xref: /openbmc/linux/arch/x86/events/core.c (revision 93f5715e)
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/mm.h>
24 #include <linux/sched/clock.h>
25 #include <linux/uaccess.h>
26 #include <linux/slab.h>
27 #include <linux/cpu.h>
28 #include <linux/bitops.h>
29 #include <linux/device.h>
30 #include <linux/nospec.h>
31 
32 #include <asm/apic.h>
33 #include <asm/stacktrace.h>
34 #include <asm/nmi.h>
35 #include <asm/smp.h>
36 #include <asm/alternative.h>
37 #include <asm/mmu_context.h>
38 #include <asm/tlbflush.h>
39 #include <asm/timer.h>
40 #include <asm/desc.h>
41 #include <asm/ldt.h>
42 #include <asm/unwind.h>
43 
44 #include "perf_event.h"
45 
46 struct x86_pmu x86_pmu __read_mostly;
47 
48 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
49 	.enabled = 1,
50 };
51 
52 DEFINE_STATIC_KEY_FALSE(rdpmc_always_available_key);
53 
54 u64 __read_mostly hw_cache_event_ids
55 				[PERF_COUNT_HW_CACHE_MAX]
56 				[PERF_COUNT_HW_CACHE_OP_MAX]
57 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
58 u64 __read_mostly hw_cache_extra_regs
59 				[PERF_COUNT_HW_CACHE_MAX]
60 				[PERF_COUNT_HW_CACHE_OP_MAX]
61 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
62 
63 /*
64  * Propagate event elapsed time into the generic event.
65  * Can only be executed on the CPU where the event is active.
66  * Returns the delta events processed.
67  */
68 u64 x86_perf_event_update(struct perf_event *event)
69 {
70 	struct hw_perf_event *hwc = &event->hw;
71 	int shift = 64 - x86_pmu.cntval_bits;
72 	u64 prev_raw_count, new_raw_count;
73 	int idx = hwc->idx;
74 	u64 delta;
75 
76 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
77 		return 0;
78 
79 	/*
80 	 * Careful: an NMI might modify the previous event value.
81 	 *
82 	 * Our tactic to handle this is to first atomically read and
83 	 * exchange a new raw count - then add that new-prev delta
84 	 * count to the generic event atomically:
85 	 */
86 again:
87 	prev_raw_count = local64_read(&hwc->prev_count);
88 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
89 
90 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
91 					new_raw_count) != prev_raw_count)
92 		goto again;
93 
94 	/*
95 	 * Now we have the new raw value and have updated the prev
96 	 * timestamp already. We can now calculate the elapsed delta
97 	 * (event-)time and add that to the generic event.
98 	 *
99 	 * Careful, not all hw sign-extends above the physical width
100 	 * of the count.
101 	 */
102 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
103 	delta >>= shift;
104 
105 	local64_add(delta, &event->count);
106 	local64_sub(delta, &hwc->period_left);
107 
108 	return new_raw_count;
109 }
110 
111 /*
112  * Find and validate any extra registers to set up.
113  */
114 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
115 {
116 	struct hw_perf_event_extra *reg;
117 	struct extra_reg *er;
118 
119 	reg = &event->hw.extra_reg;
120 
121 	if (!x86_pmu.extra_regs)
122 		return 0;
123 
124 	for (er = x86_pmu.extra_regs; er->msr; er++) {
125 		if (er->event != (config & er->config_mask))
126 			continue;
127 		if (event->attr.config1 & ~er->valid_mask)
128 			return -EINVAL;
129 		/* Check if the extra msrs can be safely accessed*/
130 		if (!er->extra_msr_access)
131 			return -ENXIO;
132 
133 		reg->idx = er->idx;
134 		reg->config = event->attr.config1;
135 		reg->reg = er->msr;
136 		break;
137 	}
138 	return 0;
139 }
140 
141 static atomic_t active_events;
142 static atomic_t pmc_refcount;
143 static DEFINE_MUTEX(pmc_reserve_mutex);
144 
145 #ifdef CONFIG_X86_LOCAL_APIC
146 
147 static bool reserve_pmc_hardware(void)
148 {
149 	int i;
150 
151 	for (i = 0; i < x86_pmu.num_counters; i++) {
152 		if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
153 			goto perfctr_fail;
154 	}
155 
156 	for (i = 0; i < x86_pmu.num_counters; i++) {
157 		if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
158 			goto eventsel_fail;
159 	}
160 
161 	return true;
162 
163 eventsel_fail:
164 	for (i--; i >= 0; i--)
165 		release_evntsel_nmi(x86_pmu_config_addr(i));
166 
167 	i = x86_pmu.num_counters;
168 
169 perfctr_fail:
170 	for (i--; i >= 0; i--)
171 		release_perfctr_nmi(x86_pmu_event_addr(i));
172 
173 	return false;
174 }
175 
176 static void release_pmc_hardware(void)
177 {
178 	int i;
179 
180 	for (i = 0; i < x86_pmu.num_counters; i++) {
181 		release_perfctr_nmi(x86_pmu_event_addr(i));
182 		release_evntsel_nmi(x86_pmu_config_addr(i));
183 	}
184 }
185 
186 #else
187 
188 static bool reserve_pmc_hardware(void) { return true; }
189 static void release_pmc_hardware(void) {}
190 
191 #endif
192 
193 static bool check_hw_exists(void)
194 {
195 	u64 val, val_fail = -1, val_new= ~0;
196 	int i, reg, reg_fail = -1, ret = 0;
197 	int bios_fail = 0;
198 	int reg_safe = -1;
199 
200 	/*
201 	 * Check to see if the BIOS enabled any of the counters, if so
202 	 * complain and bail.
203 	 */
204 	for (i = 0; i < x86_pmu.num_counters; i++) {
205 		reg = x86_pmu_config_addr(i);
206 		ret = rdmsrl_safe(reg, &val);
207 		if (ret)
208 			goto msr_fail;
209 		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
210 			bios_fail = 1;
211 			val_fail = val;
212 			reg_fail = reg;
213 		} else {
214 			reg_safe = i;
215 		}
216 	}
217 
218 	if (x86_pmu.num_counters_fixed) {
219 		reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
220 		ret = rdmsrl_safe(reg, &val);
221 		if (ret)
222 			goto msr_fail;
223 		for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
224 			if (val & (0x03 << i*4)) {
225 				bios_fail = 1;
226 				val_fail = val;
227 				reg_fail = reg;
228 			}
229 		}
230 	}
231 
232 	/*
233 	 * If all the counters are enabled, the below test will always
234 	 * fail.  The tools will also become useless in this scenario.
235 	 * Just fail and disable the hardware counters.
236 	 */
237 
238 	if (reg_safe == -1) {
239 		reg = reg_safe;
240 		goto msr_fail;
241 	}
242 
243 	/*
244 	 * Read the current value, change it and read it back to see if it
245 	 * matches, this is needed to detect certain hardware emulators
246 	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
247 	 */
248 	reg = x86_pmu_event_addr(reg_safe);
249 	if (rdmsrl_safe(reg, &val))
250 		goto msr_fail;
251 	val ^= 0xffffUL;
252 	ret = wrmsrl_safe(reg, val);
253 	ret |= rdmsrl_safe(reg, &val_new);
254 	if (ret || val != val_new)
255 		goto msr_fail;
256 
257 	/*
258 	 * We still allow the PMU driver to operate:
259 	 */
260 	if (bios_fail) {
261 		pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
262 		pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
263 			      reg_fail, val_fail);
264 	}
265 
266 	return true;
267 
268 msr_fail:
269 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
270 		pr_cont("PMU not available due to virtualization, using software events only.\n");
271 	} else {
272 		pr_cont("Broken PMU hardware detected, using software events only.\n");
273 		pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
274 		       reg, val_new);
275 	}
276 
277 	return false;
278 }
279 
280 static void hw_perf_event_destroy(struct perf_event *event)
281 {
282 	x86_release_hardware();
283 	atomic_dec(&active_events);
284 }
285 
286 void hw_perf_lbr_event_destroy(struct perf_event *event)
287 {
288 	hw_perf_event_destroy(event);
289 
290 	/* undo the lbr/bts event accounting */
291 	x86_del_exclusive(x86_lbr_exclusive_lbr);
292 }
293 
294 static inline int x86_pmu_initialized(void)
295 {
296 	return x86_pmu.handle_irq != NULL;
297 }
298 
299 static inline int
300 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
301 {
302 	struct perf_event_attr *attr = &event->attr;
303 	unsigned int cache_type, cache_op, cache_result;
304 	u64 config, val;
305 
306 	config = attr->config;
307 
308 	cache_type = (config >> 0) & 0xff;
309 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
310 		return -EINVAL;
311 	cache_type = array_index_nospec(cache_type, PERF_COUNT_HW_CACHE_MAX);
312 
313 	cache_op = (config >>  8) & 0xff;
314 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
315 		return -EINVAL;
316 	cache_op = array_index_nospec(cache_op, PERF_COUNT_HW_CACHE_OP_MAX);
317 
318 	cache_result = (config >> 16) & 0xff;
319 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
320 		return -EINVAL;
321 	cache_result = array_index_nospec(cache_result, PERF_COUNT_HW_CACHE_RESULT_MAX);
322 
323 	val = hw_cache_event_ids[cache_type][cache_op][cache_result];
324 
325 	if (val == 0)
326 		return -ENOENT;
327 
328 	if (val == -1)
329 		return -EINVAL;
330 
331 	hwc->config |= val;
332 	attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
333 	return x86_pmu_extra_regs(val, event);
334 }
335 
336 int x86_reserve_hardware(void)
337 {
338 	int err = 0;
339 
340 	if (!atomic_inc_not_zero(&pmc_refcount)) {
341 		mutex_lock(&pmc_reserve_mutex);
342 		if (atomic_read(&pmc_refcount) == 0) {
343 			if (!reserve_pmc_hardware())
344 				err = -EBUSY;
345 			else
346 				reserve_ds_buffers();
347 		}
348 		if (!err)
349 			atomic_inc(&pmc_refcount);
350 		mutex_unlock(&pmc_reserve_mutex);
351 	}
352 
353 	return err;
354 }
355 
356 void x86_release_hardware(void)
357 {
358 	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
359 		release_pmc_hardware();
360 		release_ds_buffers();
361 		mutex_unlock(&pmc_reserve_mutex);
362 	}
363 }
364 
365 /*
366  * Check if we can create event of a certain type (that no conflicting events
367  * are present).
