xref: /openbmc/linux/arch/x86/events/core.c (revision bb3982b4)
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 	hwc->config |= config;
442 
443 	return 0;
444 }
445 
446 /*
447  * check that branch_sample_type is compatible with
448  * settings needed for precise_ip > 1 which implies
449  * using the LBR to capture ALL taken branches at the
450  * priv levels of the measurement
451  */
452 static inline int precise_br_compat(struct perf_event *event)
453 {
454 	u64 m = event->attr.branch_sample_type;
455 	u64 b = 0;
456 
457 	/* must capture all branches */
458 	if (!(m & PERF_SAMPLE_BRANCH_ANY))
459 		return 0;
460 
461 	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
462 
463 	if (!event->attr.exclude_user)
464 		b |= PERF_SAMPLE_BRANCH_USER;
465 
466 	if (!event->attr.exclude_kernel)
467 		b |= PERF_SAMPLE_BRANCH_KERNEL;
468 
469 	/*
470 	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
471 	 */
472 
473 	return m == b;
474 }
475 
476 int x86_pmu_max_precise(void)
477 {
478 	int precise = 0;
479 
480 	/* Support for constant skid */
481 	if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
482 		precise++;
483 
484 		/* Support for IP fixup */
485 		if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
486 			precise++;
487 
488 		if (x86_pmu.pebs_prec_dist)
489 			precise++;
490 	}
491 	return precise;
492 }
493 
494 int x86_pmu_hw_config(struct perf_event *event)
495 {
496 	if (event->attr.precise_ip) {
497 		int precise = x86_pmu_max_precise();
498 
499 		if (event->attr.precise_ip > precise)
500 			return -EOPNOTSUPP;
501 
502 		/* There's no sense in having PEBS for non sampling events: */
503 		if (!is_sampling_event(event))
504 			return -EINVAL;
505 	}
506 	/*
507 	 * check that PEBS LBR correction does not conflict with
508 	 * whatever the user is asking with attr->branch_sample_type
509 	 */
510 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
511 		u64 *br_type = &event->attr.branch_sample_type;
512 
513 		if (has_branch_stack(event)) {
514 			if (!precise_br_compat(event))
515 				return -EOPNOTSUPP;
516 
517 			/* branch_sample_type is compatible */
518 
519 		} else {
520 			/*
521 			 * user did not specify  branch_sample_type
522 			 *
523 			 * For PEBS fixups, we capture all
524 			 * the branches at the priv level of the
525 			 * event.
526 			 */
527 			*br_type = PERF_SAMPLE_BRANCH_ANY;
528 
529 			if (!event->attr.exclude_user)
530 				*br_type |= PERF_SAMPLE_BRANCH_USER;
531 
532 			if (!event->attr.exclude_kernel)
533 				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
534 		}
535 	}
536 
537 	if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
538 		event->attach_state |= PERF_ATTACH_TASK_DATA;
539 
540 	/*
541 	 * Generate PMC IRQs:
542 	 * (keep 'enabled' bit clear for now)
543 	 */
544 	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
545 
546 	/*
547 	 * Count user and OS events unless requested not to
548 	 */
549 	if (!event->attr.exclude_user)
550 		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
551 	if (!event->attr.exclude_kernel)
552 		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
553 
554 	if (event->attr.type == PERF_TYPE_RAW)
555 		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
556 
557 	if (event->attr.sample_period && x86_pmu.limit_period) {
558 		if (x86_pmu.limit_period(event, event->attr.sample_period) >
559 				event->attr.sample_period)
560 			return -EINVAL;
561 	}
562 
563 	/* sample_regs_user never support XMM registers */
564 	if (unlikely(event->attr.sample_regs_user & PEBS_XMM_REGS))
565 		return -EINVAL;
566 	/*
567 	 * Besides the general purpose registers, XMM registers may
568 	 * be collected in PEBS on some platforms, e.g. Icelake
569 	 */
570 	if (unlikely(event->attr.sample_regs_intr & PEBS_XMM_REGS)) {
571 		if (x86_pmu.pebs_no_xmm_regs)
572 			return -EINVAL;
573 
574 		if (!event->attr.precise_ip)
575 			return -EINVAL;
576 	}
577 
578 	return x86_setup_perfctr(event);
579 }
580 
581 /*
582  * Setup the hardware configuration for a given attr_type
583  */
584 static int __x86_pmu_event_init(struct perf_event *event)
585 {
586 	int err;
587 
588 	if (!x86_pmu_initialized())
589 		return -ENODEV;
590 
591 	err = x86_reserve_hardware();
592 	if (err)
593 		return err;
594 
595 	atomic_inc(&active_events);
596 	event->destroy = hw_perf_event_destroy;
597 
598 	event->hw.idx = -1;
599 	event->hw.last_cpu = -1;
600 	event->hw.last_tag = ~0ULL;
601 
602 	/* mark unused */
603 	event->hw.extra_reg.idx = EXTRA_REG_NONE;
604 	event->hw.branch_reg.idx = EXTRA_REG_NONE;
605 
606 	return x86_pmu.hw_config(event);
607 }
608 
609 void x86_pmu_disable_all(void)
610 {
611 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
612 	int idx;
613 
614 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
615 		u64 val;
616 
617 		if (!test_bit(idx, cpuc->active_mask))
618 			continue;
619 		rdmsrl(x86_pmu_config_addr(idx), val);
620 		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
621 			continue;
622 		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
623 		wrmsrl(x86_pmu_config_addr(idx), val);
624 	}
625 }
626 
627 /*
628  * There may be PMI landing after enabled=0. The PMI hitting could be before or
629  * after disable_all.
630  *
631  * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
632  * It will not be re-enabled in the NMI handler again, because enabled=0. After
633  * handling the NMI, disable_all will be called, which will not change the
634  * state either. If PMI hits after disable_all, the PMU is already disabled
635  * before entering NMI handler. The NMI handler will not change the state
636  * either.
637  *
638  * So either situation is harmless.
639  */
640 static void x86_pmu_disable(struct pmu *pmu)
641 {
642 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
643 
644 	if (!x86_pmu_initialized())
645 		return;
646 
647 	if (!cpuc->enabled)
648 		return;
649 
650 	cpuc->n_added = 0;
651 	cpuc->enabled = 0;
652 	barrier();
653 
654 	x86_pmu.disable_all();
655 }
656 
657 void x86_pmu_enable_all(int added)
658 {
659 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
660 	int idx;
661 
662 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
663 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
664 
665 		if (!test_bit(idx, cpuc->active_mask))
666 			continue;
667 
668 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
669 	}
670 }
671 
672 static struct pmu pmu;
673 
674 static inline int is_x86_event(struct perf_event *event)
675 {
676 	return event->pmu == &pmu;
677 }
678 
679 struct pmu *x86_get_pmu(void)
680 {
681 	return &pmu;
682 }
683 /*
684  * Event scheduler state:
685  *
686  * Assign events iterating over all events and counters, beginning
687  * with events with least weights first. Keep the current iterator
688  * state in struct sched_state.
689  */
690 struct sched_state {
691 	int	weight;
692 	int	event;		/* event index */
693 	int	counter;	/* counter index */
694 	int	unassigned;	/* number of events to be assigned left */
695 	int	nr_gp;		/* number of GP counters used */
696 	unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
697 };
698 
699 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
700 #define	SCHED_STATES_MAX	2
701 
702 struct perf_sched {
703 	int			max_weight;
704 	int			max_events;
705 	int			max_gp;
706 	int			saved_states;
707 	struct event_constraint	**constraints;
708 	struct sched_state	state;
709 	struct sched_state	saved[SCHED_STATES_MAX];
710 };
711 
712 /*
713  * Initialize interator that runs through all events and counters.
