xref: /openbmc/linux/arch/powerpc/perf/core-book3s.c (revision b830f94f)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Performance event support - powerpc architecture code
4  *
5  * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
6  */
7 #include <linux/kernel.h>
8 #include <linux/sched.h>
9 #include <linux/sched/clock.h>
10 #include <linux/perf_event.h>
11 #include <linux/percpu.h>
12 #include <linux/hardirq.h>
13 #include <linux/uaccess.h>
14 #include <asm/reg.h>
15 #include <asm/pmc.h>
16 #include <asm/machdep.h>
17 #include <asm/firmware.h>
18 #include <asm/ptrace.h>
19 #include <asm/code-patching.h>
20 
21 #ifdef CONFIG_PPC64
22 #include "internal.h"
23 #endif
24 
25 #define BHRB_MAX_ENTRIES	32
26 #define BHRB_TARGET		0x0000000000000002
27 #define BHRB_PREDICTION		0x0000000000000001
28 #define BHRB_EA			0xFFFFFFFFFFFFFFFCUL
29 
30 struct cpu_hw_events {
31 	int n_events;
32 	int n_percpu;
33 	int disabled;
34 	int n_added;
35 	int n_limited;
36 	u8  pmcs_enabled;
37 	struct perf_event *event[MAX_HWEVENTS];
38 	u64 events[MAX_HWEVENTS];
39 	unsigned int flags[MAX_HWEVENTS];
40 	/*
41 	 * The order of the MMCR array is:
42 	 *  - 64-bit, MMCR0, MMCR1, MMCRA, MMCR2
43 	 *  - 32-bit, MMCR0, MMCR1, MMCR2
44 	 */
45 	unsigned long mmcr[4];
46 	struct perf_event *limited_counter[MAX_LIMITED_HWCOUNTERS];
47 	u8  limited_hwidx[MAX_LIMITED_HWCOUNTERS];
48 	u64 alternatives[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
49 	unsigned long amasks[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
50 	unsigned long avalues[MAX_HWEVENTS][MAX_EVENT_ALTERNATIVES];
51 
52 	unsigned int txn_flags;
53 	int n_txn_start;
54 
55 	/* BHRB bits */
56 	u64				bhrb_filter;	/* BHRB HW branch filter */
57 	unsigned int			bhrb_users;
58 	void				*bhrb_context;
59 	struct	perf_branch_stack	bhrb_stack;
60 	struct	perf_branch_entry	bhrb_entries[BHRB_MAX_ENTRIES];
61 	u64				ic_init;
62 };
63 
64 static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
65 
66 static struct power_pmu *ppmu;
67 
68 /*
69  * Normally, to ignore kernel events we set the FCS (freeze counters
70  * in supervisor mode) bit in MMCR0, but if the kernel runs with the
71  * hypervisor bit set in the MSR, or if we are running on a processor
72  * where the hypervisor bit is forced to 1 (as on Apple G5 processors),
73  * then we need to use the FCHV bit to ignore kernel events.
74  */
75 static unsigned int freeze_events_kernel = MMCR0_FCS;
76 
77 /*
78  * 32-bit doesn't have MMCRA but does have an MMCR2,
79  * and a few other names are different.
80  */
81 #ifdef CONFIG_PPC32
82 
83 #define MMCR0_FCHV		0
84 #define MMCR0_PMCjCE		MMCR0_PMCnCE
85 #define MMCR0_FC56		0
86 #define MMCR0_PMAO		0
87 #define MMCR0_EBE		0
88 #define MMCR0_BHRBA		0
89 #define MMCR0_PMCC		0
90 #define MMCR0_PMCC_U6		0
91 
92 #define SPRN_MMCRA		SPRN_MMCR2
93 #define MMCRA_SAMPLE_ENABLE	0
94 
95 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
96 {
97 	return 0;
98 }
99 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp) { }
100 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
101 {
102 	return 0;
103 }
104 static inline void perf_read_regs(struct pt_regs *regs)
105 {
106 	regs->result = 0;
107 }
108 static inline int perf_intr_is_nmi(struct pt_regs *regs)
109 {
110 	return 0;
111 }
112 
113 static inline int siar_valid(struct pt_regs *regs)
114 {
115 	return 1;
116 }
117 
118 static bool is_ebb_event(struct perf_event *event) { return false; }
119 static int ebb_event_check(struct perf_event *event) { return 0; }
120 static void ebb_event_add(struct perf_event *event) { }
121 static void ebb_switch_out(unsigned long mmcr0) { }
122 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
123 {
124 	return cpuhw->mmcr[0];
125 }
126 
127 static inline void power_pmu_bhrb_enable(struct perf_event *event) {}
128 static inline void power_pmu_bhrb_disable(struct perf_event *event) {}
129 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in) {}
130 static inline void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw) {}
131 static void pmao_restore_workaround(bool ebb) { }
132 #endif /* CONFIG_PPC32 */
133 
134 bool is_sier_available(void)
135 {
136 	if (ppmu->flags & PPMU_HAS_SIER)
137 		return true;
138 
139 	return false;
140 }
141 
142 static bool regs_use_siar(struct pt_regs *regs)
143 {
144 	/*
145 	 * When we take a performance monitor exception the regs are setup
146 	 * using perf_read_regs() which overloads some fields, in particular
147 	 * regs->result to tell us whether to use SIAR.
148 	 *
149 	 * However if the regs are from another exception, eg. a syscall, then
150 	 * they have not been setup using perf_read_regs() and so regs->result
151 	 * is something random.
152 	 */
153 	return ((TRAP(regs) == 0xf00) && regs->result);
154 }
155 
156 /*
157  * Things that are specific to 64-bit implementations.
158  */
159 #ifdef CONFIG_PPC64
160 
161 static inline unsigned long perf_ip_adjust(struct pt_regs *regs)
162 {
163 	unsigned long mmcra = regs->dsisr;
164 
165 	if ((ppmu->flags & PPMU_HAS_SSLOT) && (mmcra & MMCRA_SAMPLE_ENABLE)) {
166 		unsigned long slot = (mmcra & MMCRA_SLOT) >> MMCRA_SLOT_SHIFT;
167 		if (slot > 1)
168 			return 4 * (slot - 1);
169 	}
170 
171 	return 0;
172 }
173 
174 /*
175  * The user wants a data address recorded.
176  * If we're not doing instruction sampling, give them the SDAR
177  * (sampled data address).  If we are doing instruction sampling, then
178  * only give them the SDAR if it corresponds to the instruction
179  * pointed to by SIAR; this is indicated by the [POWER6_]MMCRA_SDSYNC, the
180  * [POWER7P_]MMCRA_SDAR_VALID bit in MMCRA, or the SDAR_VALID bit in SIER.
181  */
182 static inline void perf_get_data_addr(struct pt_regs *regs, u64 *addrp)
183 {
184 	unsigned long mmcra = regs->dsisr;
185 	bool sdar_valid;
186 
187 	if (ppmu->flags & PPMU_HAS_SIER)
188 		sdar_valid = regs->dar & SIER_SDAR_VALID;
189 	else {
190 		unsigned long sdsync;
191 
192 		if (ppmu->flags & PPMU_SIAR_VALID)
193 			sdsync = POWER7P_MMCRA_SDAR_VALID;
194 		else if (ppmu->flags & PPMU_ALT_SIPR)
195 			sdsync = POWER6_MMCRA_SDSYNC;
196 		else if (ppmu->flags & PPMU_NO_SIAR)
197 			sdsync = MMCRA_SAMPLE_ENABLE;
198 		else
199 			sdsync = MMCRA_SDSYNC;
200 
201 		sdar_valid = mmcra & sdsync;
202 	}
203 
204 	if (!(mmcra & MMCRA_SAMPLE_ENABLE) || sdar_valid)
205 		*addrp = mfspr(SPRN_SDAR);
206 
207 	if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN) &&
208 		is_kernel_addr(mfspr(SPRN_SDAR)))
209 		*addrp = 0;
210 }
211 
212 static bool regs_sihv(struct pt_regs *regs)
213 {
214 	unsigned long sihv = MMCRA_SIHV;
215 
216 	if (ppmu->flags & PPMU_HAS_SIER)
217 		return !!(regs->dar & SIER_SIHV);
218 
219 	if (ppmu->flags & PPMU_ALT_SIPR)
220 		sihv = POWER6_MMCRA_SIHV;
221 
222 	return !!(regs->dsisr & sihv);
223 }
224 
225 static bool regs_sipr(struct pt_regs *regs)
226 {
227 	unsigned long sipr = MMCRA_SIPR;
228 
229 	if (ppmu->flags & PPMU_HAS_SIER)
230 		return !!(regs->dar & SIER_SIPR);
231 
232 	if (ppmu->flags & PPMU_ALT_SIPR)
233 		sipr = POWER6_MMCRA_SIPR;
234 
235 	return !!(regs->dsisr & sipr);
236 }
237 
238 static inline u32 perf_flags_from_msr(struct pt_regs *regs)
239 {
240 	if (regs->msr & MSR_PR)
241 		return PERF_RECORD_MISC_USER;
242 	if ((regs->msr & MSR_HV) && freeze_events_kernel != MMCR0_FCHV)
243 		return PERF_RECORD_MISC_HYPERVISOR;
244 	return PERF_RECORD_MISC_KERNEL;
245 }
246 
247 static inline u32 perf_get_misc_flags(struct pt_regs *regs)
248 {
249 	bool use_siar = regs_use_siar(regs);
250 
251 	if (!use_siar)
252 		return perf_flags_from_msr(regs);
253 
254 	/*
255 	 * If we don't have flags in MMCRA, rather than using
256 	 * the MSR, we intuit the flags from the address in
257 	 * SIAR which should give slightly more reliable
258 	 * results
259 	 */
260 	if (ppmu->flags & PPMU_NO_SIPR) {
261 		unsigned long siar = mfspr(SPRN_SIAR);
262 		if (is_kernel_addr(siar))
263 			return PERF_RECORD_MISC_KERNEL;
264 		return PERF_RECORD_MISC_USER;
265 	}
266 
267 	/* PR has priority over HV, so order below is important */
268 	if (regs_sipr(regs))
269 		return PERF_RECORD_MISC_USER;
270 
271 	if (regs_sihv(regs) && (freeze_events_kernel != MMCR0_FCHV))
272 		return PERF_RECORD_MISC_HYPERVISOR;
273 
274 	return PERF_RECORD_MISC_KERNEL;
275 }
276 
277 /*
278  * Overload regs->dsisr to store MMCRA so we only need to read it once
279  * on each interrupt.
280  * Overload regs->dar to store SIER if we have it.
281  * Overload regs->result to specify whether we should use the MSR (result
282  * is zero) or the SIAR (result is non zero).
