xref: /linux/arch/x86/events/core.c (revision 0d3b051adbb72ed81956447d0d1e54d5943ee6f5)
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 #include <linux/static_call.h>
32 
33 #include <asm/apic.h>
34 #include <asm/stacktrace.h>
35 #include <asm/nmi.h>
36 #include <asm/smp.h>
37 #include <asm/alternative.h>
38 #include <asm/mmu_context.h>
39 #include <asm/tlbflush.h>
40 #include <asm/timer.h>
41 #include <asm/desc.h>
42 #include <asm/ldt.h>
43 #include <asm/unwind.h>
44 
45 #include "perf_event.h"
46 
47 struct x86_pmu x86_pmu __read_mostly;
48 
49 DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
50 	.enabled = 1,
51 };
52 
53 DEFINE_STATIC_KEY_FALSE(rdpmc_never_available_key);
54 DEFINE_STATIC_KEY_FALSE(rdpmc_always_available_key);
55 
56 /*
57  * This here uses DEFINE_STATIC_CALL_NULL() to get a static_call defined
58  * from just a typename, as opposed to an actual function.
59  */
60 DEFINE_STATIC_CALL_NULL(x86_pmu_handle_irq,  *x86_pmu.handle_irq);
61 DEFINE_STATIC_CALL_NULL(x86_pmu_disable_all, *x86_pmu.disable_all);
62 DEFINE_STATIC_CALL_NULL(x86_pmu_enable_all,  *x86_pmu.enable_all);
63 DEFINE_STATIC_CALL_NULL(x86_pmu_enable,	     *x86_pmu.enable);
64 DEFINE_STATIC_CALL_NULL(x86_pmu_disable,     *x86_pmu.disable);
65 
66 DEFINE_STATIC_CALL_NULL(x86_pmu_add,  *x86_pmu.add);
67 DEFINE_STATIC_CALL_NULL(x86_pmu_del,  *x86_pmu.del);
68 DEFINE_STATIC_CALL_NULL(x86_pmu_read, *x86_pmu.read);
69 
70 DEFINE_STATIC_CALL_NULL(x86_pmu_schedule_events,       *x86_pmu.schedule_events);
71 DEFINE_STATIC_CALL_NULL(x86_pmu_get_event_constraints, *x86_pmu.get_event_constraints);
72 DEFINE_STATIC_CALL_NULL(x86_pmu_put_event_constraints, *x86_pmu.put_event_constraints);
73 
74 DEFINE_STATIC_CALL_NULL(x86_pmu_start_scheduling,  *x86_pmu.start_scheduling);
75 DEFINE_STATIC_CALL_NULL(x86_pmu_commit_scheduling, *x86_pmu.commit_scheduling);
76 DEFINE_STATIC_CALL_NULL(x86_pmu_stop_scheduling,   *x86_pmu.stop_scheduling);
77 
78 DEFINE_STATIC_CALL_NULL(x86_pmu_sched_task,    *x86_pmu.sched_task);
79 DEFINE_STATIC_CALL_NULL(x86_pmu_swap_task_ctx, *x86_pmu.swap_task_ctx);
80 
81 DEFINE_STATIC_CALL_NULL(x86_pmu_drain_pebs,   *x86_pmu.drain_pebs);
82 DEFINE_STATIC_CALL_NULL(x86_pmu_pebs_aliases, *x86_pmu.pebs_aliases);
83 
84 u64 __read_mostly hw_cache_event_ids
85 				[PERF_COUNT_HW_CACHE_MAX]
86 				[PERF_COUNT_HW_CACHE_OP_MAX]
87 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
88 u64 __read_mostly hw_cache_extra_regs
89 				[PERF_COUNT_HW_CACHE_MAX]
90 				[PERF_COUNT_HW_CACHE_OP_MAX]
91 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
92 
93 /*
94  * Propagate event elapsed time into the generic event.
95  * Can only be executed on the CPU where the event is active.
96  * Returns the delta events processed.
