xref: /linux/arch/powerpc/perf/core-book3s.c (revision 7cefa5a05dbda1f0bbbd98e9d2861b09a35cc6ea)
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 <linux/uaccess.h>
17 #include <asm/reg.h>
18 #include <asm/pmc.h>
19 #include <asm/machdep.h>
20 #include <asm/firmware.h>
21 #include <asm/ptrace.h>
22 #include <asm/code-patching.h>
23 
24 #define BHRB_MAX_ENTRIES	32
25 #define BHRB_TARGET		0x0000000000000002
26 #define BHRB_PREDICTION		0x0000000000000001
27 #define BHRB_EA			0xFFFFFFFFFFFFFFFCUL
28 
29 struct cpu_hw_events {
30 	int n_events;
31 	int n_percpu;
32 	int disabled;
33 	int n_added;
34 	int n_limited;
35 	u8  pmcs_enabled;
36 	struct perf_event *event[MAX_HWEVENTS];
37 	u64 events[MAX_HWEVENTS];
38 	unsigned int flags[MAX_HWEVENTS];
39 	/*
40 	 * The order of the MMCR array is:
41 	 *  - 64-bit, MMCR0, MMCR1, MMCRA, MMCR2
42 	 *  - 32-bit, MMCR0, MMCR1, MMCR2
43 	 */
44 	unsigned long mmcr[4];
45 	struct perf_event *limited_counter[MAX_LIMITED_HWCOUNTERS];
46 	u8  limited_hwidx[MAX_LIMITED_HWCOUNTERS];
47 	u64 alternatives[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
48 	unsigned long amasks[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
49 	unsigned long avalues[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
50 
51 	unsigned int txn_flags;
52 	int n_txn_start;
53 
54 	/* BHRB bits */
55 	u64				bhrb_filter;	/* BHRB HW branch filter */
56 	unsigned int			bhrb_users;
57 	void				*bhrb_context;
58 	struct	perf_branch_stack	bhrb_stack;
59 	struct	perf_branch_entry	bhrb_entries[BHRB_MAX_ENTRIES];
60 	u64				ic_init;
61 };
62 
63 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
64 
65 static struct power_pmu *ppmu;
66 
67 /*
68  * Normally, to ignore kernel events we set the FCS (freeze counters
69  * in supervisor mode) bit in MMCR0, but if the kernel runs with the
70  * hypervisor bit set in the MSR, or if we are running on a processor
71  * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
72  * then we need to use the FCHV bit to ignore kernel events.
73  */
74 static unsigned int freeze_events_kernel = MMCR0_FCS;
75 
76 /*
77  * 32-bit doesn't have MMCRA but does have an MMCR2,
78  * and a few other names are different.
79  */
80 #ifdef CONFIG_PPC32
81 
82 #define MMCR0_FCHV		0
83 #define MMCR0_PMCjCE		MMCR0_PMCnCE
84 #define MMCR0_FC56		0
85 #define MMCR0_PMAO		0
86 #define MMCR0_EBE		0
87 #define MMCR0_BHRBA		0
88 #define MMCR0_PMCC		0
89 #define MMCR0_PMCC_U6		0
90 
91 #define SPRN_MMCRA		SPRN_MMCR2
92 #define MMCRA_SAMPLE_ENABLE	0
93 
94 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
95 {
96 	return 0;
97 }
98 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
99 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
100 {
101 	return 0;
102 }
103 static inline void perf_read_regs(struct pt_regs *regs)
104 {
105 	regs->result = 0;
106 }
107 static inline int perf_intr_is_nmi(struct pt_regs *regs)
108 {
109 	return 0;
110 }
111 
112 static inline int siar_valid(struct pt_regs *regs)
113 {
114 	return 1;
115 }
116 
117 static bool is_ebb_event(struct perf_event *event) { return false; }
118 static int ebb_event_check(struct perf_event *event) { return 0; }
119 static void ebb_event_add(struct perf_event *event) { }
120 static void ebb_switch_out(unsigned long mmcr0) { }
121 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
122 {
123 	return cpuhw->mmcr[0];
124 }
125 
126 static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
127 static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
128 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in) {}
129 static inline void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw) {}
130 static void pmao_restore_workaround(bool ebb) { }
131 static bool use_ic(u64 event)
132 {
133 	return false;
134 }
135 #endif /* CONFIG_PPC32 */
136 
137 static bool regs_use_siar(struct pt_regs *regs)
138 {
139 	/*
140 	 * When we take a performance monitor exception the regs are setup
141 	 * using perf_read_regs() which overloads some fields, in particular
142 	 * regs->result to tell us whether to use SIAR.
143 	 *
144 	 * However if the regs are from another exception, eg. a syscall, then
145 	 * they have not been setup using perf_read_regs() and so regs->result
146 	 * is something random.
147 	 */
148 	return ((TRAP(regs) == 0xf00) && regs->result);
149 }
150 
151 /*
152  * Things that are specific to 64-bit implementations.
153  */
154 #ifdef CONFIG_PPC64
155 
156 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
157 {
158 	unsigned long mmcra = regs->dsisr;
159 
160 	if ((ppmu->flags & PPMU_HAS_SSLOT) && (mmcra & MMCRA_SAMPLE_ENABLE)) {
161 		unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
162 		if (slot > 1)
163 			return 4 * (slot - 1);
164 	}
165 
166 	return 0;
167 }
168 
169 /*
170  * The user wants a data address recorded.
171  * If we're not doing instruction sampling, give them the SDAR
172  * (sampled data address).  If we are doing instruction sampling, then
173  * only give them the SDAR if it corresponds to the instruction
174  * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC, the
175  * [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA, or the SDAR_VALID bit in SIER.
176  */
177 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
178 {
179 	unsigned long mmcra = regs->dsisr;
180 	bool sdar_valid;
181 
182 	if (ppmu->flags & PPMU_HAS_SIER)
183 		sdar_valid = regs->dar & SIER_SDAR_VALID;
184 	else {
185 		unsigned long sdsync;
186 
187 		if (ppmu->flags & PPMU_SIAR_VALID)
188 			sdsync = POWER7P_MMCRA_SDAR_VALID;
189 		else if (ppmu->flags & PPMU_ALT_SIPR)
190 			sdsync = POWER6_MMCRA_SDSYNC;
191 		else
192 			sdsync = MMCRA_SDSYNC;
193 
194 		sdar_valid = mmcra & sdsync;
195 	}
196 
197 	if (!(mmcra & MMCRA_SAMPLE_ENABLE) || sdar_valid)
198 		*addrp = mfspr(SPRN_SDAR);
199 }
200 
201 static bool regs_sihv(struct pt_regs *regs)
202 {
203 	unsigned long sihv = MMCRA_SIHV;
204 
205 	if (ppmu->flags & PPMU_HAS_SIER)
206 		return !!(regs->dar & SIER_SIHV);
207 
208 	if (ppmu->flags & PPMU_ALT_SIPR)
209 		sihv = POWER6_MMCRA_SIHV;
210 
211 	return !!(regs->dsisr & sihv);
212 }
213 
214 static bool regs_sipr(struct pt_regs *regs)
215 {
216 	unsigned long sipr = MMCRA_SIPR;
217 
218 	if (ppmu->flags & PPMU_HAS_SIER)
219 		return !!(regs->dar & SIER_SIPR);
220 
221 	if (ppmu->flags & PPMU_ALT_SIPR)
222 		sipr = POWER6_MMCRA_SIPR;
223 
224 	return !!(regs->dsisr & sipr);
225 }
226 
227 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
228 {
229 	if (regs->msr & MSR_PR)
230 		return PERF_RECORD_MISC_USER;
231 	if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
232 		return PERF_RECORD_MISC_HYPERVISOR;
233 	return PERF_RECORD_MISC_KERNEL;
234 }
235 
236 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
237 {
238 	bool use_siar = regs_use_siar(regs);
239 
240 	if (!use_siar)
241 		return perf_flags_from_msr(regs);
242 
243 	/*
244 	 * If we don't have flags in MMCRA, rather than using
245 	 * the MSR, we intuit the flags from the address in
246 	 * SIAR which should give slightly more reliable
247 	 * results
248 	 */
249 	if (ppmu->flags & PPMU_NO_SIPR) {
250 		unsigned long siar = mfspr(SPRN_SIAR);
251 		if (is_kernel_addr(siar))
252 			return PERF_RECORD_MISC_KERNEL;
253 		return PERF_RECORD_MISC_USER;
254 	}
255 
256 	/* PR has priority over HV, so order below is important */
257 	if (regs_sipr(regs))
258 		return PERF_RECORD_MISC_USER;
259 
260 	if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
261 		return PERF_RECORD_MISC_HYPERVISOR;
262 
263 	return PERF_RECORD_MISC_KERNEL;
264 }
265 
266 /*
267  * Overload regs->dsisr to store MMCRA so we only need to read it once
268  * on each interrupt.
269  * Overload regs->dar to store SIER if we have it.
270  * Overload regs->result to specify whether we should use the MSR (result
271  * is zero) or the SIAR (result is non zero).
272  */
273 static inline void perf_read_regs(struct pt_regs *regs)
274 {
275 	unsigned long mmcra = mfspr(SPRN_MMCRA);
276 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
277 	int use_siar;
278 
279 	regs->dsisr = mmcra;
280 
281 	if (ppmu->flags & PPMU_HAS_SIER)
282 		regs->dar = mfspr(SPRN_SIER);
283 
284 	/*
285 	 * If this isn't a PMU exception (eg a software event) the SIAR is
286 	 * not valid. Use pt_regs.
287 	 *
288 	 * If it is a marked event use the SIAR.
289 	 *
290 	 * If the PMU doesn't update the SIAR for non marked events use
291 	 * pt_regs.