283  */
284 static inline void perf_read_regs(struct pt_regs *regs)
285 {
286 	unsigned long mmcra = mfspr(SPRN_MMCRA);
287 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
288 	int use_siar;
289 
290 	regs->dsisr = mmcra;
291 
292 	if (ppmu->flags & PPMU_HAS_SIER)
293 		regs->dar = mfspr(SPRN_SIER);
294 
295 	/*
296 	 * If this isn't a PMU exception (eg a software event) the SIAR is
297 	 * not valid. Use pt_regs.
298 	 *
299 	 * If it is a marked event use the SIAR.
300 	 *
301 	 * If the PMU doesn't update the SIAR for non marked events use
302 	 * pt_regs.
303 	 *
304 	 * If the PMU has HV/PR flags then check to see if they
305 	 * place the exception in userspace. If so, use pt_regs. In
306 	 * continuous sampling mode the SIAR and the PMU exception are
307 	 * not synchronised, so they may be many instructions apart.
308 	 * This can result in confusing backtraces. We still want
309 	 * hypervisor samples as well as samples in the kernel with
310 	 * interrupts off hence the userspace check.
311 	 */
312 	if (TRAP(regs) != 0xf00)
313 		use_siar = 0;
314 	else if ((ppmu->flags & PPMU_NO_SIAR))
315 		use_siar = 0;
316 	else if (marked)
317 		use_siar = 1;
318 	else if ((ppmu->flags & PPMU_NO_CONT_SAMPLING))
319 		use_siar = 0;
320 	else if (!(ppmu->flags & PPMU_NO_SIPR) && regs_sipr(regs))
321 		use_siar = 0;
322 	else
323 		use_siar = 1;
324 
325 	regs->result = use_siar;
326 }
327 
328 /*
329  * If interrupts were soft-disabled when a PMU interrupt occurs, treat
330  * it as an NMI.
331  */
332 static inline int perf_intr_is_nmi(struct pt_regs *regs)
333 {
334 	return (regs->softe & IRQS_DISABLED);
335 }
336 
337 /*
338  * On processors like P7+ that have the SIAR-Valid bit, marked instructions
339  * must be sampled only if the SIAR-valid bit is set.
340  *
341  * For unmarked instructions and for processors that don't have the SIAR-Valid
342  * bit, assume that SIAR is valid.
343  */
344 static inline int siar_valid(struct pt_regs *regs)
345 {
346 	unsigned long mmcra = regs->dsisr;
347 	int marked = mmcra & MMCRA_SAMPLE_ENABLE;
348 
349 	if (marked) {
350 		if (ppmu->flags & PPMU_HAS_SIER)
351 			return regs->dar & SIER_SIAR_VALID;
352 
353 		if (ppmu->flags & PPMU_SIAR_VALID)
354 			return mmcra & POWER7P_MMCRA_SIAR_VALID;
355 	}
356 
357 	return 1;
358 }
359 
360 
361 /* Reset all possible BHRB entries */
362 static void power_pmu_bhrb_reset(void)
363 {
364 	asm volatile(PPC_CLRBHRB);
365 }
366 
367 static void power_pmu_bhrb_enable(struct perf_event *event)
368 {
369 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
370 
371 	if (!ppmu->bhrb_nr)
372 		return;
373 
374 	/* Clear BHRB if we changed task context to avoid data leaks */
375 	if (event->ctx->task && cpuhw->bhrb_context != event->ctx) {
376 		power_pmu_bhrb_reset();
377 		cpuhw->bhrb_context = event->ctx;
378 	}
379 	cpuhw->bhrb_users++;
380 	perf_sched_cb_inc(event->ctx->pmu);
381 }
382 
383 static void power_pmu_bhrb_disable(struct perf_event *event)
384 {
385 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
386 
387 	if (!ppmu->bhrb_nr)
388 		return;
389 
390 	WARN_ON_ONCE(!cpuhw->bhrb_users);
391 	cpuhw->bhrb_users--;
392 	perf_sched_cb_dec(event->ctx->pmu);
393 
394 	if (!cpuhw->disabled && !cpuhw->bhrb_users) {
395 		/* BHRB cannot be turned off when other
396 		 * events are active on the PMU.
397 		 */
398 
399 		/* avoid stale pointer */
400 		cpuhw->bhrb_context = NULL;
401 	}
402 }
403 
404 /* Called from ctxsw to prevent one process's branch entries to
405  * mingle with the other process's entries during context switch.
406  */
407 static void power_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
408 {
409 	if (!ppmu->bhrb_nr)
410 		return;
411 
412 	if (sched_in)
413 		power_pmu_bhrb_reset();
414 }
415 /* Calculate the to address for a branch */
416 static __u64 power_pmu_bhrb_to(u64 addr)
417 {
418 	unsigned int instr;
419 	int ret;
420 	__u64 target;
421 
422 	if (is_kernel_addr(addr)) {
423 		if (probe_kernel_read(&instr, (void *)addr, sizeof(instr)))
424 			return 0;
425 
426 		return branch_target(&instr);
427 	}
428 
429 	/* Userspace: need copy instruction here then translate it */
430 	pagefault_disable();
431 	ret = __get_user_inatomic(instr, (unsigned int __user *)addr);
432 	if (ret) {
433 		pagefault_enable();
434 		return 0;
435 	}
436 	pagefault_enable();
437 
438 	target = branch_target(&instr);
439 	if ((!target) || (instr & BRANCH_ABSOLUTE))
440 		return target;
441 
442 	/* Translate relative branch target from kernel to user address */
443 	return target - (unsigned long)&instr + addr;
444 }
445 
446 /* Processing BHRB entries */
447 static void power_pmu_bhrb_read(struct cpu_hw_events *cpuhw)
448 {
449 	u64 val;
450 	u64 addr;
451 	int r_index, u_index, pred;
452 
453 	r_index = 0;
454 	u_index = 0;
455 	while (r_index < ppmu->bhrb_nr) {
456 		/* Assembly read function */
457 		val = read_bhrb(r_index++);
458 		if (!val)
459 			/* Terminal marker: End of valid BHRB entries */
460 			break;
461 		else {
462 			addr = val & BHRB_EA;
463 			pred = val & BHRB_PREDICTION;
464 
465 			if (!addr)
466 				/* invalid entry */
467 				continue;
468 
469 			/*
470 			 * BHRB rolling buffer could very much contain the kernel
471 			 * addresses at this point. Check the privileges before
472 			 * exporting it to userspace (avoid exposure of regions
473 			 * where we could have speculative execution)
474 			 */
475 			if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN) &&
476 				is_kernel_addr(addr))
477 				continue;
478 
479 			/* Branches are read most recent first (ie. mfbhrb 0 is
480 			 * the most recent branch).
481 			 * There are two types of valid entries:
482 			 * 1) a target entry which is the to address of a
483 			 *    computed goto like a blr,bctr,btar.  The next
484 			 *    entry read from the bhrb will be branch
485 			 *    corresponding to this target (ie. the actual
486 			 *    blr/bctr/btar instruction).
487 			 * 2) a from address which is an actual branch.  If a
488 			 *    target entry proceeds this, then this is the
489 			 *    matching branch for that target.  If this is not
490 			 *    following a target entry, then this is a branch
491 			 *    where the target is given as an immediate field
492 			 *    in the instruction (ie. an i or b form branch).
493 			 *    In this case we need to read the instruction from
494 			 *    memory to determine the target/to address.
495 			 */
496 
497 			if (val & BHRB_TARGET) {
498 				/* Target branches use two entries
499 				 * (ie. computed gotos/XL form)
500 				 */
501 				cpuhw->bhrb_entries[u_index].to = addr;
502 				cpuhw->bhrb_entries[u_index].mispred = pred;
503 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
504 
505 				/* Get from address in next entry */
506 				val = read_bhrb(r_index++);
507 				addr = val & BHRB_EA;
508 				if (val & BHRB_TARGET) {
509 					/* Shouldn't have two targets in a
510 					   row.. Reset index and try again */
511 					r_index--;
512 					addr = 0;
513 				}
514 				cpuhw->bhrb_entries[u_index].from = addr;
515 			} else {
516 				/* Branches to immediate field
517 				   (ie I or B form) */
518 				cpuhw->bhrb_entries[u_index].from = addr;
519 				cpuhw->bhrb_entries[u_index].to =
520 					power_pmu_bhrb_to(addr);
521 				cpuhw->bhrb_entries[u_index].mispred = pred;
522 				cpuhw->bhrb_entries[u_index].predicted = ~pred;
523 			}
524 			u_index++;
525 
526 		}
527 	}
528 	cpuhw->bhrb_stack.nr = u_index;
529 	return;
530 }
531 
532 static bool is_ebb_event(struct perf_event *event)
533 {
534 	/*
535 	 * This could be a per-PMU callback, but we'd rather avoid the cost. We
536 	 * check that the PMU supports EBB, meaning those that don't can still
537 	 * use bit 63 of the event code for something else if they wish.
538 	 */
539 	return (ppmu->flags & PPMU_ARCH_207S) &&
540 	       ((event->attr.config >> PERF_EVENT_CONFIG_EBB_SHIFT) & 1);
541 }
542 
543 static int ebb_event_check(struct perf_event *event)
544 {
545 	struct perf_event *leader = event->group_leader;
546 
547 	/* Event and group leader must agree on EBB */
548 	if (is_ebb_event(leader) != is_ebb_event(event))
549 		return -EINVAL;
550 
551 	if (is_ebb_event(event)) {
552 		if (!(event->attach_state & PERF_ATTACH_TASK))
553 			return -EINVAL;
554 
555 		if (!leader->attr.pinned || !leader->attr.exclusive)
556 			return -EINVAL;
557 
558 		if (event->attr.freq ||
559 		    event->attr.inherit ||
560 		    event->attr.sample_type ||
561 		    event->attr.sample_period ||
562 		    event->attr.enable_on_exec)
563 			return -EINVAL;
564 	}
565 
566 	return 0;
567 }
568 
569 static void ebb_event_add(struct perf_event *event)
570 {
571 	if (!is_ebb_event(event) || current->thread.used_ebb)
572 		return;
573 
574 	/*
575 	 * IFF this is the first time we've added an EBB event, set
576 	 * PMXE in the user MMCR0 so we can detect when it's cleared by
577 	 * userspace. We need this so that we can context switch while
578 	 * userspace is in the EBB handler (where PMXE is 0).
579 	 */
580 	current->thread.used_ebb = 1;
581 	current->thread.mmcr0 |= MMCR0_PMXE;
582 }
583 
584 static void ebb_switch_out(unsigned long mmcr0)
585 {
586 	if (!(mmcr0 & MMCR0_EBE))
587 		return;
588 
589 	current->thread.siar  = mfspr(SPRN_SIAR);
590 	current->thread.sier  = mfspr(SPRN_SIER);
591 	current->thread.sdar  = mfspr(SPRN_SDAR);
592 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
593 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
594 }
595 
596 static unsigned long ebb_switch_in(bool ebb, struct cpu_hw_events *cpuhw)
597 {
598 	unsigned long mmcr0 = cpuhw->mmcr[0];
599 
600 	if (!ebb)
601 		goto out;
602 
603 	/* Enable EBB and read/write to all 6 PMCs and BHRB for userspace */
604 	mmcr0 |= MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC_U6;
605 
606 	/*
607 	 * Add any bits from the user MMCR0, FC or PMAO. This is compatible
608 	 * with pmao_restore_workaround() because we may add PMAO but we never
609 	 * clear it here.
