xref: /linux/arch/powerpc/perf/core-book3s.c (revision c0e297dc61f8d4453e07afbea1fa8d0e67cd4a34)
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 group_flag;
52 	int n_txn_start;
53 
54 	/* BHRB bits */
55 	u64				bhrb_filter;	/* BHRB HW branch filter */
56 	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 };
61 
62 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
63 
64 static struct power_pmu *ppmu;
65 
66 /*
67  * Normally, to ignore kernel events we set the FCS (freeze counters
68  * in supervisor mode) bit in MMCR0, but if the kernel runs with the
69  * hypervisor bit set in the MSR, or if we are running on a processor
70  * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
71  * then we need to use the FCHV bit to ignore kernel events.
72  */
73 static unsigned int freeze_events_kernel = MMCR0_FCS;
74 
75 /*
76  * 32-bit doesn't have MMCRA but does have an MMCR2,
77  * and a few other names are different.
78  */
79 #ifdef CONFIG_PPC32
80 
81 #define MMCR0_FCHV		0
82 #define MMCR0_PMCjCE		MMCR0_PMCnCE
83 #define MMCR0_FC56		0
84 #define MMCR0_PMAO		0
85 #define MMCR0_EBE		0
86 #define MMCR0_BHRBA		0
87 #define MMCR0_PMCC		0
88 #define MMCR0_PMCC_U6		0
89 
90 #define SPRN_MMCRA		SPRN_MMCR2
91 #define MMCRA_SAMPLE_ENABLE	0
92 
93 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
94 {
95 	return 0;
96 }
97 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
98 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
99 {
100 	return 0;
101 }
102 static inline void perf_read_regs(struct pt_regs *regs)
103 {
104 	regs->result = 0;
105 }
106 static inline int perf_intr_is_nmi(struct pt_regs *regs)
107 {
108 	return 0;
109 }
110 
111 static inline int siar_valid(struct pt_regs *regs)
112 {
113 	return 1;
114 }
115 
116 static bool is_ebb_event(struct perf_event *event) { return false; }
117 static int ebb_event_check(struct perf_event *event) { return 0; }
118 static void ebb_event_add(struct perf_event *event) { }
119 static void ebb_switch_out(unsigned long mmcr0) { }
120 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
121 {
122 	return cpuhw->mmcr[0];
123 }
124 
125 static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
126 static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
127 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in) {}
128 static inline void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw) {}
129 static void pmao_restore_workaround(bool ebb) { }
130 #endif /* CONFIG_PPC32 */
131 
132 static bool regs_use_siar(struct pt_regs *regs)
133 {
134 	/*
135 	 * When we take a performance monitor exception the regs are setup
136 	 * using perf_read_regs() which overloads some fields, in particular
137 	 * regs->result to tell us whether to use SIAR.
138 	 *
139 	 * However if the regs are from another exception, eg. a syscall, then
140 	 * they have not been setup using perf_read_regs() and so regs->result
141 	 * is something random.
142 	 */
143 	return ((TRAP(regs) == 0xf00) && regs->result);
144 }
145 
146 /*
147  * Things that are specific to 64-bit implementations.
148  */
149 #ifdef CONFIG_PPC64
150 
151 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
152 {
153 	unsigned long mmcra = regs->dsisr;
154 
155 	if ((ppmu->flags & PPMU_HAS_SSLOT) && (mmcra & MMCRA_SAMPLE_ENABLE)) {
156 		unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
157 		if (slot > 1)
158 			return 4 * (slot - 1);
159 	}
160 
161 	return 0;
162 }
163 
164 /*
165  * The user wants a data address recorded.
166  * If we're not doing instruction sampling, give them the SDAR
167  * (sampled data address).  If we are doing instruction sampling, then
168  * only give them the SDAR if it corresponds to the instruction
169  * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC, the
170  * [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA, or the SDAR_VALID bit in SIER.
171  */
172 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
173 {
174 	unsigned long mmcra = regs->dsisr;
175 	bool sdar_valid;
176 
177 	if (ppmu->flags & PPMU_HAS_SIER)
178 		sdar_valid = regs->dar & SIER_SDAR_VALID;
179 	else {
180 		unsigned long sdsync;
181 
182 		if (ppmu->flags & PPMU_SIAR_VALID)
183 			sdsync = POWER7P_MMCRA_SDAR_VALID;
184 		else if (ppmu->flags & PPMU_ALT_SIPR)
185 			sdsync = POWER6_MMCRA_SDSYNC;
186 		else
187 			sdsync = MMCRA_SDSYNC;
188 
189 		sdar_valid = mmcra & sdsync;
190 	}
191 
192 	if (!(mmcra & MMCRA_SAMPLE_ENABLE) || sdar_valid)
193 		*addrp = mfspr(SPRN_SDAR);
194 }
195 
196 static bool regs_sihv(struct pt_regs *regs)
197 {
198 	unsigned long sihv = MMCRA_SIHV;
199 
200 	if (ppmu->flags & PPMU_HAS_SIER)
201 		return !!(regs->dar & SIER_SIHV);
202 
203 	if (ppmu->flags & PPMU_ALT_SIPR)
204 		sihv = POWER6_MMCRA_SIHV;
205 
206 	return !!(regs->dsisr & sihv);
207 }
208 
209 static bool regs_sipr(struct pt_regs *regs)
210 {
211 	unsigned long sipr = MMCRA_SIPR;
212 
213 	if (ppmu->flags & PPMU_HAS_SIER)
214 		return !!(regs->dar & SIER_SIPR);
215 
216 	if (ppmu->flags & PPMU_ALT_SIPR)
217 		sipr = POWER6_MMCRA_SIPR;
218 
219 	return !!(regs->dsisr & sipr);
220 }
221 
222 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
223 {
224 	if (regs->msr & MSR_PR)
225 		return PERF_RECORD_MISC_USER;
226 	if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
227 		return PERF_RECORD_MISC_HYPERVISOR;
228 	return PERF_RECORD_MISC_KERNEL;
229 }
230 
231 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
232 {
233 	bool use_siar = regs_use_siar(regs);
234 
235 	if (!use_siar)
236 		return perf_flags_from_msr(regs);
237 
238 	/*
239 	 * If we don't have flags in MMCRA, rather than using
240 	 * the MSR, we intuit the flags from the address in
241 	 * SIAR which should give slightly more reliable
242 	 * results
243 	 */
244 	if (ppmu->flags & PPMU_NO_SIPR) {
245 		unsigned long siar = mfspr(SPRN_SIAR);
246 		if (siar >= PAGE_OFFSET)
247 			return PERF_RECORD_MISC_KERNEL;
248 		return PERF_RECORD_MISC_USER;
249 	}
250 
251 	/* PR has priority over HV, so order below is important */
252 	if (regs_sipr(regs))
253 		return PERF_RECORD_MISC_USER;
254 
255 	if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
256 		return PERF_RECORD_MISC_HYPERVISOR;
257 
258 	return PERF_RECORD_MISC_KERNEL;
259 }
260 
261 /*
262  * Overload regs->dsisr to store MMCRA so we only need to read it once
263  * on each interrupt.
264  * Overload regs->dar to store SIER if we have it.
265  * Overload regs->result to specify whether we should use the MSR (result
266  * is zero) or the SIAR (result is non zero).
267  */
268 static inline void perf_read_regs(struct pt_regs *regs)
269 {
270 	unsigned long mmcra = mfspr(SPRN_MMCRA);
271 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
272 	int use_siar;
273 
274 	regs->dsisr = mmcra;
275 
276 	if (ppmu->flags & PPMU_HAS_SIER)
277 		regs->dar = mfspr(SPRN_SIER);
278 
279 	/*
280 	 * If this isn't a PMU exception (eg a software event) the SIAR is
281 	 * not valid. Use pt_regs.
282 	 *
283 	 * If it is a marked event use the SIAR.
284 	 *
285 	 * If the PMU doesn't update the SIAR for non marked events use
286 	 * pt_regs.
287 	 *
288 	 * If the PMU has HV/PR flags then check to see if they
289 	 * place the exception in userspace. If so, use pt_regs. In
290 	 * continuous sampling mode the SIAR and the PMU exception are
291 	 * not synchronised, so they may be many instructions apart.
292 	 * This can result in confusing backtraces. We still want
293 	 * hypervisor samples as well as samples in the kernel with
294 	 * interrupts off hence the userspace check.
295 	 */
296 	if (TRAP(regs) != 0xf00)
297 		use_siar = 0;
298 	else if (marked)
299 		use_siar = 1;
300 	else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
301 		use_siar = 0;
302 	else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
303 		use_siar = 0;
304 	else
305 		use_siar = 1;
306 
307 	regs->result = use_siar;
308 }
309 
310 /*
311  * If interrupts were soft-disabled when a PMU interrupt occurs, treat
312  * it as an NMI.
313  */
314 static inline int perf_intr_is_nmi(struct pt_regs *regs)
315 {
316 	return !regs->softe;
317 }
318 
319 /*
320  * On processors like P7+ that have the SIAR-Valid bit, marked instructions
321  * must be sampled only if the SIAR-valid bit is set.
322  *
323  * For unmarked instructions and for processors that don't have the SIAR-Valid
324  * bit, assume that SIAR is valid.
