xref: /linux/arch/x86/events/core.c (revision 071bf69a0220253a44acb8b2a27f7a262b9a46bf)
1 /*
2  * Performance events x86 architecture code
3  *
4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6  *  Copyright (C) 2009 Jaswinder Singh Rajput
7  *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
8  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra
9  *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
10  *  Copyright (C) 2009 Google, Inc., Stephane Eranian
11  *
12  *  For licencing details see kernel-base/COPYING
13  */
14 
15 #include <linux/perf_event.h>
16 #include <linux/capability.h>
17 #include <linux/notifier.h>
18 #include <linux/hardirq.h>
19 #include <linux/kprobes.h>
20 #include <linux/export.h>
21 #include <linux/init.h>
22 #include <linux/kdebug.h>
23 #include <linux/sched.h>
24 #include <linux/uaccess.h>
25 #include <linux/slab.h>
26 #include <linux/cpu.h>
27 #include <linux/bitops.h>
28 #include <linux/device.h>
29 
30 #include <asm/apic.h>
31 #include <asm/stacktrace.h>
32 #include <asm/nmi.h>
33 #include <asm/smp.h>
34 #include <asm/alternative.h>
35 #include <asm/mmu_context.h>
36 #include <asm/tlbflush.h>
37 #include <asm/timer.h>
38 #include <asm/desc.h>
39 #include <asm/ldt.h>
40 
41 #include "perf_event.h"
42 
43 struct x86_pmu x86_pmu __read_mostly;
44 
45 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
46 	.enabled = 1,
47 };
48 
49 struct static_key rdpmc_always_available = STATIC_KEY_INIT_FALSE;
50 
51 u64 __read_mostly hw_cache_event_ids
52 				[PERF_COUNT_HW_CACHE_MAX]
53 				[PERF_COUNT_HW_CACHE_OP_MAX]
54 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
55 u64 __read_mostly hw_cache_extra_regs
56 				[PERF_COUNT_HW_CACHE_MAX]
57 				[PERF_COUNT_HW_CACHE_OP_MAX]
58 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
59 
60 /*
61  * Propagate event elapsed time into the generic event.
62  * Can only be executed on the CPU where the event is active.
63  * Returns the delta events processed.
64  */
65 u64 x86_perf_event_update(struct perf_event *event)
66 {
67 	struct hw_perf_event *hwc = &event->hw;
68 	int shift = 64 - x86_pmu.cntval_bits;
69 	u64 prev_raw_count, new_raw_count;
70 	int idx = hwc->idx;
71 	s64 delta;
72 
73 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
74 		return 0;
75 
76 	/*
77 	 * Careful: an NMI might modify the previous event value.
78 	 *
79 	 * Our tactic to handle this is to first atomically read and
80 	 * exchange a new raw count - then add that new-prev delta
81 	 * count to the generic event atomically:
82 	 */
83 again:
84 	prev_raw_count = local64_read(&hwc->prev_count);
85 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
86 
87 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
88 					new_raw_count) != prev_raw_count)
89 		goto again;
90 
91 	/*
92 	 * Now we have the new raw value and have updated the prev
93 	 * timestamp already. We can now calculate the elapsed delta
94 	 * (event-)time and add that to the generic event.
95 	 *
96 	 * Careful, not all hw sign-extends above the physical width
97 	 * of the count.
98 	 */
99 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
100 	delta >>= shift;
101 
102 	local64_add(delta, &event->count);
103 	local64_sub(delta, &hwc->period_left);
104 
105 	return new_raw_count;
106 }
107 
108 /*
109  * Find and validate any extra registers to set up.
110  */
111 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
112 {
113 	struct hw_perf_event_extra *reg;
114 	struct extra_reg *er;
115 
116 	reg = &event->hw.extra_reg;
117 
118 	if (!x86_pmu.extra_regs)
119 		return 0;
120 
121 	for (er = x86_pmu.extra_regs; er->msr; er++) {
122 		if (er->event != (config & er->config_mask))
123 			continue;
124 		if (event->attr.config1 & ~er->valid_mask)
125 			return -EINVAL;
126 		/* Check if the extra msrs can be safely accessed*/
127 		if (!er->extra_msr_access)
128 			return -ENXIO;
129 
130 		reg->idx = er->idx;
131 		reg->config = event->attr.config1;
132 		reg->reg = er->msr;
133 		break;
134 	}
135 	return 0;
136 }
137 
138 static atomic_t active_events;
139 static atomic_t pmc_refcount;
140 static DEFINE_MUTEX(pmc_reserve_mutex);
141 
142 #ifdef CONFIG_X86_LOCAL_APIC
143 
144 static bool reserve_pmc_hardware(void)
145 {
146 	int i;
147 
148 	for (i = 0; i < x86_pmu.num_counters; i++) {
149 		if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
150 			goto perfctr_fail;
151 	}
152 
153 	for (i = 0; i < x86_pmu.num_counters; i++) {
154 		if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
155 			goto eventsel_fail;
156 	}
157 
158 	return true;
159 
160 eventsel_fail:
161 	for (i--; i >= 0; i--)
162 		release_evntsel_nmi(x86_pmu_config_addr(i));
163 
164 	i = x86_pmu.num_counters;
165 
166 perfctr_fail:
167 	for (i--; i >= 0; i--)
168 		release_perfctr_nmi(x86_pmu_event_addr(i));
169 
170 	return false;
171 }
172 
173 static void release_pmc_hardware(void)
174 {
175 	int i;
176 
177 	for (i = 0; i < x86_pmu.num_counters; i++) {
178 		release_perfctr_nmi(x86_pmu_event_addr(i));
179 		release_evntsel_nmi(x86_pmu_config_addr(i));
180 	}
181 }
182 
183 #else
184 
185 static bool reserve_pmc_hardware(void) { return true; }
186 static void release_pmc_hardware(void) {}
187 
188 #endif
189 
190 static bool check_hw_exists(void)
191 {
192 	u64 val, val_fail, val_new= ~0;
193 	int i, reg, reg_fail, ret = 0;
194 	int bios_fail = 0;
195 	int reg_safe = -1;
196 
197 	/*
198 	 * Check to see if the BIOS enabled any of the counters, if so
199 	 * complain and bail.
200 	 */
201 	for (i = 0; i < x86_pmu.num_counters; i++) {
202 		reg = x86_pmu_config_addr(i);
203 		ret = rdmsrl_safe(reg, &val);
204 		if (ret)
205 			goto msr_fail;
206 		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
207 			bios_fail = 1;
208 			val_fail = val;
209 			reg_fail = reg;
210 		} else {
211 			reg_safe = i;
212 		}
213 	}
214 
215 	if (x86_pmu.num_counters_fixed) {
216 		reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
217 		ret = rdmsrl_safe(reg, &val);
218 		if (ret)
219 			goto msr_fail;
220 		for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
221 			if (val & (0x03 << i*4)) {
222 				bios_fail = 1;
223 				val_fail = val;
224 				reg_fail = reg;
225 			}
226 		}
227 	}
228 
229 	/*
230 	 * If all the counters are enabled, the below test will always
231 	 * fail.  The tools will also become useless in this scenario.
232 	 * Just fail and disable the hardware counters.
233 	 */
234 
235 	if (reg_safe == -1) {
236 		reg = reg_safe;
237 		goto msr_fail;
238 	}
239 
240 	/*
241 	 * Read the current value, change it and read it back to see if it
242 	 * matches, this is needed to detect certain hardware emulators
243 	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
244 	 */
245 	reg = x86_pmu_event_addr(reg_safe);
246 	if (rdmsrl_safe(reg, &val))
247 		goto msr_fail;
248 	val ^= 0xffffUL;
249 	ret = wrmsrl_safe(reg, val);
250 	ret |= rdmsrl_safe(reg, &val_new);
251 	if (ret || val != val_new)
252 		goto msr_fail;
253 
254 	/*
255 	 * We still allow the PMU driver to operate:
256 	 */
257 	if (bios_fail) {
258 		pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
259 		pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
260 			      reg_fail, val_fail);
261 	}
262 
263 	return true;
264 
265 msr_fail:
266 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
267 		pr_cont("PMU not available due to virtualization, using software events only.\n");
268 	} else {
269 		pr_cont("Broken PMU hardware detected, using software events only.\n");
270 		pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
271 		       reg, val_new);
272 	}
273 
274 	return false;
275 }
276 
277 static void hw_perf_event_destroy(struct perf_event *event)
278 {
279 	x86_release_hardware();
280 	atomic_dec(&active_events);
281 }
282 
283 void hw_perf_lbr_event_destroy(struct perf_event *event)
284 {
285 	hw_perf_event_destroy(event);
286 
287 	/* undo the lbr/bts event accounting */
288 	x86_del_exclusive(x86_lbr_exclusive_lbr);
289 }
290 
291 static inline int x86_pmu_initialized(void)
292 {
293 	return x86_pmu.handle_irq != NULL;
294 }
295 
296 static inline int
297 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
298 {
299 	struct perf_event_attr *attr = &event->attr;
300 	unsigned int cache_type, cache_op, cache_result;
301 	u64 config, val;
302 
303 	config = attr->config;
304 
305 	cache_type = (config >>  0) & 0xff;
306 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
307 		return -EINVAL;
308 
309 	cache_op = (config >>  8) & 0xff;
310 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
311 		return -EINVAL;
312 
313 	cache_result = (config >> 16) & 0xff;
314 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
315 		return -EINVAL;
316 
317 	val = hw_cache_event_ids[cache_type][cache_op][cache_result];
318 
319 	if (val == 0)
320 		return -ENOENT;
321 
322 	if (val == -1)
323 		return -EINVAL;
324 
325 	hwc->config |= val;
326 	attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
327 	return x86_pmu_extra_regs(val, event);
328 }
329 
330 int x86_reserve_hardware(void)
331 {
332 	int err = 0;
333 
334 	if (!atomic_inc_not_zero(&pmc_refcount)) {
335 		mutex_lock(&pmc_reserve_mutex);
336 		if (atomic_read(&pmc_refcount) == 0) {
337 			if (!reserve_pmc_hardware())
338 				err = -EBUSY;
339 			else
340 				reserve_ds_buffers();
341 		}
342 		if (!err)
343 			atomic_inc(&pmc_refcount);
344 		mutex_unlock(&pmc_reserve_mutex);
345 	}
346 
347 	return err;
348 }
349 
350 void x86_release_hardware(void)
351 {
352 	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
353 		release_pmc_hardware();
354 		release_ds_buffers();
355 		mutex_unlock(&pmc_reserve_mutex);
356 	}
357 }
358 
359 /*
360  * Check if we can create event of a certain type (that no conflicting events
361  * are present).
