xref: /openbmc/linux/arch/x86/events/core.c (revision 0e17c50f)
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_never_available_key);
53 DEFINE_STATIC_KEY_FALSE(rdpmc_always_available_key);
54 
55 u64 __read_mostly hw_cache_event_ids
56 				[PERF_COUNT_HW_CACHE_MAX]
57 				[PERF_COUNT_HW_CACHE_OP_MAX]
58 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
59 u64 __read_mostly hw_cache_extra_regs
60 				[PERF_COUNT_HW_CACHE_MAX]
61 				[PERF_COUNT_HW_CACHE_OP_MAX]
62 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
63 
64 /*
65  * Propagate event elapsed time into the generic event.
66  * Can only be executed on the CPU where the event is active.
67  * Returns the delta events processed.
68  */
69 u64 x86_perf_event_update(struct perf_event *event)
70 {
71 	struct hw_perf_event *hwc = &event->hw;
72 	int shift = 64 - x86_pmu.cntval_bits;
73 	u64 prev_raw_count, new_raw_count;
74 	int idx = hwc->idx;
75 	u64 delta;
76 
77 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
78 		return 0;
79 
80 	/*
81 	 * Careful: an NMI might modify the previous event value.
82 	 *
83 	 * Our tactic to handle this is to first atomically read and
84 	 * exchange a new raw count - then add that new-prev delta
85 	 * count to the generic event atomically:
86 	 */
87 again:
88 	prev_raw_count = local64_read(&hwc->prev_count);
89 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
90 
91 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
92 					new_raw_count) != prev_raw_count)
93 		goto again;
94 
95 	/*
96 	 * Now we have the new raw value and have updated the prev
97 	 * timestamp already. We can now calculate the elapsed delta
98 	 * (event-)time and add that to the generic event.
99 	 *
100 	 * Careful, not all hw sign-extends above the physical width
101 	 * of the count.
102 	 */
103 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
104 	delta >>= shift;
105 
106 	local64_add(delta, &event->count);
107 	local64_sub(delta, &hwc->period_left);
108 
109 	return new_raw_count;
110 }
111 
112 /*
113  * Find and validate any extra registers to set up.
114  */
115 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
116 {
117 	struct hw_perf_event_extra *reg;
118 	struct extra_reg *er;
119 
120 	reg = &event->hw.extra_reg;
121 
122 	if (!x86_pmu.extra_regs)
123 		return 0;
124 
125 	for (er = x86_pmu.extra_regs; er->msr; er++) {
126 		if (er->event != (config & er->config_mask))
127 			continue;
128 		if (event->attr.config1 & ~er->valid_mask)
129 			return -EINVAL;
130 		/* Check if the extra msrs can be safely accessed*/
131 		if (!er->extra_msr_access)
132 			return -ENXIO;
133 
134 		reg->idx = er->idx;
135 		reg->config = event->attr.config1;
136 		reg->reg = er->msr;
137 		break;
138 	}
139 	return 0;
140 }
141 
142 static atomic_t active_events;
143 static atomic_t pmc_refcount;
144 static DEFINE_MUTEX(pmc_reserve_mutex);
145 
146 #ifdef CONFIG_X86_LOCAL_APIC
147 
148 static bool reserve_pmc_hardware(void)
149 {
150 	int i;
151 
152 	for (i = 0; i < x86_pmu.num_counters; i++) {
153 		if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
154 			goto perfctr_fail;
155 	}
156 
157 	for (i = 0; i < x86_pmu.num_counters; i++) {
158 		if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
159 			goto eventsel_fail;
160 	}
161 
162 	return true;
163 
164 eventsel_fail:
165 	for (i--; i >= 0; i--)
166 		release_evntsel_nmi(x86_pmu_config_addr(i));
167 
168 	i = x86_pmu.num_counters;
169 
170 perfctr_fail:
171 	for (i--; i >= 0; i--)
172 		release_perfctr_nmi(x86_pmu_event_addr(i));
173 
174 	return false;
175 }
176 
177 static void release_pmc_hardware(void)
178 {
179 	int i;
180 
181 	for (i = 0; i < x86_pmu.num_counters; i++) {
182 		release_perfctr_nmi(x86_pmu_event_addr(i));
183 		release_evntsel_nmi(x86_pmu_config_addr(i));
184 	}
185 }
186 
187 #else
188 
189 static bool reserve_pmc_hardware(void) { return true; }
190 static void release_pmc_hardware(void) {}
191 
192 #endif
193 
194 static bool check_hw_exists(void)
195 {
196 	u64 val, val_fail = -1, val_new= ~0;
197 	int i, reg, reg_fail = -1, ret = 0;
198 	int bios_fail = 0;
199 	int reg_safe = -1;
200 
201 	/*
202 	 * Check to see if the BIOS enabled any of the counters, if so
203 	 * complain and bail.
204 	 */
205 	for (i = 0; i < x86_pmu.num_counters; i++) {
206 		reg = x86_pmu_config_addr(i);
207 		ret = rdmsrl_safe(reg, &val);
208 		if (ret)
209 			goto msr_fail;
210 		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
211 			bios_fail = 1;
212 			val_fail = val;
213 			reg_fail = reg;
214 		} else {
215 			reg_safe = i;
216 		}
217 	}
218 
219 	if (x86_pmu.num_counters_fixed) {
220 		reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
221 		ret = rdmsrl_safe(reg, &val);
222 		if (ret)
223 			goto msr_fail;
224 		for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
225 			if (val & (0x03 << i*4)) {
226 				bios_fail = 1;
227 				val_fail = val;
228 				reg_fail = reg;
229 			}
230 		}
231 	}
232 
233 	/*
234 	 * If all the counters are enabled, the below test will always
235 	 * fail.  The tools will also become useless in this scenario.
236 	 * Just fail and disable the hardware counters.
237 	 */
238 
239 	if (reg_safe == -1) {
240 		reg = reg_safe;
241 		goto msr_fail;
242 	}
243 
244 	/*
245 	 * Read the current value, change it and read it back to see if it
246 	 * matches, this is needed to detect certain hardware emulators
247 	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
248 	 */
249 	reg = x86_pmu_event_addr(reg_safe);
250 	if (rdmsrl_safe(reg, &val))
251 		goto msr_fail;
252 	val ^= 0xffffUL;
253 	ret = wrmsrl_safe(reg, val);
254 	ret |= rdmsrl_safe(reg, &val_new);
255 	if (ret || val != val_new)
256 		goto msr_fail;
257 
258 	/*
259 	 * We still allow the PMU driver to operate:
260 	 */
261 	if (bios_fail) {
262 		pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
263 		pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
264 			      reg_fail, val_fail);
265 	}
266 
267 	return true;
268 
269 msr_fail:
270 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
271 		pr_cont("PMU not available due to virtualization, using software events only.\n");
272 	} else {
273 		pr_cont("Broken PMU hardware detected, using software events only.\n");
274 		pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
275 		       reg, val_new);
276 	}
277 
278 	return false;
279 }
280 
281 static void hw_perf_event_destroy(struct perf_event *event)
282 {
283 	x86_release_hardware();
284 	atomic_dec(&active_events);
285 }
286 
287 void hw_perf_lbr_event_destroy(struct perf_event *event)
288 {
289 	hw_perf_event_destroy(event);
290 
291 	/* undo the lbr/bts event accounting */
292 	x86_del_exclusive(x86_lbr_exclusive_lbr);
293 }
294 
295 static inline int x86_pmu_initialized(void)
296 {
297 	return x86_pmu.handle_irq != NULL;
298 }
299 
300 static inline int
301 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
302 {
303 	struct perf_event_attr *attr = &event->attr;
304 	unsigned int cache_type, cache_op, cache_result;
305 	u64 config, val;
306 
307 	config = attr->config;
308 
309 	cache_type = (config >> 0) & 0xff;
310 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
311 		return -EINVAL;
312 	cache_type = array_index_nospec(cache_type, PERF_COUNT_HW_CACHE_MAX);
313 
314 	cache_op = (config >>  8) & 0xff;
315 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
316 		return -EINVAL;
317 	cache_op = array_index_nospec(cache_op, PERF_COUNT_HW_CACHE_OP_MAX);
318 
319 	cache_result = (config >> 16) & 0xff;
320 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
321 		return -EINVAL;
322 	cache_result = array_index_nospec(cache_result, PERF_COUNT_HW_CACHE_RESULT_MAX);
323 
324 	val = hw_cache_event_ids[cache_type][cache_op][cache_result];
325 
326 	if (val == 0)
327 		return -ENOENT;
328 
329 	if (val == -1)
330 		return -EINVAL;
331 
332 	hwc->config |= val;
333 	attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
334 	return x86_pmu_extra_regs(val, event);
335 }
336 
337 int x86_reserve_hardware(void)
338 {
339 	int err = 0;
340 
341 	if (!atomic_inc_not_zero(&pmc_refcount)) {
342 		mutex_lock(&pmc_reserve_mutex);
343 		if (atomic_read(&pmc_refcount) == 0) {
344 			if (!reserve_pmc_hardware())
345 				err = -EBUSY;
346 			else
347 				reserve_ds_buffers();
348 		}
349 		if (!err)
350 			atomic_inc(&pmc_refcount);
351 		mutex_unlock(&pmc_reserve_mutex);
352 	}
353 
354 	return err;
355 }
356 
357 void x86_release_hardware(void)
358 {
359 	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
360 		release_pmc_hardware();
361 		release_ds_buffers();
362 		mutex_unlock(&pmc_reserve_mutex);
363 	}
364 }
365 
366 /*
367  * Check if we can create event of a certain type (that no conflicting events
368  * are present).
369  */
370 int x86_add_exclusive(unsigned int what)
371 {
372 	int i;
373 
374 	/*
375 	 * When lbr_pt_coexist we allow PT to coexist with either LBR or BTS.
376 	 * LBR and BTS are still mutually exclusive.