325  */
326 static inline int siar_valid(struct pt_regs *regs)
327 {
328 	unsigned long mmcra = regs->dsisr;
329 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
330 
331 	if (marked) {
332 		if (ppmu->flags & PPMU_HAS_SIER)
333 			return regs->dar & SIER_SIAR_VALID;
334 
335 		if (ppmu->flags & PPMU_SIAR_VALID)
336 			return mmcra & POWER7P_MMCRA_SIAR_VALID;
337 	}
338 
339 	return 1;
340 }
341 
342 
343 /* Reset all possible BHRB entries */
344 static void power_pmu_bhrb_reset(void)
345 {
346 	asm volatile(PPC_CLRBHRB);
347 }
348 
349 static void power_pmu_bhrb_enable(struct perf_event *event)
350 {
351 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
352 
353 	if (!ppmu->bhrb_nr)
354 		return;
355 
356 	/* Clear BHRB if we changed task context to avoid data leaks */
357 	if (event->ctx->task && cpuhw->bhrb_context != event->ctx) {
358 		power_pmu_bhrb_reset();
359 		cpuhw->bhrb_context = event->ctx;
360 	}
361 	cpuhw->bhrb_users++;
362 	perf_sched_cb_inc(event->ctx->pmu);
363 }
364 
365 static void power_pmu_bhrb_disable(struct perf_event *event)
366 {
367 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
368 
369 	if (!ppmu->bhrb_nr)
370 		return;
371 
372 	cpuhw->bhrb_users--;
373 	WARN_ON_ONCE(cpuhw->bhrb_users < 0);
374 	perf_sched_cb_dec(event->ctx->pmu);
375 
376 	if (!cpuhw->disabled && !cpuhw->bhrb_users) {
377 		/* BHRB cannot be turned off when other
378 		 * events are active on the PMU.
379 		 */
380 
381 		/* avoid stale pointer */
382 		cpuhw->bhrb_context = NULL;
383 	}
384 }
385 
386 /* Called from ctxsw to prevent one process's branch entries to
387  * mingle with the other process's entries during context switch.
388  */
389 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
390 {
391 	if (!ppmu->bhrb_nr)
392 		return;
393 
394 	if (sched_in)
395 		power_pmu_bhrb_reset();
396 }
397 /* Calculate the to address for a branch */
398 static __u64 power_pmu_bhrb_to(u64 addr)
399 {
400 	unsigned int instr;
401 	int ret;
402 	__u64 target;
403 
404 	if (is_kernel_addr(addr))
405 		return branch_target((unsigned int *)addr);
406 
407 	/* Userspace: need copy instruction here then translate it */
408 	pagefault_disable();
409 	ret = __get_user_inatomic(instr, (unsigned int __user *)addr);
410 	if (ret) {
411 		pagefault_enable();
412 		return 0;
413 	}
414 	pagefault_enable();
415 
416 	target = branch_target(&instr);
417 	if ((!target) || (instr & BRANCH_ABSOLUTE))
418 		return target;
419 
420 	/* Translate relative branch target from kernel to user address */
421 	return target - (unsigned long)&instr + addr;
422 }
423 
424 /* Processing BHRB entries */
425 static void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw)
426 {
427 	u64 val;
428 	u64 addr;
429 	int r_index, u_index, pred;
430 
431 	r_index = 0;
432 	u_index = 0;
433 	while (r_index < ppmu->bhrb_nr) {
434 		/* Assembly read function */
435 		val = read_bhrb(r_index++);
436 		if (!val)
437 			/* Terminal marker: End of valid BHRB entries */
438 			break;
439 		else {
440 			addr = val & BHRB_EA;
441 			pred = val & BHRB_PREDICTION;
442 
443 			if (!addr)
444 				/* invalid entry */
445 				continue;
446 
447 			/* Branches are read most recent first (ie. mfbhrb 0 is
448 			 * the most recent branch).
449 			 * There are two types of valid entries:
450 			 * 1) a target entry which is the to address of a
451 			 *    computed goto like a blr,bctr,btar.  The next
452 			 *    entry read from the bhrb will be branch
453 			 *    corresponding to this target (ie. the actual
454 			 *    blr/bctr/btar instruction).
455 			 * 2) a from address which is an actual branch.  If a
456 			 *    target entry proceeds this, then this is the
457 			 *    matching branch for that target.  If this is not
458 			 *    following a target entry, then this is a branch
459 			 *    where the target is given as an immediate field
460 			 *    in the instruction (ie. an i or b form branch).
461 			 *    In this case we need to read the instruction from
462 			 *    memory to determine the target/to address.
463 			 */
464 
465 			if (val & BHRB_TARGET) {
466 				/* Target branches use two entries
467 				 * (ie. computed gotos/XL form)
468 				 */
469 				cpuhw->bhrb_entries[u_index].to = addr;
470 				cpuhw->bhrb_entries[u_index].mispred = pred;
471 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
472 
473 				/* Get from address in next entry */
474 				val = read_bhrb(r_index++);
475 				addr = val & BHRB_EA;
476 				if (val & BHRB_TARGET) {
477 					/* Shouldn't have two targets in a
478 					   row.. Reset index and try again */
479 					r_index--;
480 					addr = 0;
481 				}
482 				cpuhw->bhrb_entries[u_index].from = addr;
483 			} else {
484 				/* Branches to immediate field
485 				   (ie I or B form) */
486 				cpuhw->bhrb_entries[u_index].from = addr;
487 				cpuhw->bhrb_entries[u_index].to =
488 					power_pmu_bhrb_to(addr);
489 				cpuhw->bhrb_entries[u_index].mispred = pred;
490 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
491 			}
492 			u_index++;
493 
494 		}
495 	}
496 	cpuhw->bhrb_stack.nr = u_index;
497 	return;
498 }
499 
500 static bool is_ebb_event(struct perf_event *event)
501 {
502 	/*
503 	 * This could be a per-PMU callback, but we'd rather avoid the cost. We
504 	 * check that the PMU supports EBB, meaning those that don't can still
505 	 * use bit 63 of the event code for something else if they wish.
506 	 */
507 	return (ppmu->flags & PPMU_ARCH_207S) &&
508 	       ((event->attr.config >> PERF_EVENT_CONFIG_EBB_SHIFT) & 1);
509 }
510 
511 static int ebb_event_check(struct perf_event *event)
512 {
513 	struct perf_event *leader = event->group_leader;
514 
515 	/* Event and group leader must agree on EBB */
516 	if (is_ebb_event(leader) != is_ebb_event(event))
517 		return -EINVAL;
518 
519 	if (is_ebb_event(event)) {
520 		if (!(event->attach_state & PERF_ATTACH_TASK))
521 			return -EINVAL;
522 
523 		if (!leader->attr.pinned || !leader->attr.exclusive)
524 			return -EINVAL;
525 
526 		if (event->attr.freq ||
527 		    event->attr.inherit ||
528 		    event->attr.sample_type ||
529 		    event->attr.sample_period ||
530 		    event->attr.enable_on_exec)
531 			return -EINVAL;
532 	}
533 
534 	return 0;
535 }
536 
537 static void ebb_event_add(struct perf_event *event)
538 {
539 	if (!is_ebb_event(event) || current->thread.used_ebb)
540 		return;
541 
542 	/*
543 	 * IFF this is the first time we've added an EBB event, set
544 	 * PMXE in the user MMCR0 so we can detect when it's cleared by
545 	 * userspace. We need this so that we can context switch while
546 	 * userspace is in the EBB handler (where PMXE is 0).
547 	 */
548 	current->thread.used_ebb = 1;
549 	current->thread.mmcr0 |= MMCR0_PMXE;
550 }
551 
552 static void ebb_switch_out(unsigned long mmcr0)
553 {
554 	if (!(mmcr0 & MMCR0_EBE))
555 		return;
556 
557 	current->thread.siar  = mfspr(SPRN_SIAR);
558 	current->thread.sier  = mfspr(SPRN_SIER);
559 	current->thread.sdar  = mfspr(SPRN_SDAR);
560 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
561 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
562 }
563 
564 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
565 {
566 	unsigned long mmcr0 = cpuhw->mmcr[0];
567 
568 	if (!ebb)
569 		goto out;
570 
571 	/* Enable EBB and read/write to all 6 PMCs and BHRB for userspace */
572 	mmcr0 |= MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC_U6;
573 
574 	/*
575 	 * Add any bits from the user MMCR0, FC or PMAO. This is compatible
576 	 * with pmao_restore_workaround() because we may add PMAO but we never
577 	 * clear it here.
578 	 */
579 	mmcr0 |= current->thread.mmcr0;
580 
581 	/*
582 	 * Be careful not to set PMXE if userspace had it cleared. This is also
583 	 * compatible with pmao_restore_workaround() because it has already
584 	 * cleared PMXE and we leave PMAO alone.
585 	 */
586 	if (!(current->thread.mmcr0 & MMCR0_PMXE))
587 		mmcr0 &= ~MMCR0_PMXE;
588 
589 	mtspr(SPRN_SIAR, current->thread.siar);
590 	mtspr(SPRN_SIER, current->thread.sier);
591 	mtspr(SPRN_SDAR, current->thread.sdar);
592 
593 	/*
594 	 * Merge the kernel & user values of MMCR2. The semantics we implement
595 	 * are that the user MMCR2 can set bits, ie. cause counters to freeze,
596 	 * but not clear bits. If a task wants to be able to clear bits, ie.