368  */
369 int x86_add_exclusive(unsigned int what)
370 {
371 	int i;
372 
373 	/*
374 	 * When lbr_pt_coexist we allow PT to coexist with either LBR or BTS.
375 	 * LBR and BTS are still mutually exclusive.
376 	 */
377 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
378 		return 0;
379 
380 	if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
381 		mutex_lock(&pmc_reserve_mutex);
382 		for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
383 			if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
384 				goto fail_unlock;
385 		}
386 		atomic_inc(&x86_pmu.lbr_exclusive[what]);
387 		mutex_unlock(&pmc_reserve_mutex);
388 	}
389 
390 	atomic_inc(&active_events);
391 	return 0;
392 
393 fail_unlock:
394 	mutex_unlock(&pmc_reserve_mutex);
395 	return -EBUSY;
396 }
397 
398 void x86_del_exclusive(unsigned int what)
399 {
400 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
401 		return;
402 
403 	atomic_dec(&x86_pmu.lbr_exclusive[what]);
404 	atomic_dec(&active_events);
405 }
406 
407 int x86_setup_perfctr(struct perf_event *event)
408 {
409 	struct perf_event_attr *attr = &event->attr;
410 	struct hw_perf_event *hwc = &event->hw;
411 	u64 config;
412 
413 	if (!is_sampling_event(event)) {
414 		hwc->sample_period = x86_pmu.max_period;
415 		hwc->last_period = hwc->sample_period;
416 		local64_set(&hwc->period_left, hwc->sample_period);
417 	}
418 
419 	if (attr->type == PERF_TYPE_RAW)
420 		return x86_pmu_extra_regs(event->attr.config, event);
421 
422 	if (attr->type == PERF_TYPE_HW_CACHE)
423 		return set_ext_hw_attr(hwc, event);
424 
425 	if (attr->config >= x86_pmu.max_events)
426 		return -EINVAL;
427 
428 	attr->config = array_index_nospec((unsigned long)attr->config, x86_pmu.max_events);
429 
430 	/*
431 	 * The generic map:
432 	 */
433 	config = x86_pmu.event_map(attr->config);
434 
435 	if (config == 0)
436 		return -ENOENT;
437 
438 	if (config == -1LL)
439 		return -EINVAL;
440 
441 	/*
442 	 * Branch tracing:
443 	 */
444 	if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
445 	    !attr->freq && hwc->sample_period == 1) {
446 		/* BTS is not supported by this architecture. */
447 		if (!x86_pmu.bts_active)
448 			return -EOPNOTSUPP;
449 
450 		/* BTS is currently only allowed for user-mode. */
451 		if (!attr->exclude_kernel)
452 			return -EOPNOTSUPP;
453 
454 		/* disallow bts if conflicting events are present */
455 		if (x86_add_exclusive(x86_lbr_exclusive_lbr))
456 			return -EBUSY;
457 
458 		event->destroy = hw_perf_lbr_event_destroy;
459 	}
460 
461 	hwc->config |= config;
462 
463 	return 0;
464 }
465 
466 /*
467  * check that branch_sample_type is compatible with
468  * settings needed for precise_ip > 1 which implies
469  * using the LBR to capture ALL taken branches at the
470  * priv levels of the measurement
471  */
472 static inline int precise_br_compat(struct perf_event *event)
473 {
474 	u64 m = event->attr.branch_sample_type;
475 	u64 b = 0;
476 
477 	/* must capture all branches */
478 	if (!(m & PERF_SAMPLE_BRANCH_ANY))
479 		return 0;
480 
481 	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
482 
483 	if (!event->attr.exclude_user)
484 		b |= PERF_SAMPLE_BRANCH_USER;
485 
486 	if (!event->attr.exclude_kernel)
487 		b |= PERF_SAMPLE_BRANCH_KERNEL;
488 
489 	/*
490 	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
491 	 */
492 
493 	return m == b;
494 }
495 
496 int x86_pmu_max_precise(void)
497 {
498 	int precise = 0;
499 
500 	/* Support for constant skid */
501 	if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
502 		precise++;
503 
504 		/* Support for IP fixup */
505 		if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
506 			precise++;
507 
508 		if (x86_pmu.pebs_prec_dist)
509 			precise++;
510 	}
511 	return precise;
512 }
513 
514 int x86_pmu_hw_config(struct perf_event *event)
515 {
516 	if (event->attr.precise_ip) {
517 		int precise = x86_pmu_max_precise();
518 
519 		if (event->attr.precise_ip > precise)
520 			return -EOPNOTSUPP;
521 
522 		/* There's no sense in having PEBS for non sampling events: */
523 		if (!is_sampling_event(event))
524 			return -EINVAL;
525 	}
526 	/*
527 	 * check that PEBS LBR correction does not conflict with
528 	 * whatever the user is asking with attr->branch_sample_type
529 	 */
530 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
531 		u64 *br_type = &event->attr.branch_sample_type;
532 
533 		if (has_branch_stack(event)) {
534 			if (!precise_br_compat(event))
535 				return -EOPNOTSUPP;
536 
537 			/* branch_sample_type is compatible */
538 
539 		} else {
540 			/*
541 			 * user did not specify  branch_sample_type
542 			 *
543 			 * For PEBS fixups, we capture all
544 			 * the branches at the priv level of the
545 			 * event.
546 			 */
547 			*br_type = PERF_SAMPLE_BRANCH_ANY;
548 
549 			if (!event->attr.exclude_user)
550 				*br_type |= PERF_SAMPLE_BRANCH_USER;
551 
552 			if (!event->attr.exclude_kernel)
553 				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
554 		}
555 	}
556 
557 	if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
558 		event->attach_state |= PERF_ATTACH_TASK_DATA;
559 
560 	/*
561 	 * Generate PMC IRQs:
562 	 * (keep 'enabled' bit clear for now)
563 	 */
564 	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
565 
566 	/*
567 	 * Count user and OS events unless requested not to
568 	 */
569 	if (!event->attr.exclude_user)
570 		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
571 	if (!event->attr.exclude_kernel)
572 		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
573 
574 	if (event->attr.type == PERF_TYPE_RAW)
575 		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
576 
577 	if (event->attr.sample_period && x86_pmu.limit_period) {
578 		if (x86_pmu.limit_period(event, event->attr.sample_period) >
579 				event->attr.sample_period)
580 			return -EINVAL;
581 	}
582 
583 	return x86_setup_perfctr(event);
584 }
585 
586 /*
587  * Setup the hardware configuration for a given attr_type
588  */
589 static int __x86_pmu_event_init(struct perf_event *event)
590 {
591 	int err;
592 
593 	if (!x86_pmu_initialized())
594 		return -ENODEV;
595 
596 	err = x86_reserve_hardware();
597 	if (err)
598 		return err;
599 
600 	atomic_inc(&active_events);
601 	event->destroy = hw_perf_event_destroy;
602 
603 	event->hw.idx = -1;
604 	event->hw.last_cpu = -1;
605 	event->hw.last_tag = ~0ULL;
606 
607 	/* mark unused */
608 	event->hw.extra_reg.idx = EXTRA_REG_NONE;
609 	event->hw.branch_reg.idx = EXTRA_REG_NONE;
610 
611 	return x86_pmu.hw_config(event);
612 }
613 
614 void x86_pmu_disable_all(void)
615 {
616 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
617 	int idx;
618 
619 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
620 		u64 val;
621 
622 		if (!test_bit(idx, cpuc->active_mask))
623 			continue;
624 		rdmsrl(x86_pmu_config_addr(idx), val);
625 		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
626 			continue;
627 		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
628 		wrmsrl(x86_pmu_config_addr(idx), val);
629 	}
630 }
631 
632 /*
633  * There may be PMI landing after enabled=0. The PMI hitting could be before or
634  * after disable_all.
635  *
636  * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
637  * It will not be re-enabled in the NMI handler again, because enabled=0. After
638  * handling the NMI, disable_all will be called, which will not change the
639  * state either. If PMI hits after disable_all, the PMU is already disabled
640  * before entering NMI handler. The NMI handler will not change the state
641  * either.
642  *
643  * So either situation is harmless.