377 	 */
378 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
379 		goto out;
380 
381 	if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
382 		mutex_lock(&pmc_reserve_mutex);
383 		for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
384 			if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
385 				goto fail_unlock;
386 		}
387 		atomic_inc(&x86_pmu.lbr_exclusive[what]);
388 		mutex_unlock(&pmc_reserve_mutex);
389 	}
390 
391 out:
392 	atomic_inc(&active_events);
393 	return 0;
394 
395 fail_unlock:
396 	mutex_unlock(&pmc_reserve_mutex);
397 	return -EBUSY;
398 }
399 
400 void x86_del_exclusive(unsigned int what)
401 {
402 	atomic_dec(&active_events);
403 
404 	/*
405 	 * See the comment in x86_add_exclusive().
406 	 */
407 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
408 		return;
409 
410 	atomic_dec(&x86_pmu.lbr_exclusive[what]);
411 }
412 
413 int x86_setup_perfctr(struct perf_event *event)
414 {
415 	struct perf_event_attr *attr = &event->attr;
416 	struct hw_perf_event *hwc = &event->hw;
417 	u64 config;
418 
419 	if (!is_sampling_event(event)) {
420 		hwc->sample_period = x86_pmu.max_period;
421 		hwc->last_period = hwc->sample_period;
422 		local64_set(&hwc->period_left, hwc->sample_period);
423 	}
424 
425 	if (attr->type == PERF_TYPE_RAW)
426 		return x86_pmu_extra_regs(event->attr.config, event);
427 
428 	if (attr->type == PERF_TYPE_HW_CACHE)
429 		return set_ext_hw_attr(hwc, event);
430 
431 	if (attr->config >= x86_pmu.max_events)
432 		return -EINVAL;
433 
434 	attr->config = array_index_nospec((unsigned long)attr->config, x86_pmu.max_events);
435 
436 	/*
437 	 * The generic map:
438 	 */
439 	config = x86_pmu.event_map(attr->config);
440 
441 	if (config == 0)
442 		return -ENOENT;
443 
444 	if (config == -1LL)
445 		return -EINVAL;
446 
447 	hwc->config |= config;
448 
449 	return 0;
450 }
451 
452 /*
453  * check that branch_sample_type is compatible with
454  * settings needed for precise_ip > 1 which implies
455  * using the LBR to capture ALL taken branches at the
456  * priv levels of the measurement
457  */
458 static inline int precise_br_compat(struct perf_event *event)
459 {
460 	u64 m = event->attr.branch_sample_type;
461 	u64 b = 0;
462 
463 	/* must capture all branches */
464 	if (!(m & PERF_SAMPLE_BRANCH_ANY))
465 		return 0;
466 
467 	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
468 
469 	if (!event->attr.exclude_user)
470 		b |= PERF_SAMPLE_BRANCH_USER;
471 
472 	if (!event->attr.exclude_kernel)
473 		b |= PERF_SAMPLE_BRANCH_KERNEL;
474 
475 	/*
476 	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
477 	 */
478 
479 	return m == b;
480 }
481 
482 int x86_pmu_max_precise(void)
483 {
484 	int precise = 0;
485 
486 	/* Support for constant skid */
487 	if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
488 		precise++;
489 
490 		/* Support for IP fixup */
491 		if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
492 			precise++;
493 
494 		if (x86_pmu.pebs_prec_dist)
495 			precise++;
496 	}
497 	return precise;
498 }
499 
500 int x86_pmu_hw_config(struct perf_event *event)
501 {
502 	if (event->attr.precise_ip) {
503 		int precise = x86_pmu_max_precise();
504 
505 		if (event->attr.precise_ip > precise)
506 			return -EOPNOTSUPP;
507 
508 		/* There's no sense in having PEBS for non sampling events: */
509 		if (!is_sampling_event(event))
510 			return -EINVAL;
511 	}
512 	/*
513 	 * check that PEBS LBR correction does not conflict with
514 	 * whatever the user is asking with attr->branch_sample_type
515 	 */
516 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
517 		u64 *br_type = &event->attr.branch_sample_type;
518 
519 		if (has_branch_stack(event)) {
520 			if (!precise_br_compat(event))
521 				return -EOPNOTSUPP;
522 
523 			/* branch_sample_type is compatible */
524 
525 		} else {
526 			/*
527 			 * user did not specify  branch_sample_type
528 			 *
529 			 * For PEBS fixups, we capture all
530 			 * the branches at the priv level of the
531 			 * event.
532 			 */
533 			*br_type = PERF_SAMPLE_BRANCH_ANY;
534 
535 			if (!event->attr.exclude_user)
536 				*br_type |= PERF_SAMPLE_BRANCH_USER;
537 
538 			if (!event->attr.exclude_kernel)
539 				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
540 		}
541 	}
542 
543 	if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
544 		event->attach_state |= PERF_ATTACH_TASK_DATA;
545 
546 	/*
547 	 * Generate PMC IRQs:
548 	 * (keep 'enabled' bit clear for now)
549 	 */
550 	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
551 
552 	/*
553 	 * Count user and OS events unless requested not to
554 	 */
555 	if (!event->attr.exclude_user)
556 		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
557 	if (!event->attr.exclude_kernel)
558 		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
559 
560 	if (event->attr.type == PERF_TYPE_RAW)
561 		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
562 
563 	if (event->attr.sample_period && x86_pmu.limit_period) {
564 		if (x86_pmu.limit_period(event, event->attr.sample_period) >
565 				event->attr.sample_period)
566 			return -EINVAL;
567 	}
568 
569 	/* sample_regs_user never support XMM registers */
570 	if (unlikely(event->attr.sample_regs_user & PERF_REG_EXTENDED_MASK))
571 		return -EINVAL;
572 	/*
573 	 * Besides the general purpose registers, XMM registers may
574 	 * be collected in PEBS on some platforms, e.g. Icelake
575 	 */
576 	if (unlikely(event->attr.sample_regs_intr & PERF_REG_EXTENDED_MASK)) {
577 		if (!(event->pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS))
578 			return -EINVAL;
579 
580 		if (!event->attr.precise_ip)
581 			return -EINVAL;
582 	}
583 
584 	return x86_setup_perfctr(event);
585 }
586 
587 /*
588  * Setup the hardware configuration for a given attr_type
589  */
590 static int __x86_pmu_event_init(struct perf_event *event)
591 {
592 	int err;
593 
594 	if (!x86_pmu_initialized())
595 		return -ENODEV;
596 
597 	err = x86_reserve_hardware();
598 	if (err)
599 		return err;
600 
601 	atomic_inc(&active_events);
602 	event->destroy = hw_perf_event_destroy;
603 
604 	event->hw.idx = -1;
605 	event->hw.last_cpu = -1;
606 	event->hw.last_tag = ~0ULL;
607 
608 	/* mark unused */
609 	event->hw.extra_reg.idx = EXTRA_REG_NONE;
610 	event->hw.branch_reg.idx = EXTRA_REG_NONE;
611 
612 	return x86_pmu.hw_config(event);
613 }
614 
615 void x86_pmu_disable_all(void)
616 {
617 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
618 	int idx;
619 
620 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
621 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
622 		u64 val;
623 
624 		if (!test_bit(idx, cpuc->active_mask))
625 			continue;
626 		rdmsrl(x86_pmu_config_addr(idx), val);
627 		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
628 			continue;
629 		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
630 		wrmsrl(x86_pmu_config_addr(idx), val);
631 		if (is_counter_pair(hwc))
632 			wrmsrl(x86_pmu_config_addr(idx + 1), 0);
633 	}
634 }
635 
636 /*
637  * There may be PMI landing after enabled=0. The PMI hitting could be before or
638  * after disable_all.
639  *
640  * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
641  * It will not be re-enabled in the NMI handler again, because enabled=0. After
642  * handling the NMI, disable_all will be called, which will not change the
643  * state either. If PMI hits after disable_all, the PMU is already disabled
644  * before entering NMI handler. The NMI handler will not change the state
645  * either.
646  *
647  * So either situation is harmless.
648  */
649 static void x86_pmu_disable(struct pmu *pmu)
650 {
651 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
652 
653 	if (!x86_pmu_initialized())
654 		return;
655 
656 	if (!cpuc->enabled)
657 		return;
658 
659 	cpuc->n_added = 0;
660 	cpuc->enabled = 0;
661 	barrier();
662 
663 	x86_pmu.disable_all();
664 }
665 
666 void x86_pmu_enable_all(int added)
667 {
668 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
669 	int idx;
670 
671 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
672 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
673 
674 		if (!test_bit(idx, cpuc->active_mask))
675 			continue;
676 
677 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
678 	}
679 }
680 
681 static struct pmu pmu;
682 
683 static inline int is_x86_event(struct perf_event *event)
684 {
685 	return event->pmu == &pmu;
686 }
687 
688 struct pmu *x86_get_pmu(void)
689 {
690 	return &pmu;
691 }
692 /*
693  * Event scheduler state:
694  *
695  * Assign events iterating over all events and counters, beginning
696  * with events with least weights first. Keep the current iterator
697  * state in struct sched_state.
698  */
699 struct sched_state {
700 	int	weight;
701 	int	event;		/* event index */
702 	int	counter;	/* counter index */
703 	int	unassigned;	/* number of events to be assigned left */
704 	int	nr_gp;		/* number of GP counters used */
705 	u64	used;
706 };
707 
708 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
709 #define	SCHED_STATES_MAX	2
710 
711 struct perf_sched {
712 	int			max_weight;
713 	int			max_events;
714 	int			max_gp;
715 	int			saved_states;
716 	struct event_constraint	**constraints;
717 	struct sched_state	state;
718 	struct sched_state	saved[SCHED_STATES_MAX];
719 };
720 
721 /*
722  * Initialize interator that runs through all events and counters.