714  */
715 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
716 			    int num, int wmin, int wmax, int gpmax)
717 {
718 	int idx;
719 
720 	memset(sched, 0, sizeof(*sched));
721 	sched->max_events	= num;
722 	sched->max_weight	= wmax;
723 	sched->max_gp		= gpmax;
724 	sched->constraints	= constraints;
725 
726 	for (idx = 0; idx < num; idx++) {
727 		if (constraints[idx]->weight == wmin)
728 			break;
729 	}
730 
731 	sched->state.event	= idx;		/* start with min weight */
732 	sched->state.weight	= wmin;
733 	sched->state.unassigned	= num;
734 }
735 
736 static void perf_sched_save_state(struct perf_sched *sched)
737 {
738 	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
739 		return;
740 
741 	sched->saved[sched->saved_states] = sched->state;
742 	sched->saved_states++;
743 }
744 
745 static bool perf_sched_restore_state(struct perf_sched *sched)
746 {
747 	if (!sched->saved_states)
748 		return false;
749 
750 	sched->saved_states--;
751 	sched->state = sched->saved[sched->saved_states];
752 
753 	/* continue with next counter: */
754 	clear_bit(sched->state.counter++, sched->state.used);
755 
756 	return true;
757 }
758 
759 /*
760  * Select a counter for the current event to schedule. Return true on
761  * success.
762  */
763 static bool __perf_sched_find_counter(struct perf_sched *sched)
764 {
765 	struct event_constraint *c;
766 	int idx;
767 
768 	if (!sched->state.unassigned)
769 		return false;
770 
771 	if (sched->state.event >= sched->max_events)
772 		return false;
773 
774 	c = sched->constraints[sched->state.event];
775 	/* Prefer fixed purpose counters */
776 	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
777 		idx = INTEL_PMC_IDX_FIXED;
778 		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
779 			if (!__test_and_set_bit(idx, sched->state.used))
780 				goto done;
781 		}
782 	}
783 
784 	/* Grab the first unused counter starting with idx */
785 	idx = sched->state.counter;
786 	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
787 		if (!__test_and_set_bit(idx, sched->state.used)) {
788 			if (sched->state.nr_gp++ >= sched->max_gp)
789 				return false;
790 
791 			goto done;
792 		}
793 	}
794 
795 	return false;
796 
797 done:
798 	sched->state.counter = idx;
799 
800 	if (c->overlap)
801 		perf_sched_save_state(sched);
802 
803 	return true;
804 }
805 
806 static bool perf_sched_find_counter(struct perf_sched *sched)
807 {
808 	while (!__perf_sched_find_counter(sched)) {
809 		if (!perf_sched_restore_state(sched))
810 			return false;
811 	}
812 
813 	return true;
814 }
815 
816 /*
817  * Go through all unassigned events and find the next one to schedule.
818  * Take events with the least weight first. Return true on success.
819  */
820 static bool perf_sched_next_event(struct perf_sched *sched)
821 {
822 	struct event_constraint *c;
823 
824 	if (!sched->state.unassigned || !--sched->state.unassigned)
825 		return false;
826 
827 	do {
828 		/* next event */
829 		sched->state.event++;
830 		if (sched->state.event >= sched->max_events) {
831 			/* next weight */
832 			sched->state.event = 0;
833 			sched->state.weight++;
834 			if (sched->state.weight > sched->max_weight)
835 				return false;
836 		}
837 		c = sched->constraints[sched->state.event];
838 	} while (c->weight != sched->state.weight);
839 
840 	sched->state.counter = 0;	/* start with first counter */
841 
842 	return true;
843 }
844 
845 /*
846  * Assign a counter for each event.
847  */
848 int perf_assign_events(struct event_constraint **constraints, int n,
849 			int wmin, int wmax, int gpmax, int *assign)
850 {
851 	struct perf_sched sched;
852 
853 	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
854 
855 	do {
856 		if (!perf_sched_find_counter(&sched))
857 			break;	/* failed */
858 		if (assign)
859 			assign[sched.state.event] = sched.state.counter;
860 	} while (perf_sched_next_event(&sched));
861 
862 	return sched.state.unassigned;
863 }
864 EXPORT_SYMBOL_GPL(perf_assign_events);
865 
866 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
867 {
868 	struct event_constraint *c;
869 	unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
870 	struct perf_event *e;
871 	int n0, i, wmin, wmax, unsched = 0;
872 	struct hw_perf_event *hwc;
873 
874 	bitmap_zero(used_mask, X86_PMC_IDX_MAX);
875 
876 	/*
877 	 * Compute the number of events already present; see x86_pmu_add(),
878 	 * validate_group() and x86_pmu_commit_txn(). For the former two
879 	 * cpuc->n_events hasn't been updated yet, while for the latter
880 	 * cpuc->n_txn contains the number of events added in the current
881 	 * transaction.
882 	 */
883 	n0 = cpuc->n_events;
884 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
885 		n0 -= cpuc->n_txn;
886 
887 	if (x86_pmu.start_scheduling)
888 		x86_pmu.start_scheduling(cpuc);
889 
890 	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
891 		c = cpuc->event_constraint[i];
892 
893 		/*
894 		 * Previously scheduled events should have a cached constraint,
895 		 * while new events should not have one.
896 		 */
897 		WARN_ON_ONCE((c && i >= n0) || (!c && i < n0));
898 
899 		/*
900 		 * Request constraints for new events; or for those events that
901 		 * have a dynamic constraint -- for those the constraint can
902 		 * change due to external factors (sibling state, allow_tfa).
903 		 */
904 		if (!c || (c->flags & PERF_X86_EVENT_DYNAMIC)) {
905 			c = x86_pmu.get_event_constraints(cpuc, i, cpuc->event_list[i]);
906 			cpuc->event_constraint[i] = c;
907 		}
908 
909 		wmin = min(wmin, c->weight);
910 		wmax = max(wmax, c->weight);
911 	}
912 
913 	/*
914 	 * fastpath, try to reuse previous register
915 	 */
916 	for (i = 0; i < n; i++) {
917 		hwc = &cpuc->event_list[i]->hw;
918 		c = cpuc->event_constraint[i];
919 
920 		/* never assigned */
921 		if (hwc->idx == -1)
922 			break;
923 
924 		/* constraint still honored */
925 		if (!test_bit(hwc->idx, c->idxmsk))
926 			break;
927 
928 		/* not already used */
929 		if (test_bit(hwc->idx, used_mask))
930 			break;
931 
932 		__set_bit(hwc->idx, used_mask);
933 		if (assign)
934 			assign[i] = hwc->idx;
935 	}
936 
937 	/* slow path */
938 	if (i != n) {
939 		int gpmax = x86_pmu.num_counters;
940 
941 		/*
942 		 * Do not allow scheduling of more than half the available
943 		 * generic counters.
944 		 *
945 		 * This helps avoid counter starvation of sibling thread by
946 		 * ensuring at most half the counters cannot be in exclusive
947 		 * mode. There is no designated counters for the limits. Any
948 		 * N/2 counters can be used. This helps with events with
949 		 * specific counter constraints.
950 		 */
951 		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
952 		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
953 			gpmax /= 2;
954 
955 		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
956 					     wmax, gpmax, assign);
957 	}
958 
959 	/*
960 	 * In case of success (unsched = 0), mark events as committed,
961 	 * so we do not put_constraint() in case new events are added
962 	 * and fail to be scheduled
963 	 *
964 	 * We invoke the lower level commit callback to lock the resource
965 	 *
966 	 * We do not need to do all of this in case we are called to
967 	 * validate an event group (assign == NULL)
968 	 */
969 	if (!unsched && assign) {
970 		for (i = 0; i < n; i++) {
971 			e = cpuc->event_list[i];
972 			if (x86_pmu.commit_scheduling)
973 				x86_pmu.commit_scheduling(cpuc, i, assign[i]);
974 		}
975 	} else {
976 		for (i = n0; i < n; i++) {
977 			e = cpuc->event_list[i];
978 
979 			/*
980 			 * release events that failed scheduling
981 			 */
982 			if (x86_pmu.put_event_constraints)
983 				x86_pmu.put_event_constraints(cpuc, e);
984 
985 			cpuc->event_constraint[i] = NULL;
986 		}
987 	}
988 
989 	if (x86_pmu.stop_scheduling)
990 		x86_pmu.stop_scheduling(cpuc);
991 
992 	return unsched ? -EINVAL : 0;
993 }
994 
995 /*
996  * dogrp: true if must collect siblings events (group)
997  * returns total number of events and error code
998  */
999 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1000 {
1001 	struct perf_event *event;
1002 	int n, max_count;
1003 
1004 	max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
1005 
1006 	/* current number of events already accepted */
1007 	n = cpuc->n_events;
1008 
1009 	if (is_x86_event(leader)) {
1010 		if (n >= max_count)
1011 			return -EINVAL;
1012 		cpuc->event_list[n] = leader;
1013 		n++;
1014 	}
1015 	if (!dogrp)
1016 		return n;
1017 
1018 	for_each_sibling_event(event, leader) {
1019 		if (!is_x86_event(event) ||
1020 		    event->state <= PERF_EVENT_STATE_OFF)
1021 			continue;
1022 
1023 		if (n >= max_count)
1024 			return -EINVAL;
1025 
1026 		cpuc->event_list[n] = event;
1027 		n++;
1028 	}
1029 	return n;
1030 }
1031 
1032 static inline void x86_assign_hw_event(struct perf_event *event,
1033 				struct cpu_hw_events *cpuc, int i)
1034 {
1035 	struct hw_perf_event *hwc = &event->hw;
1036 
1037 	hwc->idx = cpuc->assign[i];
1038 	hwc->last_cpu = smp_processor_id();
1039 	hwc->last_tag = ++cpuc->tags[i];
1040 
1041 	if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
1042 		hwc->config_base = 0;
1043 		hwc->event_base	= 0;
1044 	} else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1045 		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1046 		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
1047 		hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
1048 	} else {
1049 		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1050 		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1051 		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1052 	}
1053 }
1054 
1055 /**
1056  * x86_perf_rdpmc_index - Return PMC counter used for event
1057  * @event: the perf_event to which the PMC counter was assigned
1058  *
1059  * The counter assigned to this performance event may change if interrupts
1060  * are enabled. This counter should thus never be used while interrupts are
1061  * enabled. Before this function is used to obtain the assigned counter the
1062  * event should be checked for validity using, for example,
1063  * perf_event_read_local(), within the same interrupt disabled section in
1064  * which this counter is planned to be used.
