMC4: %08x\n",
929 		 pmcs[0], pmcs[1], pmcs[2], pmcs[3]);
930 
931 	if (ppmu->n_counter > 4)
932 		pr_info("PMC5:  %08x PMC6: %08x PMC7: %08x PMC8: %08x\n",
933 			 pmcs[4], pmcs[5], pmcs[6], pmcs[7]);
934 
935 	pr_info("MMCR0: %016lx MMCR1: %016lx MMCRA: %016lx\n",
936 		mfspr(SPRN_MMCR0), mfspr(SPRN_MMCR1), mfspr(SPRN_MMCRA));
937 
938 	sdar = sier = 0;
939 #ifdef CONFIG_PPC64
940 	sdar = mfspr(SPRN_SDAR);
941 
942 	if (ppmu->flags & PPMU_HAS_SIER)
943 		sier = mfspr(SPRN_SIER);
944 
945 	if (ppmu->flags & PPMU_ARCH_207S) {
946 		pr_info("MMCR2: %016lx EBBHR: %016lx\n",
947 			mfspr(SPRN_MMCR2), mfspr(SPRN_EBBHR));
948 		pr_info("EBBRR: %016lx BESCR: %016lx\n",
949 			mfspr(SPRN_EBBRR), mfspr(SPRN_BESCR));
950 	}
951 
952 	if (ppmu->flags & PPMU_ARCH_31) {
953 		pr_info("MMCR3: %016lx SIER2: %016lx SIER3: %016lx\n",
954 			mfspr(SPRN_MMCR3), mfspr(SPRN_SIER2), mfspr(SPRN_SIER3));
955 	}
956 #endif
957 	pr_info("SIAR:  %016lx SDAR:  %016lx SIER:  %016lx\n",
958 		mfspr(SPRN_SIAR), sdar, sier);
959 
960 	local_irq_restore(flags);
961 }
962 
963 /*
964  * Check if a set of events can all go on the PMU at once.
965  * If they can't, this will look at alternative codes for the events
966  * and see if any combination of alternative codes is feasible.
967  * The feasible set is returned in event_id[].
968  */
969 static int power_check_constraints(struct cpu_hw_events *cpuhw,
970 				   u64 event_id[], unsigned int cflags[],
971 				   int n_ev, struct perf_event **event)
972 {
973 	unsigned long mask, value, nv;
974 	unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
975 	int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
976 	int i, j;
977 	unsigned long addf = ppmu->add_fields;
978 	unsigned long tadd = ppmu->test_adder;
979 	unsigned long grp_mask = ppmu->group_constraint_mask;
980 	unsigned long grp_val = ppmu->group_constraint_val;
981 
982 	if (n_ev > ppmu->n_counter)
983 		return -1;
984 
985 	/* First see if the events will go on as-is */
986 	for (i = 0; i < n_ev; ++i) {
987 		if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
988 		    && !ppmu->limited_pmc_event(event_id[i])) {
989 			ppmu->get_alternatives(event_id[i], cflags[i],
990 					       cpuhw->alternatives[i]);
991 			event_id[i] = cpuhw->alternatives[i][0];
992 		}
993 		if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
994 					 &cpuhw->avalues[i][0], event[i]->attr.config1))
995 			return -1;
996 	}
997 	value = mask = 0;
998 	for (i = 0; i < n_ev; ++i) {
999 		nv = (value | cpuhw->avalues[i][0]) +
1000 			(value & cpuhw->avalues[i][0] & addf);
1001 
1002 		if (((((nv + tadd) ^ value) & mask) & (~grp_mask)) != 0)
1003 			break;
1004 
1005 		if (((((nv + tadd) ^ cpuhw->avalues[i][0]) & cpuhw->amasks[i][0])
1006 			& (~grp_mask)) != 0)
1007 			break;
1008 
1009 		value = nv;
1010 		mask |= cpuhw->amasks[i][0];
1011 	}
1012 	if (i == n_ev) {
1013 		if ((value & mask & grp_mask) != (mask & grp_val))
1014 			return -1;
1015 		else
1016 			return 0;	/* all OK */
1017 	}
1018 
1019 	/* doesn't work, gather alternatives... */
1020 	if (!ppmu->get_alternatives)
1021 		return -1;
1022 	for (i = 0; i < n_ev; ++i) {
1023 		choice[i] = 0;
1024 		n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
1025 						  cpuhw->alternatives[i]);
1026 		for (j = 1; j < n_alt[i]; ++j)
1027 			ppmu->get_constraint(cpuhw->alternatives[i][j],
1028 					     &cpuhw->amasks[i][j],
1029 					     &cpuhw->avalues[i][j],
1030 					     event[i]->attr.config1);
1031 	}
1032 
1033 	/* enumerate all possibilities and see if any will work */
1034 	i = 0;
1035 	j = -1;
1036 	value = mask = nv = 0;
1037 	while (i < n_ev) {
1038 		if (j >= 0) {
1039 			/* we're backtracking, restore context */
1040 			value = svalues[i];
1041 			mask = smasks[i];
1042 			j = choice[i];
1043 		}
1044 		/*
1045 		 * See if any alternative k for event_id i,
1046 		 * where k > j, will satisfy the constraints.
1047 		 */
1048 		while (++j < n_alt[i]) {
1049 			nv = (value | cpuhw->avalues[i][j]) +
1050 				(value & cpuhw->avalues[i][j] & addf);
1051 			if ((((nv + tadd) ^ value) & mask) == 0 &&
1052 			    (((nv + tadd) ^ cpuhw->avalues[i][j])
1053 			     & cpuhw->amasks[i][j]) == 0)
1054 				break;
1055 		}
1056 		if (j >= n_alt[i]) {
1057 			/*
1058 			 * No feasible alternative, backtrack
1059 			 * to event_id i-1 and continue enumerating its
1060 			 * alternatives from where we got up to.
1061 			 */
1062 			if (--i < 0)
1063 				return -1;
1064 		} else {
1065 			/*
1066 			 * Found a feasible alternative for event_id i,
1067 			 * remember where we got up to with this event_id,
1068 			 * go on to the next event_id, and start with
1069 			 * the first alternative for it.
1070 			 */
1071 			choice[i] = j;
1072 			svalues[i] = value;
1073 			smasks[i] = mask;
1074 			value = nv;
1075 			mask |= cpuhw->amasks[i][j];
1076 			++i;
1077 			j = -1;
1078 		}
1079 	}
1080 
1081 	/* OK, we have a feasible combination, tell the caller the solution */
1082 	for (i = 0; i < n_ev; ++i)
1083 		event_id[i] = cpuhw->alternatives[i][choice[i]];
1084 	return 0;
1085 }
1086 
1087 /*
1088  * Check if newly-added events have consistent settings for
1089  * exclude_{user,kernel,hv} with each other and any previously
1090  * added events.
1091  */
1092 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
1093 			  int n_prev, int n_new)
1094 {
1095 	int eu = 0, ek = 0, eh = 0;
1096 	int i, n, first;
1097 	struct perf_event *event;
1098 
1099 	/*
1100 	 * If the PMU we're on supports per event exclude settings then we
1101 	 * don't need to do any of this logic. NB. This assumes no PMU has both
1102 	 * per event exclude and limited PMCs.