597 	 * unfreeze counters, it should not set exclude_xxx in its events and
598 	 * instead manage the MMCR2 entirely by itself.
599 	 */
600 	mtspr(SPRN_MMCR2, cpuhw->mmcr[3] | current->thread.mmcr2);
601 out:
602 	return mmcr0;
603 }
604 
605 static void pmao_restore_workaround(bool ebb)
606 {
607 	unsigned pmcs[6];
608 
609 	if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
610 		return;
611 
612 	/*
613 	 * On POWER8E there is a hardware defect which affects the PMU context
614 	 * switch logic, ie. power_pmu_disable/enable().
615 	 *
616 	 * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
617 	 * by the hardware. Sometime later the actual PMU exception is
618 	 * delivered.
619 	 *
620 	 * If we context switch, or simply disable/enable, the PMU prior to the
621 	 * exception arriving, the exception will be lost when we clear PMAO.
622 	 *
623 	 * When we reenable the PMU, we will write the saved MMCR0 with PMAO
624 	 * set, and this _should_ generate an exception. However because of the
625 	 * defect no exception is generated when we write PMAO, and we get
626 	 * stuck with no counters counting but no exception delivered.
627 	 *
628 	 * The workaround is to detect this case and tweak the hardware to
629 	 * create another pending PMU exception.
630 	 *
631 	 * We do that by setting up PMC6 (cycles) for an imminent overflow and
632 	 * enabling the PMU. That causes a new exception to be generated in the
633 	 * chip, but we don't take it yet because we have interrupts hard
634 	 * disabled. We then write back the PMU state as we want it to be seen
635 	 * by the exception handler. When we reenable interrupts the exception
636 	 * handler will be called and see the correct state.
637 	 *
638 	 * The logic is the same for EBB, except that the exception is gated by
639 	 * us having interrupts hard disabled as well as the fact that we are
640 	 * not in userspace. The exception is finally delivered when we return
641 	 * to userspace.
642 	 */
643 
644 	/* Only if PMAO is set and PMAO_SYNC is clear */
645 	if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
646 		return;
647 
648 	/* If we're doing EBB, only if BESCR[GE] is set */
649 	if (ebb && !(current->thread.bescr & BESCR_GE))
650 		return;
651 
652 	/*
653 	 * We are already soft-disabled in power_pmu_enable(). We need to hard
654 	 * enable to actually prevent the PMU exception from firing.
655 	 */
656 	hard_irq_disable();
657 
658 	/*
659 	 * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
660 	 * Using read/write_pmc() in a for loop adds 12 function calls and
661 	 * almost doubles our code size.
662 	 */
663 	pmcs[0] = mfspr(SPRN_PMC1);
664 	pmcs[1] = mfspr(SPRN_PMC2);
665 	pmcs[2] = mfspr(SPRN_PMC3);
666 	pmcs[3] = mfspr(SPRN_PMC4);
667 	pmcs[4] = mfspr(SPRN_PMC5);
668 	pmcs[5] = mfspr(SPRN_PMC6);
669 
670 	/* Ensure all freeze bits are unset */
671 	mtspr(SPRN_MMCR2, 0);
672 
673 	/* Set up PMC6 to overflow in one cycle */
674 	mtspr(SPRN_PMC6, 0x7FFFFFFE);
675 
676 	/* Enable exceptions and unfreeze PMC6 */
677 	mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
678 
679 	/* Now we need to refreeze and restore the PMCs */
680 	mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
681 
682 	mtspr(SPRN_PMC1, pmcs[0]);
683 	mtspr(SPRN_PMC2, pmcs[1]);
684 	mtspr(SPRN_PMC3, pmcs[2]);
685 	mtspr(SPRN_PMC4, pmcs[3]);
686 	mtspr(SPRN_PMC5, pmcs[4]);
687 	mtspr(SPRN_PMC6, pmcs[5]);
688 }
689 #endif /* CONFIG_PPC64 */
690 
691 static void perf_event_interrupt(struct pt_regs *regs);
692 
693 /*
694  * Read one performance monitor counter (PMC).
695  */
696 static unsigned long read_pmc(int idx)
697 {
698 	unsigned long val;
699 
700 	switch (idx) {
701 	case 1:
702 		val = mfspr(SPRN_PMC1);
703 		break;
704 	case 2:
705 		val = mfspr(SPRN_PMC2);
706 		break;
707 	case 3:
708 		val = mfspr(SPRN_PMC3);
709 		break;
710 	case 4:
711 		val = mfspr(SPRN_PMC4);
712 		break;
713 	case 5:
714 		val = mfspr(SPRN_PMC5);
715 		break;
716 	case 6:
717 		val = mfspr(SPRN_PMC6);
718 		break;
719 #ifdef CONFIG_PPC64
720 	case 7:
721 		val = mfspr(SPRN_PMC7);
722 		break;
723 	case 8:
724 		val = mfspr(SPRN_PMC8);
725 		break;
726 #endif /* CONFIG_PPC64 */
727 	default:
728 		printk(KERN_ERR "oops trying to read PMC%d\n", idx);
729 		val = 0;
730 	}
731 	return val;
732 }
733 
734 /*
735  * Write one PMC.
736  */
737 static void write_pmc(int idx, unsigned long val)
738 {
739 	switch (idx) {
740 	case 1:
741 		mtspr(SPRN_PMC1, val);
742 		break;
743 	case 2:
744 		mtspr(SPRN_PMC2, val);
745 		break;
746 	case 3:
747 		mtspr(SPRN_PMC3, val);
748 		break;
749 	case 4:
750 		mtspr(SPRN_PMC4, val);
751 		break;
752 	case 5:
753 		mtspr(SPRN_PMC5, val);
754 		break;
755 	case 6:
756 		mtspr(SPRN_PMC6, val);
757 		break;
758 #ifdef CONFIG_PPC64
759 	case 7:
760 		mtspr(SPRN_PMC7, val);
761 		break;
762 	case 8:
763 		mtspr(SPRN_PMC8, val);
764 		break;
765 #endif /* CONFIG_PPC64 */
766 	default:
767 		printk(KERN_ERR "oops trying to write PMC%d\n", idx);
768 	}
769 }
770 
771 /* Called from sysrq_handle_showregs() */
772 void perf_event_print_debug(void)
773 {
774 	unsigned long sdar, sier, flags;
775 	u32 pmcs[MAX_HWEVENTS];
776 	int i;
777 
778 	if (!ppmu->n_counter)
779 		return;
780 
781 	local_irq_save(flags);
782 
783 	pr_info("CPU: %d PMU registers, ppmu = %s n_counters = %d",
784 		 smp_processor_id(), ppmu->name, ppmu->n_counter);
785 
786 	for (i = 0; i < ppmu->n_counter; i++)
787 		pmcs[i] = read_pmc(i + 1);
788 
789 	for (; i < MAX_HWEVENTS; i++)
790 		pmcs[i] = 0xdeadbeef;
791 
792 	pr_info("PMC1:  %08x PMC2: %08x PMC3: %08x PMC4: %08x\n",
793 		 pmcs[0], pmcs[1], pmcs[2], pmcs[3]);
794 
795 	if (ppmu->n_counter > 4)
796 		pr_info("PMC5:  %08x PMC6: %08x PMC7: %08x PMC8: %08x\n",
797 			 pmcs[4], pmcs[5], pmcs[6], pmcs[7]);
798 
799 	pr_info("MMCR0: %016lx MMCR1: %016lx MMCRA: %016lx\n",
800 		mfspr(SPRN_MMCR0), mfspr(SPRN_MMCR1), mfspr(SPRN_MMCRA));
801 
802 	sdar = sier = 0;
803 #ifdef CONFIG_PPC64
804 	sdar = mfspr(SPRN_SDAR);
805 
806 	if (ppmu->flags & PPMU_HAS_SIER)
807 		sier = mfspr(SPRN_SIER);
808 
809 	if (ppmu->flags & PPMU_ARCH_207S) {
810 		pr_info("MMCR2: %016lx EBBHR: %016lx\n",
811 			mfspr(SPRN_MMCR2), mfspr(SPRN_EBBHR));
812 		pr_info("EBBRR: %016lx BESCR: %016lx\n",
813 			mfspr(SPRN_EBBRR), mfspr(SPRN_BESCR));
814 	}
815 #endif
816 	pr_info("SIAR:  %016lx SDAR:  %016lx SIER:  %016lx\n",
817 		mfspr(SPRN_SIAR), sdar, sier);
818 
819 	local_irq_restore(flags);
820 }
821 
822 /*
823  * Check if a set of events can all go on the PMU at once.
824  * If they can't, this will look at alternative codes for the events
825  * and see if any combination of alternative codes is feasible.
826  * The feasible set is returned in event_id[].