610 	 */
611 	mmcr0 |= current->thread.mmcr0;
612 
613 	/*
614 	 * Be careful not to set PMXE if userspace had it cleared. This is also
615 	 * compatible with pmao_restore_workaround() because it has already
616 	 * cleared PMXE and we leave PMAO alone.
617 	 */
618 	if (!(current->thread.mmcr0 & MMCR0_PMXE))
619 		mmcr0 &= ~MMCR0_PMXE;
620 
621 	mtspr(SPRN_SIAR, current->thread.siar);
622 	mtspr(SPRN_SIER, current->thread.sier);
623 	mtspr(SPRN_SDAR, current->thread.sdar);
624 
625 	/*
626 	 * Merge the kernel & user values of MMCR2. The semantics we implement
627 	 * are that the user MMCR2 can set bits, ie. cause counters to freeze,
628 	 * but not clear bits. If a task wants to be able to clear bits, ie.
629 	 * unfreeze counters, it should not set exclude_xxx in its events and
630 	 * instead manage the MMCR2 entirely by itself.
631 	 */
632 	mtspr(SPRN_MMCR2, cpuhw->mmcr[3] | current->thread.mmcr2);
633 out:
634 	return mmcr0;
635 }
636 
637 static void pmao_restore_workaround(bool ebb)
638 {
639 	unsigned pmcs[6];
640 
641 	if (!cpu_has_feature(CPU_FTR_PMAO_BUG))
642 		return;
643 
644 	/*
645 	 * On POWER8E there is a hardware defect which affects the PMU context
646 	 * switch logic, ie. power_pmu_disable/enable().
647 	 *
648 	 * When a counter overflows PMXE is cleared and FC/PMAO is set in MMCR0
649 	 * by the hardware. Sometime later the actual PMU exception is
650 	 * delivered.
651 	 *
652 	 * If we context switch, or simply disable/enable, the PMU prior to the
653 	 * exception arriving, the exception will be lost when we clear PMAO.
654 	 *
655 	 * When we reenable the PMU, we will write the saved MMCR0 with PMAO
656 	 * set, and this _should_ generate an exception. However because of the
657 	 * defect no exception is generated when we write PMAO, and we get
658 	 * stuck with no counters counting but no exception delivered.
659 	 *
660 	 * The workaround is to detect this case and tweak the hardware to
661 	 * create another pending PMU exception.
662 	 *
663 	 * We do that by setting up PMC6 (cycles) for an imminent overflow and
664 	 * enabling the PMU. That causes a new exception to be generated in the
665 	 * chip, but we don't take it yet because we have interrupts hard
666 	 * disabled. We then write back the PMU state as we want it to be seen
667 	 * by the exception handler. When we reenable interrupts the exception
668 	 * handler will be called and see the correct state.
669 	 *
670 	 * The logic is the same for EBB, except that the exception is gated by
671 	 * us having interrupts hard disabled as well as the fact that we are
672 	 * not in userspace. The exception is finally delivered when we return
673 	 * to userspace.
674 	 */
675 
676 	/* Only if PMAO is set and PMAO_SYNC is clear */
677 	if ((current->thread.mmcr0 & (MMCR0_PMAO | MMCR0_PMAO_SYNC)) != MMCR0_PMAO)
678 		return;
679 
680 	/* If we're doing EBB, only if BESCR[GE] is set */
681 	if (ebb && !(current->thread.bescr & BESCR_GE))
682 		return;
683 
684 	/*
685 	 * We are already soft-disabled in power_pmu_enable(). We need to hard
686 	 * disable to actually prevent the PMU exception from firing.
687 	 */
688 	hard_irq_disable();
689 
690 	/*
691 	 * This is a bit gross, but we know we're on POWER8E and have 6 PMCs.
692 	 * Using read/write_pmc() in a for loop adds 12 function calls and
693 	 * almost doubles our code size.
694 	 */
695 	pmcs[0] = mfspr(SPRN_PMC1);
696 	pmcs[1] = mfspr(SPRN_PMC2);
697 	pmcs[2] = mfspr(SPRN_PMC3);
698 	pmcs[3] = mfspr(SPRN_PMC4);
699 	pmcs[4] = mfspr(SPRN_PMC5);
700 	pmcs[5] = mfspr(SPRN_PMC6);
701 
702 	/* Ensure all freeze bits are unset */
703 	mtspr(SPRN_MMCR2, 0);
704 
705 	/* Set up PMC6 to overflow in one cycle */
706 	mtspr(SPRN_PMC6, 0x7FFFFFFE);
707 
708 	/* Enable exceptions and unfreeze PMC6 */
709 	mtspr(SPRN_MMCR0, MMCR0_PMXE | MMCR0_PMCjCE | MMCR0_PMAO);
710 
711 	/* Now we need to refreeze and restore the PMCs */
712 	mtspr(SPRN_MMCR0, MMCR0_FC | MMCR0_PMAO);
713 
714 	mtspr(SPRN_PMC1, pmcs[0]);
715 	mtspr(SPRN_PMC2, pmcs[1]);
716 	mtspr(SPRN_PMC3, pmcs[2]);
717 	mtspr(SPRN_PMC4, pmcs[3]);
718 	mtspr(SPRN_PMC5, pmcs[4]);
719 	mtspr(SPRN_PMC6, pmcs[5]);
720 }
721 
722 #endif /* CONFIG_PPC64 */
723 
724 static void perf_event_interrupt(struct pt_regs *regs);
725 
726 /*
727  * Read one performance monitor counter (PMC).
728  */
729 static unsigned long read_pmc(int idx)
730 {
731 	unsigned long val;
732 
733 	switch (idx) {
734 	case 1:
735 		val = mfspr(SPRN_PMC1);
736 		break;
737 	case 2:
738 		val = mfspr(SPRN_PMC2);
739 		break;
740 	case 3:
741 		val = mfspr(SPRN_PMC3);
742 		break;
743 	case 4:
744 		val = mfspr(SPRN_PMC4);
745 		break;
746 	case 5:
747 		val = mfspr(SPRN_PMC5);
748 		break;
749 	case 6:
750 		val = mfspr(SPRN_PMC6);
751 		break;
752 #ifdef CONFIG_PPC64
753 	case 7:
754 		val = mfspr(SPRN_PMC7);
755 		break;
756 	case 8:
757 		val = mfspr(SPRN_PMC8);
758 		break;
759 #endif /* CONFIG_PPC64 */
760 	default:
761 		printk(KERN_ERR "oops trying to read PMC%d\n", idx);
762 		val = 0;
763 	}
764 	return val;
765 }
766 
767 /*
768  * Write one PMC.
769  */
770 static void write_pmc(int idx, unsigned long val)
771 {
772 	switch (idx) {
773 	case 1:
774 		mtspr(SPRN_PMC1, val);
775 		break;
776 	case 2:
777 		mtspr(SPRN_PMC2, val);
778 		break;
779 	case 3:
780 		mtspr(SPRN_PMC3, val);
781 		break;
782 	case 4:
783 		mtspr(SPRN_PMC4, val);
784 		break;
785 	case 5:
786 		mtspr(SPRN_PMC5, val);
787 		break;
788 	case 6:
789 		mtspr(SPRN_PMC6, val);
790 		break;
791 #ifdef CONFIG_PPC64
792 	case 7:
793 		mtspr(SPRN_PMC7, val);
794 		break;
795 	case 8:
796 		mtspr(SPRN_PMC8, val);
797 		break;
798 #endif /* CONFIG_PPC64 */
799 	default:
800 		printk(KERN_ERR "oops trying to write PMC%d\n", idx);
801 	}
802 }
803 
804 /* Called from sysrq_handle_showregs() */
805 void perf_event_print_debug(void)
806 {
807 	unsigned long sdar, sier, flags;
808 	u32 pmcs[MAX_HWEVENTS];
809 	int i;
810 
811 	if (!ppmu) {
812 		pr_info("Performance monitor hardware not registered.\n");
813 		return;
814 	}
815 
816 	if (!ppmu->n_counter)
817 		return;
818 
819 	local_irq_save(flags);
820 
821 	pr_info("CPU: %d PMU registers, ppmu = %s n_counters = %d",
822 		 smp_processor_id(), ppmu->name, ppmu->n_counter);
823 
824 	for (i = 0; i < ppmu->n_counter; i++)
825 		pmcs[i] = read_pmc(i + 1);
826 
827 	for (; i < MAX_HWEVENTS; i++)
828 		pmcs[i] = 0xdeadbeef;
829 
830 	pr_info("PMC1:  %08x PMC2: %08x PMC3: %08x PMC4: %08x\n",
831 		 pmcs[0], pmcs[1], pmcs[2], pmcs[3]);
832 
833 	if (ppmu->n_counter > 4)
834 		pr_info("PMC5:  %08x PMC6: %08x PMC7: %08x PMC8: %08x\n",
835 			 pmcs[4], pmcs[5], pmcs[6], pmcs[7]);
836 
837 	pr_info("MMCR0: %016lx MMCR1: %016lx MMCRA: %016lx\n",
838 		mfspr(SPRN_MMCR0), mfspr(SPRN_MMCR1), mfspr(SPRN_MMCRA));
839 
840 	sdar = sier = 0;
841 #ifdef CONFIG_PPC64
842 	sdar = mfspr(SPRN_SDAR);
843 
844 	if (ppmu->flags & PPMU_HAS_SIER)
845 		sier = mfspr(SPRN_SIER);
846 
847 	if (ppmu->flags & PPMU_ARCH_207S) {
848 		pr_info("MMCR2: %016lx EBBHR: %016lx\n",
849 			mfspr(SPRN_MMCR2), mfspr(SPRN_EBBHR));
850 		pr_info("EBBRR: %016lx BESCR: %016lx\n",
851 			mfspr(SPRN_EBBRR), mfspr(SPRN_BESCR));
852 	}
853 #endif
854 	pr_info("SIAR:  %016lx SDAR:  %016lx SIER:  %016lx\n",
855 		mfspr(SPRN_SIAR), sdar, sier);
856 
857 	local_irq_restore(flags);
858 }
859 
860 /*
861  * Check if a set of events can all go on the PMU at once.
862  * If they can't, this will look at alternative codes for the events
863  * and see if any combination of alternative codes is feasible.
864  * The feasible set is returned in event_id[].