1103 	 */
1104 	if (ppmu->flags & PPMU_ARCH_207S)
1105 		return 0;
1106 
1107 	n = n_prev + n_new;
1108 	if (n <= 1)
1109 		return 0;
1110 
1111 	first = 1;
1112 	for (i = 0; i < n; ++i) {
1113 		if (cflags[i] & PPMU_LIMITED_PMC_OK) {
1114 			cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
1115 			continue;
1116 		}
1117 		event = ctrs[i];
1118 		if (first) {
1119 			eu = event->attr.exclude_user;
1120 			ek = event->attr.exclude_kernel;
1121 			eh = event->attr.exclude_hv;
1122 			first = 0;
1123 		} else if (event->attr.exclude_user != eu ||
1124 			   event->attr.exclude_kernel != ek ||
1125 			   event->attr.exclude_hv != eh) {
1126 			return -EAGAIN;
1127 		}
1128 	}
1129 
1130 	if (eu || ek || eh)
1131 		for (i = 0; i < n; ++i)
1132 			if (cflags[i] & PPMU_LIMITED_PMC_OK)
1133 				cflags[i] |= PPMU_LIMITED_PMC_REQD;
1134 
1135 	return 0;
1136 }
1137 
1138 static u64 check_and_compute_delta(u64 prev, u64 val)
1139 {
1140 	u64 delta = (val - prev) & 0xfffffffful;
1141 
1142 	/*
1143 	 * POWER7 can roll back counter values, if the new value is smaller
1144 	 * than the previous value it will cause the delta and the counter to
1145 	 * have bogus values unless we rolled a counter over.  If a counter is
1146 	 * rolled back, it will be smaller, but within 256, which is the maximum
1147 	 * number of events to rollback at once.  If we detect a rollback
1148 	 * return 0.  This can lead to a small lack of precision in the
1149 	 * counters.
1150 	 */
1151 	if (prev > val && (prev - val) < 256)
1152 		delta = 0;
1153 
1154 	return delta;
1155 }
1156 
1157 static void power_pmu_read(struct perf_event *event)
1158 {
1159 	s64 val, delta, prev;
1160 
1161 	if (event->hw.state & PERF_HES_STOPPED)
1162 		return;
1163 
1164 	if (!event->hw.idx)
1165 		return;
1166 
1167 	if (is_ebb_event(event)) {
1168 		val = read_pmc(event->hw.idx);
1169 		local64_set(&event->hw.prev_count, val);
1170 		return;
1171 	}
1172 
1173 	/*
1174 	 * Performance monitor interrupts come even when interrupts
1175 	 * are soft-disabled, as long as interrupts are hard-enabled.
1176 	 * Therefore we treat them like NMIs.
1177 	 */
1178 	do {
1179 		prev = local64_read(&event->hw.prev_count);
1180 		barrier();
1181 		val = read_pmc(event->hw.idx);
1182 		delta = check_and_compute_delta(prev, val);
1183 		if (!delta)
1184 			return;
1185 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
1186 
1187 	local64_add(delta, &event->count);
1188 
1189 	/*
1190 	 * A number of places program the PMC with (0x80000000 - period_left).
1191 	 * We never want period_left to be less than 1 because we will program
1192 	 * the PMC with a value >= 0x800000000 and an edge detected PMC will
1193 	 * roll around to 0 before taking an exception. We have seen this
1194 	 * on POWER8.
1195 	 *
1196 	 * To fix this, clamp the minimum value of period_left to 1.
1197 	 */
1198 	do {
1199 		prev = local64_read(&event->hw.period_left);
1200 		val = prev - delta;
1201 		if (val < 1)
1202 			val = 1;
1203 	} while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
1204 }
1205 
1206 /*
1207  * On some machines, PMC5 and PMC6 can't be written, don't respect
1208  * the freeze conditions, and don't generate interrupts.  This tells
1209  * us if `event' is using such a PMC.
1210  */
1211 static int is_limited_pmc(int pmcnum)
1212 {
1213 	return (ppmu->flags & PPMU_LIMITED_PMC5_6)
1214 		&& (pmcnum == 5 || pmcnum == 6);
1215 }
1216 
1217 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
1218 				    unsigned long pmc5, unsigned long pmc6)
1219 {
1220 	struct perf_event *event;
1221 	u64 val, prev, delta;
1222 	int i;
1223 
1224 	for (i = 0; i < cpuhw->n_limited; ++i) {
1225 		event = cpuhw->limited_counter[i];
1226 		if (!event->hw.idx)
1227 			continue;
1228 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1229 		prev = local64_read(&event->hw.prev_count);
1230 		event->hw.idx = 0;
1231 		delta = check_and_compute_delta(prev, val);
1232 		if (delta)
1233 			local64_add(delta, &event->count);
1234 	}
1235 }
1236 
1237 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
1238 				  unsigned long pmc5, unsigned long pmc6)
1239 {
1240 	struct perf_event *event;
1241 	u64 val, prev;
1242 	int i;
1243 
1244 	for (i = 0; i < cpuhw->n_limited; ++i) {
1245 		event = cpuhw->limited_counter[i];
1246 		event->hw.idx = cpuhw->limited_hwidx[i];
1247 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1248 		prev = local64_read(&event->hw.prev_count);
1249 		if (check_and_compute_delta(prev, val))
1250 			local64_set(&event->hw.prev_count, val);
1251 		perf_event_update_userpage(event);
1252 	}
1253 }
1254 
1255 /*
1256  * Since limited events don't respect the freeze conditions, we
1257  * have to read them immediately after freezing or unfreezing the
1258  * other events.  We try to keep the values from the limited
1259  * events as consistent as possible by keeping the delay (in
1260  * cycles and instructions) between freezing/unfreezing and reading
1261  * the limited events as small and consistent as possible.
1262  * Therefore, if any limited events are in use, we read them
1263  * both, and always in the same order, to minimize variability,
1264  * and do it inside the same asm that writes MMCR0.
1265  */
1266 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
1267 {
1268 	unsigned long pmc5, pmc6;
1269 
1270 	if (!cpuhw->n_limited) {
1271 		mtspr(SPRN_MMCR0, mmcr0);
1272 		return;
1273 	}
1274 
1275 	/*
1276 	 * Write MMCR0, then read PMC5 and PMC6 immediately.
1277 	 * To ensure we don't get a performance monitor interrupt
1278 	 * between writing MMCR0 and freezing/thawing the limited
1279 	 * events, we first write MMCR0 with the event overflow
1280 	 * interrupt enable bits turned off.
1281 	 */
1282 	asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
1283 		     : "=&r" (pmc5), "=&r" (pmc6)
1284 		     : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
1285 		       "i" (SPRN_MMCR0),
1286 		       "i" (SPRN_PMC5), "i" (SPRN_PMC6));
1287 
1288 	if (mmcr0 & MMCR0_FC)
1289 		freeze_limited_counters(cpuhw, pmc5, pmc6);
1290 	else
1291 		thaw_limited_counters(cpuhw, pmc5, pmc6);
1292 
1293 	/*
1294 	 * Write the full MMCR0 including the event overflow interrupt
1295 	 * enable bits, if necessary.
1296 	 */
1297 	if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
1298 		mtspr(SPRN_MMCR0, mmcr0);
1299 }
1300 
1301 /*
1302  * Disable all events to prevent PMU interrupts and to allow
1303  * events to be added or removed.
1304  */
1305 static void power_pmu_disable(struct pmu *pmu)
1306 {
1307 	struct cpu_hw_events *cpuhw;
1308 	unsigned long flags, mmcr0, val, mmcra;
1309 
1310 	if (!ppmu)
1311 		return;
1312 	local_irq_save(flags);
1313 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1314 
1315 	if (!cpuhw->disabled) {
1316 		/*
1317 		 * Check if we ever enabled the PMU on this cpu.
1318 		 */
1319 		if (!cpuhw->pmcs_enabled) {
1320 			ppc_enable_pmcs();
1321 			cpuhw->pmcs_enabled = 1;
1322 		}
1323 
1324 		/*
1325 		 * Set the 'freeze counters' bit, clear EBE/BHRBA/PMCC/PMAO/FC56
1326 		 * Also clear PMXE to disable PMI's getting triggered in some
1327 		 * corner cases during PMU disable.
