xref: /openbmc/linux/arch/powerpc/perf/core-book3s.c (revision a9a08845)
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 if (ppmu->flags & PPMU_NO_SIAR)
192 			sdsync = MMCRA_SAMPLE_ENABLE;
193 		else
194 			sdsync = MMCRA_SDSYNC;
195 
196 		sdar_valid = mmcra & sdsync;
197 	}
198 
199 	if (!(mmcra & MMCRA_SAMPLE_ENABLE) || sdar_valid)
200 		*addrp = mfspr(SPRN_SDAR);
201 }
202 
203 static bool regs_sihv(struct pt_regs *regs)
204 {
205 	unsigned long sihv = MMCRA_SIHV;
206 
207 	if (ppmu->flags & PPMU_HAS_SIER)
208 		return !!(regs->dar & SIER_SIHV);
209 
210 	if (ppmu->flags & PPMU_ALT_SIPR)
211 		sihv = POWER6_MMCRA_SIHV;
212 
213 	return !!(regs->dsisr & sihv);
214 }
215 
216 static bool regs_sipr(struct pt_regs *regs)
217 {
218 	unsigned long sipr = MMCRA_SIPR;
219 
220 	if (ppmu->flags & PPMU_HAS_SIER)
221 		return !!(regs->dar & SIER_SIPR);
222 
223 	if (ppmu->flags & PPMU_ALT_SIPR)
224 		sipr = POWER6_MMCRA_SIPR;
225 
226 	return !!(regs->dsisr & sipr);
227 }
228 
229 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
230 {
231 	if (regs->msr & MSR_PR)
232 		return PERF_RECORD_MISC_USER;
233 	if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
234 		return PERF_RECORD_MISC_HYPERVISOR;
235 	return PERF_RECORD_MISC_KERNEL;
236 }
237 
238 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
239 {
240 	bool use_siar = regs_use_siar(regs);
241 
242 	if (!use_siar)
243 		return perf_flags_from_msr(regs);
244 
245 	/*
246 	 * If we don't have flags in MMCRA, rather than using
247 	 * the MSR, we intuit the flags from the address in
248 	 * SIAR which should give slightly more reliable
249 	 * results
250 	 */
251 	if (ppmu->flags & PPMU_NO_SIPR) {
252 		unsigned long siar = mfspr(SPRN_SIAR);
253 		if (is_kernel_addr(siar))
254 			return PERF_RECORD_MISC_KERNEL;
255 		return PERF_RECORD_MISC_USER;
256 	}
257 
258 	/* PR has priority over HV, so order below is important */
259 	if (regs_sipr(regs))
260 		return PERF_RECORD_MISC_USER;
261 
262 	if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
263 		return PERF_RECORD_MISC_HYPERVISOR;
264 
265 	return PERF_RECORD_MISC_KERNEL;
266 }
267 
268 /*
269  * Overload regs->dsisr to store MMCRA so we only need to read it once
270  * on each interrupt.
271  * Overload regs->dar to store SIER if we have it.
272  * Overload regs->result to specify whether we should use the MSR (result
273  * is zero) or the SIAR (result is non zero).
274  */
275 static inline void perf_read_regs(struct pt_regs *regs)
276 {
277 	unsigned long mmcra = mfspr(SPRN_MMCRA);
278 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
279 	int use_siar;
280 
281 	regs->dsisr = mmcra;
282 
283 	if (ppmu->flags & PPMU_HAS_SIER)
284 		regs->dar = mfspr(SPRN_SIER);
285 
286 	/*
287 	 * If this isn't a PMU exception (eg a software event) the SIAR is
288 	 * not valid. Use pt_regs.
289 	 *
290 	 * If it is a marked event use the SIAR.
291 	 *
292 	 * If the PMU doesn't update the SIAR for non marked events use
293 	 * pt_regs.
294 	 *
295 	 * If the PMU has HV/PR flags then check to see if they
296 	 * place the exception in userspace. If so, use pt_regs. In
297 	 * continuous sampling mode the SIAR and the PMU exception are
298 	 * not synchronised, so they may be many instructions apart.
299 	 * This can result in confusing backtraces. We still want
300 	 * hypervisor samples as well as samples in the kernel with
301 	 * interrupts off hence the userspace check.
302 	 */
303 	if (TRAP(regs) != 0xf00)
304 		use_siar = 0;
305 	else if ((ppmu->flags & PPMU_NO_SIAR))
306 		use_siar = 0;
307 	else if (marked)
308 		use_siar = 1;
309 	else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
310 		use_siar = 0;
311 	else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
312 		use_siar = 0;
313 	else
314 		use_siar = 1;
315 
316 	regs->result = use_siar;
317 }
318 
319 /*
320  * If interrupts were soft-disabled when a PMU interrupt occurs, treat
321  * it as an NMI.
322  */
323 static inline int perf_intr_is_nmi(struct pt_regs *regs)
324 {
325 	return (regs->softe & IRQS_DISABLED);
326 }
327 
328 /*
329  * On processors like P7+ that have the SIAR-Valid bit, marked instructions
330  * must be sampled only if the SIAR-valid bit is set.
331  *
332  * For unmarked instructions and for processors that don't have the SIAR-Valid
333  * bit, assume that SIAR is valid.
334  */
335 static inline int siar_valid(struct pt_regs *regs)
336 {
337 	unsigned long mmcra = regs->dsisr;
338 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
339 
340 	if (marked) {
341 		if (ppmu->flags & PPMU_HAS_SIER)
342 			return regs->dar & SIER_SIAR_VALID;
343 
344 		if (ppmu->flags & PPMU_SIAR_VALID)
345 			return mmcra & POWER7P_MMCRA_SIAR_VALID;
346 	}
347 
348 	return 1;
349 }
350 
351 
352 /* Reset all possible BHRB entries */
353 static void power_pmu_bhrb_reset(void)
354 {
355 	asm volatile(PPC_CLRBHRB);
356 }
357 
358 static void power_pmu_bhrb_enable(struct perf_event *event)
359 {
360 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
361 
362 	if (!ppmu->bhrb_nr)
363 		return;
364 
365 	/* Clear BHRB if we changed task context to avoid data leaks */
366 	if (event->ctx->task && cpuhw->bhrb_context != event->ctx) {
367 		power_pmu_bhrb_reset();
368 		cpuhw->bhrb_context = event->ctx;
369 	}
370 	cpuhw->bhrb_users++;
371 	perf_sched_cb_inc(event->ctx->pmu);
372 }
373 
374 static void power_pmu_bhrb_disable(struct perf_event *event)
375 {
376 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
377 
378 	if (!ppmu->bhrb_nr)
379 		return;
380 
381 	WARN_ON_ONCE(!cpuhw->bhrb_users);
382 	cpuhw->bhrb_users--;
383 	perf_sched_cb_dec(event->ctx->pmu);
384 
385 	if (!cpuhw->disabled && !cpuhw->bhrb_users) {
386 		/* BHRB cannot be turned off when other
387 		 * events are active on the PMU.
388 		 */
389 
390 		/* avoid stale pointer */
391 		cpuhw->bhrb_context = NULL;
392 	}
393 }
394 
395 /* Called from ctxsw to prevent one process's branch entries to
396  * mingle with the other process's entries during context switch.
397  */
398 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
399 {
400 	if (!ppmu->bhrb_nr)
401 		return;
402 
403 	if (sched_in)
404 		power_pmu_bhrb_reset();
405 }
406 /* Calculate the to address for a branch */
407 static __u64 power_pmu_bhrb_to(u64 addr)
408 {
409 	unsigned int instr;
410 	int ret;
411 	__u64 target;
412 
413 	if (is_kernel_addr(addr)) {
414 		if (probe_kernel_read(&instr, (void *)addr, sizeof(instr)))
415 			return 0;
416 
417 		return branch_target(&instr);
418 	}
419 
420 	/* Userspace: need copy instruction here then translate it */
421 	pagefault_disable();
422 	ret = __get_user_inatomic(instr, (unsigned int __user *)addr);
423 	if (ret) {
424 		pagefault_enable();
425 		return 0;
426 	}
427 	pagefault_enable();
428 
429 	target = branch_target(&instr);
430 	if ((!target) || (instr & BRANCH_ABSOLUTE))
431 		return target;
432 
433 	/* Translate relative branch target from kernel to user address */
434 	return target - (unsigned long)&instr + addr;
435 }
436 
437 /* Processing BHRB entries */
438 static void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw)
439 {
440 	u64 val;
441 	u64 addr;
442 	int r_index, u_index, pred;
443 
444 	r_index = 0;
445 	u_index = 0;
446 	while (r_index < ppmu->bhrb_nr) {
447 		/* Assembly read function */
448 		val = read_bhrb(r_index++);
449 		if (!val)
450 			/* Terminal marker: End of valid BHRB entries */
451 			break;
452 		else {
453 			addr = val & BHRB_EA;
454 			pred = val & BHRB_PREDICTION;
455 
456 			if (!addr)
457 				/* invalid entry */
458 				continue;
459 
460 			/* Branches are read most recent first (ie. mfbhrb 0 is
461 			 * the most recent branch).