865  */
866 static int power_check_constraints(struct cpu_hw_events *cpuhw,
867 				   u64 event_id[], unsigned int cflags[],
868 				   int n_ev)
869 {
870 	unsigned long mask, value, nv;
871 	unsigned long smasks[MAX_HWEVENTS], svalues[MAX_HWEVENTS];
872 	int n_alt[MAX_HWEVENTS], choice[MAX_HWEVENTS];
873 	int i, j;
874 	unsigned long addf = ppmu->add_fields;
875 	unsigned long tadd = ppmu->test_adder;
876 	unsigned long grp_mask = ppmu->group_constraint_mask;
877 	unsigned long grp_val = ppmu->group_constraint_val;
878 
879 	if (n_ev > ppmu->n_counter)
880 		return -1;
881 
882 	/* First see if the events will go on as-is */
883 	for (i = 0; i < n_ev; ++i) {
884 		if ((cflags[i] & PPMU_LIMITED_PMC_REQD)
885 		    && !ppmu->limited_pmc_event(event_id[i])) {
886 			ppmu->get_alternatives(event_id[i], cflags[i],
887 					       cpuhw->alternatives[i]);
888 			event_id[i] = cpuhw->alternatives[i][0];
889 		}
890 		if (ppmu->get_constraint(event_id[i], &cpuhw->amasks[i][0],
891 					 &cpuhw->avalues[i][0]))
892 			return -1;
893 	}
894 	value = mask = 0;
895 	for (i = 0; i < n_ev; ++i) {
896 		nv = (value | cpuhw->avalues[i][0]) +
897 			(value & cpuhw->avalues[i][0] & addf);
898 
899 		if (((((nv + tadd) ^ value) & mask) & (~grp_mask)) != 0)
900 			break;
901 
902 		if (((((nv + tadd) ^ cpuhw->avalues[i][0]) & cpuhw->amasks[i][0])
903 			& (~grp_mask)) != 0)
904 			break;
905 
906 		value = nv;
907 		mask |= cpuhw->amasks[i][0];
908 	}
909 	if (i == n_ev) {
910 		if ((value & mask & grp_mask) != (mask & grp_val))
911 			return -1;
912 		else
913 			return 0;	/* all OK */
914 	}
915 
916 	/* doesn't work, gather alternatives... */
917 	if (!ppmu->get_alternatives)
918 		return -1;
919 	for (i = 0; i < n_ev; ++i) {
920 		choice[i] = 0;
921 		n_alt[i] = ppmu->get_alternatives(event_id[i], cflags[i],
922 						  cpuhw->alternatives[i]);
923 		for (j = 1; j < n_alt[i]; ++j)
924 			ppmu->get_constraint(cpuhw->alternatives[i][j],
925 					     &cpuhw->amasks[i][j],
926 					     &cpuhw->avalues[i][j]);
927 	}
928 
929 	/* enumerate all possibilities and see if any will work */
930 	i = 0;
931 	j = -1;
932 	value = mask = nv = 0;
933 	while (i < n_ev) {
934 		if (j >= 0) {
935 			/* we're backtracking, restore context */
936 			value = svalues[i];
937 			mask = smasks[i];
938 			j = choice[i];
939 		}
940 		/*
941 		 * See if any alternative k for event_id i,
942 		 * where k > j, will satisfy the constraints.
943 		 */
944 		while (++j < n_alt[i]) {
945 			nv = (value | cpuhw->avalues[i][j]) +
946 				(value & cpuhw->avalues[i][j] & addf);
947 			if ((((nv + tadd) ^ value) & mask) == 0 &&
948 			    (((nv + tadd) ^ cpuhw->avalues[i][j])
949 			     & cpuhw->amasks[i][j]) == 0)
950 				break;
951 		}
952 		if (j >= n_alt[i]) {
953 			/*
954 			 * No feasible alternative, backtrack
955 			 * to event_id i-1 and continue enumerating its
956 			 * alternatives from where we got up to.
957 			 */
958 			if (--i < 0)
959 				return -1;
960 		} else {
961 			/*
962 			 * Found a feasible alternative for event_id i,
963 			 * remember where we got up to with this event_id,
964 			 * go on to the next event_id, and start with
965 			 * the first alternative for it.
966 			 */
967 			choice[i] = j;
968 			svalues[i] = value;
969 			smasks[i] = mask;
970 			value = nv;
971 			mask |= cpuhw->amasks[i][j];
972 			++i;
973 			j = -1;
974 		}
975 	}
976 
977 	/* OK, we have a feasible combination, tell the caller the solution */
978 	for (i = 0; i < n_ev; ++i)
979 		event_id[i] = cpuhw->alternatives[i][choice[i]];
980 	return 0;
981 }
982 
983 /*
984  * Check if newly-added events have consistent settings for
985  * exclude_{user,kernel,hv} with each other and any previously
986  * added events.
987  */
988 static int check_excludes(struct perf_event **ctrs, unsigned int cflags[],
989 			  int n_prev, int n_new)
990 {
991 	int eu = 0, ek = 0, eh = 0;
992 	int i, n, first;
993 	struct perf_event *event;
994 
995 	/*
996 	 * If the PMU we're on supports per event exclude settings then we
997 	 * don't need to do any of this logic. NB. This assumes no PMU has both
998 	 * per event exclude and limited PMCs.
999 	 */
1000 	if (ppmu->flags & PPMU_ARCH_207S)
1001 		return 0;
1002 
1003 	n = n_prev + n_new;
1004 	if (n <= 1)
1005 		return 0;
1006 
1007 	first = 1;
1008 	for (i = 0; i < n; ++i) {
1009 		if (cflags[i] & PPMU_LIMITED_PMC_OK) {
1010 			cflags[i] &= ~PPMU_LIMITED_PMC_REQD;
1011 			continue;
1012 		}
1013 		event = ctrs[i];
1014 		if (first) {
1015 			eu = event->attr.exclude_user;
1016 			ek = event->attr.exclude_kernel;
1017 			eh = event->attr.exclude_hv;
1018 			first = 0;
1019 		} else if (event->attr.exclude_user != eu ||
1020 			   event->attr.exclude_kernel != ek ||
1021 			   event->attr.exclude_hv != eh) {
1022 			return -EAGAIN;
1023 		}
1024 	}
1025 
1026 	if (eu || ek || eh)
1027 		for (i = 0; i < n; ++i)
1028 			if (cflags[i] & PPMU_LIMITED_PMC_OK)
1029 				cflags[i] |= PPMU_LIMITED_PMC_REQD;
1030 
1031 	return 0;
1032 }
1033 
1034 static u64 check_and_compute_delta(u64 prev, u64 val)
1035 {
1036 	u64 delta = (val - prev) & 0xfffffffful;
1037 
1038 	/*
1039 	 * POWER7 can roll back counter values, if the new value is smaller
1040 	 * than the previous value it will cause the delta and the counter to
1041 	 * have bogus values unless we rolled a counter over.  If a coutner is
1042 	 * rolled back, it will be smaller, but within 256, which is the maximum
1043 	 * number of events to rollback at once.  If we detect a rollback
1044 	 * return 0.  This can lead to a small lack of precision in the
1045 	 * counters.
1046 	 */
1047 	if (prev > val && (prev - val) < 256)
1048 		delta = 0;
1049 
1050 	return delta;
1051 }
1052 
1053 static void power_pmu_read(struct perf_event *event)
1054 {
1055 	s64 val, delta, prev;
1056 
1057 	if (event->hw.state & PERF_HES_STOPPED)
1058 		return;
1059 
1060 	if (!event->hw.idx)
1061 		return;
1062 
1063 	if (is_ebb_event(event)) {
1064 		val = read_pmc(event->hw.idx);
1065 		local64_set(&event->hw.prev_count, val);
1066 		return;
1067 	}
1068 
1069 	/*
1070 	 * Performance monitor interrupts come even when interrupts
1071 	 * are soft-disabled, as long as interrupts are hard-enabled.
1072 	 * Therefore we treat them like NMIs.
1073 	 */
1074 	do {
1075 		prev = local64_read(&event->hw.prev_count);
1076 		barrier();
1077 		val = read_pmc(event->hw.idx);
1078 		delta = check_and_compute_delta(prev, val);
1079 		if (!delta)
1080 			return;
1081 	} while (local64_cmpxchg(&event->hw.prev_count, prev, val) != prev);
1082 
1083 	local64_add(delta, &event->count);
1084 
1085 	/*
1086 	 * A number of places program the PMC with (0x80000000 - period_left).
1087 	 * We never want period_left to be less than 1 because we will program
1088 	 * the PMC with a value >= 0x800000000 and an edge detected PMC will
1089 	 * roll around to 0 before taking an exception. We have seen this
1090 	 * on POWER8.
1091 	 *
1092 	 * To fix this, clamp the minimum value of period_left to 1.
1093 	 */
1094 	do {
1095 		prev = local64_read(&event->hw.period_left);
1096 		val = prev - delta;
1097 		if (val < 1)
1098 			val = 1;
1099 	} while (local64_cmpxchg(&event->hw.period_left, prev, val) != prev);
1100 }
1101 
1102 /*
1103  * On some machines, PMC5 and PMC6 can't be written, don't respect
1104  * the freeze conditions, and don't generate interrupts.  This tells
1105  * us if `event' is using such a PMC.
1106  */
1107 static int is_limited_pmc(int pmcnum)
1108 {
1109 	return (ppmu->flags & PPMU_LIMITED_PMC5_6)
1110 		&& (pmcnum == 5 || pmcnum == 6);
1111 }
1112 
1113 static void freeze_limited_counters(struct cpu_hw_events *cpuhw,
1114 				    unsigned long pmc5, unsigned long pmc6)
1115 {
1116 	struct perf_event *event;
1117 	u64 val, prev, delta;
1118 	int i;
1119 
1120 	for (i = 0; i < cpuhw->n_limited; ++i) {
1121 		event = cpuhw->limited_counter[i];
1122 		if (!event->hw.idx)
1123 			continue;
1124 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1125 		prev = local64_read(&event->hw.prev_count);
1126 		event->hw.idx = 0;
1127 		delta = check_and_compute_delta(prev, val);
1128 		if (delta)
1129 			local64_add(delta, &event->count);
1130 	}
1131 }
1132 
1133 static void thaw_limited_counters(struct cpu_hw_events *cpuhw,
1134 				  unsigned long pmc5, unsigned long pmc6)
1135 {
1136 	struct perf_event *event;
1137 	u64 val, prev;
1138 	int i;
1139 
1140 	for (i = 0; i < cpuhw->n_limited; ++i) {
1141 		event = cpuhw->limited_counter[i];
1142 		event->hw.idx = cpuhw->limited_hwidx[i];
1143 		val = (event->hw.idx == 5) ? pmc5 : pmc6;
1144 		prev = local64_read(&event->hw.prev_count);
1145 		if (check_and_compute_delta(prev, val))
1146 			local64_set(&event->hw.prev_count, val);
1147 		perf_event_update_userpage(event);
1148 	}
1149 }
1150 
1151 /*
1152  * Since limited events don't respect the freeze conditions, we
1153  * have to read them immediately after freezing or unfreezing the
1154  * other events.  We try to keep the values from the limited
1155  * events as consistent as possible by keeping the delay (in
1156  * cycles and instructions) between freezing/unfreezing and reading
1157  * the limited events as small and consistent as possible.