97  */
98 u64 x86_perf_event_update(struct perf_event *event)
99 {
100 	struct hw_perf_event *hwc = &event->hw;
101 	int shift = 64 - x86_pmu.cntval_bits;
102 	u64 prev_raw_count, new_raw_count;
103 	u64 delta;
104 
105 	if (unlikely(!hwc->event_base))
106 		return 0;
107 
108 	if (unlikely(is_topdown_count(event)) && x86_pmu.update_topdown_event)
109 		return x86_pmu.update_topdown_event(event);
110 
111 	/*
112 	 * Careful: an NMI might modify the previous event value.
113 	 *
114 	 * Our tactic to handle this is to first atomically read and
115 	 * exchange a new raw count - then add that new-prev delta
116 	 * count to the generic event atomically:
117 	 */
118 again:
119 	prev_raw_count = local64_read(&hwc->prev_count);
120 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
121 
122 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
123 					new_raw_count) != prev_raw_count)
124 		goto again;
125 
126 	/*
127 	 * Now we have the new raw value and have updated the prev
128 	 * timestamp already. We can now calculate the elapsed delta
129 	 * (event-)time and add that to the generic event.
130 	 *
131 	 * Careful, not all hw sign-extends above the physical width
132 	 * of the count.
133 	 */
134 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
135 	delta >>= shift;
136 
137 	local64_add(delta, &event->count);
138 	local64_sub(delta, &hwc->period_left);
139 
140 	return new_raw_count;
141 }
142 
143 /*
144  * Find and validate any extra registers to set up.
145  */
146 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
147 {
148 	struct hw_perf_event_extra *reg;
149 	struct extra_reg *er;
150 
151 	reg = &event->hw.extra_reg;
152 
153 	if (!x86_pmu.extra_regs)
154 		return 0;
155 
156 	for (er = x86_pmu.extra_regs; er->msr; er++) {
157 		if (er->event != (config & er->config_mask))
158 			continue;
159 		if (event->attr.config1 & ~er->valid_mask)
160 			return -EINVAL;
161 		/* Check if the extra msrs can be safely accessed*/
162 		if (!er->extra_msr_access)
163 			return -ENXIO;
164 
165 		reg->idx = er->idx;
166 		reg->config = event->attr.config1;
167 		reg->reg = er->msr;
168 		break;
169 	}
170 	return 0;
171 }
172 
173 static atomic_t active_events;
174 static atomic_t pmc_refcount;
175 static DEFINE_MUTEX(pmc_reserve_mutex);
176 
177 #ifdef CONFIG_X86_LOCAL_APIC
178 
179 static bool reserve_pmc_hardware(void)
180 {
181 	int i;
182 
183 	for (i = 0; i < x86_pmu.num_counters; i++) {
184 		if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
185 			goto perfctr_fail;
186 	}
187 
188 	for (i = 0; i < x86_pmu.num_counters; i++) {
189 		if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
190 			goto eventsel_fail;
191 	}
192 
193 	return true;
194 
195 eventsel_fail:
196 	for (i--; i >= 0; i--)
197 		release_evntsel_nmi(x86_pmu_config_addr(i));
198 
199 	i = x86_pmu.num_counters;
200 
201 perfctr_fail:
202 	for (i--; i >= 0; i--)
203 		release_perfctr_nmi(x86_pmu_event_addr(i));
204 
205 	return false;
206 }
207 
208 static void release_pmc_hardware(void)
209 {
210 	int i;
211 
212 	for (i = 0; i < x86_pmu.num_counters; i++) {
213 		release_perfctr_nmi(x86_pmu_event_addr(i));
214 		release_evntsel_nmi(x86_pmu_config_addr(i));
215 	}
216 }
217 
218 #else
219 
220 static bool reserve_pmc_hardware(void) { return true; }
221 static void release_pmc_hardware(void) {}
222 
223 #endif
224 
225 static bool check_hw_exists(void)
226 {
227 	u64 val, val_fail = -1, val_new= ~0;
228 	int i, reg, reg_fail = -1, ret = 0;
229 	int bios_fail = 0;
230 	int reg_safe = -1;
231 
232 	/*
233 	 * Check to see if the BIOS enabled any of the counters, if so
234 	 * complain and bail.