292 	 *
293 	 * If the PMU has HV/PR flags then check to see if they
294 	 * place the exception in userspace. If so, use pt_regs. In
295 	 * continuous sampling mode the SIAR and the PMU exception are
296 	 * not synchronised, so they may be many instructions apart.
297 	 * This can result in confusing backtraces. We still want
298 	 * hypervisor samples as well as samples in the kernel with
299 	 * interrupts off hence the userspace check.
300 	 */
301 	if (TRAP(regs) != 0xf00)
302 		use_siar = 0;
303 	else if ((ppmu->flags & PPMU_NO_SIAR))
304 		use_siar = 0;
305 	else if (marked)
306 		use_siar = 1;
307 	else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
308 		use_siar = 0;
309 	else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
310 		use_siar = 0;
311 	else
312 		use_siar = 1;
313 
314 	regs->result = use_siar;
315 }
316 
317 /*
318  * If interrupts were soft-disabled when a PMU interrupt occurs, treat
319  * it as an NMI.
320  */
321 static inline int perf_intr_is_nmi(struct pt_regs *regs)
322 {
323 	return !regs->softe;
324 }
325 
326 /*
327  * On processors like P7+ that have the SIAR-Valid bit, marked instructions
328  * must be sampled only if the SIAR-valid bit is set.
329  *
330  * For unmarked instructions and for processors that don't have the SIAR-Valid
331  * bit, assume that SIAR is valid.
332  */
333 static inline int siar_valid(struct pt_regs *regs)
334 {
335 	unsigned long mmcra = regs->dsisr;
336 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
337 
338 	if (marked) {
339 		if (ppmu->flags & PPMU_HAS_SIER)
340 			return regs->dar & SIER_SIAR_VALID;
341 
342 		if (ppmu->flags & PPMU_SIAR_VALID)
343 			return mmcra & POWER7P_MMCRA_SIAR_VALID;
344 	}
345 
346 	return 1;
347 }
348 
349 
350 /* Reset all possible BHRB entries */
351 static void power_pmu_bhrb_reset(void)
352 {
353 	asm volatile(PPC_CLRBHRB);
354 }
355 
356 static void power_pmu_bhrb_enable(struct perf_event *event)
357 {
358 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
359 
360 	if (!ppmu->bhrb_nr)
361 		return;
362 
363 	/* Clear BHRB if we changed task context to avoid data leaks */
364 	if (event->ctx->task && cpuhw->bhrb_context != event->ctx) {
365 		power_pmu_bhrb_reset();
366 		cpuhw->bhrb_context = event->ctx;
367 	}
368 	cpuhw->bhrb_users++;
369 	perf_sched_cb_inc(event->ctx->pmu);
370 }
371 
372 static void power_pmu_bhrb_disable(struct perf_event *event)
373 {
374 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
375 
376 	if (!ppmu->bhrb_nr)
377 		return;
378 
379 	WARN_ON_ONCE(!cpuhw->bhrb_users);
380 	cpuhw->bhrb_users--;
381 	perf_sched_cb_dec(event->ctx->pmu);
382 
383 	if (!cpuhw->disabled && !cpuhw->bhrb_users) {
384 		/* BHRB cannot be turned off when other
385 		 * events are active on the PMU.
386 		 */
387 
388 		/* avoid stale pointer */
389 		cpuhw->bhrb_context = NULL;
390 	}
391 }
392 
393 /* Called from ctxsw to prevent one process's branch entries to
394  * mingle with the other process's entries during context switch.
395  */
396 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
397 {
398 	if (!ppmu->bhrb_nr)
399 		return;
400 
401 	if (sched_in)
402 		power_pmu_bhrb_reset();
403 }
404 /* Calculate the to address for a branch */
405 static __u64 power_pmu_bhrb_to(u64 addr)
406 {
407 	unsigned int instr;
408 	int ret;
409 	__u64 target;
410 
411 	if (is_kernel_addr(addr))
412 		return branch_target((unsigned int *)addr);
413 
414 	/* Userspace: need copy instruction here then translate it */
415 	pagefault_disable();
416 	ret = __get_user_inatomic(instr, (unsigned int __user *)addr);
417 	if (ret) {
418 		pagefault_enable();
419 		return 0;
420 	}
421 	pagefault_enable();
422 
423 	target = branch_target(&instr);
424 	if ((!target) || (instr & BRANCH_ABSOLUTE))
425 		return target;
426 
427 	/* Translate relative branch target from kernel to user address */
428 	return target - (unsigned long)&instr + addr;
429 }
430 
431 /* Processing BHRB entries */
432 static void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw)
433 {
434 	u64 val;
435 	u64 addr;
436 	int r_index, u_index, pred;
437 
438 	r_index = 0;
439 	u_index = 0;
440 	while (r_index < ppmu->bhrb_nr) {
441 		/* Assembly read function */
442 		val = read_bhrb(r_index++);
443 		if (!val)
444 			/* Terminal marker: End of valid BHRB entries */
445 			break;
446 		else {
447 			addr = val & BHRB_EA;
448 			pred = val & BHRB_PREDICTION;
449 
450 			if (!addr)
451 				/* invalid entry */
452 				continue;
453 
454 			/* Branches are read most recent first (ie. mfbhrb 0 is
455 			 * the most recent branch).
456 			 * There are two types of valid entries:
457 			 * 1) a target entry which is the to address of a
458 			 *    computed goto like a blr,bctr,btar.  The next
459 			 *    entry read from the bhrb will be branch
460 			 *    corresponding to this target (ie. the actual
461 			 *    blr/bctr/btar instruction).
462 			 * 2) a from address which is an actual branch.  If a
463 			 *    target entry proceeds this, then this is the
464 			 *    matching branch for that target.  If this is not
465 			 *    following a target entry, then this is a branch
466 			 *    where the target is given as an immediate field
467 			 *    in the instruction (ie. an i or b form branch).
468 			 *    In this case we need to read the instruction from
469 			 *    memory to determine the target/to address.
470 			 */
471 
472 			if (val & BHRB_TARGET) {
473 				/* Target branches use two entries
474 				 * (ie. computed gotos/XL form)
475 				 */
476 				cpuhw->bhrb_entries[u_index].to = addr;
477 				cpuhw->bhrb_entries[u_index].mispred = pred;
478 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
479 
480 				/* Get from address in next entry */
481 				val = read_bhrb(r_index++);
482 				addr = val & BHRB_EA;
483 				if (val & BHRB_TARGET) {
484 					/* Shouldn't have two targets in a
485 					   row.. Reset index and try again */
486 					r_index--;
487 					addr = 0;
488 				}
489 				cpuhw->bhrb_entries[u_index].from = addr;
490 			} else {
491 				/* Branches to immediate field
492 				   (ie I or B form) */
493 				cpuhw->bhrb_entries[u_index].from = addr;
494 				cpuhw->bhrb_entries[u_index].to =
495 					power_pmu_bhrb_to(addr);
496 				cpuhw->bhrb_entries[u_index].mispred = pred;
497 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
498 			}
499 			u_index++;
500 
501 		}
502 	}
503 	cpuhw->bhrb_stack.nr = u_index;
504 	return;
505 }
506 
507 static bool is_ebb_event(struct perf_event *event)
508 {
509 	/*
510 	 * This could be a per-PMU callback, but we'd rather avoid the cost. We
511 	 * check that the PMU supports EBB, meaning those that don't can still
512 	 * use bit 63 of the event code for something else if they wish.
513 	 */
514 	return (ppmu->flags & PPMU_ARCH_207S) &&
515 	       ((event->attr.config >> PERF_EVENT_CONFIG_EBB_SHIFT) & 1);
516 }
517 
518 static int ebb_event_check(struct perf_event *event)
519 {
520 	struct perf_event *leader = event->group_leader;
521 
522 	/* Event and group leader must agree on EBB */
523 	if (is_ebb_event(leader) != is_ebb_event(event))
524 		return -EINVAL;
525 
526 	if (is_ebb_event(event)) {
527 		if (!(event->attach_state & PERF_ATTACH_TASK))
528 			return -EINVAL;
529 
530 		if (!leader->attr.pinned || !leader->attr.exclusive)
531 			return -EINVAL;
532 
533 		if (event->attr.freq ||
534 		    event->attr.inherit ||
535 		    event->attr.sample_type ||
536 		    event->attr.sample_period ||
537 		    event->attr.enable_on_exec)
538 			return -EINVAL;
539 	}
540 
541 	return 0;
542 }
543 
544 static void ebb_event_add(struct perf_event *event)
545 {
546 	if (!is_ebb_event(event) || current->thread.used_ebb)
547 		return;
548 
549 	/*
550 	 * IFF this is the first time we've added an EBB event, set
551 	 * PMXE in the user MMCR0 so we can detect when it's cleared by
552 	 * userspace. We need this so that we can context switch while
553 	 * userspace is in the EBB handler (where PMXE is 0).
554 	 */
555 	current->thread.used_ebb = 1;
556 	current->thread.mmcr0 |= MMCR0_PMXE;
557 }
558 
559 static void ebb_switch_out(unsigned long mmcr0)
560 {
561 	if (!(mmcr0 & MMCR0_EBE))
562 		return;
563 
564 	current->thread.siar  = mfspr(SPRN_SIAR);
565 	current->thread.sier  = mfspr(SPRN_SIER);
566 	current->thread.sdar  = mfspr(SPRN_SDAR);
567 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
568 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
569 }
570 
571 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
572 {
573 	unsigned long mmcr0 = cpuhw->mmcr[0];
574 
575 	if (!ebb)
576 		goto out;
577 
578 	/* Enable EBB and read/write to all 6 PMCs and BHRB for userspace */
579 	mmcr0 |= MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC_U6;
580 
581 	/*
582 	 * Add any bits from the user MMCR0, FC or PMAO. This is compatible
583 	 * with pmao_restore_workaround() because we may add PMAO but we never
584 	 * clear it here.