362  */
363 int x86_add_exclusive(unsigned int what)
364 {
365 	int i;
366 
367 	if (x86_pmu.lbr_pt_coexist)
368 		return 0;
369 
370 	if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
371 		mutex_lock(&pmc_reserve_mutex);
372 		for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
373 			if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
374 				goto fail_unlock;
375 		}
376 		atomic_inc(&x86_pmu.lbr_exclusive[what]);
377 		mutex_unlock(&pmc_reserve_mutex);
378 	}
379 
380 	atomic_inc(&active_events);
381 	return 0;
382 
383 fail_unlock:
384 	mutex_unlock(&pmc_reserve_mutex);
385 	return -EBUSY;
386 }
387 
388 void x86_del_exclusive(unsigned int what)
389 {
390 	if (x86_pmu.lbr_pt_coexist)
391 		return;
392 
393 	atomic_dec(&x86_pmu.lbr_exclusive[what]);
394 	atomic_dec(&active_events);
395 }
396 
397 int x86_setup_perfctr(struct perf_event *event)
398 {
399 	struct perf_event_attr *attr = &event->attr;
400 	struct hw_perf_event *hwc = &event->hw;
401 	u64 config;
402 
403 	if (!is_sampling_event(event)) {
404 		hwc->sample_period = x86_pmu.max_period;
405 		hwc->last_period = hwc->sample_period;
406 		local64_set(&hwc->period_left, hwc->sample_period);
407 	}
408 
409 	if (attr->type == PERF_TYPE_RAW)
410 		return x86_pmu_extra_regs(event->attr.config, event);
411 
412 	if (attr->type == PERF_TYPE_HW_CACHE)
413 		return set_ext_hw_attr(hwc, event);
414 
415 	if (attr->config >= x86_pmu.max_events)
416 		return -EINVAL;
417 
418 	/*
419 	 * The generic map:
420 	 */
421 	config = x86_pmu.event_map(attr->config);
422 
423 	if (config == 0)
424 		return -ENOENT;
425 
426 	if (config == -1LL)
427 		return -EINVAL;
428 
429 	/*
430 	 * Branch tracing:
431 	 */
432 	if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
433 	    !attr->freq && hwc->sample_period == 1) {
434 		/* BTS is not supported by this architecture. */
435 		if (!x86_pmu.bts_active)
436 			return -EOPNOTSUPP;
437 
438 		/* BTS is currently only allowed for user-mode. */
439 		if (!attr->exclude_kernel)
440 			return -EOPNOTSUPP;
441 
442 		/* disallow bts if conflicting events are present */
443 		if (x86_add_exclusive(x86_lbr_exclusive_lbr))
444 			return -EBUSY;
445 
446 		event->destroy = hw_perf_lbr_event_destroy;
447 	}
448 
449 	hwc->config |= config;
450 
451 	return 0;
452 }
453 
454 /*
455  * check that branch_sample_type is compatible with
456  * settings needed for precise_ip > 1 which implies
457  * using the LBR to capture ALL taken branches at the
458  * priv levels of the measurement
459  */
460 static inline int precise_br_compat(struct perf_event *event)
461 {
462 	u64 m = event->attr.branch_sample_type;
463 	u64 b = 0;
464 
465 	/* must capture all branches */
466 	if (!(m & PERF_SAMPLE_BRANCH_ANY))
467 		return 0;
468 
469 	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
470 
471 	if (!event->attr.exclude_user)
472 		b |= PERF_SAMPLE_BRANCH_USER;
473 
474 	if (!event->attr.exclude_kernel)
475 		b |= PERF_SAMPLE_BRANCH_KERNEL;
476 
477 	/*
478 	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
479 	 */
480 
481 	return m == b;
482 }
483 
484 int x86_pmu_hw_config(struct perf_event *event)
485 {
486 	if (event->attr.precise_ip) {
487 		int precise = 0;
488 
489 		/* Support for constant skid */
490 		if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
491 			precise++;
492 
493 			/* Support for IP fixup */
494 			if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
495 				precise++;
496 
497 			if (x86_pmu.pebs_prec_dist)
498 				precise++;
499 		}
500 
501 		if (event->attr.precise_ip > precise)
502 			return -EOPNOTSUPP;
503 	}
504 	/*
505 	 * check that PEBS LBR correction does not conflict with
506 	 * whatever the user is asking with attr->branch_sample_type
507 	 */
508 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
509 		u64 *br_type = &event->attr.branch_sample_type;
510 
511 		if (has_branch_stack(event)) {
512 			if (!precise_br_compat(event))
513 				return -EOPNOTSUPP;
514 
515 			/* branch_sample_type is compatible */
516 
517 		} else {
518 			/*
519 			 * user did not specify  branch_sample_type
520 			 *
521 			 * For PEBS fixups, we capture all
522 			 * the branches at the priv level of the
523 			 * event.
524 			 */
525 			*br_type = PERF_SAMPLE_BRANCH_ANY;
526 
527 			if (!event->attr.exclude_user)
528 				*br_type |= PERF_SAMPLE_BRANCH_USER;
529 
530 			if (!event->attr.exclude_kernel)
531 				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
532 		}
533 	}
534 
535 	if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
536 		event->attach_state |= PERF_ATTACH_TASK_DATA;
537 
538 	/*
539 	 * Generate PMC IRQs:
540 	 * (keep 'enabled' bit clear for now)
541 	 */
542 	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
543 
544 	/*
545 	 * Count user and OS events unless requested not to
546 	 */
547 	if (!event->attr.exclude_user)
548 		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
549 	if (!event->attr.exclude_kernel)
550 		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
551 
552 	if (event->attr.type == PERF_TYPE_RAW)
553 		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
554 
555 	if (event->attr.sample_period && x86_pmu.limit_period) {
556 		if (x86_pmu.limit_period(event, event->attr.sample_period) >
557 				event->attr.sample_period)
558 			return -EINVAL;
559 	}
560 
561 	return x86_setup_perfctr(event);
562 }
563 
564 /*
565  * Setup the hardware configuration for a given attr_type
566  */
567 static int __x86_pmu_event_init(struct perf_event *event)
568 {
569 	int err;
570 
571 	if (!x86_pmu_initialized())
572 		return -ENODEV;
573 
574 	err = x86_reserve_hardware();
575 	if (err)
576 		return err;
577 
578 	atomic_inc(&active_events);
579 	event->destroy = hw_perf_event_destroy;
580 
581 	event->hw.idx = -1;
582 	event->hw.last_cpu = -1;
583 	event->hw.last_tag = ~0ULL;
584 
585 	/* mark unused */
586 	event->hw.extra_reg.idx = EXTRA_REG_NONE;
587 	event->hw.branch_reg.idx = EXTRA_REG_NONE;
588 
589 	return x86_pmu.hw_config(event);
590 }
591 
592 void x86_pmu_disable_all(void)
593 {
594 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
595 	int idx;
596 
597 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
598 		u64 val;
599 
600 		if (!test_bit(idx, cpuc->active_mask))
601 			continue;
602 		rdmsrl(x86_pmu_config_addr(idx), val);
603 		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
604 			continue;
605 		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
606 		wrmsrl(x86_pmu_config_addr(idx), val);
607 	}
608 }
609 
610 /*
611  * There may be PMI landing after enabled=0. The PMI hitting could be before or
612  * after disable_all.
613  *
614  * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
615  * It will not be re-enabled in the NMI handler again, because enabled=0. After
616  * handling the NMI, disable_all will be called, which will not change the
617  * state either. If PMI hits after disable_all, the PMU is already disabled
618  * before entering NMI handler. The NMI handler will not change the state
619  * either.
620  *
621  * So either situation is harmless.