644  */
645 static void x86_pmu_disable(struct pmu *pmu)
646 {
647 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
648 
649 	if (!x86_pmu_initialized())
650 		return;
651 
652 	if (!cpuc->enabled)
653 		return;
654 
655 	cpuc->n_added = 0;
656 	cpuc->enabled = 0;
657 	barrier();
658 
659 	x86_pmu.disable_all();
660 }
661 
662 void x86_pmu_enable_all(int added)
663 {
664 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
665 	int idx;
666 
667 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
668 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
669 
670 		if (!test_bit(idx, cpuc->active_mask))
671 			continue;
672 
673 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
674 	}
675 }
676 
677 static struct pmu pmu;
678 
679 static inline int is_x86_event(struct perf_event *event)
680 {
681 	return event->pmu == &pmu;
682 }
683 
684 /*
685  * Event scheduler state:
686  *
687  * Assign events iterating over all events and counters, beginning
688  * with events with least weights first. Keep the current iterator
689  * state in struct sched_state.
690  */
691 struct sched_state {
692 	int	weight;
693 	int	event;		/* event index */
694 	int	counter;	/* counter index */
695 	int	unassigned;	/* number of events to be assigned left */
696 	int	nr_gp;		/* number of GP counters used */
697 	unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
698 };
699 
700 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
701 #define	SCHED_STATES_MAX	2
702 
703 struct perf_sched {
704 	int			max_weight;
705 	int			max_events;
706 	int			max_gp;
707 	int			saved_states;
708 	struct event_constraint	**constraints;
709 	struct sched_state	state;
710 	struct sched_state	saved[SCHED_STATES_MAX];
711 };
712 
713 /*
714  * Initialize interator that runs through all events and counters.
715  */
716 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
717 			    int num, int wmin, int wmax, int gpmax)
718 {
719 	int idx;
720 
721 	memset(sched, 0, sizeof(*sched));
722 	sched->max_events	= num;
723 	sched->max_weight	= wmax;
724 	sched->max_gp		= gpmax;
725 	sched->constraints	= constraints;
726 
727 	for (idx = 0; idx < num; idx++) {
728 		if (constraints[idx]->weight == wmin)
729 			break;
730 	}
731 
732 	sched->state.event	= idx;		/* start with min weight */
733 	sched->state.weight	= wmin;
734 	sched->state.unassigned	= num;
735 }
736 
737 static void perf_sched_save_state(struct perf_sched *sched)
738 {
739 	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
740 		return;
741 
742 	sched->saved[sched->saved_states] = sched->state;
743 	sched->saved_states++;
744 }
745 
746 static bool perf_sched_restore_state(struct perf_sched *sched)
747 {
748 	if (!sched->saved_states)
749 		return false;
750 
751 	sched->saved_states--;
752 	sched->state = sched->saved[sched->saved_states];
753 
754 	/* continue with next counter: */
755 	clear_bit(sched->state.counter++, sched->state.used);
756 
757 	return true;
758 }
759 
760 /*
761  * Select a counter for the current event to schedule. Return true on
762  * success.
763  */
764 static bool __perf_sched_find_counter(struct perf_sched *sched)
765 {
766 	struct event_constraint *c;
767 	int idx;
768 
769 	if (!sched->state.unassigned)
770 		return false;
771 
772 	if (sched->state.event >= sched->max_events)
773 		return false;
774 
775 	c = sched->constraints[sched->state.event];
776 	/* Prefer fixed purpose counters */
777 	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
778 		idx = INTEL_PMC_IDX_FIXED;
779 		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
780 			if (!__test_and_set_bit(idx, sched->state.used))
781 				goto done;
782 		}
783 	}
784 
785 	/* Grab the first unused counter starting with idx */
786 	idx = sched->state.counter;
787 	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
788 		if (!__test_and_set_bit(idx, sched->state.used)) {
789 			if (sched->state.nr_gp++ >= sched->max_gp)
790 				return false;
791 
792 			goto done;
793 		}
794 	}
795 
796 	return false;
797 
798 done:
799 	sched->state.counter = idx;
800 
801 	if (c->overlap)
802 		perf_sched_save_state(sched);
803 
804 	return true;
805 }
806 
807 static bool perf_sched_find_counter(struct perf_sched *sched)
808 {
809 	while (!__perf_sched_find_counter(sched)) {
810 		if (!perf_sched_restore_state(sched))
811 			return false;
812 	}
813 
814 	return true;
815 }
816 
817 /*
818  * Go through all unassigned events and find the next one to schedule.
819  * Take events with the least weight first. Return true on success.
820  */
821 static bool perf_sched_next_event(struct perf_sched *sched)
822 {
823 	struct event_constraint *c;
824 
825 	if (!sched->state.unassigned || !--sched->state.unassigned)
826 		return false;
827 
828 	do {
829 		/* next event */
830 		sched->state.event++;
831 		if (sched->state.event >= sched->max_events) {
832 			/* next weight */
833 			sched->state.event = 0;
834 			sched->state.weight++;
835 			if (sched->state.weight > sched->max_weight)
836 				return false;
837 		}
838 		c = sched->constraints[sched->state.event];
839 	} while (c->weight != sched->state.weight);
840 
841 	sched->state.counter = 0;	/* start with first counter */
842 
843 	return true;
844 }
845 
846 /*
847  * Assign a counter for each event.
848  */
849 int perf_assign_events(struct event_constraint **constraints, int n,
850 			int wmin, int wmax, int gpmax, int *assign)
851 {
852 	struct perf_sched sched;
853 
854 	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
855 
856 	do {
857 		if (!perf_sched_find_counter(&sched))
858 			break;	/* failed */
859 		if (assign)
860 			assign[sched.state.event] = sched.state.counter;
861 	} while (perf_sched_next_event(&sched));
862 
863 	return sched.state.unassigned;
864 }
865 EXPORT_SYMBOL_GPL(perf_assign_events);
866 
867 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
868 {
869 	struct event_constraint *c;
870 	unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
871 	struct perf_event *e;
872 	int i, wmin, wmax, unsched = 0;
873 	struct hw_perf_event *hwc;
874 
875 	bitmap_zero(used_mask, X86_PMC_IDX_MAX);
876 
877 	if (x86_pmu.start_scheduling)
878 		x86_pmu.start_scheduling(cpuc);
879 
880 	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
881 		cpuc->event_constraint[i] = NULL;
882 		c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
883 		cpuc->event_constraint[i] = c;
884 
885 		wmin = min(wmin, c->weight);
886 		wmax = max(wmax, c->weight);
887 	}
888 
889 	/*
890 	 * fastpath, try to reuse previous register
891 	 */
892 	for (i = 0; i < n; i++) {
893 		hwc = &cpuc->event_list[i]->hw;
894 		c = cpuc->event_constraint[i];
895 
896 		/* never assigned */
897 		if (hwc->idx == -1)
898 			break;
899 
900 		/* constraint still honored */
901 		if (!test_bit(hwc->idx, c->idxmsk))
902 			break;
903 
904 		/* not already used */
905 		if (test_bit(hwc->idx, used_mask))
906 			break;
907 
908 		__set_bit(hwc->idx, used_mask);
909 		if (assign)
910 			assign[i] = hwc->idx;
911 	}
912 
913 	/* slow path */
914 	if (i != n) {
915 		int gpmax = x86_pmu.num_counters;
916 
917 		/*
918 		 * Do not allow scheduling of more than half the available
919 		 * generic counters.
920 		 *
921 		 * This helps avoid counter starvation of sibling thread by
922 		 * ensuring at most half the counters cannot be in exclusive
923 		 * mode. There is no designated counters for the limits. Any
924 		 * N/2 counters can be used. This helps with events with
925 		 * specific counter constraints.