585 	 */
586 	mmcr0 |= current->thread.mmcr0;
587 
588 	/*
589 	 * Be careful not to set PMXE if userspace had it cleared. This is also
590 	 * compatible with pmao_restore_workaround() because it has already
591 	 * cleared PMXE and we leave PMAO alone.
592 	 */
593 	if (!(current->thread.mmcr0 & MMCR0_PMXE))
594 		mmcr0 &= ~MMCR0_PMXE;
595 
596 	mtspr(SPRN_SIAR, current->thread.siar);
597 	mtspr(SPRN_SIER, current->thread.sier);
598 	mtspr(SPRN_SDAR, current->thread.sdar);
599 
600 	/*
601 	 * Merge the kernel & user values of MMCR2. The semantics we implement
602 	 * are that the user MMCR2 can set bits, ie. cause counters to freeze,
603 	 * but not clear bits. If a task wants to be able to clear bits, ie.
604 	 * unfreeze counters, it should not set exclude_xxx in its events and
605 	 * instead manage the MMCR2 entirely by itself.
606 	 */
607 	mtspr(SPRN_MMCR2, cpuhw->mmcr[3] | current->thread.mmcr2);
608 out:
609 	return mmcr0;
610 }
611 
612 static void pmao_restore_workaround(bool ebb)
613 {
614 	unsigned pmcs[6];
615 
616 	if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
617 		return;
618 
619 	/*
620 	 * On POWER8E there is a hardware defect which affects the PMU context
621 	 * switch logic, ie. power_pmu_disable/enable().
622 	 *
623 	 * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
624 	 * by the hardware. Sometime later the actual PMU exception is
625 	 * delivered.
626 	 *
627 	 * If we context switch, or simply disable/enable, the PMU prior to the
628 	 * exception arriving, the exception will be lost when we clear PMAO.
629 	 *
630 	 * When we reenable the PMU, we will write the saved MMCR0 with PMAO
631 	 * set, and this _should_ generate an exception. However because of the
632 	 * defect no exception is generated when we write PMAO, and we get
633 	 * stuck with no counters counting but no exception delivered.
634 	 *
635 	 * The workaround is to detect this case and tweak the hardware to
636 	 * create another pending PMU exception.
637 	 *
638 	 * We do that by setting up PMC6 (cycles) for an imminent overflow and
639 	 * enabling the PMU. That causes a new exception to be generated in the
640 	 * chip, but we don't take it yet because we have interrupts hard
641 	 * disabled. We then write back the PMU state as we want it to be seen
642 	 * by the exception handler. When we reenable interrupts the exception
643 	 * handler will be called and see the correct state.
644 	 *
645 	 * The logic is the same for EBB, except that the exception is gated by
646 	 * us having interrupts hard disabled as well as the fact that we are
647 	 * not in userspace. The exception is finally delivered when we return
648 	 * to userspace.
649 	 */
650 
651 	/* Only if PMAO is set and PMAO_SYNC is clear */
652 	if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
653 		return;
654 
655 	/* If we're doing EBB, only if BESCR[GE] is set */
656 	if (ebb && !(current->thread.bescr & BESCR_GE))
657 		return;
658 
659 	/*
660 	 * We are already soft-disabled in power_pmu_enable(). We need to hard
661 	 * disable to actually prevent the PMU exception from firing.
662 	 */
663 	hard_irq_disable();
664 
665 	/*
666 	 * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
667 	 * Using read/write_pmc() in a for loop adds 12 function calls and
668 	 * almost doubles our code size.
669 	 */
670 	pmcs[0] = mfspr(SPRN_PMC1);
671 	pmcs[1] = mfspr(SPRN_PMC2);
672 	pmcs[2] = mfspr(SPRN_PMC3);
673 	pmcs[3] = mfspr(SPRN_PMC4);
674 	pmcs[4] = mfspr(SPRN_PMC5);
675 	pmcs[5] = mfspr(SPRN_PMC6);
676 
677 	/* Ensure all freeze bits are unset */
678 	mtspr(SPRN_MMCR2, 0);
679 
680 	/* Set up PMC6 to overflow in one cycle */
681 	mtspr(SPRN_PMC6, 0x7FFFFFFE);
682 
683 	/* Enable exceptions and unfreeze PMC6 */
684 	mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
685 
686 	/* Now we need to refreeze and restore the PMCs */
687 	mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
688 
689 	mtspr(SPRN_PMC1, pmcs[0]);
690 	mtspr(SPRN_PMC2, pmcs[1]);
691 	mtspr(SPRN_PMC3, pmcs[2]);
692 	mtspr(SPRN_PMC4, pmcs[3]);
693 	mtspr(SPRN_PMC5, pmcs[4]);
694 	mtspr(SPRN_PMC6, pmcs[5]);
695 }
696 
697 static bool use_ic(u64 event)
698 {
699 	if (cpu_has_feature(CPU_FTR_POWER9_DD1) &&
700 			(event == 0x200f2 || event == 0x300f2))
701 		return true;
702 
703 	return false;
704 }
705 #endif /* CONFIG_PPC64 */
706 
707 static void perf_event_interrupt(struct pt_regs *regs);
708 
709 /*
710  * Read one performance monitor counter (PMC).
711  */
712 static unsigned long read_pmc(int idx)
713 {
714 	unsigned long val;
715 
716 	switch (idx) {
717 	case 1:
718 		val = mfspr(SPRN_PMC1);
719 		break;
720 	case 2:
721 		val = mfspr(SPRN_PMC2);
722 		break;
723 	case 3:
724 		val = mfspr(SPRN_PMC3);
725 		break;
726 	case 4:
727 		val = mfspr(SPRN_PMC4);
728 		break;
729 	case 5:
730 		val = mfspr(SPRN_PMC5);
731 		break;
732 	case 6:
733 		val = mfspr(SPRN_PMC6);
734 		break;
735 #ifdef CONFIG_PPC64
736 	case 7:
737 		val = mfspr(SPRN_PMC7);
738 		break;
739 	case 8:
740 		val = mfspr(SPRN_PMC8);
741 		break;
742 #endif /* CONFIG_PPC64 */
743 	default:
744 		printk(KERN_ERR "oops trying to read PMC%d\n", idx);
745 		val = 0;
746 	}
747 	return val;
748 }
749 
750 /*
751  * Write one PMC.
752  */
753 static void write_pmc(int idx, unsigned long val)
754 {
755 	switch (idx) {
756 	case 1:
757 		mtspr(SPRN_PMC1, val);
758 		break;
759 	case 2:
760 		mtspr(SPRN_PMC2, val);
761 		break;
762 	case 3:
763 		mtspr(SPRN_PMC3, val);
764 		break;
765 	case 4:
766 		mtspr(SPRN_PMC4, val);
767 		break;
768 	case 5:
769 		mtspr(SPRN_PMC5, val);
770 		break;
771 	case 6:
772 		mtspr(SPRN_PMC6, val);
773 		break;
774 #ifdef CONFIG_PPC64
775 	case 7:
776 		mtspr(SPRN_PMC7, val);
777 		break;
778 	case 8:
779 		mtspr(SPRN_PMC8, val);
780 		break;
781 #endif /* CONFIG_PPC64 */
782 	default:
783 		printk(KERN_ERR "oops trying to write PMC%d\n", idx);
784 	}
785 }
786 
787 /* Called from sysrq_handle_showregs() */
788 void perf_event_print_debug(void)
789 {
790 	unsigned long sdar, sier, flags;
791 	u32 pmcs[MAX_HWEVENTS];
792 	int i;
793 
794 	if (!ppmu->n_counter)
795 		return;
796 
797 	local_irq_save(flags);
798 
799 	pr_info("CPU: %d PMU registers, ppmu = %s n_counters = %d",
800 		 smp_processor_id(), ppmu->name, ppmu->n_counter);
801 
802 	for (i = 0; i < ppmu->n_counter; i++)
803 		pmcs[i] = read_pmc(i + 1);
804 
805 	for (; i < MAX_HWEVENTS; i++)
806 		pmcs[i] = 0xdeadbeef;
807 
808 	pr_info("PMC1:  %08x PMC2: %08x PMC3: %08x PMC4: %08x\n",
809 		 pmcs[0], pmcs[1], pmcs[2], pmcs[3]);
810 
811 	if (ppmu->n_counter > 4)
812 		pr_info("PMC5:  %08x PMC6: %08x PMC7: %08x PMC8: %08x\n",
813 			 pmcs[4], pmcs[5], pmcs[6], pmcs[7]);
814 
815 	pr_info("MMCR0: %016lx MMCR1: %016lx MMCRA: %016lx\n",
816 		mfspr(SPRN_MMCR0), mfspr(SPRN_MMCR1), mfspr(SPRN_MMCRA));
817 
818 	sdar = sier = 0;
819 #ifdef CONFIG_PPC64
820 	sdar = mfspr(SPRN_SDAR);
821 
822 	if (ppmu->flags & PPMU_HAS_SIER)
823 		sier = mfspr(SPRN_SIER);
824 
825 	if (ppmu->flags & PPMU_ARCH_207S) {
826 		pr_info("MMCR2: %016lx EBBHR: %016lx\n",
827 			mfspr(SPRN_MMCR2), mfspr(SPRN_EBBHR));
828 		pr_info("EBBRR: %016lx BESCR: %016lx\n",
829 			mfspr(SPRN_EBBRR), mfspr(SPRN_BESCR));
830 	}
831 #endif
832 	pr_info("SIAR:  %016lx SDAR:  %016lx SIER:  %016lx\n",
833 		mfspr(SPRN_SIAR), sdar, sier);
834 
835 	local_irq_restore(flags);
836 }
837 
838 /*
839  * Check if a set of events can all go on the PMU at once.