1328 		 */
1329 		val  = mmcr0 = mfspr(SPRN_MMCR0);
1330 		val |= MMCR0_FC;
1331 		val &= ~(MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC | MMCR0_PMAO |
1332 			 MMCR0_PMXE | MMCR0_FC56);
1333 		/* Set mmcr0 PMCCEXT for p10 */
1334 		if (ppmu->flags & PPMU_ARCH_31)
1335 			val |= MMCR0_PMCCEXT;
1336 
1337 		/*
1338 		 * The barrier is to make sure the mtspr has been
1339 		 * executed and the PMU has frozen the events etc.
1340 		 * before we return.
1341 		 */
1342 		write_mmcr0(cpuhw, val);
1343 		mb();
1344 		isync();
1345 
1346 		/*
1347 		 * Some corner cases could clear the PMU counter overflow
1348 		 * while a masked PMI is pending. One such case is when
1349 		 * a PMI happens during interrupt replay and perf counter
1350 		 * values are cleared by PMU callbacks before replay.
1351 		 *
1352 		 * If any PMC corresponding to the active PMU events are
1353 		 * overflown, disable the interrupt by clearing the paca
1354 		 * bit for PMI since we are disabling the PMU now.
1355 		 * Otherwise provide a warning if there is PMI pending, but
1356 		 * no counter is found overflown.
1357 		 */
1358 		if (any_pmc_overflown(cpuhw)) {
1359 			/*
1360 			 * Since power_pmu_disable runs under local_irq_save, it
1361 			 * could happen that code hits a PMC overflow without PMI
1362 			 * pending in paca. Hence only clear PMI pending if it was
1363 			 * set.
1364 			 *
1365 			 * If a PMI is pending, then MSR[EE] must be disabled (because
1366 			 * the masked PMI handler disabling EE). So it is safe to
1367 			 * call clear_pmi_irq_pending().
1368 			 */
1369 			if (pmi_irq_pending())
1370 				clear_pmi_irq_pending();
1371 		} else
1372 			WARN_ON(pmi_irq_pending());
1373 
1374 		val = mmcra = cpuhw->mmcr.mmcra;
1375 
1376 		/*
1377 		 * Disable instruction sampling if it was enabled
1378 		 */
1379 		if (cpuhw->mmcr.mmcra & MMCRA_SAMPLE_ENABLE)
1380 			val &= ~MMCRA_SAMPLE_ENABLE;
1381 
1382 		/* Disable BHRB via mmcra (BHRBRD) for p10 */
1383 		if (ppmu->flags & PPMU_ARCH_31)
1384 			val |= MMCRA_BHRB_DISABLE;
1385 
1386 		/*
1387 		 * Write SPRN_MMCRA if mmcra has either disabled
1388 		 * instruction sampling or BHRB.
1389 		 */
1390 		if (val != mmcra) {
1391 			mtspr(SPRN_MMCRA, mmcra);
1392 			mb();
1393 			isync();
1394 		}
1395 
1396 		cpuhw->disabled = 1;
1397 		cpuhw->n_added = 0;
1398 
1399 		ebb_switch_out(mmcr0);
1400 
1401 #ifdef CONFIG_PPC64
1402 		/*
1403 		 * These are readable by userspace, may contain kernel
1404 		 * addresses and are not switched by context switch, so clear
1405 		 * them now to avoid leaking anything to userspace in general
1406 		 * including to another process.
1407 		 */
1408 		if (ppmu->flags & PPMU_ARCH_207S) {
1409 			mtspr(SPRN_SDAR, 0);
1410 			mtspr(SPRN_SIAR, 0);
1411 		}
1412 #endif
1413 	}
1414 
1415 	local_irq_restore(flags);
1416 }
1417 
1418 /*
1419  * Re-enable all events if disable == 0.
1420  * If we were previously disabled and events were added, then
1421  * put the new config on the PMU.
1422  */
1423 static void power_pmu_enable(struct pmu *pmu)
1424 {
1425 	struct perf_event *event;
1426 	struct cpu_hw_events *cpuhw;
1427 	unsigned long flags;
1428 	long i;
1429 	unsigned long val, mmcr0;
1430 	s64 left;
1431 	unsigned int hwc_index[MAX_HWEVENTS];
1432 	int n_lim;
1433 	int idx;
1434 	bool ebb;
1435 
1436 	if (!ppmu)
1437 		return;
1438 	local_irq_save(flags);
1439 
1440 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1441 	if (!cpuhw->disabled)
1442 		goto out;
1443 
1444 	if (cpuhw->n_events == 0) {
1445 		ppc_set_pmu_inuse(0);
1446 		goto out;
1447 	}
1448 
1449 	cpuhw->disabled = 0;
1450 
1451 	/*
1452 	 * EBB requires an exclusive group and all events must have the EBB
1453 	 * flag set, or not set, so we can just check a single event. Also we
1454 	 * know we have at least one event.
1455 	 */
1456 	ebb = is_ebb_event(cpuhw->event[0]);
1457 
1458 	/*
1459 	 * If we didn't change anything, or only removed events,
1460 	 * no need to recalculate MMCR* settings and reset the PMCs.
1461 	 * Just reenable the PMU with the current MMCR* settings
1462 	 * (possibly updated for removal of events).
1463 	 */
1464 	if (!cpuhw->n_added) {
1465 		/*
1466 		 * If there is any active event with an overflown PMC
1467 		 * value, set back PACA_IRQ_PMI which would have been
1468 		 * cleared in power_pmu_disable().
1469 		 */
1470 		hard_irq_disable();
1471 		if (any_pmc_overflown(cpuhw))
1472 			set_pmi_irq_pending();
1473 
1474 		mtspr(SPRN_MMCRA, cpuhw->mmcr.mmcra & ~MMCRA_SAMPLE_ENABLE);
1475 		mtspr(SPRN_MMCR1, cpuhw->mmcr.mmcr1);
1476 		if (ppmu->flags & PPMU_ARCH_31)
1477 			mtspr(SPRN_MMCR3, cpuhw->mmcr.mmcr3);
1478 		goto out_enable;
1479 	}
1480 
1481 	/*
1482 	 * Clear all MMCR settings and recompute them for the new set of events.
1483 	 */
1484 	memset(&cpuhw->mmcr, 0, sizeof(cpuhw->mmcr));
1485 
1486 	if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
1487 			       &cpuhw->mmcr, cpuhw->event, ppmu->flags)) {
1488 		/* shouldn't ever get here */
1489 		printk(KERN_ERR "oops compute_mmcr failed\n");
1490 		goto out;
1491 	}
1492 
1493 	if (!(ppmu->flags & PPMU_ARCH_207S)) {
1494 		/*
1495 		 * Add in MMCR0 freeze bits corresponding to the attr.exclude_*
1496 		 * bits for the first event. We have already checked that all
1497 		 * events have the same value for these bits as the first event.
1498 		 */
1499 		event = cpuhw->event[0];
1500 		if (event->attr.exclude_user)
1501 			cpuhw->mmcr.mmcr0 |= MMCR0_FCP;
1502 		if (event->attr.exclude_kernel)
1503 			cpuhw->mmcr.mmcr0 |= freeze_events_kernel;
1504 		if (event->attr.exclude_hv)
1505 			cpuhw->mmcr.mmcr0 |= MMCR0_FCHV;
1506 	}
1507 
1508 	/*
1509 	 * Write the new configuration to MMCR* with the freeze
1510 	 * bit set and set the hardware events to their initial values.
1511 	 * Then unfreeze the events.