462 			 * There are two types of valid entries:
463 			 * 1) a target entry which is the to address of a
464 			 *    computed goto like a blr,bctr,btar.  The next
465 			 *    entry read from the bhrb will be branch
466 			 *    corresponding to this target (ie. the actual
467 			 *    blr/bctr/btar instruction).
468 			 * 2) a from address which is an actual branch.  If a
469 			 *    target entry proceeds this, then this is the
470 			 *    matching branch for that target.  If this is not
471 			 *    following a target entry, then this is a branch
472 			 *    where the target is given as an immediate field
473 			 *    in the instruction (ie. an i or b form branch).
474 			 *    In this case we need to read the instruction from
475 			 *    memory to determine the target/to address.
476 			 */
477 
478 			if (val & BHRB_TARGET) {
479 				/* Target branches use two entries
480 				 * (ie. computed gotos/XL form)
481 				 */
482 				cpuhw->bhrb_entries[u_index].to = addr;
483 				cpuhw->bhrb_entries[u_index].mispred = pred;
484 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
485 
486 				/* Get from address in next entry */
487 				val = read_bhrb(r_index++);
488 				addr = val & BHRB_EA;
489 				if (val & BHRB_TARGET) {
490 					/* Shouldn't have two targets in a
491 					   row.. Reset index and try again */
492 					r_index--;
493 					addr = 0;
494 				}
495 				cpuhw->bhrb_entries[u_index].from = addr;
496 			} else {
497 				/* Branches to immediate field
498 				   (ie I or B form) */
499 				cpuhw->bhrb_entries[u_index].from = addr;
500 				cpuhw->bhrb_entries[u_index].to =
501 					power_pmu_bhrb_to(addr);
502 				cpuhw->bhrb_entries[u_index].mispred = pred;
503 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
504 			}
505 			u_index++;
506 
507 		}
508 	}
509 	cpuhw->bhrb_stack.nr = u_index;
510 	return;
511 }
512 
513 static bool is_ebb_event(struct perf_event *event)
514 {
515 	/*
516 	 * This could be a per-PMU callback, but we'd rather avoid the cost. We
517 	 * check that the PMU supports EBB, meaning those that don't can still
518 	 * use bit 63 of the event code for something else if they wish.
519 	 */
520 	return (ppmu->flags & PPMU_ARCH_207S) &&
521 	       ((event->attr.config >> PERF_EVENT_CONFIG_EBB_SHIFT) & 1);
522 }
523 
524 static int ebb_event_check(struct perf_event *event)
525 {
526 	struct perf_event *leader = event->group_leader;
527 
528 	/* Event and group leader must agree on EBB */
529 	if (is_ebb_event(leader) != is_ebb_event(event))
530 		return -EINVAL;
531 
532 	if (is_ebb_event(event)) {
533 		if (!(event->attach_state & PERF_ATTACH_TASK))
534 			return -EINVAL;
535 
536 		if (!leader->attr.pinned || !leader->attr.exclusive)
537 			return -EINVAL;
538 
539 		if (event->attr.freq ||
540 		    event->attr.inherit ||
541 		    event->attr.sample_type ||
542 		    event->attr.sample_period ||
543 		    event->attr.enable_on_exec)
544 			return -EINVAL;
545 	}
546 
547 	return 0;
548 }
549 
550 static void ebb_event_add(struct perf_event *event)
551 {
552 	if (!is_ebb_event(event) || current->thread.used_ebb)
553 		return;
554 
555 	/*
556 	 * IFF this is the first time we've added an EBB event, set
557 	 * PMXE in the user MMCR0 so we can detect when it's cleared by
558 	 * userspace. We need this so that we can context switch while
559 	 * userspace is in the EBB handler (where PMXE is 0).
560 	 */
561 	current->thread.used_ebb = 1;
562 	current->thread.mmcr0 |= MMCR0_PMXE;
563 }
564 
565 static void ebb_switch_out(unsigned long mmcr0)
566 {
567 	if (!(mmcr0 & MMCR0_EBE))
568 		return;
569 
570 	current->thread.siar  = mfspr(SPRN_SIAR);
571 	current->thread.sier  = mfspr(SPRN_SIER);
572 	current->thread.sdar  = mfspr(SPRN_SDAR);
573 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
574 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
575 }
576 
577 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
578 {
579 	unsigned long mmcr0 = cpuhw->mmcr[0];
580 
581 	if (!ebb)
582 		goto out;
583 
584 	/* Enable EBB and read/write to all 6 PMCs and BHRB for userspace */
585 	mmcr0 |= MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC_U6;
586 
587 	/*
588 	 * Add any bits from the user MMCR0, FC or PMAO. This is compatible
589 	 * with pmao_restore_workaround() because we may add PMAO but we never
590 	 * clear it here.
591 	 */
592 	mmcr0 |= current->thread.mmcr0;
593 
594 	/*
595 	 * Be careful not to set PMXE if userspace had it cleared. This is also
596 	 * compatible with pmao_restore_workaround() because it has already
597 	 * cleared PMXE and we leave PMAO alone.
598 	 */
599 	if (!(current->thread.mmcr0 & MMCR0_PMXE))
600 		mmcr0 &= ~MMCR0_PMXE;
601 
602 	mtspr(SPRN_SIAR, current->thread.siar);
603 	mtspr(SPRN_SIER, current->thread.sier);
604 	mtspr(SPRN_SDAR, current->thread.sdar);
605 
606 	/*
607 	 * Merge the kernel & user values of MMCR2. The semantics we implement
608 	 * are that the user MMCR2 can set bits, ie. cause counters to freeze,
609 	 * but not clear bits. If a task wants to be able to clear bits, ie.
610 	 * unfreeze counters, it should not set exclude_xxx in its events and
611 	 * instead manage the MMCR2 entirely by itself.
612 	 */
613 	mtspr(SPRN_MMCR2, cpuhw->mmcr[3] | current->thread.mmcr2);
614 out:
615 	return mmcr0;
616 }
617 
618 static void pmao_restore_workaround(bool ebb)
619 {
620 	unsigned pmcs[6];
621 
622 	if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
623 		return;
624 
625 	/*
626 	 * On POWER8E there is a hardware defect which affects the PMU context
627 	 * switch logic, ie. power_pmu_disable/enable().
628 	 *
629 	 * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
630 	 * by the hardware. Sometime later the actual PMU exception is
631 	 * delivered.
632 	 *
633 	 * If we context switch, or simply disable/enable, the PMU prior to the
634 	 * exception arriving, the exception will be lost when we clear PMAO.
635 	 *
636 	 * When we reenable the PMU, we will write the saved MMCR0 with PMAO
637 	 * set, and this _should_ generate an exception. However because of the
638 	 * defect no exception is generated when we write PMAO, and we get
639 	 * stuck with no counters counting but no exception delivered.
640 	 *
641 	 * The workaround is to detect this case and tweak the hardware to
642 	 * create another pending PMU exception.
643 	 *
644 	 * We do that by setting up PMC6 (cycles) for an imminent overflow and
645 	 * enabling the PMU. That causes a new exception to be generated in the
646 	 * chip, but we don't take it yet because we have interrupts hard
647 	 * disabled. We then write back the PMU state as we want it to be seen
648 	 * by the exception handler. When we reenable interrupts the exception
649 	 * handler will be called and see the correct state.
650 	 *
651 	 * The logic is the same for EBB, except that the exception is gated by
652 	 * us having interrupts hard disabled as well as the fact that we are
653 	 * not in userspace. The exception is finally delivered when we return
654 	 * to userspace.
655 	 */
656 
657 	/* Only if PMAO is set and PMAO_SYNC is clear */
658 	if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
659 		return;
660 
661 	/* If we're doing EBB, only if BESCR[GE] is set */
662 	if (ebb && !(current->thread.bescr & BESCR_GE))
663 		return;
664 
665 	/*
666 	 * We are already soft-disabled in power_pmu_enable(). We need to hard
667 	 * disable to actually prevent the PMU exception from firing.
668 	 */
669 	hard_irq_disable();
670 
671 	/*
672 	 * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
673 	 * Using read/write_pmc() in a for loop adds 12 function calls and
674 	 * almost doubles our code size.
675 	 */
676 	pmcs[0] = mfspr(SPRN_PMC1);
677 	pmcs[1] = mfspr(SPRN_PMC2);
678 	pmcs[2] = mfspr(SPRN_PMC3);
679 	pmcs[3] = mfspr(SPRN_PMC4);
680 	pmcs[4] = mfspr(SPRN_PMC5);
681 	pmcs[5] = mfspr(SPRN_PMC6);
682 
683 	/* Ensure all freeze bits are unset */
684 	mtspr(SPRN_MMCR2, 0);
685 
686 	/* Set up PMC6 to overflow in one cycle */
687 	mtspr(SPRN_PMC6, 0x7FFFFFFE);
688 
689 	/* Enable exceptions and unfreeze PMC6 */
690 	mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
691 
692 	/* Now we need to refreeze and restore the PMCs */
693 	mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
694 
695 	mtspr(SPRN_PMC1, pmcs[0]);
696 	mtspr(SPRN_PMC2, pmcs[1]);
697 	mtspr(SPRN_PMC3, pmcs[2]);
698 	mtspr(SPRN_PMC4, pmcs[3]);
699 	mtspr(SPRN_PMC5, pmcs[4]);
700 	mtspr(SPRN_PMC6, pmcs[5]);
701 }
702 
703 static bool use_ic(u64 event)
704 {
705 	if (cpu_has_feature(CPU_FTR_POWER9_DD1) &&
706 			(event == 0x200f2 || event == 0x300f2))
707 		return true;
708 
709 	return false;
710 }
711 #endif /* CONFIG_PPC64 */
712 
713 static void perf_event_interrupt(struct pt_regs *regs);
714 
715 /*
716  * Read one performance monitor counter (PMC).