622  */
623 static void x86_pmu_disable(struct pmu *pmu)
624 {
625 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
626 
627 	if (!x86_pmu_initialized())
628 		return;
629 
630 	if (!cpuc->enabled)
631 		return;
632 
633 	cpuc->n_added = 0;
634 	cpuc->enabled = 0;
635 	barrier();
636 
637 	x86_pmu.disable_all();
638 }
639 
640 void x86_pmu_enable_all(int added)
641 {
642 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
643 	int idx;
644 
645 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
646 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
647 
648 		if (!test_bit(idx, cpuc->active_mask))
649 			continue;
650 
651 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
652 	}
653 }
654 
655 static struct pmu pmu;
656 
657 static inline int is_x86_event(struct perf_event *event)
658 {
659 	return event->pmu == &pmu;
660 }
661 
662 /*
663  * Event scheduler state:
664  *
665  * Assign events iterating over all events and counters, beginning
666  * with events with least weights first. Keep the current iterator
667  * state in struct sched_state.
668  */
669 struct sched_state {
670 	int	weight;
671 	int	event;		/* event index */
672 	int	counter;	/* counter index */
673 	int	unassigned;	/* number of events to be assigned left */
674 	int	nr_gp;		/* number of GP counters used */
675 	unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
676 };
677 
678 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
679 #define	SCHED_STATES_MAX	2
680 
681 struct perf_sched {
682 	int			max_weight;
683 	int			max_events;
684 	int			max_gp;
685 	int			saved_states;
686 	struct event_constraint	**constraints;
687 	struct sched_state	state;
688 	struct sched_state	saved[SCHED_STATES_MAX];
689 };
690 
691 /*
692  * Initialize interator that runs through all events and counters.
693  */
694 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
695 			    int num, int wmin, int wmax, int gpmax)
696 {
697 	int idx;
698 
699 	memset(sched, 0, sizeof(*sched));
700 	sched->max_events	= num;
701 	sched->max_weight	= wmax;
702 	sched->max_gp		= gpmax;
703 	sched->constraints	= constraints;
704 
705 	for (idx = 0; idx < num; idx++) {
706 		if (constraints[idx]->weight == wmin)
707 			break;
708 	}
709 
710 	sched->state.event	= idx;		/* start with min weight */
711 	sched->state.weight	= wmin;
712 	sched->state.unassigned	= num;
713 }
714 
715 static void perf_sched_save_state(struct perf_sched *sched)
716 {
717 	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
718 		return;
719 
720 	sched->saved[sched->saved_states] = sched->state;
721 	sched->saved_states++;
722 }
723 
724 static bool perf_sched_restore_state(struct perf_sched *sched)
725 {
726 	if (!sched->saved_states)
727 		return false;
728 
729 	sched->saved_states--;
730 	sched->state = sched->saved[sched->saved_states];
731 
732 	/* continue with next counter: */
733 	clear_bit(sched->state.counter++, sched->state.used);
734 
735 	return true;
736 }
737 
738 /*
739  * Select a counter for the current event to schedule. Return true on
740  * success.
741  */
742 static bool __perf_sched_find_counter(struct perf_sched *sched)
743 {
744 	struct event_constraint *c;
745 	int idx;
746 
747 	if (!sched->state.unassigned)
748 		return false;
749 
750 	if (sched->state.event >= sched->max_events)
751 		return false;
752 
753 	c = sched->constraints[sched->state.event];
754 	/* Prefer fixed purpose counters */
755 	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
756 		idx = INTEL_PMC_IDX_FIXED;
757 		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
758 			if (!__test_and_set_bit(idx, sched->state.used))
759 				goto done;
760 		}
761 	}
762 
763 	/* Grab the first unused counter starting with idx */
764 	idx = sched->state.counter;
765 	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
766 		if (!__test_and_set_bit(idx, sched->state.used)) {
767 			if (sched->state.nr_gp++ >= sched->max_gp)
768 				return false;
769 
770 			goto done;
771 		}
772 	}
773 
774 	return false;
775 
776 done:
777 	sched->state.counter = idx;
778 
779 	if (c->overlap)
780 		perf_sched_save_state(sched);
781 
782 	return true;
783 }
784 
785 static bool perf_sched_find_counter(struct perf_sched *sched)
786 {
787 	while (!__perf_sched_find_counter(sched)) {
788 		if (!perf_sched_restore_state(sched))
789 			return false;
790 	}
791 
792 	return true;
793 }
794 
795 /*
796  * Go through all unassigned events and find the next one to schedule.
797  * Take events with the least weight first. Return true on success.
798  */
799 static bool perf_sched_next_event(struct perf_sched *sched)
800 {
801 	struct event_constraint *c;
802 
803 	if (!sched->state.unassigned || !--sched->state.unassigned)
804 		return false;
805 
806 	do {
807 		/* next event */
808 		sched->state.event++;
809 		if (sched->state.event >= sched->max_events) {
810 			/* next weight */
811 			sched->state.event = 0;
812 			sched->state.weight++;
813 			if (sched->state.weight > sched->max_weight)
814 				return false;
815 		}
816 		c = sched->constraints[sched->state.event];
817 	} while (c->weight != sched->state.weight);
818 
819 	sched->state.counter = 0;	/* start with first counter */
820 
821 	return true;
822 }
823 
824 /*
825  * Assign a counter for each event.
826  */
827 int perf_assign_events(struct event_constraint **constraints, int n,
828 			int wmin, int wmax, int gpmax, int *assign)
829 {
830 	struct perf_sched sched;
831 
832 	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
833 
834 	do {
835 		if (!perf_sched_find_counter(&sched))
836 			break;	/* failed */
837 		if (assign)
838 			assign[sched.state.event] = sched.state.counter;
839 	} while (perf_sched_next_event(&sched));
840 
841 	return sched.state.unassigned;
842 }
843 EXPORT_SYMBOL_GPL(perf_assign_events);
844 
845 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
846 {
847 	struct event_constraint *c;
848 	unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
849 	struct perf_event *e;
850 	int i, wmin, wmax, unsched = 0;
851 	struct hw_perf_event *hwc;
852 
853 	bitmap_zero(used_mask, X86_PMC_IDX_MAX);
854 
855 	if (x86_pmu.start_scheduling)
856 		x86_pmu.start_scheduling(cpuc);
857 
858 	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
859 		cpuc->event_constraint[i] = NULL;
860 		c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
861 		cpuc->event_constraint[i] = c;
862 
863 		wmin = min(wmin, c->weight);
864 		wmax = max(wmax, c->weight);
865 	}
866 
867 	/*
868 	 * fastpath, try to reuse previous register
869 	 */
870 	for (i = 0; i < n; i++) {
871 		hwc = &cpuc->event_list[i]->hw;
872 		c = cpuc->event_constraint[i];
873 
874 		/* never assigned */
875 		if (hwc->idx == -1)
876 			break;
877 
878 		/* constraint still honored */
879 		if (!test_bit(hwc->idx, c->idxmsk))
880 			break;
881 
882 		/* not already used */
883 		if (test_bit(hwc->idx, used_mask))
884 			break;
885 
886 		__set_bit(hwc->idx, used_mask);
887 		if (assign)
888 			assign[i] = hwc->idx;
889 	}
890 
891 	/* slow path */
892 	if (i != n) {
893 		int gpmax = x86_pmu.num_counters;
894 
895 		/*
896 		 * Do not allow scheduling of more than half the available
897 		 * generic counters.
898 		 *
899 		 * This helps avoid counter starvation of sibling thread by
900 		 * ensuring at most half the counters cannot be in exclusive
901 		 * mode. There is no designated counters for the limits. Any
902 		 * N/2 counters can be used. This helps with events with
903 		 * specific counter constraints.