1065  *
1066  * Return: The index of the performance monitoring counter assigned to
1067  * @perf_event.
1068  */
1069 int x86_perf_rdpmc_index(struct perf_event *event)
1070 {
1071 	lockdep_assert_irqs_disabled();
1072 
1073 	return event->hw.event_base_rdpmc;
1074 }
1075 
1076 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1077 					struct cpu_hw_events *cpuc,
1078 					int i)
1079 {
1080 	return hwc->idx == cpuc->assign[i] &&
1081 		hwc->last_cpu == smp_processor_id() &&
1082 		hwc->last_tag == cpuc->tags[i];
1083 }
1084 
1085 static void x86_pmu_start(struct perf_event *event, int flags);
1086 
1087 static void x86_pmu_enable(struct pmu *pmu)
1088 {
1089 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1090 	struct perf_event *event;
1091 	struct hw_perf_event *hwc;
1092 	int i, added = cpuc->n_added;
1093 
1094 	if (!x86_pmu_initialized())
1095 		return;
1096 
1097 	if (cpuc->enabled)
1098 		return;
1099 
1100 	if (cpuc->n_added) {
1101 		int n_running = cpuc->n_events - cpuc->n_added;
1102 		/*
1103 		 * apply assignment obtained either from
1104 		 * hw_perf_group_sched_in() or x86_pmu_enable()
1105 		 *
1106 		 * step1: save events moving to new counters
1107 		 */
1108 		for (i = 0; i < n_running; i++) {
1109 			event = cpuc->event_list[i];
1110 			hwc = &event->hw;
1111 
1112 			/*
1113 			 * we can avoid reprogramming counter if:
1114 			 * - assigned same counter as last time
1115 			 * - running on same CPU as last time
1116 			 * - no other event has used the counter since
1117 			 */
1118 			if (hwc->idx == -1 ||
1119 			    match_prev_assignment(hwc, cpuc, i))
1120 				continue;
1121 
1122 			/*
1123 			 * Ensure we don't accidentally enable a stopped
1124 			 * counter simply because we rescheduled.
1125 			 */
1126 			if (hwc->state & PERF_HES_STOPPED)
1127 				hwc->state |= PERF_HES_ARCH;
1128 
1129 			x86_pmu_stop(event, PERF_EF_UPDATE);
1130 		}
1131 
1132 		/*
1133 		 * step2: reprogram moved events into new counters
1134 		 */
1135 		for (i = 0; i < cpuc->n_events; i++) {
1136 			event = cpuc->event_list[i];
1137 			hwc = &event->hw;
1138 
1139 			if (!match_prev_assignment(hwc, cpuc, i))
1140 				x86_assign_hw_event(event, cpuc, i);
1141 			else if (i < n_running)
1142 				continue;
1143 
1144 			if (hwc->state & PERF_HES_ARCH)
1145 				continue;
1146 
1147 			x86_pmu_start(event, PERF_EF_RELOAD);
1148 		}
1149 		cpuc->n_added = 0;
1150 		perf_events_lapic_init();
1151 	}
1152 
1153 	cpuc->enabled = 1;
1154 	barrier();
1155 
1156 	x86_pmu.enable_all(added);
1157 }
1158 
1159 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1160 
1161 /*
1162  * Set the next IRQ period, based on the hwc->period_left value.
1163  * To be called with the event disabled in hw:
1164  */
1165 int x86_perf_event_set_period(struct perf_event *event)
1166 {
1167 	struct hw_perf_event *hwc = &event->hw;
1168 	s64 left = local64_read(&hwc->period_left);
1169 	s64 period = hwc->sample_period;
1170 	int ret = 0, idx = hwc->idx;
1171 
1172 	if (idx == INTEL_PMC_IDX_FIXED_BTS)
1173 		return 0;
1174 
1175 	/*
1176 	 * If we are way outside a reasonable range then just skip forward:
1177 	 */
1178 	if (unlikely(left <= -period)) {
1179 		left = period;
1180 		local64_set(&hwc->period_left, left);
1181 		hwc->last_period = period;
1182 		ret = 1;
1183 	}
1184 
1185 	if (unlikely(left <= 0)) {
1186 		left += period;
1187 		local64_set(&hwc->period_left, left);
1188 		hwc->last_period = period;
1189 		ret = 1;
1190 	}
1191 	/*
1192 	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1193 	 */
1194 	if (unlikely(left < 2))
1195 		left = 2;
1196 
1197 	if (left > x86_pmu.max_period)
1198 		left = x86_pmu.max_period;
1199 
1200 	if (x86_pmu.limit_period)
1201 		left = x86_pmu.limit_period(event, left);
1202 
1203 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1204 
1205 	/*
1206 	 * The hw event starts counting from this event offset,
1207 	 * mark it to be able to extra future deltas:
1208 	 */
1209 	local64_set(&hwc->prev_count, (u64)-left);
1210 
1211 	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1212 
1213 	/*
1214 	 * Due to erratum on certan cpu we need
1215 	 * a second write to be sure the register
1216 	 * is updated properly
1217 	 */
1218 	if (x86_pmu.perfctr_second_write) {
1219 		wrmsrl(hwc->event_base,
1220 			(u64)(-left) & x86_pmu.cntval_mask);
1221 	}
1222 
1223 	perf_event_update_userpage(event);
1224 
1225 	return ret;
1226 }
1227 
1228 void x86_pmu_enable_event(struct perf_event *event)
1229 {
1230 	if (__this_cpu_read(cpu_hw_events.enabled))
1231 		__x86_pmu_enable_event(&event->hw,
1232 				       ARCH_PERFMON_EVENTSEL_ENABLE);
1233 }
1234 
1235 /*
1236  * Add a single event to the PMU.
1237  *
1238  * The event is added to the group of enabled events
1239  * but only if it can be scehduled with existing events.
1240  */
1241 static int x86_pmu_add(struct perf_event *event, int flags)
1242 {
1243 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1244 	struct hw_perf_event *hwc;
1245 	int assign[X86_PMC_IDX_MAX];
1246 	int n, n0, ret;
1247 
1248 	hwc = &event->hw;
1249 
1250 	n0 = cpuc->n_events;
1251 	ret = n = collect_events(cpuc, event, false);
1252 	if (ret < 0)
1253 		goto out;
1254 
1255 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1256 	if (!(flags & PERF_EF_START))
1257 		hwc->state |= PERF_HES_ARCH;
1258 
1259 	/*
1260 	 * If group events scheduling transaction was started,
1261 	 * skip the schedulability test here, it will be performed
1262 	 * at commit time (->commit_txn) as a whole.
1263 	 *
1264 	 * If commit fails, we'll call ->del() on all events
1265 	 * for which ->add() was called.