717  */
718 static unsigned long read_pmc(int idx)
719 {
720 	unsigned long val;
721 
722 	switch (idx) {
723 	case 1:
724 		val = mfspr(SPRN_PMC1);
725 		break;
726 	case 2:
727 		val = mfspr(SPRN_PMC2);
728 		break;
729 	case 3:
730 		val = mfspr(SPRN_PMC3);
731 		break;
732 	case 4:
733 		val = mfspr(SPRN_PMC4);
734 		break;
735 	case 5:
736 		val = mfspr(SPRN_PMC5);
737 		break;
738 	case 6:
739 		val = mfspr(SPRN_PMC6);
740 		break;
741 #ifdef CONFIG_PPC64
742 	case 7:
743 		val = mfspr(SPRN_PMC7);
744 		break;
745 	case 8:
746 		val = mfspr(SPRN_PMC8);
747 		break;
748 #endif /* CONFIG_PPC64 */
749 	default:
750 		printk(KERN_ERR "oops trying to read PMC%d\n", idx);
751 		val = 0;
752 	}
753 	return val;
754 }
755 
756 /*
757  * Write one PMC.
758  */
759 static void write_pmc(int idx, unsigned long val)
760 {
761 	switch (idx) {
762 	case 1:
763 		mtspr(SPRN_PMC1, val);
764 		break;
765 	case 2:
766 		mtspr(SPRN_PMC2, val);
767 		break;
768 	case 3:
769 		mtspr(SPRN_PMC3, val);
770 		break;
771 	case 4:
772 		mtspr(SPRN_PMC4, val);
773 		break;
774 	case 5:
775 		mtspr(SPRN_PMC5, val);
776 		break;
777 	case 6:
778 		mtspr(SPRN_PMC6, val);
779 		break;
780 #ifdef CONFIG_PPC64
781 	case 7:
782 		mtspr(SPRN_PMC7, val);
783 		break;
784 	case 8:
785 		mtspr(SPRN_PMC8, val);
786 		break;
787 #endif /* CONFIG_PPC64 */
788 	default:
789 		printk(KERN_ERR "oops trying to write PMC%d\n", idx);
790 	}
791 }
792 
793 /* Called from sysrq_handle_showregs() */
794 void perf_event_print_debug(void)
795 {
796 	unsigned long sdar, sier, flags;
797 	u32 pmcs[MAX_HWEVENTS];
798 	int i;
799 
800 	if (!ppmu) {
801 		pr_info("Performance monitor hardware not registered.\n");
802 		return;
803 	}
804 
805 	if (!ppmu->n_counter)
806 		return;
807 
808 	local_irq_save(flags);
809 
810 	pr_info("CPU: %d PMU registers, ppmu = %s n_counters = %d",
811 		 smp_processor_id(), ppmu->name, ppmu->n_counter);
812 
813 	for (i = 0; i < ppmu->n_counter; i++)
814 		pmcs[i] = read_pmc(i + 1);
815 
816 	for (; i < MAX_HWEVENTS; i++)
817 		pmcs[i] = 0xdeadbeef;
818 
819 	pr_info("PMC1:  %08x PMC2: %08x PMC3: %08x PMC4: %08x\n",
820 		 pmcs[0], pmcs[1], pmcs[2], pmcs[3]);
821 
822 	if (ppmu->n_counter > 4)
823 		pr_info("PMC5:  %08x PMC6: %08x PMC7: %08x PMC8: %08x\n",
824 			 pmcs[4], pmcs[5], pmcs[6], pmcs[7]);
825 
826 	pr_info("MMCR0: %016lx MMCR1: %016lx MMCRA: %016lx\n",
827 		mfspr(SPRN_MMCR0), mfspr(SPRN_MMCR1), mfspr(SPRN_MMCRA));
828 
829 	sdar = sier = 0;
830 #ifdef CONFIG_PPC64
831 	sdar = mfspr(SPRN_SDAR);
832 
833 	if (ppmu->flags & PPMU_HAS_SIER)
834 		sier = mfspr(SPRN_SIER);
835 
836 	if (ppmu->flags & PPMU_ARCH_207S) {
837 		pr_info("MMCR2: %016lx EBBHR: %016lx\n",
838 			mfspr(SPRN_MMCR2), mfspr(SPRN_EBBHR));
839 		pr_info("EBBRR: %016lx BESCR: %016lx\n",
840 			mfspr(SPRN_EBBRR), mfspr(SPRN_BESCR));
841 	}
842 #endif
843 	pr_info("SIAR:  %016lx SDAR:  %016lx SIER:  %016lx\n",
844 		mfspr(SPRN_SIAR), sdar, sier);
845 
846 	local_irq_restore(flags);
847 }
848 
849 /*
850  * Check if a set of events can all go on the PMU at once.
851  * If they can't, this will look at alternative codes for the events
852  * and see if any combination of alternative codes is feasible.
853  * The feasible set is returned in event_id[].
854  */
855 static int power_check_constraints(struct cpu_hw_events *cpuhw,
856 				   u64 event_id[], unsigned int cflags[],
857 				   int n_ev)
858 {
859 	unsigned long mask, value, nv;
860 	unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
861 	int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
862 	int i, j;
863 	unsigned long addf = ppmu->add_fields;
864 	unsigned long tadd = ppmu->test_adder;
865 
866 	if (n_ev > ppmu->n_counter)
867 		return -1;
868 
869 	/* First see if the events will go on as-is */
870 	for (i = 0; i < n_ev; ++i) {
871 		if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
872 		    && !ppmu->limited_pmc_event(event_id[i])) {
873 			ppmu->get_alternatives(event_id[i], cflags[i],
874 					       cpuhw->alternatives[i]);
875 			event_id[i] = cpuhw->alternatives[i][0];
876 		}
877 		if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
878 					 &cpuhw->avalues[i][0]))
879 			return -1;
880 	}
881 	value = mask = 0;
882 	for (i = 0; i < n_ev; ++i) {
883 		nv = (value | cpuhw->avalues[i][0]) +
884 			(value & cpuhw->avalues[i][0] & addf);
885 		if ((((nv + tadd) ^ value) & mask) != 0 ||
886 		    (((nv + tadd) ^ cpuhw->avalues[i][0]) &
887 		     cpuhw->amasks[i][0]) != 0)
888 			break;
889 		value = nv;
890 		mask |= cpuhw->amasks[i][0];
891 	}
892 	if (i == n_ev)
893 		return 0;	/* all OK */
894 
895 	/* doesn't work, gather alternatives... */
896 	if (!ppmu->get_alternatives)
897 		return -1;
898 	for (i = 0; i < n_ev; ++i) {
899 		choice[i] = 0;
900 		n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
901 						  cpuhw->alternatives[i]);
902 		for (j = 1; j < n_alt[i]; ++j)
903 			ppmu->get_constraint(cpuhw->alternatives[i][j],
904 					     &cpuhw->amasks[i][j],
905 					     &cpuhw->avalues[i][j]);
906 	}
907 
908 	/* enumerate all possibilities and see if any will work */
909 	i = 0;
910 	j = -1;
911 	value = mask = nv = 0;
912 	while (i < n_ev) {
913 		if (j >= 0) {
914 			/* we're backtracking, restore context */
915 			value = svalues[i];
916 			mask = smasks[i];
917 			j = choice[i];
918 		}
919 		/*
920 		 * See if any alternative k for event_id i,
921 		 * where k > j, will satisfy the constraints.
922 		 */
923 		while (++j < n_alt[i]) {
924 			nv = (value | cpuhw->avalues[i][j]) +
925 				(value & cpuhw->avalues[i][j] & addf);
926 			if ((((nv + tadd) ^ value) & mask) == 0 &&
927 			    (((nv + tadd) ^ cpuhw->avalues[i][j])
928 			     & cpuhw->amasks[i][j]) == 0)
929 				break;
930 		}
931 		if (j >= n_alt[i]) {
932 			/*
933 			 * No feasible alternative, backtrack
934 			 * to event_id i-1 and continue enumerating its
935 			 * alternatives from where we got up to.
936 			 */
937 			if (--i < 0)
938 				return -1;
939 		} else {
940 			/*
941 			 * Found a feasible alternative for event_id i,
942 			 * remember where we got up to with this event_id,
943 			 * go on to the next event_id, and start with
944 			 * the first alternative for it.
