xref: /openbmc/linux/arch/x86/events/core.c (revision 6d99a79c)
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 /**
1037  * x86_perf_rdpmc_index - Return PMC counter used for event
1038  * @event: the perf_event to which the PMC counter was assigned
1039  *
1040  * The counter assigned to this performance event may change if interrupts
1041  * are enabled. This counter should thus never be used while interrupts are
1042  * enabled. Before this function is used to obtain the assigned counter the
1043  * event should be checked for validity using, for example,
1044  * perf_event_read_local(), within the same interrupt disabled section in
1045  * which this counter is planned to be used.
1046  *
1047  * Return: The index of the performance monitoring counter assigned to
1048  * @perf_event.
1049  */
1050 int x86_perf_rdpmc_index(struct perf_event *event)
1051 {
1052 	lockdep_assert_irqs_disabled();
1053 
1054 	return event->hw.event_base_rdpmc;
1055 }
1056 
1057 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1058 					struct cpu_hw_events *cpuc,
1059 					int i)
1060 {
1061 	return hwc->idx == cpuc->assign[i] &&
1062 		hwc->last_cpu == smp_processor_id() &&
1063 		hwc->last_tag == cpuc->tags[i];
1064 }
1065 
1066 static void x86_pmu_start(struct perf_event *event, int flags);
1067 
1068 static void x86_pmu_enable(struct pmu *pmu)
1069 {
1070 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1071 	struct perf_event *event;
1072 	struct hw_perf_event *hwc;
1073 	int i, added = cpuc->n_added;
1074 
1075 	if (!x86_pmu_initialized())
1076 		return;
1077 
1078 	if (cpuc->enabled)
1079 		return;
1080 
1081 	if (cpuc->n_added) {
1082 		int n_running = cpuc->n_events - cpuc->n_added;
1083 		/*
1084 		 * apply assignment obtained either from
1085 		 * hw_perf_group_sched_in() or x86_pmu_enable()
1086 		 *
1087 		 * step1: save events moving to new counters
1088 		 */
1089 		for (i = 0; i < n_running; i++) {
1090 			event = cpuc->event_list[i];
1091 			hwc = &event->hw;
1092 
1093 			/*
1094 			 * we can avoid reprogramming counter if:
1095 			 * - assigned same counter as last time
1096 			 * - running on same CPU as last time
1097 			 * - no other event has used the counter since
1098 			 */
1099 			if (hwc->idx == -1 ||
1100 			    match_prev_assignment(hwc, cpuc, i))
1101 				continue;
1102 
1103 			/*
1104 			 * Ensure we don't accidentally enable a stopped
1105 			 * counter simply because we rescheduled.
1106 			 */
1107 			if (hwc->state & PERF_HES_STOPPED)
1108 				hwc->state |= PERF_HES_ARCH;
1109 
1110 			x86_pmu_stop(event, PERF_EF_UPDATE);
1111 		}
1112 
1113 		/*
1114 		 * step2: reprogram moved events into new counters
1115 		 */
1116 		for (i = 0; i < cpuc->n_events; i++) {
1117 			event = cpuc->event_list[i];
1118 			hwc = &event->hw;
1119 
1120 			if (!match_prev_assignment(hwc, cpuc, i))
1121 				x86_assign_hw_event(event, cpuc, i);
1122 			else if (i < n_running)
1123 				continue;
1124 
1125 			if (hwc->state & PERF_HES_ARCH)
1126 				continue;
1127 
1128 			x86_pmu_start(event, PERF_EF_RELOAD);
1129 		}
1130 		cpuc->n_added = 0;
1131 		perf_events_lapic_init();
1132 	}
1133 
1134 	cpuc->enabled = 1;
1135 	barrier();
1136 
1137 	x86_pmu.enable_all(added);
1138 }
1139 
1140 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1141 
1142 /*
1143  * Set the next IRQ period, based on the hwc->period_left value.
1144  * To be called with the event disabled in hw:
1145  */
1146 int x86_perf_event_set_period(struct perf_event *event)
1147 {
1148 	struct hw_perf_event *hwc = &event->hw;
1149 	s64 left = local64_read(&hwc->period_left);
1150 	s64 period = hwc->sample_period;
1151 	int ret = 0, idx = hwc->idx;
1152 
1153 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
1154 		return 0;
1155 
1156 	/*
1157 	 * If we are way outside a reasonable range then just skip forward:
1158 	 */
1159 	if (unlikely(left <= -period)) {
1160 		left = period;
1161 		local64_set(&hwc->period_left, left);
1162 		hwc->last_period = period;
1163 		ret = 1;
1164 	}
1165 
1166 	if (unlikely(left <= 0)) {
1167 		left += period;
1168 		local64_set(&hwc->period_left, left);
1169 		hwc->last_period = period;
1170 		ret = 1;
1171 	}
1172 	/*
1173 	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1174 	 */
1175 	if (unlikely(left < 2))
1176 		left = 2;
1177 
1178 	if (left > x86_pmu.max_period)
1179 		left = x86_pmu.max_period;
1180 
1181 	if (x86_pmu.limit_period)
1182 		left = x86_pmu.limit_period(event, left);
1183 
1184 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1185 
1186 	/*
1187 	 * The hw event starts counting from this event offset,
1188 	 * mark it to be able to extra future deltas:
1189 	 */
1190 	local64_set(&hwc->prev_count, (u64)-left);
1191 
1192 	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1193 
1194 	/*
1195 	 * Due to erratum on certan cpu we need
1196 	 * a second write to be sure the register
1197 	 * is updated properly
1198 	 */
1199 	if (x86_pmu.perfctr_second_write) {
1200 		wrmsrl(hwc->event_base,
1201 			(u64)(-left) & x86_pmu.cntval_mask);
1202 	}
1203 
1204 	perf_event_update_userpage(event);
1205 
1206 	return ret;
1207 }
1208 
1209 void x86_pmu_enable_event(struct perf_event *event)
1210 {
1211 	if (__this_cpu_read(cpu_hw_events.enabled))
1212 		__x86_pmu_enable_event(&event->hw,
1213 				       ARCH_PERFMON_EVENTSEL_ENABLE);
1214 }
1215 
1216 /*
1217  * Add a single event to the PMU.