1512 	 */
1513 	ppc_set_pmu_inuse(1);
1514 	mtspr(SPRN_MMCRA, cpuhw->mmcr.mmcra & ~MMCRA_SAMPLE_ENABLE);
1515 	mtspr(SPRN_MMCR1, cpuhw->mmcr.mmcr1);
1516 	mtspr(SPRN_MMCR0, (cpuhw->mmcr.mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
1517 				| MMCR0_FC);
1518 	if (ppmu->flags & PPMU_ARCH_207S)
1519 		mtspr(SPRN_MMCR2, cpuhw->mmcr.mmcr2);
1520 
1521 	if (ppmu->flags & PPMU_ARCH_31)
1522 		mtspr(SPRN_MMCR3, cpuhw->mmcr.mmcr3);
1523 
1524 	/*
1525 	 * Read off any pre-existing events that need to move
1526 	 * to another PMC.
1527 	 */
1528 	for (i = 0; i < cpuhw->n_events; ++i) {
1529 		event = cpuhw->event[i];
1530 		if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
1531 			power_pmu_read(event);
1532 			write_pmc(event->hw.idx, 0);
1533 			event->hw.idx = 0;
1534 		}
1535 	}
1536 
1537 	/*
1538 	 * Initialize the PMCs for all the new and moved events.
1539 	 */
1540 	cpuhw->n_limited = n_lim = 0;
1541 	for (i = 0; i < cpuhw->n_events; ++i) {
1542 		event = cpuhw->event[i];
1543 		if (event->hw.idx)
1544 			continue;
1545 		idx = hwc_index[i] + 1;
1546 		if (is_limited_pmc(idx)) {
1547 			cpuhw->limited_counter[n_lim] = event;
1548 			cpuhw->limited_hwidx[n_lim] = idx;
1549 			++n_lim;
1550 			continue;
1551 		}
1552 
1553 		if (ebb)
1554 			val = local64_read(&event->hw.prev_count);
1555 		else {
1556 			val = 0;
1557 			if (event->hw.sample_period) {
1558 				left = local64_read(&event->hw.period_left);
1559 				if (left < 0x80000000L)
1560 					val = 0x80000000L - left;
1561 			}
1562 			local64_set(&event->hw.prev_count, val);
1563 		}
1564 
1565 		event->hw.idx = idx;
1566 		if (event->hw.state & PERF_HES_STOPPED)
1567 			val = 0;
1568 		write_pmc(idx, val);
1569 
1570 		perf_event_update_userpage(event);
1571 	}
1572 	cpuhw->n_limited = n_lim;
1573 	cpuhw->mmcr.mmcr0 |= MMCR0_PMXE | MMCR0_FCECE;
1574 
1575  out_enable:
1576 	pmao_restore_workaround(ebb);
1577 
1578 	mmcr0 = ebb_switch_in(ebb, cpuhw);
1579 
1580 	mb();
1581 	if (cpuhw->bhrb_users)
1582 		ppmu->config_bhrb(cpuhw->bhrb_filter);
1583 
1584 	write_mmcr0(cpuhw, mmcr0);
1585 
1586 	/*
1587 	 * Enable instruction sampling if necessary
1588 	 */
1589 	if (cpuhw->mmcr.mmcra & MMCRA_SAMPLE_ENABLE) {
1590 		mb();
1591 		mtspr(SPRN_MMCRA, cpuhw->mmcr.mmcra);
1592 	}
1593 
1594  out:
1595 
1596 	local_irq_restore(flags);
1597 }
1598 
1599 static int collect_events(struct perf_event *group, int max_count,
1600 			  struct perf_event *ctrs[], u64 *events,
1601 			  unsigned int *flags)
1602 {
1603 	int n = 0;
1604 	struct perf_event *event;
1605 
1606 	if (group->pmu->task_ctx_nr == perf_hw_context) {
1607 		if (n >= max_count)
1608 			return -1;
1609 		ctrs[n] = group;
1610 		flags[n] = group->hw.event_base;
1611 		events[n++] = group->hw.config;
1612 	}
1613 	for_each_sibling_event(event, group) {
1614 		if (event->pmu->task_ctx_nr == perf_hw_context &&
1615 		    event->state != PERF_EVENT_STATE_OFF) {
1616 			if (n >= max_count)
1617 				return -1;
1618 			ctrs[n] = event;
1619 			flags[n] = event->hw.event_base;
1620 			events[n++] = event->hw.config;
1621 		}
1622 	}
1623 	return n;
1624 }
1625 
1626 /*
1627  * Add an event to the PMU.
1628  * If all events are not already frozen, then we disable and
1629  * re-enable the PMU in order to get hw_perf_enable to do the
1630  * actual work of reconfiguring the PMU.
1631  */
1632 static int power_pmu_add(struct perf_event *event, int ef_flags)
1633 {
1634 	struct cpu_hw_events *cpuhw;
1635 	unsigned long flags;
1636 	int n0;
1637 	int ret = -EAGAIN;
1638 
1639 	local_irq_save(flags);
1640 	perf_pmu_disable(event->pmu);
1641 
1642 	/*
1643 	 * Add the event to the list (if there is room)
1644 	 * and check whether the total set is still feasible.
1645 	 */
1646 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1647 	n0 = cpuhw->n_events;
1648 	if (n0 >= ppmu->n_counter)
1649 		goto out;
1650 	cpuhw->event[n0] = event;
1651 	cpuhw->events[n0] = event->hw.config;
1652 	cpuhw->flags[n0] = event->hw.event_base;
1653 
1654 	/*
1655 	 * This event may have been disabled/stopped in record_and_restart()
1656 	 * because we exceeded the ->event_limit. If re-starting the event,
1657 	 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
1658 	 * notification is re-enabled.
1659 	 */
1660 	if (!(ef_flags & PERF_EF_START))
1661 		event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1662 	else
1663 		event->hw.state = 0;
1664 
1665 	/*
1666 	 * If group events scheduling transaction was started,
1667 	 * skip the schedulability test here, it will be performed
1668 	 * at commit time(->commit_txn) as a whole
1669 	 */
1670 	if (cpuhw->txn_flags & PERF_PMU_TXN_ADD)
1671 		goto nocheck;
1672 
1673 	if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
1674 		goto out;
1675 	if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1, cpuhw->event))
1676 		goto out;
1677 	event->hw.config = cpuhw->events[n0];
1678 
1679 nocheck:
1680 	ebb_event_add(event);
1681 
1682 	++cpuhw->n_events;
1683 	++cpuhw->n_added;
1684 
1685 	ret = 0;
1686  out:
1687 	if (has_branch_stack(event)) {
1688 		u64 bhrb_filter = -1;
1689 
1690 		if (ppmu->bhrb_filter_map)
1691 			bhrb_filter = ppmu->bhrb_filter_map(
1692 				event->attr.branch_sample_type);
1693 
1694 		if (bhrb_filter != -1) {
1695 			cpuhw->bhrb_filter = bhrb_filter;
1696 			power_pmu_bhrb_enable(event);
1697 		}
1698 	}
1699 
1700 	perf_pmu_enable(event->pmu);
1701 	local_irq_restore(flags);
1702 	return ret;
1703 }
1704 
1705 /*
1706  * Remove an event from the PMU.