945 			 */
946 			choice[i] = j;
947 			svalues[i] = value;
948 			smasks[i] = mask;
949 			value = nv;
950 			mask |= cpuhw->amasks[i][j];
951 			++i;
952 			j = -1;
953 		}
954 	}
955 
956 	/* OK, we have a feasible combination, tell the caller the solution */
957 	for (i = 0; i < n_ev; ++i)
958 		event_id[i] = cpuhw->alternatives[i][choice[i]];
959 	return 0;
960 }
961 
962 /*
963  * Check if newly-added events have consistent settings for
964  * exclude_{user,kernel,hv} with each other and any previously
965  * added events.
966  */
967 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
968 			  int n_prev, int n_new)
969 {
970 	int eu = 0, ek = 0, eh = 0;
971 	int i, n, first;
972 	struct perf_event *event;
973 
974 	/*
975 	 * If the PMU we're on supports per event exclude settings then we
976 	 * don't need to do any of this logic. NB. This assumes no PMU has both
977 	 * per event exclude and limited PMCs.
978 	 */
979 	if (ppmu->flags & PPMU_ARCH_207S)
980 		return 0;
981 
982 	n = n_prev + n_new;
983 	if (n <= 1)
984 		return 0;
985 
986 	first = 1;
987 	for (i = 0; i < n; ++i) {
988 		if (cflags[i] & PPMU_LIMITED_PMC_OK) {
989 			cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
990 			continue;
991 		}
992 		event = ctrs[i];
993 		if (first) {
994 			eu = event->attr.exclude_user;
995 			ek = event->attr.exclude_kernel;
996 			eh = event->attr.exclude_hv;
997 			first = 0;
998 		} else if (event->attr.exclude_user != eu ||
999 			   event->attr.exclude_kernel != ek ||
1000 			   event->attr.exclude_hv != eh) {
1001 			return -EAGAIN;
1002 		}
1003 	}
1004 
1005 	if (eu || ek || eh)
1006 		for (i = 0; i < n; ++i)
1007 			if (cflags[i] & PPMU_LIMITED_PMC_OK)
1008 				cflags[i] |= PPMU_LIMITED_PMC_REQD;
1009 
1010 	return 0;
1011 }
1012 
1013 static u64 check_and_compute_delta(u64 prev, u64 val)
1014 {
1015 	u64 delta = (val - prev) & 0xfffffffful;
1016 
1017 	/*
1018 	 * POWER7 can roll back counter values, if the new value is smaller
1019 	 * than the previous value it will cause the delta and the counter to
1020 	 * have bogus values unless we rolled a counter over.  If a coutner is
1021 	 * rolled back, it will be smaller, but within 256, which is the maximum
1022 	 * number of events to rollback at once.  If we detect a rollback
1023 	 * return 0.  This can lead to a small lack of precision in the
1024 	 * counters.
1025 	 */
1026 	if (prev > val && (prev - val) < 256)
1027 		delta = 0;
1028 
1029 	return delta;
1030 }
1031 
1032 static void power_pmu_read(struct perf_event *event)
1033 {
1034 	s64 val, delta, prev;
1035 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1036 
1037 	if (event->hw.state & PERF_HES_STOPPED)
1038 		return;
1039 
1040 	if (!event->hw.idx)
1041 		return;
1042 
1043 	if (is_ebb_event(event)) {
1044 		val = read_pmc(event->hw.idx);
1045 		if (use_ic(event->attr.config)) {
1046 			val = mfspr(SPRN_IC);
1047 			if (val > cpuhw->ic_init)
1048 				val = val - cpuhw->ic_init;
1049 			else
1050 				val = val + (0 - cpuhw->ic_init);
1051 		}
1052 		local64_set(&event->hw.prev_count, val);
1053 		return;
1054 	}
1055 
1056 	/*
1057 	 * Performance monitor interrupts come even when interrupts
1058 	 * are soft-disabled, as long as interrupts are hard-enabled.
1059 	 * Therefore we treat them like NMIs.
1060 	 */
1061 	do {
1062 		prev = local64_read(&event->hw.prev_count);
1063 		barrier();
1064 		val = read_pmc(event->hw.idx);
1065 		if (use_ic(event->attr.config)) {
1066 			val = mfspr(SPRN_IC);
1067 			if (val > cpuhw->ic_init)
1068 				val = val - cpuhw->ic_init;
1069 			else
1070 				val = val + (0 - cpuhw->ic_init);
1071 		}
1072 		delta = check_and_compute_delta(prev, val);
1073 		if (!delta)
1074 			return;
1075 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
1076 
1077 	local64_add(delta, &event->count);
1078 
1079 	/*
1080 	 * A number of places program the PMC with (0x80000000 - period_left).
1081 	 * We never want period_left to be less than 1 because we will program
1082 	 * the PMC with a value >= 0x800000000 and an edge detected PMC will
1083 	 * roll around to 0 before taking an exception. We have seen this
1084 	 * on POWER8.
1085 	 *
1086 	 * To fix this, clamp the minimum value of period_left to 1.
1087 	 */
1088 	do {
1089 		prev = local64_read(&event->hw.period_left);
1090 		val = prev - delta;
1091 		if (val < 1)
1092 			val = 1;
1093 	} while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
1094 }
1095 
1096 /*
1097  * On some machines, PMC5 and PMC6 can't be written, don't respect
1098  * the freeze conditions, and don't generate interrupts.  This tells
1099  * us if `event' is using such a PMC.
1100  */
1101 static int is_limited_pmc(int pmcnum)
1102 {
1103 	return (ppmu->flags & PPMU_LIMITED_PMC5_6)
1104 		&& (pmcnum == 5 || pmcnum == 6);
1105 }
1106 
1107 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
1108 				    unsigned long pmc5, unsigned long pmc6)
1109 {
1110 	struct perf_event *event;
1111 	u64 val, prev, delta;
1112 	int i;
1113 
1114 	for (i = 0; i < cpuhw->n_limited; ++i) {
1115 		event = cpuhw->limited_counter[i];
1116 		if (!event->hw.idx)
1117 			continue;
1118 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1119 		prev = local64_read(&event->hw.prev_count);
1120 		event->hw.idx = 0;
1121 		delta = check_and_compute_delta(prev, val);
1122 		if (delta)
1123 			local64_add(delta, &event->count);
1124 	}
1125 }
1126 
1127 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
1128 				  unsigned long pmc5, unsigned long pmc6)
1129 {
1130 	struct perf_event *event;
1131 	u64 val, prev;
1132 	int i;
1133 
1134 	for (i = 0; i < cpuhw->n_limited; ++i) {
1135 		event = cpuhw->limited_counter[i];
1136 		event->hw.idx = cpuhw->limited_hwidx[i];
1137 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1138 		prev = local64_read(&event->hw.prev_count);
1139 		if (check_and_compute_delta(prev, val))
1140 			local64_set(&event->hw.prev_count, val);
1141 		perf_event_update_userpage(event);
1142 	}
1143 }
1144 
1145 /*
1146  * Since limited events don't respect the freeze conditions, we
1147  * have to read them immediately after freezing or unfreezing the
1148  * other events.  We try to keep the values from the limited
1149  * events as consistent as possible by keeping the delay (in
1150  * cycles and instructions) between freezing/unfreezing and reading
1151  * the limited events as small and consistent as possible.
1152  * Therefore, if any limited events are in use, we read them
1153  * both, and always in the same order, to minimize variability,
1154  * and do it inside the same asm that writes MMCR0.
1155  */
1156 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
1157 {
1158 	unsigned long pmc5, pmc6;
1159 
1160 	if (!cpuhw->n_limited) {
1161 		mtspr(SPRN_MMCR0, mmcr0);
1162 		return;
1163 	}
1164 
1165 	/*
1166 	 * Write MMCR0, then read PMC5 and PMC6 immediately.
1167 	 * To ensure we don't get a performance monitor interrupt
1168 	 * between writing MMCR0 and freezing/thawing the limited
1169 	 * events, we first write MMCR0 with the event overflow
1170 	 * interrupt enable bits turned off.
1171 	 */
1172 	asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
1173 		     : "=&r" (pmc5), "=&r" (pmc6)
1174 		     : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
1175 		       "i" (SPRN_MMCR0),
1176 		       "i" (SPRN_PMC5), "i" (SPRN_PMC6));
1177 
1178 	if (mmcr0 & MMCR0_FC)
1179 		freeze_limited_counters(cpuhw, pmc5, pmc6);
1180 	else
1181 		thaw_limited_counters(cpuhw, pmc5, pmc6);
1182 
1183 	/*
1184 	 * Write the full MMCR0 including the event overflow interrupt
1185 	 * enable bits, if necessary.
1186 	 */
1187 	if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
1188 		mtspr(SPRN_MMCR0, mmcr0);
1189 }
1190 
1191 /*
1192  * Disable all events to prevent PMU interrupts and to allow
1193  * events to be added or removed.