827  */
828 static int power_check_constraints(struct cpu_hw_events *cpuhw,
829 				   u64 event_id[], unsigned int cflags[],
830 				   int n_ev)
831 {
832 	unsigned long mask, value, nv;
833 	unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
834 	int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
835 	int i, j;
836 	unsigned long addf = ppmu->add_fields;
837 	unsigned long tadd = ppmu->test_adder;
838 
839 	if (n_ev > ppmu->n_counter)
840 		return -1;
841 
842 	/* First see if the events will go on as-is */
843 	for (i = 0; i < n_ev; ++i) {
844 		if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
845 		    && !ppmu->limited_pmc_event(event_id[i])) {
846 			ppmu->get_alternatives(event_id[i], cflags[i],
847 					       cpuhw->alternatives[i]);
848 			event_id[i] = cpuhw->alternatives[i][0];
849 		}
850 		if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
851 					 &cpuhw->avalues[i][0]))
852 			return -1;
853 	}
854 	value = mask = 0;
855 	for (i = 0; i < n_ev; ++i) {
856 		nv = (value | cpuhw->avalues[i][0]) +
857 			(value & cpuhw->avalues[i][0] & addf);
858 		if ((((nv + tadd) ^ value) & mask) != 0 ||
859 		    (((nv + tadd) ^ cpuhw->avalues[i][0]) &
860 		     cpuhw->amasks[i][0]) != 0)
861 			break;
862 		value = nv;
863 		mask |= cpuhw->amasks[i][0];
864 	}
865 	if (i == n_ev)
866 		return 0;	/* all OK */
867 
868 	/* doesn't work, gather alternatives... */
869 	if (!ppmu->get_alternatives)
870 		return -1;
871 	for (i = 0; i < n_ev; ++i) {
872 		choice[i] = 0;
873 		n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
874 						  cpuhw->alternatives[i]);
875 		for (j = 1; j < n_alt[i]; ++j)
876 			ppmu->get_constraint(cpuhw->alternatives[i][j],
877 					     &cpuhw->amasks[i][j],
878 					     &cpuhw->avalues[i][j]);
879 	}
880 
881 	/* enumerate all possibilities and see if any will work */
882 	i = 0;
883 	j = -1;
884 	value = mask = nv = 0;
885 	while (i < n_ev) {
886 		if (j >= 0) {
887 			/* we're backtracking, restore context */
888 			value = svalues[i];
889 			mask = smasks[i];
890 			j = choice[i];
891 		}
892 		/*
893 		 * See if any alternative k for event_id i,
894 		 * where k > j, will satisfy the constraints.
895 		 */
896 		while (++j < n_alt[i]) {
897 			nv = (value | cpuhw->avalues[i][j]) +
898 				(value & cpuhw->avalues[i][j] & addf);
899 			if ((((nv + tadd) ^ value) & mask) == 0 &&
900 			    (((nv + tadd) ^ cpuhw->avalues[i][j])
901 			     & cpuhw->amasks[i][j]) == 0)
902 				break;
903 		}
904 		if (j >= n_alt[i]) {
905 			/*
906 			 * No feasible alternative, backtrack
907 			 * to event_id i-1 and continue enumerating its
908 			 * alternatives from where we got up to.
909 			 */
910 			if (--i < 0)
911 				return -1;
912 		} else {
913 			/*
914 			 * Found a feasible alternative for event_id i,
915 			 * remember where we got up to with this event_id,
916 			 * go on to the next event_id, and start with
917 			 * the first alternative for it.
918 			 */
919 			choice[i] = j;
920 			svalues[i] = value;
921 			smasks[i] = mask;
922 			value = nv;
923 			mask |= cpuhw->amasks[i][j];
924 			++i;
925 			j = -1;
926 		}
927 	}
928 
929 	/* OK, we have a feasible combination, tell the caller the solution */
930 	for (i = 0; i < n_ev; ++i)
931 		event_id[i] = cpuhw->alternatives[i][choice[i]];
932 	return 0;
933 }
934 
935 /*
936  * Check if newly-added events have consistent settings for
937  * exclude_{user,kernel,hv} with each other and any previously
938  * added events.
939  */
940 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
941 			  int n_prev, int n_new)
942 {
943 	int eu = 0, ek = 0, eh = 0;
944 	int i, n, first;
945 	struct perf_event *event;
946 
947 	/*
948 	 * If the PMU we're on supports per event exclude settings then we
949 	 * don't need to do any of this logic. NB. This assumes no PMU has both
950 	 * per event exclude and limited PMCs.
951 	 */
952 	if (ppmu->flags & PPMU_ARCH_207S)
953 		return 0;
954 
955 	n = n_prev + n_new;
956 	if (n <= 1)
957 		return 0;
958 
959 	first = 1;
960 	for (i = 0; i < n; ++i) {
961 		if (cflags[i] & PPMU_LIMITED_PMC_OK) {
962 			cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
963 			continue;
964 		}
965 		event = ctrs[i];
966 		if (first) {
967 			eu = event->attr.exclude_user;
968 			ek = event->attr.exclude_kernel;
969 			eh = event->attr.exclude_hv;
970 			first = 0;
971 		} else if (event->attr.exclude_user != eu ||
972 			   event->attr.exclude_kernel != ek ||
973 			   event->attr.exclude_hv != eh) {
974 			return -EAGAIN;
975 		}
976 	}
977 
978 	if (eu || ek || eh)
979 		for (i = 0; i < n; ++i)
980 			if (cflags[i] & PPMU_LIMITED_PMC_OK)
981 				cflags[i] |= PPMU_LIMITED_PMC_REQD;
982 
983 	return 0;
984 }
985 
986 static u64 check_and_compute_delta(u64 prev, u64 val)
987 {
988 	u64 delta = (val - prev) & 0xfffffffful;
989 
990 	/*
991 	 * POWER7 can roll back counter values, if the new value is smaller
992 	 * than the previous value it will cause the delta and the counter to
993 	 * have bogus values unless we rolled a counter over.  If a coutner is
994 	 * rolled back, it will be smaller, but within 256, which is the maximum
995 	 * number of events to rollback at once.  If we dectect a rollback
996 	 * return 0.  This can lead to a small lack of precision in the
997 	 * counters.
998 	 */
999 	if (prev > val && (prev - val) < 256)
1000 		delta = 0;
1001 
1002 	return delta;
1003 }
1004 
1005 static void power_pmu_read(struct perf_event *event)
1006 {
1007 	s64 val, delta, prev;
1008 
1009 	if (event->hw.state & PERF_HES_STOPPED)
1010 		return;
1011 
1012 	if (!event->hw.idx)
1013 		return;
1014 
1015 	if (is_ebb_event(event)) {
1016 		val = read_pmc(event->hw.idx);
1017 		local64_set(&event->hw.prev_count, val);
1018 		return;
1019 	}
1020 
1021 	/*
1022 	 * Performance monitor interrupts come even when interrupts
1023 	 * are soft-disabled, as long as interrupts are hard-enabled.
1024 	 * Therefore we treat them like NMIs.
1025 	 */
1026 	do {
1027 		prev = local64_read(&event->hw.prev_count);
1028 		barrier();
1029 		val = read_pmc(event->hw.idx);
1030 		delta = check_and_compute_delta(prev, val);
1031 		if (!delta)
1032 			return;
1033 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
1034 
1035 	local64_add(delta, &event->count);
1036 
1037 	/*
1038 	 * A number of places program the PMC with (0x80000000 - period_left).
1039 	 * We never want period_left to be less than 1 because we will program
1040 	 * the PMC with a value >= 0x800000000 and an edge detected PMC will
1041 	 * roll around to 0 before taking an exception. We have seen this
1042 	 * on POWER8.
1043 	 *
1044 	 * To fix this, clamp the minimum value of period_left to 1.
1045 	 */
1046 	do {
1047 		prev = local64_read(&event->hw.period_left);
1048 		val = prev - delta;
1049 		if (val < 1)
1050 			val = 1;
1051 	} while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
1052 }
1053 
1054 /*
1055  * On some machines, PMC5 and PMC6 can't be written, don't respect
1056  * the freeze conditions, and don't generate interrupts.  This tells
1057  * us if `event' is using such a PMC.
1058  */
1059 static int is_limited_pmc(int pmcnum)
1060 {
1061 	return (ppmu->flags & PPMU_LIMITED_PMC5_6)
1062 		&& (pmcnum == 5 || pmcnum == 6);
1063 }
1064 
1065 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
1066 				    unsigned long pmc5, unsigned long pmc6)
1067 {
1068 	struct perf_event *event;
1069 	u64 val, prev, delta;
1070 	int i;
1071 
1072 	for (i = 0; i < cpuhw->n_limited; ++i) {
1073 		event = cpuhw->limited_counter[i];
1074 		if (!event->hw.idx)
1075 			continue;
1076 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1077 		prev = local64_read(&event->hw.prev_count);
1078 		event->hw.idx = 0;
1079 		delta = check_and_compute_delta(prev, val);
1080 		if (delta)
1081 			local64_add(delta, &event->count);
1082 	}
1083 }
1084 
1085 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
1086 				  unsigned long pmc5, unsigned long pmc6)
1087 {
1088 	struct perf_event *event;
1089 	u64 val, prev;
1090 	int i;
1091 
1092 	for (i = 0; i < cpuhw->n_limited; ++i) {
1093 		event = cpuhw->limited_counter[i];
1094 		event->hw.idx = cpuhw->limited_hwidx[i];
1095 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1096 		prev = local64_read(&event->hw.prev_count);
1097 		if (check_and_compute_delta(prev, val))
1098 			local64_set(&event->hw.prev_count, val);
1099 		perf_event_update_userpage(event);
1100 	}
1101 }
1102 
1103 /*
1104  * Since limited events don't respect the freeze conditions, we
1105  * have to read them immediately after freezing or unfreezing the
1106  * other events.  We try to keep the values from the limited
1107  * events as consistent as possible by keeping the delay (in
1108  * cycles and instructions) between freezing/unfreezing and reading
1109  * the limited events as small and consistent as possible.