840  * If they can't, this will look at alternative codes for the events
841  * and see if any combination of alternative codes is feasible.
842  * The feasible set is returned in event_id[].
843  */
844 static int power_check_constraints(struct cpu_hw_events *cpuhw,
845 				   u64 event_id[], unsigned int cflags[],
846 				   int n_ev)
847 {
848 	unsigned long mask, value, nv;
849 	unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
850 	int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
851 	int i, j;
852 	unsigned long addf = ppmu->add_fields;
853 	unsigned long tadd = ppmu->test_adder;
854 
855 	if (n_ev > ppmu->n_counter)
856 		return -1;
857 
858 	/* First see if the events will go on as-is */
859 	for (i = 0; i < n_ev; ++i) {
860 		if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
861 		    && !ppmu->limited_pmc_event(event_id[i])) {
862 			ppmu->get_alternatives(event_id[i], cflags[i],
863 					       cpuhw->alternatives[i]);
864 			event_id[i] = cpuhw->alternatives[i][0];
865 		}
866 		if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
867 					 &cpuhw->avalues[i][0]))
868 			return -1;
869 	}
870 	value = mask = 0;
871 	for (i = 0; i < n_ev; ++i) {
872 		nv = (value | cpuhw->avalues[i][0]) +
873 			(value & cpuhw->avalues[i][0] & addf);
874 		if ((((nv + tadd) ^ value) & mask) != 0 ||
875 		    (((nv + tadd) ^ cpuhw->avalues[i][0]) &
876 		     cpuhw->amasks[i][0]) != 0)
877 			break;
878 		value = nv;
879 		mask |= cpuhw->amasks[i][0];
880 	}
881 	if (i == n_ev)
882 		return 0;	/* all OK */
883 
884 	/* doesn't work, gather alternatives... */
885 	if (!ppmu->get_alternatives)
886 		return -1;
887 	for (i = 0; i < n_ev; ++i) {
888 		choice[i] = 0;
889 		n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
890 						  cpuhw->alternatives[i]);
891 		for (j = 1; j < n_alt[i]; ++j)
892 			ppmu->get_constraint(cpuhw->alternatives[i][j],
893 					     &cpuhw->amasks[i][j],
894 					     &cpuhw->avalues[i][j]);
895 	}
896 
897 	/* enumerate all possibilities and see if any will work */
898 	i = 0;
899 	j = -1;
900 	value = mask = nv = 0;
901 	while (i < n_ev) {
902 		if (j >= 0) {
903 			/* we're backtracking, restore context */
904 			value = svalues[i];
905 			mask = smasks[i];
906 			j = choice[i];
907 		}
908 		/*
909 		 * See if any alternative k for event_id i,
910 		 * where k > j, will satisfy the constraints.
911 		 */
912 		while (++j < n_alt[i]) {
913 			nv = (value | cpuhw->avalues[i][j]) +
914 				(value & cpuhw->avalues[i][j] & addf);
915 			if ((((nv + tadd) ^ value) & mask) == 0 &&
916 			    (((nv + tadd) ^ cpuhw->avalues[i][j])
917 			     & cpuhw->amasks[i][j]) == 0)
918 				break;
919 		}
920 		if (j >= n_alt[i]) {
921 			/*
922 			 * No feasible alternative, backtrack
923 			 * to event_id i-1 and continue enumerating its
924 			 * alternatives from where we got up to.
925 			 */
926 			if (--i < 0)
927 				return -1;
928 		} else {
929 			/*
930 			 * Found a feasible alternative for event_id i,
931 			 * remember where we got up to with this event_id,
932 			 * go on to the next event_id, and start with
933 			 * the first alternative for it.
934 			 */
935 			choice[i] = j;
936 			svalues[i] = value;
937 			smasks[i] = mask;
938 			value = nv;
939 			mask |= cpuhw->amasks[i][j];
940 			++i;
941 			j = -1;
942 		}
943 	}
944 
945 	/* OK, we have a feasible combination, tell the caller the solution */
946 	for (i = 0; i < n_ev; ++i)
947 		event_id[i] = cpuhw->alternatives[i][choice[i]];
948 	return 0;
949 }
950 
951 /*
952  * Check if newly-added events have consistent settings for
953  * exclude_{user,kernel,hv} with each other and any previously
954  * added events.
955  */
956 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
957 			  int n_prev, int n_new)
958 {
959 	int eu = 0, ek = 0, eh = 0;
960 	int i, n, first;
961 	struct perf_event *event;
962 
963 	/*
964 	 * If the PMU we're on supports per event exclude settings then we
965 	 * don't need to do any of this logic. NB. This assumes no PMU has both
966 	 * per event exclude and limited PMCs.
967 	 */
968 	if (ppmu->flags & PPMU_ARCH_207S)
969 		return 0;
970 
971 	n = n_prev + n_new;
972 	if (n <= 1)
973 		return 0;
974 
975 	first = 1;
976 	for (i = 0; i < n; ++i) {
977 		if (cflags[i] & PPMU_LIMITED_PMC_OK) {
978 			cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
979 			continue;
980 		}
981 		event = ctrs[i];
982 		if (first) {
983 			eu = event->attr.exclude_user;
984 			ek = event->attr.exclude_kernel;
985 			eh = event->attr.exclude_hv;
986 			first = 0;
987 		} else if (event->attr.exclude_user != eu ||
988 			   event->attr.exclude_kernel != ek ||
989 			   event->attr.exclude_hv != eh) {
990 			return -EAGAIN;
991 		}
992 	}
993 
994 	if (eu || ek || eh)
995 		for (i = 0; i < n; ++i)
996 			if (cflags[i] & PPMU_LIMITED_PMC_OK)
997 				cflags[i] |= PPMU_LIMITED_PMC_REQD;
998 
999 	return 0;
1000 }
1001 
1002 static u64 check_and_compute_delta(u64 prev, u64 val)
1003 {
1004 	u64 delta = (val - prev) & 0xfffffffful;
1005 
1006 	/*
1007 	 * POWER7 can roll back counter values, if the new value is smaller
1008 	 * than the previous value it will cause the delta and the counter to
1009 	 * have bogus values unless we rolled a counter over.  If a coutner is
1010 	 * rolled back, it will be smaller, but within 256, which is the maximum
1011 	 * number of events to rollback at once.  If we detect a rollback
1012 	 * return 0.  This can lead to a small lack of precision in the
1013 	 * counters.
1014 	 */
1015 	if (prev > val && (prev - val) < 256)
1016 		delta = 0;
1017 
1018 	return delta;
1019 }
1020 
1021 static void power_pmu_read(struct perf_event *event)
1022 {
1023 	s64 val, delta, prev;
1024 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1025 
1026 	if (event->hw.state & PERF_HES_STOPPED)
1027 		return;
1028 
1029 	if (!event->hw.idx)
1030 		return;
1031 
1032 	if (is_ebb_event(event)) {
1033 		val = read_pmc(event->hw.idx);
1034 		if (use_ic(event->attr.config)) {
1035 			val = mfspr(SPRN_IC);
1036 			if (val > cpuhw->ic_init)
1037 				val = val - cpuhw->ic_init;
1038 			else
1039 				val = val + (0 - cpuhw->ic_init);
1040 		}
1041 		local64_set(&event->hw.prev_count, val);
1042 		return;
1043 	}
1044 
1045 	/*
1046 	 * Performance monitor interrupts come even when interrupts
1047 	 * are soft-disabled, as long as interrupts are hard-enabled.
1048 	 * Therefore we treat them like NMIs.
1049 	 */
1050 	do {
1051 		prev = local64_read(&event->hw.prev_count);
1052 		barrier();
1053 		val = read_pmc(event->hw.idx);
1054 		if (use_ic(event->attr.config)) {
1055 			val = mfspr(SPRN_IC);
1056 			if (val > cpuhw->ic_init)
1057 				val = val - cpuhw->ic_init;
1058 			else
1059 				val = val + (0 - cpuhw->ic_init);
1060 		}
1061 		delta = check_and_compute_delta(prev, val);
1062 		if (!delta)
1063 			return;
1064 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
1065 
1066 	local64_add(delta, &event->count);
1067 
1068 	/*
1069 	 * A number of places program the PMC with (0x80000000 - period_left).
1070 	 * We never want period_left to be less than 1 because we will program
1071 	 * the PMC with a value >= 0x800000000 and an edge detected PMC will
1072 	 * roll around to 0 before taking an exception. We have seen this
1073 	 * on POWER8.
1074 	 *
1075 	 * To fix this, clamp the minimum value of period_left to 1.
1076 	 */
1077 	do {
1078 		prev = local64_read(&event->hw.period_left);
1079 		val = prev - delta;
1080 		if (val < 1)
1081 			val = 1;
1082 	} while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
1083 }
1084 
1085 /*
1086  * On some machines, PMC5 and PMC6 can't be written, don't respect
1087  * the freeze conditions, and don't generate interrupts.  This tells
1088  * us if `event' is using such a PMC.