904 		 */
905 		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
906 		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
907 			gpmax /= 2;
908 
909 		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
910 					     wmax, gpmax, assign);
911 	}
912 
913 	/*
914 	 * In case of success (unsched = 0), mark events as committed,
915 	 * so we do not put_constraint() in case new events are added
916 	 * and fail to be scheduled
917 	 *
918 	 * We invoke the lower level commit callback to lock the resource
919 	 *
920 	 * We do not need to do all of this in case we are called to
921 	 * validate an event group (assign == NULL)
922 	 */
923 	if (!unsched && assign) {
924 		for (i = 0; i < n; i++) {
925 			e = cpuc->event_list[i];
926 			e->hw.flags |= PERF_X86_EVENT_COMMITTED;
927 			if (x86_pmu.commit_scheduling)
928 				x86_pmu.commit_scheduling(cpuc, i, assign[i]);
929 		}
930 	} else {
931 		for (i = 0; i < n; i++) {
932 			e = cpuc->event_list[i];
933 			/*
934 			 * do not put_constraint() on comitted events,
935 			 * because they are good to go
936 			 */
937 			if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
938 				continue;
939 
940 			/*
941 			 * release events that failed scheduling
942 			 */
943 			if (x86_pmu.put_event_constraints)
944 				x86_pmu.put_event_constraints(cpuc, e);
945 		}
946 	}
947 
948 	if (x86_pmu.stop_scheduling)
949 		x86_pmu.stop_scheduling(cpuc);
950 
951 	return unsched ? -EINVAL : 0;
952 }
953 
954 /*
955  * dogrp: true if must collect siblings events (group)
956  * returns total number of events and error code
957  */
958 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
959 {
960 	struct perf_event *event;
961 	int n, max_count;
962 
963 	max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
964 
965 	/* current number of events already accepted */
966 	n = cpuc->n_events;
967 
968 	if (is_x86_event(leader)) {
969 		if (n >= max_count)
970 			return -EINVAL;
971 		cpuc->event_list[n] = leader;
972 		n++;
973 	}
974 	if (!dogrp)
975 		return n;
976 
977 	list_for_each_entry(event, &leader->sibling_list, group_entry) {
978 		if (!is_x86_event(event) ||
979 		    event->state <= PERF_EVENT_STATE_OFF)
980 			continue;
981 
982 		if (n >= max_count)
983 			return -EINVAL;
984 
985 		cpuc->event_list[n] = event;
986 		n++;
987 	}
988 	return n;
989 }
990 
991 static inline void x86_assign_hw_event(struct perf_event *event,
992 				struct cpu_hw_events *cpuc, int i)
993 {
994 	struct hw_perf_event *hwc = &event->hw;
995 
996 	hwc->idx = cpuc->assign[i];
997 	hwc->last_cpu = smp_processor_id();
998 	hwc->last_tag = ++cpuc->tags[i];
999 
1000 	if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
1001 		hwc->config_base = 0;
1002 		hwc->event_base	= 0;
1003 	} else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1004 		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1005 		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1006 		hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1007 	} else {
1008 		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1009 		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1010 		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1011 	}
1012 }
1013 
1014 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1015 					struct cpu_hw_events *cpuc,
1016 					int i)
1017 {
1018 	return hwc->idx == cpuc->assign[i] &&
1019 		hwc->last_cpu == smp_processor_id() &&
1020 		hwc->last_tag == cpuc->tags[i];
1021 }
1022 
1023 static void x86_pmu_start(struct perf_event *event, int flags);
1024 
1025 static void x86_pmu_enable(struct pmu *pmu)
1026 {
1027 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1028 	struct perf_event *event;
1029 	struct hw_perf_event *hwc;
1030 	int i, added = cpuc->n_added;
1031 
1032 	if (!x86_pmu_initialized())
1033 		return;
1034 
1035 	if (cpuc->enabled)
1036 		return;
1037 
1038 	if (cpuc->n_added) {
1039 		int n_running = cpuc->n_events - cpuc->n_added;
1040 		/*
1041 		 * apply assignment obtained either from
1042 		 * hw_perf_group_sched_in() or x86_pmu_enable()
1043 		 *
1044 		 * step1: save events moving to new counters
1045 		 */
1046 		for (i = 0; i < n_running; i++) {
1047 			event = cpuc->event_list[i];
1048 			hwc = &event->hw;
1049 
1050 			/*
1051 			 * we can avoid reprogramming counter if:
1052 			 * - assigned same counter as last time
1053 			 * - running on same CPU as last time
1054 			 * - no other event has used the counter since
1055 			 */
1056 			if (hwc->idx == -1 ||
1057 			    match_prev_assignment(hwc, cpuc, i))
1058 				continue;
1059 
1060 			/*
1061 			 * Ensure we don't accidentally enable a stopped
1062 			 * counter simply because we rescheduled.
1063 			 */
1064 			if (hwc->state & PERF_HES_STOPPED)
1065 				hwc->state |= PERF_HES_ARCH;
1066 
1067 			x86_pmu_stop(event, PERF_EF_UPDATE);
1068 		}
1069 
1070 		/*
1071 		 * step2: reprogram moved events into new counters
1072 		 */
1073 		for (i = 0; i < cpuc->n_events; i++) {
1074 			event = cpuc->event_list[i];
1075 			hwc = &event->hw;
1076 
1077 			if (!match_prev_assignment(hwc, cpuc, i))
1078 				x86_assign_hw_event(event, cpuc, i);
1079 			else if (i < n_running)
1080 				continue;
1081 
1082 			if (hwc->state & PERF_HES_ARCH)
1083 				continue;
1084 
1085 			x86_pmu_start(event, PERF_EF_RELOAD);
1086 		}
1087 		cpuc->n_added = 0;
1088 		perf_events_lapic_init();
1089 	}
1090 
1091 	cpuc->enabled = 1;
1092 	barrier();
1093 
1094 	x86_pmu.enable_all(added);
1095 }
1096 
1097 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1098 
1099 /*
1100  * Set the next IRQ period, based on the hwc->period_left value.
1101  * To be called with the event disabled in hw:
1102  */
1103 int x86_perf_event_set_period(struct perf_event *event)
1104 {
1105 	struct hw_perf_event *hwc = &event->hw;
1106 	s64 left = local64_read(&hwc->period_left);
1107 	s64 period = hwc->sample_period;
1108 	int ret = 0, idx = hwc->idx;
1109 
1110 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
1111 		return 0;
1112 
1113 	/*
1114 	 * If we are way outside a reasonable range then just skip forward:
1115 	 */
1116 	if (unlikely(left <= -period)) {
1117 		left = period;
1118 		local64_set(&hwc->period_left, left);
1119 		hwc->last_period = period;
1120 		ret = 1;
1121 	}
1122 
1123 	if (unlikely(left <= 0)) {
1124 		left += period;
1125 		local64_set(&hwc->period_left, left);
1126 		hwc->last_period = period;
1127 		ret = 1;
1128 	}
1129 	/*
1130 	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1131 	 */
1132 	if (unlikely(left < 2))
1133 		left = 2;
1134 
1135 	if (left > x86_pmu.max_period)
1136 		left = x86_pmu.max_period;
1137 
1138 	if (x86_pmu.limit_period)
1139 		left = x86_pmu.limit_period(event, left);
1140 
1141 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1142 
1143 	if (!(hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) ||
1144 	    local64_read(&hwc->prev_count) != (u64)-left) {
1145 		/*
1146 		 * The hw event starts counting from this event offset,
1147 		 * mark it to be able to extra future deltas:
1148 		 */
1149 		local64_set(&hwc->prev_count, (u64)-left);
1150 
1151 		wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1152 	}
1153 
1154 	/*
1155 	 * Due to erratum on certan cpu we need
1156 	 * a second write to be sure the register
1157 	 * is updated properly
1158 	 */
1159 	if (x86_pmu.perfctr_second_write) {
1160 		wrmsrl(hwc->event_base,
1161 			(u64)(-left) & x86_pmu.cntval_mask);
1162 	}
1163 
1164 	perf_event_update_userpage(event);
1165 
1166 	return ret;
1167 }
1168 
1169 void x86_pmu_enable_event(struct perf_event *event)
1170 {
1171 	if (__this_cpu_read(cpu_hw_events.enabled))
1172 		__x86_pmu_enable_event(&event->hw,
1173 				       ARCH_PERFMON_EVENTSEL_ENABLE);
1174 }
1175 
1176 /*
1177  * Add a single event to the PMU.
1178  *
1179  * The event is added to the group of enabled events
1180  * but only if it can be scehduled with existing events.
1181  */
1182 static int x86_pmu_add(struct perf_event *event, int flags)
1183 {
1184 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1185 	struct hw_perf_event *hwc;
1186 	int assign[X86_PMC_IDX_MAX];
1187 	int n, n0, ret;
1188 
1189 	hwc = &event->hw;
1190 
1191 	n0 = cpuc->n_events;
1192 	ret = n = collect_events(cpuc, event, false);
1193 	if (ret < 0)
1194 		goto out;
1195 
1196 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1197 	if (!(flags & PERF_EF_START))
1198 		hwc->state |= PERF_HES_ARCH;
1199 
1200 	/*
1201 	 * If group events scheduling transaction was started,
1202 	 * skip the schedulability test here, it will be performed
1203 	 * at commit time (->commit_txn) as a whole.
1204 	 */
1205 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1206 		goto done_collect;
1207 
1208 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1209 	if (ret)
1210 		goto out;
1211 	/*
1212 	 * copy new assignment, now we know it is possible
1213 	 * will be used by hw_perf_enable()
1214 	 */
1215 	memcpy(cpuc->assign, assign, n*sizeof(int));
1216 
1217 done_collect:
1218 	/*
1219 	 * Commit the collect_events() state. See x86_pmu_del() and
1220 	 * x86_pmu_*_txn().