235 	 */
236 	for (i = 0; i < x86_pmu.num_counters; i++) {
237 		reg = x86_pmu_config_addr(i);
238 		ret = rdmsrl_safe(reg, &val);
239 		if (ret)
240 			goto msr_fail;
241 		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
242 			bios_fail = 1;
243 			val_fail = val;
244 			reg_fail = reg;
245 		} else {
246 			reg_safe = i;
247 		}
248 	}
249 
250 	if (x86_pmu.num_counters_fixed) {
251 		reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
252 		ret = rdmsrl_safe(reg, &val);
253 		if (ret)
254 			goto msr_fail;
255 		for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
256 			if (val & (0x03 << i*4)) {
257 				bios_fail = 1;
258 				val_fail = val;
259 				reg_fail = reg;
260 			}
261 		}
262 	}
263 
264 	/*
265 	 * If all the counters are enabled, the below test will always
266 	 * fail.  The tools will also become useless in this scenario.
267 	 * Just fail and disable the hardware counters.
268 	 */
269 
270 	if (reg_safe == -1) {
271 		reg = reg_safe;
272 		goto msr_fail;
273 	}
274 
275 	/*
276 	 * Read the current value, change it and read it back to see if it
277 	 * matches, this is needed to detect certain hardware emulators
278 	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
279 	 */
280 	reg = x86_pmu_event_addr(reg_safe);
281 	if (rdmsrl_safe(reg, &val))
282 		goto msr_fail;
283 	val ^= 0xffffUL;
284 	ret = wrmsrl_safe(reg, val);
285 	ret |= rdmsrl_safe(reg, &val_new);
286 	if (ret || val != val_new)
287 		goto msr_fail;
288 
289 	/*
290 	 * We still allow the PMU driver to operate:
291 	 */
292 	if (bios_fail) {
293 		pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
294 		pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
295 			      reg_fail, val_fail);
296 	}
297 
298 	return true;
299 
300 msr_fail:
301 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
302 		pr_cont("PMU not available due to virtualization, using software events only.\n");
303 	} else {
304 		pr_cont("Broken PMU hardware detected, using software events only.\n");
305 		pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
306 		       reg, val_new);
307 	}
308 
309 	return false;
310 }
311 
312 static void hw_perf_event_destroy(struct perf_event *event)
313 {
314 	x86_release_hardware();
315 	atomic_dec(&active_events);
316 }
317 
318 void hw_perf_lbr_event_destroy(struct perf_event *event)
319 {
320 	hw_perf_event_destroy(event);
321 
322 	/* undo the lbr/bts event accounting */
323 	x86_del_exclusive(x86_lbr_exclusive_lbr);
324 }
325 
326 static inline int x86_pmu_initialized(void)
327 {
328 	return x86_pmu.handle_irq != NULL;
329 }
330 
331 static inline int
332 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
333 {
334 	struct perf_event_attr *attr = &event->attr;
335 	unsigned int cache_type, cache_op, cache_result;
336 	u64 config, val;
337 
338 	config = attr->config;
339 
340 	cache_type = (config >> 0) & 0xff;
341 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
342 		return -EINVAL;
343 	cache_type = array_index_nospec(cache_type, PERF_COUNT_HW_CACHE_MAX);
344 
345 	cache_op = (config >>  8) & 0xff;
346 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
347 		return -EINVAL;
348 	cache_op = array_index_nospec(cache_op, PERF_COUNT_HW_CACHE_OP_MAX);
349 
350 	cache_result = (config >> 16) & 0xff;
351 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
352 		return -EINVAL;
353 	cache_result = array_index_nospec(cache_result, PERF_COUNT_HW_CACHE_RESULT_MAX);
354 
355 	val = hw_cache_event_ids[cache_type][cache_op][cache_result];
356 
357 	if (val == 0)
358 		return -ENOENT;
359 
360 	if (val == -1)
361 		return -EINVAL;
362 
363 	hwc->config |= val;
364 	attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
365 	return x86_pmu_extra_regs(val, event);
366 }
367 
368 int x86_reserve_hardware(void)
369 {
370 	int err = 0;
371 
372 	if (!atomic_inc_not_zero(&pmc_refcount)) {
373 		mutex_lock(&pmc_reserve_mutex);
374 		if (atomic_read(&pmc_refcount) == 0) {
375 			if (!reserve_pmc_hardware())
376 				err = -EBUSY;
377 			else
378 				reserve_ds_buffers();
379 		}
380 		if (!err)
381 			atomic_inc(&pmc_refcount);
382 		mutex_unlock(&pmc_reserve_mutex);
383 	}
384 
385 	return err;
386 }
387 
388 void x86_release_hardware(void)
389 {
390 	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
391 		release_pmc_hardware();
392 		release_ds_buffers();
393 		release_lbr_buffers();
394 		mutex_unlock(&pmc_reserve_mutex);
395 	}
396 }
397 
398 /*
399  * Check if we can create event of a certain type (that no conflicting events
400  * are present).