926 		 */
927 		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
928 		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
929 			gpmax /= 2;
930 
931 		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
932 					     wmax, gpmax, assign);
933 	}
934 
935 	/*
936 	 * In case of success (unsched = 0), mark events as committed,
937 	 * so we do not put_constraint() in case new events are added
938 	 * and fail to be scheduled
939 	 *
940 	 * We invoke the lower level commit callback to lock the resource
941 	 *
942 	 * We do not need to do all of this in case we are called to
943 	 * validate an event group (assign == NULL)
944 	 */
945 	if (!unsched && assign) {
946 		for (i = 0; i < n; i++) {
947 			e = cpuc->event_list[i];
948 			e->hw.flags |= PERF_X86_EVENT_COMMITTED;
949 			if (x86_pmu.commit_scheduling)
950 				x86_pmu.commit_scheduling(cpuc, i, assign[i]);
951 		}
952 	} else {
953 		for (i = 0; i < n; i++) {
954 			e = cpuc->event_list[i];
955 			/*
956 			 * do not put_constraint() on comitted events,
957 			 * because they are good to go
958 			 */
959 			if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
960 				continue;
961 
962 			/*
963 			 * release events that failed scheduling
964 			 */
965 			if (x86_pmu.put_event_constraints)
966 				x86_pmu.put_event_constraints(cpuc, e);
967 		}
968 	}
969 
970 	if (x86_pmu.stop_scheduling)
971 		x86_pmu.stop_scheduling(cpuc);
972 
973 	return unsched ? -EINVAL : 0;
974 }
975 
976 /*
977  * dogrp: true if must collect siblings events (group)
978  * returns total number of events and error code
979  */
980 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
981 {
982 	struct perf_event *event;
983 	int n, max_count;
984 
985 	max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
986 
987 	/* current number of events already accepted */
988 	n = cpuc->n_events;
989 
990 	if (is_x86_event(leader)) {
991 		if (n >= max_count)
992 			return -EINVAL;
993 		cpuc->event_list[n] = leader;
994 		n++;
995 	}
996 	if (!dogrp)
997 		return n;
998 
999 	for_each_sibling_event(event, leader) {
1000 		if (!is_x86_event(event) ||
1001 		    event->state <= PERF_EVENT_STATE_OFF)
1002 			continue;
1003 
1004 		if (n >= max_count)
1005 			return -EINVAL;
1006 
1007 		cpuc->event_list[n] = event;
1008 		n++;
1009 	}
1010 	return n;
1011 }
1012 
1013 static inline void x86_assign_hw_event(struct perf_event *event,
1014 				struct cpu_hw_events *cpuc, int i)
1015 {
1016 	struct hw_perf_event *hwc = &event->hw;
1017 
1018 	hwc->idx = cpuc->assign[i];
1019 	hwc->last_cpu = smp_processor_id();
1020 	hwc->last_tag = ++cpuc->tags[i];
1021 
1022 	if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
1023 		hwc->config_base = 0;
1024 		hwc->event_base	= 0;
1025 	} else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1026 		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1027 		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1028 		hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1029 	} else {
1030 		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1031 		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1032 		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1033 	}
1034 }
1035 
1036 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1037 					struct cpu_hw_events *cpuc,
1038 					int i)
1039 {
1040 	return hwc->idx == cpuc->assign[i] &&
1041 		hwc->last_cpu == smp_processor_id() &&
1042 		hwc->last_tag == cpuc->tags[i];
1043 }
1044 
1045 static void x86_pmu_start(struct perf_event *event, int flags);
1046 
1047 static void x86_pmu_enable(struct pmu *pmu)
1048 {
1049 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1050 	struct perf_event *event;
1051 	struct hw_perf_event *hwc;
1052 	int i, added = cpuc->n_added;
1053 
1054 	if (!x86_pmu_initialized())
1055 		return;
1056 
1057 	if (cpuc->enabled)
1058 		return;
1059 
1060 	if (cpuc->n_added) {
1061 		int n_running = cpuc->n_events - cpuc->n_added;
1062 		/*
1063 		 * apply assignment obtained either from
1064 		 * hw_perf_group_sched_in() or x86_pmu_enable()
1065 		 *
1066 		 * step1: save events moving to new counters
1067 		 */
1068 		for (i = 0; i < n_running; i++) {
1069 			event = cpuc->event_list[i];
1070 			hwc = &event->hw;
1071 
1072 			/*
1073 			 * we can avoid reprogramming counter if:
1074 			 * - assigned same counter as last time
1075 			 * - running on same CPU as last time
1076 			 * - no other event has used the counter since
1077 			 */
1078 			if (hwc->idx == -1 ||
1079 			    match_prev_assignment(hwc, cpuc, i))
1080 				continue;
1081 
1082 			/*
1083 			 * Ensure we don't accidentally enable a stopped
1084 			 * counter simply because we rescheduled.
1085 			 */
1086 			if (hwc->state & PERF_HES_STOPPED)
1087 				hwc->state |= PERF_HES_ARCH;
1088 
1089 			x86_pmu_stop(event, PERF_EF_UPDATE);
1090 		}
1091 
1092 		/*
1093 		 * step2: reprogram moved events into new counters
1094 		 */
1095 		for (i = 0; i < cpuc->n_events; i++) {
1096 			event = cpuc->event_list[i];
1097 			hwc = &event->hw;
1098 
1099 			if (!match_prev_assignment(hwc, cpuc, i))
1100 				x86_assign_hw_event(event, cpuc, i);
1101 			else if (i < n_running)
1102 				continue;
1103 
1104 			if (hwc->state & PERF_HES_ARCH)
1105 				continue;
1106 
1107 			x86_pmu_start(event, PERF_EF_RELOAD);
1108 		}
1109 		cpuc->n_added = 0;
1110 		perf_events_lapic_init();
1111 	}
1112 
1113 	cpuc->enabled = 1;
1114 	barrier();
1115 
1116 	x86_pmu.enable_all(added);
1117 }
1118 
1119 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1120 
1121 /*
1122  * Set the next IRQ period, based on the hwc->period_left value.
1123  * To be called with the event disabled in hw:
1124  */
1125 int x86_perf_event_set_period(struct perf_event *event)
1126 {
1127 	struct hw_perf_event *hwc = &event->hw;
1128 	s64 left = local64_read(&hwc->period_left);
1129 	s64 period = hwc->sample_period;
1130 	int ret = 0, idx = hwc->idx;
1131 
1132 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
1133 		return 0;
1134 
1135 	/*
1136 	 * If we are way outside a reasonable range then just skip forward:
1137 	 */
1138 	if (unlikely(left <= -period)) {
1139 		left = period;
1140 		local64_set(&hwc->period_left, left);
1141 		hwc->last_period = period;
1142 		ret = 1;
1143 	}
1144 
1145 	if (unlikely(left <= 0)) {
1146 		left += period;
1147 		local64_set(&hwc->period_left, left);
1148 		hwc->last_period = period;
1149 		ret = 1;
1150 	}
1151 	/*
1152 	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1153 	 */
1154 	if (unlikely(left < 2))
1155 		left = 2;
1156 
1157 	if (left > x86_pmu.max_period)
1158 		left = x86_pmu.max_period;
1159 
1160 	if (x86_pmu.limit_period)
1161 		left = x86_pmu.limit_period(event, left);
1162 
1163 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1164 
1165 	/*
1166 	 * The hw event starts counting from this event offset,
1167 	 * mark it to be able to extra future deltas:
1168 	 */
1169 	local64_set(&hwc->prev_count, (u64)-left);
1170 
1171 	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1172 
1173 	/*
1174 	 * Due to erratum on certan cpu we need
1175 	 * a second write to be sure the register
1176 	 * is updated properly
1177 	 */
1178 	if (x86_pmu.perfctr_second_write) {
1179 		wrmsrl(hwc->event_base,
1180 			(u64)(-left) & x86_pmu.cntval_mask);
1181 	}
1182 
1183 	perf_event_update_userpage(event);
1184 
1185 	return ret;
1186 }
1187 
1188 void x86_pmu_enable_event(struct perf_event *event)
1189 {
1190 	if (__this_cpu_read(cpu_hw_events.enabled))
1191 		__x86_pmu_enable_event(&event->hw,
1192 				       ARCH_PERFMON_EVENTSEL_ENABLE);
1193 }
1194 
1195 /*
1196  * Add a single event to the PMU.
1197  *
1198  * The event is added to the group of enabled events
1199  * but only if it can be scehduled with existing events.
1200  */
1201 static int x86_pmu_add(struct perf_event *event, int flags)
1202 {
1203 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1204 	struct hw_perf_event *hwc;
1205 	int assign[X86_PMC_IDX_MAX];
1206 	int n, n0, ret;
1207 
1208 	hwc = &event->hw;
1209 
1210 	n0 = cpuc->n_events;
1211 	ret = n = collect_events(cpuc, event, false);
1212 	if (ret < 0)
1213 		goto out;
1214 
1215 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1216 	if (!(flags & PERF_EF_START))
1217 		hwc->state |= PERF_HES_ARCH;
1218 
1219 	/*
1220 	 * If group events scheduling transaction was started,
1221 	 * skip the schedulability test here, it will be performed
1222 	 * at commit time (->commit_txn) as a whole.
1223 	 *
1224 	 * If commit fails, we'll call ->del() on all events
1225 	 * for which ->add() was called.
1226 	 */
1227 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1228 		goto done_collect;
1229 
1230 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1231 	if (ret)
1232 		goto out;
1233 	/*
1234 	 * copy new assignment, now we know it is possible
1235 	 * will be used by hw_perf_enable()
1236 	 */
1237 	memcpy(cpuc->assign, assign, n*sizeof(int));
1238 
1239 done_collect:
1240 	/*
1241 	 * Commit the collect_events() state. See x86_pmu_del() and
1242 	 * x86_pmu_*_txn().
1243 	 */
1244 	cpuc->n_events = n;
1245 	cpuc->n_added += n - n0;
1246 	cpuc->n_txn += n - n0;
1247 
1248 	if (x86_pmu.add) {
1249 		/*
1250 		 * This is before x86_pmu_enable() will call x86_pmu_start(),
1251 		 * so we enable LBRs before an event needs them etc..