1158  * Therefore, if any limited events are in use, we read them
1159  * both, and always in the same order, to minimize variability,
1160  * and do it inside the same asm that writes MMCR0.
1161  */
1162 static void write_mmcr0(struct cpu_hw_events *cpuhw, unsigned long mmcr0)
1163 {
1164 	unsigned long pmc5, pmc6;
1165 
1166 	if (!cpuhw->n_limited) {
1167 		mtspr(SPRN_MMCR0, mmcr0);
1168 		return;
1169 	}
1170 
1171 	/*
1172 	 * Write MMCR0, then read PMC5 and PMC6 immediately.
1173 	 * To ensure we don't get a performance monitor interrupt
1174 	 * between writing MMCR0 and freezing/thawing the limited
1175 	 * events, we first write MMCR0 with the event overflow
1176 	 * interrupt enable bits turned off.
1177 	 */
1178 	asm volatile("mtspr %3,%2; mfspr %0,%4; mfspr %1,%5"
1179 		     : "=&r" (pmc5), "=&r" (pmc6)
1180 		     : "r" (mmcr0 & ~(MMCR0_PMC1CE | MMCR0_PMCjCE)),
1181 		       "i" (SPRN_MMCR0),
1182 		       "i" (SPRN_PMC5), "i" (SPRN_PMC6));
1183 
1184 	if (mmcr0 & MMCR0_FC)
1185 		freeze_limited_counters(cpuhw, pmc5, pmc6);
1186 	else
1187 		thaw_limited_counters(cpuhw, pmc5, pmc6);
1188 
1189 	/*
1190 	 * Write the full MMCR0 including the event overflow interrupt
1191 	 * enable bits, if necessary.
1192 	 */
1193 	if (mmcr0 & (MMCR0_PMC1CE | MMCR0_PMCjCE))
1194 		mtspr(SPRN_MMCR0, mmcr0);
1195 }
1196 
1197 /*
1198  * Disable all events to prevent PMU interrupts and to allow
1199  * events to be added or removed.
1200  */
1201 static void power_pmu_disable(struct pmu *pmu)
1202 {
1203 	struct cpu_hw_events *cpuhw;
1204 	unsigned long flags, mmcr0, val;
1205 
1206 	if (!ppmu)
1207 		return;
1208 	local_irq_save(flags);
1209 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1210 
1211 	if (!cpuhw->disabled) {
1212 		/*
1213 		 * Check if we ever enabled the PMU on this cpu.
1214 		 */
1215 		if (!cpuhw->pmcs_enabled) {
1216 			ppc_enable_pmcs();
1217 			cpuhw->pmcs_enabled = 1;
1218 		}
1219 
1220 		/*
1221 		 * Set the 'freeze counters' bit, clear EBE/BHRBA/PMCC/PMAO/FC56
1222 		 */
1223 		val  = mmcr0 = mfspr(SPRN_MMCR0);
1224 		val |= MMCR0_FC;
1225 		val &= ~(MMCR0_EBE | MMCR0_BHRBA | MMCR0_PMCC | MMCR0_PMAO |
1226 			 MMCR0_FC56);
1227 
1228 		/*
1229 		 * The barrier is to make sure the mtspr has been
1230 		 * executed and the PMU has frozen the events etc.
1231 		 * before we return.
1232 		 */
1233 		write_mmcr0(cpuhw, val);
1234 		mb();
1235 		isync();
1236 
1237 		/*
1238 		 * Disable instruction sampling if it was enabled
1239 		 */
1240 		if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1241 			mtspr(SPRN_MMCRA,
1242 			      cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1243 			mb();
1244 			isync();
1245 		}
1246 
1247 		cpuhw->disabled = 1;
1248 		cpuhw->n_added = 0;
1249 
1250 		ebb_switch_out(mmcr0);
1251 
1252 #ifdef CONFIG_PPC64
1253 		/*
1254 		 * These are readable by userspace, may contain kernel
1255 		 * addresses and are not switched by context switch, so clear
1256 		 * them now to avoid leaking anything to userspace in general
1257 		 * including to another process.
1258 		 */
1259 		if (ppmu->flags & PPMU_ARCH_207S) {
1260 			mtspr(SPRN_SDAR, 0);
1261 			mtspr(SPRN_SIAR, 0);
1262 		}
1263 #endif
1264 	}
1265 
1266 	local_irq_restore(flags);
1267 }
1268 
1269 /*
1270  * Re-enable all events if disable == 0.
1271  * If we were previously disabled and events were added, then
1272  * put the new config on the PMU.
1273  */
1274 static void power_pmu_enable(struct pmu *pmu)
1275 {
1276 	struct perf_event *event;
1277 	struct cpu_hw_events *cpuhw;
1278 	unsigned long flags;
1279 	long i;
1280 	unsigned long val, mmcr0;
1281 	s64 left;
1282 	unsigned int hwc_index[MAX_HWEVENTS];
1283 	int n_lim;
1284 	int idx;
1285 	bool ebb;
1286 
1287 	if (!ppmu)
1288 		return;
1289 	local_irq_save(flags);
1290 
1291 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1292 	if (!cpuhw->disabled)
1293 		goto out;
1294 
1295 	if (cpuhw->n_events == 0) {
1296 		ppc_set_pmu_inuse(0);
1297 		goto out;
1298 	}
1299 
1300 	cpuhw->disabled = 0;
1301 
1302 	/*
1303 	 * EBB requires an exclusive group and all events must have the EBB
1304 	 * flag set, or not set, so we can just check a single event. Also we
1305 	 * know we have at least one event.
1306 	 */
1307 	ebb = is_ebb_event(cpuhw->event[0]);
1308 
1309 	/*
1310 	 * If we didn't change anything, or only removed events,
1311 	 * no need to recalculate MMCR* settings and reset the PMCs.
1312 	 * Just reenable the PMU with the current MMCR* settings
1313 	 * (possibly updated for removal of events).
1314 	 */
1315 	if (!cpuhw->n_added) {
1316 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1317 		mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1318 		goto out_enable;
1319 	}
1320 
1321 	/*
1322 	 * Clear all MMCR settings and recompute them for the new set of events.
1323 	 */
1324 	memset(cpuhw->mmcr, 0, sizeof(cpuhw->mmcr));
1325 
1326 	if (ppmu->compute_mmcr(cpuhw->events, cpuhw->n_events, hwc_index,
1327 			       cpuhw->mmcr, cpuhw->event)) {
1328 		/* shouldn't ever get here */
1329 		printk(KERN_ERR "oops compute_mmcr failed\n");
1330 		goto out;
1331 	}
1332 
1333 	if (!(ppmu->flags & PPMU_ARCH_207S)) {
1334 		/*
1335 		 * Add in MMCR0 freeze bits corresponding to the attr.exclude_*
1336 		 * bits for the first event. We have already checked that all
1337 		 * events have the same value for these bits as the first event.
1338 		 */
1339 		event = cpuhw->event[0];
1340 		if (event->attr.exclude_user)
1341 			cpuhw->mmcr[0] |= MMCR0_FCP;
1342 		if (event->attr.exclude_kernel)
1343 			cpuhw->mmcr[0] |= freeze_events_kernel;
1344 		if (event->attr.exclude_hv)
1345 			cpuhw->mmcr[0] |= MMCR0_FCHV;
1346 	}
1347 
1348 	/*
1349 	 * Write the new configuration to MMCR* with the freeze
1350 	 * bit set and set the hardware events to their initial values.
1351 	 * Then unfreeze the events.
1352 	 */
1353 	ppc_set_pmu_inuse(1);
1354 	mtspr(SPRN_MMCRA, cpuhw->mmcr[2] & ~MMCRA_SAMPLE_ENABLE);
1355 	mtspr(SPRN_MMCR1, cpuhw->mmcr[1]);
1356 	mtspr(SPRN_MMCR0, (cpuhw->mmcr[0] & ~(MMCR0_PMC1CE | MMCR0_PMCjCE))
1357 				| MMCR0_FC);
1358 	if (ppmu->flags & PPMU_ARCH_207S)
1359 		mtspr(SPRN_MMCR2, cpuhw->mmcr[3]);
1360 
1361 	/*
1362 	 * Read off any pre-existing events that need to move
1363 	 * to another PMC.
1364 	 */
1365 	for (i = 0; i < cpuhw->n_events; ++i) {
1366 		event = cpuhw->event[i];
1367 		if (event->hw.idx && event->hw.idx != hwc_index[i] + 1) {
1368 			power_pmu_read(event);
1369 			write_pmc(event->hw.idx, 0);
1370 			event->hw.idx = 0;
1371 		}
1372 	}
1373 
1374 	/*
1375 	 * Initialize the PMCs for all the new and moved events.
1376 	 */
1377 	cpuhw->n_limited = n_lim = 0;
1378 	for (i = 0; i < cpuhw->n_events; ++i) {
1379 		event = cpuhw->event[i];
1380 		if (event->hw.idx)
1381 			continue;
1382 		idx = hwc_index[i] + 1;
1383 		if (is_limited_pmc(idx)) {
1384 			cpuhw->limited_counter[n_lim] = event;
1385 			cpuhw->limited_hwidx[n_lim] = idx;
1386 			++n_lim;
1387 			continue;
1388 		}
1389 
1390 		if (ebb)
1391 			val = local64_read(&event->hw.prev_count);
1392 		else {
1393 			val = 0;
1394 			if (event->hw.sample_period) {
1395 				left = local64_read(&event->hw.period_left);
1396 				if (left < 0x80000000L)
1397 					val = 0x80000000L - left;
1398 			}
1399 			local64_set(&event->hw.prev_count, val);
1400 		}
1401 
1402 		event->hw.idx = idx;
1403 		if (event->hw.state & PERF_HES_STOPPED)
1404 			val = 0;
1405 		write_pmc(idx, val);
1406 
1407 		perf_event_update_userpage(event);
1408 	}
1409 	cpuhw->n_limited = n_lim;
1410 	cpuhw->mmcr[0] |= MMCR0_PMXE | MMCR0_FCECE;
1411 
1412  out_enable:
1413 	pmao_restore_workaround(ebb);
1414 
1415 	mmcr0 = ebb_switch_in(ebb, cpuhw);
1416 
1417 	mb();
1418 	if (cpuhw->bhrb_users)
1419 		ppmu->config_bhrb(cpuhw->bhrb_filter);
1420 
1421 	write_mmcr0(cpuhw, mmcr0);
1422 
1423 	/*
1424 	 * Enable instruction sampling if necessary
1425 	 */
1426 	if (cpuhw->mmcr[2] & MMCRA_SAMPLE_ENABLE) {
1427 		mb();
1428 		mtspr(SPRN_MMCRA, cpuhw->mmcr[2]);
1429 	}
1430 
1431  out:
1432 
1433 	local_irq_restore(flags);
1434 }
1435 
1436 static int collect_events(struct perf_event *group, int max_count,
1437 			  struct perf_event *ctrs[], u64 *events,
1438 			  unsigned int *flags)
1439 {
1440 	int n = 0;
1441 	struct perf_event *event;
1442 
1443 	if (group->pmu->task_ctx_nr == perf_hw_context) {
1444 		if (n >= max_count)
1445 			return -1;
1446 		ctrs[n] = group;
1447 		flags[n] = group->hw.event_base;
1448 		events[n++] = group->hw.config;
1449 	}
1450 	for_each_sibling_event(event, group) {
1451 		if (event->pmu->task_ctx_nr == perf_hw_context &&
1452 		    event->state != PERF_EVENT_STATE_OFF) {
1453 			if (n >= max_count)
1454 				return -1;
1455 			ctrs[n] = event;
1456 			flags[n] = event->hw.event_base;
1457 			events[n++] = event->hw.config;
1458 		}
1459 	}
1460 	return n;
1461 }
1462 
1463 /*
1464  * Add an event to the PMU.