1266 	 */
1267 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1268 		goto done_collect;
1269 
1270 	ret = x86_pmu.schedule_events(cpuc, n, assign);
1271 	if (ret)
1272 		goto out;
1273 	/*
1274 	 * copy new assignment, now we know it is possible
1275 	 * will be used by hw_perf_enable()
1276 	 */
1277 	memcpy(cpuc->assign, assign, n*sizeof(int));
1278 
1279 done_collect:
1280 	/*
1281 	 * Commit the collect_events() state. See x86_pmu_del() and
1282 	 * x86_pmu_*_txn().
1283 	 */
1284 	cpuc->n_events = n;
1285 	cpuc->n_added += n - n0;
1286 	cpuc->n_txn += n - n0;
1287 
1288 	if (x86_pmu.add) {
1289 		/*
1290 		 * This is before x86_pmu_enable() will call x86_pmu_start(),
1291 		 * so we enable LBRs before an event needs them etc..
1292 		 */
1293 		x86_pmu.add(event);
1294 	}
1295 
1296 	ret = 0;
1297 out:
1298 	return ret;
1299 }
1300 
1301 static void x86_pmu_start(struct perf_event *event, int flags)
1302 {
1303 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1304 	int idx = event->hw.idx;
1305 
1306 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1307 		return;
1308 
1309 	if (WARN_ON_ONCE(idx == -1))
1310 		return;
1311 
1312 	if (flags & PERF_EF_RELOAD) {
1313 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1314 		x86_perf_event_set_period(event);
1315 	}
1316 
1317 	event->hw.state = 0;
1318 
1319 	cpuc->events[idx] = event;
1320 	__set_bit(idx, cpuc->active_mask);
1321 	__set_bit(idx, cpuc->running);
1322 	x86_pmu.enable(event);
1323 	perf_event_update_userpage(event);
1324 }
1325 
1326 void perf_event_print_debug(void)
1327 {
1328 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1329 	u64 pebs, debugctl;
1330 	struct cpu_hw_events *cpuc;
1331 	unsigned long flags;
1332 	int cpu, idx;
1333 
1334 	if (!x86_pmu.num_counters)
1335 		return;
1336 
1337 	local_irq_save(flags);
1338 
1339 	cpu = smp_processor_id();
1340 	cpuc = &per_cpu(cpu_hw_events, cpu);
1341 
1342 	if (x86_pmu.version >= 2) {
1343 		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1344 		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1345 		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1346 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1347 
1348 		pr_info("\n");
1349 		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1350 		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1351 		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1352 		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1353 		if (x86_pmu.pebs_constraints) {
1354 			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1355 			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1356 		}
1357 		if (x86_pmu.lbr_nr) {
1358 			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1359 			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1360 		}
1361 	}
1362 	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1363 
1364 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1365 		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1366 		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1367 
1368 		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1369 
1370 		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1371 			cpu, idx, pmc_ctrl);
1372 		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1373 			cpu, idx, pmc_count);
1374 		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1375 			cpu, idx, prev_left);
1376 	}
1377 	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1378 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1379 
1380 		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1381 			cpu, idx, pmc_count);
1382 	}
1383 	local_irq_restore(flags);
1384 }
1385 
1386 void x86_pmu_stop(struct perf_event *event, int flags)
1387 {
1388 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1389 	struct hw_perf_event *hwc = &event->hw;
1390 
1391 	if (test_bit(hwc->idx, cpuc->active_mask)) {
1392 		x86_pmu.disable(event);
1393 		__clear_bit(hwc->idx, cpuc->active_mask);
1394 		cpuc->events[hwc->idx] = NULL;
1395 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1396 		hwc->state |= PERF_HES_STOPPED;
1397 	}
1398 
1399 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1400 		/*
1401 		 * Drain the remaining delta count out of a event
1402 		 * that we are disabling:
1403 		 */
1404 		x86_perf_event_update(event);
1405 		hwc->state |= PERF_HES_UPTODATE;
1406 	}
1407 }
1408 
1409 static void x86_pmu_del(struct perf_event *event, int flags)
1410 {
1411 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1412 	int i;
1413 
1414 	/*
1415 	 * If we're called during a txn, we only need to undo x86_pmu.add.
1416 	 * The events never got scheduled and ->cancel_txn will truncate
1417 	 * the event_list.
1418 	 *
1419 	 * XXX assumes any ->del() called during a TXN will only be on
1420 	 * an event added during that same TXN.
1421 	 */
1422 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1423 		goto do_del;
1424 
1425 	/*
1426 	 * Not a TXN, therefore cleanup properly.
1427 	 */
1428 	x86_pmu_stop(event, PERF_EF_UPDATE);
1429 
1430 	for (i = 0; i < cpuc->n_events; i++) {
1431 		if (event == cpuc->event_list[i])
1432 			break;
1433 	}
1434 
1435 	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1436 		return;
1437 
1438 	/* If we have a newly added event; make sure to decrease n_added. */
1439 	if (i >= cpuc->n_events - cpuc->n_added)
1440 		--cpuc->n_added;
1441 
1442 	if (x86_pmu.put_event_constraints)
1443 		x86_pmu.put_event_constraints(cpuc, event);
1444 
1445 	/* Delete the array entry. */
1446 	while (++i < cpuc->n_events) {
1447 		cpuc->event_list[i-1] = cpuc->event_list[i];
1448 		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1449 	}
1450 	cpuc->event_constraint[i-1] = NULL;
1451 	--cpuc->n_events;
1452 
1453 	perf_event_update_userpage(event);
1454 
1455 do_del:
1456 	if (x86_pmu.del) {
1457 		/*
1458 		 * This is after x86_pmu_stop(); so we disable LBRs after any
1459 		 * event can need them etc..
1460 		 */
1461 		x86_pmu.del(event);
1462 	}
1463 }
1464 
1465 int x86_pmu_handle_irq(struct pt_regs *regs)
1466 {
1467 	struct perf_sample_data data;
1468 	struct cpu_hw_events *cpuc;
1469 	struct perf_event *event;
1470 	int idx, handled = 0;
1471 	u64 val;
1472 
1473 	cpuc = this_cpu_ptr(&cpu_hw_events);
1474 
1475 	/*
1476 	 * Some chipsets need to unmask the LVTPC in a particular spot
1477 	 * inside the nmi handler.  As a result, the unmasking was pushed
1478 	 * into all the nmi handlers.
1479 	 *
1480 	 * This generic handler doesn't seem to have any issues where the
1481 	 * unmasking occurs so it was left at the top.
1482 	 */
1483 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1484 
1485 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1486 		if (!test_bit(idx, cpuc->active_mask))
1487 			continue;
1488 
1489 		event = cpuc->events[idx];
1490 
1491 		val = x86_perf_event_update(event);
1492 		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1493 			continue;
1494 
1495 		/*
1496 		 * event overflow
1497 		 */
1498 		handled++;
1499 		perf_sample_data_init(&data, 0, event->hw.last_period);
1500 
1501 		if (!x86_perf_event_set_period(event))
1502 			continue;
1503 
1504 		if (perf_event_overflow(event, &data, regs))
1505 			x86_pmu_stop(event, 0);
1506 	}
1507 
1508 	if (handled)
1509 		inc_irq_stat(apic_perf_irqs);
1510 
1511 	return handled;
1512 }
1513 
1514 void perf_events_lapic_init(void)
1515 {
1516 	if (!x86_pmu.apic || !x86_pmu_initialized())
1517 		return;
1518 
1519 	/*
1520 	 * Always use NMI for PMU
1521 	 */
1522 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1523 }
1524 
1525 static int
1526 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1527 {
1528 	u64 start_clock;
1529 	u64 finish_clock;
1530 	int ret;
1531 
1532 	/*
1533 	 * All PMUs/events that share this PMI handler should make sure to
1534 	 * increment active_events for their events.