1252 		 */
1253 		x86_pmu.add(event);
1254 	}
1255 
1256 	ret = 0;
1257 out:
1258 	return ret;
1259 }
1260 
1261 static void x86_pmu_start(struct perf_event *event, int flags)
1262 {
1263 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1264 	int idx = event->hw.idx;
1265 
1266 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1267 		return;
1268 
1269 	if (WARN_ON_ONCE(idx == -1))
1270 		return;
1271 
1272 	if (flags & PERF_EF_RELOAD) {
1273 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1274 		x86_perf_event_set_period(event);
1275 	}
1276 
1277 	event->hw.state = 0;
1278 
1279 	cpuc->events[idx] = event;
1280 	__set_bit(idx, cpuc->active_mask);
1281 	__set_bit(idx, cpuc->running);
1282 	x86_pmu.enable(event);
1283 	perf_event_update_userpage(event);
1284 }
1285 
1286 void perf_event_print_debug(void)
1287 {
1288 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1289 	u64 pebs, debugctl;
1290 	struct cpu_hw_events *cpuc;
1291 	unsigned long flags;
1292 	int cpu, idx;
1293 
1294 	if (!x86_pmu.num_counters)
1295 		return;
1296 
1297 	local_irq_save(flags);
1298 
1299 	cpu = smp_processor_id();
1300 	cpuc = &per_cpu(cpu_hw_events, cpu);
1301 
1302 	if (x86_pmu.version >= 2) {
1303 		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1304 		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1305 		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1306 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1307 
1308 		pr_info("\n");
1309 		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1310 		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1311 		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1312 		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1313 		if (x86_pmu.pebs_constraints) {
1314 			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1315 			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1316 		}
1317 		if (x86_pmu.lbr_nr) {
1318 			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1319 			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1320 		}
1321 	}
1322 	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1323 
1324 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1325 		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1326 		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1327 
1328 		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1329 
1330 		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1331 			cpu, idx, pmc_ctrl);
1332 		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1333 			cpu, idx, pmc_count);
1334 		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1335 			cpu, idx, prev_left);
1336 	}
1337 	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1338 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1339 
1340 		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1341 			cpu, idx, pmc_count);
1342 	}
1343 	local_irq_restore(flags);
1344 }
1345 
1346 void x86_pmu_stop(struct perf_event *event, int flags)
1347 {
1348 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1349 	struct hw_perf_event *hwc = &event->hw;
1350 
1351 	if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1352 		x86_pmu.disable(event);
1353 		cpuc->events[hwc->idx] = NULL;
1354 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1355 		hwc->state |= PERF_HES_STOPPED;
1356 	}
1357 
1358 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1359 		/*
1360 		 * Drain the remaining delta count out of a event
1361 		 * that we are disabling:
1362 		 */
1363 		x86_perf_event_update(event);
1364 		hwc->state |= PERF_HES_UPTODATE;
1365 	}
1366 }
1367 
1368 static void x86_pmu_del(struct perf_event *event, int flags)
1369 {
1370 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1371 	int i;
1372 
1373 	/*
1374 	 * event is descheduled
1375 	 */
1376 	event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1377 
1378 	/*
1379 	 * If we're called during a txn, we only need to undo x86_pmu.add.
1380 	 * The events never got scheduled and ->cancel_txn will truncate
1381 	 * the event_list.
1382 	 *
1383 	 * XXX assumes any ->del() called during a TXN will only be on
1384 	 * an event added during that same TXN.
1385 	 */
1386 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1387 		goto do_del;
1388 
1389 	/*
1390 	 * Not a TXN, therefore cleanup properly.
1391 	 */
1392 	x86_pmu_stop(event, PERF_EF_UPDATE);
1393 
1394 	for (i = 0; i < cpuc->n_events; i++) {
1395 		if (event == cpuc->event_list[i])
1396 			break;
1397 	}
1398 
1399 	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1400 		return;
1401 
1402 	/* If we have a newly added event; make sure to decrease n_added. */
1403 	if (i >= cpuc->n_events - cpuc->n_added)
1404 		--cpuc->n_added;
1405 
1406 	if (x86_pmu.put_event_constraints)
1407 		x86_pmu.put_event_constraints(cpuc, event);
1408 
1409 	/* Delete the array entry. */
1410 	while (++i < cpuc->n_events) {
1411 		cpuc->event_list[i-1] = cpuc->event_list[i];
1412 		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1413 	}
1414 	--cpuc->n_events;
1415 
1416 	perf_event_update_userpage(event);
1417 
1418 do_del:
1419 	if (x86_pmu.del) {
1420 		/*
1421 		 * This is after x86_pmu_stop(); so we disable LBRs after any
1422 		 * event can need them etc..
1423 		 */
1424 		x86_pmu.del(event);
1425 	}
1426 }
1427 
1428 int x86_pmu_handle_irq(struct pt_regs *regs)
1429 {
1430 	struct perf_sample_data data;
1431 	struct cpu_hw_events *cpuc;
1432 	struct perf_event *event;
1433 	int idx, handled = 0;
1434 	u64 val;
1435 
1436 	cpuc = this_cpu_ptr(&cpu_hw_events);
1437 
1438 	/*
1439 	 * Some chipsets need to unmask the LVTPC in a particular spot
1440 	 * inside the nmi handler.  As a result, the unmasking was pushed
1441 	 * into all the nmi handlers.
1442 	 *
1443 	 * This generic handler doesn't seem to have any issues where the
1444 	 * unmasking occurs so it was left at the top.
1445 	 */
1446 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1447 
1448 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1449 		if (!test_bit(idx, cpuc->active_mask)) {
1450 			/*
1451 			 * Though we deactivated the counter some cpus
1452 			 * might still deliver spurious interrupts still
1453 			 * in flight. Catch them:
1454 			 */
1455 			if (__test_and_clear_bit(idx, cpuc->running))
1456 				handled++;
1457 			continue;
1458 		}
1459 
1460 		event = cpuc->events[idx];
1461 
1462 		val = x86_perf_event_update(event);
1463 		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1464 			continue;
1465 
1466 		/*
1467 		 * event overflow
1468 		 */
1469 		handled++;
1470 		perf_sample_data_init(&data, 0, event->hw.last_period);
1471 
1472 		if (!x86_perf_event_set_period(event))
1473 			continue;
1474 
1475 		if (perf_event_overflow(event, &data, regs))
1476 			x86_pmu_stop(event, 0);
1477 	}
1478 
1479 	if (handled)
1480 		inc_irq_stat(apic_perf_irqs);
1481 
1482 	return handled;
1483 }
1484 
1485 void perf_events_lapic_init(void)
1486 {
1487 	if (!x86_pmu.apic || !x86_pmu_initialized())
1488 		return;
1489 
1490 	/*
1491 	 * Always use NMI for PMU
1492 	 */
1493 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1494 }
1495 
1496 static int
1497 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1498 {
1499 	u64 start_clock;
1500 	u64 finish_clock;
1501 	int ret;
1502 
1503 	/*
1504 	 * All PMUs/events that share this PMI handler should make sure to
1505 	 * increment active_events for their events.
1506 	 */
1507 	if (!atomic_read(&active_events))
1508 		return NMI_DONE;
1509 
1510 	start_clock = sched_clock();
1511 	ret = x86_pmu.handle_irq(regs);
1512 	finish_clock = sched_clock();
1513 
1514 	perf_sample_event_took(finish_clock - start_clock);
1515 
1516 	return ret;
1517 }
1518 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1519 
1520 struct event_constraint emptyconstraint;
1521 struct event_constraint unconstrained;
1522 
1523 static int x86_pmu_prepare_cpu(unsigned int cpu)
1524 {
1525 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1526 	int i;
1527 
1528 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1529 		cpuc->kfree_on_online[i] = NULL;
1530 	if (x86_pmu.cpu_prepare)
1531 		return x86_pmu.cpu_prepare(cpu);
1532 	return 0;
1533 }
1534 
1535 static int x86_pmu_dead_cpu(unsigned int cpu)
1536 {
1537 	if (x86_pmu.cpu_dead)
1538 		x86_pmu.cpu_dead(cpu);
1539 	return 0;
1540 }
1541 
1542 static int x86_pmu_online_cpu(unsigned int cpu)
1543 {
1544 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1545 	int i;
1546 
1547 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1548 		kfree(cpuc->kfree_on_online[i]);
1549 		cpuc->kfree_on_online[i] = NULL;
1550 	}
1551 	return 0;
1552 }
1553 
1554 static int x86_pmu_starting_cpu(unsigned int cpu)
1555 {
1556 	if (x86_pmu.cpu_starting)
1557 		x86_pmu.cpu_starting(cpu);
1558 	return 0;
1559 }
1560 
1561 static int x86_pmu_dying_cpu(unsigned int cpu)
1562 {
1563 	if (x86_pmu.cpu_dying)
1564 		x86_pmu.cpu_dying(cpu);
1565 	return 0;
1566 }
1567 
1568 static void __init pmu_check_apic(void)
1569 {
1570 	if (boot_cpu_has(X86_FEATURE_APIC))
1571 		return;
1572 
1573 	x86_pmu.apic = 0;
1574 	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1575 	pr_info("no hardware sampling interrupt available.\n");
1576 
1577 	/*
1578 	 * If we have a PMU initialized but no APIC
1579 	 * interrupts, we cannot sample hardware
1580 	 * events (user-space has to fall back and
1581 	 * sample via a hrtimer based software event):
1582 	 */
1583 	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1584 
1585 }
1586 
1587 static struct attribute_group x86_pmu_format_group = {
1588 	.name = "format",
1589 	.attrs = NULL,
1590 };
1591 
1592 /*
1593  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1594  * out of events_attr attributes.