723  */
724 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
725 			    int num, int wmin, int wmax, int gpmax)
726 {
727 	int idx;
728 
729 	memset(sched, 0, sizeof(*sched));
730 	sched->max_events	= num;
731 	sched->max_weight	= wmax;
732 	sched->max_gp		= gpmax;
733 	sched->constraints	= constraints;
734 
735 	for (idx = 0; idx < num; idx++) {
736 		if (constraints[idx]->weight == wmin)
737 			break;
738 	}
739 
740 	sched->state.event	= idx;		/* start with min weight */
741 	sched->state.weight	= wmin;
742 	sched->state.unassigned	= num;
743 }
744 
745 static void perf_sched_save_state(struct perf_sched *sched)
746 {
747 	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
748 		return;
749 
750 	sched->saved[sched->saved_states] = sched->state;
751 	sched->saved_states++;
752 }
753 
754 static bool perf_sched_restore_state(struct perf_sched *sched)
755 {
756 	if (!sched->saved_states)
757 		return false;
758 
759 	sched->saved_states--;
760 	sched->state = sched->saved[sched->saved_states];
761 
762 	/* this assignment didn't work out */
763 	/* XXX broken vs EVENT_PAIR */
764 	sched->state.used &= ~BIT_ULL(sched->state.counter);
765 
766 	/* try the next one */
767 	sched->state.counter++;
768 
769 	return true;
770 }
771 
772 /*
773  * Select a counter for the current event to schedule. Return true on
774  * success.
775  */
776 static bool __perf_sched_find_counter(struct perf_sched *sched)
777 {
778 	struct event_constraint *c;
779 	int idx;
780 
781 	if (!sched->state.unassigned)
782 		return false;
783 
784 	if (sched->state.event >= sched->max_events)
785 		return false;
786 
787 	c = sched->constraints[sched->state.event];
788 	/* Prefer fixed purpose counters */
789 	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
790 		idx = INTEL_PMC_IDX_FIXED;
791 		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
792 			u64 mask = BIT_ULL(idx);
793 
794 			if (sched->state.used & mask)
795 				continue;
796 
797 			sched->state.used |= mask;
798 			goto done;
799 		}
800 	}
801 
802 	/* Grab the first unused counter starting with idx */
803 	idx = sched->state.counter;
804 	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
805 		u64 mask = BIT_ULL(idx);
806 
807 		if (c->flags & PERF_X86_EVENT_PAIR)
808 			mask |= mask << 1;
809 
810 		if (sched->state.used & mask)
811 			continue;
812 
813 		if (sched->state.nr_gp++ >= sched->max_gp)
814 			return false;
815 
816 		sched->state.used |= mask;
817 		goto done;
818 	}
819 
820 	return false;
821 
822 done:
823 	sched->state.counter = idx;
824 
825 	if (c->overlap)
826 		perf_sched_save_state(sched);
827 
828 	return true;
829 }
830 
831 static bool perf_sched_find_counter(struct perf_sched *sched)
832 {
833 	while (!__perf_sched_find_counter(sched)) {
834 		if (!perf_sched_restore_state(sched))
835 			return false;
836 	}
837 
838 	return true;
839 }
840 
841 /*
842  * Go through all unassigned events and find the next one to schedule.
843  * Take events with the least weight first. Return true on success.
844  */
845 static bool perf_sched_next_event(struct perf_sched *sched)
846 {
847 	struct event_constraint *c;
848 
849 	if (!sched->state.unassigned || !--sched->state.unassigned)
850 		return false;
851 
852 	do {
853 		/* next event */
854 		sched->state.event++;
855 		if (sched->state.event >= sched->max_events) {
856 			/* next weight */
857 			sched->state.event = 0;
858 			sched->state.weight++;
859 			if (sched->state.weight > sched->max_weight)
860 				return false;
861 		}
862 		c = sched->constraints[sched->state.event];
863 	} while (c->weight != sched->state.weight);
864 
865 	sched->state.counter = 0;	/* start with first counter */
866 
867 	return true;
868 }
869 
870 /*
871  * Assign a counter for each event.
872  */
873 int perf_assign_events(struct event_constraint **constraints, int n,
874 			int wmin, int wmax, int gpmax, int *assign)
875 {
876 	struct perf_sched sched;
877 
878 	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
879 
880 	do {
881 		if (!perf_sched_find_counter(&sched))
882 			break;	/* failed */
883 		if (assign)
884 			assign[sched.state.event] = sched.state.counter;
885 	} while (perf_sched_next_event(&sched));
886 
887 	return sched.state.unassigned;
888 }
889 EXPORT_SYMBOL_GPL(perf_assign_events);
890 
891 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
892 {
893 	struct event_constraint *c;
894 	struct perf_event *e;
895 	int n0, i, wmin, wmax, unsched = 0;
896 	struct hw_perf_event *hwc;
897 	u64 used_mask = 0;
898 
899 	/*
900 	 * Compute the number of events already present; see x86_pmu_add(),
901 	 * validate_group() and x86_pmu_commit_txn(). For the former two
902 	 * cpuc->n_events hasn't been updated yet, while for the latter
903 	 * cpuc->n_txn contains the number of events added in the current
904 	 * transaction.
905 	 */
906 	n0 = cpuc->n_events;
907 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
908 		n0 -= cpuc->n_txn;
909 
910 	if (x86_pmu.start_scheduling)
911 		x86_pmu.start_scheduling(cpuc);
912 
913 	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
914 		c = cpuc->event_constraint[i];
915 
916 		/*
917 		 * Previously scheduled events should have a cached constraint,
918 		 * while new events should not have one.
919 		 */
920 		WARN_ON_ONCE((c && i >= n0) || (!c && i < n0));
921 
922 		/*
923 		 * Request constraints for new events; or for those events that
924 		 * have a dynamic constraint -- for those the constraint can
925 		 * change due to external factors (sibling state, allow_tfa).
926 		 */
927 		if (!c || (c->flags & PERF_X86_EVENT_DYNAMIC)) {
928 			c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
929 			cpuc->event_constraint[i] = c;
930 		}
931 
932 		wmin = min(wmin, c->weight);
933 		wmax = max(wmax, c->weight);
934 	}
935 
936 	/*
937 	 * fastpath, try to reuse previous register
938 	 */
939 	for (i = 0; i < n; i++) {
940 		u64 mask;
941 
942 		hwc = &cpuc->event_list[i]->hw;
943 		c = cpuc->event_constraint[i];
944 
945 		/* never assigned */
946 		if (hwc->idx == -1)
947 			break;
948 
949 		/* constraint still honored */
950 		if (!test_bit(hwc->idx, c->idxmsk))
951 			break;
952 
953 		mask = BIT_ULL(hwc->idx);
954 		if (is_counter_pair(hwc))
955 			mask |= mask << 1;
956 
957 		/* not already used */
958 		if (used_mask & mask)
959 			break;
960 
961 		used_mask |= mask;
962 
963 		if (assign)
964 			assign[i] = hwc->idx;
965 	}
966 
967 	/* slow path */
968 	if (i != n) {
969 		int gpmax = x86_pmu.num_counters;
970 
971 		/*
972 		 * Do not allow scheduling of more than half the available
973 		 * generic counters.
974 		 *
975 		 * This helps avoid counter starvation of sibling thread by
976 		 * ensuring at most half the counters cannot be in exclusive
977 		 * mode. There is no designated counters for the limits. Any
978 		 * N/2 counters can be used. This helps with events with
979 		 * specific counter constraints.
980 		 */
981 		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
982 		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
983 			gpmax /= 2;
984 
985 		/*
986 		 * Reduce the amount of available counters to allow fitting
987 		 * the extra Merge events needed by large increment events.
988 		 */
989 		if (x86_pmu.flags & PMU_FL_PAIR) {
990 			gpmax = x86_pmu.num_counters - cpuc->n_pair;
991 			WARN_ON(gpmax <= 0);
992 		}
993 
994 		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
995 					     wmax, gpmax, assign);
996 	}
997 
998 	/*
999 	 * In case of success (unsched = 0), mark events as committed,
1000 	 * so we do not put_constraint() in case new events are added
1001 	 * and fail to be scheduled
1002 	 *
1003 	 * We invoke the lower level commit callback to lock the resource
1004 	 *
1005 	 * We do not need to do all of this in case we are called to
1006 	 * validate an event group (assign == NULL)
1007 	 */
1008 	if (!unsched && assign) {
1009 		for (i = 0; i < n; i++) {
1010 			e = cpuc->event_list[i];
1011 			if (x86_pmu.commit_scheduling)
1012 				x86_pmu.commit_scheduling(cpuc, i, assign[i]);
1013 		}
1014 	} else {
1015 		for (i = n0; i < n; i++) {
1016 			e = cpuc->event_list[i];
1017 
1018 			/*
1019 			 * release events that failed scheduling
1020 			 */
1021 			if (x86_pmu.put_event_constraints)
1022 				x86_pmu.put_event_constraints(cpuc, e);
1023 
1024 			cpuc->event_constraint[i] = NULL;
1025 		}
1026 	}
1027 
1028 	if (x86_pmu.stop_scheduling)
1029 		x86_pmu.stop_scheduling(cpuc);
1030 
1031 	return unsched ? -EINVAL : 0;
1032 }
1033 
1034 /*
1035  * dogrp: true if must collect siblings events (group)
1036  * returns total number of events and error code
1037  */
1038 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1039 {
1040 	struct perf_event *event;
1041 	int n, max_count;
1042 
1043 	max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
1044 
1045 	/* current number of events already accepted */
1046 	n = cpuc->n_events;
1047 	if (!cpuc->n_events)
1048 		cpuc->pebs_output = 0;
1049 
1050 	if (!cpuc->is_fake && leader->attr.precise_ip) {
1051 		/*
1052 		 * For PEBS->PT, if !aux_event, the group leader (PT) went
1053 		 * away, the group was broken down and this singleton event
1054 		 * can't schedule any more.