1218  *
1219  * The event is added to the group of enabled events
1220  * but only if it can be scehduled with existing events.
1221  */
1222 static int x86_pmu_add(struct perf_event *event, int flags)
1223 {
1224 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1225 	struct hw_perf_event *hwc;
1226 	int assign[X86_PMC_IDX_MAX];
1227 	int n, n0, ret;
1228 
1229 	hwc = &event->hw;
1230 
1231 	n0 = cpuc->n_events;
1232 	ret = n = collect_events(cpuc, event, false);
1233 	if (ret < 0)
1234 		goto out;
1235 
1236 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1237 	if (!(flags & PERF_EF_START))
1238 		hwc->state |= PERF_HES_ARCH;
1239 
1240 	/*
1241 	 * If group events scheduling transaction was started,
1242 	 * skip the schedulability test here, it will be performed
1243 	 * at commit time (->commit_txn) as a whole.
1244 	 *
1245 	 * If commit fails, we'll call ->del() on all events
1246 	 * for which ->add() was called.
1247 	 */
1248 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1249 		goto done_collect;
1250 
1251 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1252 	if (ret)
1253 		goto out;
1254 	/*
1255 	 * copy new assignment, now we know it is possible
1256 	 * will be used by hw_perf_enable()
1257 	 */
1258 	memcpy(cpuc->assign, assign, n*sizeof(int));
1259 
1260 done_collect:
1261 	/*
1262 	 * Commit the collect_events() state. See x86_pmu_del() and
1263 	 * x86_pmu_*_txn().
1264 	 */
1265 	cpuc->n_events = n;
1266 	cpuc->n_added += n - n0;
1267 	cpuc->n_txn += n - n0;
1268 
1269 	if (x86_pmu.add) {
1270 		/*
1271 		 * This is before x86_pmu_enable() will call x86_pmu_start(),
1272 		 * so we enable LBRs before an event needs them etc..
1273 		 */
1274 		x86_pmu.add(event);
1275 	}
1276 
1277 	ret = 0;
1278 out:
1279 	return ret;
1280 }
1281 
1282 static void x86_pmu_start(struct perf_event *event, int flags)
1283 {
1284 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1285 	int idx = event->hw.idx;
1286 
1287 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1288 		return;
1289 
1290 	if (WARN_ON_ONCE(idx == -1))
1291 		return;
1292 
1293 	if (flags & PERF_EF_RELOAD) {
1294 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1295 		x86_perf_event_set_period(event);
1296 	}
1297 
1298 	event->hw.state = 0;
1299 
1300 	cpuc->events[idx] = event;
1301 	__set_bit(idx, cpuc->active_mask);
1302 	__set_bit(idx, cpuc->running);
1303 	x86_pmu.enable(event);
1304 	perf_event_update_userpage(event);
1305 }
1306 
1307 void perf_event_print_debug(void)
1308 {
1309 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1310 	u64 pebs, debugctl;
1311 	struct cpu_hw_events *cpuc;
1312 	unsigned long flags;
1313 	int cpu, idx;
1314 
1315 	if (!x86_pmu.num_counters)
1316 		return;
1317 
1318 	local_irq_save(flags);
1319 
1320 	cpu = smp_processor_id();
1321 	cpuc = &per_cpu(cpu_hw_events, cpu);
1322 
1323 	if (x86_pmu.version >= 2) {
1324 		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1325 		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1326 		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1327 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1328 
1329 		pr_info("\n");
1330 		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1331 		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1332 		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1333 		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1334 		if (x86_pmu.pebs_constraints) {
1335 			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1336 			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1337 		}
1338 		if (x86_pmu.lbr_nr) {
1339 			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1340 			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1341 		}
1342 	}
1343 	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1344 
1345 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1346 		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1347 		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1348 
1349 		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1350 
1351 		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1352 			cpu, idx, pmc_ctrl);
1353 		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1354 			cpu, idx, pmc_count);
1355 		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1356 			cpu, idx, prev_left);
1357 	}
1358 	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1359 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1360 
1361 		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1362 			cpu, idx, pmc_count);
1363 	}
1364 	local_irq_restore(flags);
1365 }
1366 
1367 void x86_pmu_stop(struct perf_event *event, int flags)
1368 {
1369 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1370 	struct hw_perf_event *hwc = &event->hw;
1371 
1372 	if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1373 		x86_pmu.disable(event);
1374 		cpuc->events[hwc->idx] = NULL;
1375 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1376 		hwc->state |= PERF_HES_STOPPED;
1377 	}
1378 
1379 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1380 		/*
1381 		 * Drain the remaining delta count out of a event
1382 		 * that we are disabling:
1383 		 */
1384 		x86_perf_event_update(event);
1385 		hwc->state |= PERF_HES_UPTODATE;
1386 	}
1387 }
1388 
1389 static void x86_pmu_del(struct perf_event *event, int flags)
1390 {
1391 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1392 	int i;
1393 
1394 	/*
1395 	 * event is descheduled
1396 	 */
1397 	event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1398 
1399 	/*
1400 	 * If we're called during a txn, we only need to undo x86_pmu.add.
1401 	 * The events never got scheduled and ->cancel_txn will truncate
1402 	 * the event_list.
1403 	 *
1404 	 * XXX assumes any ->del() called during a TXN will only be on
1405 	 * an event added during that same TXN.
1406 	 */
1407 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1408 		goto do_del;
1409 
1410 	/*
1411 	 * Not a TXN, therefore cleanup properly.