1535 	 */
1536 	if (!atomic_read(&active_events))
1537 		return NMI_DONE;
1538 
1539 	start_clock = sched_clock();
1540 	ret = x86_pmu.handle_irq(regs);
1541 	finish_clock = sched_clock();
1542 
1543 	perf_sample_event_took(finish_clock - start_clock);
1544 
1545 	return ret;
1546 }
1547 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1548 
1549 struct event_constraint emptyconstraint;
1550 struct event_constraint unconstrained;
1551 
1552 static int x86_pmu_prepare_cpu(unsigned int cpu)
1553 {
1554 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1555 	int i;
1556 
1557 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1558 		cpuc->kfree_on_online[i] = NULL;
1559 	if (x86_pmu.cpu_prepare)
1560 		return x86_pmu.cpu_prepare(cpu);
1561 	return 0;
1562 }
1563 
1564 static int x86_pmu_dead_cpu(unsigned int cpu)
1565 {
1566 	if (x86_pmu.cpu_dead)
1567 		x86_pmu.cpu_dead(cpu);
1568 	return 0;
1569 }
1570 
1571 static int x86_pmu_online_cpu(unsigned int cpu)
1572 {
1573 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1574 	int i;
1575 
1576 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1577 		kfree(cpuc->kfree_on_online[i]);
1578 		cpuc->kfree_on_online[i] = NULL;
1579 	}
1580 	return 0;
1581 }
1582 
1583 static int x86_pmu_starting_cpu(unsigned int cpu)
1584 {
1585 	if (x86_pmu.cpu_starting)
1586 		x86_pmu.cpu_starting(cpu);
1587 	return 0;
1588 }
1589 
1590 static int x86_pmu_dying_cpu(unsigned int cpu)
1591 {
1592 	if (x86_pmu.cpu_dying)
1593 		x86_pmu.cpu_dying(cpu);
1594 	return 0;
1595 }
1596 
1597 static void __init pmu_check_apic(void)
1598 {
1599 	if (boot_cpu_has(X86_FEATURE_APIC))
1600 		return;
1601 
1602 	x86_pmu.apic = 0;
1603 	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1604 	pr_info("no hardware sampling interrupt available.\n");
1605 
1606 	/*
1607 	 * If we have a PMU initialized but no APIC
1608 	 * interrupts, we cannot sample hardware
1609 	 * events (user-space has to fall back and
1610 	 * sample via a hrtimer based software event):
1611 	 */
1612 	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1613 
1614 }
1615 
1616 static struct attribute_group x86_pmu_format_group __ro_after_init = {
1617 	.name = "format",
1618 	.attrs = NULL,
1619 };
1620 
1621 /*
1622  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1623  * out of events_attr attributes.
1624  */
1625 static void __init filter_events(struct attribute **attrs)
1626 {
1627 	struct device_attribute *d;
1628 	struct perf_pmu_events_attr *pmu_attr;
1629 	int offset = 0;
1630 	int i, j;
1631 
1632 	for (i = 0; attrs[i]; i++) {
1633 		d = (struct device_attribute *)attrs[i];
1634 		pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1635 		/* str trumps id */
1636 		if (pmu_attr->event_str)
1637 			continue;
1638 		if (x86_pmu.event_map(i + offset))
1639 			continue;
1640 
1641 		for (j = i; attrs[j]; j++)
1642 			attrs[j] = attrs[j + 1];
1643 
1644 		/* Check the shifted attr. */
1645 		i--;
1646 
1647 		/*
1648 		 * event_map() is index based, the attrs array is organized
1649 		 * by increasing event index. If we shift the events, then
1650 		 * we need to compensate for the event_map(), otherwise
1651 		 * we are looking up the wrong event in the map
1652 		 */
1653 		offset++;
1654 	}
1655 }
1656 
1657 /* Merge two pointer arrays */
1658 __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1659 {
1660 	struct attribute **new;
1661 	int j, i;
1662 
1663 	for (j = 0; a && a[j]; j++)
1664 		;
1665 	for (i = 0; b && b[i]; i++)
1666 		j++;
1667 	j++;
1668 
1669 	new = kmalloc_array(j, sizeof(struct attribute *), GFP_KERNEL);
1670 	if (!new)
1671 		return NULL;
1672 
1673 	j = 0;
1674 	for (i = 0; a && a[i]; i++)
1675 		new[j++] = a[i];
1676 	for (i = 0; b && b[i]; i++)
1677 		new[j++] = b[i];
1678 	new[j] = NULL;
1679 
1680 	return new;
1681 }
1682 
1683 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1684 {
1685 	struct perf_pmu_events_attr *pmu_attr = \
1686 		container_of(attr, struct perf_pmu_events_attr, attr);
1687 	u64 config = x86_pmu.event_map(pmu_attr->id);
1688 
1689 	/* string trumps id */
1690 	if (pmu_attr->event_str)
1691 		return sprintf(page, "%s", pmu_attr->event_str);
1692 
1693 	return x86_pmu.events_sysfs_show(page, config);
1694 }
1695 EXPORT_SYMBOL_GPL(events_sysfs_show);
1696 
1697 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1698 			  char *page)
1699 {
1700 	struct perf_pmu_events_ht_attr *pmu_attr =
1701 		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1702 
1703 	/*
1704 	 * Report conditional events depending on Hyper-Threading.
1705 	 *
1706 	 * This is overly conservative as usually the HT special
1707 	 * handling is not needed if the other CPU thread is idle.
1708 	 *
1709 	 * Note this does not (and cannot) handle the case when thread
1710 	 * siblings are invisible, for example with virtualization
1711 	 * if they are owned by some other guest.  The user tool
1712 	 * has to re-read when a thread sibling gets onlined later.
1713 	 */
1714 	return sprintf(page, "%s",
1715 			topology_max_smt_threads() > 1 ?
1716 			pmu_attr->event_str_ht :
1717 			pmu_attr->event_str_noht);
1718 }
1719 
1720 EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1721 EVENT_ATTR(instructions,		INSTRUCTIONS		);
1722 EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1723 EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1724 EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1725 EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1726 EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1727 EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1728 EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1729 EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1730 
1731 static struct attribute *empty_attrs;
1732 
1733 static struct attribute *events_attr[] = {
1734 	EVENT_PTR(CPU_CYCLES),
1735 	EVENT_PTR(INSTRUCTIONS),
1736 	EVENT_PTR(CACHE_REFERENCES),
1737 	EVENT_PTR(CACHE_MISSES),
1738 	EVENT_PTR(BRANCH_INSTRUCTIONS),
1739 	EVENT_PTR(BRANCH_MISSES),
1740 	EVENT_PTR(BUS_CYCLES),
1741 	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1742 	EVENT_PTR(STALLED_CYCLES_BACKEND),
1743 	EVENT_PTR(REF_CPU_CYCLES),
1744 	NULL,
1745 };
1746 
1747 static struct attribute_group x86_pmu_events_group __ro_after_init = {
1748 	.name = "events",
1749 	.attrs = events_attr,
1750 };
1751 
1752 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1753 {
1754 	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1755 	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1756 	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1757 	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1758 	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1759 	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1760 	ssize_t ret;
1761 
1762 	/*
1763 	* We have whole page size to spend and just little data
1764 	* to write, so we can safely use sprintf.