1055 		 */
1056 		if (is_pebs_pt(leader) && !leader->aux_event)
1057 			return -EINVAL;
1058 
1059 		/*
1060 		 * pebs_output: 0: no PEBS so far, 1: PT, 2: DS
1061 		 */
1062 		if (cpuc->pebs_output &&
1063 		    cpuc->pebs_output != is_pebs_pt(leader) + 1)
1064 			return -EINVAL;
1065 
1066 		cpuc->pebs_output = is_pebs_pt(leader) + 1;
1067 	}
1068 
1069 	if (is_x86_event(leader)) {
1070 		if (n >= max_count)
1071 			return -EINVAL;
1072 		cpuc->event_list[n] = leader;
1073 		n++;
1074 		if (is_counter_pair(&leader->hw))
1075 			cpuc->n_pair++;
1076 	}
1077 	if (!dogrp)
1078 		return n;
1079 
1080 	for_each_sibling_event(event, leader) {
1081 		if (!is_x86_event(event) ||
1082 		    event->state <= PERF_EVENT_STATE_OFF)
1083 			continue;
1084 
1085 		if (n >= max_count)
1086 			return -EINVAL;
1087 
1088 		cpuc->event_list[n] = event;
1089 		n++;
1090 		if (is_counter_pair(&event->hw))
1091 			cpuc->n_pair++;
1092 	}
1093 	return n;
1094 }
1095 
1096 static inline void x86_assign_hw_event(struct perf_event *event,
1097 				struct cpu_hw_events *cpuc, int i)
1098 {
1099 	struct hw_perf_event *hwc = &event->hw;
1100 
1101 	hwc->idx = cpuc->assign[i];
1102 	hwc->last_cpu = smp_processor_id();
1103 	hwc->last_tag = ++cpuc->tags[i];
1104 
1105 	if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
1106 		hwc->config_base = 0;
1107 		hwc->event_base	= 0;
1108 	} else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1109 		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1110 		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1111 		hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1112 	} else {
1113 		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1114 		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1115 		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1116 	}
1117 }
1118 
1119 /**
1120  * x86_perf_rdpmc_index - Return PMC counter used for event
1121  * @event: the perf_event to which the PMC counter was assigned
1122  *
1123  * The counter assigned to this performance event may change if interrupts
1124  * are enabled. This counter should thus never be used while interrupts are
1125  * enabled. Before this function is used to obtain the assigned counter the
1126  * event should be checked for validity using, for example,
1127  * perf_event_read_local(), within the same interrupt disabled section in
1128  * which this counter is planned to be used.
1129  *
1130  * Return: The index of the performance monitoring counter assigned to
1131  * @perf_event.
1132  */
1133 int x86_perf_rdpmc_index(struct perf_event *event)
1134 {
1135 	lockdep_assert_irqs_disabled();
1136 
1137 	return event->hw.event_base_rdpmc;
1138 }
1139 
1140 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1141 					struct cpu_hw_events *cpuc,
1142 					int i)
1143 {
1144 	return hwc->idx == cpuc->assign[i] &&
1145 		hwc->last_cpu == smp_processor_id() &&
1146 		hwc->last_tag == cpuc->tags[i];
1147 }
1148 
1149 static void x86_pmu_start(struct perf_event *event, int flags);
1150 
1151 static void x86_pmu_enable(struct pmu *pmu)
1152 {
1153 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1154 	struct perf_event *event;
1155 	struct hw_perf_event *hwc;
1156 	int i, added = cpuc->n_added;
1157 
1158 	if (!x86_pmu_initialized())
1159 		return;
1160 
1161 	if (cpuc->enabled)
1162 		return;
1163 
1164 	if (cpuc->n_added) {
1165 		int n_running = cpuc->n_events - cpuc->n_added;
1166 		/*
1167 		 * apply assignment obtained either from
1168 		 * hw_perf_group_sched_in() or x86_pmu_enable()
1169 		 *
1170 		 * step1: save events moving to new counters
1171 		 */
1172 		for (i = 0; i < n_running; i++) {
1173 			event = cpuc->event_list[i];
1174 			hwc = &event->hw;
1175 
1176 			/*
1177 			 * we can avoid reprogramming counter if:
1178 			 * - assigned same counter as last time
1179 			 * - running on same CPU as last time
1180 			 * - no other event has used the counter since
1181 			 */
1182 			if (hwc->idx == -1 ||
1183 			    match_prev_assignment(hwc, cpuc, i))
1184 				continue;
1185 
1186 			/*
1187 			 * Ensure we don't accidentally enable a stopped
1188 			 * counter simply because we rescheduled.
1189 			 */
1190 			if (hwc->state & PERF_HES_STOPPED)
1191 				hwc->state |= PERF_HES_ARCH;
1192 
1193 			x86_pmu_stop(event, PERF_EF_UPDATE);
1194 		}
1195 
1196 		/*
1197 		 * step2: reprogram moved events into new counters
1198 		 */
1199 		for (i = 0; i < cpuc->n_events; i++) {
1200 			event = cpuc->event_list[i];
1201 			hwc = &event->hw;
1202 
1203 			if (!match_prev_assignment(hwc, cpuc, i))
1204 				x86_assign_hw_event(event, cpuc, i);
1205 			else if (i < n_running)
1206 				continue;
1207 
1208 			if (hwc->state & PERF_HES_ARCH)
1209 				continue;
1210 
1211 			x86_pmu_start(event, PERF_EF_RELOAD);
1212 		}
1213 		cpuc->n_added = 0;
1214 		perf_events_lapic_init();
1215 	}
1216 
1217 	cpuc->enabled = 1;
1218 	barrier();
1219 
1220 	x86_pmu.enable_all(added);
1221 }
1222 
1223 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1224 
1225 /*
1226  * Set the next IRQ period, based on the hwc->period_left value.
1227  * To be called with the event disabled in hw:
1228  */
1229 int x86_perf_event_set_period(struct perf_event *event)
1230 {
1231 	struct hw_perf_event *hwc = &event->hw;
1232 	s64 left = local64_read(&hwc->period_left);
1233 	s64 period = hwc->sample_period;
1234 	int ret = 0, idx = hwc->idx;
1235 
1236 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
1237 		return 0;
1238 
1239 	/*
1240 	 * If we are way outside a reasonable range then just skip forward:
1241 	 */
1242 	if (unlikely(left <= -period)) {
1243 		left = period;
1244 		local64_set(&hwc->period_left, left);
1245 		hwc->last_period = period;
1246 		ret = 1;
1247 	}
1248 
1249 	if (unlikely(left <= 0)) {
1250 		left += period;
1251 		local64_set(&hwc->period_left, left);
1252 		hwc->last_period = period;
1253 		ret = 1;
1254 	}
1255 	/*
1256 	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1257 	 */
1258 	if (unlikely(left < 2))
1259 		left = 2;
1260 
1261 	if (left > x86_pmu.max_period)
1262 		left = x86_pmu.max_period;
1263 
1264 	if (x86_pmu.limit_period)
1265 		left = x86_pmu.limit_period(event, left);
1266 
1267 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1268 
1269 	/*
1270 	 * The hw event starts counting from this event offset,
1271 	 * mark it to be able to extra future deltas:
1272 	 */
1273 	local64_set(&hwc->prev_count, (u64)-left);
1274 
1275 	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1276 
1277 	/*
1278 	 * Clear the Merge event counter's upper 16 bits since
1279 	 * we currently declare a 48-bit counter width
1280 	 */
1281 	if (is_counter_pair(hwc))
1282 		wrmsrl(x86_pmu_event_addr(idx + 1), 0);
1283 
1284 	/*
1285 	 * Due to erratum on certan cpu we need
1286 	 * a second write to be sure the register
1287 	 * is updated properly
1288 	 */
1289 	if (x86_pmu.perfctr_second_write) {
1290 		wrmsrl(hwc->event_base,
1291 			(u64)(-left) & x86_pmu.cntval_mask);
1292 	}
1293 
1294 	perf_event_update_userpage(event);
1295 
1296 	return ret;
1297 }
1298 
1299 void x86_pmu_enable_event(struct perf_event *event)
1300 {
1301 	if (__this_cpu_read(cpu_hw_events.enabled))
1302 		__x86_pmu_enable_event(&event->hw,
1303 				       ARCH_PERFMON_EVENTSEL_ENABLE);
1304 }
1305 
1306 /*
1307  * Add a single event to the PMU.
1308  *
1309  * The event is added to the group of enabled events
1310  * but only if it can be scheduled with existing events.
1311  */
1312 static int x86_pmu_add(struct perf_event *event, int flags)
1313 {
1314 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1315 	struct hw_perf_event *hwc;
1316 	int assign[X86_PMC_IDX_MAX];
1317 	int n, n0, ret;
1318 
1319 	hwc = &event->hw;
1320 
1321 	n0 = cpuc->n_events;
1322 	ret = n = collect_events(cpuc, event, false);
1323 	if (ret < 0)
1324 		goto out;
1325 
1326 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1327 	if (!(flags & PERF_EF_START))
1328 		hwc->state |= PERF_HES_ARCH;
1329 
1330 	/*
1331 	 * If group events scheduling transaction was started,
1332 	 * skip the schedulability test here, it will be performed
1333 	 * at commit time (->commit_txn) as a whole.
1334 	 *
1335 	 * If commit fails, we'll call ->del() on all events
1336 	 * for which ->add() was called.
1337 	 */
1338 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1339 		goto done_collect;
1340 
1341 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1342 	if (ret)
1343 		goto out;
1344 	/*
1345 	 * copy new assignment, now we know it is possible
1346 	 * will be used by hw_perf_enable()
1347 	 */
1348 	memcpy(cpuc->assign, assign, n*sizeof(int));
1349 
1350 done_collect:
1351 	/*
1352 	 * Commit the collect_events() state. See x86_pmu_del() and
1353 	 * x86_pmu_*_txn().