1194  */
1195 static void power_pmu_disable(struct pmu *pmu)
1196 {
1197 	struct cpu_hw_events *cpuhw;
1198 	unsigned long flags, mmcr0, val;
1199 
1200 	if (!ppmu)
1201 		return;
1202 	local_irq_save(flags);
1203 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1204 
1205 	if (!cpuhw->disabled) {
1206 		/*
1207 		 * Check if we ever enabled the PMU on this cpu.
1208 		 */
1209 		if (!cpuhw->pmcs_enabled) {
1210 			ppc_enable_pmcs();
1211 			cpuhw->pmcs_enabled = 1;
1212 		}
1213 
1214 		/*
1215 		 * Set the 'freeze counters' bit, clear EBE/BHRBA/PMCC/PMAO/FC56
1216 		 */
1217 		val  = mmcr0 = mfspr(SPRN_MMCR0);
1218 		val |= MMCR0_FC;
1219 		val &= ~(MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC | MMCR0_PMAO |
1220 			 MMCR0_FC56);
1221 
1222 		/*
1223 		 * The barrier is to make sure the mtspr has been
1224 		 * executed and the PMU has frozen the events etc.
1225 		 * before we return.
1226 		 */
1227 		write_mmcr0(cpuhw, val);
1228 		mb();
1229 
1230 		/*
1231 		 * Disable instruction sampling if it was enabled
1232 		 */
1233 		if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1234 			mtspr(SPRN_MMCRA,
1235 			      cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1236 			mb();
1237 		}
1238 
1239 		cpuhw->disabled = 1;
1240 		cpuhw->n_added = 0;
1241 
1242 		ebb_switch_out(mmcr0);
1243 	}
1244 
1245 	local_irq_restore(flags);
1246 }
1247 
1248 /*
1249  * Re-enable all events if disable == 0.
1250  * If we were previously disabled and events were added, then
1251  * put the new config on the PMU.
1252  */
1253 static void power_pmu_enable(struct pmu *pmu)
1254 {
1255 	struct perf_event *event;
1256 	struct cpu_hw_events *cpuhw;
1257 	unsigned long flags;
1258 	long i;
1259 	unsigned long val, mmcr0;
1260 	s64 left;
1261 	unsigned int hwc_index[MAX_HWEVENTS];
1262 	int n_lim;
1263 	int idx;
1264 	bool ebb;
1265 
1266 	if (!ppmu)
1267 		return;
1268 	local_irq_save(flags);
1269 
1270 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1271 	if (!cpuhw->disabled)
1272 		goto out;
1273 
1274 	if (cpuhw->n_events == 0) {
1275 		ppc_set_pmu_inuse(0);
1276 		goto out;
1277 	}
1278 
1279 	cpuhw->disabled = 0;
1280 
1281 	/*
1282 	 * EBB requires an exclusive group and all events must have the EBB
1283 	 * flag set, or not set, so we can just check a single event. Also we
1284 	 * know we have at least one event.
1285 	 */
1286 	ebb = is_ebb_event(cpuhw->event[0]);
1287 
1288 	/*
1289 	 * If we didn't change anything, or only removed events,
1290 	 * no need to recalculate MMCR* settings and reset the PMCs.
1291 	 * Just reenable the PMU with the current MMCR* settings
1292 	 * (possibly updated for removal of events).
1293 	 */
1294 	if (!cpuhw->n_added) {
1295 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1296 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1297 		goto out_enable;
1298 	}
1299 
1300 	/*
1301 	 * Clear all MMCR settings and recompute them for the new set of events.
1302 	 */
1303 	memset(cpuhw->mmcr, 0, sizeof(cpuhw->mmcr));
1304 
1305 	if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
1306 			       cpuhw->mmcr, cpuhw->event)) {
1307 		/* shouldn't ever get here */
1308 		printk(KERN_ERR "oops compute_mmcr failed\n");
1309 		goto out;
1310 	}
1311 
1312 	if (!(ppmu->flags & PPMU_ARCH_207S)) {
1313 		/*
1314 		 * Add in MMCR0 freeze bits corresponding to the attr.exclude_*
1315 		 * bits for the first event. We have already checked that all
1316 		 * events have the same value for these bits as the first event.
1317 		 */
1318 		event = cpuhw->event[0];
1319 		if (event->attr.exclude_user)
1320 			cpuhw->mmcr[0] |= MMCR0_FCP;
1321 		if (event->attr.exclude_kernel)
1322 			cpuhw->mmcr[0] |= freeze_events_kernel;
1323 		if (event->attr.exclude_hv)
1324 			cpuhw->mmcr[0] |= MMCR0_FCHV;
1325 	}
1326 
1327 	/*
1328 	 * Write the new configuration to MMCR* with the freeze
1329 	 * bit set and set the hardware events to their initial values.
1330 	 * Then unfreeze the events.
1331 	 */
1332 	ppc_set_pmu_inuse(1);
1333 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1334 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1335 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
1336 				| MMCR0_FC);
1337 	if (ppmu->flags & PPMU_ARCH_207S)
1338 		mtspr(SPRN_MMCR2, cpuhw->mmcr[3]);
1339 
1340 	/*
1341 	 * Read off any pre-existing events that need to move
1342 	 * to another PMC.
1343 	 */
1344 	for (i = 0; i < cpuhw->n_events; ++i) {
1345 		event = cpuhw->event[i];
1346 		if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
1347 			power_pmu_read(event);
1348 			write_pmc(event->hw.idx, 0);
1349 			event->hw.idx = 0;
1350 		}
1351 	}
1352 
1353 	/*
1354 	 * Initialize the PMCs for all the new and moved events.
1355 	 */
1356 	cpuhw->n_limited = n_lim = 0;
1357 	for (i = 0; i < cpuhw->n_events; ++i) {
1358 		event = cpuhw->event[i];
1359 		if (event->hw.idx)
1360 			continue;
1361 		idx = hwc_index[i] + 1;
1362 		if (is_limited_pmc(idx)) {
1363 			cpuhw->limited_counter[n_lim] = event;
1364 			cpuhw->limited_hwidx[n_lim] = idx;
1365 			++n_lim;
1366 			continue;
1367 		}
1368 
1369 		if (ebb)
1370 			val = local64_read(&event->hw.prev_count);
1371 		else {
1372 			val = 0;
1373 			if (event->hw.sample_period) {
1374 				left = local64_read(&event->hw.period_left);
1375 				if (left < 0x80000000L)
1376 					val = 0x80000000L - left;
1377 			}
1378 			local64_set(&event->hw.prev_count, val);
1379 		}
1380 
1381 		event->hw.idx = idx;
1382 		if (event->hw.state & PERF_HES_STOPPED)
1383 			val = 0;
1384 		write_pmc(idx, val);
1385 
1386 		perf_event_update_userpage(event);
1387 	}
1388 	cpuhw->n_limited = n_lim;
1389 	cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
1390 
1391  out_enable:
1392 	pmao_restore_workaround(ebb);
1393 
1394 	mmcr0 = ebb_switch_in(ebb, cpuhw);
1395 
1396 	mb();
1397 	if (cpuhw->bhrb_users)
1398 		ppmu->config_bhrb(cpuhw->bhrb_filter);
1399 
1400 	write_mmcr0(cpuhw, mmcr0);
1401 
1402 	/*
1403 	 * Enable instruction sampling if necessary
1404 	 */
1405 	if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1406 		mb();
1407 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
1408 	}
1409 
1410  out:
1411 
1412 	local_irq_restore(flags);
1413 }
1414 
1415 static int collect_events(struct perf_event *group, int max_count,
1416 			  struct perf_event *ctrs[], u64 *events,
1417 			  unsigned int *flags)
1418 {
1419 	int n = 0;
1420 	struct perf_event *event;
1421 
1422 	if (group->pmu->task_ctx_nr == perf_hw_context) {
1423 		if (n >= max_count)
1424 			return -1;
1425 		ctrs[n] = group;
1426 		flags[n] = group->hw.event_base;
1427 		events[n++] = group->hw.config;
1428 	}
1429 	list_for_each_entry(event, &group->sibling_list, group_entry) {
1430 		if (event->pmu->task_ctx_nr == perf_hw_context &&
1431 		    event->state != PERF_EVENT_STATE_OFF) {
1432 			if (n >= max_count)
1433 				return -1;
1434 			ctrs[n] = event;
1435 			flags[n] = event->hw.event_base;
1436 			events[n++] = event->hw.config;
1437 		}
1438 	}
1439 	return n;
1440 }
1441 
1442 /*
1443  * Add a event to the PMU.
1444  * If all events are not already frozen, then we disable and
1445  * re-enable the PMU in order to get hw_perf_enable to do the
1446  * actual work of reconfiguring the PMU.
1447  */
1448 static int power_pmu_add(struct perf_event *event, int ef_flags)
1449 {
1450 	struct cpu_hw_events *cpuhw;
1451 	unsigned long flags;
1452 	int n0;
1453 	int ret = -EAGAIN;
1454 
1455 	local_irq_save(flags);
1456 	perf_pmu_disable(event->pmu);
1457 
1458 	/*
1459 	 * Add the event to the list (if there is room)
1460 	 * and check whether the total set is still feasible.