1765 	*/
1766 	ret = sprintf(page, "event=0x%02llx", event);
1767 
1768 	if (umask)
1769 		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1770 
1771 	if (edge)
1772 		ret += sprintf(page + ret, ",edge");
1773 
1774 	if (pc)
1775 		ret += sprintf(page + ret, ",pc");
1776 
1777 	if (any)
1778 		ret += sprintf(page + ret, ",any");
1779 
1780 	if (inv)
1781 		ret += sprintf(page + ret, ",inv");
1782 
1783 	if (cmask)
1784 		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1785 
1786 	ret += sprintf(page + ret, "\n");
1787 
1788 	return ret;
1789 }
1790 
1791 static struct attribute_group x86_pmu_attr_group;
1792 static struct attribute_group x86_pmu_caps_group;
1793 
1794 static int __init init_hw_perf_events(void)
1795 {
1796 	struct x86_pmu_quirk *quirk;
1797 	int err;
1798 
1799 	pr_info("Performance Events: ");
1800 
1801 	switch (boot_cpu_data.x86_vendor) {
1802 	case X86_VENDOR_INTEL:
1803 		err = intel_pmu_init();
1804 		break;
1805 	case X86_VENDOR_AMD:
1806 		err = amd_pmu_init();
1807 		break;
1808 	case X86_VENDOR_HYGON:
1809 		err = amd_pmu_init();
1810 		x86_pmu.name = "HYGON";
1811 		break;
1812 	default:
1813 		err = -ENOTSUPP;
1814 	}
1815 	if (err != 0) {
1816 		pr_cont("no PMU driver, software events only.\n");
1817 		return 0;
1818 	}
1819 
1820 	pmu_check_apic();
1821 
1822 	/* sanity check that the hardware exists or is emulated */
1823 	if (!check_hw_exists())
1824 		return 0;
1825 
1826 	pr_cont("%s PMU driver.\n", x86_pmu.name);
1827 
1828 	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1829 
1830 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1831 		quirk->func();
1832 
1833 	if (!x86_pmu.intel_ctrl)
1834 		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1835 
1836 	perf_events_lapic_init();
1837 	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1838 
1839 	unconstrained = (struct event_constraint)
1840 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1841 				   0, x86_pmu.num_counters, 0, 0);
1842 
1843 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1844 
1845 	if (x86_pmu.caps_attrs) {
1846 		struct attribute **tmp;
1847 
1848 		tmp = merge_attr(x86_pmu_caps_group.attrs, x86_pmu.caps_attrs);
1849 		if (!WARN_ON(!tmp))
1850 			x86_pmu_caps_group.attrs = tmp;
1851 	}
1852 
1853 	if (x86_pmu.event_attrs)
1854 		x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1855 
1856 	if (!x86_pmu.events_sysfs_show)
1857 		x86_pmu_events_group.attrs = &empty_attrs;
1858 	else
1859 		filter_events(x86_pmu_events_group.attrs);
1860 
1861 	if (x86_pmu.cpu_events) {
1862 		struct attribute **tmp;
1863 
1864 		tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1865 		if (!WARN_ON(!tmp))
1866 			x86_pmu_events_group.attrs = tmp;
1867 	}
1868 
1869 	if (x86_pmu.attrs) {
1870 		struct attribute **tmp;
1871 
1872 		tmp = merge_attr(x86_pmu_attr_group.attrs, x86_pmu.attrs);
1873 		if (!WARN_ON(!tmp))
1874 			x86_pmu_attr_group.attrs = tmp;
1875 	}
1876 
1877 	pr_info("... version:                %d\n",     x86_pmu.version);
1878 	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1879 	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1880 	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1881 	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1882 	pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1883 	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
1884 
1885 	/*
1886 	 * Install callbacks. Core will call them for each online
1887 	 * cpu.
1888 	 */
1889 	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
1890 				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
1891 	if (err)
1892 		return err;
1893 
1894 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
1895 				"perf/x86:starting", x86_pmu_starting_cpu,
1896 				x86_pmu_dying_cpu);
1897 	if (err)
1898 		goto out;
1899 
1900 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
1901 				x86_pmu_online_cpu, NULL);
1902 	if (err)
1903 		goto out1;
1904 
1905 	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1906 	if (err)
1907 		goto out2;
1908 
1909 	return 0;
1910 
1911 out2:
1912 	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
1913 out1:
1914 	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
1915 out:
1916 	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
1917 	return err;
1918 }
1919 early_initcall(init_hw_perf_events);
1920 
1921 static inline void x86_pmu_read(struct perf_event *event)
1922 {
1923 	if (x86_pmu.read)
1924 		return x86_pmu.read(event);
1925 	x86_perf_event_update(event);
1926 }
1927 
1928 /*
1929  * Start group events scheduling transaction
1930  * Set the flag to make pmu::enable() not perform the
1931  * schedulability test, it will be performed at commit time
1932  *
1933  * We only support PERF_PMU_TXN_ADD transactions. Save the
1934  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
1935  * transactions.
1936  */
1937 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
1938 {
1939 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1940 
1941 	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
1942 
1943 	cpuc->txn_flags = txn_flags;
1944 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1945 		return;
1946 
1947 	perf_pmu_disable(pmu);
1948 	__this_cpu_write(cpu_hw_events.n_txn, 0);
1949 }
1950 
1951 /*
1952  * Stop group events scheduling transaction
1953  * Clear the flag and pmu::enable() will perform the
1954  * schedulability test.
1955  */
1956 static void x86_pmu_cancel_txn(struct pmu *pmu)
1957 {
1958 	unsigned int txn_flags;
1959 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1960 
1961 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1962 
1963 	txn_flags = cpuc->txn_flags;
1964 	cpuc->txn_flags = 0;
1965 	if (txn_flags & ~PERF_PMU_TXN_ADD)
1966 		return;
1967 
1968 	/*
1969 	 * Truncate collected array by the number of events added in this
1970 	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1971 	 */
1972 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1973 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1974 	perf_pmu_enable(pmu);
1975 }
1976 
1977 /*
1978  * Commit group events scheduling transaction
1979  * Perform the group schedulability test as a whole
1980  * Return 0 if success
1981  *
1982  * Does not cancel the transaction on failure; expects the caller to do this.
1983  */
1984 static int x86_pmu_commit_txn(struct pmu *pmu)
1985 {
1986 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1987 	int assign[X86_PMC_IDX_MAX];
1988 	int n, ret;
1989 
1990 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
1991 
1992 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
1993 		cpuc->txn_flags = 0;
1994 		return 0;
1995 	}
1996 
1997 	n = cpuc->n_events;
1998 
1999 	if (!x86_pmu_initialized())
2000 		return -EAGAIN;
2001 
2002 	ret = x86_pmu.schedule_events(cpuc, n, assign);
2003 	if (ret)
2004 		return ret;
2005 
2006 	/*
2007 	 * copy new assignment, now we know it is possible
2008 	 * will be used by hw_perf_enable()
2009 	 */
2010 	memcpy(cpuc->assign, assign, n*sizeof(int));
2011 
2012 	cpuc->txn_flags = 0;
2013 	perf_pmu_enable(pmu);
2014 	return 0;
2015 }
2016 /*
2017  * a fake_cpuc is used to validate event groups. Due to
2018  * the extra reg logic, we need to also allocate a fake
2019  * per_core and per_cpu structure. Otherwise, group events
2020  * using extra reg may conflict without the kernel being
2021  * able to catch this when the last event gets added to
2022  * the group.
2023  */
2024 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
2025 {
2026 	intel_cpuc_finish(cpuc);
2027 	kfree(cpuc);
2028 }
2029 
2030 static struct cpu_hw_events *allocate_fake_cpuc(void)
2031 {
2032 	struct cpu_hw_events *cpuc;
2033 	int cpu = raw_smp_processor_id();
2034 
2035 	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
2036 	if (!cpuc)
2037 		return ERR_PTR(-ENOMEM);
2038 	cpuc->is_fake = 1;
2039 
2040 	if (intel_cpuc_prepare(cpuc, cpu))
2041 		goto error;
2042 
2043 	return cpuc;
2044 error:
2045 	free_fake_cpuc(cpuc);
2046 	return ERR_PTR(-ENOMEM);
2047 }
2048 
2049 /*
2050  * validate that we can schedule this event
2051  */
2052 static int validate_event(struct perf_event *event)
2053 {
2054 	struct cpu_hw_events *fake_cpuc;
2055 	struct event_constraint *c;
2056 	int ret = 0;
2057 
2058 	fake_cpuc = allocate_fake_cpuc();
2059 	if (IS_ERR(fake_cpuc))
2060 		return PTR_ERR(fake_cpuc);
2061 
2062 	c = x86_pmu.get_event_constraints(fake_cpuc, 0, event);
2063 
2064 	if (!c || !c->weight)
2065 		ret = -EINVAL;
2066 
2067 	if (x86_pmu.put_event_constraints)
2068 		x86_pmu.put_event_constraints(fake_cpuc, event);
2069 
2070 	free_fake_cpuc(fake_cpuc);
2071 
2072 	return ret;
2073 }
2074 
2075 /*
2076  * validate a single event group
2077  *
2078  * validation include:
2079  *	- check events are compatible which each other
2080  *	- events do not compete for the same counter
2081  *	- number of events <= number of counters
2082  *
2083  * validation ensures the group can be loaded onto the
2084  * PMU if it was the only group available.