401  */
402 int x86_add_exclusive(unsigned int what)
403 {
404 	int i;
405 
406 	/*
407 	 * When lbr_pt_coexist we allow PT to coexist with either LBR or BTS.
408 	 * LBR and BTS are still mutually exclusive.
409 	 */
410 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
411 		goto out;
412 
413 	if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
414 		mutex_lock(&pmc_reserve_mutex);
415 		for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
416 			if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
417 				goto fail_unlock;
418 		}
419 		atomic_inc(&x86_pmu.lbr_exclusive[what]);
420 		mutex_unlock(&pmc_reserve_mutex);
421 	}
422 
423 out:
424 	atomic_inc(&active_events);
425 	return 0;
426 
427 fail_unlock:
428 	mutex_unlock(&pmc_reserve_mutex);
429 	return -EBUSY;
430 }
431 
432 void x86_del_exclusive(unsigned int what)
433 {
434 	atomic_dec(&active_events);
435 
436 	/*
437 	 * See the comment in x86_add_exclusive().
438 	 */
439 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
440 		return;
441 
442 	atomic_dec(&x86_pmu.lbr_exclusive[what]);
443 }
444 
445 int x86_setup_perfctr(struct perf_event *event)
446 {
447 	struct perf_event_attr *attr = &event->attr;
448 	struct hw_perf_event *hwc = &event->hw;
449 	u64 config;
450 
451 	if (!is_sampling_event(event)) {
452 		hwc->sample_period = x86_pmu.max_period;
453 		hwc->last_period = hwc->sample_period;
454 		local64_set(&hwc->period_left, hwc->sample_period);
455 	}
456 
457 	if (attr->type == PERF_TYPE_RAW)
458 		return x86_pmu_extra_regs(event->attr.config, event);
459 
460 	if (attr->type == PERF_TYPE_HW_CACHE)
461 		return set_ext_hw_attr(hwc, event);
462 
463 	if (attr->config >= x86_pmu.max_events)
464 		return -EINVAL;
465 
466 	attr->config = array_index_nospec((unsigned long)attr->config, x86_pmu.max_events);
467 
468 	/*
469 	 * The generic map:
470 	 */
471 	config = x86_pmu.event_map(attr->config);
472 
473 	if (config == 0)
474 		return -ENOENT;
475 
476 	if (config == -1LL)
477 		return -EINVAL;
478 
479 	hwc->config |= config;
480 
481 	return 0;
482 }
483 
484 /*
485  * check that branch_sample_type is compatible with
486  * settings needed for precise_ip > 1 which implies
487  * using the LBR to capture ALL taken branches at the
488  * priv levels of the measurement
489  */
490 static inline int precise_br_compat(struct perf_event *event)
491 {
492 	u64 m = event->attr.branch_sample_type;
493 	u64 b = 0;
494 
495 	/* must capture all branches */
496 	if (!(m & PERF_SAMPLE_BRANCH_ANY))
497 		return 0;
498 
499 	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
500 
501 	if (!event->attr.exclude_user)
502 		b |= PERF_SAMPLE_BRANCH_USER;
503 
504 	if (!event->attr.exclude_kernel)
505 		b |= PERF_SAMPLE_BRANCH_KERNEL;
506 
507 	/*
508 	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
509 	 */
510 
511 	return m == b;
512 }
513 
514 int x86_pmu_max_precise(void)
515 {
516 	int precise = 0;
517 
518 	/* Support for constant skid */
519 	if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
520 		precise++;
521 
522 		/* Support for IP fixup */
523 		if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
524 			precise++;
525 
526 		if (x86_pmu.pebs_prec_dist)
527 			precise++;
528 	}
529 	return precise;
530 }
531 
532 int x86_pmu_hw_config(struct perf_event *event)
533 {
534 	if (event->attr.precise_ip) {
535 		int precise = x86_pmu_max_precise();
536 
537 		if (event->attr.precise_ip > precise)
538 			return -EOPNOTSUPP;
539 
540 		/* There's no sense in having PEBS for non sampling events: */
541 		if (!is_sampling_event(event))
542 			return -EINVAL;
543 	}
544 	/*
545 	 * check that PEBS LBR correction does not conflict with