1707  */
1708 static void power_pmu_del(struct perf_event *event, int ef_flags)
1709 {
1710 	struct cpu_hw_events *cpuhw;
1711 	long i;
1712 	unsigned long flags;
1713 
1714 	local_irq_save(flags);
1715 	perf_pmu_disable(event->pmu);
1716 
1717 	power_pmu_read(event);
1718 
1719 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1720 	for (i = 0; i < cpuhw->n_events; ++i) {
1721 		if (event == cpuhw->event[i]) {
1722 			while (++i < cpuhw->n_events) {
1723 				cpuhw->event[i-1] = cpuhw->event[i];
1724 				cpuhw->events[i-1] = cpuhw->events[i];
1725 				cpuhw->flags[i-1] = cpuhw->flags[i];
1726 			}
1727 			--cpuhw->n_events;
1728 			ppmu->disable_pmc(event->hw.idx - 1, &cpuhw->mmcr);
1729 			if (event->hw.idx) {
1730 				write_pmc(event->hw.idx, 0);
1731 				event->hw.idx = 0;
1732 			}
1733 			perf_event_update_userpage(event);
1734 			break;
1735 		}
1736 	}
1737 	for (i = 0; i < cpuhw->n_limited; ++i)
1738 		if (event == cpuhw->limited_counter[i])
1739 			break;
1740 	if (i < cpuhw->n_limited) {
1741 		while (++i < cpuhw->n_limited) {
1742 			cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
1743 			cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
1744 		}
1745 		--cpuhw->n_limited;
1746 	}
1747 	if (cpuhw->n_events == 0) {
1748 		/* disable exceptions if no events are running */
1749 		cpuhw->mmcr.mmcr0 &= ~(MMCR0_PMXE | MMCR0_FCECE);
1750 	}
1751 
1752 	if (has_branch_stack(event))
1753 		power_pmu_bhrb_disable(event);
1754 
1755 	perf_pmu_enable(event->pmu);
1756 	local_irq_restore(flags);
1757 }
1758 
1759 /*
1760  * POWER-PMU does not support disabling individual counters, hence
1761  * program their cycle counter to their max value and ignore the interrupts.
1762  */
1763 
1764 static void power_pmu_start(struct perf_event *event, int ef_flags)
1765 {
1766 	unsigned long flags;
1767 	s64 left;
1768 	unsigned long val;
1769 
1770 	if (!event->hw.idx || !event->hw.sample_period)
1771 		return;
1772 
1773 	if (!(event->hw.state & PERF_HES_STOPPED))
1774 		return;
1775 
1776 	if (ef_flags & PERF_EF_RELOAD)
1777 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1778 
1779 	local_irq_save(flags);
1780 	perf_pmu_disable(event->pmu);
1781 
1782 	event->hw.state = 0;
1783 	left = local64_read(&event->hw.period_left);
1784 
1785 	val = 0;
1786 	if (left < 0x80000000L)
1787 		val = 0x80000000L - left;
1788 
1789 	write_pmc(event->hw.idx, val);
1790 
1791 	perf_event_update_userpage(event);
1792 	perf_pmu_enable(event->pmu);
1793 	local_irq_restore(flags);
1794 }
1795 
1796 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1797 {
1798 	unsigned long flags;
1799 
1800 	if (!event->hw.idx || !event->hw.sample_period)
1801 		return;
1802 
1803 	if (event->hw.state & PERF_HES_STOPPED)
1804 		return;
1805 
1806 	local_irq_save(flags);
1807 	perf_pmu_disable(event->pmu);
1808 
1809 	power_pmu_read(event);
1810 	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1811 	write_pmc(event->hw.idx, 0);
1812 
1813 	perf_event_update_userpage(event);
1814 	perf_pmu_enable(event->pmu);
1815 	local_irq_restore(flags);
1816 }
1817 
1818 /*
1819  * Start group events scheduling transaction
1820  * Set the flag to make pmu::enable() not perform the
1821  * schedulability test, it will be performed at commit time
1822  *
1823  * We only support PERF_PMU_TXN_ADD transactions. Save the
1824  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1825  * transactions.
1826  */
1827 static void power_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1828 {
1829 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1830 
1831 	WARN_ON_ONCE(cpuhw->txn_flags);		/* txn already in flight */
1832 
1833 	cpuhw->txn_flags = txn_flags;
1834 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1835 		return;
1836 
1837 	perf_pmu_disable(pmu);
1838 	cpuhw->n_txn_start = cpuhw->n_events;
1839 }
1840 
1841 /*
1842  * Stop group events scheduling transaction
1843  * Clear the flag and pmu::enable() will perform the
1844  * schedulability test.
1845  */
1846 static void power_pmu_cancel_txn(struct pmu *pmu)
1847 {
1848 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1849 	unsigned int txn_flags;
1850 
1851 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1852 
1853 	txn_flags = cpuhw->txn_flags;
1854 	cpuhw->txn_flags = 0;
1855 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1856 		return;
1857 
1858 	perf_pmu_enable(pmu);
1859 }
1860 
1861 /*
1862  * Commit group events scheduling transaction
1863  * Perform the group schedulability test as a whole
1864  * Return 0 if success
1865  */
1866 static int power_pmu_commit_txn(struct pmu *pmu)
1867 {
1868 	struct cpu_hw_events *cpuhw;
1869 	long i, n;
1870 
1871 	if (!ppmu)
1872 		return -EAGAIN;
1873 
1874 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1875 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1876 
1877 	if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
1878 		cpuhw->txn_flags = 0;
1879 		return 0;
1880 	}
1881 
1882 	n = cpuhw->n_events;
1883 	if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1884 		return -EAGAIN;
1885 	i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n, cpuhw->event);
1886 	if (i < 0)
1887 		return -EAGAIN;
1888 
1889 	for (i = cpuhw->n_txn_start; i < n; ++i)
1890 		cpuhw->event[i]->hw.config = cpuhw->events[i];
1891 
1892 	cpuhw->txn_flags = 0;
1893 	perf_pmu_enable(pmu);
1894 	return 0;
1895 }
1896 
1897 /*
1898  * Return 1 if we might be able to put event on a limited PMC,
1899  * or 0 if not.
1900  * An event can only go on a limited PMC if it counts something
1901  * that a limited PMC can count, doesn't require interrupts, and
1902  * doesn't exclude any processor mode.
1903  */
1904 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1905 				 unsigned int flags)
1906 {
1907 	int n;
1908 	u64 alt[MAX_EVENT_ALTERNATIVES];
1909 
1910 	if (event->attr.exclude_user
1911 	    || event->attr.exclude_kernel
1912 	    || event->attr.exclude_hv
1913 	    || event->attr.sample_period)
1914 		return 0;
1915 
1916 	if (ppmu->limited_pmc_event(ev))
1917 		return 1;
1918 
1919 	/*
1920 	 * The requested event_id isn't on a limited PMC already;
1921 	 * see if any alternative code goes on a limited PMC.
1922 	 */
1923 	if (!ppmu->get_alternatives)
1924 		return 0;
1925 
1926 	flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1927 	n = ppmu->get_alternatives(ev, flags, alt);
1928 
1929 	return n > 0;
1930 }
1931 
1932 /*
1933  * Find an alternative event_id that goes on a normal PMC, if possible,
1934  * and return the event_id code, or 0 if there is no such alternative.
1935  * (Note: event_id code 0 is "don't count" on all machines.)
1936  */
1937 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1938 {
1939 	u64 alt[MAX_EVENT_ALTERNATIVES];
1940 	int n;
1941 
1942 	flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1943 	n = ppmu->get_alternatives(ev, flags, alt);
1944 	if (!n)
1945 		return 0;
1946 	return alt[0];
1947 }
1948 
1949 /* Number of perf_events counting hardware events */
1950 static atomic_t num_events;
1951 /* Used to avoid races in calling reserve/release_pmc_hardware */
1952 static DEFINE_MUTEX(pmc_reserve_mutex);
1953 
1954 /*
1955  * Release the PMU if this is the last perf_event.