2085  */
2086 static int validate_group(struct perf_event *event)
2087 {
2088 	struct perf_event *leader = event->group_leader;
2089 	struct cpu_hw_events *fake_cpuc;
2090 	int ret = -EINVAL, n;
2091 
2092 	fake_cpuc = allocate_fake_cpuc();
2093 	if (IS_ERR(fake_cpuc))
2094 		return PTR_ERR(fake_cpuc);
2095 	/*
2096 	 * the event is not yet connected with its
2097 	 * siblings therefore we must first collect
2098 	 * existing siblings, then add the new event
2099 	 * before we can simulate the scheduling
2100 	 */
2101 	n = collect_events(fake_cpuc, leader, true);
2102 	if (n < 0)
2103 		goto out;
2104 
2105 	fake_cpuc->n_events = n;
2106 	n = collect_events(fake_cpuc, event, false);
2107 	if (n < 0)
2108 		goto out;
2109 
2110 	fake_cpuc->n_events = 0;
2111 	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2112 
2113 out:
2114 	free_fake_cpuc(fake_cpuc);
2115 	return ret;
2116 }
2117 
2118 static int x86_pmu_event_init(struct perf_event *event)
2119 {
2120 	struct pmu *tmp;
2121 	int err;
2122 
2123 	switch (event->attr.type) {
2124 	case PERF_TYPE_RAW:
2125 	case PERF_TYPE_HARDWARE:
2126 	case PERF_TYPE_HW_CACHE:
2127 		break;
2128 
2129 	default:
2130 		return -ENOENT;
2131 	}
2132 
2133 	err = __x86_pmu_event_init(event);
2134 	if (!err) {
2135 		/*
2136 		 * we temporarily connect event to its pmu
2137 		 * such that validate_group() can classify
2138 		 * it as an x86 event using is_x86_event()
2139 		 */
2140 		tmp = event->pmu;
2141 		event->pmu = &pmu;
2142 
2143 		if (event->group_leader != event)
2144 			err = validate_group(event);
2145 		else
2146 			err = validate_event(event);
2147 
2148 		event->pmu = tmp;
2149 	}
2150 	if (err) {
2151 		if (event->destroy)
2152 			event->destroy(event);
2153 	}
2154 
2155 	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2156 	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2157 		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2158 
2159 	return err;
2160 }
2161 
2162 static void refresh_pce(void *ignored)
2163 {
2164 	load_mm_cr4(this_cpu_read(cpu_tlbstate.loaded_mm));
2165 }
2166 
2167 static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2168 {
2169 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2170 		return;
2171 
2172 	/*
2173 	 * This function relies on not being called concurrently in two
2174 	 * tasks in the same mm.  Otherwise one task could observe
2175 	 * perf_rdpmc_allowed > 1 and return all the way back to
2176 	 * userspace with CR4.PCE clear while another task is still
2177 	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2178 	 *
2179 	 * For now, this can't happen because all callers hold mmap_sem
2180 	 * for write.  If this changes, we'll need a different solution.
2181 	 */
2182 	lockdep_assert_held_exclusive(&mm->mmap_sem);
2183 
2184 	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2185 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2186 }
2187 
2188 static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2189 {
2190 
2191 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2192 		return;
2193 
2194 	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2195 		on_each_cpu_mask(mm_cpumask(mm), refresh_pce, NULL, 1);
2196 }
2197 
2198 static int x86_pmu_event_idx(struct perf_event *event)
2199 {
2200 	int idx = event->hw.idx;
2201 
2202 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2203 		return 0;
2204 
2205 	if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
2206 		idx -= INTEL_PMC_IDX_FIXED;
2207 		idx |= 1 << 30;
2208 	}
2209 
2210 	return idx + 1;
2211 }
2212 
2213 static ssize_t get_attr_rdpmc(struct device *cdev,
2214 			      struct device_attribute *attr,
2215 			      char *buf)
2216 {
2217 	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2218 }
2219 
2220 static ssize_t set_attr_rdpmc(struct device *cdev,
2221 			      struct device_attribute *attr,
2222 			      const char *buf, size_t count)
2223 {
2224 	unsigned long val;
2225 	ssize_t ret;
2226 
2227 	ret = kstrtoul(buf, 0, &val);
2228 	if (ret)
2229 		return ret;
2230 
2231 	if (val > 2)
2232 		return -EINVAL;
2233 
2234 	if (x86_pmu.attr_rdpmc_broken)
2235 		return -ENOTSUPP;
2236 
2237 	if ((val == 2) != (x86_pmu.attr_rdpmc == 2)) {
2238 		/*
2239 		 * Changing into or out of always available, aka
2240 		 * perf-event-bypassing mode.  This path is extremely slow,
2241 		 * but only root can trigger it, so it's okay.
2242 		 */
2243 		if (val == 2)
2244 			static_branch_inc(&rdpmc_always_available_key);
2245 		else
2246 			static_branch_dec(&rdpmc_always_available_key);
2247 		on_each_cpu(refresh_pce, NULL, 1);
2248 	}
2249 
2250 	x86_pmu.attr_rdpmc = val;
2251 
2252 	return count;
2253 }
2254 
2255 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2256 
2257 static struct attribute *x86_pmu_attrs[] = {
2258 	&dev_attr_rdpmc.attr,
2259 	NULL,
2260 };
2261 
2262 static struct attribute_group x86_pmu_attr_group __ro_after_init = {
2263 	.attrs = x86_pmu_attrs,
2264 };
2265 
2266 static ssize_t max_precise_show(struct device *cdev,
2267 				  struct device_attribute *attr,
2268 				  char *buf)
2269 {
2270 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2271 }
2272 
2273 static DEVICE_ATTR_RO(max_precise);
2274 
2275 static struct attribute *x86_pmu_caps_attrs[] = {
2276 	&dev_attr_max_precise.attr,
2277 	NULL
2278 };
2279 
2280 static struct attribute_group x86_pmu_caps_group __ro_after_init = {
2281 	.name = "caps",
2282 	.attrs = x86_pmu_caps_attrs,
2283 };
2284 
2285 static const struct attribute_group *x86_pmu_attr_groups[] = {
2286 	&x86_pmu_attr_group,
2287 	&x86_pmu_format_group,
2288 	&x86_pmu_events_group,
2289 	&x86_pmu_caps_group,
2290 	NULL,
2291 };
2292 
2293 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2294 {
2295 	if (x86_pmu.sched_task)
2296 		x86_pmu.sched_task(ctx, sched_in);
2297 }
2298 
2299 void perf_check_microcode(void)
2300 {
2301 	if (x86_pmu.check_microcode)
2302 		x86_pmu.check_microcode();
2303 }
2304 
2305 static int x86_pmu_check_period(struct perf_event *event, u64 value)
2306 {
2307 	if (x86_pmu.check_period && x86_pmu.check_period(event, value))
2308 		return -EINVAL;
2309 
2310 	if (value && x86_pmu.limit_period) {
2311 		if (x86_pmu.limit_period(event, value) > value)
2312 			return -EINVAL;
2313 	}
2314 
2315 	return 0;
2316 }
2317 
2318 static struct pmu pmu = {
2319 	.pmu_enable		= x86_pmu_enable,
2320 	.pmu_disable		= x86_pmu_disable,
2321 
2322 	.attr_groups		= x86_pmu_attr_groups,
2323 
2324 	.event_init		= x86_pmu_event_init,
2325 
2326 	.event_mapped		= x86_pmu_event_mapped,
2327 	.event_unmapped		= x86_pmu_event_unmapped,
2328 
2329 	.add			= x86_pmu_add,
2330 	.del			= x86_pmu_del,
2331 	.start			= x86_pmu_start,
2332 	.stop			= x86_pmu_stop,
2333 	.read			= x86_pmu_read,
2334 
2335 	.start_txn		= x86_pmu_start_txn,
2336 	.cancel_txn		= x86_pmu_cancel_txn,
2337 	.commit_txn		= x86_pmu_commit_txn,
2338 
2339 	.event_idx		= x86_pmu_event_idx,
2340 	.sched_task		= x86_pmu_sched_task,
2341 	.task_ctx_size          = sizeof(struct x86_perf_task_context),
2342 	.check_period		= x86_pmu_check_period,
2343 };
2344 
2345 void arch_perf_update_userpage(struct perf_event *event,
2346 			       struct perf_event_mmap_page *userpg, u64 now)
2347 {
2348 	struct cyc2ns_data data;
2349 	u64 offset;
2350 
2351 	userpg->cap_user_time = 0;
2352 	userpg->cap_user_time_zero = 0;
2353 	userpg->cap_user_rdpmc =
2354 		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2355 	userpg->pmc_width = x86_pmu.cntval_bits;
2356 
2357 	if (!using_native_sched_clock() || !sched_clock_stable())
2358 		return;
2359 
2360 	cyc2ns_read_begin(&data);
2361 
2362 	offset = data.cyc2ns_offset + __sched_clock_offset;
2363 
2364 	/*
2365 	 * Internal timekeeping for enabled/running/stopped times
2366 	 * is always in the local_clock domain.