1412 	 */
1413 	x86_pmu_stop(event, PERF_EF_UPDATE);
1414 
1415 	for (i = 0; i < cpuc->n_events; i++) {
1416 		if (event == cpuc->event_list[i])
1417 			break;
1418 	}
1419 
1420 	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1421 		return;
1422 
1423 	/* If we have a newly added event; make sure to decrease n_added. */
1424 	if (i >= cpuc->n_events - cpuc->n_added)
1425 		--cpuc->n_added;
1426 
1427 	if (x86_pmu.put_event_constraints)
1428 		x86_pmu.put_event_constraints(cpuc, event);
1429 
1430 	/* Delete the array entry. */
1431 	while (++i < cpuc->n_events) {
1432 		cpuc->event_list[i-1] = cpuc->event_list[i];
1433 		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1434 	}
1435 	--cpuc->n_events;
1436 
1437 	perf_event_update_userpage(event);
1438 
1439 do_del:
1440 	if (x86_pmu.del) {
1441 		/*
1442 		 * This is after x86_pmu_stop(); so we disable LBRs after any
1443 		 * event can need them etc..
1444 		 */
1445 		x86_pmu.del(event);
1446 	}
1447 }
1448 
1449 int x86_pmu_handle_irq(struct pt_regs *regs)
1450 {
1451 	struct perf_sample_data data;
1452 	struct cpu_hw_events *cpuc;
1453 	struct perf_event *event;
1454 	int idx, handled = 0;
1455 	u64 val;
1456 
1457 	cpuc = this_cpu_ptr(&cpu_hw_events);
1458 
1459 	/*
1460 	 * Some chipsets need to unmask the LVTPC in a particular spot
1461 	 * inside the nmi handler.  As a result, the unmasking was pushed
1462 	 * into all the nmi handlers.
1463 	 *
1464 	 * This generic handler doesn't seem to have any issues where the
1465 	 * unmasking occurs so it was left at the top.
1466 	 */
1467 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1468 
1469 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1470 		if (!test_bit(idx, cpuc->active_mask)) {
1471 			/*
1472 			 * Though we deactivated the counter some cpus
1473 			 * might still deliver spurious interrupts still
1474 			 * in flight. Catch them:
1475 			 */
1476 			if (__test_and_clear_bit(idx, cpuc->running))
1477 				handled++;
1478 			continue;
1479 		}
1480 
1481 		event = cpuc->events[idx];
1482 
1483 		val = x86_perf_event_update(event);
1484 		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1485 			continue;
1486 
1487 		/*
1488 		 * event overflow
1489 		 */
1490 		handled++;
1491 		perf_sample_data_init(&data, 0, event->hw.last_period);
1492 
1493 		if (!x86_perf_event_set_period(event))
1494 			continue;
1495 
1496 		if (perf_event_overflow(event, &data, regs))
1497 			x86_pmu_stop(event, 0);
1498 	}
1499 
1500 	if (handled)
1501 		inc_irq_stat(apic_perf_irqs);
1502 
1503 	return handled;
1504 }
1505 
1506 void perf_events_lapic_init(void)
1507 {
1508 	if (!x86_pmu.apic || !x86_pmu_initialized())
1509 		return;
1510 
1511 	/*
1512 	 * Always use NMI for PMU
1513 	 */
1514 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1515 }
1516 
1517 static int
1518 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1519 {
1520 	u64 start_clock;
1521 	u64 finish_clock;
1522 	int ret;
1523 
1524 	/*
1525 	 * All PMUs/events that share this PMI handler should make sure to
1526 	 * increment active_events for their events.
1527 	 */
1528 	if (!atomic_read(&active_events))
1529 		return NMI_DONE;
1530 
1531 	start_clock = sched_clock();
1532 	ret = x86_pmu.handle_irq(regs);
1533 	finish_clock = sched_clock();
1534 
1535 	perf_sample_event_took(finish_clock - start_clock);
1536 
1537 	return ret;
1538 }
1539 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1540 
1541 struct event_constraint emptyconstraint;
1542 struct event_constraint unconstrained;
1543 
1544 static int x86_pmu_prepare_cpu(unsigned int cpu)
1545 {
1546 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1547 	int i;
1548 
1549 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1550 		cpuc->kfree_on_online[i] = NULL;
1551 	if (x86_pmu.cpu_prepare)
1552 		return x86_pmu.cpu_prepare(cpu);
1553 	return 0;
1554 }
1555 
1556 static int x86_pmu_dead_cpu(unsigned int cpu)
1557 {
1558 	if (x86_pmu.cpu_dead)
1559 		x86_pmu.cpu_dead(cpu);
1560 	return 0;
1561 }
1562 
1563 static int x86_pmu_online_cpu(unsigned int cpu)
1564 {
1565 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1566 	int i;
1567 
1568 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1569 		kfree(cpuc->kfree_on_online[i]);
1570 		cpuc->kfree_on_online[i] = NULL;
1571 	}
1572 	return 0;
1573 }
1574 
1575 static int x86_pmu_starting_cpu(unsigned int cpu)
1576 {
1577 	if (x86_pmu.cpu_starting)
1578 		x86_pmu.cpu_starting(cpu);
1579 	return 0;
1580 }
1581 
1582 static int x86_pmu_dying_cpu(unsigned int cpu)
1583 {
1584 	if (x86_pmu.cpu_dying)
1585 		x86_pmu.cpu_dying(cpu);
1586 	return 0;
1587 }
1588 
1589 static void __init pmu_check_apic(void)
1590 {
1591 	if (boot_cpu_has(X86_FEATURE_APIC))
1592 		return;
1593 
1594 	x86_pmu.apic = 0;
1595 	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1596 	pr_info("no hardware sampling interrupt available.\n");
1597 
1598 	/*
1599 	 * If we have a PMU initialized but no APIC
1600 	 * interrupts, we cannot sample hardware
1601 	 * events (user-space has to fall back and
1602 	 * sample via a hrtimer based software event):
1603 	 */
1604 	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1605 
1606 }
1607 
1608 static struct attribute_group x86_pmu_format_group __ro_after_init = {
1609 	.name = "format",
1610 	.attrs = NULL,
1611 };
1612 
1613 /*
1614  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1615  * out of events_attr attributes.