1354 	 */
1355 	cpuc->n_events = n;
1356 	cpuc->n_added += n - n0;
1357 	cpuc->n_txn += n - n0;
1358 
1359 	if (x86_pmu.add) {
1360 		/*
1361 		 * This is before x86_pmu_enable() will call x86_pmu_start(),
1362 		 * so we enable LBRs before an event needs them etc..
1363 		 */
1364 		x86_pmu.add(event);
1365 	}
1366 
1367 	ret = 0;
1368 out:
1369 	return ret;
1370 }
1371 
1372 static void x86_pmu_start(struct perf_event *event, int flags)
1373 {
1374 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1375 	int idx = event->hw.idx;
1376 
1377 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1378 		return;
1379 
1380 	if (WARN_ON_ONCE(idx == -1))
1381 		return;
1382 
1383 	if (flags & PERF_EF_RELOAD) {
1384 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1385 		x86_perf_event_set_period(event);
1386 	}
1387 
1388 	event->hw.state = 0;
1389 
1390 	cpuc->events[idx] = event;
1391 	__set_bit(idx, cpuc->active_mask);
1392 	__set_bit(idx, cpuc->running);
1393 	x86_pmu.enable(event);
1394 	perf_event_update_userpage(event);
1395 }
1396 
1397 void perf_event_print_debug(void)
1398 {
1399 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1400 	u64 pebs, debugctl;
1401 	struct cpu_hw_events *cpuc;
1402 	unsigned long flags;
1403 	int cpu, idx;
1404 
1405 	if (!x86_pmu.num_counters)
1406 		return;
1407 
1408 	local_irq_save(flags);
1409 
1410 	cpu = smp_processor_id();
1411 	cpuc = &per_cpu(cpu_hw_events, cpu);
1412 
1413 	if (x86_pmu.version >= 2) {
1414 		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1415 		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1416 		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1417 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1418 
1419 		pr_info("\n");
1420 		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1421 		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1422 		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1423 		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1424 		if (x86_pmu.pebs_constraints) {
1425 			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1426 			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1427 		}
1428 		if (x86_pmu.lbr_nr) {
1429 			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1430 			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1431 		}
1432 	}
1433 	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1434 
1435 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1436 		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1437 		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1438 
1439 		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1440 
1441 		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1442 			cpu, idx, pmc_ctrl);
1443 		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1444 			cpu, idx, pmc_count);
1445 		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1446 			cpu, idx, prev_left);
1447 	}
1448 	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1449 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1450 
1451 		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1452 			cpu, idx, pmc_count);
1453 	}
1454 	local_irq_restore(flags);
1455 }
1456 
1457 void x86_pmu_stop(struct perf_event *event, int flags)
1458 {
1459 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1460 	struct hw_perf_event *hwc = &event->hw;
1461 
1462 	if (test_bit(hwc->idx, cpuc->active_mask)) {
1463 		x86_pmu.disable(event);
1464 		__clear_bit(hwc->idx, cpuc->active_mask);
1465 		cpuc->events[hwc->idx] = NULL;
1466 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1467 		hwc->state |= PERF_HES_STOPPED;
1468 	}
1469 
1470 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1471 		/*
1472 		 * Drain the remaining delta count out of a event
1473 		 * that we are disabling:
1474 		 */
1475 		x86_perf_event_update(event);
1476 		hwc->state |= PERF_HES_UPTODATE;
1477 	}
1478 }
1479 
1480 static void x86_pmu_del(struct perf_event *event, int flags)
1481 {
1482 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1483 	int i;
1484 
1485 	/*
1486 	 * If we're called during a txn, we only need to undo x86_pmu.add.
1487 	 * The events never got scheduled and ->cancel_txn will truncate
1488 	 * the event_list.
1489 	 *
1490 	 * XXX assumes any ->del() called during a TXN will only be on
1491 	 * an event added during that same TXN.
1492 	 */
1493 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1494 		goto do_del;
1495 
1496 	/*
1497 	 * Not a TXN, therefore cleanup properly.
1498 	 */
1499 	x86_pmu_stop(event, PERF_EF_UPDATE);
1500 
1501 	for (i = 0; i < cpuc->n_events; i++) {
1502 		if (event == cpuc->event_list[i])
1503 			break;
1504 	}
1505 
1506 	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1507 		return;
1508 
1509 	/* If we have a newly added event; make sure to decrease n_added. */
1510 	if (i >= cpuc->n_events - cpuc->n_added)
1511 		--cpuc->n_added;
1512 
1513 	if (x86_pmu.put_event_constraints)
1514 		x86_pmu.put_event_constraints(cpuc, event);
1515 
1516 	/* Delete the array entry. */
1517 	while (++i < cpuc->n_events) {
1518 		cpuc->event_list[i-1] = cpuc->event_list[i];
1519 		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1520 	}
1521 	cpuc->event_constraint[i-1] = NULL;
1522 	--cpuc->n_events;
1523 
1524 	perf_event_update_userpage(event);
1525 
1526 do_del:
1527 	if (x86_pmu.del) {
1528 		/*
1529 		 * This is after x86_pmu_stop(); so we disable LBRs after any
1530 		 * event can need them etc..
1531 		 */
1532 		x86_pmu.del(event);
1533 	}
1534 }
1535 
1536 int x86_pmu_handle_irq(struct pt_regs *regs)
1537 {
1538 	struct perf_sample_data data;
1539 	struct cpu_hw_events *cpuc;
1540 	struct perf_event *event;
1541 	int idx, handled = 0;
1542 	u64 val;
1543 
1544 	cpuc = this_cpu_ptr(&cpu_hw_events);
1545 
1546 	/*
1547 	 * Some chipsets need to unmask the LVTPC in a particular spot
1548 	 * inside the nmi handler.  As a result, the unmasking was pushed
1549 	 * into all the nmi handlers.
1550 	 *
1551 	 * This generic handler doesn't seem to have any issues where the
1552 	 * unmasking occurs so it was left at the top.
1553 	 */
1554 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1555 
1556 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1557 		if (!test_bit(idx, cpuc->active_mask))
1558 			continue;
1559 
1560 		event = cpuc->events[idx];
1561 
1562 		val = x86_perf_event_update(event);
1563 		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1564 			continue;
1565 
1566 		/*
1567 		 * event overflow
1568 		 */
1569 		handled++;
1570 		perf_sample_data_init(&data, 0, event->hw.last_period);
1571 
1572 		if (!x86_perf_event_set_period(event))
1573 			continue;
1574 
1575 		if (perf_event_overflow(event, &data, regs))
1576 			x86_pmu_stop(event, 0);
1577 	}
1578 
1579 	if (handled)
1580 		inc_irq_stat(apic_perf_irqs);
1581 
1582 	return handled;
1583 }
1584 
1585 void perf_events_lapic_init(void)
1586 {
1587 	if (!x86_pmu.apic || !x86_pmu_initialized())
1588 		return;
1589 
1590 	/*
1591 	 * Always use NMI for PMU
1592 	 */
1593 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1594 }
1595 
1596 static int
1597 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1598 {
1599 	u64 start_clock;
1600 	u64 finish_clock;
1601 	int ret;
1602 
1603 	/*
1604 	 * All PMUs/events that share this PMI handler should make sure to
1605 	 * increment active_events for their events.
1606 	 */
1607 	if (!atomic_read(&active_events))
1608 		return NMI_DONE;
1609 
1610 	start_clock = sched_clock();
1611 	ret = x86_pmu.handle_irq(regs);
1612 	finish_clock = sched_clock();
1613 
1614 	perf_sample_event_took(finish_clock - start_clock);
1615 
1616 	return ret;
1617 }
1618 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1619 
1620 struct event_constraint emptyconstraint;
1621 struct event_constraint unconstrained;
1622 
1623 static int x86_pmu_prepare_cpu(unsigned int cpu)
1624 {
1625 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1626 	int i;
1627 
1628 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1629 		cpuc->kfree_on_online[i] = NULL;
1630 	if (x86_pmu.cpu_prepare)
1631 		return x86_pmu.cpu_prepare(cpu);
1632 	return 0;
1633 }
1634 
1635 static int x86_pmu_dead_cpu(unsigned int cpu)
1636 {
1637 	if (x86_pmu.cpu_dead)
1638 		x86_pmu.cpu_dead(cpu);
1639 	return 0;
1640 }
1641 
1642 static int x86_pmu_online_cpu(unsigned int cpu)
1643 {
1644 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1645 	int i;
1646 
1647 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1648 		kfree(cpuc->kfree_on_online[i]);
1649 		cpuc->kfree_on_online[i] = NULL;
1650 	}
1651 	return 0;
1652 }
1653 
1654 static int x86_pmu_starting_cpu(unsigned int cpu)
1655 {
1656 	if (x86_pmu.cpu_starting)
1657 		x86_pmu.cpu_starting(cpu);
1658 	return 0;
1659 }
1660 
1661 static int x86_pmu_dying_cpu(unsigned int cpu)
1662 {
1663 	if (x86_pmu.cpu_dying)
1664 		x86_pmu.cpu_dying(cpu);
1665 	return 0;
1666 }
1667 
1668 static void __init pmu_check_apic(void)
1669 {
1670 	if (boot_cpu_has(X86_FEATURE_APIC))
1671 		return;
1672 
1673 	x86_pmu.apic = 0;
1674 	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1675 	pr_info("no hardware sampling interrupt available.\n");
1676 
1677 	/*
1678 	 * If we have a PMU initialized but no APIC
1679 	 * interrupts, we cannot sample hardware
1680 	 * events (user-space has to fall back and
1681 	 * sample via a hrtimer based software event):
1682 	 */
1683 	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1684 
1685 }
1686 
1687 static struct attribute_group x86_pmu_format_group __ro_after_init = {
1688 	.name = "format",
1689 	.attrs = NULL,
1690 };
1691 
1692 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1693 {
1694 	struct perf_pmu_events_attr *pmu_attr =
1695 		container_of(attr, struct perf_pmu_events_attr, attr);
1696 	u64 config = 0;
1697 
1698 	if (pmu_attr->id < x86_pmu.max_events)
1699 		config = x86_pmu.event_map(pmu_attr->id);
1700 
1701 	/* string trumps id */
1702 	if (pmu_attr->event_str)
1703 		return sprintf(page, "%s", pmu_attr->event_str);
1704 
1705 	return x86_pmu.events_sysfs_show(page, config);
1706 }
1707 EXPORT_SYMBOL_GPL(events_sysfs_show);
1708 
1709 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1710 			  char *page)
1711 {
1712 	struct perf_pmu_events_ht_attr *pmu_attr =
1713 		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1714 
1715 	/*
1716 	 * Report conditional events depending on Hyper-Threading.