1595  */
1596 static void __init filter_events(struct attribute **attrs)
1597 {
1598 	struct device_attribute *d;
1599 	struct perf_pmu_events_attr *pmu_attr;
1600 	int offset = 0;
1601 	int i, j;
1602 
1603 	for (i = 0; attrs[i]; i++) {
1604 		d = (struct device_attribute *)attrs[i];
1605 		pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1606 		/* str trumps id */
1607 		if (pmu_attr->event_str)
1608 			continue;
1609 		if (x86_pmu.event_map(i + offset))
1610 			continue;
1611 
1612 		for (j = i; attrs[j]; j++)
1613 			attrs[j] = attrs[j + 1];
1614 
1615 		/* Check the shifted attr. */
1616 		i--;
1617 
1618 		/*
1619 		 * event_map() is index based, the attrs array is organized
1620 		 * by increasing event index. If we shift the events, then
1621 		 * we need to compensate for the event_map(), otherwise
1622 		 * we are looking up the wrong event in the map
1623 		 */
1624 		offset++;
1625 	}
1626 }
1627 
1628 /* Merge two pointer arrays */
1629 __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1630 {
1631 	struct attribute **new;
1632 	int j, i;
1633 
1634 	for (j = 0; a[j]; j++)
1635 		;
1636 	for (i = 0; b[i]; i++)
1637 		j++;
1638 	j++;
1639 
1640 	new = kmalloc_array(j, sizeof(struct attribute *), GFP_KERNEL);
1641 	if (!new)
1642 		return NULL;
1643 
1644 	j = 0;
1645 	for (i = 0; a[i]; i++)
1646 		new[j++] = a[i];
1647 	for (i = 0; b[i]; i++)
1648 		new[j++] = b[i];
1649 	new[j] = NULL;
1650 
1651 	return new;
1652 }
1653 
1654 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1655 {
1656 	struct perf_pmu_events_attr *pmu_attr = \
1657 		container_of(attr, struct perf_pmu_events_attr, attr);
1658 	u64 config = x86_pmu.event_map(pmu_attr->id);
1659 
1660 	/* string trumps id */
1661 	if (pmu_attr->event_str)
1662 		return sprintf(page, "%s", pmu_attr->event_str);
1663 
1664 	return x86_pmu.events_sysfs_show(page, config);
1665 }
1666 EXPORT_SYMBOL_GPL(events_sysfs_show);
1667 
1668 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1669 			  char *page)
1670 {
1671 	struct perf_pmu_events_ht_attr *pmu_attr =
1672 		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1673 
1674 	/*
1675 	 * Report conditional events depending on Hyper-Threading.
1676 	 *
1677 	 * This is overly conservative as usually the HT special
1678 	 * handling is not needed if the other CPU thread is idle.
1679 	 *
1680 	 * Note this does not (and cannot) handle the case when thread
1681 	 * siblings are invisible, for example with virtualization
1682 	 * if they are owned by some other guest.  The user tool
1683 	 * has to re-read when a thread sibling gets onlined later.
1684 	 */
1685 	return sprintf(page, "%s",
1686 			topology_max_smt_threads() > 1 ?
1687 			pmu_attr->event_str_ht :
1688 			pmu_attr->event_str_noht);
1689 }
1690 
1691 EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1692 EVENT_ATTR(instructions,		INSTRUCTIONS		);
1693 EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1694 EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1695 EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1696 EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1697 EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1698 EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1699 EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1700 EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1701 
1702 static struct attribute *empty_attrs;
1703 
1704 static struct attribute *events_attr[] = {
1705 	EVENT_PTR(CPU_CYCLES),
1706 	EVENT_PTR(INSTRUCTIONS),
1707 	EVENT_PTR(CACHE_REFERENCES),
1708 	EVENT_PTR(CACHE_MISSES),
1709 	EVENT_PTR(BRANCH_INSTRUCTIONS),
1710 	EVENT_PTR(BRANCH_MISSES),
1711 	EVENT_PTR(BUS_CYCLES),
1712 	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1713 	EVENT_PTR(STALLED_CYCLES_BACKEND),
1714 	EVENT_PTR(REF_CPU_CYCLES),
1715 	NULL,
1716 };
1717 
1718 static struct attribute_group x86_pmu_events_group = {
1719 	.name = "events",
1720 	.attrs = events_attr,
1721 };
1722 
1723 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1724 {
1725 	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1726 	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1727 	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1728 	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1729 	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1730 	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1731 	ssize_t ret;
1732 
1733 	/*
1734 	* We have whole page size to spend and just little data
1735 	* to write, so we can safely use sprintf.
1736 	*/
1737 	ret = sprintf(page, "event=0x%02llx", event);
1738 
1739 	if (umask)
1740 		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1741 
1742 	if (edge)
1743 		ret += sprintf(page + ret, ",edge");
1744 
1745 	if (pc)
1746 		ret += sprintf(page + ret, ",pc");
1747 
1748 	if (any)
1749 		ret += sprintf(page + ret, ",any");
1750 
1751 	if (inv)
1752 		ret += sprintf(page + ret, ",inv");
1753 
1754 	if (cmask)
1755 		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1756 
1757 	ret += sprintf(page + ret, "\n");
1758 
1759 	return ret;
1760 }
1761 
1762 static struct attribute_group x86_pmu_attr_group;
1763 static struct attribute_group x86_pmu_caps_group;
1764 
1765 static int __init init_hw_perf_events(void)
1766 {
1767 	struct x86_pmu_quirk *quirk;
1768 	int err;
1769 
1770 	pr_info("Performance Events: ");
1771 
1772 	switch (boot_cpu_data.x86_vendor) {
1773 	case X86_VENDOR_INTEL:
1774 		err = intel_pmu_init();
1775 		break;
1776 	case X86_VENDOR_AMD:
1777 		err = amd_pmu_init();
1778 		break;
1779 	default:
1780 		err = -ENOTSUPP;
1781 	}
1782 	if (err != 0) {
1783 		pr_cont("no PMU driver, software events only.\n");
1784 		return 0;
1785 	}
1786 
1787 	pmu_check_apic();
1788 
1789 	/* sanity check that the hardware exists or is emulated */
1790 	if (!check_hw_exists())
1791 		return 0;
1792 
1793 	pr_cont("%s PMU driver.\n", x86_pmu.name);
1794 
1795 	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1796 
1797 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1798 		quirk->func();
1799 
1800 	if (!x86_pmu.intel_ctrl)
1801 		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1802 
1803 	perf_events_lapic_init();
1804 	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1805 
1806 	unconstrained = (struct event_constraint)
1807 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1808 				   0, x86_pmu.num_counters, 0, 0);
1809 
1810 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1811 
1812 	if (x86_pmu.caps_attrs) {
1813 		struct attribute **tmp;
1814 
1815 		tmp = merge_attr(x86_pmu_caps_group.attrs, x86_pmu.caps_attrs);
1816 		if (!WARN_ON(!tmp))
1817 			x86_pmu_caps_group.attrs = tmp;
1818 	}
1819 
1820 	if (x86_pmu.event_attrs)
1821 		x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1822 
1823 	if (!x86_pmu.events_sysfs_show)
1824 		x86_pmu_events_group.attrs = &empty_attrs;
1825 	else
1826 		filter_events(x86_pmu_events_group.attrs);
1827 
1828 	if (x86_pmu.cpu_events) {
1829 		struct attribute **tmp;
1830 
1831 		tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1832 		if (!WARN_ON(!tmp))
1833 			x86_pmu_events_group.attrs = tmp;
1834 	}
1835 
1836 	if (x86_pmu.attrs) {
1837 		struct attribute **tmp;
1838 
1839 		tmp = merge_attr(x86_pmu_attr_group.attrs, x86_pmu.attrs);
1840 		if (!WARN_ON(!tmp))
1841 			x86_pmu_attr_group.attrs = tmp;
1842 	}
1843 
1844 	pr_info("... version:                %d\n",     x86_pmu.version);
1845 	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1846 	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1847 	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1848 	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1849 	pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1850 	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1851 
1852 	/*
1853 	 * Install callbacks. Core will call them for each online
1854 	 * cpu.
1855 	 */
1856 	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
1857 				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1858 	if (err)
1859 		return err;
1860 
1861 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1862 				"perf/x86:starting", x86_pmu_starting_cpu,
1863 				x86_pmu_dying_cpu);
1864 	if (err)
1865 		goto out;
1866 
1867 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
1868 				x86_pmu_online_cpu, NULL);
1869 	if (err)
1870 		goto out1;
1871 
1872 	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1873 	if (err)
1874 		goto out2;
1875 
1876 	return 0;
1877 
1878 out2:
1879 	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1880 out1:
1881 	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1882 out:
1883 	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
1884 	return err;
1885 }
1886 early_initcall(init_hw_perf_events);
1887 
1888 static inline void x86_pmu_read(struct perf_event *event)
1889 {
1890 	if (x86_pmu.read)
1891 		return x86_pmu.read(event);
1892 	x86_perf_event_update(event);
1893 }
1894 
1895 /*
1896  * Start group events scheduling transaction
1897  * Set the flag to make pmu::enable() not perform the
1898  * schedulability test, it will be performed at commit time
1899  *
1900  * We only support PERF_PMU_TXN_ADD transactions. Save the
1901  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1902  * transactions.
1903  */
1904 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1905 {
1906 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1907 
1908 	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
1909 
1910 	cpuc->txn_flags = txn_flags;
1911 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1912 		return;
1913 
1914 	perf_pmu_disable(pmu);
1915 	__this_cpu_write(cpu_hw_events.n_txn, 0);
1916 }
1917 
1918 /*
1919  * Stop group events scheduling transaction
1920  * Clear the flag and pmu::enable() will perform the
1921  * schedulability test.
1922  */
1923 static void x86_pmu_cancel_txn(struct pmu *pmu)
1924 {
1925 	unsigned int txn_flags;
1926 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1927 
1928 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1929 
1930 	txn_flags = cpuc->txn_flags;
1931 	cpuc->txn_flags = 0;
1932 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1933 		return;
1934 
1935 	/*
1936 	 * Truncate collected array by the number of events added in this
1937 	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1938 	 */
1939 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1940 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1941 	perf_pmu_enable(pmu);
1942 }
1943 
1944 /*
1945  * Commit group events scheduling transaction
1946  * Perform the group schedulability test as a whole
1947  * Return 0 if success
1948  *
1949  * Does not cancel the transaction on failure; expects the caller to do this.