1089  */
1090 static int is_limited_pmc(int pmcnum)
1091 {
1092 	return (ppmu->flags & PPMU_LIMITED_PMC5_6)
1093 		&& (pmcnum == 5 || pmcnum == 6);
1094 }
1095 
1096 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
1097 				    unsigned long pmc5, unsigned long pmc6)
1098 {
1099 	struct perf_event *event;
1100 	u64 val, prev, delta;
1101 	int i;
1102 
1103 	for (i = 0; i < cpuhw->n_limited; ++i) {
1104 		event = cpuhw->limited_counter[i];
1105 		if (!event->hw.idx)
1106 			continue;
1107 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1108 		prev = local64_read(&event->hw.prev_count);
1109 		event->hw.idx = 0;
1110 		delta = check_and_compute_delta(prev, val);
1111 		if (delta)
1112 			local64_add(delta, &event->count);
1113 	}
1114 }
1115 
1116 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
1117 				  unsigned long pmc5, unsigned long pmc6)
1118 {
1119 	struct perf_event *event;
1120 	u64 val, prev;
1121 	int i;
1122 
1123 	for (i = 0; i < cpuhw->n_limited; ++i) {
1124 		event = cpuhw->limited_counter[i];
1125 		event->hw.idx = cpuhw->limited_hwidx[i];
1126 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1127 		prev = local64_read(&event->hw.prev_count);
1128 		if (check_and_compute_delta(prev, val))
1129 			local64_set(&event->hw.prev_count, val);
1130 		perf_event_update_userpage(event);
1131 	}
1132 }
1133 
1134 /*
1135  * Since limited events don't respect the freeze conditions, we
1136  * have to read them immediately after freezing or unfreezing the
1137  * other events.  We try to keep the values from the limited
1138  * events as consistent as possible by keeping the delay (in
1139  * cycles and instructions) between freezing/unfreezing and reading
1140  * the limited events as small and consistent as possible.
1141  * Therefore, if any limited events are in use, we read them
1142  * both, and always in the same order, to minimize variability,
1143  * and do it inside the same asm that writes MMCR0.
1144  */
1145 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
1146 {
1147 	unsigned long pmc5, pmc6;
1148 
1149 	if (!cpuhw->n_limited) {
1150 		mtspr(SPRN_MMCR0, mmcr0);
1151 		return;
1152 	}
1153 
1154 	/*
1155 	 * Write MMCR0, then read PMC5 and PMC6 immediately.
1156 	 * To ensure we don't get a performance monitor interrupt
1157 	 * between writing MMCR0 and freezing/thawing the limited
1158 	 * events, we first write MMCR0 with the event overflow
1159 	 * interrupt enable bits turned off.
1160 	 */
1161 	asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
1162 		     : "=&r" (pmc5), "=&r" (pmc6)
1163 		     : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
1164 		       "i" (SPRN_MMCR0),
1165 		       "i" (SPRN_PMC5), "i" (SPRN_PMC6));
1166 
1167 	if (mmcr0 & MMCR0_FC)
1168 		freeze_limited_counters(cpuhw, pmc5, pmc6);
1169 	else
1170 		thaw_limited_counters(cpuhw, pmc5, pmc6);
1171 
1172 	/*
1173 	 * Write the full MMCR0 including the event overflow interrupt
1174 	 * enable bits, if necessary.
1175 	 */
1176 	if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
1177 		mtspr(SPRN_MMCR0, mmcr0);
1178 }
1179 
1180 /*
1181  * Disable all events to prevent PMU interrupts and to allow
1182  * events to be added or removed.
1183  */
1184 static void power_pmu_disable(struct pmu *pmu)
1185 {
1186 	struct cpu_hw_events *cpuhw;
1187 	unsigned long flags, mmcr0, val;
1188 
1189 	if (!ppmu)
1190 		return;
1191 	local_irq_save(flags);
1192 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1193 
1194 	if (!cpuhw->disabled) {
1195 		/*
1196 		 * Check if we ever enabled the PMU on this cpu.
1197 		 */
1198 		if (!cpuhw->pmcs_enabled) {
1199 			ppc_enable_pmcs();
1200 			cpuhw->pmcs_enabled = 1;
1201 		}
1202 
1203 		/*
1204 		 * Set the 'freeze counters' bit, clear EBE/BHRBA/PMCC/PMAO/FC56
1205 		 */
1206 		val  = mmcr0 = mfspr(SPRN_MMCR0);
1207 		val |= MMCR0_FC;
1208 		val &= ~(MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC | MMCR0_PMAO |
1209 			 MMCR0_FC56);
1210 
1211 		/*
1212 		 * The barrier is to make sure the mtspr has been
1213 		 * executed and the PMU has frozen the events etc.
1214 		 * before we return.
1215 		 */
1216 		write_mmcr0(cpuhw, val);
1217 		mb();
1218 
1219 		/*
1220 		 * Disable instruction sampling if it was enabled
1221 		 */
1222 		if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1223 			mtspr(SPRN_MMCRA,
1224 			      cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1225 			mb();
1226 		}
1227 
1228 		cpuhw->disabled = 1;
1229 		cpuhw->n_added = 0;
1230 
1231 		ebb_switch_out(mmcr0);
1232 	}
1233 
1234 	local_irq_restore(flags);
1235 }
1236 
1237 /*
1238  * Re-enable all events if disable == 0.
1239  * If we were previously disabled and events were added, then
1240  * put the new config on the PMU.
1241  */
1242 static void power_pmu_enable(struct pmu *pmu)
1243 {
1244 	struct perf_event *event;
1245 	struct cpu_hw_events *cpuhw;
1246 	unsigned long flags;
1247 	long i;
1248 	unsigned long val, mmcr0;
1249 	s64 left;
1250 	unsigned int hwc_index[MAX_HWEVENTS];
1251 	int n_lim;
1252 	int idx;
1253 	bool ebb;
1254 
1255 	if (!ppmu)
1256 		return;
1257 	local_irq_save(flags);
1258 
1259 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1260 	if (!cpuhw->disabled)
1261 		goto out;
1262 
1263 	if (cpuhw->n_events == 0) {
1264 		ppc_set_pmu_inuse(0);
1265 		goto out;
1266 	}
1267 
1268 	cpuhw->disabled = 0;
1269 
1270 	/*
1271 	 * EBB requires an exclusive group and all events must have the EBB
1272 	 * flag set, or not set, so we can just check a single event. Also we
1273 	 * know we have at least one event.
1274 	 */
1275 	ebb = is_ebb_event(cpuhw->event[0]);
1276 
1277 	/*
1278 	 * If we didn't change anything, or only removed events,
1279 	 * no need to recalculate MMCR* settings and reset the PMCs.
1280 	 * Just reenable the PMU with the current MMCR* settings
1281 	 * (possibly updated for removal of events).
1282 	 */
1283 	if (!cpuhw->n_added) {
1284 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1285 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1286 		goto out_enable;
1287 	}
1288 
1289 	/*
1290 	 * Clear all MMCR settings and recompute them for the new set of events.
1291 	 */
1292 	memset(cpuhw->mmcr, 0, sizeof(cpuhw->mmcr));
1293 
1294 	if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
1295 			       cpuhw->mmcr, cpuhw->event)) {
1296 		/* shouldn't ever get here */
1297 		printk(KERN_ERR "oops compute_mmcr failed\n");
1298 		goto out;
1299 	}
1300 
1301 	if (!(ppmu->flags & PPMU_ARCH_207S)) {
1302 		/*
1303 		 * Add in MMCR0 freeze bits corresponding to the attr.exclude_*
1304 		 * bits for the first event. We have already checked that all
1305 		 * events have the same value for these bits as the first event.
1306 		 */
1307 		event = cpuhw->event[0];
1308 		if (event->attr.exclude_user)
1309 			cpuhw->mmcr[0] |= MMCR0_FCP;
1310 		if (event->attr.exclude_kernel)
1311 			cpuhw->mmcr[0] |= freeze_events_kernel;
1312 		if (event->attr.exclude_hv)
1313 			cpuhw->mmcr[0] |= MMCR0_FCHV;
1314 	}
1315 
1316 	/*
1317 	 * Write the new configuration to MMCR* with the freeze
1318 	 * bit set and set the hardware events to their initial values.
1319 	 * Then unfreeze the events.
1320 	 */
1321 	ppc_set_pmu_inuse(1);
1322 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1323 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1324 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
1325 				| MMCR0_FC);
1326 	if (ppmu->flags & PPMU_ARCH_207S)
1327 		mtspr(SPRN_MMCR2, cpuhw->mmcr[3]);
1328 
1329 	/*
1330 	 * Read off any pre-existing events that need to move
1331 	 * to another PMC.
1332 	 */
1333 	for (i = 0; i < cpuhw->n_events; ++i) {
1334 		event = cpuhw->event[i];
1335 		if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
1336 			power_pmu_read(event);
1337 			write_pmc(event->hw.idx, 0);
1338 			event->hw.idx = 0;
1339 		}
1340 	}
1341 
1342 	/*
1343 	 * Initialize the PMCs for all the new and moved events.