1717 	 *
1718 	 * This is overly conservative as usually the HT special
1719 	 * handling is not needed if the other CPU thread is idle.
1720 	 *
1721 	 * Note this does not (and cannot) handle the case when thread
1722 	 * siblings are invisible, for example with virtualization
1723 	 * if they are owned by some other guest.  The user tool
1724 	 * has to re-read when a thread sibling gets onlined later.
1725 	 */
1726 	return sprintf(page, "%s",
1727 			topology_max_smt_threads() > 1 ?
1728 			pmu_attr->event_str_ht :
1729 			pmu_attr->event_str_noht);
1730 }
1731 
1732 EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1733 EVENT_ATTR(instructions,		INSTRUCTIONS		);
1734 EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1735 EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1736 EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1737 EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1738 EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1739 EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1740 EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1741 EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1742 
1743 static struct attribute *empty_attrs;
1744 
1745 static struct attribute *events_attr[] = {
1746 	EVENT_PTR(CPU_CYCLES),
1747 	EVENT_PTR(INSTRUCTIONS),
1748 	EVENT_PTR(CACHE_REFERENCES),
1749 	EVENT_PTR(CACHE_MISSES),
1750 	EVENT_PTR(BRANCH_INSTRUCTIONS),
1751 	EVENT_PTR(BRANCH_MISSES),
1752 	EVENT_PTR(BUS_CYCLES),
1753 	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1754 	EVENT_PTR(STALLED_CYCLES_BACKEND),
1755 	EVENT_PTR(REF_CPU_CYCLES),
1756 	NULL,
1757 };
1758 
1759 /*
1760  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1761  * out of events_attr attributes.
1762  */
1763 static umode_t
1764 is_visible(struct kobject *kobj, struct attribute *attr, int idx)
1765 {
1766 	struct perf_pmu_events_attr *pmu_attr;
1767 
1768 	if (idx >= x86_pmu.max_events)
1769 		return 0;
1770 
1771 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
1772 	/* str trumps id */
1773 	return pmu_attr->event_str || x86_pmu.event_map(idx) ? attr->mode : 0;
1774 }
1775 
1776 static struct attribute_group x86_pmu_events_group __ro_after_init = {
1777 	.name = "events",
1778 	.attrs = events_attr,
1779 	.is_visible = is_visible,
1780 };
1781 
1782 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1783 {
1784 	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1785 	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1786 	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1787 	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1788 	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1789 	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1790 	ssize_t ret;
1791 
1792 	/*
1793 	* We have whole page size to spend and just little data
1794 	* to write, so we can safely use sprintf.
1795 	*/
1796 	ret = sprintf(page, "event=0x%02llx", event);
1797 
1798 	if (umask)
1799 		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1800 
1801 	if (edge)
1802 		ret += sprintf(page + ret, ",edge");
1803 
1804 	if (pc)
1805 		ret += sprintf(page + ret, ",pc");
1806 
1807 	if (any)
1808 		ret += sprintf(page + ret, ",any");
1809 
1810 	if (inv)
1811 		ret += sprintf(page + ret, ",inv");
1812 
1813 	if (cmask)
1814 		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1815 
1816 	ret += sprintf(page + ret, "\n");
1817 
1818 	return ret;
1819 }
1820 
1821 static struct attribute_group x86_pmu_attr_group;
1822 static struct attribute_group x86_pmu_caps_group;
1823 
1824 static int __init init_hw_perf_events(void)
1825 {
1826 	struct x86_pmu_quirk *quirk;
1827 	int err;
1828 
1829 	pr_info("Performance Events: ");
1830 
1831 	switch (boot_cpu_data.x86_vendor) {
1832 	case X86_VENDOR_INTEL:
1833 		err = intel_pmu_init();
1834 		break;
1835 	case X86_VENDOR_AMD:
1836 		err = amd_pmu_init();
1837 		break;
1838 	case X86_VENDOR_HYGON:
1839 		err = amd_pmu_init();
1840 		x86_pmu.name = "HYGON";
1841 		break;
1842 	default:
1843 		err = -ENOTSUPP;
1844 	}
1845 	if (err != 0) {
1846 		pr_cont("no PMU driver, software events only.\n");
1847 		return 0;
1848 	}
1849 
1850 	pmu_check_apic();
1851 
1852 	/* sanity check that the hardware exists or is emulated */
1853 	if (!check_hw_exists())
1854 		return 0;
1855 
1856 	pr_cont("%s PMU driver.\n", x86_pmu.name);
1857 
1858 	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1859 
1860 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1861 		quirk->func();
1862 
1863 	if (!x86_pmu.intel_ctrl)
1864 		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1865 
1866 	perf_events_lapic_init();
1867 	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1868 
1869 	unconstrained = (struct event_constraint)
1870 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1871 				   0, x86_pmu.num_counters, 0, 0);
1872 
1873 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1874 
1875 	if (!x86_pmu.events_sysfs_show)
1876 		x86_pmu_events_group.attrs = &empty_attrs;
1877 
1878 	pmu.attr_update = x86_pmu.attr_update;
1879 
1880 	pr_info("... version:                %d\n",     x86_pmu.version);
1881 	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1882 	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1883 	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1884 	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1885 	pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1886 	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1887 
1888 	/*
1889 	 * Install callbacks. Core will call them for each online
1890 	 * cpu.
1891 	 */
1892 	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
1893 				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1894 	if (err)
1895 		return err;
1896 
1897 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1898 				"perf/x86:starting", x86_pmu_starting_cpu,
1899 				x86_pmu_dying_cpu);
1900 	if (err)
1901 		goto out;
1902 
1903 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
1904 				x86_pmu_online_cpu, NULL);
1905 	if (err)
1906 		goto out1;
1907 
1908 	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1909 	if (err)
1910 		goto out2;
1911 
1912 	return 0;
1913 
1914 out2:
1915 	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1916 out1:
1917 	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1918 out:
1919 	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
1920 	return err;
1921 }
1922 early_initcall(init_hw_perf_events);
1923 
1924 static inline void x86_pmu_read(struct perf_event *event)
1925 {
1926 	if (x86_pmu.read)
1927 		return x86_pmu.read(event);
1928 	x86_perf_event_update(event);
1929 }
1930 
1931 /*
1932  * Start group events scheduling transaction
1933  * Set the flag to make pmu::enable() not perform the
1934  * schedulability test, it will be performed at commit time
1935  *
1936  * We only support PERF_PMU_TXN_ADD transactions. Save the
1937  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1938  * transactions.
1939  */
1940 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1941 {
1942 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1943 
1944 	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
1945 
1946 	cpuc->txn_flags = txn_flags;
1947 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1948 		return;
1949 
1950 	perf_pmu_disable(pmu);
1951 	__this_cpu_write(cpu_hw_events.n_txn, 0);
1952 }
1953 
1954 /*
1955  * Stop group events scheduling transaction
1956  * Clear the flag and pmu::enable() will perform the
1957  * schedulability test.
1958  */
1959 static void x86_pmu_cancel_txn(struct pmu *pmu)
1960 {
1961 	unsigned int txn_flags;
1962 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1963 
1964 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1965 
1966 	txn_flags = cpuc->txn_flags;
1967 	cpuc->txn_flags = 0;
1968 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1969 		return;
1970 
1971 	/*
1972 	 * Truncate collected array by the number of events added in this
1973 	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1974 	 */
1975 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1976 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1977 	perf_pmu_enable(pmu);
1978 }
1979 
1980 /*
1981  * Commit group events scheduling transaction
1982  * Perform the group schedulability test as a whole
1983  * Return 0 if success
1984  *
1985  * Does not cancel the transaction on failure; expects the caller to do this.
1986  */
1987 static int x86_pmu_commit_txn(struct pmu *pmu)
1988 {
1989 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1990 	int assign[X86_PMC_IDX_MAX];
1991 	int n, ret;
1992 
1993 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1994 
1995 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1996 		cpuc->txn_flags = 0;
1997 		return 0;
1998 	}
1999 
2000 	n = cpuc->n_events;
2001 
2002 	if (!x86_pmu_initialized())
2003 		return -EAGAIN;
2004 
2005 	ret = x86_pmu.schedule_events(cpuc, n, assign);
2006 	if (ret)
2007 		return ret;
2008 
2009 	/*
2010 	 * copy new assignment, now we know it is possible
2011 	 * will be used by hw_perf_enable()
2012 	 */
2013 	memcpy(cpuc->assign, assign, n*sizeof(int));
2014 
2015 	cpuc->txn_flags = 0;
2016 	perf_pmu_enable(pmu);
2017 	return 0;
2018 }
2019 /*
2020  * a fake_cpuc is used to validate event groups. Due to
2021  * the extra reg logic, we need to also allocate a fake
2022  * per_core and per_cpu structure. Otherwise, group events
2023  * using extra reg may conflict without the kernel being
2024  * able to catch this when the last event gets added to
2025  * the group.