1950  */
1951 static int x86_pmu_commit_txn(struct pmu *pmu)
1952 {
1953 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1954 	int assign[X86_PMC_IDX_MAX];
1955 	int n, ret;
1956 
1957 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1958 
1959 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1960 		cpuc->txn_flags = 0;
1961 		return 0;
1962 	}
1963 
1964 	n = cpuc->n_events;
1965 
1966 	if (!x86_pmu_initialized())
1967 		return -EAGAIN;
1968 
1969 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1970 	if (ret)
1971 		return ret;
1972 
1973 	/*
1974 	 * copy new assignment, now we know it is possible
1975 	 * will be used by hw_perf_enable()
1976 	 */
1977 	memcpy(cpuc->assign, assign, n*sizeof(int));
1978 
1979 	cpuc->txn_flags = 0;
1980 	perf_pmu_enable(pmu);
1981 	return 0;
1982 }
1983 /*
1984  * a fake_cpuc is used to validate event groups. Due to
1985  * the extra reg logic, we need to also allocate a fake
1986  * per_core and per_cpu structure. Otherwise, group events
1987  * using extra reg may conflict without the kernel being
1988  * able to catch this when the last event gets added to
1989  * the group.
1990  */
1991 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1992 {
1993 	kfree(cpuc->shared_regs);
1994 	kfree(cpuc);
1995 }
1996 
1997 static struct cpu_hw_events *allocate_fake_cpuc(void)
1998 {
1999 	struct cpu_hw_events *cpuc;
2000 	int cpu = raw_smp_processor_id();
2001 
2002 	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
2003 	if (!cpuc)
2004 		return ERR_PTR(-ENOMEM);
2005 
2006 	/* only needed, if we have extra_regs */
2007 	if (x86_pmu.extra_regs) {
2008 		cpuc->shared_regs = allocate_shared_regs(cpu);
2009 		if (!cpuc->shared_regs)
2010 			goto error;
2011 	}
2012 	cpuc->is_fake = 1;
2013 	return cpuc;
2014 error:
2015 	free_fake_cpuc(cpuc);
2016 	return ERR_PTR(-ENOMEM);
2017 }
2018 
2019 /*
2020  * validate that we can schedule this event
2021  */
2022 static int validate_event(struct perf_event *event)
2023 {
2024 	struct cpu_hw_events *fake_cpuc;
2025 	struct event_constraint *c;
2026 	int ret = 0;
2027 
2028 	fake_cpuc = allocate_fake_cpuc();
2029 	if (IS_ERR(fake_cpuc))
2030 		return PTR_ERR(fake_cpuc);
2031 
2032 	c = x86_pmu.get_event_constraints(fake_cpuc, -1, event);
2033 
2034 	if (!c || !c->weight)
2035 		ret = -EINVAL;
2036 
2037 	if (x86_pmu.put_event_constraints)
2038 		x86_pmu.put_event_constraints(fake_cpuc, event);
2039 
2040 	free_fake_cpuc(fake_cpuc);
2041 
2042 	return ret;
2043 }
2044 
2045 /*
2046  * validate a single event group
2047  *
2048  * validation include:
2049  *	- check events are compatible which each other
2050  *	- events do not compete for the same counter
2051  *	- number of events <= number of counters
2052  *
2053  * validation ensures the group can be loaded onto the
2054  * PMU if it was the only group available.
2055  */
2056 static int validate_group(struct perf_event *event)
2057 {
2058 	struct perf_event *leader = event->group_leader;
2059 	struct cpu_hw_events *fake_cpuc;
2060 	int ret = -EINVAL, n;
2061 
2062 	fake_cpuc = allocate_fake_cpuc();
2063 	if (IS_ERR(fake_cpuc))
2064 		return PTR_ERR(fake_cpuc);
2065 	/*
2066 	 * the event is not yet connected with its
2067 	 * siblings therefore we must first collect
2068 	 * existing siblings, then add the new event
2069 	 * before we can simulate the scheduling
2070 	 */
2071 	n = collect_events(fake_cpuc, leader, true);
2072 	if (n < 0)
2073 		goto out;
2074 
2075 	fake_cpuc->n_events = n;
2076 	n = collect_events(fake_cpuc, event, false);
2077 	if (n < 0)
2078 		goto out;
2079 
2080 	fake_cpuc->n_events = n;
2081 
2082 	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2083 
2084 out:
2085 	free_fake_cpuc(fake_cpuc);
2086 	return ret;
2087 }
2088 
2089 static int x86_pmu_event_init(struct perf_event *event)
2090 {
2091 	struct pmu *tmp;
2092 	int err;
2093 
2094 	switch (event->attr.type) {
2095 	case PERF_TYPE_RAW:
2096 	case PERF_TYPE_HARDWARE:
2097 	case PERF_TYPE_HW_CACHE:
2098 		break;
2099 
2100 	default:
2101 		return -ENOENT;
2102 	}
2103 
2104 	err = __x86_pmu_event_init(event);
2105 	if (!err) {
2106 		/*
2107 		 * we temporarily connect event to its pmu
2108 		 * such that validate_group() can classify
2109 		 * it as an x86 event using is_x86_event()
2110 		 */
2111 		tmp = event->pmu;
2112 		event->pmu = &pmu;
2113 
2114 		if (event->group_leader != event)
2115 			err = validate_group(event);
2116 		else
2117 			err = validate_event(event);
2118 
2119 		event->pmu = tmp;
2120 	}
2121 	if (err) {
2122 		if (event->destroy)
2123 			event->destroy(event);
2124 	}
2125 
2126 	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2127 	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2128 		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2129 
2130 	return err;
2131 }
2132 
2133 static void refresh_pce(void *ignored)
2134 {
2135 	load_mm_cr4(this_cpu_read(cpu_tlbstate.loaded_mm));
2136 }
2137 
2138 static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2139 {
2140 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2141 		return;
2142 
2143 	/*
2144 	 * This function relies on not being called concurrently in two
2145 	 * tasks in the same mm.  Otherwise one task could observe
2146 	 * perf_rdpmc_allowed > 1 and return all the way back to
2147 	 * userspace with CR4.PCE clear while another task is still
2148 	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2149 	 *
2150 	 * For now, this can't happen because all callers hold mmap_sem
2151 	 * for write.  If this changes, we'll need a different solution.
2152 	 */
2153 	lockdep_assert_held_exclusive(&mm->mmap_sem);
2154 
2155 	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2156 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2157 }
2158 
2159 static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2160 {
2161 
2162 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2163 		return;
2164 
2165 	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2166 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2167 }
2168 
2169 static int x86_pmu_event_idx(struct perf_event *event)
2170 {
2171 	int idx = event->hw.idx;
2172 
2173 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2174 		return 0;
2175 
2176 	if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2177 		idx -= INTEL_PMC_IDX_FIXED;
2178 		idx |= 1 << 30;
2179 	}
2180 
2181 	return idx + 1;
2182 }
2183 
2184 static ssize_t get_attr_rdpmc(struct device *cdev,
2185 			      struct device_attribute *attr,
2186 			      char *buf)
2187 {
2188 	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2189 }
2190 
2191 static ssize_t set_attr_rdpmc(struct device *cdev,
2192 			      struct device_attribute *attr,
2193 			      const char *buf, size_t count)
2194 {
2195 	unsigned long val;
2196 	ssize_t ret;
2197 
2198 	ret = kstrtoul(buf, 0, &val);
2199 	if (ret)
2200 		return ret;
2201 
2202 	if (val > 2)
2203 		return -EINVAL;
2204 
2205 	if (x86_pmu.attr_rdpmc_broken)
2206 		return -ENOTSUPP;
2207 
2208 	if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2209 		/*
2210 		 * Changing into or out of always available, aka
2211 		 * perf-event-bypassing mode.  This path is extremely slow,
2212 		 * but only root can trigger it, so it's okay.