1344 	 */
1345 	cpuhw->n_limited = n_lim = 0;
1346 	for (i = 0; i < cpuhw->n_events; ++i) {
1347 		event = cpuhw->event[i];
1348 		if (event->hw.idx)
1349 			continue;
1350 		idx = hwc_index[i] + 1;
1351 		if (is_limited_pmc(idx)) {
1352 			cpuhw->limited_counter[n_lim] = event;
1353 			cpuhw->limited_hwidx[n_lim] = idx;
1354 			++n_lim;
1355 			continue;
1356 		}
1357 
1358 		if (ebb)
1359 			val = local64_read(&event->hw.prev_count);
1360 		else {
1361 			val = 0;
1362 			if (event->hw.sample_period) {
1363 				left = local64_read(&event->hw.period_left);
1364 				if (left < 0x80000000L)
1365 					val = 0x80000000L - left;
1366 			}
1367 			local64_set(&event->hw.prev_count, val);
1368 		}
1369 
1370 		event->hw.idx = idx;
1371 		if (event->hw.state & PERF_HES_STOPPED)
1372 			val = 0;
1373 		write_pmc(idx, val);
1374 
1375 		perf_event_update_userpage(event);
1376 	}
1377 	cpuhw->n_limited = n_lim;
1378 	cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
1379 
1380  out_enable:
1381 	pmao_restore_workaround(ebb);
1382 
1383 	mmcr0 = ebb_switch_in(ebb, cpuhw);
1384 
1385 	mb();
1386 	if (cpuhw->bhrb_users)
1387 		ppmu->config_bhrb(cpuhw->bhrb_filter);
1388 
1389 	write_mmcr0(cpuhw, mmcr0);
1390 
1391 	/*
1392 	 * Enable instruction sampling if necessary
1393 	 */
1394 	if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1395 		mb();
1396 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
1397 	}
1398 
1399  out:
1400 
1401 	local_irq_restore(flags);
1402 }
1403 
1404 static int collect_events(struct perf_event *group, int max_count,
1405 			  struct perf_event *ctrs[], u64 *events,
1406 			  unsigned int *flags)
1407 {
1408 	int n = 0;
1409 	struct perf_event *event;
1410 
1411 	if (!is_software_event(group)) {
1412 		if (n >= max_count)
1413 			return -1;
1414 		ctrs[n] = group;
1415 		flags[n] = group->hw.event_base;
1416 		events[n++] = group->hw.config;
1417 	}
1418 	list_for_each_entry(event, &group->sibling_list, group_entry) {
1419 		if (!is_software_event(event) &&
1420 		    event->state != PERF_EVENT_STATE_OFF) {
1421 			if (n >= max_count)
1422 				return -1;
1423 			ctrs[n] = event;
1424 			flags[n] = event->hw.event_base;
1425 			events[n++] = event->hw.config;
1426 		}
1427 	}
1428 	return n;
1429 }
1430 
1431 /*
1432  * Add a event to the PMU.
1433  * If all events are not already frozen, then we disable and
1434  * re-enable the PMU in order to get hw_perf_enable to do the
1435  * actual work of reconfiguring the PMU.
1436  */
1437 static int power_pmu_add(struct perf_event *event, int ef_flags)
1438 {
1439 	struct cpu_hw_events *cpuhw;
1440 	unsigned long flags;
1441 	int n0;
1442 	int ret = -EAGAIN;
1443 
1444 	local_irq_save(flags);
1445 	perf_pmu_disable(event->pmu);
1446 
1447 	/*
1448 	 * Add the event to the list (if there is room)
1449 	 * and check whether the total set is still feasible.
1450 	 */
1451 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1452 	n0 = cpuhw->n_events;
1453 	if (n0 >= ppmu->n_counter)
1454 		goto out;
1455 	cpuhw->event[n0] = event;
1456 	cpuhw->events[n0] = event->hw.config;
1457 	cpuhw->flags[n0] = event->hw.event_base;
1458 
1459 	/*
1460 	 * This event may have been disabled/stopped in record_and_restart()
1461 	 * because we exceeded the ->event_limit. If re-starting the event,
1462 	 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
1463 	 * notification is re-enabled.
1464 	 */
1465 	if (!(ef_flags & PERF_EF_START))
1466 		event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1467 	else
1468 		event->hw.state = 0;
1469 
1470 	/*
1471 	 * If group events scheduling transaction was started,
1472 	 * skip the schedulability test here, it will be performed
1473 	 * at commit time(->commit_txn) as a whole
1474 	 */
1475 	if (cpuhw->txn_flags & PERF_PMU_TXN_ADD)
1476 		goto nocheck;
1477 
1478 	if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
1479 		goto out;
1480 	if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
1481 		goto out;
1482 	event->hw.config = cpuhw->events[n0];
1483 
1484 nocheck:
1485 	ebb_event_add(event);
1486 
1487 	++cpuhw->n_events;
1488 	++cpuhw->n_added;
1489 
1490 	ret = 0;
1491  out:
1492 	if (has_branch_stack(event)) {
1493 		power_pmu_bhrb_enable(event);
1494 		cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1495 					event->attr.branch_sample_type);
1496 	}
1497 
1498 	/*
1499 	 * Workaround for POWER9 DD1 to use the Instruction Counter
1500 	 * register value for instruction counting
1501 	 */
1502 	if (use_ic(event->attr.config))
1503 		cpuhw->ic_init = mfspr(SPRN_IC);
1504 
1505 	perf_pmu_enable(event->pmu);
1506 	local_irq_restore(flags);
1507 	return ret;
1508 }
1509 
1510 /*
1511  * Remove a event from the PMU.
1512  */
1513 static void power_pmu_del(struct perf_event *event, int ef_flags)
1514 {
1515 	struct cpu_hw_events *cpuhw;
1516 	long i;
1517 	unsigned long flags;
1518 
1519 	local_irq_save(flags);
1520 	perf_pmu_disable(event->pmu);
1521 
1522 	power_pmu_read(event);
1523 
1524 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1525 	for (i = 0; i < cpuhw->n_events; ++i) {
1526 		if (event == cpuhw->event[i]) {
1527 			while (++i < cpuhw->n_events) {
1528 				cpuhw->event[i-1] = cpuhw->event[i];
1529 				cpuhw->events[i-1] = cpuhw->events[i];
1530 				cpuhw->flags[i-1] = cpuhw->flags[i];
1531 			}
1532 			--cpuhw->n_events;
1533 			ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
1534 			if (event->hw.idx) {
1535 				write_pmc(event->hw.idx, 0);
1536 				event->hw.idx = 0;
1537 			}
1538 			perf_event_update_userpage(event);
1539 			break;
1540 		}
1541 	}
1542 	for (i = 0; i < cpuhw->n_limited; ++i)
1543 		if (event == cpuhw->limited_counter[i])
1544 			break;
1545 	if (i < cpuhw->n_limited) {
1546 		while (++i < cpuhw->n_limited) {
1547 			cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
1548 			cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
1549 		}
1550 		--cpuhw->n_limited;
1551 	}
1552 	if (cpuhw->n_events == 0) {
1553 		/* disable exceptions if no events are running */
1554 		cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
1555 	}
1556 
1557 	if (has_branch_stack(event))
1558 		power_pmu_bhrb_disable(event);
1559 
1560 	perf_pmu_enable(event->pmu);
1561 	local_irq_restore(flags);
1562 }
1563 
1564 /*
1565  * POWER-PMU does not support disabling individual counters, hence
1566  * program their cycle counter to their max value and ignore the interrupts.
1567  */
1568 
1569 static void power_pmu_start(struct perf_event *event, int ef_flags)
1570 {
1571 	unsigned long flags;
1572 	s64 left;
1573 	unsigned long val;
1574 
1575 	if (!event->hw.idx || !event->hw.sample_period)
1576 		return;
1577 
1578 	if (!(event->hw.state & PERF_HES_STOPPED))
1579 		return;
1580 
1581 	if (ef_flags & PERF_EF_RELOAD)
1582 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1583 
1584 	local_irq_save(flags);
1585 	perf_pmu_disable(event->pmu);
1586 
1587 	event->hw.state = 0;
1588 	left = local64_read(&event->hw.period_left);
1589 
1590 	val = 0;
1591 	if (left < 0x80000000L)
1592 		val = 0x80000000L - left;
1593 
1594 	write_pmc(event->hw.idx, val);
1595 
1596 	perf_event_update_userpage(event);
1597 	perf_pmu_enable(event->pmu);
1598 	local_irq_restore(flags);
1599 }
1600 
1601 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1602 {
1603 	unsigned long flags;
1604 
1605 	if (!event->hw.idx || !event->hw.sample_period)
1606 		return;
1607 
1608 	if (event->hw.state & PERF_HES_STOPPED)
1609 		return;
1610 
1611 	local_irq_save(flags);
1612 	perf_pmu_disable(event->pmu);
1613 
1614 	power_pmu_read(event);
1615 	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1616 	write_pmc(event->hw.idx, 0);
1617 
1618 	perf_event_update_userpage(event);
1619 	perf_pmu_enable(event->pmu);
1620 	local_irq_restore(flags);
1621 }
1622 
1623 /*
1624  * Start group events scheduling transaction
1625  * Set the flag to make pmu::enable() not perform the
1626  * schedulability test, it will be performed at commit time
1627  *
1628  * We only support PERF_PMU_TXN_ADD transactions. Save the
1629  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1630  * transactions.
1631  */
1632 static void power_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1633 {
1634 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1635 
1636 	WARN_ON_ONCE(cpuhw->txn_flags);		/* txn already in flight */
1637 
1638 	cpuhw->txn_flags = txn_flags;
1639 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1640 		return;
1641 
1642 	perf_pmu_disable(pmu);
1643 	cpuhw->n_txn_start = cpuhw->n_events;
1644 }
1645 
1646 /*
1647  * Stop group events scheduling transaction
1648  * Clear the flag and pmu::enable() will perform the
1649  * schedulability test.
1650  */
1651 static void power_pmu_cancel_txn(struct pmu *pmu)
1652 {
1653 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1654 	unsigned int txn_flags;
1655 
1656 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1657 
1658 	txn_flags = cpuhw->txn_flags;
1659 	cpuhw->txn_flags = 0;
1660 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1661 		return;
1662 
1663 	perf_pmu_enable(pmu);
1664 }
1665 
1666 /*
1667  * Commit group events scheduling transaction
1668  * Perform the group schedulability test as a whole
1669  * Return 0 if success
1670  */
1671 static int power_pmu_commit_txn(struct pmu *pmu)
1672 {
1673 	struct cpu_hw_events *cpuhw;
1674 	long i, n;
1675 
1676 	if (!ppmu)
1677 		return -EAGAIN;
1678 
1679 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1680 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1681 
1682 	if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
1683 		cpuhw->txn_flags = 0;
1684 		return 0;
1685 	}
1686 
1687 	n = cpuhw->n_events;
1688 	if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1689 		return -EAGAIN;
1690 	i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1691 	if (i < 0)
1692 		return -EAGAIN;
1693 
1694 	for (i = cpuhw->n_txn_start; i < n; ++i)
1695 		cpuhw->event[i]->hw.config = cpuhw->events[i];
1696 
1697 	cpuhw->txn_flags = 0;
1698 	perf_pmu_enable(pmu);
1699 	return 0;
1700 }
1701 
1702 /*
1703  * Return 1 if we might be able to put event on a limited PMC,
1704  * or 0 if not.