546 	 * whatever the user is asking with attr->branch_sample_type
547 	 */
548 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
549 		u64 *br_type = &event->attr.branch_sample_type;
550 
551 		if (has_branch_stack(event)) {
552 			if (!precise_br_compat(event))
553 				return -EOPNOTSUPP;
554 
555 			/* branch_sample_type is compatible */
556 
557 		} else {
558 			/*
559 			 * user did not specify  branch_sample_type
560 			 *
561 			 * For PEBS fixups, we capture all
562 			 * the branches at the priv level of the
563 			 * event.
564 			 */
565 			*br_type = PERF_SAMPLE_BRANCH_ANY;
566 
567 			if (!event->attr.exclude_user)
568 				*br_type |= PERF_SAMPLE_BRANCH_USER;
569 
570 			if (!event->attr.exclude_kernel)
571 				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
572 		}
573 	}
574 
575 	if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
576 		event->attach_state |= PERF_ATTACH_TASK_DATA;
577 
578 	/*
579 	 * Generate PMC IRQs:
580 	 * (keep 'enabled' bit clear for now)
581 	 */
582 	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
583 
584 	/*
585 	 * Count user and OS events unless requested not to
586 	 */
587 	if (!event->attr.exclude_user)
588 		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
589 	if (!event->attr.exclude_kernel)
590 		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
591 
592 	if (event->attr.type == PERF_TYPE_RAW)
593 		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
594 
595 	if (event->attr.sample_period && x86_pmu.limit_period) {
596 		if (x86_pmu.limit_period(event, event->attr.sample_period) >
597 				event->attr.sample_period)
598 			return -EINVAL;
599 	}
600 
601 	/* sample_regs_user never support XMM registers */
602 	if (unlikely(event->attr.sample_regs_user & PERF_REG_EXTENDED_MASK))
603 		return -EINVAL;
604 	/*
605 	 * Besides the general purpose registers, XMM registers may
606 	 * be collected in PEBS on some platforms, e.g. Icelake
607 	 */
608 	if (unlikely(event->attr.sample_regs_intr & PERF_REG_EXTENDED_MASK)) {
609 		if (!(event->pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS))
610 			return -EINVAL;
611 
612 		if (!event->attr.precise_ip)
613 			return -EINVAL;
614 	}
615 
616 	return x86_setup_perfctr(event);
617 }
618 
619 /*
620  * Setup the hardware configuration for a given attr_type
621  */
622 static int __x86_pmu_event_init(struct perf_event *event)
623 {
624 	int err;
625 
626 	if (!x86_pmu_initialized())
627 		return -ENODEV;
628 
629 	err = x86_reserve_hardware();
630 	if (err)
631 		return err;
632 
633 	atomic_inc(&active_events);
634 	event->destroy = hw_perf_event_destroy;
635 
636 	event->hw.idx = -1;
637 	event->hw.last_cpu = -1;
638 	event->hw.last_tag = ~0ULL;
639 
640 	/* mark unused */
641 	event->hw.extra_reg.idx = EXTRA_REG_NONE;
642 	event->hw.branch_reg.idx = EXTRA_REG_NONE;
643 
644 	return x86_pmu.hw_config(event);
645 }
646 
647 void x86_pmu_disable_all(void)
648 {
649 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
650 	int idx;
651 
652 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
653 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
654 		u64 val;
655 
656 		if (!test_bit(idx, cpuc->active_mask))
657 			continue;
658 		rdmsrl(x86_pmu_config_addr(idx), val);
659 		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
660 			continue;
661 		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
662 		wrmsrl(x86_pmu_config_addr(idx), val);
663 		if (is_counter_pair(hwc))
664 			wrmsrl(x86_pmu_config_addr(idx + 1), 0);
665 	}
666 }
667 
668 /*
669  * There may be PMI landing after enabled=0. The PMI hitting could be before or
670  * after disable_all.