1465  * If all events are not already frozen, then we disable and
1466  * re-enable the PMU in order to get hw_perf_enable to do the
1467  * actual work of reconfiguring the PMU.
1468  */
1469 static int power_pmu_add(struct perf_event *event, int ef_flags)
1470 {
1471 	struct cpu_hw_events *cpuhw;
1472 	unsigned long flags;
1473 	int n0;
1474 	int ret = -EAGAIN;
1475 
1476 	local_irq_save(flags);
1477 	perf_pmu_disable(event->pmu);
1478 
1479 	/*
1480 	 * Add the event to the list (if there is room)
1481 	 * and check whether the total set is still feasible.
1482 	 */
1483 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1484 	n0 = cpuhw->n_events;
1485 	if (n0 >= ppmu->n_counter)
1486 		goto out;
1487 	cpuhw->event[n0] = event;
1488 	cpuhw->events[n0] = event->hw.config;
1489 	cpuhw->flags[n0] = event->hw.event_base;
1490 
1491 	/*
1492 	 * This event may have been disabled/stopped in record_and_restart()
1493 	 * because we exceeded the ->event_limit. If re-starting the event,
1494 	 * clear the ->hw.state (STOPPED and UPTODATE flags), so the user
1495 	 * notification is re-enabled.
1496 	 */
1497 	if (!(ef_flags & PERF_EF_START))
1498 		event->hw.state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1499 	else
1500 		event->hw.state = 0;
1501 
1502 	/*
1503 	 * If group events scheduling transaction was started,
1504 	 * skip the schedulability test here, it will be performed
1505 	 * at commit time(->commit_txn) as a whole
1506 	 */
1507 	if (cpuhw->txn_flags & PERF_PMU_TXN_ADD)
1508 		goto nocheck;
1509 
1510 	if (check_excludes(cpuhw->event, cpuhw->flags, n0, 1))
1511 		goto out;
1512 	if (power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n0 + 1))
1513 		goto out;
1514 	event->hw.config = cpuhw->events[n0];
1515 
1516 nocheck:
1517 	ebb_event_add(event);
1518 
1519 	++cpuhw->n_events;
1520 	++cpuhw->n_added;
1521 
1522 	ret = 0;
1523  out:
1524 	if (has_branch_stack(event)) {
1525 		power_pmu_bhrb_enable(event);
1526 		cpuhw->bhrb_filter = ppmu->bhrb_filter_map(
1527 					event->attr.branch_sample_type);
1528 	}
1529 
1530 	perf_pmu_enable(event->pmu);
1531 	local_irq_restore(flags);
1532 	return ret;
1533 }
1534 
1535 /*
1536  * Remove an event from the PMU.
1537  */
1538 static void power_pmu_del(struct perf_event *event, int ef_flags)
1539 {
1540 	struct cpu_hw_events *cpuhw;
1541 	long i;
1542 	unsigned long flags;
1543 
1544 	local_irq_save(flags);
1545 	perf_pmu_disable(event->pmu);
1546 
1547 	power_pmu_read(event);
1548 
1549 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1550 	for (i = 0; i < cpuhw->n_events; ++i) {
1551 		if (event == cpuhw->event[i]) {
1552 			while (++i < cpuhw->n_events) {
1553 				cpuhw->event[i-1] = cpuhw->event[i];
1554 				cpuhw->events[i-1] = cpuhw->events[i];
1555 				cpuhw->flags[i-1] = cpuhw->flags[i];
1556 			}
1557 			--cpuhw->n_events;
1558 			ppmu->disable_pmc(event->hw.idx - 1, cpuhw->mmcr);
1559 			if (event->hw.idx) {
1560 				write_pmc(event->hw.idx, 0);
1561 				event->hw.idx = 0;
1562 			}
1563 			perf_event_update_userpage(event);
1564 			break;
1565 		}
1566 	}
1567 	for (i = 0; i < cpuhw->n_limited; ++i)
1568 		if (event == cpuhw->limited_counter[i])
1569 			break;
1570 	if (i < cpuhw->n_limited) {
1571 		while (++i < cpuhw->n_limited) {
1572 			cpuhw->limited_counter[i-1] = cpuhw->limited_counter[i];
1573 			cpuhw->limited_hwidx[i-1] = cpuhw->limited_hwidx[i];
1574 		}
1575 		--cpuhw->n_limited;
1576 	}
1577 	if (cpuhw->n_events == 0) {
1578 		/* disable exceptions if no events are running */
1579 		cpuhw->mmcr[0] &= ~(MMCR0_PMXE | MMCR0_FCECE);
1580 	}
1581 
1582 	if (has_branch_stack(event))
1583 		power_pmu_bhrb_disable(event);
1584 
1585 	perf_pmu_enable(event->pmu);
1586 	local_irq_restore(flags);
1587 }
1588 
1589 /*
1590  * POWER-PMU does not support disabling individual counters, hence
1591  * program their cycle counter to their max value and ignore the interrupts.
1592  */
1593 
1594 static void power_pmu_start(struct perf_event *event, int ef_flags)
1595 {
1596 	unsigned long flags;
1597 	s64 left;
1598 	unsigned long val;
1599 
1600 	if (!event->hw.idx || !event->hw.sample_period)
1601 		return;
1602 
1603 	if (!(event->hw.state & PERF_HES_STOPPED))
1604 		return;
1605 
1606 	if (ef_flags & PERF_EF_RELOAD)
1607 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1608 
1609 	local_irq_save(flags);
1610 	perf_pmu_disable(event->pmu);
1611 
1612 	event->hw.state = 0;
1613 	left = local64_read(&event->hw.period_left);
1614 
1615 	val = 0;
1616 	if (left < 0x80000000L)
1617 		val = 0x80000000L - left;
1618 
1619 	write_pmc(event->hw.idx, val);
1620 
1621 	perf_event_update_userpage(event);
1622 	perf_pmu_enable(event->pmu);
1623 	local_irq_restore(flags);
1624 }
1625 
1626 static void power_pmu_stop(struct perf_event *event, int ef_flags)
1627 {
1628 	unsigned long flags;
1629 
1630 	if (!event->hw.idx || !event->hw.sample_period)
1631 		return;
1632 
1633 	if (event->hw.state & PERF_HES_STOPPED)
1634 		return;
1635 
1636 	local_irq_save(flags);
1637 	perf_pmu_disable(event->pmu);
1638 
1639 	power_pmu_read(event);
1640 	event->hw.state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1641 	write_pmc(event->hw.idx, 0);
1642 
1643 	perf_event_update_userpage(event);
1644 	perf_pmu_enable(event->pmu);
1645 	local_irq_restore(flags);
1646 }
1647 
1648 /*
1649  * Start group events scheduling transaction
1650  * Set the flag to make pmu::enable() not perform the
1651  * schedulability test, it will be performed at commit time
1652  *
1653  * We only support PERF_PMU_TXN_ADD transactions. Save the
1654  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1655  * transactions.
1656  */
1657 static void power_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1658 {
1659 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1660 
1661 	WARN_ON_ONCE(cpuhw->txn_flags);		/* txn already in flight */
1662 
1663 	cpuhw->txn_flags = txn_flags;
1664 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1665 		return;
1666 
1667 	perf_pmu_disable(pmu);
1668 	cpuhw->n_txn_start = cpuhw->n_events;
1669 }
1670 
1671 /*
1672  * Stop group events scheduling transaction
1673  * Clear the flag and pmu::enable() will perform the
1674  * schedulability test.
1675  */
1676 static void power_pmu_cancel_txn(struct pmu *pmu)
1677 {
1678 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
1679 	unsigned int txn_flags;
1680 
1681 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1682 
1683 	txn_flags = cpuhw->txn_flags;
1684 	cpuhw->txn_flags = 0;
1685 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1686 		return;
1687 
1688 	perf_pmu_enable(pmu);
1689 }
1690 
1691 /*
1692  * Commit group events scheduling transaction
1693  * Perform the group schedulability test as a whole
1694  * Return 0 if success
1695  */
1696 static int power_pmu_commit_txn(struct pmu *pmu)
1697 {
1698 	struct cpu_hw_events *cpuhw;
1699 	long i, n;
1700 
1701 	if (!ppmu)
1702 		return -EAGAIN;
1703 
1704 	cpuhw = this_cpu_ptr(&cpu_hw_events);
1705 	WARN_ON_ONCE(!cpuhw->txn_flags);	/* no txn in flight */
1706 
1707 	if (cpuhw->txn_flags & ~PERF_PMU_TXN_ADD) {
1708 		cpuhw->txn_flags = 0;
1709 		return 0;
1710 	}
1711 
1712 	n = cpuhw->n_events;
1713 	if (check_excludes(cpuhw->event, cpuhw->flags, 0, n))
1714 		return -EAGAIN;
1715 	i = power_check_constraints(cpuhw, cpuhw->events, cpuhw->flags, n);
1716 	if (i < 0)
1717 		return -EAGAIN;
1718 
1719 	for (i = cpuhw->n_txn_start; i < n; ++i)
1720 		cpuhw->event[i]->hw.config = cpuhw->events[i];
1721 
1722 	cpuhw->txn_flags = 0;
1723 	perf_pmu_enable(pmu);
1724 	return 0;
1725 }
1726 
1727 /*
1728  * Return 1 if we might be able to put event on a limited PMC,
1729  * or 0 if not.
1730  * An event can only go on a limited PMC if it counts something
1731  * that a limited PMC can count, doesn't require interrupts, and
1732  * doesn't exclude any processor mode.