2026  */
2027 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
2028 {
2029 	intel_cpuc_finish(cpuc);
2030 	kfree(cpuc);
2031 }
2032 
2033 static struct cpu_hw_events *allocate_fake_cpuc(void)
2034 {
2035 	struct cpu_hw_events *cpuc;
2036 	int cpu = raw_smp_processor_id();
2037 
2038 	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
2039 	if (!cpuc)
2040 		return ERR_PTR(-ENOMEM);
2041 	cpuc->is_fake = 1;
2042 
2043 	if (intel_cpuc_prepare(cpuc, cpu))
2044 		goto error;
2045 
2046 	return cpuc;
2047 error:
2048 	free_fake_cpuc(cpuc);
2049 	return ERR_PTR(-ENOMEM);
2050 }
2051 
2052 /*
2053  * validate that we can schedule this event
2054  */
2055 static int validate_event(struct perf_event *event)
2056 {
2057 	struct cpu_hw_events *fake_cpuc;
2058 	struct event_constraint *c;
2059 	int ret = 0;
2060 
2061 	fake_cpuc = allocate_fake_cpuc();
2062 	if (IS_ERR(fake_cpuc))
2063 		return PTR_ERR(fake_cpuc);
2064 
2065 	c = x86_pmu.get_event_constraints(fake_cpuc, 0, event);
2066 
2067 	if (!c || !c->weight)
2068 		ret = -EINVAL;
2069 
2070 	if (x86_pmu.put_event_constraints)
2071 		x86_pmu.put_event_constraints(fake_cpuc, event);
2072 
2073 	free_fake_cpuc(fake_cpuc);
2074 
2075 	return ret;
2076 }
2077 
2078 /*
2079  * validate a single event group
2080  *
2081  * validation include:
2082  *	- check events are compatible which each other
2083  *	- events do not compete for the same counter
2084  *	- number of events <= number of counters
2085  *
2086  * validation ensures the group can be loaded onto the
2087  * PMU if it was the only group available.
2088  */
2089 static int validate_group(struct perf_event *event)
2090 {
2091 	struct perf_event *leader = event->group_leader;
2092 	struct cpu_hw_events *fake_cpuc;
2093 	int ret = -EINVAL, n;
2094 
2095 	fake_cpuc = allocate_fake_cpuc();
2096 	if (IS_ERR(fake_cpuc))
2097 		return PTR_ERR(fake_cpuc);
2098 	/*
2099 	 * the event is not yet connected with its
2100 	 * siblings therefore we must first collect
2101 	 * existing siblings, then add the new event
2102 	 * before we can simulate the scheduling
2103 	 */
2104 	n = collect_events(fake_cpuc, leader, true);
2105 	if (n < 0)
2106 		goto out;
2107 
2108 	fake_cpuc->n_events = n;
2109 	n = collect_events(fake_cpuc, event, false);
2110 	if (n < 0)
2111 		goto out;
2112 
2113 	fake_cpuc->n_events = 0;
2114 	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2115 
2116 out:
2117 	free_fake_cpuc(fake_cpuc);
2118 	return ret;
2119 }
2120 
2121 static int x86_pmu_event_init(struct perf_event *event)
2122 {
2123 	struct pmu *tmp;
2124 	int err;
2125 
2126 	switch (event->attr.type) {
2127 	case PERF_TYPE_RAW:
2128 	case PERF_TYPE_HARDWARE:
2129 	case PERF_TYPE_HW_CACHE:
2130 		break;
2131 
2132 	default:
2133 		return -ENOENT;
2134 	}
2135 
2136 	err = __x86_pmu_event_init(event);
2137 	if (!err) {
2138 		/*
2139 		 * we temporarily connect event to its pmu
2140 		 * such that validate_group() can classify
2141 		 * it as an x86 event using is_x86_event()
2142 		 */
2143 		tmp = event->pmu;
2144 		event->pmu = &pmu;
2145 
2146 		if (event->group_leader != event)
2147 			err = validate_group(event);
2148 		else
2149 			err = validate_event(event);
2150 
2151 		event->pmu = tmp;
2152 	}
2153 	if (err) {
2154 		if (event->destroy)
2155 			event->destroy(event);
2156 	}
2157 
2158 	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2159 	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2160 		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2161 
2162 	return err;
2163 }
2164 
2165 static void refresh_pce(void *ignored)
2166 {
2167 	load_mm_cr4_irqsoff(this_cpu_read(cpu_tlbstate.loaded_mm));
2168 }
2169 
2170 static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2171 {
2172 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2173 		return;
2174 
2175 	/*
2176 	 * This function relies on not being called concurrently in two
2177 	 * tasks in the same mm.  Otherwise one task could observe
2178 	 * perf_rdpmc_allowed > 1 and return all the way back to
2179 	 * userspace with CR4.PCE clear while another task is still
2180 	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2181 	 *
2182 	 * For now, this can't happen because all callers hold mmap_sem
2183 	 * for write.  If this changes, we'll need a different solution.
2184 	 */
2185 	lockdep_assert_held_write(&mm->mmap_sem);
2186 
2187 	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2188 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2189 }
2190 
2191 static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2192 {
2193 
2194 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2195 		return;
2196 
2197 	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2198 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2199 }
2200 
2201 static int x86_pmu_event_idx(struct perf_event *event)
2202 {
2203 	int idx = event->hw.idx;
2204 
2205 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2206 		return 0;
2207 
2208 	if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2209 		idx -= INTEL_PMC_IDX_FIXED;
2210 		idx |= 1 << 30;
2211 	}
2212 
2213 	return idx + 1;
2214 }
2215 
2216 static ssize_t get_attr_rdpmc(struct device *cdev,
2217 			      struct device_attribute *attr,
2218 			      char *buf)
2219 {
2220 	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2221 }
2222 
2223 static ssize_t set_attr_rdpmc(struct device *cdev,
2224 			      struct device_attribute *attr,
2225 			      const char *buf, size_t count)
2226 {
2227 	unsigned long val;
2228 	ssize_t ret;
2229 
2230 	ret = kstrtoul(buf, 0, &val);
2231 	if (ret)
2232 		return ret;
2233 
2234 	if (val > 2)
2235 		return -EINVAL;
2236 
2237 	if (x86_pmu.attr_rdpmc_broken)
2238 		return -ENOTSUPP;
2239 
2240 	if (val != x86_pmu.attr_rdpmc) {
2241 		/*
2242 		 * Changing into or out of never available or always available,
2243 		 * aka perf-event-bypassing mode. This path is extremely slow,
2244 		 * but only root can trigger it, so it's okay.
2245 		 */
2246 		if (val == 0)
2247 			static_branch_inc(&rdpmc_never_available_key);
2248 		else if (x86_pmu.attr_rdpmc == 0)
2249 			static_branch_dec(&rdpmc_never_available_key);
2250 
2251 		if (val == 2)
2252 			static_branch_inc(&rdpmc_always_available_key);
2253 		else if (x86_pmu.attr_rdpmc == 2)
2254 			static_branch_dec(&rdpmc_always_available_key);
2255 
2256 		on_each_cpu(refresh_pce, NULL, 1);
2257 		x86_pmu.attr_rdpmc = val;
2258 	}
2259 
2260 	return count;
2261 }
2262 
2263 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2264 
2265 static struct attribute *x86_pmu_attrs[] = {
2266 	&dev_attr_rdpmc.attr,
2267 	NULL,
2268 };
2269 
2270 static struct attribute_group x86_pmu_attr_group __ro_after_init = {
2271 	.attrs = x86_pmu_attrs,
2272 };
2273 
2274 static ssize_t max_precise_show(struct device *cdev,
2275 				  struct device_attribute *attr,
2276 				  char *buf)
2277 {
2278 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2279 }
2280 
2281 static DEVICE_ATTR_RO(max_precise);
2282 
2283 static struct attribute *x86_pmu_caps_attrs[] = {
2284 	&dev_attr_max_precise.attr,
2285 	NULL
2286 };
2287 
2288 static struct attribute_group x86_pmu_caps_group __ro_after_init = {
2289 	.name = "caps",
2290 	.attrs = x86_pmu_caps_attrs,
2291 };
2292 
2293 static const struct attribute_group *x86_pmu_attr_groups[] = {
2294 	&x86_pmu_attr_group,
2295 	&x86_pmu_format_group,
2296 	&x86_pmu_events_group,
2297 	&x86_pmu_caps_group,
2298 	NULL,
2299 };
2300 
2301 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2302 {
2303 	if (x86_pmu.sched_task)
2304 		x86_pmu.sched_task(ctx, sched_in);
2305 }
2306 
2307 static void x86_pmu_swap_task_ctx(struct perf_event_context *prev,
2308 				  struct perf_event_context *next)
2309 {
2310 	if (x86_pmu.swap_task_ctx)
2311 		x86_pmu.swap_task_ctx(prev, next);
2312 }
2313 
2314 void perf_check_microcode(void)
2315 {
2316 	if (x86_pmu.check_microcode)
2317 		x86_pmu.check_microcode();
2318 }
2319 
2320 static int x86_pmu_check_period(struct perf_event *event, u64 value)
2321 {
2322 	if (x86_pmu.check_period && x86_pmu.check_period(event, value))
2323 		return -EINVAL;
2324 
2325 	if (value && x86_pmu.limit_period) {
2326 		if (x86_pmu.limit_period(event, value) > value)
2327 			return -EINVAL;
2328 	}
2329 
2330 	return 0;
2331 }
2332 
2333 static int x86_pmu_aux_output_match(struct perf_event *event)
2334 {
2335 	if (!(pmu.capabilities & PERF_PMU_CAP_AUX_OUTPUT))
2336 		return 0;
2337 
2338 	if (x86_pmu.aux_output_match)
2339 		return x86_pmu.aux_output_match(event);
2340 
2341 	return 0;
2342 }
2343 
2344 static struct pmu pmu = {
2345 	.pmu_enable		= x86_pmu_enable,
2346 	.pmu_disable		= x86_pmu_disable,
2347 
2348 	.attr_groups		= x86_pmu_attr_groups,
2349 
2350 	.event_init		= x86_pmu_event_init,
2351 
2352 	.event_mapped		= x86_pmu_event_mapped,
2353 	.event_unmapped		= x86_pmu_event_unmapped,
2354 
2355 	.add			= x86_pmu_add,
2356 	.del			= x86_pmu_del,
2357 	.start			= x86_pmu_start,
2358 	.stop			= x86_pmu_stop,
2359 	.read			= x86_pmu_read,
2360 
2361 	.start_txn		= x86_pmu_start_txn,
2362 	.cancel_txn		= x86_pmu_cancel_txn,
2363 	.commit_txn		= x86_pmu_commit_txn,
2364 
2365 	.event_idx		= x86_pmu_event_idx,
2366 	.sched_task		= x86_pmu_sched_task,
2367 	.task_ctx_size          = sizeof(struct x86_perf_task_context),
2368 	.swap_task_ctx		= x86_pmu_swap_task_ctx,
2369 	.check_period		= x86_pmu_check_period,
2370 
2371 	.aux_output_match	= x86_pmu_aux_output_match,
2372 };
2373 
2374 void arch_perf_update_userpage(struct perf_event *event,
2375 			       struct perf_event_mmap_page *userpg, u64 now)
2376 {
2377 	struct cyc2ns_data data;
2378 	u64 offset;
2379 
2380 	userpg->cap_user_time = 0;
2381 	userpg->cap_user_time_zero = 0;
2382 	userpg->cap_user_rdpmc =
2383 		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2384 	userpg->pmc_width = x86_pmu.cntval_bits;
2385 
2386 	if (!using_native_sched_clock() || !sched_clock_stable())
2387 		return;
2388 
2389 	cyc2ns_read_begin(&data);
2390 
2391 	offset = data.cyc2ns_offset + __sched_clock_offset;
2392 
2393 	/*
2394 	 * Internal timekeeping for enabled/running/stopped times
2395 	 * is always in the local_clock domain.