671  *
672  * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
673  * It will not be re-enabled in the NMI handler again, because enabled=0. After
674  * handling the NMI, disable_all will be called, which will not change the
675  * state either. If PMI hits after disable_all, the PMU is already disabled
676  * before entering NMI handler. The NMI handler will not change the state
677  * either.
678  *
679  * So either situation is harmless.
680  */
681 static void x86_pmu_disable(struct pmu *pmu)
682 {
683 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
684 
685 	if (!x86_pmu_initialized())
686 		return;
687 
688 	if (!cpuc->enabled)
689 		return;
690 
691 	cpuc->n_added = 0;
692 	cpuc->enabled = 0;
693 	barrier();
694 
695 	static_call(x86_pmu_disable_all)();
696 }
697 
698 void x86_pmu_enable_all(int added)
699 {
700 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
701 	int idx;
702 
703 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
704 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
705 
706 		if (!test_bit(idx, cpuc->active_mask))
707 			continue;
708 
709 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
710 	}
711 }
712 
713 static struct pmu pmu;
714 
715 static inline int is_x86_event(struct perf_event *event)
716 {
717 	return event->pmu == &pmu;
718 }
719 
720 struct pmu *x86_get_pmu(void)
721 {
722 	return &pmu;
723 }
724 /*
725  * Event scheduler state:
726  *
727  * Assign events iterating over all events and counters, beginning
728  * with events with least weights first. Keep the current iterator
729  * state in struct sched_state.
730  */
731 struct sched_state {
732 	int	weight;
733 	int	event;		/* event index */
734 	int	counter;	/* counter index */
735 	int	unassigned;	/* number of events to be assigned left */
736 	int	nr_gp;		/* number of GP counters used */
737 	u64	used;
738 };
739 
740 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
741 #define	SCHED_STATES_MAX	2
742 
743 struct perf_sched {
744 	int			max_weight;
745 	int			max_events;
746 	int			max_gp;
747 	int			saved_states;
748 	struct event_constraint	**constraints;
749 	struct sched_state	state;
750 	struct sched_state	saved[SCHED_STATES_MAX];
751 };
752 
753 /*
754  * Initialize interator that runs through all events and counters.
755  */
756 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
757 			    int num, int wmin, int wmax, int gpmax)
758 {
759 	int idx;
760 
761 	memset(sched, 0, sizeof(*sched));
762 	sched->max_events	= num;
763 	sched->max_weight	= wmax;
764 	sched->max_gp		= gpmax;
765 	sched->constraints	= constraints;
766 
767 	for (idx = 0; idx < num; idx++) {
768 		if (constraints[idx]->weight == wmin)
769 			break;
770 	}
771 
772 	sched->state.event	= idx;		/* start with min weight */
773 	sched->state.weight	= wmin;
774 	sched->state.unassigned	= num;
775 }
776 
777 static void perf_sched_save_state(struct perf_sched *sched)
778 {
779 	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
780 		return;
781 
782 	sched->saved[sched->saved_states] = sched->state;
783 	sched->saved_states++;
784 }
785 
786 static bool perf_sched_restore_state(struct perf_sched *sched)
787 {
788 	if (!sched->saved_states)
789 		return false;
790 
791 	sched->saved_states--;
792 	sched->state = sched->saved[sched->saved_states];
793 
794 	/* this assignment didn't work out */
795 	/* XXX broken vs EVENT_PAIR */
796 	sched->state.used &= ~BIT_ULL(sched->state.counter);
797 
798 	/* try the next one */
799 	sched->state.counter++;
800 
801 	return true;
802 }
803 
804 /*
805  * Select a counter for the current event to schedule. Return true on
806  * success.