1221 	 */
1222 	cpuc->n_events = n;
1223 	cpuc->n_added += n - n0;
1224 	cpuc->n_txn += n - n0;
1225 
1226 	ret = 0;
1227 out:
1228 	return ret;
1229 }
1230 
1231 static void x86_pmu_start(struct perf_event *event, int flags)
1232 {
1233 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1234 	int idx = event->hw.idx;
1235 
1236 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1237 		return;
1238 
1239 	if (WARN_ON_ONCE(idx == -1))
1240 		return;
1241 
1242 	if (flags & PERF_EF_RELOAD) {
1243 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1244 		x86_perf_event_set_period(event);
1245 	}
1246 
1247 	event->hw.state = 0;
1248 
1249 	cpuc->events[idx] = event;
1250 	__set_bit(idx, cpuc->active_mask);
1251 	__set_bit(idx, cpuc->running);
1252 	x86_pmu.enable(event);
1253 	perf_event_update_userpage(event);
1254 }
1255 
1256 void perf_event_print_debug(void)
1257 {
1258 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1259 	u64 pebs, debugctl;
1260 	struct cpu_hw_events *cpuc;
1261 	unsigned long flags;
1262 	int cpu, idx;
1263 
1264 	if (!x86_pmu.num_counters)
1265 		return;
1266 
1267 	local_irq_save(flags);
1268 
1269 	cpu = smp_processor_id();
1270 	cpuc = &per_cpu(cpu_hw_events, cpu);
1271 
1272 	if (x86_pmu.version >= 2) {
1273 		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1274 		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1275 		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1276 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1277 
1278 		pr_info("\n");
1279 		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1280 		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1281 		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1282 		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1283 		if (x86_pmu.pebs_constraints) {
1284 			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1285 			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1286 		}
1287 		if (x86_pmu.lbr_nr) {
1288 			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1289 			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1290 		}
1291 	}
1292 	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1293 
1294 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1295 		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1296 		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1297 
1298 		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1299 
1300 		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1301 			cpu, idx, pmc_ctrl);
1302 		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1303 			cpu, idx, pmc_count);
1304 		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1305 			cpu, idx, prev_left);
1306 	}
1307 	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1308 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1309 
1310 		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1311 			cpu, idx, pmc_count);
1312 	}
1313 	local_irq_restore(flags);
1314 }
1315 
1316 void x86_pmu_stop(struct perf_event *event, int flags)
1317 {
1318 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1319 	struct hw_perf_event *hwc = &event->hw;
1320 
1321 	if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1322 		x86_pmu.disable(event);
1323 		cpuc->events[hwc->idx] = NULL;
1324 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1325 		hwc->state |= PERF_HES_STOPPED;
1326 	}
1327 
1328 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1329 		/*
1330 		 * Drain the remaining delta count out of a event
1331 		 * that we are disabling:
1332 		 */
1333 		x86_perf_event_update(event);
1334 		hwc->state |= PERF_HES_UPTODATE;
1335 	}
1336 }
1337 
1338 static void x86_pmu_del(struct perf_event *event, int flags)
1339 {
1340 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1341 	int i;
1342 
1343 	/*
1344 	 * event is descheduled
1345 	 */
1346 	event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1347 
1348 	/*
1349 	 * If we're called during a txn, we don't need to do anything.
1350 	 * The events never got scheduled and ->cancel_txn will truncate
1351 	 * the event_list.
1352 	 *
1353 	 * XXX assumes any ->del() called during a TXN will only be on
1354 	 * an event added during that same TXN.
1355 	 */
1356 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1357 		return;
1358 
1359 	/*
1360 	 * Not a TXN, therefore cleanup properly.
1361 	 */
1362 	x86_pmu_stop(event, PERF_EF_UPDATE);
1363 
1364 	for (i = 0; i < cpuc->n_events; i++) {
1365 		if (event == cpuc->event_list[i])
1366 			break;
1367 	}
1368 
1369 	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1370 		return;
1371 
1372 	/* If we have a newly added event; make sure to decrease n_added. */
1373 	if (i >= cpuc->n_events - cpuc->n_added)
1374 		--cpuc->n_added;
1375 
1376 	if (x86_pmu.put_event_constraints)
1377 		x86_pmu.put_event_constraints(cpuc, event);
1378 
1379 	/* Delete the array entry. */
1380 	while (++i < cpuc->n_events) {
1381 		cpuc->event_list[i-1] = cpuc->event_list[i];
1382 		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1383 	}
1384 	--cpuc->n_events;
1385 
1386 	perf_event_update_userpage(event);
1387 }
1388 
1389 int x86_pmu_handle_irq(struct pt_regs *regs)
1390 {
1391 	struct perf_sample_data data;
1392 	struct cpu_hw_events *cpuc;
1393 	struct perf_event *event;
1394 	int idx, handled = 0;
1395 	u64 val;
1396 
1397 	cpuc = this_cpu_ptr(&cpu_hw_events);
1398 
1399 	/*
1400 	 * Some chipsets need to unmask the LVTPC in a particular spot
1401 	 * inside the nmi handler.  As a result, the unmasking was pushed
1402 	 * into all the nmi handlers.
1403 	 *
1404 	 * This generic handler doesn't seem to have any issues where the
1405 	 * unmasking occurs so it was left at the top.
1406 	 */
1407 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1408 
1409 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1410 		if (!test_bit(idx, cpuc->active_mask)) {
1411 			/*
1412 			 * Though we deactivated the counter some cpus
1413 			 * might still deliver spurious interrupts still
1414 			 * in flight. Catch them:
1415 			 */
1416 			if (__test_and_clear_bit(idx, cpuc->running))
1417 				handled++;
1418 			continue;
1419 		}
1420 
1421 		event = cpuc->events[idx];
1422 
1423 		val = x86_perf_event_update(event);
1424 		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1425 			continue;
1426 
1427 		/*
1428 		 * event overflow
1429 		 */
1430 		handled++;
1431 		perf_sample_data_init(&data, 0, event->hw.last_period);
1432 
1433 		if (!x86_perf_event_set_period(event))
1434 			continue;
1435 
1436 		if (perf_event_overflow(event, &data, regs))
1437 			x86_pmu_stop(event, 0);
1438 	}
1439 
1440 	if (handled)
1441 		inc_irq_stat(apic_perf_irqs);
1442 
1443 	return handled;
1444 }
1445 
1446 void perf_events_lapic_init(void)
1447 {
1448 	if (!x86_pmu.apic || !x86_pmu_initialized())
1449 		return;
1450 
1451 	/*
1452 	 * Always use NMI for PMU
1453 	 */
1454 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1455 }
1456 
1457 static int
1458 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1459 {
1460 	u64 start_clock;
1461 	u64 finish_clock;
1462 	int ret;
1463 
1464 	/*
1465 	 * All PMUs/events that share this PMI handler should make sure to
1466 	 * increment active_events for their events.
1467 	 */
1468 	if (!atomic_read(&active_events))
1469 		return NMI_DONE;
1470 
1471 	start_clock = sched_clock();
1472 	ret = x86_pmu.handle_irq(regs);
1473 	finish_clock = sched_clock();
1474 
1475 	perf_sample_event_took(finish_clock - start_clock);
1476 
1477 	return ret;
1478 }
1479 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1480 
1481 struct event_constraint emptyconstraint;
1482 struct event_constraint unconstrained;
1483 
1484 static int x86_pmu_prepare_cpu(unsigned int cpu)
1485 {
1486 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1487 	int i;
1488 
1489 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1490 		cpuc->kfree_on_online[i] = NULL;
1491 	if (x86_pmu.cpu_prepare)
1492 		return x86_pmu.cpu_prepare(cpu);
1493 	return 0;
1494 }
1495 
1496 static int x86_pmu_dead_cpu(unsigned int cpu)
1497 {
1498 	if (x86_pmu.cpu_dead)
1499 		x86_pmu.cpu_dead(cpu);
1500 	return 0;
1501 }
1502 
1503 static int x86_pmu_online_cpu(unsigned int cpu)
1504 {
1505 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1506 	int i;
1507 
1508 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1509 		kfree(cpuc->kfree_on_online[i]);
1510 		cpuc->kfree_on_online[i] = NULL;
1511 	}
1512 	return 0;
1513 }
1514 
1515 static int x86_pmu_starting_cpu(unsigned int cpu)
1516 {
1517 	if (x86_pmu.cpu_starting)
1518 		x86_pmu.cpu_starting(cpu);
1519 	return 0;
1520 }
1521 
1522 static int x86_pmu_dying_cpu(unsigned int cpu)
1523 {
1524 	if (x86_pmu.cpu_dying)
1525 		x86_pmu.cpu_dying(cpu);
1526 	return 0;
1527 }
1528 
1529 static void __init pmu_check_apic(void)
1530 {
1531 	if (boot_cpu_has(X86_FEATURE_APIC))
1532 		return;
1533 
1534 	x86_pmu.apic = 0;
1535 	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1536 	pr_info("no hardware sampling interrupt available.\n");
1537 
1538 	/*
1539 	 * If we have a PMU initialized but no APIC
1540 	 * interrupts, we cannot sample hardware
1541 	 * events (user-space has to fall back and
1542 	 * sample via a hrtimer based software event):
1543 	 */
1544 	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1545 
1546 }
1547 
1548 static struct attribute_group x86_pmu_format_group = {
1549 	.name = "format",
1550 	.attrs = NULL,
1551 };
1552 
1553 /*
1554  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1555  * out of events_attr attributes.