1461 	 */
1462 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1463 	n0 = cpuhw->n_events;
1464 	if (n0 >= ppmu->n_counter)
1465 		goto out;
1466 	cpuhw->event[n0] = event;
1467 	cpuhw->events[n0] = event->hw.config;
1468 	cpuhw->flags[n0] = event->hw.event_base;
1469 
1470 	/*
1471 	 * This event may have been disabled/stopped in record_and_restart()
1472 	 * because we exceeded the ->event_limit. If re-starting the event,
1473 	 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
1474 	 * notification is re-enabled.
1475 	 */
1476 	if (!(ef_flags & PERF_EF_START))
1477 		event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1478 	else
1479 		event->hw.state = 0;
1480 
1481 	/*
1482 	 * If group events scheduling transaction was started,
1483 	 * skip the schedulability test here, it will be performed
1484 	 * at commit time(->commit_txn) as a whole
1485 	 */
1486 	if (cpuhw->txn_flags & PERF_PMU_TXN_ADD)
1487 		goto nocheck;
1488 
1489 	if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
1490 		goto out;
1491 	if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
1492 		goto out;
1493 	event->hw.config = cpuhw->events[n0];
1494 
1495 nocheck:
1496 	ebb_event_add(event);
1497 
1498 	++cpuhw->n_events;
1499 	++cpuhw->n_added;
1500 
1501 	ret = 0;
1502  out:
1503 	if (has_branch_stack(event)) {
1504 		power_pmu_bhrb_enable(event);
1505 		cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1506 					event->attr.branch_sample_type);
1507 	}
1508 
1509 	/*
1510 	 * Workaround for POWER9 DD1 to use the Instruction Counter
1511 	 * register value for instruction counting
1512 	 */
1513 	if (use_ic(event->attr.config))
1514 		cpuhw->ic_init = mfspr(SPRN_IC);
1515 
1516 	perf_pmu_enable(event->pmu);
1517 	local_irq_restore(flags);
1518 	return ret;
1519 }
1520 
1521 /*
1522  * Remove a event from the PMU.
1523  */
1524 static void power_pmu_del(struct perf_event *event, int ef_flags)
1525 {
1526 	struct cpu_hw_events *cpuhw;
1527 	long i;
1528 	unsigned long flags;
1529 
1530 	local_irq_save(flags);
1531 	perf_pmu_disable(event->pmu);
1532 
1533 	power_pmu_read(event);
1534 
1535 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1536 	for (i = 0; i < cpuhw->n_events; ++i) {
1537 		if (event == cpuhw->event[i]) {
1538 			while (++i < cpuhw->n_events) {
1539 				cpuhw->event[i-1] = cpuhw->event[i];
1540 				cpuhw->events[i-1] = cpuhw->events[i];
1541 				cpuhw->flags[i-1] = cpuhw->flags[i];
1542 			}
1543 			--cpuhw->n_events;
1544 			ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
1545 			if (event->hw.idx) {
1546 				write_pmc(event->hw.idx, 0);
1547 				event->hw.idx = 0;
1548 			}
1549 			perf_event_update_userpage(event);
1550 			break;
1551 		}
1552 	}
1553 	for (i = 0; i < cpuhw->n_limited; ++i)
1554 		if (event == cpuhw->limited_counter[i])
1555 			break;
1556 	if (i < cpuhw->n_limited) {
1557 		while (++i < cpuhw->n_limited) {
1558 			cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
1559 			cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
1560 		}
1561 		--cpuhw->n_limited;
1562 	}
1563 	if (cpuhw->n_events == 0) {
1564 		/* disable exceptions if no events are running */
1565 		cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
1566 	}
1567 
1568 	if (has_branch_stack(event))
1569 		power_pmu_bhrb_disable(event);
1570 
1571 	perf_pmu_enable(event->pmu);
1572 	local_irq_restore(flags);
1573 }
1574 
1575 /*
1576  * POWER-PMU does not support disabling individual counters, hence
1577  * program their cycle counter to their max value and ignore the interrupts.
1578  */
1579 
1580 static void power_pmu_start(struct perf_event *event, int ef_flags)
1581 {
1582 	unsigned long flags;
1583 	s64 left;
1584 	unsigned long val;
1585 
1586 	if (!event->hw.idx || !event->hw.sample_period)
1587 		return;
1588 
1589 	if (!(event->hw.state & PERF_HES_STOPPED))
1590 		return;
1591 
1592 	if (ef_flags & PERF_EF_RELOAD)
1593 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1594 
1595 	local_irq_save(flags);
1596 	perf_pmu_disable(event->pmu);
1597 
1598 	event->hw.state = 0;
1599 	left = local64_read(&event->hw.period_left);
1600 
1601 	val = 0;
1602 	if (left < 0x80000000L)
1603 		val = 0x80000000L - left;
1604 
1605 	write_pmc(event->hw.idx, val);
1606 
1607 	perf_event_update_userpage(event);
1608 	perf_pmu_enable(event->pmu);
1609 	local_irq_restore(flags);
1610 }
1611 
1612 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1613 {
1614 	unsigned long flags;
1615 
1616 	if (!event->hw.idx || !event->hw.sample_period)
1617 		return;
1618 
1619 	if (event->hw.state & PERF_HES_STOPPED)
1620 		return;
1621 
1622 	local_irq_save(flags);
1623 	perf_pmu_disable(event->pmu);
1624 
1625 	power_pmu_read(event);
1626 	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1627 	write_pmc(event->hw.idx, 0);
1628 
1629 	perf_event_update_userpage(event);
1630 	perf_pmu_enable(event->pmu);
1631 	local_irq_restore(flags);
1632 }
1633 
1634 /*
1635  * Start group events scheduling transaction
1636  * Set the flag to make pmu::enable() not perform the
1637  * schedulability test, it will be performed at commit time
1638  *
1639  * We only support PERF_PMU_TXN_ADD transactions. Save the
1640  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1641  * transactions.
1642  */
1643 static void power_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1644 {
1645 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1646 
1647 	WARN_ON_ONCE(cpuhw->txn_flags);		/* txn already in flight */
1648 
1649 	cpuhw->txn_flags = txn_flags;
1650 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1651 		return;
1652 
1653 	perf_pmu_disable(pmu);
1654 	cpuhw->n_txn_start = cpuhw->n_events;
1655 }
1656 
1657 /*
1658  * Stop group events scheduling transaction
1659  * Clear the flag and pmu::enable() will perform the
1660  * schedulability test.
1661  */
1662 static void power_pmu_cancel_txn(struct pmu *pmu)
1663 {
1664 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1665 	unsigned int txn_flags;
1666 
1667 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1668 
1669 	txn_flags = cpuhw->txn_flags;
1670 	cpuhw->txn_flags = 0;
1671 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1672 		return;
1673 
1674 	perf_pmu_enable(pmu);
1675 }
1676 
1677 /*
1678  * Commit group events scheduling transaction
1679  * Perform the group schedulability test as a whole
1680  * Return 0 if success
1681  */
1682 static int power_pmu_commit_txn(struct pmu *pmu)
1683 {
1684 	struct cpu_hw_events *cpuhw;
1685 	long i, n;
1686 
1687 	if (!ppmu)
1688 		return -EAGAIN;
1689 
1690 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1691 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1692 
1693 	if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
1694 		cpuhw->txn_flags = 0;
1695 		return 0;
1696 	}
1697 
1698 	n = cpuhw->n_events;
1699 	if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1700 		return -EAGAIN;
1701 	i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1702 	if (i < 0)
1703 		return -EAGAIN;
1704 
1705 	for (i = cpuhw->n_txn_start; i < n; ++i)
1706 		cpuhw->event[i]->hw.config = cpuhw->events[i];
1707 
1708 	cpuhw->txn_flags = 0;
1709 	perf_pmu_enable(pmu);
1710 	return 0;
1711 }
1712 
1713 /*
1714  * Return 1 if we might be able to put event on a limited PMC,
1715  * or 0 if not.
1716  * A event can only go on a limited PMC if it counts something
1717  * that a limited PMC can count, doesn't require interrupts, and
1718  * doesn't exclude any processor mode.
1719  */
1720 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1721 				 unsigned int flags)
1722 {
1723 	int n;
1724 	u64 alt[MAX_EVENT_ALTERNATIVES];
1725 
1726 	if (event->attr.exclude_user
1727 	    || event->attr.exclude_kernel
1728 	    || event->attr.exclude_hv
1729 	    || event->attr.sample_period)
1730 		return 0;
1731 
1732 	if (ppmu->limited_pmc_event(ev))
1733 		return 1;
1734 
1735 	/*
1736 	 * The requested event_id isn't on a limited PMC already;
1737 	 * see if any alternative code goes on a limited PMC.
1738 	 */
1739 	if (!ppmu->get_alternatives)
1740 		return 0;
1741 
1742 	flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1743 	n = ppmu->get_alternatives(ev, flags, alt);
1744 
1745 	return n > 0;
1746 }
1747 
1748 /*
1749  * Find an alternative event_id that goes on a normal PMC, if possible,
1750  * and return the event_id code, or 0 if there is no such alternative.