1733  */
1734 static int can_go_on_limited_pmc(struct perf_event *event, u64 ev,
1735 				 unsigned int flags)
1736 {
1737 	int n;
1738 	u64 alt[MAX_EVENT_ALTERNATIVES];
1739 
1740 	if (event->attr.exclude_user
1741 	    || event->attr.exclude_kernel
1742 	    || event->attr.exclude_hv
1743 	    || event->attr.sample_period)
1744 		return 0;
1745 
1746 	if (ppmu->limited_pmc_event(ev))
1747 		return 1;
1748 
1749 	/*
1750 	 * The requested event_id isn't on a limited PMC already;
1751 	 * see if any alternative code goes on a limited PMC.
1752 	 */
1753 	if (!ppmu->get_alternatives)
1754 		return 0;
1755 
1756 	flags |= PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD;
1757 	n = ppmu->get_alternatives(ev, flags, alt);
1758 
1759 	return n > 0;
1760 }
1761 
1762 /*
1763  * Find an alternative event_id that goes on a normal PMC, if possible,
1764  * and return the event_id code, or 0 if there is no such alternative.
1765  * (Note: event_id code 0 is "don't count" on all machines.)
1766  */
1767 static u64 normal_pmc_alternative(u64 ev, unsigned long flags)
1768 {
1769 	u64 alt[MAX_EVENT_ALTERNATIVES];
1770 	int n;
1771 
1772 	flags &= ~(PPMU_LIMITED_PMC_OK | PPMU_LIMITED_PMC_REQD);
1773 	n = ppmu->get_alternatives(ev, flags, alt);
1774 	if (!n)
1775 		return 0;
1776 	return alt[0];
1777 }
1778 
1779 /* Number of perf_events counting hardware events */
1780 static atomic_t num_events;
1781 /* Used to avoid races in calling reserve/release_pmc_hardware */
1782 static DEFINE_MUTEX(pmc_reserve_mutex);
1783 
1784 /*
1785  * Release the PMU if this is the last perf_event.
1786  */
1787 static void hw_perf_event_destroy(struct perf_event *event)
1788 {
1789 	if (!atomic_add_unless(&num_events, -1, 1)) {
1790 		mutex_lock(&pmc_reserve_mutex);
1791 		if (atomic_dec_return(&num_events) == 0)
1792 			release_pmc_hardware();
1793 		mutex_unlock(&pmc_reserve_mutex);
1794 	}
1795 }
1796 
1797 /*
1798  * Translate a generic cache event_id config to a raw event_id code.
1799  */
1800 static int hw_perf_cache_event(u64 config, u64 *eventp)
1801 {
1802 	unsigned long type, op, result;
1803 	int ev;
1804 
1805 	if (!ppmu->cache_events)
1806 		return -EINVAL;
1807 
1808 	/* unpack config */
1809 	type = config & 0xff;
1810 	op = (config >> 8) & 0xff;
1811 	result = (config >> 16) & 0xff;
1812 
1813 	if (type >= PERF_COUNT_HW_CACHE_MAX ||
1814 	    op >= PERF_COUNT_HW_CACHE_OP_MAX ||
1815 	    result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
1816 		return -EINVAL;
1817 
1818 	ev = (*ppmu->cache_events)[type][op][result];
1819 	if (ev == 0)
1820 		return -EOPNOTSUPP;
1821 	if (ev == -1)
1822 		return -EINVAL;
1823 	*eventp = ev;
1824 	return 0;
1825 }
1826 
1827 static bool is_event_blacklisted(u64 ev)
1828 {
1829 	int i;
1830 
1831 	for (i=0; i < ppmu->n_blacklist_ev; i++) {
1832 		if (ppmu->blacklist_ev[i] == ev)
1833 			return true;
1834 	}
1835 
1836 	return false;
1837 }
1838 
1839 static int power_pmu_event_init(struct perf_event *event)
1840 {
1841 	u64 ev;
1842 	unsigned long flags;
1843 	struct perf_event *ctrs[MAX_HWEVENTS];
1844 	u64 events[MAX_HWEVENTS];
1845 	unsigned int cflags[MAX_HWEVENTS];
1846 	int n;
1847 	int err;
1848 	struct cpu_hw_events *cpuhw;
1849 	u64 bhrb_filter;
1850 
1851 	if (!ppmu)
1852 		return -ENOENT;
1853 
1854 	if (has_branch_stack(event)) {
1855 	        /* PMU has BHRB enabled */
1856 		if (!(ppmu->flags & PPMU_ARCH_207S))
1857 			return -EOPNOTSUPP;
1858 	}
1859 
1860 	switch (event->attr.type) {
1861 	case PERF_TYPE_HARDWARE:
1862 		ev = event->attr.config;
1863 		if (ev >= ppmu->n_generic || ppmu->generic_events[ev] == 0)
1864 			return -EOPNOTSUPP;
1865 
1866 		if (ppmu->blacklist_ev && is_event_blacklisted(ev))
1867 			return -EINVAL;
1868 		ev = ppmu->generic_events[ev];
1869 		break;
1870 	case PERF_TYPE_HW_CACHE:
1871 		err = hw_perf_cache_event(event->attr.config, &ev);
1872 		if (err)
1873 			return err;
1874 
1875 		if (ppmu->blacklist_ev && is_event_blacklisted(ev))
1876 			return -EINVAL;
1877 		break;
1878 	case PERF_TYPE_RAW:
1879 		ev = event->attr.config;
1880 
1881 		if (ppmu->blacklist_ev && is_event_blacklisted(ev))
1882 			return -EINVAL;
1883 		break;
1884 	default:
1885 		return -ENOENT;
1886 	}
1887 
1888 	event->hw.config_base = ev;
1889 	event->hw.idx = 0;
1890 
1891 	/*
1892 	 * If we are not running on a hypervisor, force the
1893 	 * exclude_hv bit to 0 so that we don't care what
1894 	 * the user set it to.
1895 	 */
1896 	if (!firmware_has_feature(FW_FEATURE_LPAR))
1897 		event->attr.exclude_hv = 0;
1898 
1899 	/*
1900 	 * If this is a per-task event, then we can use
1901 	 * PM_RUN_* events interchangeably with their non RUN_*
1902 	 * equivalents, e.g. PM_RUN_CYC instead of PM_CYC.
1903 	 * XXX we should check if the task is an idle task.
1904 	 */
1905 	flags = 0;
1906 	if (event->attach_state & PERF_ATTACH_TASK)
1907 		flags |= PPMU_ONLY_COUNT_RUN;
1908 
1909 	/*
1910 	 * If this machine has limited events, check whether this
1911 	 * event_id could go on a limited event.
1912 	 */
1913 	if (ppmu->flags & PPMU_LIMITED_PMC5_6) {
1914 		if (can_go_on_limited_pmc(event, ev, flags)) {
1915 			flags |= PPMU_LIMITED_PMC_OK;
1916 		} else if (ppmu->limited_pmc_event(ev)) {
1917 			/*
1918 			 * The requested event_id is on a limited PMC,
1919 			 * but we can't use a limited PMC; see if any
1920 			 * alternative goes on a normal PMC.
1921 			 */
1922 			ev = normal_pmc_alternative(ev, flags);
1923 			if (!ev)
1924 				return -EINVAL;
1925 		}
1926 	}
1927 
1928 	/* Extra checks for EBB */
1929 	err = ebb_event_check(event);
1930 	if (err)
1931 		return err;
1932 
1933 	/*
1934 	 * If this is in a group, check if it can go on with all the
1935 	 * other hardware events in the group.  We assume the event
1936 	 * hasn't been linked into its leader's sibling list at this point.
1937 	 */
1938 	n = 0;
1939 	if (event->group_leader != event) {
1940 		n = collect_events(event->group_leader, ppmu->n_counter - 1,
1941 				   ctrs, events, cflags);
1942 		if (n < 0)
1943 			return -EINVAL;
1944 	}
1945 	events[n] = ev;
1946 	ctrs[n] = event;
1947 	cflags[n] = flags;
1948 	if (check_excludes(ctrs, cflags, n, 1))
1949 		return -EINVAL;
1950 
1951 	cpuhw = &get_cpu_var(cpu_hw_events);
1952 	err = power_check_constraints(cpuhw, events, cflags, n + 1);
1953 
1954 	if (has_branch_stack(event)) {
1955 		bhrb_filter = ppmu->bhrb_filter_map(
1956 					event->attr.branch_sample_type);
1957 
1958 		if (bhrb_filter == -1) {
1959 			put_cpu_var(cpu_hw_events);
1960 			return -EOPNOTSUPP;
1961 		}
1962 		cpuhw->bhrb_filter = bhrb_filter;
1963 	}
1964 
1965 	put_cpu_var(cpu_hw_events);
1966 	if (err)
1967 		return -EINVAL;
1968 
1969 	event->hw.config = events[n];
1970 	event->hw.event_base = cflags[n];
1971 	event->hw.last_period = event->hw.sample_period;
1972 	local64_set(&event->hw.period_left, event->hw.last_period);
1973 
1974 	/*
1975 	 * For EBB events we just context switch the PMC value, we don't do any
1976 	 * of the sample_period logic. We use hw.prev_count for this.
1977 	 */
1978 	if (is_ebb_event(event))
1979 		local64_set(&event->hw.prev_count, 0);
1980 
1981 	/*
1982 	 * See if we need to reserve the PMU.
1983 	 * If no events are currently in use, then we have to take a
1984 	 * mutex to ensure that we don't race with another task doing
1985 	 * reserve_pmc_hardware or release_pmc_hardware.
1986 	 */
1987 	err = 0;
1988 	if (!atomic_inc_not_zero(&num_events)) {
1989 		mutex_lock(&pmc_reserve_mutex);
1990 		if (atomic_read(&num_events) == 0 &&
1991 		    reserve_pmc_hardware(perf_event_interrupt))
1992 			err = -EBUSY;
1993 		else
1994 			atomic_inc(&num_events);
1995 		mutex_unlock(&pmc_reserve_mutex);
1996 	}
1997 	event->destroy = hw_perf_event_destroy;
1998 
1999 	return err;
2000 }
2001 
2002 static int power_pmu_event_idx(struct perf_event *event)
2003 {
2004 	return event->hw.idx;
2005 }
2006 
2007 ssize_t power_events_sysfs_show(struct device *dev,
2008 				struct device_attribute *attr, char *page)
2009 {
2010 	struct perf_pmu_events_attr *pmu_attr;
2011 
2012 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
2013 
2014 	return sprintf(page, "event=0x%02llx\n", pmu_attr->id);
2015 }
2016 
2017 static struct pmu power_pmu = {
2018 	.pmu_enable	= power_pmu_enable,
2019 	.pmu_disable	= power_pmu_disable,
2020 	.event_init	= power_pmu_event_init,
2021 	.add		= power_pmu_add,
2022 	.del		= power_pmu_del,
2023 	.start		= power_pmu_start,
2024 	.stop		= power_pmu_stop,
2025 	.read		= power_pmu_read,
2026 	.start_txn	= power_pmu_start_txn,
2027 	.cancel_txn	= power_pmu_cancel_txn,
2028 	.commit_txn	= power_pmu_commit_txn,
2029 	.event_idx	= power_pmu_event_idx,
2030 	.sched_task	= power_pmu_sched_task,
2031 };
2032 
2033 /*
2034  * A counter has overflowed; update its count and record
2035  * things if requested.  Note that interrupts are hard-disabled
2036  * here so there is no possibility of being interrupted.