1956  */
1957 static void hw_perf_event_destroy(struct perf_event *event)
1958 {
1959 	if (!atomic_add_unless(&num_events, -1, 1)) {
1960 		mutex_lock(&pmc_reserve_mutex);
1961 		if (atomic_dec_return(&num_events) == 0)
1962 			release_pmc_hardware();
1963 		mutex_unlock(&pmc_reserve_mutex);
1964 	}
1965 }
1966 
1967 /*
1968  * Translate a generic cache event_id config to a raw event_id code.
1969  */
1970 static int hw_perf_cache_event(u64 config, u64 *eventp)
1971 {
1972 	unsigned long type, op, result;
1973 	u64 ev;
1974 
1975 	if (!ppmu->cache_events)
1976 		return -EINVAL;
1977 
1978 	/* unpack config */
1979 	type = config & 0xff;
1980 	op = (config >> 8) & 0xff;
1981 	result = (config >> 16) & 0xff;
1982 
1983 	if (type >= PERF_COUNT_HW_CACHE_MAX ||
1984 	    op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1985 	    result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1986 		return -EINVAL;
1987 
1988 	ev = (*ppmu->cache_events)[type][op][result];
1989 	if (ev == 0)
1990 		return -EOPNOTSUPP;
1991 	if (ev == -1)
1992 		return -EINVAL;
1993 	*eventp = ev;
1994 	return 0;
1995 }
1996 
1997 static bool is_event_blacklisted(u64 ev)
1998 {
1999 	int i;
2000 
2001 	for (i=0; i < ppmu->n_blacklist_ev; i++) {
2002 		if (ppmu->blacklist_ev[i] == ev)
2003 			return true;
2004 	}
2005 
2006 	return false;
2007 }
2008 
2009 static int power_pmu_event_init(struct perf_event *event)
2010 {
2011 	u64 ev;
2012 	unsigned long flags, irq_flags;
2013 	struct perf_event *ctrs[MAX_HWEVENTS];
2014 	u64 events[MAX_HWEVENTS];
2015 	unsigned int cflags[MAX_HWEVENTS];
2016 	int n;
2017 	int err;
2018 	struct cpu_hw_events *cpuhw;
2019 
2020 	if (!ppmu)
2021 		return -ENOENT;
2022 
2023 	if (has_branch_stack(event)) {
2024 	        /* PMU has BHRB enabled */
2025 		if (!(ppmu->flags & PPMU_ARCH_207S))
2026 			return -EOPNOTSUPP;
2027 	}
2028 
2029 	switch (event->attr.type) {
2030 	case PERF_TYPE_HARDWARE:
2031 		ev = event->attr.config;
2032 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
2033 			return -EOPNOTSUPP;
2034 
2035 		if (ppmu->blacklist_ev && is_event_blacklisted(ev))
2036 			return -EINVAL;
2037 		ev = ppmu->generic_events[ev];
2038 		break;
2039 	case PERF_TYPE_HW_CACHE:
2040 		err = hw_perf_cache_event(event->attr.config, &ev);
2041 		if (err)
2042 			return err;
2043 
2044 		if (ppmu->blacklist_ev && is_event_blacklisted(ev))
2045 			return -EINVAL;
2046 		break;
2047 	case PERF_TYPE_RAW:
2048 		ev = event->attr.config;
2049 
2050 		if (ppmu->blacklist_ev && is_event_blacklisted(ev))
2051 			return -EINVAL;
2052 		break;
2053 	default:
2054 		return -ENOENT;
2055 	}
2056 
2057 	/*
2058 	 * PMU config registers have fields that are
2059 	 * reserved and some specific values for bit fields are reserved.
2060 	 * For ex., MMCRA[61:62] is Random Sampling Mode (SM)
2061 	 * and value of 0b11 to this field is reserved.
2062 	 * Check for invalid values in attr.config.
2063 	 */
2064 	if (ppmu->check_attr_config &&
2065 	    ppmu->check_attr_config(event))
2066 		return -EINVAL;
2067 
2068 	event->hw.config_base = ev;
2069 	event->hw.idx = 0;
2070 
2071 	/*
2072 	 * If we are not running on a hypervisor, force the
2073 	 * exclude_hv bit to 0 so that we don't care what
2074 	 * the user set it to.
2075 	 */
2076 	if (!firmware_has_feature(FW_FEATURE_LPAR))
2077 		event->attr.exclude_hv = 0;
2078 
2079 	/*
2080 	 * If this is a per-task event, then we can use
2081 	 * PM_RUN_* events interchangeably with their non RUN_*
2082 	 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
2083 	 * XXX we should check if the task is an idle task.
2084 	 */
2085 	flags = 0;
2086 	if (event->attach_state & PERF_ATTACH_TASK)
2087 		flags |= PPMU_ONLY_COUNT_RUN;
2088 
2089 	/*
2090 	 * If this machine has limited events, check whether this
2091 	 * event_id could go on a limited event.
2092 	 */
2093 	if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
2094 		if (can_go_on_limited_pmc(event, ev, flags)) {
2095 			flags |= PPMU_LIMITED_PMC_OK;
2096 		} else if (ppmu->limited_pmc_event(ev)) {
2097 			/*
2098 			 * The requested event_id is on a limited PMC,
2099 			 * but we can't use a limited PMC; see if any
2100 			 * alternative goes on a normal PMC.
2101 			 */
2102 			ev = normal_pmc_alternative(ev, flags);
2103 			if (!ev)
2104 				return -EINVAL;
2105 		}
2106 	}
2107 
2108 	/* Extra checks for EBB */
2109 	err = ebb_event_check(event);
2110 	if (err)
2111 		return err;
2112 
2113 	/*
2114 	 * If this is in a group, check if it can go on with all the
2115 	 * other hardware events in the group.  We assume the event
2116 	 * hasn't been linked into its leader's sibling list at this point.
2117 	 */
2118 	n = 0;
2119 	if (event->group_leader != event) {
2120 		n = collect_events(event->group_leader, ppmu->n_counter - 1,
2121 				   ctrs, events, cflags);
2122 		if (n < 0)
2123 			return -EINVAL;
2124 	}
2125 	events[n] = ev;
2126 	ctrs[n] = event;
2127 	cflags[n] = flags;
2128 	if (check_excludes(ctrs, cflags, n, 1))
2129 		return -EINVAL;
2130 
2131 	local_irq_save(irq_flags);
2132 	cpuhw = this_cpu_ptr(&cpu_hw_events);
2133 
2134 	err = power_check_constraints(cpuhw, events, cflags, n + 1, ctrs);
2135 
2136 	if (has_branch_stack(event)) {
2137 		u64 bhrb_filter = -1;
2138 
2139 		if (ppmu->bhrb_filter_map)
2140 			bhrb_filter = ppmu->bhrb_filter_map(
2141 					event->attr.branch_sample_type);
2142 
2143 		if (bhrb_filter == -1) {
2144 			local_irq_restore(irq_flags);
2145 			return -EOPNOTSUPP;
2146 		}
2147 		cpuhw->bhrb_filter = bhrb_filter;
2148 	}
2149 
2150 	local_irq_restore(irq_flags);
2151 	if (err)
2152 		return -EINVAL;
2153 
2154 	event->hw.config = events[n];
2155 	event->hw.event_base = cflags[n];
2156 	event->hw.last_period = event->hw.sample_period;
2157 	local64_set(&event->hw.period_left, event->hw.last_period);
2158 
2159 	/*
2160 	 * For EBB events we just context switch the PMC value, we don't do any
2161 	 * of the sample_period logic. We use hw.prev_count for this.
2162 	 */
2163 	if (is_ebb_event(event))
2164 		local64_set(&event->hw.prev_count, 0);
2165 
2166 	/*
2167 	 * See if we need to reserve the PMU.