1751  * (Note: event_id code 0 is "don't count" on all machines.)
1752  */
1753 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1754 {
1755 	u64 alt[MAX_EVENT_ALTERNATIVES];
1756 	int n;
1757 
1758 	flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1759 	n = ppmu->get_alternatives(ev, flags, alt);
1760 	if (!n)
1761 		return 0;
1762 	return alt[0];
1763 }
1764 
1765 /* Number of perf_events counting hardware events */
1766 static atomic_t num_events;
1767 /* Used to avoid races in calling reserve/release_pmc_hardware */
1768 static DEFINE_MUTEX(pmc_reserve_mutex);
1769 
1770 /*
1771  * Release the PMU if this is the last perf_event.
1772  */
1773 static void hw_perf_event_destroy(struct perf_event *event)
1774 {
1775 	if (!atomic_add_unless(&num_events, -1, 1)) {
1776 		mutex_lock(&pmc_reserve_mutex);
1777 		if (atomic_dec_return(&num_events) == 0)
1778 			release_pmc_hardware();
1779 		mutex_unlock(&pmc_reserve_mutex);
1780 	}
1781 }
1782 
1783 /*
1784  * Translate a generic cache event_id config to a raw event_id code.
1785  */
1786 static int hw_perf_cache_event(u64 config, u64 *eventp)
1787 {
1788 	unsigned long type, op, result;
1789 	int ev;
1790 
1791 	if (!ppmu->cache_events)
1792 		return -EINVAL;
1793 
1794 	/* unpack config */
1795 	type = config & 0xff;
1796 	op = (config >> 8) & 0xff;
1797 	result = (config >> 16) & 0xff;
1798 
1799 	if (type >= PERF_COUNT_HW_CACHE_MAX ||
1800 	    op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1801 	    result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1802 		return -EINVAL;
1803 
1804 	ev = (*ppmu->cache_events)[type][op][result];
1805 	if (ev == 0)
1806 		return -EOPNOTSUPP;
1807 	if (ev == -1)
1808 		return -EINVAL;
1809 	*eventp = ev;
1810 	return 0;
1811 }
1812 
1813 static int power_pmu_event_init(struct perf_event *event)
1814 {
1815 	u64 ev;
1816 	unsigned long flags;
1817 	struct perf_event *ctrs[MAX_HWEVENTS];
1818 	u64 events[MAX_HWEVENTS];
1819 	unsigned int cflags[MAX_HWEVENTS];
1820 	int n;
1821 	int err;
1822 	struct cpu_hw_events *cpuhw;
1823 
1824 	if (!ppmu)
1825 		return -ENOENT;
1826 
1827 	if (has_branch_stack(event)) {
1828 	        /* PMU has BHRB enabled */
1829 		if (!(ppmu->flags & PPMU_ARCH_207S))
1830 			return -EOPNOTSUPP;
1831 	}
1832 
1833 	switch (event->attr.type) {
1834 	case PERF_TYPE_HARDWARE:
1835 		ev = event->attr.config;
1836 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1837 			return -EOPNOTSUPP;
1838 		ev = ppmu->generic_events[ev];
1839 		break;
1840 	case PERF_TYPE_HW_CACHE:
1841 		err = hw_perf_cache_event(event->attr.config, &ev);
1842 		if (err)
1843 			return err;
1844 		break;
1845 	case PERF_TYPE_RAW:
1846 		ev = event->attr.config;
1847 		break;
1848 	default:
1849 		return -ENOENT;
1850 	}
1851 
1852 	event->hw.config_base = ev;
1853 	event->hw.idx = 0;
1854 
1855 	/*
1856 	 * If we are not running on a hypervisor, force the
1857 	 * exclude_hv bit to 0 so that we don't care what
1858 	 * the user set it to.
1859 	 */
1860 	if (!firmware_has_feature(FW_FEATURE_LPAR))
1861 		event->attr.exclude_hv = 0;
1862 
1863 	/*
1864 	 * If this is a per-task event, then we can use
1865 	 * PM_RUN_* events interchangeably with their non RUN_*
1866 	 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1867 	 * XXX we should check if the task is an idle task.
1868 	 */
1869 	flags = 0;
1870 	if (event->attach_state & PERF_ATTACH_TASK)
1871 		flags |= PPMU_ONLY_COUNT_RUN;
1872 
1873 	/*
1874 	 * If this machine has limited events, check whether this
1875 	 * event_id could go on a limited event.
1876 	 */
1877 	if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1878 		if (can_go_on_limited_pmc(event, ev, flags)) {
1879 			flags |= PPMU_LIMITED_PMC_OK;
1880 		} else if (ppmu->limited_pmc_event(ev)) {
1881 			/*
1882 			 * The requested event_id is on a limited PMC,
1883 			 * but we can't use a limited PMC; see if any
1884 			 * alternative goes on a normal PMC.
1885 			 */
1886 			ev = normal_pmc_alternative(ev, flags);
1887 			if (!ev)
1888 				return -EINVAL;
1889 		}
1890 	}
1891 
1892 	/* Extra checks for EBB */
1893 	err = ebb_event_check(event);
1894 	if (err)
1895 		return err;
1896 
1897 	/*
1898 	 * If this is in a group, check if it can go on with all the
1899 	 * other hardware events in the group.  We assume the event
1900 	 * hasn't been linked into its leader's sibling list at this point.
1901 	 */
1902 	n = 0;
1903 	if (event->group_leader != event) {
1904 		n = collect_events(event->group_leader, ppmu->n_counter - 1,
1905 				   ctrs, events, cflags);
1906 		if (n < 0)
1907 			return -EINVAL;
1908 	}
1909 	events[n] = ev;
1910 	ctrs[n] = event;
1911 	cflags[n] = flags;
1912 	if (check_excludes(ctrs, cflags, n, 1))
1913 		return -EINVAL;
1914 
1915 	cpuhw = &get_cpu_var(cpu_hw_events);
1916 	err = power_check_constraints(cpuhw, events, cflags, n + 1);
1917 
1918 	if (has_branch_stack(event)) {
1919 		cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1920 					event->attr.branch_sample_type);
1921 
1922 		if (cpuhw->bhrb_filter == -1) {
1923 			put_cpu_var(cpu_hw_events);
1924 			return -EOPNOTSUPP;
1925 		}
1926 	}
1927 
1928 	put_cpu_var(cpu_hw_events);
1929 	if (err)
1930 		return -EINVAL;
1931 
1932 	event->hw.config = events[n];
1933 	event->hw.event_base = cflags[n];
1934 	event->hw.last_period = event->hw.sample_period;
1935 	local64_set(&event->hw.period_left, event->hw.last_period);
1936 
1937 	/*
1938 	 * For EBB events we just context switch the PMC value, we don't do any
1939 	 * of the sample_period logic. We use hw.prev_count for this.
1940 	 */
1941 	if (is_ebb_event(event))
1942 		local64_set(&event->hw.prev_count, 0);
1943 
1944 	/*
1945 	 * See if we need to reserve the PMU.
1946 	 * If no events are currently in use, then we have to take a
1947 	 * mutex to ensure that we don't race with another task doing
1948 	 * reserve_pmc_hardware or release_pmc_hardware.
1949 	 */
1950 	err = 0;
1951 	if (!atomic_inc_not_zero(&num_events)) {
1952 		mutex_lock(&pmc_reserve_mutex);
1953 		if (atomic_read(&num_events) == 0 &&
1954 		    reserve_pmc_hardware(perf_event_interrupt))
1955 			err = -EBUSY;
1956 		else
1957 			atomic_inc(&num_events);
1958 		mutex_unlock(&pmc_reserve_mutex);
1959 	}
1960 	event->destroy = hw_perf_event_destroy;
1961 
1962 	return err;
1963 }
1964 
1965 static int power_pmu_event_idx(struct perf_event *event)
1966 {
1967 	return event->hw.idx;
1968 }
1969 
1970 ssize_t power_events_sysfs_show(struct device *dev,
1971 				struct device_attribute *attr, char *page)
1972 {
1973 	struct perf_pmu_events_attr *pmu_attr;
1974 
1975 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
1976 
1977 	return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
1978 }
1979 
1980 static struct pmu power_pmu = {
1981 	.pmu_enable	= power_pmu_enable,
1982 	.pmu_disable	= power_pmu_disable,
1983 	.event_init	= power_pmu_event_init,
1984 	.add		= power_pmu_add,
1985 	.del		= power_pmu_del,
1986 	.start		= power_pmu_start,
1987 	.stop		= power_pmu_stop,
1988 	.read		= power_pmu_read,
1989 	.start_txn	= power_pmu_start_txn,
1990 	.cancel_txn	= power_pmu_cancel_txn,
1991 	.commit_txn	= power_pmu_commit_txn,
1992 	.event_idx	= power_pmu_event_idx,
1993 	.sched_task	= power_pmu_sched_task,
1994 };
1995 
1996 /*
1997  * A counter has overflowed; update its count and record
1998  * things if requested.  Note that interrupts are hard-disabled
1999  * here so there is no possibility of being interrupted.