2396 	 */
2397 	userpg->cap_user_time = 1;
2398 	userpg->time_mult = data.cyc2ns_mul;
2399 	userpg->time_shift = data.cyc2ns_shift;
2400 	userpg->time_offset = offset - now;
2401 
2402 	/*
2403 	 * cap_user_time_zero doesn't make sense when we're using a different
2404 	 * time base for the records.
2405 	 */
2406 	if (!event->attr.use_clockid) {
2407 		userpg->cap_user_time_zero = 1;
2408 		userpg->time_zero = offset;
2409 	}
2410 
2411 	cyc2ns_read_end();
2412 }
2413 
2414 /*
2415  * Determine whether the regs were taken from an irq/exception handler rather
2416  * than from perf_arch_fetch_caller_regs().
2417  */
2418 static bool perf_hw_regs(struct pt_regs *regs)
2419 {
2420 	return regs->flags & X86_EFLAGS_FIXED;
2421 }
2422 
2423 void
2424 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2425 {
2426 	struct unwind_state state;
2427 	unsigned long addr;
2428 
2429 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2430 		/* TODO: We don't support guest os callchain now */
2431 		return;
2432 	}
2433 
2434 	if (perf_callchain_store(entry, regs->ip))
2435 		return;
2436 
2437 	if (perf_hw_regs(regs))
2438 		unwind_start(&state, current, regs, NULL);
2439 	else
2440 		unwind_start(&state, current, NULL, (void *)regs->sp);
2441 
2442 	for (; !unwind_done(&state); unwind_next_frame(&state)) {
2443 		addr = unwind_get_return_address(&state);
2444 		if (!addr || perf_callchain_store(entry, addr))
2445 			return;
2446 	}
2447 }
2448 
2449 static inline int
2450 valid_user_frame(const void __user *fp, unsigned long size)
2451 {
2452 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2453 }
2454 
2455 static unsigned long get_segment_base(unsigned int segment)
2456 {
2457 	struct desc_struct *desc;
2458 	unsigned int idx = segment >> 3;
2459 
2460 	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2461 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2462 		struct ldt_struct *ldt;
2463 
2464 		/* IRQs are off, so this synchronizes with smp_store_release */
2465 		ldt = READ_ONCE(current->active_mm->context.ldt);
2466 		if (!ldt || idx >= ldt->nr_entries)
2467 			return 0;
2468 
2469 		desc = &ldt->entries[idx];
2470 #else
2471 		return 0;
2472 #endif
2473 	} else {
2474 		if (idx >= GDT_ENTRIES)
2475 			return 0;
2476 
2477 		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2478 	}
2479 
2480 	return get_desc_base(desc);
2481 }
2482 
2483 #ifdef CONFIG_IA32_EMULATION
2484 
2485 #include <linux/compat.h>
2486 
2487 static inline int
2488 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2489 {
2490 	/* 32-bit process in 64-bit kernel. */
2491 	unsigned long ss_base, cs_base;
2492 	struct stack_frame_ia32 frame;
2493 	const void __user *fp;
2494 
2495 	if (!test_thread_flag(TIF_IA32))
2496 		return 0;
2497 
2498 	cs_base = get_segment_base(regs->cs);
2499 	ss_base = get_segment_base(regs->ss);
2500 
2501 	fp = compat_ptr(ss_base + regs->bp);
2502 	pagefault_disable();
2503 	while (entry->nr < entry->max_stack) {
2504 		unsigned long bytes;
2505 		frame.next_frame     = 0;
2506 		frame.return_address = 0;
2507 
2508 		if (!valid_user_frame(fp, sizeof(frame)))
2509 			break;
2510 
2511 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2512 		if (bytes != 0)
2513 			break;
2514 		bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2515 		if (bytes != 0)
2516 			break;
2517 
2518 		perf_callchain_store(entry, cs_base + frame.return_address);
2519 		fp = compat_ptr(ss_base + frame.next_frame);
2520 	}
2521 	pagefault_enable();
2522 	return 1;
2523 }
2524 #else
2525 static inline int
2526 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2527 {
2528     return 0;
2529 }
2530 #endif
2531 
2532 void
2533 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2534 {
2535 	struct stack_frame frame;
2536 	const unsigned long __user *fp;
2537 
2538 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2539 		/* TODO: We don't support guest os callchain now */
2540 		return;
2541 	}
2542 
2543 	/*
2544 	 * We don't know what to do with VM86 stacks.. ignore them for now.
2545 	 */
2546 	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2547 		return;
2548 
2549 	fp = (unsigned long __user *)regs->bp;
2550 
2551 	perf_callchain_store(entry, regs->ip);
2552 
2553 	if (!nmi_uaccess_okay())
2554 		return;
2555 
2556 	if (perf_callchain_user32(regs, entry))
2557 		return;
2558 
2559 	pagefault_disable();
2560 	while (entry->nr < entry->max_stack) {
2561 		unsigned long bytes;
2562 
2563 		frame.next_frame	     = NULL;
2564 		frame.return_address = 0;
2565 
2566 		if (!valid_user_frame(fp, sizeof(frame)))
2567 			break;
2568 
2569 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2570 		if (bytes != 0)
2571 			break;
2572 		bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2573 		if (bytes != 0)
2574 			break;
2575 
2576 		perf_callchain_store(entry, frame.return_address);
2577 		fp = (void __user *)frame.next_frame;
2578 	}
2579 	pagefault_enable();
2580 }
2581 
2582 /*
2583  * Deal with code segment offsets for the various execution modes:
2584  *
2585  *   VM86 - the good olde 16 bit days, where the linear address is
2586  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2587  *
2588  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2589  *          to figure out what the 32bit base address is.
2590  *
2591  *    X32 - has TIF_X32 set, but is running in x86_64
2592  *
2593  * X86_64 - CS,DS,SS,ES are all zero based.
2594  */
2595 static unsigned long code_segment_base(struct pt_regs *regs)
2596 {
2597 	/*
2598 	 * For IA32 we look at the GDT/LDT segment base to convert the
2599 	 * effective IP to a linear address.
2600 	 */
2601 
2602 #ifdef CONFIG_X86_32
2603 	/*
2604 	 * If we are in VM86 mode, add the segment offset to convert to a
2605 	 * linear address.
2606 	 */
2607 	if (regs->flags & X86_VM_MASK)
2608 		return 0x10 * regs->cs;
2609 
2610 	if (user_mode(regs) && regs->cs != __USER_CS)
2611 		return get_segment_base(regs->cs);
2612 #else
2613 	if (user_mode(regs) && !user_64bit_mode(regs) &&
2614 	    regs->cs != __USER32_CS)
2615 		return get_segment_base(regs->cs);
2616 #endif
2617 	return 0;
2618 }
2619 
2620 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2621 {
2622 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2623 		return perf_guest_cbs->get_guest_ip();
2624 
2625 	return regs->ip + code_segment_base(regs);
2626 }
2627 
2628 unsigned long perf_misc_flags(struct pt_regs *regs)
2629 {
2630 	int misc = 0;
2631 
2632 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2633 		if (perf_guest_cbs->is_user_mode())
2634 			misc |= PERF_RECORD_MISC_GUEST_USER;
2635 		else
2636 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2637 	} else {
2638 		if (user_mode(regs))
2639 			misc |= PERF_RECORD_MISC_USER;
2640 		else
2641 			misc |= PERF_RECORD_MISC_KERNEL;
2642 	}
2643 
2644 	if (regs->flags & PERF_EFLAGS_EXACT)
2645 		misc |= PERF_RECORD_MISC_EXACT_IP;
2646 
2647 	return misc;
2648 }
2649 
2650 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2651 {
2652 	cap->version		= x86_pmu.version;
2653 	cap->num_counters_gp	= x86_pmu.num_counters;
2654 	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2655 	cap->bit_width_gp	= x86_pmu.cntval_bits;
2656 	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2657 	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2658 	cap->events_mask_len	= x86_pmu.events_mask_len;
2659 }
2660 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
2661