1556  */
1557 static void __init filter_events(struct attribute **attrs)
1558 {
1559 	struct device_attribute *d;
1560 	struct perf_pmu_events_attr *pmu_attr;
1561 	int offset = 0;
1562 	int i, j;
1563 
1564 	for (i = 0; attrs[i]; i++) {
1565 		d = (struct device_attribute *)attrs[i];
1566 		pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1567 		/* str trumps id */
1568 		if (pmu_attr->event_str)
1569 			continue;
1570 		if (x86_pmu.event_map(i + offset))
1571 			continue;
1572 
1573 		for (j = i; attrs[j]; j++)
1574 			attrs[j] = attrs[j + 1];
1575 
1576 		/* Check the shifted attr. */
1577 		i--;
1578 
1579 		/*
1580 		 * event_map() is index based, the attrs array is organized
1581 		 * by increasing event index. If we shift the events, then
1582 		 * we need to compensate for the event_map(), otherwise
1583 		 * we are looking up the wrong event in the map
1584 		 */
1585 		offset++;
1586 	}
1587 }
1588 
1589 /* Merge two pointer arrays */
1590 __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1591 {
1592 	struct attribute **new;
1593 	int j, i;
1594 
1595 	for (j = 0; a[j]; j++)
1596 		;
1597 	for (i = 0; b[i]; i++)
1598 		j++;
1599 	j++;
1600 
1601 	new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1602 	if (!new)
1603 		return NULL;
1604 
1605 	j = 0;
1606 	for (i = 0; a[i]; i++)
1607 		new[j++] = a[i];
1608 	for (i = 0; b[i]; i++)
1609 		new[j++] = b[i];
1610 	new[j] = NULL;
1611 
1612 	return new;
1613 }
1614 
1615 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1616 {
1617 	struct perf_pmu_events_attr *pmu_attr = \
1618 		container_of(attr, struct perf_pmu_events_attr, attr);
1619 	u64 config = x86_pmu.event_map(pmu_attr->id);
1620 
1621 	/* string trumps id */
1622 	if (pmu_attr->event_str)
1623 		return sprintf(page, "%s", pmu_attr->event_str);
1624 
1625 	return x86_pmu.events_sysfs_show(page, config);
1626 }
1627 EXPORT_SYMBOL_GPL(events_sysfs_show);
1628 
1629 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1630 			  char *page)
1631 {
1632 	struct perf_pmu_events_ht_attr *pmu_attr =
1633 		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1634 
1635 	/*
1636 	 * Report conditional events depending on Hyper-Threading.
1637 	 *
1638 	 * This is overly conservative as usually the HT special
1639 	 * handling is not needed if the other CPU thread is idle.
1640 	 *
1641 	 * Note this does not (and cannot) handle the case when thread
1642 	 * siblings are invisible, for example with virtualization
1643 	 * if they are owned by some other guest.  The user tool
1644 	 * has to re-read when a thread sibling gets onlined later.
1645 	 */
1646 	return sprintf(page, "%s",
1647 			topology_max_smt_threads() > 1 ?
1648 			pmu_attr->event_str_ht :
1649 			pmu_attr->event_str_noht);
1650 }
1651 
1652 EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1653 EVENT_ATTR(instructions,		INSTRUCTIONS		);
1654 EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1655 EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1656 EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1657 EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1658 EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1659 EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1660 EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1661 EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1662 
1663 static struct attribute *empty_attrs;
1664 
1665 static struct attribute *events_attr[] = {
1666 	EVENT_PTR(CPU_CYCLES),
1667 	EVENT_PTR(INSTRUCTIONS),
1668 	EVENT_PTR(CACHE_REFERENCES),
1669 	EVENT_PTR(CACHE_MISSES),
1670 	EVENT_PTR(BRANCH_INSTRUCTIONS),
1671 	EVENT_PTR(BRANCH_MISSES),
1672 	EVENT_PTR(BUS_CYCLES),
1673 	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1674 	EVENT_PTR(STALLED_CYCLES_BACKEND),
1675 	EVENT_PTR(REF_CPU_CYCLES),
1676 	NULL,
1677 };
1678 
1679 static struct attribute_group x86_pmu_events_group = {
1680 	.name = "events",
1681 	.attrs = events_attr,
1682 };
1683 
1684 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1685 {
1686 	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1687 	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1688 	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1689 	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1690 	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1691 	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1692 	ssize_t ret;
1693 
1694 	/*
1695 	* We have whole page size to spend and just little data
1696 	* to write, so we can safely use sprintf.
1697 	*/
1698 	ret = sprintf(page, "event=0x%02llx", event);
1699 
1700 	if (umask)
1701 		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1702 
1703 	if (edge)
1704 		ret += sprintf(page + ret, ",edge");
1705 
1706 	if (pc)
1707 		ret += sprintf(page + ret, ",pc");
1708 
1709 	if (any)
1710 		ret += sprintf(page + ret, ",any");
1711 
1712 	if (inv)
1713 		ret += sprintf(page + ret, ",inv");
1714 
1715 	if (cmask)
1716 		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1717 
1718 	ret += sprintf(page + ret, "\n");
1719 
1720 	return ret;
1721 }
1722 
1723 static int __init init_hw_perf_events(void)
1724 {
1725 	struct x86_pmu_quirk *quirk;
1726 	int err;
1727 
1728 	pr_info("Performance Events: ");
1729 
1730 	switch (boot_cpu_data.x86_vendor) {
1731 	case X86_VENDOR_INTEL:
1732 		err = intel_pmu_init();
1733 		break;
1734 	case X86_VENDOR_AMD:
1735 		err = amd_pmu_init();
1736 		break;
1737 	default:
1738 		err = -ENOTSUPP;
1739 	}
1740 	if (err != 0) {
1741 		pr_cont("no PMU driver, software events only.\n");
1742 		return 0;
1743 	}
1744 
1745 	pmu_check_apic();
1746 
1747 	/* sanity check that the hardware exists or is emulated */
1748 	if (!check_hw_exists())
1749 		return 0;
1750 
1751 	pr_cont("%s PMU driver.\n", x86_pmu.name);
1752 
1753 	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1754 
1755 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1756 		quirk->func();
1757 
1758 	if (!x86_pmu.intel_ctrl)
1759 		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1760 
1761 	perf_events_lapic_init();
1762 	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1763 
1764 	unconstrained = (struct event_constraint)
1765 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1766 				   0, x86_pmu.num_counters, 0, 0);
1767 
1768 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1769 
1770 	if (x86_pmu.event_attrs)
1771 		x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1772 
1773 	if (!x86_pmu.events_sysfs_show)
1774 		x86_pmu_events_group.attrs = &empty_attrs;
1775 	else
1776 		filter_events(x86_pmu_events_group.attrs);
1777 
1778 	if (x86_pmu.cpu_events) {
1779 		struct attribute **tmp;
1780 
1781 		tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1782 		if (!WARN_ON(!tmp))
1783 			x86_pmu_events_group.attrs = tmp;
1784 	}
1785 
1786 	pr_info("... version:                %d\n",     x86_pmu.version);
1787 	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1788 	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1789 	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1790 	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1791 	pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1792 	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1793 
1794 	/*
1795 	 * Install callbacks. Core will call them for each online
1796 	 * cpu.
1797 	 */
1798 	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "PERF_X86_PREPARE",
1799 				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1800 	if (err)
1801 		return err;
1802 
1803 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1804 				"AP_PERF_X86_STARTING", x86_pmu_starting_cpu,
1805 				x86_pmu_dying_cpu);
1806 	if (err)
1807 		goto out;
1808 
1809 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "AP_PERF_X86_ONLINE",
1810 				x86_pmu_online_cpu, NULL);
1811 	if (err)
1812 		goto out1;
1813 
1814 	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1815 	if (err)
1816 		goto out2;
1817 
1818 	return 0;
1819 
1820 out2:
1821 	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1822 out1:
1823 	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1824 out:
1825 	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
1826 	return err;
1827 }
1828 early_initcall(init_hw_perf_events);
1829 
1830 static inline void x86_pmu_read(struct perf_event *event)
1831 {
1832 	x86_perf_event_update(event);
1833 }
1834 
1835 /*
1836  * Start group events scheduling transaction
1837  * Set the flag to make pmu::enable() not perform the
1838  * schedulability test, it will be performed at commit time
1839  *
1840  * We only support PERF_PMU_TXN_ADD transactions. Save the
1841  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1842  * transactions.
1843  */
1844 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1845 {
1846 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1847 
1848 	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
1849 
1850 	cpuc->txn_flags = txn_flags;
1851 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1852 		return;
1853 
1854 	perf_pmu_disable(pmu);
1855 	__this_cpu_write(cpu_hw_events.n_txn, 0);
1856 }
1857 
1858 /*
1859  * Stop group events scheduling transaction
1860  * Clear the flag and pmu::enable() will perform the
1861  * schedulability test.
1862  */
1863 static void x86_pmu_cancel_txn(struct pmu *pmu)
1864 {
1865 	unsigned int txn_flags;
1866 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1867 
1868 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1869 
1870 	txn_flags = cpuc->txn_flags;
1871 	cpuc->txn_flags = 0;
1872 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1873 		return;
1874 
1875 	/*
1876 	 * Truncate collected array by the number of events added in this
1877 	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1878 	 */
1879 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1880 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1881 	perf_pmu_enable(pmu);
1882 }
1883 
1884 /*
1885  * Commit group events scheduling transaction
1886  * Perform the group schedulability test as a whole
1887  * Return 0 if success
1888  *
1889  * Does not cancel the transaction on failure; expects the caller to do this.