2213 		 */
2214 		if (val == 2)
2215 			static_branch_inc(&rdpmc_always_available_key);
2216 		else
2217 			static_branch_dec(&rdpmc_always_available_key);
2218 		on_each_cpu(refresh_pce, NULL, 1);
2219 	}
2220 
2221 	x86_pmu.attr_rdpmc = val;
2222 
2223 	return count;
2224 }
2225 
2226 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2227 
2228 static struct attribute *x86_pmu_attrs[] = {
2229 	&dev_attr_rdpmc.attr,
2230 	NULL,
2231 };
2232 
2233 static struct attribute_group x86_pmu_attr_group = {
2234 	.attrs = x86_pmu_attrs,
2235 };
2236 
2237 static ssize_t max_precise_show(struct device *cdev,
2238 				  struct device_attribute *attr,
2239 				  char *buf)
2240 {
2241 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2242 }
2243 
2244 static DEVICE_ATTR_RO(max_precise);
2245 
2246 static struct attribute *x86_pmu_caps_attrs[] = {
2247 	&dev_attr_max_precise.attr,
2248 	NULL
2249 };
2250 
2251 static struct attribute_group x86_pmu_caps_group = {
2252 	.name = "caps",
2253 	.attrs = x86_pmu_caps_attrs,
2254 };
2255 
2256 static const struct attribute_group *x86_pmu_attr_groups[] = {
2257 	&x86_pmu_attr_group,
2258 	&x86_pmu_format_group,
2259 	&x86_pmu_events_group,
2260 	&x86_pmu_caps_group,
2261 	NULL,
2262 };
2263 
2264 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2265 {
2266 	if (x86_pmu.sched_task)
2267 		x86_pmu.sched_task(ctx, sched_in);
2268 }
2269 
2270 void perf_check_microcode(void)
2271 {
2272 	if (x86_pmu.check_microcode)
2273 		x86_pmu.check_microcode();
2274 }
2275 
2276 static struct pmu pmu = {
2277 	.pmu_enable		= x86_pmu_enable,
2278 	.pmu_disable		= x86_pmu_disable,
2279 
2280 	.attr_groups		= x86_pmu_attr_groups,
2281 
2282 	.event_init		= x86_pmu_event_init,
2283 
2284 	.event_mapped		= x86_pmu_event_mapped,
2285 	.event_unmapped		= x86_pmu_event_unmapped,
2286 
2287 	.add			= x86_pmu_add,
2288 	.del			= x86_pmu_del,
2289 	.start			= x86_pmu_start,
2290 	.stop			= x86_pmu_stop,
2291 	.read			= x86_pmu_read,
2292 
2293 	.start_txn		= x86_pmu_start_txn,
2294 	.cancel_txn		= x86_pmu_cancel_txn,
2295 	.commit_txn		= x86_pmu_commit_txn,
2296 
2297 	.event_idx		= x86_pmu_event_idx,
2298 	.sched_task		= x86_pmu_sched_task,
2299 	.task_ctx_size          = sizeof(struct x86_perf_task_context),
2300 };
2301 
2302 void arch_perf_update_userpage(struct perf_event *event,
2303 			       struct perf_event_mmap_page *userpg, u64 now)
2304 {
2305 	struct cyc2ns_data data;
2306 	u64 offset;
2307 
2308 	userpg->cap_user_time = 0;
2309 	userpg->cap_user_time_zero = 0;
2310 	userpg->cap_user_rdpmc =
2311 		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2312 	userpg->pmc_width = x86_pmu.cntval_bits;
2313 
2314 	if (!using_native_sched_clock() || !sched_clock_stable())
2315 		return;
2316 
2317 	cyc2ns_read_begin(&data);
2318 
2319 	offset = data.cyc2ns_offset + __sched_clock_offset;
2320 
2321 	/*
2322 	 * Internal timekeeping for enabled/running/stopped times
2323 	 * is always in the local_clock domain.
2324 	 */
2325 	userpg->cap_user_time = 1;
2326 	userpg->time_mult = data.cyc2ns_mul;
2327 	userpg->time_shift = data.cyc2ns_shift;
2328 	userpg->time_offset = offset - now;
2329 
2330 	/*
2331 	 * cap_user_time_zero doesn't make sense when we're using a different
2332 	 * time base for the records.
2333 	 */
2334 	if (!event->attr.use_clockid) {
2335 		userpg->cap_user_time_zero = 1;
2336 		userpg->time_zero = offset;
2337 	}
2338 
2339 	cyc2ns_read_end();
2340 }
2341 
2342 void
2343 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2344 {
2345 	struct unwind_state state;
2346 	unsigned long addr;
2347 
2348 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2349 		/* TODO: We don't support guest os callchain now */
2350 		return;
2351 	}
2352 
2353 	if (perf_callchain_store(entry, regs->ip))
2354 		return;
2355 
2356 	for (unwind_start(&state, current, regs, NULL); !unwind_done(&state);
2357 	     unwind_next_frame(&state)) {
2358 		addr = unwind_get_return_address(&state);
2359 		if (!addr || perf_callchain_store(entry, addr))
2360 			return;
2361 	}
2362 }
2363 
2364 static inline int
2365 valid_user_frame(const void __user *fp, unsigned long size)
2366 {
2367 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2368 }
2369 
2370 static unsigned long get_segment_base(unsigned int segment)
2371 {
2372 	struct desc_struct *desc;
2373 	unsigned int idx = segment >> 3;
2374 
2375 	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2376 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2377 		struct ldt_struct *ldt;
2378 
2379 		/* IRQs are off, so this synchronizes with smp_store_release */
2380 		ldt = READ_ONCE(current->active_mm->context.ldt);
2381 		if (!ldt || idx >= ldt->nr_entries)
2382 			return 0;
2383 
2384 		desc = &ldt->entries[idx];
2385 #else
2386 		return 0;
2387 #endif
2388 	} else {
2389 		if (idx >= GDT_ENTRIES)
2390 			return 0;
2391 
2392 		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2393 	}
2394 
2395 	return get_desc_base(desc);
2396 }
2397 
2398 #ifdef CONFIG_IA32_EMULATION
2399 
2400 #include <linux/compat.h>
2401 
2402 static inline int
2403 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2404 {
2405 	/* 32-bit process in 64-bit kernel. */
2406 	unsigned long ss_base, cs_base;
2407 	struct stack_frame_ia32 frame;
2408 	const void __user *fp;
2409 
2410 	if (!test_thread_flag(TIF_IA32))
2411 		return 0;
2412 
2413 	cs_base = get_segment_base(regs->cs);
2414 	ss_base = get_segment_base(regs->ss);
2415 
2416 	fp = compat_ptr(ss_base + regs->bp);
2417 	pagefault_disable();
2418 	while (entry->nr < entry->max_stack) {
2419 		unsigned long bytes;
2420 		frame.next_frame     = 0;
2421 		frame.return_address = 0;
2422 
2423 		if (!valid_user_frame(fp, sizeof(frame)))
2424 			break;
2425 
2426 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2427 		if (bytes != 0)
2428 			break;
2429 		bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2430 		if (bytes != 0)
2431 			break;
2432 
2433 		perf_callchain_store(entry, cs_base + frame.return_address);
2434 		fp = compat_ptr(ss_base + frame.next_frame);
2435 	}
2436 	pagefault_enable();
2437 	return 1;
2438 }
2439 #else
2440 static inline int
2441 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2442 {
2443     return 0;
2444 }
2445 #endif
2446 
2447 void
2448 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2449 {
2450 	struct stack_frame frame;
2451 	const unsigned long __user *fp;
2452 
2453 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2454 		/* TODO: We don't support guest os callchain now */
2455 		return;
2456 	}
2457 
2458 	/*
2459 	 * We don't know what to do with VM86 stacks.. ignore them for now.
2460 	 */
2461 	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2462 		return;
2463 
2464 	fp = (unsigned long __user *)regs->bp;
2465 
2466 	perf_callchain_store(entry, regs->ip);
2467 
2468 	if (!current->mm)
2469 		return;
2470 
2471 	if (perf_callchain_user32(regs, entry))
2472 		return;
2473 
2474 	pagefault_disable();
2475 	while (entry->nr < entry->max_stack) {
2476 		unsigned long bytes;
2477 
2478 		frame.next_frame	     = NULL;
2479 		frame.return_address = 0;
2480 
2481 		if (!valid_user_frame(fp, sizeof(frame)))
2482 			break;
2483 
2484 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2485 		if (bytes != 0)
2486 			break;
2487 		bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2488 		if (bytes != 0)
2489 			break;
2490 
2491 		perf_callchain_store(entry, frame.return_address);
2492 		fp = (void __user *)frame.next_frame;
2493 	}
2494 	pagefault_enable();
2495 }
2496 
2497 /*
2498  * Deal with code segment offsets for the various execution modes:
2499  *
2500  *   VM86 - the good olde 16 bit days, where the linear address is
2501  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2502  *
2503  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2504  *          to figure out what the 32bit base address is.
2505  *
2506  *    X32 - has TIF_X32 set, but is running in x86_64
2507  *
2508  * X86_64 - CS,DS,SS,ES are all zero based.
2509  */
2510 static unsigned long code_segment_base(struct pt_regs *regs)
2511 {
2512 	/*
2513 	 * For IA32 we look at the GDT/LDT segment base to convert the
2514 	 * effective IP to a linear address.
2515 	 */
2516 
2517 #ifdef CONFIG_X86_32
2518 	/*
2519 	 * If we are in VM86 mode, add the segment offset to convert to a
2520 	 * linear address.
2521 	 */
2522 	if (regs->flags & X86_VM_MASK)
2523 		return 0x10 * regs->cs;
2524 
2525 	if (user_mode(regs) && regs->cs != __USER_CS)
2526 		return get_segment_base(regs->cs);
2527 #else
2528 	if (user_mode(regs) && !user_64bit_mode(regs) &&
2529 	    regs->cs != __USER32_CS)
2530 		return get_segment_base(regs->cs);
2531 #endif
2532 	return 0;
2533 }
2534 
2535 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2536 {
2537 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2538 		return perf_guest_cbs->get_guest_ip();
2539 
2540 	return regs->ip + code_segment_base(regs);
2541 }
2542 
2543 unsigned long perf_misc_flags(struct pt_regs *regs)
2544 {
2545 	int misc = 0;
2546 
2547 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2548 		if (perf_guest_cbs->is_user_mode())
2549 			misc |= PERF_RECORD_MISC_GUEST_USER;
2550 		else
2551 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2552 	} else {
2553 		if (user_mode(regs))
2554 			misc |= PERF_RECORD_MISC_USER;
2555 		else
2556 			misc |= PERF_RECORD_MISC_KERNEL;
2557 	}
2558 
2559 	if (regs->flags & PERF_EFLAGS_EXACT)
2560 		misc |= PERF_RECORD_MISC_EXACT_IP;
2561 
2562 	return misc;
2563 }
2564 
2565 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2566 {
2567 	cap->version		= x86_pmu.version;
2568 	cap->num_counters_gp	= x86_pmu.num_counters;
2569 	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2570 	cap->bit_width_gp	= x86_pmu.cntval_bits;
2571 	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2572 	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2573 	cap->events_mask_len	= x86_pmu.events_mask_len;
2574 }
2575 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
2576