1110  * Therefore, if any limited events are in use, we read them
1111  * both, and always in the same order, to minimize variability,
1112  * and do it inside the same asm that writes MMCR0.
1113  */
1114 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
1115 {
1116 	unsigned long pmc5, pmc6;
1117 
1118 	if (!cpuhw->n_limited) {
1119 		mtspr(SPRN_MMCR0, mmcr0);
1120 		return;
1121 	}
1122 
1123 	/*
1124 	 * Write MMCR0, then read PMC5 and PMC6 immediately.
1125 	 * To ensure we don't get a performance monitor interrupt
1126 	 * between writing MMCR0 and freezing/thawing the limited
1127 	 * events, we first write MMCR0 with the event overflow
1128 	 * interrupt enable bits turned off.
1129 	 */
1130 	asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
1131 		     : "=&r" (pmc5), "=&r" (pmc6)
1132 		     : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
1133 		       "i" (SPRN_MMCR0),
1134 		       "i" (SPRN_PMC5), "i" (SPRN_PMC6));
1135 
1136 	if (mmcr0 & MMCR0_FC)
1137 		freeze_limited_counters(cpuhw, pmc5, pmc6);
1138 	else
1139 		thaw_limited_counters(cpuhw, pmc5, pmc6);
1140 
1141 	/*
1142 	 * Write the full MMCR0 including the event overflow interrupt
1143 	 * enable bits, if necessary.
1144 	 */
1145 	if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
1146 		mtspr(SPRN_MMCR0, mmcr0);
1147 }
1148 
1149 /*
1150  * Disable all events to prevent PMU interrupts and to allow
1151  * events to be added or removed.
1152  */
1153 static void power_pmu_disable(struct pmu *pmu)
1154 {
1155 	struct cpu_hw_events *cpuhw;
1156 	unsigned long flags, mmcr0, val;
1157 
1158 	if (!ppmu)
1159 		return;
1160 	local_irq_save(flags);
1161 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1162 
1163 	if (!cpuhw->disabled) {
1164 		/*
1165 		 * Check if we ever enabled the PMU on this cpu.
1166 		 */
1167 		if (!cpuhw->pmcs_enabled) {
1168 			ppc_enable_pmcs();
1169 			cpuhw->pmcs_enabled = 1;
1170 		}
1171 
1172 		/*
1173 		 * Set the 'freeze counters' bit, clear EBE/BHRBA/PMCC/PMAO/FC56
1174 		 */
1175 		val  = mmcr0 = mfspr(SPRN_MMCR0);
1176 		val |= MMCR0_FC;
1177 		val &= ~(MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC | MMCR0_PMAO |
1178 			 MMCR0_FC56);
1179 
1180 		/*
1181 		 * The barrier is to make sure the mtspr has been
1182 		 * executed and the PMU has frozen the events etc.
1183 		 * before we return.
1184 		 */
1185 		write_mmcr0(cpuhw, val);
1186 		mb();
1187 
1188 		/*
1189 		 * Disable instruction sampling if it was enabled
1190 		 */
1191 		if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1192 			mtspr(SPRN_MMCRA,
1193 			      cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1194 			mb();
1195 		}
1196 
1197 		cpuhw->disabled = 1;
1198 		cpuhw->n_added = 0;
1199 
1200 		ebb_switch_out(mmcr0);
1201 	}
1202 
1203 	local_irq_restore(flags);
1204 }
1205 
1206 /*
1207  * Re-enable all events if disable == 0.
1208  * If we were previously disabled and events were added, then
1209  * put the new config on the PMU.
1210  */
1211 static void power_pmu_enable(struct pmu *pmu)
1212 {
1213 	struct perf_event *event;
1214 	struct cpu_hw_events *cpuhw;
1215 	unsigned long flags;
1216 	long i;
1217 	unsigned long val, mmcr0;
1218 	s64 left;
1219 	unsigned int hwc_index[MAX_HWEVENTS];
1220 	int n_lim;
1221 	int idx;
1222 	bool ebb;
1223 
1224 	if (!ppmu)
1225 		return;
1226 	local_irq_save(flags);
1227 
1228 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1229 	if (!cpuhw->disabled)
1230 		goto out;
1231 
1232 	if (cpuhw->n_events == 0) {
1233 		ppc_set_pmu_inuse(0);
1234 		goto out;
1235 	}
1236 
1237 	cpuhw->disabled = 0;
1238 
1239 	/*
1240 	 * EBB requires an exclusive group and all events must have the EBB
1241 	 * flag set, or not set, so we can just check a single event. Also we
1242 	 * know we have at least one event.
1243 	 */
1244 	ebb = is_ebb_event(cpuhw->event[0]);
1245 
1246 	/*
1247 	 * If we didn't change anything, or only removed events,
1248 	 * no need to recalculate MMCR* settings and reset the PMCs.
1249 	 * Just reenable the PMU with the current MMCR* settings
1250 	 * (possibly updated for removal of events).
1251 	 */
1252 	if (!cpuhw->n_added) {
1253 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1254 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1255 		goto out_enable;
1256 	}
1257 
1258 	/*
1259 	 * Clear all MMCR settings and recompute them for the new set of events.
1260 	 */
1261 	memset(cpuhw->mmcr, 0, sizeof(cpuhw->mmcr));
1262 
1263 	if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
1264 			       cpuhw->mmcr, cpuhw->event)) {
1265 		/* shouldn't ever get here */
1266 		printk(KERN_ERR "oops compute_mmcr failed\n");
1267 		goto out;
1268 	}
1269 
1270 	if (!(ppmu->flags & PPMU_ARCH_207S)) {
1271 		/*
1272 		 * Add in MMCR0 freeze bits corresponding to the attr.exclude_*
1273 		 * bits for the first event. We have already checked that all
1274 		 * events have the same value for these bits as the first event.
1275 		 */
1276 		event = cpuhw->event[0];
1277 		if (event->attr.exclude_user)
1278 			cpuhw->mmcr[0] |= MMCR0_FCP;
1279 		if (event->attr.exclude_kernel)
1280 			cpuhw->mmcr[0] |= freeze_events_kernel;
1281 		if (event->attr.exclude_hv)
1282 			cpuhw->mmcr[0] |= MMCR0_FCHV;
1283 	}
1284 
1285 	/*
1286 	 * Write the new configuration to MMCR* with the freeze
1287 	 * bit set and set the hardware events to their initial values.
1288 	 * Then unfreeze the events.
1289 	 */
1290 	ppc_set_pmu_inuse(1);
1291 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1292 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1293 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
1294 				| MMCR0_FC);
1295 	if (ppmu->flags & PPMU_ARCH_207S)
1296 		mtspr(SPRN_MMCR2, cpuhw->mmcr[3]);
1297 
1298 	/*
1299 	 * Read off any pre-existing events that need to move
1300 	 * to another PMC.
1301 	 */
1302 	for (i = 0; i < cpuhw->n_events; ++i) {
1303 		event = cpuhw->event[i];
1304 		if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
1305 			power_pmu_read(event);
1306 			write_pmc(event->hw.idx, 0);
1307 			event->hw.idx = 0;
1308 		}
1309 	}
1310 
1311 	/*
1312 	 * Initialize the PMCs for all the new and moved events.
1313 	 */
1314 	cpuhw->n_limited = n_lim = 0;
1315 	for (i = 0; i < cpuhw->n_events; ++i) {
1316 		event = cpuhw->event[i];
1317 		if (event->hw.idx)
1318 			continue;
1319 		idx = hwc_index[i] + 1;
1320 		if (is_limited_pmc(idx)) {
1321 			cpuhw->limited_counter[n_lim] = event;
1322 			cpuhw->limited_hwidx[n_lim] = idx;
1323 			++n_lim;
1324 			continue;
1325 		}
1326 
1327 		if (ebb)
1328 			val = local64_read(&event->hw.prev_count);
1329 		else {
1330 			val = 0;
1331 			if (event->hw.sample_period) {
1332 				left = local64_read(&event->hw.period_left);
1333 				if (left < 0x80000000L)
1334 					val = 0x80000000L - left;
1335 			}
1336 			local64_set(&event->hw.prev_count, val);
1337 		}
1338 
1339 		event->hw.idx = idx;
1340 		if (event->hw.state & PERF_HES_STOPPED)
1341 			val = 0;
1342 		write_pmc(idx, val);
1343 
1344 		perf_event_update_userpage(event);
1345 	}
1346 	cpuhw->n_limited = n_lim;
1347 	cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
1348 
1349  out_enable:
1350 	pmao_restore_workaround(ebb);
1351 
1352 	mmcr0 = ebb_switch_in(ebb, cpuhw);
1353 
1354 	mb();
1355 	if (cpuhw->bhrb_users)
1356 		ppmu->config_bhrb(cpuhw->bhrb_filter);
1357 
1358 	write_mmcr0(cpuhw, mmcr0);
1359 
1360 	/*
1361 	 * Enable instruction sampling if necessary
1362 	 */
1363 	if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1364 		mb();
1365 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
1366 	}
1367 
1368  out:
1369 
1370 	local_irq_restore(flags);
1371 }
1372 
1373 static int collect_events(struct perf_event *group, int max_count,
1374 			  struct perf_event *ctrs[], u64 *events,
1375 			  unsigned int *flags)
1376 {
1377 	int n = 0;
1378 	struct perf_event *event;
1379 
1380 	if (!is_software_event(group)) {
1381 		if (n >= max_count)
1382 			return -1;
1383 		ctrs[n] = group;
1384 		flags[n] = group->hw.event_base;
1385 		events[n++] = group->hw.config;
1386 	}
1387 	list_for_each_entry(event, &group->sibling_list, group_entry) {
1388 		if (!is_software_event(event) &&
1389 		    event->state != PERF_EVENT_STATE_OFF) {
1390 			if (n >= max_count)
1391 				return -1;
1392 			ctrs[n] = event;
1393 			flags[n] = event->hw.event_base;
1394 			events[n++] = event->hw.config;
1395 		}
1396 	}
1397 	return n;
1398 }
1399 
1400 /*
1401  * Add a event to the PMU.