1705  * A event can only go on a limited PMC if it counts something
1706  * that a limited PMC can count, doesn't require interrupts, and
1707  * doesn't exclude any processor mode.
1708  */
1709 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1710 				 unsigned int flags)
1711 {
1712 	int n;
1713 	u64 alt[MAX_EVENT_ALTERNATIVES];
1714 
1715 	if (event->attr.exclude_user
1716 	    || event->attr.exclude_kernel
1717 	    || event->attr.exclude_hv
1718 	    || event->attr.sample_period)
1719 		return 0;
1720 
1721 	if (ppmu->limited_pmc_event(ev))
1722 		return 1;
1723 
1724 	/*
1725 	 * The requested event_id isn't on a limited PMC already;
1726 	 * see if any alternative code goes on a limited PMC.
1727 	 */
1728 	if (!ppmu->get_alternatives)
1729 		return 0;
1730 
1731 	flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1732 	n = ppmu->get_alternatives(ev, flags, alt);
1733 
1734 	return n > 0;
1735 }
1736 
1737 /*
1738  * Find an alternative event_id that goes on a normal PMC, if possible,
1739  * and return the event_id code, or 0 if there is no such alternative.
1740  * (Note: event_id code 0 is "don't count" on all machines.)
1741  */
1742 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1743 {
1744 	u64 alt[MAX_EVENT_ALTERNATIVES];
1745 	int n;
1746 
1747 	flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1748 	n = ppmu->get_alternatives(ev, flags, alt);
1749 	if (!n)
1750 		return 0;
1751 	return alt[0];
1752 }
1753 
1754 /* Number of perf_events counting hardware events */
1755 static atomic_t num_events;
1756 /* Used to avoid races in calling reserve/release_pmc_hardware */
1757 static DEFINE_MUTEX(pmc_reserve_mutex);
1758 
1759 /*
1760  * Release the PMU if this is the last perf_event.
1761  */
1762 static void hw_perf_event_destroy(struct perf_event *event)
1763 {
1764 	if (!atomic_add_unless(&num_events, -1, 1)) {
1765 		mutex_lock(&pmc_reserve_mutex);
1766 		if (atomic_dec_return(&num_events) == 0)
1767 			release_pmc_hardware();
1768 		mutex_unlock(&pmc_reserve_mutex);
1769 	}
1770 }
1771 
1772 /*
1773  * Translate a generic cache event_id config to a raw event_id code.
1774  */
1775 static int hw_perf_cache_event(u64 config, u64 *eventp)
1776 {
1777 	unsigned long type, op, result;
1778 	int ev;
1779 
1780 	if (!ppmu->cache_events)
1781 		return -EINVAL;
1782 
1783 	/* unpack config */
1784 	type = config & 0xff;
1785 	op = (config >> 8) & 0xff;
1786 	result = (config >> 16) & 0xff;
1787 
1788 	if (type >= PERF_COUNT_HW_CACHE_MAX ||
1789 	    op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1790 	    result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1791 		return -EINVAL;
1792 
1793 	ev = (*ppmu->cache_events)[type][op][result];
1794 	if (ev == 0)
1795 		return -EOPNOTSUPP;
1796 	if (ev == -1)
1797 		return -EINVAL;
1798 	*eventp = ev;
1799 	return 0;
1800 }
1801 
1802 static int power_pmu_event_init(struct perf_event *event)
1803 {
1804 	u64 ev;
1805 	unsigned long flags;
1806 	struct perf_event *ctrs[MAX_HWEVENTS];
1807 	u64 events[MAX_HWEVENTS];
1808 	unsigned int cflags[MAX_HWEVENTS];
1809 	int n;
1810 	int err;
1811 	struct cpu_hw_events *cpuhw;
1812 
1813 	if (!ppmu)
1814 		return -ENOENT;
1815 
1816 	if (has_branch_stack(event)) {
1817 	        /* PMU has BHRB enabled */
1818 		if (!(ppmu->flags & PPMU_ARCH_207S))
1819 			return -EOPNOTSUPP;
1820 	}
1821 
1822 	switch (event->attr.type) {
1823 	case PERF_TYPE_HARDWARE:
1824 		ev = event->attr.config;
1825 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1826 			return -EOPNOTSUPP;
1827 		ev = ppmu->generic_events[ev];
1828 		break;
1829 	case PERF_TYPE_HW_CACHE:
1830 		err = hw_perf_cache_event(event->attr.config, &ev);
1831 		if (err)
1832 			return err;
1833 		break;
1834 	case PERF_TYPE_RAW:
1835 		ev = event->attr.config;
1836 		break;
1837 	default:
1838 		return -ENOENT;
1839 	}
1840 
1841 	event->hw.config_base = ev;
1842 	event->hw.idx = 0;
1843 
1844 	/*
1845 	 * If we are not running on a hypervisor, force the
1846 	 * exclude_hv bit to 0 so that we don't care what
1847 	 * the user set it to.
1848 	 */
1849 	if (!firmware_has_feature(FW_FEATURE_LPAR))
1850 		event->attr.exclude_hv = 0;
1851 
1852 	/*
1853 	 * If this is a per-task event, then we can use
1854 	 * PM_RUN_* events interchangeably with their non RUN_*
1855 	 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1856 	 * XXX we should check if the task is an idle task.
1857 	 */
1858 	flags = 0;
1859 	if (event->attach_state & PERF_ATTACH_TASK)
1860 		flags |= PPMU_ONLY_COUNT_RUN;
1861 
1862 	/*
1863 	 * If this machine has limited events, check whether this
1864 	 * event_id could go on a limited event.
1865 	 */
1866 	if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1867 		if (can_go_on_limited_pmc(event, ev, flags)) {
1868 			flags |= PPMU_LIMITED_PMC_OK;
1869 		} else if (ppmu->limited_pmc_event(ev)) {
1870 			/*
1871 			 * The requested event_id is on a limited PMC,
1872 			 * but we can't use a limited PMC; see if any
1873 			 * alternative goes on a normal PMC.
1874 			 */
1875 			ev = normal_pmc_alternative(ev, flags);
1876 			if (!ev)
1877 				return -EINVAL;
1878 		}
1879 	}
1880 
1881 	/* Extra checks for EBB */
1882 	err = ebb_event_check(event);
1883 	if (err)
1884 		return err;
1885 
1886 	/*
1887 	 * If this is in a group, check if it can go on with all the
1888 	 * other hardware events in the group.  We assume the event
1889 	 * hasn't been linked into its leader's sibling list at this point.
1890 	 */
1891 	n = 0;
1892 	if (event->group_leader != event) {
1893 		n = collect_events(event->group_leader, ppmu->n_counter - 1,
1894 				   ctrs, events, cflags);
1895 		if (n < 0)
1896 			return -EINVAL;
1897 	}
1898 	events[n] = ev;
1899 	ctrs[n] = event;
1900 	cflags[n] = flags;
1901 	if (check_excludes(ctrs, cflags, n, 1))
1902 		return -EINVAL;
1903 
1904 	cpuhw = &get_cpu_var(cpu_hw_events);
1905 	err = power_check_constraints(cpuhw, events, cflags, n + 1);
1906 
1907 	if (has_branch_stack(event)) {
1908 		cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1909 					event->attr.branch_sample_type);
1910 
1911 		if (cpuhw->bhrb_filter == -1) {
1912 			put_cpu_var(cpu_hw_events);
1913 			return -EOPNOTSUPP;
1914 		}
1915 	}
1916 
1917 	put_cpu_var(cpu_hw_events);
1918 	if (err)
1919 		return -EINVAL;
1920 
1921 	event->hw.config = events[n];
1922 	event->hw.event_base = cflags[n];
1923 	event->hw.last_period = event->hw.sample_period;
1924 	local64_set(&event->hw.period_left, event->hw.last_period);
1925 
1926 	/*
1927 	 * For EBB events we just context switch the PMC value, we don't do any
1928 	 * of the sample_period logic. We use hw.prev_count for this.
1929 	 */
1930 	if (is_ebb_event(event))
1931 		local64_set(&event->hw.prev_count, 0);
1932 
1933 	/*
1934 	 * See if we need to reserve the PMU.
1935 	 * If no events are currently in use, then we have to take a
1936 	 * mutex to ensure that we don't race with another task doing
1937 	 * reserve_pmc_hardware or release_pmc_hardware.