1616  */
1617 static void __init filter_events(struct attribute **attrs)
1618 {
1619 	struct device_attribute *d;
1620 	struct perf_pmu_events_attr *pmu_attr;
1621 	int offset = 0;
1622 	int i, j;
1623 
1624 	for (i = 0; attrs[i]; i++) {
1625 		d = (struct device_attribute *)attrs[i];
1626 		pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1627 		/* str trumps id */
1628 		if (pmu_attr->event_str)
1629 			continue;
1630 		if (x86_pmu.event_map(i + offset))
1631 			continue;
1632 
1633 		for (j = i; attrs[j]; j++)
1634 			attrs[j] = attrs[j + 1];
1635 
1636 		/* Check the shifted attr. */
1637 		i--;
1638 
1639 		/*
1640 		 * event_map() is index based, the attrs array is organized
1641 		 * by increasing event index. If we shift the events, then
1642 		 * we need to compensate for the event_map(), otherwise
1643 		 * we are looking up the wrong event in the map
1644 		 */
1645 		offset++;
1646 	}
1647 }
1648 
1649 /* Merge two pointer arrays */
1650 __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1651 {
1652 	struct attribute **new;
1653 	int j, i;
1654 
1655 	for (j = 0; a && a[j]; j++)
1656 		;
1657 	for (i = 0; b && b[i]; i++)
1658 		j++;
1659 	j++;
1660 
1661 	new = kmalloc_array(j, sizeof(struct attribute *), GFP_KERNEL);
1662 	if (!new)
1663 		return NULL;
1664 
1665 	j = 0;
1666 	for (i = 0; a && a[i]; i++)
1667 		new[j++] = a[i];
1668 	for (i = 0; b && b[i]; i++)
1669 		new[j++] = b[i];
1670 	new[j] = NULL;
1671 
1672 	return new;
1673 }
1674 
1675 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1676 {
1677 	struct perf_pmu_events_attr *pmu_attr = \
1678 		container_of(attr, struct perf_pmu_events_attr, attr);
1679 	u64 config = x86_pmu.event_map(pmu_attr->id);
1680 
1681 	/* string trumps id */
1682 	if (pmu_attr->event_str)
1683 		return sprintf(page, "%s", pmu_attr->event_str);
1684 
1685 	return x86_pmu.events_sysfs_show(page, config);
1686 }
1687 EXPORT_SYMBOL_GPL(events_sysfs_show);
1688 
1689 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1690 			  char *page)
1691 {
1692 	struct perf_pmu_events_ht_attr *pmu_attr =
1693 		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1694 
1695 	/*
1696 	 * Report conditional events depending on Hyper-Threading.
1697 	 *
1698 	 * This is overly conservative as usually the HT special
1699 	 * handling is not needed if the other CPU thread is idle.
1700 	 *
1701 	 * Note this does not (and cannot) handle the case when thread
1702 	 * siblings are invisible, for example with virtualization
1703 	 * if they are owned by some other guest.  The user tool
1704 	 * has to re-read when a thread sibling gets onlined later.
1705 	 */
1706 	return sprintf(page, "%s",
1707 			topology_max_smt_threads() > 1 ?
1708 			pmu_attr->event_str_ht :
1709 			pmu_attr->event_str_noht);
1710 }
1711 
1712 EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1713 EVENT_ATTR(instructions,		INSTRUCTIONS		);
1714 EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1715 EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1716 EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1717 EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1718 EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1719 EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1720 EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1721 EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1722 
1723 static struct attribute *empty_attrs;
1724 
1725 static struct attribute *events_attr[] = {
1726 	EVENT_PTR(CPU_CYCLES),
1727 	EVENT_PTR(INSTRUCTIONS),
1728 	EVENT_PTR(CACHE_REFERENCES),
1729 	EVENT_PTR(CACHE_MISSES),
1730 	EVENT_PTR(BRANCH_INSTRUCTIONS),
1731 	EVENT_PTR(BRANCH_MISSES),
1732 	EVENT_PTR(BUS_CYCLES),
1733 	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1734 	EVENT_PTR(STALLED_CYCLES_BACKEND),
1735 	EVENT_PTR(REF_CPU_CYCLES),
1736 	NULL,
1737 };
1738 
1739 static struct attribute_group x86_pmu_events_group __ro_after_init = {
1740 	.name = "events",
1741 	.attrs = events_attr,
1742 };
1743 
1744 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1745 {
1746 	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1747 	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1748 	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1749 	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1750 	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1751 	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1752 	ssize_t ret;
1753 
1754 	/*
1755 	* We have whole page size to spend and just little data
1756 	* to write, so we can safely use sprintf.