2037  */
2038 static void record_and_restart(struct perf_event *event, unsigned long val,
2039 			       struct pt_regs *regs)
2040 {
2041 	u64 period = event->hw.sample_period;
2042 	s64 prev, delta, left;
2043 	int record = 0;
2044 
2045 	if (event->hw.state & PERF_HES_STOPPED) {
2046 		write_pmc(event->hw.idx, 0);
2047 		return;
2048 	}
2049 
2050 	/* we don't have to worry about interrupts here */
2051 	prev = local64_read(&event->hw.prev_count);
2052 	delta = check_and_compute_delta(prev, val);
2053 	local64_add(delta, &event->count);
2054 
2055 	/*
2056 	 * See if the total period for this event has expired,
2057 	 * and update for the next period.
2058 	 */
2059 	val = 0;
2060 	left = local64_read(&event->hw.period_left) - delta;
2061 	if (delta == 0)
2062 		left++;
2063 	if (period) {
2064 		if (left <= 0) {
2065 			left += period;
2066 			if (left <= 0)
2067 				left = period;
2068 			record = siar_valid(regs);
2069 			event->hw.last_period = event->hw.sample_period;
2070 		}
2071 		if (left < 0x80000000LL)
2072 			val = 0x80000000LL - left;
2073 	}
2074 
2075 	write_pmc(event->hw.idx, val);
2076 	local64_set(&event->hw.prev_count, val);
2077 	local64_set(&event->hw.period_left, left);
2078 	perf_event_update_userpage(event);
2079 
2080 	/*
2081 	 * Finally record data if requested.
2082 	 */
2083 	if (record) {
2084 		struct perf_sample_data data;
2085 
2086 		perf_sample_data_init(&data, ~0ULL, event->hw.last_period);
2087 
2088 		if (event->attr.sample_type &
2089 		    (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR))
2090 			perf_get_data_addr(regs, &data.addr);
2091 
2092 		if (event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK) {
2093 			struct cpu_hw_events *cpuhw;
2094 			cpuhw = this_cpu_ptr(&cpu_hw_events);
2095 			power_pmu_bhrb_read(cpuhw);
2096 			data.br_stack = &cpuhw->bhrb_stack;
2097 		}
2098 
2099 		if (event->attr.sample_type & PERF_SAMPLE_DATA_SRC &&
2100 						ppmu->get_mem_data_src)
2101 			ppmu->get_mem_data_src(&data.data_src, ppmu->flags, regs);
2102 
2103 		if (event->attr.sample_type & PERF_SAMPLE_WEIGHT &&
2104 						ppmu->get_mem_weight)
2105 			ppmu->get_mem_weight(&data.weight);
2106 
2107 		if (perf_event_overflow(event, &data, regs))
2108 			power_pmu_stop(event, 0);
2109 	}
2110 }
2111 
2112 /*
2113  * Called from generic code to get the misc flags (i.e. processor mode)
2114  * for an event_id.
2115  */
2116 unsigned long perf_misc_flags(struct pt_regs *regs)
2117 {
2118 	u32 flags = perf_get_misc_flags(regs);
2119 
2120 	if (flags)
2121 		return flags;
2122 	return user_mode(regs) ? PERF_RECORD_MISC_USER :
2123 		PERF_RECORD_MISC_KERNEL;
2124 }
2125 
2126 /*
2127  * Called from generic code to get the instruction pointer
2128  * for an event_id.
2129  */
2130 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2131 {
2132 	bool use_siar = regs_use_siar(regs);
2133 
2134 	if (use_siar && siar_valid(regs))
2135 		return mfspr(SPRN_SIAR) + perf_ip_adjust(regs);
2136 	else if (use_siar)
2137 		return 0;		// no valid instruction pointer
2138 	else
2139 		return regs->nip;
2140 }
2141 
2142 static bool pmc_overflow_power7(unsigned long val)
2143 {
2144 	/*
2145 	 * Events on POWER7 can roll back if a speculative event doesn't
2146 	 * eventually complete. Unfortunately in some rare cases they will
2147 	 * raise a performance monitor exception. We need to catch this to
2148 	 * ensure we reset the PMC. In all cases the PMC will be 256 or less
2149 	 * cycles from overflow.
2150 	 *
2151 	 * We only do this if the first pass fails to find any overflowing
2152 	 * PMCs because a user might set a period of less than 256 and we
2153 	 * don't want to mistakenly reset them.
2154 	 */
2155 	if ((0x80000000 - val) <= 256)
2156 		return true;
2157 
2158 	return false;
2159 }
2160 
2161 static bool pmc_overflow(unsigned long val)
2162 {
2163 	if ((int)val < 0)
2164 		return true;
2165 
2166 	return false;
2167 }
2168 
2169 /*
2170  * Performance monitor interrupt stuff
2171  */
2172 static void __perf_event_interrupt(struct pt_regs *regs)
2173 {
2174 	int i, j;
2175 	struct cpu_hw_events *cpuhw = this_cpu_ptr(&cpu_hw_events);
2176 	struct perf_event *event;
2177 	unsigned long val[8];
2178 	int found, active;
2179 	int nmi;
2180 
2181 	if (cpuhw->n_limited)
2182 		freeze_limited_counters(cpuhw, mfspr(SPRN_PMC5),
2183 					mfspr(SPRN_PMC6));
2184 
2185 	perf_read_regs(regs);
2186 
2187 	nmi = perf_intr_is_nmi(regs);
2188 	if (nmi)
2189 		nmi_enter();
2190 	else
2191 		irq_enter();
2192 
2193 	/* Read all the PMCs since we'll need them a bunch of times */
2194 	for (i = 0; i < ppmu->n_counter; ++i)
2195 		val[i] = read_pmc(i + 1);
2196 
2197 	/* Try to find what caused the IRQ */
2198 	found = 0;
2199 	for (i = 0; i < ppmu->n_counter; ++i) {
2200 		if (!pmc_overflow(val[i]))
2201 			continue;
2202 		if (is_limited_pmc(i + 1))
2203 			continue; /* these won't generate IRQs */
2204 		/*
2205 		 * We've found one that's overflowed.  For active
2206 		 * counters we need to log this.  For inactive
2207 		 * counters, we need to reset it anyway
2208 		 */
2209 		found = 1;
2210 		active = 0;
2211 		for (j = 0; j < cpuhw->n_events; ++j) {
2212 			event = cpuhw->event[j];
2213 			if (event->hw.idx == (i + 1)) {
2214 				active = 1;
2215 				record_and_restart(event, val[i], regs);
2216 				break;
2217 			}
2218 		}
2219 		if (!active)
2220 			/* reset non active counters that have overflowed */
2221 			write_pmc(i + 1, 0);
2222 	}
2223 	if (!found && pvr_version_is(PVR_POWER7)) {
2224 		/* check active counters for special buggy p7 overflow */
2225 		for (i = 0; i < cpuhw->n_events; ++i) {
2226 			event = cpuhw->event[i];
2227 			if (!event->hw.idx || is_limited_pmc(event->hw.idx))
2228 				continue;
2229 			if (pmc_overflow_power7(val[event->hw.idx - 1])) {
2230 				/* event has overflowed in a buggy way*/
2231 				found = 1;
2232 				record_and_restart(event,
2233 						   val[event->hw.idx - 1],
2234 						   regs);
2235 			}
2236 		}
2237 	}
2238 	if (!found && !nmi && printk_ratelimit())
2239 		printk(KERN_WARNING "Can't find PMC that caused IRQ\n");
2240 
2241 	/*
2242 	 * Reset MMCR0 to its normal value.  This will set PMXE and
2243 	 * clear FC (freeze counters) and PMAO (perf mon alert occurred)
2244 	 * and thus allow interrupts to occur again.
2245 	 * XXX might want to use MSR.PM to keep the events frozen until
2246 	 * we get back out of this interrupt.
2247 	 */
2248 	write_mmcr0(cpuhw, cpuhw->mmcr[0]);
2249 
2250 	if (nmi)
2251 		nmi_exit();
2252 	else
2253 		irq_exit();
2254 }
2255 
2256 static void perf_event_interrupt(struct pt_regs *regs)
2257 {
2258 	u64 start_clock = sched_clock();
2259 
2260 	__perf_event_interrupt(regs);
2261 	perf_sample_event_took(sched_clock() - start_clock);
2262 }
2263 
2264 static int power_pmu_prepare_cpu(unsigned int cpu)
2265 {
2266 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
2267 
2268 	if (ppmu) {
2269 		memset(cpuhw, 0, sizeof(*cpuhw));
2270 		cpuhw->mmcr[0] = MMCR0_FC;
2271 	}
2272 	return 0;
2273 }
2274 
2275 int register_power_pmu(struct power_pmu *pmu)
2276 {
2277 	if (ppmu)
2278 		return -EBUSY;		/* something's already registered */
2279 
2280 	ppmu = pmu;
2281 	pr_info("%s performance monitor hardware support registered\n",
2282 		pmu->name);
2283 
2284 	power_pmu.attr_groups = ppmu->attr_groups;
2285 
2286 #ifdef MSR_HV
2287 	/*
2288 	 * Use FCHV to ignore kernel events if MSR.HV is set.
2289 	 */
2290 	if (mfmsr() & MSR_HV)
2291 		freeze_events_kernel = MMCR0_FCHV;
2292 #endif /* CONFIG_PPC64 */
2293 
2294 	perf_pmu_register(&power_pmu, "cpu", PERF_TYPE_RAW);
2295 	cpuhp_setup_state(CPUHP_PERF_POWER, "perf/powerpc:prepare",
2296 			  power_pmu_prepare_cpu, NULL);
2297 	return 0;
2298 }
2299 
2300 #ifdef CONFIG_PPC64
2301 static int __init init_ppc64_pmu(void)
2302 {
2303 	/* run through all the pmu drivers one at a time */
2304 	if (!init_power5_pmu())
2305 		return 0;
2306 	else if (!init_power5p_pmu())
2307 		return 0;
2308 	else if (!init_power6_pmu())
2309 		return 0;
2310 	else if (!init_power7_pmu())
2311 		return 0;
2312 	else if (!init_power8_pmu())
2313 		return 0;
2314 	else if (!init_power9_pmu())
2315 		return 0;
2316 	else if (!init_ppc970_pmu())
2317 		return 0;
2318 	else
2319 		return init_generic_compat_pmu();
2320 }
2321 early_initcall(init_ppc64_pmu);
2322 #endif
2323