1402  * If all events are not already frozen, then we disable and
1403  * re-enable the PMU in order to get hw_perf_enable to do the
1404  * actual work of reconfiguring the PMU.
1405  */
1406 static int power_pmu_add(struct perf_event *event, int ef_flags)
1407 {
1408 	struct cpu_hw_events *cpuhw;
1409 	unsigned long flags;
1410 	int n0;
1411 	int ret = -EAGAIN;
1412 
1413 	local_irq_save(flags);
1414 	perf_pmu_disable(event->pmu);
1415 
1416 	/*
1417 	 * Add the event to the list (if there is room)
1418 	 * and check whether the total set is still feasible.
1419 	 */
1420 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1421 	n0 = cpuhw->n_events;
1422 	if (n0 >= ppmu->n_counter)
1423 		goto out;
1424 	cpuhw->event[n0] = event;
1425 	cpuhw->events[n0] = event->hw.config;
1426 	cpuhw->flags[n0] = event->hw.event_base;
1427 
1428 	/*
1429 	 * This event may have been disabled/stopped in record_and_restart()
1430 	 * because we exceeded the ->event_limit. If re-starting the event,
1431 	 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
1432 	 * notification is re-enabled.
1433 	 */
1434 	if (!(ef_flags & PERF_EF_START))
1435 		event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1436 	else
1437 		event->hw.state = 0;
1438 
1439 	/*
1440 	 * If group events scheduling transaction was started,
1441 	 * skip the schedulability test here, it will be performed
1442 	 * at commit time(->commit_txn) as a whole
1443 	 */
1444 	if (cpuhw->group_flag & PERF_EVENT_TXN)
1445 		goto nocheck;
1446 
1447 	if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
1448 		goto out;
1449 	if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
1450 		goto out;
1451 	event->hw.config = cpuhw->events[n0];
1452 
1453 nocheck:
1454 	ebb_event_add(event);
1455 
1456 	++cpuhw->n_events;
1457 	++cpuhw->n_added;
1458 
1459 	ret = 0;
1460  out:
1461 	if (has_branch_stack(event)) {
1462 		power_pmu_bhrb_enable(event);
1463 		cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1464 					event->attr.branch_sample_type);
1465 	}
1466 
1467 	perf_pmu_enable(event->pmu);
1468 	local_irq_restore(flags);
1469 	return ret;
1470 }
1471 
1472 /*
1473  * Remove a event from the PMU.
1474  */
1475 static void power_pmu_del(struct perf_event *event, int ef_flags)
1476 {
1477 	struct cpu_hw_events *cpuhw;
1478 	long i;
1479 	unsigned long flags;
1480 
1481 	local_irq_save(flags);
1482 	perf_pmu_disable(event->pmu);
1483 
1484 	power_pmu_read(event);
1485 
1486 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1487 	for (i = 0; i < cpuhw->n_events; ++i) {
1488 		if (event == cpuhw->event[i]) {
1489 			while (++i < cpuhw->n_events) {
1490 				cpuhw->event[i-1] = cpuhw->event[i];
1491 				cpuhw->events[i-1] = cpuhw->events[i];
1492 				cpuhw->flags[i-1] = cpuhw->flags[i];
1493 			}
1494 			--cpuhw->n_events;
1495 			ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
1496 			if (event->hw.idx) {
1497 				write_pmc(event->hw.idx, 0);
1498 				event->hw.idx = 0;
1499 			}
1500 			perf_event_update_userpage(event);
1501 			break;
1502 		}
1503 	}
1504 	for (i = 0; i < cpuhw->n_limited; ++i)
1505 		if (event == cpuhw->limited_counter[i])
1506 			break;
1507 	if (i < cpuhw->n_limited) {
1508 		while (++i < cpuhw->n_limited) {
1509 			cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
1510 			cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
1511 		}
1512 		--cpuhw->n_limited;
1513 	}
1514 	if (cpuhw->n_events == 0) {
1515 		/* disable exceptions if no events are running */
1516 		cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
1517 	}
1518 
1519 	if (has_branch_stack(event))
1520 		power_pmu_bhrb_disable(event);
1521 
1522 	perf_pmu_enable(event->pmu);
1523 	local_irq_restore(flags);
1524 }
1525 
1526 /*
1527  * POWER-PMU does not support disabling individual counters, hence
1528  * program their cycle counter to their max value and ignore the interrupts.
1529  */
1530 
1531 static void power_pmu_start(struct perf_event *event, int ef_flags)
1532 {
1533 	unsigned long flags;
1534 	s64 left;
1535 	unsigned long val;
1536 
1537 	if (!event->hw.idx || !event->hw.sample_period)
1538 		return;
1539 
1540 	if (!(event->hw.state & PERF_HES_STOPPED))
1541 		return;
1542 
1543 	if (ef_flags & PERF_EF_RELOAD)
1544 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1545 
1546 	local_irq_save(flags);
1547 	perf_pmu_disable(event->pmu);
1548 
1549 	event->hw.state = 0;
1550 	left = local64_read(&event->hw.period_left);
1551 
1552 	val = 0;
1553 	if (left < 0x80000000L)
1554 		val = 0x80000000L - left;
1555 
1556 	write_pmc(event->hw.idx, val);
1557 
1558 	perf_event_update_userpage(event);
1559 	perf_pmu_enable(event->pmu);
1560 	local_irq_restore(flags);
1561 }
1562 
1563 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1564 {
1565 	unsigned long flags;
1566 
1567 	if (!event->hw.idx || !event->hw.sample_period)
1568 		return;
1569 
1570 	if (event->hw.state & PERF_HES_STOPPED)
1571 		return;
1572 
1573 	local_irq_save(flags);
1574 	perf_pmu_disable(event->pmu);
1575 
1576 	power_pmu_read(event);
1577 	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1578 	write_pmc(event->hw.idx, 0);
1579 
1580 	perf_event_update_userpage(event);
1581 	perf_pmu_enable(event->pmu);
1582 	local_irq_restore(flags);
1583 }
1584 
1585 /*
1586  * Start group events scheduling transaction
1587  * Set the flag to make pmu::enable() not perform the
1588  * schedulability test, it will be performed at commit time
1589  */
1590 static void power_pmu_start_txn(struct pmu *pmu)
1591 {
1592 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1593 
1594 	perf_pmu_disable(pmu);
1595 	cpuhw->group_flag |= PERF_EVENT_TXN;
1596 	cpuhw->n_txn_start = cpuhw->n_events;
1597 }
1598 
1599 /*
1600  * Stop group events scheduling transaction
1601  * Clear the flag and pmu::enable() will perform the
1602  * schedulability test.
1603  */
1604 static void power_pmu_cancel_txn(struct pmu *pmu)
1605 {
1606 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1607 
1608 	cpuhw->group_flag &= ~PERF_EVENT_TXN;
1609 	perf_pmu_enable(pmu);
1610 }
1611 
1612 /*
1613  * Commit group events scheduling transaction
1614  * Perform the group schedulability test as a whole
1615  * Return 0 if success
1616  */
1617 static int power_pmu_commit_txn(struct pmu *pmu)
1618 {
1619 	struct cpu_hw_events *cpuhw;
1620 	long i, n;
1621 
1622 	if (!ppmu)
1623 		return -EAGAIN;
1624 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1625 	n = cpuhw->n_events;
1626 	if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1627 		return -EAGAIN;
1628 	i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1629 	if (i < 0)
1630 		return -EAGAIN;
1631 
1632 	for (i = cpuhw->n_txn_start; i < n; ++i)
1633 		cpuhw->event[i]->hw.config = cpuhw->events[i];
1634 
1635 	cpuhw->group_flag &= ~PERF_EVENT_TXN;
1636 	perf_pmu_enable(pmu);
1637 	return 0;
1638 }
1639 
1640 /*
1641  * Return 1 if we might be able to put event on a limited PMC,
1642  * or 0 if not.
1643  * A event can only go on a limited PMC if it counts something
1644  * that a limited PMC can count, doesn't require interrupts, and
1645  * doesn't exclude any processor mode.