1757 	*/
1758 	ret = sprintf(page, "event=0x%02llx", event);
1759 
1760 	if (umask)
1761 		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1762 
1763 	if (edge)
1764 		ret += sprintf(page + ret, ",edge");
1765 
1766 	if (pc)
1767 		ret += sprintf(page + ret, ",pc");
1768 
1769 	if (any)
1770 		ret += sprintf(page + ret, ",any");
1771 
1772 	if (inv)
1773 		ret += sprintf(page + ret, ",inv");
1774 
1775 	if (cmask)
1776 		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1777 
1778 	ret += sprintf(page + ret, "\n");
1779 
1780 	return ret;
1781 }
1782 
1783 static struct attribute_group x86_pmu_attr_group;
1784 static struct attribute_group x86_pmu_caps_group;
1785 
1786 static int __init init_hw_perf_events(void)
1787 {
1788 	struct x86_pmu_quirk *quirk;
1789 	int err;
1790 
1791 	pr_info("Performance Events: ");
1792 
1793 	switch (boot_cpu_data.x86_vendor) {
1794 	case X86_VENDOR_INTEL:
1795 		err = intel_pmu_init();
1796 		break;
1797 	case X86_VENDOR_AMD:
1798 		err = amd_pmu_init();
1799 		break;
1800 	case X86_VENDOR_HYGON:
1801 		err = amd_pmu_init();
1802 		x86_pmu.name = "HYGON";
1803 		break;
1804 	default:
1805 		err = -ENOTSUPP;
1806 	}
1807 	if (err != 0) {
1808 		pr_cont("no PMU driver, software events only.\n");
1809 		return 0;
1810 	}
1811 
1812 	pmu_check_apic();
1813 
1814 	/* sanity check that the hardware exists or is emulated */
1815 	if (!check_hw_exists())
1816 		return 0;
1817 
1818 	pr_cont("%s PMU driver.\n", x86_pmu.name);
1819 
1820 	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1821 
1822 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1823 		quirk->func();
1824 
1825 	if (!x86_pmu.intel_ctrl)
1826 		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1827 
1828 	perf_events_lapic_init();
1829 	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1830 
1831 	unconstrained = (struct event_constraint)
1832 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1833 				   0, x86_pmu.num_counters, 0, 0);
1834 
1835 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1836 
1837 	if (x86_pmu.caps_attrs) {
1838 		struct attribute **tmp;
1839 
1840 		tmp = merge_attr(x86_pmu_caps_group.attrs, x86_pmu.caps_attrs);
1841 		if (!WARN_ON(!tmp))
1842 			x86_pmu_caps_group.attrs = tmp;
1843 	}
1844 
1845 	if (x86_pmu.event_attrs)
1846 		x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1847 
1848 	if (!x86_pmu.events_sysfs_show)
1849 		x86_pmu_events_group.attrs = &empty_attrs;
1850 	else
1851 		filter_events(x86_pmu_events_group.attrs);
1852 
1853 	if (x86_pmu.cpu_events) {
1854 		struct attribute **tmp;
1855 
1856 		tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1857 		if (!WARN_ON(!tmp))
1858 			x86_pmu_events_group.attrs = tmp;
1859 	}
1860 
1861 	if (x86_pmu.attrs) {
1862 		struct attribute **tmp;
1863 
1864 		tmp = merge_attr(x86_pmu_attr_group.attrs, x86_pmu.attrs);
1865 		if (!WARN_ON(!tmp))
1866 			x86_pmu_attr_group.attrs = tmp;
1867 	}
1868 
1869 	pr_info("... version:                %d\n",     x86_pmu.version);
1870 	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1871 	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1872 	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1873 	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1874 	pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1875 	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1876 
1877 	/*
1878 	 * Install callbacks. Core will call them for each online
1879 	 * cpu.
1880 	 */
1881 	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
1882 				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1883 	if (err)
1884 		return err;
1885 
1886 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1887 				"perf/x86:starting", x86_pmu_starting_cpu,
1888 				x86_pmu_dying_cpu);
1889 	if (err)
1890 		goto out;
1891 
1892 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
1893 				x86_pmu_online_cpu, NULL);
1894 	if (err)
1895 		goto out1;
1896 
1897 	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1898 	if (err)
1899 		goto out2;
1900 
1901 	return 0;
1902 
1903 out2:
1904 	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1905 out1:
1906 	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1907 out:
1908 	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
1909 	return err;
1910 }
1911 early_initcall(init_hw_perf_events);
1912 
1913 static inline void x86_pmu_read(struct perf_event *event)
1914 {
1915 	if (x86_pmu.read)
1916 		return x86_pmu.read(event);
1917 	x86_perf_event_update(event);
1918 }
1919 
1920 /*
1921  * Start group events scheduling transaction
1922  * Set the flag to make pmu::enable() not perform the
1923  * schedulability test, it will be performed at commit time
1924  *
1925  * We only support PERF_PMU_TXN_ADD transactions. Save the
1926  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1927  * transactions.
1928  */
1929 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1930 {
1931 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1932 
1933 	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
1934 
1935 	cpuc->txn_flags = txn_flags;
1936 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1937 		return;
1938 
1939 	perf_pmu_disable(pmu);
1940 	__this_cpu_write(cpu_hw_events.n_txn, 0);
1941 }
1942 
1943 /*
1944  * Stop group events scheduling transaction
1945  * Clear the flag and pmu::enable() will perform the
1946  * schedulability test.
1947  */
1948 static void x86_pmu_cancel_txn(struct pmu *pmu)
1949 {
1950 	unsigned int txn_flags;
1951 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1952 
1953 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1954 
1955 	txn_flags = cpuc->txn_flags;
1956 	cpuc->txn_flags = 0;
1957 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1958 		return;
1959 
1960 	/*
1961 	 * Truncate collected array by the number of events added in this
1962 	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1963 	 */
1964 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1965 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1966 	perf_pmu_enable(pmu);
1967 }
1968 
1969 /*
1970  * Commit group events scheduling transaction
1971  * Perform the group schedulability test as a whole
1972  * Return 0 if success
1973  *
1974  * Does not cancel the transaction on failure; expects the caller to do this.
1975  */
1976 static int x86_pmu_commit_txn(struct pmu *pmu)
1977 {
1978 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1979 	int assign[X86_PMC_IDX_MAX];
1980 	int n, ret;
1981 
1982 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1983 
1984 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1985 		cpuc->txn_flags = 0;
1986 		return 0;
1987 	}
1988 
1989 	n = cpuc->n_events;
1990 
1991 	if (!x86_pmu_initialized())
1992 		return -EAGAIN;
1993 
1994 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1995 	if (ret)
1996 		return ret;
1997 
1998 	/*
1999 	 * copy new assignment, now we know it is possible
2000 	 * will be used by hw_perf_enable()
2001 	 */
2002 	memcpy(cpuc->assign, assign, n*sizeof(int));
2003 
2004 	cpuc->txn_flags = 0;
2005 	perf_pmu_enable(pmu);
2006 	return 0;
2007 }
2008 /*
2009  * a fake_cpuc is used to validate event groups. Due to
2010  * the extra reg logic, we need to also allocate a fake
2011  * per_core and per_cpu structure. Otherwise, group events
2012  * using extra reg may conflict without the kernel being
2013  * able to catch this when the last event gets added to
2014  * the group.