2000  */
2001 static void record_and_restart(struct perf_event *event, unsigned long val,
2002 			       struct pt_regs *regs)
2003 {
2004 	u64 period = event->hw.sample_period;
2005 	s64 prev, delta, left;
2006 	int record = 0;
2007 
2008 	if (event->hw.state & PERF_HES_STOPPED) {
2009 		write_pmc(event->hw.idx, 0);
2010 		return;
2011 	}
2012 
2013 	/* we don't have to worry about interrupts here */
2014 	prev = local64_read(&event->hw.prev_count);
2015 	delta = check_and_compute_delta(prev, val);
2016 	local64_add(delta, &event->count);
2017 
2018 	/*
2019 	 * See if the total period for this event has expired,
2020 	 * and update for the next period.
2021 	 */
2022 	val = 0;
2023 	left = local64_read(&event->hw.period_left) - delta;
2024 	if (delta == 0)
2025 		left++;
2026 	if (period) {
2027 		if (left <= 0) {
2028 			left += period;
2029 			if (left <= 0)
2030 				left = period;
2031 			record = siar_valid(regs);
2032 			event->hw.last_period = event->hw.sample_period;
2033 		}
2034 		if (left < 0x80000000LL)
2035 			val = 0x80000000LL - left;
2036 	}
2037 
2038 	write_pmc(event->hw.idx, val);
2039 	local64_set(&event->hw.prev_count, val);
2040 	local64_set(&event->hw.period_left, left);
2041 	perf_event_update_userpage(event);
2042 
2043 	/*
2044 	 * Finally record data if requested.
2045 	 */
2046 	if (record) {
2047 		struct perf_sample_data data;
2048 
2049 		perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
2050 
2051 		if (event->attr.sample_type &
2052 		    (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR))
2053 			perf_get_data_addr(regs, &data.addr);
2054 
2055 		if (event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK) {
2056 			struct cpu_hw_events *cpuhw;
2057 			cpuhw = this_cpu_ptr(&cpu_hw_events);
2058 			power_pmu_bhrb_read(cpuhw);
2059 			data.br_stack = &cpuhw->bhrb_stack;
2060 		}
2061 
2062 		if (event->attr.sample_type & PERF_SAMPLE_DATA_SRC &&
2063 						ppmu->get_mem_data_src)
2064 			ppmu->get_mem_data_src(&data.data_src, ppmu->flags, regs);
2065 
2066 		if (event->attr.sample_type & PERF_SAMPLE_WEIGHT &&
2067 						ppmu->get_mem_weight)
2068 			ppmu->get_mem_weight(&data.weight);
2069 
2070 		if (perf_event_overflow(event, &data, regs))
2071 			power_pmu_stop(event, 0);
2072 	}
2073 }
2074 
2075 /*
2076  * Called from generic code to get the misc flags (i.e. processor mode)
2077  * for an event_id.
2078  */
2079 unsigned long perf_misc_flags(struct pt_regs *regs)
2080 {
2081 	u32 flags = perf_get_misc_flags(regs);
2082 
2083 	if (flags)
2084 		return flags;
2085 	return user_mode(regs) ? PERF_RECORD_MISC_USER :
2086 		PERF_RECORD_MISC_KERNEL;
2087 }
2088 
2089 /*
2090  * Called from generic code to get the instruction pointer
2091  * for an event_id.
2092  */
2093 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2094 {
2095 	bool use_siar = regs_use_siar(regs);
2096 
2097 	if (use_siar && siar_valid(regs))
2098 		return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
2099 	else if (use_siar)
2100 		return 0;		// no valid instruction pointer
2101 	else
2102 		return regs->nip;
2103 }
2104 
2105 static bool pmc_overflow_power7(unsigned long val)
2106 {
2107 	/*
2108 	 * Events on POWER7 can roll back if a speculative event doesn't
2109 	 * eventually complete. Unfortunately in some rare cases they will
2110 	 * raise a performance monitor exception. We need to catch this to
2111 	 * ensure we reset the PMC. In all cases the PMC will be 256 or less
2112 	 * cycles from overflow.
2113 	 *
2114 	 * We only do this if the first pass fails to find any overflowing
2115 	 * PMCs because a user might set a period of less than 256 and we
2116 	 * don't want to mistakenly reset them.
2117 	 */
2118 	if ((0x80000000 - val) <= 256)
2119 		return true;
2120 
2121 	return false;
2122 }
2123 
2124 static bool pmc_overflow(unsigned long val)
2125 {
2126 	if ((int)val < 0)
2127 		return true;
2128 
2129 	return false;
2130 }
2131 
2132 /*
2133  * Performance monitor interrupt stuff
2134  */
2135 static void perf_event_interrupt(struct pt_regs *regs)
2136 {
2137 	int i, j;
2138 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
2139 	struct perf_event *event;
2140 	unsigned long val[8];
2141 	int found, active;
2142 	int nmi;
2143 
2144 	if (cpuhw->n_limited)
2145 		freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
2146 					mfspr(SPRN_PMC6));
2147 
2148 	perf_read_regs(regs);
2149 
2150 	nmi = perf_intr_is_nmi(regs);
2151 	if (nmi)
2152 		nmi_enter();
2153 	else
2154 		irq_enter();
2155 
2156 	/* Read all the PMCs since we'll need them a bunch of times */
2157 	for (i = 0; i < ppmu->n_counter; ++i)
2158 		val[i] = read_pmc(i + 1);
2159 
2160 	/* Try to find what caused the IRQ */
2161 	found = 0;
2162 	for (i = 0; i < ppmu->n_counter; ++i) {
2163 		if (!pmc_overflow(val[i]))
2164 			continue;
2165 		if (is_limited_pmc(i + 1))
2166 			continue; /* these won't generate IRQs */
2167 		/*
2168 		 * We've found one that's overflowed.  For active
2169 		 * counters we need to log this.  For inactive
2170 		 * counters, we need to reset it anyway
2171 		 */
2172 		found = 1;
2173 		active = 0;
2174 		for (j = 0; j < cpuhw->n_events; ++j) {
2175 			event = cpuhw->event[j];
2176 			if (event->hw.idx == (i + 1)) {
2177 				active = 1;
2178 				record_and_restart(event, val[i], regs);
2179 				break;
2180 			}
2181 		}
2182 		if (!active)
2183 			/* reset non active counters that have overflowed */
2184 			write_pmc(i + 1, 0);
2185 	}
2186 	if (!found && pvr_version_is(PVR_POWER7)) {
2187 		/* check active counters for special buggy p7 overflow */
2188 		for (i = 0; i < cpuhw->n_events; ++i) {
2189 			event = cpuhw->event[i];
2190 			if (!event->hw.idx || is_limited_pmc(event->hw.idx))
2191 				continue;
2192 			if (pmc_overflow_power7(val[event->hw.idx - 1])) {
2193 				/* event has overflowed in a buggy way*/
2194 				found = 1;
2195 				record_and_restart(event,
2196 						   val[event->hw.idx - 1],
2197 						   regs);
2198 			}
2199 		}
2200 	}
2201 	if (!found && !nmi && printk_ratelimit())
2202 		printk(KERN_WARNING "Can't find PMC that caused IRQ\n");
2203 
2204 	/*
2205 	 * Reset MMCR0 to its normal value.  This will set PMXE and
2206 	 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
2207 	 * and thus allow interrupts to occur again.
2208 	 * XXX might want to use MSR.PM to keep the events frozen until
2209 	 * we get back out of this interrupt.
2210 	 */
2211 	write_mmcr0(cpuhw, cpuhw->mmcr[0]);
2212 
2213 	if (nmi)
2214 		nmi_exit();
2215 	else
2216 		irq_exit();
2217 }
2218 
2219 static int power_pmu_prepare_cpu(unsigned int cpu)
2220 {
2221 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
2222 
2223 	if (ppmu) {
2224 		memset(cpuhw, 0, sizeof(*cpuhw));
2225 		cpuhw->mmcr[0] = MMCR0_FC;
2226 	}
2227 	return 0;
2228 }
2229 
2230 int register_power_pmu(struct power_pmu *pmu)
2231 {
2232 	if (ppmu)
2233 		return -EBUSY;		/* something's already registered */
2234 
2235 	ppmu = pmu;
2236 	pr_info("%s performance monitor hardware support registered\n",
2237 		pmu->name);
2238 
2239 	power_pmu.attr_groups = ppmu->attr_groups;
2240 
2241 #ifdef MSR_HV
2242 	/*
2243 	 * Use FCHV to ignore kernel events if MSR.HV is set.
2244 	 */
2245 	if (mfmsr() & MSR_HV)
2246 		freeze_events_kernel = MMCR0_FCHV;
2247 #endif /* CONFIG_PPC64 */
2248 
2249 	perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
2250 	cpuhp_setup_state(CPUHP_PERF_POWER, "perf/powerpc:prepare",
2251 			  power_pmu_prepare_cpu, NULL);
2252 	return 0;
2253 }
2254