2168 	 * If no events are currently in use, then we have to take a
2169 	 * mutex to ensure that we don't race with another task doing
2170 	 * reserve_pmc_hardware or release_pmc_hardware.
2171 	 */
2172 	err = 0;
2173 	if (!atomic_inc_not_zero(&num_events)) {
2174 		mutex_lock(&pmc_reserve_mutex);
2175 		if (atomic_read(&num_events) == 0 &&
2176 		    reserve_pmc_hardware(perf_event_interrupt))
2177 			err = -EBUSY;
2178 		else
2179 			atomic_inc(&num_events);
2180 		mutex_unlock(&pmc_reserve_mutex);
2181 	}
2182 	event->destroy = hw_perf_event_destroy;
2183 
2184 	return err;
2185 }
2186 
2187 static int power_pmu_event_idx(struct perf_event *event)
2188 {
2189 	return event->hw.idx;
2190 }
2191 
2192 ssize_t power_events_sysfs_show(struct device *dev,
2193 				struct device_attribute *attr, char *page)
2194 {
2195 	struct perf_pmu_events_attr *pmu_attr;
2196 
2197 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
2198 
2199 	return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
2200 }
2201 
2202 static struct pmu power_pmu = {
2203 	.pmu_enable	= power_pmu_enable,
2204 	.pmu_disable	= power_pmu_disable,
2205 	.event_init	= power_pmu_event_init,
2206 	.add		= power_pmu_add,
2207 	.del		= power_pmu_del,
2208 	.start		= power_pmu_start,
2209 	.stop		= power_pmu_stop,
2210 	.read		= power_pmu_read,
2211 	.start_txn	= power_pmu_start_txn,
2212 	.cancel_txn	= power_pmu_cancel_txn,
2213 	.commit_txn	= power_pmu_commit_txn,
2214 	.event_idx	= power_pmu_event_idx,
2215 	.sched_task	= power_pmu_sched_task,
2216 };
2217 
2218 #define PERF_SAMPLE_ADDR_TYPE  (PERF_SAMPLE_ADDR |		\
2219 				PERF_SAMPLE_PHYS_ADDR |		\
2220 				PERF_SAMPLE_DATA_PAGE_SIZE)
2221 /*
2222  * A counter has overflowed; update its count and record
2223  * things if requested.  Note that interrupts are hard-disabled
2224  * here so there is no possibility of being interrupted.
2225  */
2226 static void record_and_restart(struct perf_event *event, unsigned long val,
2227 			       struct pt_regs *regs)
2228 {
2229 	u64 period = event->hw.sample_period;
2230 	s64 prev, delta, left;
2231 	int record = 0;
2232 
2233 	if (event->hw.state & PERF_HES_STOPPED) {
2234 		write_pmc(event->hw.idx, 0);
2235 		return;
2236 	}
2237 
2238 	/* we don't have to worry about interrupts here */
2239 	prev = local64_read(&event->hw.prev_count);
2240 	delta = check_and_compute_delta(prev, val);
2241 	local64_add(delta, &event->count);
2242 
2243 	/*
2244 	 * See if the total period for this event has expired,
2245 	 * and update for the next period.
2246 	 */
2247 	val = 0;
2248 	left = local64_read(&event->hw.period_left) - delta;
2249 	if (delta == 0)
2250 		left++;
2251 	if (period) {
2252 		if (left <= 0) {
2253 			left += period;
2254 			if (left <= 0)
2255 				left = period;
2256 
2257 			/*
2258 			 * If address is not requested in the sample via
2259 			 * PERF_SAMPLE_IP, just record that sample irrespective
2260 			 * of SIAR valid check.
2261 			 */
2262 			if (event->attr.sample_type & PERF_SAMPLE_IP)
2263 				record = siar_valid(regs);
2264 			else
2265 				record = 1;
2266 
2267 			event->hw.last_period = event->hw.sample_period;
2268 		}
2269 		if (left < 0x80000000LL)
2270 			val = 0x80000000LL - left;
2271 	}
2272 
2273 	write_pmc(event->hw.idx, val);
2274 	local64_set(&event->hw.prev_count, val);
2275 	local64_set(&event->hw.period_left, left);
2276 	perf_event_update_userpage(event);
2277 
2278 	/*
2279 	 * Due to hardware limitation, sometimes SIAR could sample a kernel
2280 	 * address even when freeze on supervisor state (kernel) is set in
2281 	 * MMCR2. Check attr.exclude_kernel and address to drop the sample in
2282 	 * these cases.
2283 	 */
2284 	if (event->attr.exclude_kernel &&
2285 	    (event->attr.sample_type & PERF_SAMPLE_IP) &&
2286 	    is_kernel_addr(mfspr(SPRN_SIAR)))
2287 		record = 0;
2288 
2289 	/*
2290 	 * Finally record data if requested.
2291 	 */
2292 	if (record) {
2293 		struct perf_sample_data data;
2294 
2295 		perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
2296 
2297 		if (event->attr.sample_type & PERF_SAMPLE_ADDR_TYPE)
2298 			perf_get_data_addr(event, regs, &data.addr);
2299 
2300 		if (event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK) {
2301 			struct cpu_hw_events *cpuhw;
2302 			cpuhw = this_cpu_ptr(&cpu_hw_events);
2303 			power_pmu_bhrb_read(event, cpuhw);
2304 			data.br_stack = &cpuhw->bhrb_stack;
2305 		}
2306 
2307 		if (event->attr.sample_type & PERF_SAMPLE_DATA_SRC &&
2308 						ppmu->get_mem_data_src)
2309 			ppmu->get_mem_data_src(&data.data_src, ppmu->flags, regs);
2310 
2311 		if (event->attr.sample_type & PERF_SAMPLE_WEIGHT_TYPE &&
2312 						ppmu->get_mem_weight)
2313 			ppmu->get_mem_weight(&data.weight.full, event->attr.sample_type);
2314 
2315 		if (perf_event_overflow(event, &data, regs))
2316 			power_pmu_stop(event, 0);
2317 	} else if (period) {
2318 		/* Account for interrupt in case of invalid SIAR */
2319 		if (perf_event_account_interrupt(event))
2320 			power_pmu_stop(event, 0);
2321 	}
2322 }
2323 
2324 /*
2325  * Called from generic code to get the misc flags (i.e. processor mode)
2326  * for an event_id.
2327  */
2328 unsigned long perf_misc_flags(struct pt_regs *regs)
2329 {
2330 	u32 flags = perf_get_misc_flags(regs);
2331 
2332 	if (flags)
2333 		return flags;
2334 	return user_mode(regs) ? PERF_RECORD_MISC_USER :
2335 		PERF_RECORD_MISC_KERNEL;
2336 }
2337 
2338 /*
2339  * Called from generic code to get the instruction pointer
2340  * for an event_id.
2341  */
2342 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2343 {
2344 	unsigned long siar = mfspr(SPRN_SIAR);
2345 
2346 	if (regs_use_siar(regs) && siar_valid(regs) && siar)
2347 		return siar + perf_ip_adjust(regs);
2348 	else
2349 		return regs->nip;
2350 }
2351 
2352 static bool pmc_overflow_power7(unsigned long val)
2353 {
2354 	/*
2355 	 * Events on POWER7 can roll back if a speculative event doesn't
2356 	 * eventually complete. Unfortunately in some rare cases they will
2357 	 * raise a performance monitor exception. We need to catch this to
2358 	 * ensure we reset the PMC. In all cases the PMC will be 256 or less
2359 	 * cycles from overflow.
2360 	 *
2361 	 * We only do this if the first pass fails to find any overflowing
2362 	 * PMCs because a user might set a period of less than 256 and we
2363 	 * don't want to mistakenly reset them.