2015  */
2016 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
2017 {
2018 	kfree(cpuc->shared_regs);
2019 	kfree(cpuc);
2020 }
2021 
2022 static struct cpu_hw_events *allocate_fake_cpuc(void)
2023 {
2024 	struct cpu_hw_events *cpuc;
2025 	int cpu = raw_smp_processor_id();
2026 
2027 	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
2028 	if (!cpuc)
2029 		return ERR_PTR(-ENOMEM);
2030 
2031 	/* only needed, if we have extra_regs */
2032 	if (x86_pmu.extra_regs) {
2033 		cpuc->shared_regs = allocate_shared_regs(cpu);
2034 		if (!cpuc->shared_regs)
2035 			goto error;
2036 	}
2037 	cpuc->is_fake = 1;
2038 	return cpuc;
2039 error:
2040 	free_fake_cpuc(cpuc);
2041 	return ERR_PTR(-ENOMEM);
2042 }
2043 
2044 /*
2045  * validate that we can schedule this event
2046  */
2047 static int validate_event(struct perf_event *event)
2048 {
2049 	struct cpu_hw_events *fake_cpuc;
2050 	struct event_constraint *c;
2051 	int ret = 0;
2052 
2053 	fake_cpuc = allocate_fake_cpuc();
2054 	if (IS_ERR(fake_cpuc))
2055 		return PTR_ERR(fake_cpuc);
2056 
2057 	c = x86_pmu.get_event_constraints(fake_cpuc, -1, event);
2058 
2059 	if (!c || !c->weight)
2060 		ret = -EINVAL;
2061 
2062 	if (x86_pmu.put_event_constraints)
2063 		x86_pmu.put_event_constraints(fake_cpuc, event);
2064 
2065 	free_fake_cpuc(fake_cpuc);
2066 
2067 	return ret;
2068 }
2069 
2070 /*
2071  * validate a single event group
2072  *
2073  * validation include:
2074  *	- check events are compatible which each other
2075  *	- events do not compete for the same counter
2076  *	- number of events <= number of counters
2077  *
2078  * validation ensures the group can be loaded onto the
2079  * PMU if it was the only group available.
2080  */
2081 static int validate_group(struct perf_event *event)
2082 {
2083 	struct perf_event *leader = event->group_leader;
2084 	struct cpu_hw_events *fake_cpuc;
2085 	int ret = -EINVAL, n;
2086 
2087 	fake_cpuc = allocate_fake_cpuc();
2088 	if (IS_ERR(fake_cpuc))
2089 		return PTR_ERR(fake_cpuc);
2090 	/*
2091 	 * the event is not yet connected with its
2092 	 * siblings therefore we must first collect
2093 	 * existing siblings, then add the new event
2094 	 * before we can simulate the scheduling
2095 	 */
2096 	n = collect_events(fake_cpuc, leader, true);
2097 	if (n < 0)
2098 		goto out;
2099 
2100 	fake_cpuc->n_events = n;
2101 	n = collect_events(fake_cpuc, event, false);
2102 	if (n < 0)
2103 		goto out;
2104 
2105 	fake_cpuc->n_events = n;
2106 
2107 	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2108 
2109 out:
2110 	free_fake_cpuc(fake_cpuc);
2111 	return ret;
2112 }
2113 
2114 static int x86_pmu_event_init(struct perf_event *event)
2115 {
2116 	struct pmu *tmp;
2117 	int err;
2118 
2119 	switch (event->attr.type) {
2120 	case PERF_TYPE_RAW:
2121 	case PERF_TYPE_HARDWARE:
2122 	case PERF_TYPE_HW_CACHE:
2123 		break;
2124 
2125 	default:
2126 		return -ENOENT;
2127 	}
2128 
2129 	err = __x86_pmu_event_init(event);
2130 	if (!err) {
2131 		/*
2132 		 * we temporarily connect event to its pmu
2133 		 * such that validate_group() can classify
2134 		 * it as an x86 event using is_x86_event()
2135 		 */
2136 		tmp = event->pmu;
2137 		event->pmu = &pmu;
2138 
2139 		if (event->group_leader != event)
2140 			err = validate_group(event);
2141 		else
2142 			err = validate_event(event);
2143 
2144 		event->pmu = tmp;
2145 	}
2146 	if (err) {
2147 		if (event->destroy)
2148 			event->destroy(event);
2149 	}
2150 
2151 	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2152 	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2153 		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2154 
2155 	return err;
2156 }
2157 
2158 static void refresh_pce(void *ignored)
2159 {
2160 	load_mm_cr4(this_cpu_read(cpu_tlbstate.loaded_mm));
2161 }
2162 
2163 static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2164 {
2165 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2166 		return;
2167 
2168 	/*
2169 	 * This function relies on not being called concurrently in two
2170 	 * tasks in the same mm.  Otherwise one task could observe
2171 	 * perf_rdpmc_allowed > 1 and return all the way back to
2172 	 * userspace with CR4.PCE clear while another task is still
2173 	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2174 	 *
2175 	 * For now, this can't happen because all callers hold mmap_sem
2176 	 * for write.  If this changes, we'll need a different solution.
2177 	 */
2178 	lockdep_assert_held_exclusive(&mm->mmap_sem);
2179 
2180 	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2181 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2182 }
2183 
2184 static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2185 {
2186 
2187 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2188 		return;
2189 
2190 	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2191 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2192 }
2193 
2194 static int x86_pmu_event_idx(struct perf_event *event)
2195 {
2196 	int idx = event->hw.idx;
2197 
2198 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2199 		return 0;
2200 
2201 	if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2202 		idx -= INTEL_PMC_IDX_FIXED;
2203 		idx |= 1 << 30;
2204 	}
2205 
2206 	return idx + 1;
2207 }
2208 
2209 static ssize_t get_attr_rdpmc(struct device *cdev,
2210 			      struct device_attribute *attr,
2211 			      char *buf)
2212 {
2213 	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2214 }
2215 
2216 static ssize_t set_attr_rdpmc(struct device *cdev,
2217 			      struct device_attribute *attr,
2218 			      const char *buf, size_t count)
2219 {
2220 	unsigned long val;
2221 	ssize_t ret;
2222 
2223 	ret = kstrtoul(buf, 0, &val);
2224 	if (ret)
2225 		return ret;
2226 
2227 	if (val > 2)
2228 		return -EINVAL;
2229 
2230 	if (x86_pmu.attr_rdpmc_broken)
2231 		return -ENOTSUPP;
2232 
2233 	if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2234 		/*
2235 		 * Changing into or out of always available, aka
2236 		 * perf-event-bypassing mode.  This path is extremely slow,
2237 		 * but only root can trigger it, so it's okay.