1646  */
1647 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1648 				 unsigned int flags)
1649 {
1650 	int n;
1651 	u64 alt[MAX_EVENT_ALTERNATIVES];
1652 
1653 	if (event->attr.exclude_user
1654 	    || event->attr.exclude_kernel
1655 	    || event->attr.exclude_hv
1656 	    || event->attr.sample_period)
1657 		return 0;
1658 
1659 	if (ppmu->limited_pmc_event(ev))
1660 		return 1;
1661 
1662 	/*
1663 	 * The requested event_id isn't on a limited PMC already;
1664 	 * see if any alternative code goes on a limited PMC.
1665 	 */
1666 	if (!ppmu->get_alternatives)
1667 		return 0;
1668 
1669 	flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1670 	n = ppmu->get_alternatives(ev, flags, alt);
1671 
1672 	return n > 0;
1673 }
1674 
1675 /*
1676  * Find an alternative event_id that goes on a normal PMC, if possible,
1677  * and return the event_id code, or 0 if there is no such alternative.
1678  * (Note: event_id code 0 is "don't count" on all machines.)
1679  */
1680 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1681 {
1682 	u64 alt[MAX_EVENT_ALTERNATIVES];
1683 	int n;
1684 
1685 	flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1686 	n = ppmu->get_alternatives(ev, flags, alt);
1687 	if (!n)
1688 		return 0;
1689 	return alt[0];
1690 }
1691 
1692 /* Number of perf_events counting hardware events */
1693 static atomic_t num_events;
1694 /* Used to avoid races in calling reserve/release_pmc_hardware */
1695 static DEFINE_MUTEX(pmc_reserve_mutex);
1696 
1697 /*
1698  * Release the PMU if this is the last perf_event.
1699  */
1700 static void hw_perf_event_destroy(struct perf_event *event)
1701 {
1702 	if (!atomic_add_unless(&num_events, -1, 1)) {
1703 		mutex_lock(&pmc_reserve_mutex);
1704 		if (atomic_dec_return(&num_events) == 0)
1705 			release_pmc_hardware();
1706 		mutex_unlock(&pmc_reserve_mutex);
1707 	}
1708 }
1709 
1710 /*
1711  * Translate a generic cache event_id config to a raw event_id code.
1712  */
1713 static int hw_perf_cache_event(u64 config, u64 *eventp)
1714 {
1715 	unsigned long type, op, result;
1716 	int ev;
1717 
1718 	if (!ppmu->cache_events)
1719 		return -EINVAL;
1720 
1721 	/* unpack config */
1722 	type = config & 0xff;
1723 	op = (config >> 8) & 0xff;
1724 	result = (config >> 16) & 0xff;
1725 
1726 	if (type >= PERF_COUNT_HW_CACHE_MAX ||
1727 	    op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1728 	    result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1729 		return -EINVAL;
1730 
1731 	ev = (*ppmu->cache_events)[type][op][result];
1732 	if (ev == 0)
1733 		return -EOPNOTSUPP;
1734 	if (ev == -1)
1735 		return -EINVAL;
1736 	*eventp = ev;
1737 	return 0;
1738 }
1739 
1740 static int power_pmu_event_init(struct perf_event *event)
1741 {
1742 	u64 ev;
1743 	unsigned long flags;
1744 	struct perf_event *ctrs[MAX_HWEVENTS];
1745 	u64 events[MAX_HWEVENTS];
1746 	unsigned int cflags[MAX_HWEVENTS];
1747 	int n;
1748 	int err;
1749 	struct cpu_hw_events *cpuhw;
1750 
1751 	if (!ppmu)
1752 		return -ENOENT;
1753 
1754 	if (has_branch_stack(event)) {
1755 	        /* PMU has BHRB enabled */
1756 		if (!(ppmu->flags & PPMU_ARCH_207S))
1757 			return -EOPNOTSUPP;
1758 	}
1759 
1760 	switch (event->attr.type) {
1761 	case PERF_TYPE_HARDWARE:
1762 		ev = event->attr.config;
1763 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1764 			return -EOPNOTSUPP;
1765 		ev = ppmu->generic_events[ev];
1766 		break;
1767 	case PERF_TYPE_HW_CACHE:
1768 		err = hw_perf_cache_event(event->attr.config, &ev);
1769 		if (err)
1770 			return err;
1771 		break;
1772 	case PERF_TYPE_RAW:
1773 		ev = event->attr.config;
1774 		break;
1775 	default:
1776 		return -ENOENT;
1777 	}
1778 
1779 	event->hw.config_base = ev;
1780 	event->hw.idx = 0;
1781 
1782 	/*
1783 	 * If we are not running on a hypervisor, force the
1784 	 * exclude_hv bit to 0 so that we don't care what
1785 	 * the user set it to.
1786 	 */
1787 	if (!firmware_has_feature(FW_FEATURE_LPAR))
1788 		event->attr.exclude_hv = 0;
1789 
1790 	/*
1791 	 * If this is a per-task event, then we can use
1792 	 * PM_RUN_* events interchangeably with their non RUN_*
1793 	 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1794 	 * XXX we should check if the task is an idle task.
1795 	 */
1796 	flags = 0;
1797 	if (event->attach_state & PERF_ATTACH_TASK)
1798 		flags |= PPMU_ONLY_COUNT_RUN;
1799 
1800 	/*
1801 	 * If this machine has limited events, check whether this
1802 	 * event_id could go on a limited event.
1803 	 */
1804 	if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1805 		if (can_go_on_limited_pmc(event, ev, flags)) {
1806 			flags |= PPMU_LIMITED_PMC_OK;
1807 		} else if (ppmu->limited_pmc_event(ev)) {
1808 			/*
1809 			 * The requested event_id is on a limited PMC,
1810 			 * but we can't use a limited PMC; see if any
1811 			 * alternative goes on a normal PMC.
1812 			 */
1813 			ev = normal_pmc_alternative(ev, flags);
1814 			if (!ev)
1815 				return -EINVAL;
1816 		}
1817 	}
1818 
1819 	/* Extra checks for EBB */
1820 	err = ebb_event_check(event);
1821 	if (err)
1822 		return err;
1823 
1824 	/*
1825 	 * If this is in a group, check if it can go on with all the
1826 	 * other hardware events in the group.  We assume the event
1827 	 * hasn't been linked into its leader's sibling list at this point.
1828 	 */
1829 	n = 0;
1830 	if (event->group_leader != event) {
1831 		n = collect_events(event->group_leader, ppmu->n_counter - 1,
1832 				   ctrs, events, cflags);
1833 		if (n < 0)
1834 			return -EINVAL;
1835 	}
1836 	events[n] = ev;
1837 	ctrs[n] = event;
1838 	cflags[n] = flags;
1839 	if (check_excludes(ctrs, cflags, n, 1))
1840 		return -EINVAL;
1841 
1842 	cpuhw = &get_cpu_var(cpu_hw_events);
1843 	err = power_check_constraints(cpuhw, events, cflags, n + 1);
1844 
1845 	if (has_branch_stack(event)) {
1846 		cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1847 					event->attr.branch_sample_type);
1848 
1849 		if (cpuhw->bhrb_filter == -1) {
1850 			put_cpu_var(cpu_hw_events);
1851 			return -EOPNOTSUPP;
1852 		}
1853 	}
1854 
1855 	put_cpu_var(cpu_hw_events);
1856 	if (err)
1857 		return -EINVAL;
1858 
1859 	event->hw.config = events[n];
1860 	event->hw.event_base = cflags[n];
1861 	event->hw.last_period = event->hw.sample_period;
1862 	local64_set(&event->hw.period_left, event->hw.last_period);
1863 
1864 	/*
1865 	 * For EBB events we just context switch the PMC value, we don't do any
1866 	 * of the sample_period logic. We use hw.prev_count for this.
1867 	 */
1868 	if (is_ebb_event(event))
1869 		local64_set(&event->hw.prev_count, 0);
1870 
1871 	/*
1872 	 * See if we need to reserve the PMU.
1873 	 * If no events are currently in use, then we have to take a
1874 	 * mutex to ensure that we don't race with another task doing
1875 	 * reserve_pmc_hardware or release_pmc_hardware.
1876 	 */
1877 	err = 0;
1878 	if (!atomic_inc_not_zero(&num_events)) {
1879 		mutex_lock(&pmc_reserve_mutex);
1880 		if (atomic_read(&num_events) == 0 &&
1881 		    reserve_pmc_hardware(perf_event_interrupt))
1882 			err = -EBUSY;
1883 		else
1884 			atomic_inc(&num_events);
1885 		mutex_unlock(&pmc_reserve_mutex);
1886 	}
1887 	event->destroy = hw_perf_event_destroy;
1888 
1889 	return err;
1890 }
1891 
1892 static int power_pmu_event_idx(struct perf_event *event)
1893 {
1894 	return event->hw.idx;
1895 }
1896 
1897 ssize_t power_events_sysfs_show(struct device *dev,
1898 				struct device_attribute *attr, char *page)
1899 {
1900 	struct perf_pmu_events_attr *pmu_attr;
1901 
1902 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
1903 
1904 	return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
1905 }
1906 
1907 static struct pmu power_pmu = {
1908 	.pmu_enable	= power_pmu_enable,
1909 	.pmu_disable	= power_pmu_disable,
1910 	.event_init	= power_pmu_event_init,
1911 	.add		= power_pmu_add,
1912 	.del		= power_pmu_del,
1913 	.start		= power_pmu_start,
1914 	.stop		= power_pmu_stop,
1915 	.read		= power_pmu_read,
1916 	.start_txn	= power_pmu_start_txn,
1917 	.cancel_txn	= power_pmu_cancel_txn,
1918 	.commit_txn	= power_pmu_commit_txn,
1919 	.event_idx	= power_pmu_event_idx,
1920 	.sched_task	= power_pmu_sched_task,
1921 };
1922 
1923 /*
1924  * A counter has overflowed; update its count and record
1925  * things if requested.  Note that interrupts are hard-disabled
1926  * here so there is no possibility of being interrupted.