1890  */
1891 static int x86_pmu_commit_txn(struct pmu *pmu)
1892 {
1893 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1894 	int assign[X86_PMC_IDX_MAX];
1895 	int n, ret;
1896 
1897 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1898 
1899 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1900 		cpuc->txn_flags = 0;
1901 		return 0;
1902 	}
1903 
1904 	n = cpuc->n_events;
1905 
1906 	if (!x86_pmu_initialized())
1907 		return -EAGAIN;
1908 
1909 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1910 	if (ret)
1911 		return ret;
1912 
1913 	/*
1914 	 * copy new assignment, now we know it is possible
1915 	 * will be used by hw_perf_enable()
1916 	 */
1917 	memcpy(cpuc->assign, assign, n*sizeof(int));
1918 
1919 	cpuc->txn_flags = 0;
1920 	perf_pmu_enable(pmu);
1921 	return 0;
1922 }
1923 /*
1924  * a fake_cpuc is used to validate event groups. Due to
1925  * the extra reg logic, we need to also allocate a fake
1926  * per_core and per_cpu structure. Otherwise, group events
1927  * using extra reg may conflict without the kernel being
1928  * able to catch this when the last event gets added to
1929  * the group.
1930  */
1931 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1932 {
1933 	kfree(cpuc->shared_regs);
1934 	kfree(cpuc);
1935 }
1936 
1937 static struct cpu_hw_events *allocate_fake_cpuc(void)
1938 {
1939 	struct cpu_hw_events *cpuc;
1940 	int cpu = raw_smp_processor_id();
1941 
1942 	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1943 	if (!cpuc)
1944 		return ERR_PTR(-ENOMEM);
1945 
1946 	/* only needed, if we have extra_regs */
1947 	if (x86_pmu.extra_regs) {
1948 		cpuc->shared_regs = allocate_shared_regs(cpu);
1949 		if (!cpuc->shared_regs)
1950 			goto error;
1951 	}
1952 	cpuc->is_fake = 1;
1953 	return cpuc;
1954 error:
1955 	free_fake_cpuc(cpuc);
1956 	return ERR_PTR(-ENOMEM);
1957 }
1958 
1959 /*
1960  * validate that we can schedule this event
1961  */
1962 static int validate_event(struct perf_event *event)
1963 {
1964 	struct cpu_hw_events *fake_cpuc;
1965 	struct event_constraint *c;
1966 	int ret = 0;
1967 
1968 	fake_cpuc = allocate_fake_cpuc();
1969 	if (IS_ERR(fake_cpuc))
1970 		return PTR_ERR(fake_cpuc);
1971 
1972 	c = x86_pmu.get_event_constraints(fake_cpuc, -1, event);
1973 
1974 	if (!c || !c->weight)
1975 		ret = -EINVAL;
1976 
1977 	if (x86_pmu.put_event_constraints)
1978 		x86_pmu.put_event_constraints(fake_cpuc, event);
1979 
1980 	free_fake_cpuc(fake_cpuc);
1981 
1982 	return ret;
1983 }
1984 
1985 /*
1986  * validate a single event group
1987  *
1988  * validation include:
1989  *	- check events are compatible which each other
1990  *	- events do not compete for the same counter
1991  *	- number of events <= number of counters
1992  *
1993  * validation ensures the group can be loaded onto the
1994  * PMU if it was the only group available.
1995  */
1996 static int validate_group(struct perf_event *event)
1997 {
1998 	struct perf_event *leader = event->group_leader;
1999 	struct cpu_hw_events *fake_cpuc;
2000 	int ret = -EINVAL, n;
2001 
2002 	fake_cpuc = allocate_fake_cpuc();
2003 	if (IS_ERR(fake_cpuc))
2004 		return PTR_ERR(fake_cpuc);
2005 	/*
2006 	 * the event is not yet connected with its
2007 	 * siblings therefore we must first collect
2008 	 * existing siblings, then add the new event
2009 	 * before we can simulate the scheduling
2010 	 */
2011 	n = collect_events(fake_cpuc, leader, true);
2012 	if (n < 0)
2013 		goto out;
2014 
2015 	fake_cpuc->n_events = n;
2016 	n = collect_events(fake_cpuc, event, false);
2017 	if (n < 0)
2018 		goto out;
2019 
2020 	fake_cpuc->n_events = n;
2021 
2022 	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2023 
2024 out:
2025 	free_fake_cpuc(fake_cpuc);
2026 	return ret;
2027 }
2028 
2029 static int x86_pmu_event_init(struct perf_event *event)
2030 {
2031 	struct pmu *tmp;
2032 	int err;
2033 
2034 	switch (event->attr.type) {
2035 	case PERF_TYPE_RAW:
2036 	case PERF_TYPE_HARDWARE:
2037 	case PERF_TYPE_HW_CACHE:
2038 		break;
2039 
2040 	default:
2041 		return -ENOENT;
2042 	}
2043 
2044 	err = __x86_pmu_event_init(event);
2045 	if (!err) {
2046 		/*
2047 		 * we temporarily connect event to its pmu
2048 		 * such that validate_group() can classify
2049 		 * it as an x86 event using is_x86_event()
2050 		 */
2051 		tmp = event->pmu;
2052 		event->pmu = &pmu;
2053 
2054 		if (event->group_leader != event)
2055 			err = validate_group(event);
2056 		else
2057 			err = validate_event(event);
2058 
2059 		event->pmu = tmp;
2060 	}
2061 	if (err) {
2062 		if (event->destroy)
2063 			event->destroy(event);
2064 	}
2065 
2066 	if (ACCESS_ONCE(x86_pmu.attr_rdpmc))
2067 		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2068 
2069 	return err;
2070 }
2071 
2072 static void refresh_pce(void *ignored)
2073 {
2074 	if (current->mm)
2075 		load_mm_cr4(current->mm);
2076 }
2077 
2078 static void x86_pmu_event_mapped(struct perf_event *event)
2079 {
2080 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2081 		return;
2082 
2083 	if (atomic_inc_return(&current->mm->context.perf_rdpmc_allowed) == 1)
2084 		on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2085 }
2086 
2087 static void x86_pmu_event_unmapped(struct perf_event *event)
2088 {
2089 	if (!current->mm)
2090 		return;
2091 
2092 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2093 		return;
2094 
2095 	if (atomic_dec_and_test(&current->mm->context.perf_rdpmc_allowed))
2096 		on_each_cpu_mask(mm_cpumask(current->mm), refresh_pce, NULL, 1);
2097 }
2098 
2099 static int x86_pmu_event_idx(struct perf_event *event)
2100 {
2101 	int idx = event->hw.idx;
2102 
2103 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2104 		return 0;
2105 
2106 	if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2107 		idx -= INTEL_PMC_IDX_FIXED;
2108 		idx |= 1 << 30;
2109 	}
2110 
2111 	return idx + 1;
2112 }
2113 
2114 static ssize_t get_attr_rdpmc(struct device *cdev,
2115 			      struct device_attribute *attr,
2116 			      char *buf)
2117 {
2118 	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2119 }
2120 
2121 static ssize_t set_attr_rdpmc(struct device *cdev,
2122 			      struct device_attribute *attr,
2123 			      const char *buf, size_t count)
2124 {
2125 	unsigned long val;
2126 	ssize_t ret;
2127 
2128 	ret = kstrtoul(buf, 0, &val);
2129 	if (ret)
2130 		return ret;
2131 
2132 	if (val > 2)
2133 		return -EINVAL;
2134 
2135 	if (x86_pmu.attr_rdpmc_broken)
2136 		return -ENOTSUPP;
2137 
2138 	if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2139 		/*
2140 		 * Changing into or out of always available, aka
2141 		 * perf-event-bypassing mode.  This path is extremely slow,
2142 		 * but only root can trigger it, so it's okay.
2143 		 */
2144 		if (val == 2)
2145 			static_key_slow_inc(&rdpmc_always_available);
2146 		else
2147 			static_key_slow_dec(&rdpmc_always_available);
2148 		on_each_cpu(refresh_pce, NULL, 1);
2149 	}
2150 
2151 	x86_pmu.attr_rdpmc = val;
2152 
2153 	return count;
2154 }
2155 
2156 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2157 
2158 static struct attribute *x86_pmu_attrs[] = {
2159 	&dev_attr_rdpmc.attr,
2160 	NULL,
2161 };
2162 
2163 static struct attribute_group x86_pmu_attr_group = {
2164 	.attrs = x86_pmu_attrs,
2165 };
2166 
2167 static const struct attribute_group *x86_pmu_attr_groups[] = {
2168 	&x86_pmu_attr_group,
2169 	&x86_pmu_format_group,
2170 	&x86_pmu_events_group,
2171 	NULL,
2172 };
2173 
2174 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2175 {
2176 	if (x86_pmu.sched_task)
2177 		x86_pmu.sched_task(ctx, sched_in);
2178 }
2179 
2180 void perf_check_microcode(void)
2181 {
2182 	if (x86_pmu.check_microcode)
2183 		x86_pmu.check_microcode();
2184 }
2185 EXPORT_SYMBOL_GPL(perf_check_microcode);
2186 
2187 static struct pmu pmu = {
2188 	.pmu_enable		= x86_pmu_enable,
2189 	.pmu_disable		= x86_pmu_disable,
2190 
2191 	.attr_groups		= x86_pmu_attr_groups,
2192 
2193 	.event_init		= x86_pmu_event_init,
2194 
2195 	.event_mapped		= x86_pmu_event_mapped,
2196 	.event_unmapped		= x86_pmu_event_unmapped,
2197 
2198 	.add			= x86_pmu_add,
2199 	.del			= x86_pmu_del,
2200 	.start			= x86_pmu_start,
2201 	.stop			= x86_pmu_stop,
2202 	.read			= x86_pmu_read,
2203 
2204 	.start_txn		= x86_pmu_start_txn,
2205 	.cancel_txn		= x86_pmu_cancel_txn,
2206 	.commit_txn		= x86_pmu_commit_txn,
2207 
2208 	.event_idx		= x86_pmu_event_idx,
2209 	.sched_task		= x86_pmu_sched_task,
2210 	.task_ctx_size          = sizeof(struct x86_perf_task_context),
2211 };
2212 
2213 void arch_perf_update_userpage(struct perf_event *event,
2214 			       struct perf_event_mmap_page *userpg, u64 now)
2215 {
2216 	struct cyc2ns_data *data;
2217 
2218 	userpg->cap_user_time = 0;
2219 	userpg->cap_user_time_zero = 0;
2220 	userpg->cap_user_rdpmc =
2221 		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2222 	userpg->pmc_width = x86_pmu.cntval_bits;
2223 
2224 	if (!sched_clock_stable())
2225 		return;
2226 
2227 	data = cyc2ns_read_begin();
2228 
2229 	/*
2230 	 * Internal timekeeping for enabled/running/stopped times
2231 	 * is always in the local_clock domain.