2238 		 */
2239 		if (val == 2)
2240 			static_branch_inc(&rdpmc_always_available_key);
2241 		else
2242 			static_branch_dec(&rdpmc_always_available_key);
2243 		on_each_cpu(refresh_pce, NULL, 1);
2244 	}
2245 
2246 	x86_pmu.attr_rdpmc = val;
2247 
2248 	return count;
2249 }
2250 
2251 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2252 
2253 static struct attribute *x86_pmu_attrs[] = {
2254 	&dev_attr_rdpmc.attr,
2255 	NULL,
2256 };
2257 
2258 static struct attribute_group x86_pmu_attr_group __ro_after_init = {
2259 	.attrs = x86_pmu_attrs,
2260 };
2261 
2262 static ssize_t max_precise_show(struct device *cdev,
2263 				  struct device_attribute *attr,
2264 				  char *buf)
2265 {
2266 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2267 }
2268 
2269 static DEVICE_ATTR_RO(max_precise);
2270 
2271 static struct attribute *x86_pmu_caps_attrs[] = {
2272 	&dev_attr_max_precise.attr,
2273 	NULL
2274 };
2275 
2276 static struct attribute_group x86_pmu_caps_group __ro_after_init = {
2277 	.name = "caps",
2278 	.attrs = x86_pmu_caps_attrs,
2279 };
2280 
2281 static const struct attribute_group *x86_pmu_attr_groups[] = {
2282 	&x86_pmu_attr_group,
2283 	&x86_pmu_format_group,
2284 	&x86_pmu_events_group,
2285 	&x86_pmu_caps_group,
2286 	NULL,
2287 };
2288 
2289 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2290 {
2291 	if (x86_pmu.sched_task)
2292 		x86_pmu.sched_task(ctx, sched_in);
2293 }
2294 
2295 void perf_check_microcode(void)
2296 {
2297 	if (x86_pmu.check_microcode)
2298 		x86_pmu.check_microcode();
2299 }
2300 
2301 static struct pmu pmu = {
2302 	.pmu_enable		= x86_pmu_enable,
2303 	.pmu_disable		= x86_pmu_disable,
2304 
2305 	.attr_groups		= x86_pmu_attr_groups,
2306 
2307 	.event_init		= x86_pmu_event_init,
2308 
2309 	.event_mapped		= x86_pmu_event_mapped,
2310 	.event_unmapped		= x86_pmu_event_unmapped,
2311 
2312 	.add			= x86_pmu_add,
2313 	.del			= x86_pmu_del,
2314 	.start			= x86_pmu_start,
2315 	.stop			= x86_pmu_stop,
2316 	.read			= x86_pmu_read,
2317 
2318 	.start_txn		= x86_pmu_start_txn,
2319 	.cancel_txn		= x86_pmu_cancel_txn,
2320 	.commit_txn		= x86_pmu_commit_txn,
2321 
2322 	.event_idx		= x86_pmu_event_idx,
2323 	.sched_task		= x86_pmu_sched_task,
2324 	.task_ctx_size          = sizeof(struct x86_perf_task_context),
2325 };
2326 
2327 void arch_perf_update_userpage(struct perf_event *event,
2328 			       struct perf_event_mmap_page *userpg, u64 now)
2329 {
2330 	struct cyc2ns_data data;
2331 	u64 offset;
2332 
2333 	userpg->cap_user_time = 0;
2334 	userpg->cap_user_time_zero = 0;
2335 	userpg->cap_user_rdpmc =
2336 		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2337 	userpg->pmc_width = x86_pmu.cntval_bits;
2338 
2339 	if (!using_native_sched_clock() || !sched_clock_stable())
2340 		return;
2341 
2342 	cyc2ns_read_begin(&data);
2343 
2344 	offset = data.cyc2ns_offset + __sched_clock_offset;
2345 
2346 	/*
2347 	 * Internal timekeeping for enabled/running/stopped times
2348 	 * is always in the local_clock domain.
2349 	 */
2350 	userpg->cap_user_time = 1;
2351 	userpg->time_mult = data.cyc2ns_mul;
2352 	userpg->time_shift = data.cyc2ns_shift;
2353 	userpg->time_offset = offset - now;
2354 
2355 	/*
2356 	 * cap_user_time_zero doesn't make sense when we're using a different
2357 	 * time base for the records.