1927  */
1928 static void record_and_restart(struct perf_event *event, unsigned long val,
1929 			       struct pt_regs *regs)
1930 {
1931 	u64 period = event->hw.sample_period;
1932 	s64 prev, delta, left;
1933 	int record = 0;
1934 
1935 	if (event->hw.state & PERF_HES_STOPPED) {
1936 		write_pmc(event->hw.idx, 0);
1937 		return;
1938 	}
1939 
1940 	/* we don't have to worry about interrupts here */
1941 	prev = local64_read(&event->hw.prev_count);
1942 	delta = check_and_compute_delta(prev, val);
1943 	local64_add(delta, &event->count);
1944 
1945 	/*
1946 	 * See if the total period for this event has expired,
1947 	 * and update for the next period.
1948 	 */
1949 	val = 0;
1950 	left = local64_read(&event->hw.period_left) - delta;
1951 	if (delta == 0)
1952 		left++;
1953 	if (period) {
1954 		if (left <= 0) {
1955 			left += period;
1956 			if (left <= 0)
1957 				left = period;
1958 			record = siar_valid(regs);
1959 			event->hw.last_period = event->hw.sample_period;
1960 		}
1961 		if (left < 0x80000000LL)
1962 			val = 0x80000000LL - left;
1963 	}
1964 
1965 	write_pmc(event->hw.idx, val);
1966 	local64_set(&event->hw.prev_count, val);
1967 	local64_set(&event->hw.period_left, left);
1968 	perf_event_update_userpage(event);
1969 
1970 	/*
1971 	 * Finally record data if requested.
1972 	 */
1973 	if (record) {
1974 		struct perf_sample_data data;
1975 
1976 		perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
1977 
1978 		if (event->attr.sample_type & PERF_SAMPLE_ADDR)
1979 			perf_get_data_addr(regs, &data.addr);
1980 
1981 		if (event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK) {
1982 			struct cpu_hw_events *cpuhw;
1983 			cpuhw = this_cpu_ptr(&cpu_hw_events);
1984 			power_pmu_bhrb_read(cpuhw);
1985 			data.br_stack = &cpuhw->bhrb_stack;
1986 		}
1987 
1988 		if (perf_event_overflow(event, &data, regs))
1989 			power_pmu_stop(event, 0);
1990 	}
1991 }
1992 
1993 /*
1994  * Called from generic code to get the misc flags (i.e. processor mode)
1995  * for an event_id.
1996  */
1997 unsigned long perf_misc_flags(struct pt_regs *regs)
1998 {
1999 	u32 flags = perf_get_misc_flags(regs);
2000 
2001 	if (flags)
2002 		return flags;
2003 	return user_mode(regs) ? PERF_RECORD_MISC_USER :
2004 		PERF_RECORD_MISC_KERNEL;
2005 }
2006 
2007 /*
2008  * Called from generic code to get the instruction pointer
2009  * for an event_id.
2010  */
2011 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2012 {
2013 	bool use_siar = regs_use_siar(regs);
2014 
2015 	if (use_siar && siar_valid(regs))
2016 		return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
2017 	else if (use_siar)
2018 		return 0;		// no valid instruction pointer
2019 	else
2020 		return regs->nip;
2021 }
2022 
2023 static bool pmc_overflow_power7(unsigned long val)
2024 {
2025 	/*
2026 	 * Events on POWER7 can roll back if a speculative event doesn't
2027 	 * eventually complete. Unfortunately in some rare cases they will
2028 	 * raise a performance monitor exception. We need to catch this to
2029 	 * ensure we reset the PMC. In all cases the PMC will be 256 or less
2030 	 * cycles from overflow.
2031 	 *
2032 	 * We only do this if the first pass fails to find any overflowing
2033 	 * PMCs because a user might set a period of less than 256 and we
2034 	 * don't want to mistakenly reset them.
2035 	 */
2036 	if ((0x80000000 - val) <= 256)
2037 		return true;
2038 
2039 	return false;
2040 }
2041 
2042 static bool pmc_overflow(unsigned long val)
2043 {
2044 	if ((int)val < 0)
2045 		return true;
2046 
2047 	return false;
2048 }
2049 
2050 /*
2051  * Performance monitor interrupt stuff
2052  */
2053 static void perf_event_interrupt(struct pt_regs *regs)
2054 {
2055 	int i, j;
2056 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
2057 	struct perf_event *event;
2058 	unsigned long val[8];
2059 	int found, active;
2060 	int nmi;
2061 
2062 	if (cpuhw->n_limited)
2063 		freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
2064 					mfspr(SPRN_PMC6));
2065 
2066 	perf_read_regs(regs);
2067 
2068 	nmi = perf_intr_is_nmi(regs);
2069 	if (nmi)
2070 		nmi_enter();
2071 	else
2072 		irq_enter();
2073 
2074 	/* Read all the PMCs since we'll need them a bunch of times */
2075 	for (i = 0; i < ppmu->n_counter; ++i)
2076 		val[i] = read_pmc(i + 1);
2077 
2078 	/* Try to find what caused the IRQ */
2079 	found = 0;
2080 	for (i = 0; i < ppmu->n_counter; ++i) {
2081 		if (!pmc_overflow(val[i]))
2082 			continue;
2083 		if (is_limited_pmc(i + 1))
2084 			continue; /* these won't generate IRQs */
2085 		/*
2086 		 * We've found one that's overflowed.  For active
2087 		 * counters we need to log this.  For inactive
2088 		 * counters, we need to reset it anyway
2089 		 */
2090 		found = 1;
2091 		active = 0;
2092 		for (j = 0; j < cpuhw->n_events; ++j) {
2093 			event = cpuhw->event[j];
2094 			if (event->hw.idx == (i + 1)) {
2095 				active = 1;
2096 				record_and_restart(event, val[i], regs);
2097 				break;
2098 			}
2099 		}
2100 		if (!active)
2101 			/* reset non active counters that have overflowed */
2102 			write_pmc(i + 1, 0);
2103 	}
2104 	if (!found && pvr_version_is(PVR_POWER7)) {
2105 		/* check active counters for special buggy p7 overflow */
2106 		for (i = 0; i < cpuhw->n_events; ++i) {
2107 			event = cpuhw->event[i];
2108 			if (!event->hw.idx || is_limited_pmc(event->hw.idx))
2109 				continue;
2110 			if (pmc_overflow_power7(val[event->hw.idx - 1])) {
2111 				/* event has overflowed in a buggy way*/
2112 				found = 1;
2113 				record_and_restart(event,
2114 						   val[event->hw.idx - 1],
2115 						   regs);
2116 			}
2117 		}
2118 	}
2119 	if (!found && !nmi && printk_ratelimit())
2120 		printk(KERN_WARNING "Can't find PMC that caused IRQ\n");
2121 
2122 	/*
2123 	 * Reset MMCR0 to its normal value.  This will set PMXE and
2124 	 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
2125 	 * and thus allow interrupts to occur again.
2126 	 * XXX might want to use MSR.PM to keep the events frozen until
2127 	 * we get back out of this interrupt.
2128 	 */
2129 	write_mmcr0(cpuhw, cpuhw->mmcr[0]);
2130 
2131 	if (nmi)
2132 		nmi_exit();
2133 	else
2134 		irq_exit();
2135 }
2136 
2137 static void power_pmu_setup(int cpu)
2138 {
2139 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
2140 
2141 	if (!ppmu)
2142 		return;
2143 	memset(cpuhw, 0, sizeof(*cpuhw));
2144 	cpuhw->mmcr[0] = MMCR0_FC;
2145 }
2146 
2147 static int
2148 power_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
2149 {
2150 	unsigned int cpu = (long)hcpu;
2151 
2152 	switch (action & ~CPU_TASKS_FROZEN) {
2153 	case CPU_UP_PREPARE:
2154 		power_pmu_setup(cpu);
2155 		break;
2156 
2157 	default:
2158 		break;
2159 	}
2160 
2161 	return NOTIFY_OK;
2162 }
2163 
2164 int register_power_pmu(struct power_pmu *pmu)
2165 {
2166 	if (ppmu)
2167 		return -EBUSY;		/* something's already registered */
2168 
2169 	ppmu = pmu;
2170 	pr_info("%s performance monitor hardware support registered\n",
2171 		pmu->name);
2172 
2173 	power_pmu.attr_groups = ppmu->attr_groups;
2174 
2175 #ifdef MSR_HV
2176 	/*
2177 	 * Use FCHV to ignore kernel events if MSR.HV is set.
2178 	 */
2179 	if (mfmsr() & MSR_HV)
2180 		freeze_events_kernel = MMCR0_FCHV;
2181 #endif /* CONFIG_PPC64 */
2182 
2183 	perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
2184 	perf_cpu_notifier(power_pmu_notifier);
2185 
2186 	return 0;
2187 }
2188