2364 	 */
2365 	if ((0x80000000 - val) <= 256)
2366 		return true;
2367 
2368 	return false;
2369 }
2370 
2371 static bool pmc_overflow(unsigned long val)
2372 {
2373 	if ((int)val < 0)
2374 		return true;
2375 
2376 	return false;
2377 }
2378 
2379 /*
2380  * Performance monitor interrupt stuff
2381  */
2382 static void __perf_event_interrupt(struct pt_regs *regs)
2383 {
2384 	int i, j;
2385 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
2386 	struct perf_event *event;
2387 	int found, active;
2388 
2389 	if (cpuhw->n_limited)
2390 		freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
2391 					mfspr(SPRN_PMC6));
2392 
2393 	perf_read_regs(regs);
2394 
2395 	/* Read all the PMCs since we'll need them a bunch of times */
2396 	for (i = 0; i < ppmu->n_counter; ++i)
2397 		cpuhw->pmcs[i] = read_pmc(i + 1);
2398 
2399 	/* Try to find what caused the IRQ */
2400 	found = 0;
2401 	for (i = 0; i < ppmu->n_counter; ++i) {
2402 		if (!pmc_overflow(cpuhw->pmcs[i]))
2403 			continue;
2404 		if (is_limited_pmc(i + 1))
2405 			continue; /* these won't generate IRQs */
2406 		/*
2407 		 * We've found one that's overflowed.  For active
2408 		 * counters we need to log this.  For inactive
2409 		 * counters, we need to reset it anyway
2410 		 */
2411 		found = 1;
2412 		active = 0;
2413 		for (j = 0; j < cpuhw->n_events; ++j) {
2414 			event = cpuhw->event[j];
2415 			if (event->hw.idx == (i + 1)) {
2416 				active = 1;
2417 				record_and_restart(event, cpuhw->pmcs[i], regs);
2418 				break;
2419 			}
2420 		}
2421 
2422 		/*
2423 		 * Clear PACA_IRQ_PMI in case it was set by
2424 		 * set_pmi_irq_pending() when PMU was enabled
2425 		 * after accounting for interrupts.
2426 		 */
2427 		clear_pmi_irq_pending();
2428 
2429 		if (!active)
2430 			/* reset non active counters that have overflowed */
2431 			write_pmc(i + 1, 0);
2432 	}
2433 	if (!found && pvr_version_is(PVR_POWER7)) {
2434 		/* check active counters for special buggy p7 overflow */
2435 		for (i = 0; i < cpuhw->n_events; ++i) {
2436 			event = cpuhw->event[i];
2437 			if (!event->hw.idx || is_limited_pmc(event->hw.idx))
2438 				continue;
2439 			if (pmc_overflow_power7(cpuhw->pmcs[event->hw.idx - 1])) {
2440 				/* event has overflowed in a buggy way*/
2441 				found = 1;
2442 				record_and_restart(event,
2443 						   cpuhw->pmcs[event->hw.idx - 1],
2444 						   regs);
2445 			}
2446 		}
2447 	}
2448 
2449 	/*
2450 	 * During system wide profiling or while specific CPU is monitored for an
2451 	 * event, some corner cases could cause PMC to overflow in idle path. This
2452 	 * will trigger a PMI after waking up from idle. Since counter values are _not_
2453 	 * saved/restored in idle path, can lead to below "Can't find PMC" message.
2454 	 */
2455 	if (unlikely(!found) && !arch_irq_disabled_regs(regs))
2456 		printk_ratelimited(KERN_WARNING "Can't find PMC that caused IRQ\n");
2457 
2458 	/*
2459 	 * Reset MMCR0 to its normal value.  This will set PMXE and
2460 	 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
2461 	 * and thus allow interrupts to occur again.
2462 	 * XXX might want to use MSR.PM to keep the events frozen until
2463 	 * we get back out of this interrupt.
2464 	 */
2465 	write_mmcr0(cpuhw, cpuhw->mmcr.mmcr0);
2466 
2467 	/* Clear the cpuhw->pmcs */
2468 	memset(&cpuhw->pmcs, 0, sizeof(cpuhw->pmcs));
2469 
2470 }
2471 
2472 static void perf_event_interrupt(struct pt_regs *regs)
2473 {
2474 	u64 start_clock = sched_clock();
2475 
2476 	__perf_event_interrupt(regs);
2477 	perf_sample_event_took(sched_clock() - start_clock);
2478 }
2479 
2480 static int power_pmu_prepare_cpu(unsigned int cpu)
2481 {
2482 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
2483 
2484 	if (ppmu) {
2485 		memset(cpuhw, 0, sizeof(*cpuhw));
2486 		cpuhw->mmcr.mmcr0 = MMCR0_FC;
2487 	}
2488 	return 0;
2489 }
2490 
2491 int __init register_power_pmu(struct power_pmu *pmu)
2492 {
2493 	if (ppmu)
2494 		return -EBUSY;		/* something's already registered */
2495 
2496 	ppmu = pmu;
2497 	pr_info("%s performance monitor hardware support registered\n",
2498 		pmu->name);
2499 
2500 	power_pmu.attr_groups = ppmu->attr_groups;
2501 	power_pmu.capabilities |= (ppmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS);
2502 
2503 #ifdef MSR_HV
2504 	/*
2505 	 * Use FCHV to ignore kernel events if MSR.HV is set.
2506 	 */
2507 	if (mfmsr() & MSR_HV)
2508 		freeze_events_kernel = MMCR0_FCHV;
2509 #endif /* CONFIG_PPC64 */
2510 
2511 	perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
2512 	cpuhp_setup_state(CPUHP_PERF_POWER, "perf/powerpc:prepare",
2513 			  power_pmu_prepare_cpu, NULL);
2514 	return 0;
2515 }
2516 
2517 #ifdef CONFIG_PPC64
2518 static bool pmu_override = false;
2519 static unsigned long pmu_override_val;
2520 static void do_pmu_override(void *data)
2521 {
2522 	ppc_set_pmu_inuse(1);
2523 	if (pmu_override_val)
2524 		mtspr(SPRN_MMCR1, pmu_override_val);
2525 	mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_FC);
2526 }
2527 
2528 static int __init init_ppc64_pmu(void)
2529 {
2530 	if (cpu_has_feature(CPU_FTR_HVMODE) && pmu_override) {
2531 		pr_warn("disabling perf due to pmu_override= command line option.\n");
2532 		on_each_cpu(do_pmu_override, NULL, 1);
2533 		return 0;
2534 	}
2535 
2536 	/* run through all the pmu drivers one at a time */
2537 	if (!init_power5_pmu())
2538 		return 0;
2539 	else if (!init_power5p_pmu())
2540 		return 0;
2541 	else if (!init_power6_pmu())
2542 		return 0;
2543 	else if (!init_power7_pmu())
2544 		return 0;
2545 	else if (!init_power8_pmu())
2546 		return 0;
2547 	else if (!init_power9_pmu())
2548 		return 0;
2549 	else if (!init_power10_pmu())
2550 		return 0;
2551 	else if (!init_ppc970_pmu())
2552 		return 0;
2553 	else
2554 		return init_generic_compat_pmu();
2555 }
2556 early_initcall(init_ppc64_pmu);
2557 
2558 static int __init pmu_setup(char *str)
2559 {
2560 	unsigned long val;
2561 
2562 	if (!early_cpu_has_feature(CPU_FTR_HVMODE))
2563 		return 0;
2564 
2565 	pmu_override = true;
2566 
2567 	if (kstrtoul(str, 0, &val))
2568 		val = 0;
2569 
2570 	pmu_override_val = val;
2571 
2572 	return 1;
2573 }
2574 __setup("pmu_override=", pmu_setup);
2575 
2576 #endif
2577