807  */
808 static bool __perf_sched_find_counter(struct perf_sched *sched)
809 {
810 	struct event_constraint *c;
811 	int idx;
812 
813 	if (!sched->state.unassigned)
814 		return false;
815 
816 	if (sched->state.event >= sched->max_events)
817 		return false;
818 
819 	c = sched->constraints[sched->state.event];
820 	/* Prefer fixed purpose counters */
821 	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
822 		idx = INTEL_PMC_IDX_FIXED;
823 		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
824 			u64 mask = BIT_ULL(idx);
825 
826 			if (sched->state.used & mask)
827 				continue;
828 
829 			sched->state.used |= mask;
830 			goto done;
831 		}
832 	}
833 
834 	/* Grab the first unused counter starting with idx */
835 	idx = sched->state.counter;
836 	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
837 		u64 mask = BIT_ULL(idx);
838 
839 		if (c->flags & PERF_X86_EVENT_PAIR)
840 			mask |= mask << 1;
841 
842 		if (sched->state.used & mask)
843 			continue;
844 
845 		if (sched->state.nr_gp++ >= sched->max_gp)
846 			return false;
847 
848 		sched->state.used |= mask;
849 		goto done;
850 	}
851 
852 	return false;
853 
854 done:
855 	sched->state.counter = idx;
856 
857 	if (c->overlap)
858 		perf_sched_save_state(sched);
859 
860 	return true;
861 }
862 
863 static bool perf_sched_find_counter(struct perf_sched *sched)
864 {
865 	while (!__perf_sched_find_counter(sched)) {
866 		if (!perf_sched_restore_state(sched))
867 			return false;
868 	}
869 
870 	return true;
871 }
872 
873 /*
874  * Go through all unassigned events and find the next one to schedule.
875  * Take events with the least weight first. Return true on success.
876  */
877 static bool perf_sched_next_event(struct perf_sched *sched)
878 {
879 	struct event_constraint *c;
880 
881 	if (!sched->state.unassigned || !--sched->state.unassigned)
882 		return false;
883 
884 	do {
885 		/* next event */
886 		sched->state.event++;
887 		if (sched->state.event >= sched->max_events) {
888 			/* next weight */
889 			sched->state.event = 0;
890 			sched->state.weight++;
891 			if (sched->state.weight > sched->max_weight)
892 				return false;
893 		}
894 		c = sched->constraints[sched->state.event];
895 	} while (c->weight != sched->state.weight);
896 
897 	sched->state.counter = 0;	/* start with first counter */
898 
899 	return true;
900 }
901 
902 /*
903  * Assign a counter for each event.
904  */
905 int perf_assign_events(struct event_constraint **constraints, int n,
906 			int wmin, int wmax, int gpmax, int *assign)
907 {
908 	struct perf_sched sched;
909 
910 	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
911 
912 	do {
913 		if (!perf_sched_find_counter(&sched))
914 			break;	/* failed */
915 		if (assign)
916 			assign[sched.state.event] = sched.state.counter;
917 	} while (perf_sched_next_event(&sched));
918 
919 	return sched.state.unassigned;
920 }
921 EXPORT_SYMBOL_GPL(perf_assign_events);
922 
923 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
924 {
925 	struct event_constraint *c;
926 	struct perf_event *e;
927 	int n0, i, wmin, wmax, unsched = 0;
928 	struct hw_perf_event *hwc;
929 	u64 used_mask = 0;
930 
931 	/*
932 	 * Compute the number of events already present; see x86_pmu_add(),
933 	 * validate_group() and x86_pmu_commit_txn(). For the former two
934 	 * cpuc->n_events hasn't been updated yet, while for the latter
935 	 * cpuc->n_txn contains the number of events added in the current
936 	 * transaction.