1938 	 */
1939 	err = 0;
1940 	if (!atomic_inc_not_zero(&num_events)) {
1941 		mutex_lock(&pmc_reserve_mutex);
1942 		if (atomic_read(&num_events) == 0 &&
1943 		    reserve_pmc_hardware(perf_event_interrupt))
1944 			err = -EBUSY;
1945 		else
1946 			atomic_inc(&num_events);
1947 		mutex_unlock(&pmc_reserve_mutex);
1948 	}
1949 	event->destroy = hw_perf_event_destroy;
1950 
1951 	return err;
1952 }
1953 
1954 static int power_pmu_event_idx(struct perf_event *event)
1955 {
1956 	return event->hw.idx;
1957 }
1958 
1959 ssize_t power_events_sysfs_show(struct device *dev,
1960 				struct device_attribute *attr, char *page)
1961 {
1962 	struct perf_pmu_events_attr *pmu_attr;
1963 
1964 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
1965 
1966 	return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
1967 }
1968 
1969 static struct pmu power_pmu = {
1970 	.pmu_enable	= power_pmu_enable,
1971 	.pmu_disable	= power_pmu_disable,
1972 	.event_init	= power_pmu_event_init,
1973 	.add		= power_pmu_add,
1974 	.del		= power_pmu_del,
1975 	.start		= power_pmu_start,
1976 	.stop		= power_pmu_stop,
1977 	.read		= power_pmu_read,
1978 	.start_txn	= power_pmu_start_txn,
1979 	.cancel_txn	= power_pmu_cancel_txn,
1980 	.commit_txn	= power_pmu_commit_txn,
1981 	.event_idx	= power_pmu_event_idx,
1982 	.sched_task	= power_pmu_sched_task,
1983 };
1984 
1985 /*
1986  * A counter has overflowed; update its count and record
1987  * things if requested.  Note that interrupts are hard-disabled
1988  * here so there is no possibility of being interrupted.
1989  */
1990 static void record_and_restart(struct perf_event *event, unsigned long val,
1991 			       struct pt_regs *regs)
1992 {
1993 	u64 period = event->hw.sample_period;
1994 	s64 prev, delta, left;
1995 	int record = 0;
1996 
1997 	if (event->hw.state & PERF_HES_STOPPED) {
1998 		write_pmc(event->hw.idx, 0);
1999 		return;
2000 	}
2001 
2002 	/* we don't have to worry about interrupts here */
2003 	prev = local64_read(&event->hw.prev_count);
2004 	delta = check_and_compute_delta(prev, val);
2005 	local64_add(delta, &event->count);
2006 
2007 	/*
2008 	 * See if the total period for this event has expired,
2009 	 * and update for the next period.
2010 	 */
2011 	val = 0;
2012 	left = local64_read(&event->hw.period_left) - delta;
2013 	if (delta == 0)
2014 		left++;
2015 	if (period) {
2016 		if (left <= 0) {
2017 			left += period;
2018 			if (left <= 0)
2019 				left = period;
2020 			record = siar_valid(regs);
2021 			event->hw.last_period = event->hw.sample_period;
2022 		}
2023 		if (left < 0x80000000LL)
2024 			val = 0x80000000LL - left;
2025 	}
2026 
2027 	write_pmc(event->hw.idx, val);
2028 	local64_set(&event->hw.prev_count, val);
2029 	local64_set(&event->hw.period_left, left);
2030 	perf_event_update_userpage(event);
2031 
2032 	/*
2033 	 * Finally record data if requested.
2034 	 */
2035 	if (record) {
2036 		struct perf_sample_data data;
2037 
2038 		perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
2039 
2040 		if (event->attr.sample_type & PERF_SAMPLE_ADDR)
2041 			perf_get_data_addr(regs, &data.addr);
2042 
2043 		if (event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK) {
2044 			struct cpu_hw_events *cpuhw;
2045 			cpuhw = this_cpu_ptr(&cpu_hw_events);
2046 			power_pmu_bhrb_read(cpuhw);
2047 			data.br_stack = &cpuhw->bhrb_stack;
2048 		}
2049 
2050 		if (perf_event_overflow(event, &data, regs))
2051 			power_pmu_stop(event, 0);
2052 	}
2053 }
2054 
2055 /*
2056  * Called from generic code to get the misc flags (i.e. processor mode)
2057  * for an event_id.
2058  */
2059 unsigned long perf_misc_flags(struct pt_regs *regs)
2060 {
2061 	u32 flags = perf_get_misc_flags(regs);
2062 
2063 	if (flags)
2064 		return flags;
2065 	return user_mode(regs) ? PERF_RECORD_MISC_USER :
2066 		PERF_RECORD_MISC_KERNEL;
2067 }
2068 
2069 /*
2070  * Called from generic code to get the instruction pointer
2071  * for an event_id.
2072  */
2073 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2074 {
2075 	bool use_siar = regs_use_siar(regs);
2076 
2077 	if (use_siar && siar_valid(regs))
2078 		return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
2079 	else if (use_siar)
2080 		return 0;		// no valid instruction pointer
2081 	else
2082 		return regs->nip;
2083 }
2084 
2085 static bool pmc_overflow_power7(unsigned long val)
2086 {
2087 	/*
2088 	 * Events on POWER7 can roll back if a speculative event doesn't
2089 	 * eventually complete. Unfortunately in some rare cases they will
2090 	 * raise a performance monitor exception. We need to catch this to
2091 	 * ensure we reset the PMC. In all cases the PMC will be 256 or less
2092 	 * cycles from overflow.
2093 	 *
2094 	 * We only do this if the first pass fails to find any overflowing
2095 	 * PMCs because a user might set a period of less than 256 and we
2096 	 * don't want to mistakenly reset them.
2097 	 */
2098 	if ((0x80000000 - val) <= 256)
2099 		return true;
2100 
2101 	return false;
2102 }
2103 
2104 static bool pmc_overflow(unsigned long val)
2105 {
2106 	if ((int)val < 0)
2107 		return true;
2108 
2109 	return false;
2110 }
2111 
2112 /*
2113  * Performance monitor interrupt stuff
2114  */
2115 static void perf_event_interrupt(struct pt_regs *regs)
2116 {
2117 	int i, j;
2118 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
2119 	struct perf_event *event;
2120 	unsigned long val[8];
2121 	int found, active;
2122 	int nmi;
2123 
2124 	if (cpuhw->n_limited)
2125 		freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
2126 					mfspr(SPRN_PMC6));
2127 
2128 	perf_read_regs(regs);
2129 
2130 	nmi = perf_intr_is_nmi(regs);
2131 	if (nmi)
2132 		nmi_enter();
2133 	else
2134 		irq_enter();
2135 
2136 	/* Read all the PMCs since we'll need them a bunch of times */
2137 	for (i = 0; i < ppmu->n_counter; ++i)
2138 		val[i] = read_pmc(i + 1);
2139 
2140 	/* Try to find what caused the IRQ */
2141 	found = 0;
2142 	for (i = 0; i < ppmu->n_counter; ++i) {
2143 		if (!pmc_overflow(val[i]))
2144 			continue;
2145 		if (is_limited_pmc(i + 1))
2146 			continue; /* these won't generate IRQs */
2147 		/*
2148 		 * We've found one that's overflowed.  For active
2149 		 * counters we need to log this.  For inactive
2150 		 * counters, we need to reset it anyway
2151 		 */
2152 		found = 1;
2153 		active = 0;
2154 		for (j = 0; j < cpuhw->n_events; ++j) {
2155 			event = cpuhw->event[j];
2156 			if (event->hw.idx == (i + 1)) {
2157 				active = 1;
2158 				record_and_restart(event, val[i], regs);
2159 				break;
2160 			}
2161 		}
2162 		if (!active)
2163 			/* reset non active counters that have overflowed */
2164 			write_pmc(i + 1, 0);
2165 	}
2166 	if (!found && pvr_version_is(PVR_POWER7)) {
2167 		/* check active counters for special buggy p7 overflow */
2168 		for (i = 0; i < cpuhw->n_events; ++i) {
2169 			event = cpuhw->event[i];
2170 			if (!event->hw.idx || is_limited_pmc(event->hw.idx))
2171 				continue;
2172 			if (pmc_overflow_power7(val[event->hw.idx - 1])) {
2173 				/* event has overflowed in a buggy way*/
2174 				found = 1;
2175 				record_and_restart(event,
2176 						   val[event->hw.idx - 1],
2177 						   regs);
2178 			}
2179 		}
2180 	}
2181 	if (!found && !nmi && printk_ratelimit())
2182 		printk(KERN_WARNING "Can't find PMC that caused IRQ\n");
2183 
2184 	/*
2185 	 * Reset MMCR0 to its normal value.  This will set PMXE and
2186 	 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
2187 	 * and thus allow interrupts to occur again.
2188 	 * XXX might want to use MSR.PM to keep the events frozen until
2189 	 * we get back out of this interrupt.
2190 	 */
2191 	write_mmcr0(cpuhw, cpuhw->mmcr[0]);
2192 
2193 	if (nmi)
2194 		nmi_exit();
2195 	else
2196 		irq_exit();
2197 }
2198 
2199 static int power_pmu_prepare_cpu(unsigned int cpu)
2200 {
2201 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
2202 
2203 	if (ppmu) {
2204 		memset(cpuhw, 0, sizeof(*cpuhw));
2205 		cpuhw->mmcr[0] = MMCR0_FC;
2206 	}
2207 	return 0;
2208 }
2209 
2210 int register_power_pmu(struct power_pmu *pmu)
2211 {
2212 	if (ppmu)
2213 		return -EBUSY;		/* something's already registered */
2214 
2215 	ppmu = pmu;
2216 	pr_info("%s performance monitor hardware support registered\n",
2217 		pmu->name);
2218 
2219 	power_pmu.attr_groups = ppmu->attr_groups;
2220 
2221 #ifdef MSR_HV
2222 	/*
2223 	 * Use FCHV to ignore kernel events if MSR.HV is set.
2224 	 */
2225 	if (mfmsr() & MSR_HV)
2226 		freeze_events_kernel = MMCR0_FCHV;
2227 #endif /* CONFIG_PPC64 */
2228 
2229 	perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
2230 	cpuhp_setup_state(CPUHP_PERF_POWER, "perf/powerpc:prepare",
2231 			  power_pmu_prepare_cpu, NULL);
2232 	return 0;
2233 }
2234