2367 	 */
2368 	userpg->cap_user_time = 1;
2369 	userpg->time_mult = data.cyc2ns_mul;
2370 	userpg->time_shift = data.cyc2ns_shift;
2371 	userpg->time_offset = offset - now;
2372 
2373 	/*
2374 	 * cap_user_time_zero doesn't make sense when we're using a different
2375 	 * time base for the records.
2376 	 */
2377 	if (!event->attr.use_clockid) {
2378 		userpg->cap_user_time_zero = 1;
2379 		userpg->time_zero = offset;
2380 	}
2381 
2382 	cyc2ns_read_end();
2383 }
2384 
2385 /*
2386  * Determine whether the regs were taken from an irq/exception handler rather
2387  * than from perf_arch_fetch_caller_regs().
2388  */
2389 static bool perf_hw_regs(struct pt_regs *regs)
2390 {
2391 	return regs->flags & X86_EFLAGS_FIXED;
2392 }
2393 
2394 void
2395 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2396 {
2397 	struct unwind_state state;
2398 	unsigned long addr;
2399 
2400 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2401 		/* TODO: We don't support guest os callchain now */
2402 		return;
2403 	}
2404 
2405 	if (perf_hw_regs(regs)) {
2406 		if (perf_callchain_store(entry, regs->ip))
2407 			return;
2408 		unwind_start(&state, current, regs, NULL);
2409 	} else {
2410 		unwind_start(&state, current, NULL, (void *)regs->sp);
2411 	}
2412 
2413 	for (; !unwind_done(&state); unwind_next_frame(&state)) {
2414 		addr = unwind_get_return_address(&state);
2415 		if (!addr || perf_callchain_store(entry, addr))
2416 			return;
2417 	}
2418 }
2419 
2420 static inline int
2421 valid_user_frame(const void __user *fp, unsigned long size)
2422 {
2423 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2424 }
2425 
2426 static unsigned long get_segment_base(unsigned int segment)
2427 {
2428 	struct desc_struct *desc;
2429 	unsigned int idx = segment >> 3;
2430 
2431 	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2432 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2433 		struct ldt_struct *ldt;
2434 
2435 		/* IRQs are off, so this synchronizes with smp_store_release */
2436 		ldt = READ_ONCE(current->active_mm->context.ldt);
2437 		if (!ldt || idx >= ldt->nr_entries)
2438 			return 0;
2439 
2440 		desc = &ldt->entries[idx];
2441 #else
2442 		return 0;
2443 #endif
2444 	} else {
2445 		if (idx >= GDT_ENTRIES)
2446 			return 0;
2447 
2448 		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2449 	}
2450 
2451 	return get_desc_base(desc);
2452 }
2453 
2454 #ifdef CONFIG_IA32_EMULATION
2455 
2456 #include <linux/compat.h>
2457 
2458 static inline int
2459 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2460 {
2461 	/* 32-bit process in 64-bit kernel. */
2462 	unsigned long ss_base, cs_base;
2463 	struct stack_frame_ia32 frame;
2464 	const void __user *fp;
2465 
2466 	if (!test_thread_flag(TIF_IA32))
2467 		return 0;
2468 
2469 	cs_base = get_segment_base(regs->cs);
2470 	ss_base = get_segment_base(regs->ss);
2471 
2472 	fp = compat_ptr(ss_base + regs->bp);
2473 	pagefault_disable();
2474 	while (entry->nr < entry->max_stack) {
2475 		unsigned long bytes;
2476 		frame.next_frame     = 0;
2477 		frame.return_address = 0;
2478 
2479 		if (!valid_user_frame(fp, sizeof(frame)))
2480 			break;
2481 
2482 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, 4);
2483 		if (bytes != 0)
2484 			break;
2485 		bytes = __copy_from_user_nmi(&frame.return_address, fp+4, 4);
2486 		if (bytes != 0)
2487 			break;
2488 
2489 		perf_callchain_store(entry, cs_base + frame.return_address);
2490 		fp = compat_ptr(ss_base + frame.next_frame);
2491 	}
2492 	pagefault_enable();
2493 	return 1;
2494 }
2495 #else
2496 static inline int
2497 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2498 {
2499     return 0;
2500 }
2501 #endif
2502 
2503 void
2504 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2505 {
2506 	struct stack_frame frame;
2507 	const unsigned long __user *fp;
2508 
2509 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2510 		/* TODO: We don't support guest os callchain now */
2511 		return;
2512 	}
2513 
2514 	/*
2515 	 * We don't know what to do with VM86 stacks.. ignore them for now.
2516 	 */
2517 	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2518 		return;
2519 
2520 	fp = (unsigned long __user *)regs->bp;
2521 
2522 	perf_callchain_store(entry, regs->ip);
2523 
2524 	if (!nmi_uaccess_okay())
2525 		return;
2526 
2527 	if (perf_callchain_user32(regs, entry))
2528 		return;
2529 
2530 	pagefault_disable();
2531 	while (entry->nr < entry->max_stack) {
2532 		unsigned long bytes;
2533 
2534 		frame.next_frame	     = NULL;
2535 		frame.return_address = 0;
2536 
2537 		if (!valid_user_frame(fp, sizeof(frame)))
2538 			break;
2539 
2540 		bytes = __copy_from_user_nmi(&frame.next_frame, fp, sizeof(*fp));
2541 		if (bytes != 0)
2542 			break;
2543 		bytes = __copy_from_user_nmi(&frame.return_address, fp + 1, sizeof(*fp));
2544 		if (bytes != 0)
2545 			break;
2546 
2547 		perf_callchain_store(entry, frame.return_address);
2548 		fp = (void __user *)frame.next_frame;
2549 	}
2550 	pagefault_enable();
2551 }
2552 
2553 /*
2554  * Deal with code segment offsets for the various execution modes:
2555  *
2556  *   VM86 - the good olde 16 bit days, where the linear address is
2557  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2558  *
2559  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2560  *          to figure out what the 32bit base address is.
2561  *
2562  *    X32 - has TIF_X32 set, but is running in x86_64
2563  *
2564  * X86_64 - CS,DS,SS,ES are all zero based.
2565  */
2566 static unsigned long code_segment_base(struct pt_regs *regs)
2567 {
2568 	/*
2569 	 * For IA32 we look at the GDT/LDT segment base to convert the
2570 	 * effective IP to a linear address.
2571 	 */
2572 
2573 #ifdef CONFIG_X86_32
2574 	/*
2575 	 * If we are in VM86 mode, add the segment offset to convert to a
2576 	 * linear address.
2577 	 */
2578 	if (regs->flags & X86_VM_MASK)
2579 		return 0x10 * regs->cs;
2580 
2581 	if (user_mode(regs) && regs->cs != __USER_CS)
2582 		return get_segment_base(regs->cs);
2583 #else
2584 	if (user_mode(regs) && !user_64bit_mode(regs) &&
2585 	    regs->cs != __USER32_CS)
2586 		return get_segment_base(regs->cs);
2587 #endif
2588 	return 0;
2589 }
2590 
2591 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2592 {
2593 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2594 		return perf_guest_cbs->get_guest_ip();
2595 
2596 	return regs->ip + code_segment_base(regs);
2597 }
2598 
2599 unsigned long perf_misc_flags(struct pt_regs *regs)
2600 {
2601 	int misc = 0;
2602 
2603 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2604 		if (perf_guest_cbs->is_user_mode())
2605 			misc |= PERF_RECORD_MISC_GUEST_USER;
2606 		else
2607 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2608 	} else {
2609 		if (user_mode(regs))
2610 			misc |= PERF_RECORD_MISC_USER;
2611 		else
2612 			misc |= PERF_RECORD_MISC_KERNEL;
2613 	}
2614 
2615 	if (regs->flags & PERF_EFLAGS_EXACT)
2616 		misc |= PERF_RECORD_MISC_EXACT_IP;
2617 
2618 	return misc;
2619 }
2620 
2621 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2622 {
2623 	cap->version		= x86_pmu.version;
2624 	cap->num_counters_gp	= x86_pmu.num_counters;
2625 	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2626 	cap->bit_width_gp	= x86_pmu.cntval_bits;
2627 	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2628 	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2629 	cap->events_mask_len	= x86_pmu.events_mask_len;
2630 }
2631 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
2632