2232 	 */
2233 	userpg->cap_user_time = 1;
2234 	userpg->time_mult = data->cyc2ns_mul;
2235 	userpg->time_shift = data->cyc2ns_shift;
2236 	userpg->time_offset = data->cyc2ns_offset - now;
2237 
2238 	/*
2239 	 * cap_user_time_zero doesn't make sense when we're using a different
2240 	 * time base for the records.
2241 	 */
2242 	if (!event->attr.use_clockid) {
2243 		userpg->cap_user_time_zero = 1;
2244 		userpg->time_zero = data->cyc2ns_offset;
2245 	}
2246 
2247 	cyc2ns_read_end(data);
2248 }
2249 
2250 /*
2251  * callchain support
2252  */
2253 
2254 static int backtrace_stack(void *data, char *name)
2255 {
2256 	return 0;
2257 }
2258 
2259 static int backtrace_address(void *data, unsigned long addr, int reliable)
2260 {
2261 	struct perf_callchain_entry_ctx *entry = data;
2262 
2263 	return perf_callchain_store(entry, addr);
2264 }
2265 
2266 static const struct stacktrace_ops backtrace_ops = {
2267 	.stack			= backtrace_stack,
2268 	.address		= backtrace_address,
2269 	.walk_stack		= print_context_stack_bp,
2270 };
2271 
2272 void
2273 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2274 {
2275 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2276 		/* TODO: We don't support guest os callchain now */
2277 		return;
2278 	}
2279 
2280 	perf_callchain_store(entry, regs->ip);
2281 
2282 	dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
2283 }
2284 
2285 static inline int
2286 valid_user_frame(const void __user *fp, unsigned long size)
2287 {
2288 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2289 }
2290 
2291 static unsigned long get_segment_base(unsigned int segment)
2292 {
2293 	struct desc_struct *desc;
2294 	int idx = segment >> 3;
2295 
2296 	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2297 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2298 		struct ldt_struct *ldt;
2299 
2300 		if (idx > LDT_ENTRIES)
2301 			return 0;
2302 
2303 		/* IRQs are off, so this synchronizes with smp_store_release */
2304 		ldt = lockless_dereference(current->active_mm->context.ldt);
2305 		if (!ldt || idx > ldt->size)
2306 			return 0;
2307 
2308 		desc = &ldt->entries[idx];
2309 #else
2310 		return 0;
2311 #endif
2312 	} else {
2313 		if (idx > GDT_ENTRIES)
2314 			return 0;
2315 
2316 		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2317 	}
2318 
2319 	return get_desc_base(desc);
2320 }
2321 
2322 #ifdef CONFIG_IA32_EMULATION
2323 
2324 #include <asm/compat.h>
2325 
2326 static inline int
2327 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2328 {
2329 	/* 32-bit process in 64-bit kernel. */
2330 	unsigned long ss_base, cs_base;
2331 	struct stack_frame_ia32 frame;
2332 	const void __user *fp;
2333 
2334 	if (!test_thread_flag(TIF_IA32))
2335 		return 0;
2336 
2337 	cs_base = get_segment_base(regs->cs);
2338 	ss_base = get_segment_base(regs->ss);
2339 
2340 	fp = compat_ptr(ss_base + regs->bp);
2341 	pagefault_disable();
2342 	while (entry->nr < entry->max_stack) {
2343 		unsigned long bytes;
2344 		frame.next_frame     = 0;
2345 		frame.return_address = 0;
2346 
2347 		if (!access_ok(VERIFY_READ, fp, 8))
2348 			break;
2349 
2350 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2351 		if (bytes != 0)
2352 			break;
2353 		bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2354 		if (bytes != 0)
2355 			break;
2356 
2357 		if (!valid_user_frame(fp, sizeof(frame)))
2358 			break;
2359 
2360 		perf_callchain_store(entry, cs_base + frame.return_address);
2361 		fp = compat_ptr(ss_base + frame.next_frame);
2362 	}
2363 	pagefault_enable();
2364 	return 1;
2365 }
2366 #else
2367 static inline int
2368 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2369 {
2370     return 0;
2371 }
2372 #endif
2373 
2374 void
2375 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2376 {
2377 	struct stack_frame frame;
2378 	const unsigned long __user *fp;
2379 
2380 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2381 		/* TODO: We don't support guest os callchain now */
2382 		return;
2383 	}
2384 
2385 	/*
2386 	 * We don't know what to do with VM86 stacks.. ignore them for now.
2387 	 */
2388 	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2389 		return;
2390 
2391 	fp = (unsigned long __user *)regs->bp;
2392 
2393 	perf_callchain_store(entry, regs->ip);
2394 
2395 	if (!current->mm)
2396 		return;
2397 
2398 	if (perf_callchain_user32(regs, entry))
2399 		return;
2400 
2401 	pagefault_disable();
2402 	while (entry->nr < entry->max_stack) {
2403 		unsigned long bytes;
2404 
2405 		frame.next_frame	     = NULL;
2406 		frame.return_address = 0;
2407 
2408 		if (!access_ok(VERIFY_READ, fp, sizeof(*fp) * 2))
2409 			break;
2410 
2411 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2412 		if (bytes != 0)
2413 			break;
2414 		bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2415 		if (bytes != 0)
2416 			break;
2417 
2418 		if (!valid_user_frame(fp, sizeof(frame)))
2419 			break;
2420 
2421 		perf_callchain_store(entry, frame.return_address);
2422 		fp = (void __user *)frame.next_frame;
2423 	}
2424 	pagefault_enable();
2425 }
2426 
2427 /*
2428  * Deal with code segment offsets for the various execution modes:
2429  *
2430  *   VM86 - the good olde 16 bit days, where the linear address is
2431  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2432  *
2433  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2434  *          to figure out what the 32bit base address is.
2435  *
2436  *    X32 - has TIF_X32 set, but is running in x86_64
2437  *
2438  * X86_64 - CS,DS,SS,ES are all zero based.
2439  */
2440 static unsigned long code_segment_base(struct pt_regs *regs)
2441 {
2442 	/*
2443 	 * For IA32 we look at the GDT/LDT segment base to convert the
2444 	 * effective IP to a linear address.
2445 	 */
2446 
2447 #ifdef CONFIG_X86_32
2448 	/*
2449 	 * If we are in VM86 mode, add the segment offset to convert to a
2450 	 * linear address.
2451 	 */
2452 	if (regs->flags & X86_VM_MASK)
2453 		return 0x10 * regs->cs;
2454 
2455 	if (user_mode(regs) && regs->cs != __USER_CS)
2456 		return get_segment_base(regs->cs);
2457 #else
2458 	if (user_mode(regs) && !user_64bit_mode(regs) &&
2459 	    regs->cs != __USER32_CS)
2460 		return get_segment_base(regs->cs);
2461 #endif
2462 	return 0;
2463 }
2464 
2465 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2466 {
2467 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2468 		return perf_guest_cbs->get_guest_ip();
2469 
2470 	return regs->ip + code_segment_base(regs);
2471 }
2472 
2473 unsigned long perf_misc_flags(struct pt_regs *regs)
2474 {
2475 	int misc = 0;
2476 
2477 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2478 		if (perf_guest_cbs->is_user_mode())
2479 			misc |= PERF_RECORD_MISC_GUEST_USER;
2480 		else
2481 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2482 	} else {
2483 		if (user_mode(regs))
2484 			misc |= PERF_RECORD_MISC_USER;
2485 		else
2486 			misc |= PERF_RECORD_MISC_KERNEL;
2487 	}
2488 
2489 	if (regs->flags & PERF_EFLAGS_EXACT)
2490 		misc |= PERF_RECORD_MISC_EXACT_IP;
2491 
2492 	return misc;
2493 }
2494 
2495 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2496 {
2497 	cap->version		= x86_pmu.version;
2498 	cap->num_counters_gp	= x86_pmu.num_counters;
2499 	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2500 	cap->bit_width_gp	= x86_pmu.cntval_bits;
2501 	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2502 	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2503 	cap->events_mask_len	= x86_pmu.events_mask_len;
2504 }
2505 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
2506