2358 	 */
2359 	if (!event->attr.use_clockid) {
2360 		userpg->cap_user_time_zero = 1;
2361 		userpg->time_zero = offset;
2362 	}
2363 
2364 	cyc2ns_read_end();
2365 }
2366 
2367 void
2368 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2369 {
2370 	struct unwind_state state;
2371 	unsigned long addr;
2372 
2373 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2374 		/* TODO: We don't support guest os callchain now */
2375 		return;
2376 	}
2377 
2378 	if (perf_callchain_store(entry, regs->ip))
2379 		return;
2380 
2381 	for (unwind_start(&state, current, regs, NULL); !unwind_done(&state);
2382 	     unwind_next_frame(&state)) {
2383 		addr = unwind_get_return_address(&state);
2384 		if (!addr || perf_callchain_store(entry, addr))
2385 			return;
2386 	}
2387 }
2388 
2389 static inline int
2390 valid_user_frame(const void __user *fp, unsigned long size)
2391 {
2392 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2393 }
2394 
2395 static unsigned long get_segment_base(unsigned int segment)
2396 {
2397 	struct desc_struct *desc;
2398 	unsigned int idx = segment >> 3;
2399 
2400 	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2401 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2402 		struct ldt_struct *ldt;
2403 
2404 		/* IRQs are off, so this synchronizes with smp_store_release */
2405 		ldt = READ_ONCE(current->active_mm->context.ldt);
2406 		if (!ldt || idx >= ldt->nr_entries)
2407 			return 0;
2408 
2409 		desc = &ldt->entries[idx];
2410 #else
2411 		return 0;
2412 #endif
2413 	} else {
2414 		if (idx >= GDT_ENTRIES)
2415 			return 0;
2416 
2417 		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2418 	}
2419 
2420 	return get_desc_base(desc);
2421 }
2422 
2423 #ifdef CONFIG_IA32_EMULATION
2424 
2425 #include <linux/compat.h>
2426 
2427 static inline int
2428 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2429 {
2430 	/* 32-bit process in 64-bit kernel. */
2431 	unsigned long ss_base, cs_base;
2432 	struct stack_frame_ia32 frame;
2433 	const void __user *fp;
2434 
2435 	if (!test_thread_flag(TIF_IA32))
2436 		return 0;
2437 
2438 	cs_base = get_segment_base(regs->cs);
2439 	ss_base = get_segment_base(regs->ss);
2440 
2441 	fp = compat_ptr(ss_base + regs->bp);
2442 	pagefault_disable();
2443 	while (entry->nr < entry->max_stack) {
2444 		unsigned long bytes;
2445 		frame.next_frame     = 0;
2446 		frame.return_address = 0;
2447 
2448 		if (!valid_user_frame(fp, sizeof(frame)))
2449 			break;
2450 
2451 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2452 		if (bytes != 0)
2453 			break;
2454 		bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2455 		if (bytes != 0)
2456 			break;
2457 
2458 		perf_callchain_store(entry, cs_base + frame.return_address);
2459 		fp = compat_ptr(ss_base + frame.next_frame);
2460 	}
2461 	pagefault_enable();
2462 	return 1;
2463 }
2464 #else
2465 static inline int
2466 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2467 {
2468     return 0;
2469 }
2470 #endif
2471 
2472 void
2473 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2474 {
2475 	struct stack_frame frame;
2476 	const unsigned long __user *fp;
2477 
2478 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2479 		/* TODO: We don't support guest os callchain now */
2480 		return;
2481 	}
2482 
2483 	/*
2484 	 * We don't know what to do with VM86 stacks.. ignore them for now.
2485 	 */
2486 	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2487 		return;
2488 
2489 	fp = (unsigned long __user *)regs->bp;
2490 
2491 	perf_callchain_store(entry, regs->ip);
2492 
2493 	if (!nmi_uaccess_okay())
2494 		return;
2495 
2496 	if (perf_callchain_user32(regs, entry))
2497 		return;
2498 
2499 	pagefault_disable();
2500 	while (entry->nr < entry->max_stack) {
2501 		unsigned long bytes;
2502 
2503 		frame.next_frame	     = NULL;
2504 		frame.return_address = 0;
2505 
2506 		if (!valid_user_frame(fp, sizeof(frame)))
2507 			break;
2508 
2509 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2510 		if (bytes != 0)
2511 			break;
2512 		bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2513 		if (bytes != 0)
2514 			break;
2515 
2516 		perf_callchain_store(entry, frame.return_address);
2517 		fp = (void __user *)frame.next_frame;
2518 	}
2519 	pagefault_enable();
2520 }
2521 
2522 /*
2523  * Deal with code segment offsets for the various execution modes:
2524  *
2525  *   VM86 - the good olde 16 bit days, where the linear address is
2526  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2527  *
2528  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2529  *          to figure out what the 32bit base address is.
2530  *
2531  *    X32 - has TIF_X32 set, but is running in x86_64
2532  *
2533  * X86_64 - CS,DS,SS,ES are all zero based.
2534  */
2535 static unsigned long code_segment_base(struct pt_regs *regs)
2536 {
2537 	/*
2538 	 * For IA32 we look at the GDT/LDT segment base to convert the
2539 	 * effective IP to a linear address.
2540 	 */
2541 
2542 #ifdef CONFIG_X86_32
2543 	/*
2544 	 * If we are in VM86 mode, add the segment offset to convert to a
2545 	 * linear address.
2546 	 */
2547 	if (regs->flags & X86_VM_MASK)
2548 		return 0x10 * regs->cs;
2549 
2550 	if (user_mode(regs) && regs->cs != __USER_CS)
2551 		return get_segment_base(regs->cs);
2552 #else
2553 	if (user_mode(regs) && !user_64bit_mode(regs) &&
2554 	    regs->cs != __USER32_CS)
2555 		return get_segment_base(regs->cs);
2556 #endif
2557 	return 0;
2558 }
2559 
2560 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2561 {
2562 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2563 		return perf_guest_cbs->get_guest_ip();
2564 
2565 	return regs->ip + code_segment_base(regs);
2566 }
2567 
2568 unsigned long perf_misc_flags(struct pt_regs *regs)
2569 {
2570 	int misc = 0;
2571 
2572 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2573 		if (perf_guest_cbs->is_user_mode())
2574 			misc |= PERF_RECORD_MISC_GUEST_USER;
2575 		else
2576 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2577 	} else {
2578 		if (user_mode(regs))
2579 			misc |= PERF_RECORD_MISC_USER;
2580 		else
2581 			misc |= PERF_RECORD_MISC_KERNEL;
2582 	}
2583 
2584 	if (regs->flags & PERF_EFLAGS_EXACT)
2585 		misc |= PERF_RECORD_MISC_EXACT_IP;
2586 
2587 	return misc;
2588 }
2589 
2590 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2591 {
2592 	cap->version		= x86_pmu.version;
2593 	cap->num_counters_gp	= x86_pmu.num_counters;
2594 	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2595 	cap->bit_width_gp	= x86_pmu.cntval_bits;
2596 	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2597 	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2598 	cap->events_mask_len	= x86_pmu.events_mask_len;
2599 }
2600 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
2601