937 	 */
938 	n0 = cpuc->n_events;
939 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
940 		n0 -= cpuc->n_txn;
941 
942 	static_call_cond(x86_pmu_start_scheduling)(cpuc);
943 
944 	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
945 		c = cpuc->event_constraint[i];
946 
947 		/*
948 		 * Previously scheduled events should have a cached constraint,
949 		 * while new events should not have one.
950 		 */
951 		WARN_ON_ONCE((c && i >= n0) || (!c && i < n0));
952 
953 		/*
954 		 * Request constraints for new events; or for those events that
955 		 * have a dynamic constraint -- for those the constraint can
956 		 * change due to external factors (sibling state, allow_tfa).
957 		 */
958 		if (!c || (c->flags & PERF_X86_EVENT_DYNAMIC)) {
959 			c = static_call(x86_pmu_get_event_constraints)(cpuc, i, cpuc->event_list[i]);
960 			cpuc->event_constraint[i] = c;
961 		}
962 
963 		wmin = min(wmin, c->weight);
964 		wmax = max(wmax, c->weight);
965 	}
966 
967 	/*
968 	 * fastpath, try to reuse previous register
969 	 */
970 	for (i = 0; i < n; i++) {
971 		u64 mask;
972 
973 		hwc = &cpuc->event_list[i]->hw;
974 		c = cpuc->event_constraint[i];
975 
976 		/* never assigned */
977 		if (hwc->idx == -1)
978 			break;
979 
980 		/* constraint still honored */
981 		if (!test_bit(hwc->idx, c->idxmsk))
982 			break;
983 
984 		mask = BIT_ULL(hwc->idx);
985 		if (is_counter_pair(hwc))
986 			mask |= mask << 1;
987 
988 		/* not already used */
989 		if (used_mask & mask)
990 			break;
991 
992 		used_mask |= mask;
993 
994 		if (assign)
995 			assign[i] = hwc->idx;
996 	}
997 
998 	/* slow path */
999 	if (i != n) {
1000 		int gpmax = x86_pmu.num_counters;
1001 
1002 		/*
1003 		 * Do not allow scheduling of more than half the available
1004 		 * generic counters.
1005 		 *
1006 		 * This helps avoid counter starvation of sibling thread by
1007 		 * ensuring at most half the counters cannot be in exclusive
1008 		 * mode. There is no designated counters for the limits. Any
1009 		 * N/2 counters can be used. This helps with events with
1010 		 * specific counter constraints.
1011 		 */
1012 		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
1013 		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
1014 			gpmax /= 2;
1015 
1016 		/*
1017 		 * Reduce the amount of available counters to allow fitting
1018 		 * the extra Merge events needed by large increment events.
1019 		 */
1020 		if (x86_pmu.flags & PMU_FL_PAIR) {
1021 			gpmax = x86_pmu.num_counters - cpuc->n_pair;
1022 			WARN_ON(gpmax <= 0);
1023 		}
1024 
1025 		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
1026 					     wmax, gpmax, assign);
1027 	}
1028 
1029 	/*
1030 	 * In case of success (unsched = 0), mark events as committed,
1031 	 * so we do not put_constraint() in case new events are added
1032 	 * and fail to be scheduled
1033 	 *
1034 	 * We invoke the lower level commit callback to lock the resource
1035 	 *
1036 	 * We do not need to do all of this in case we are called to
1037 	 * validate an event group (assign == NULL)
1038 	 */
1039 	if (!unsched && assign) {
1040 		for (i = 0; i < n; i++) {
1041 			e = cpuc->event_list[i];
1042 			static_call_cond(x86_pmu_commit_scheduling)(cpuc, i, assign[i]);
1043 		}
1044 	} else {
1045 		for (i = n0; i < n; i++) {
1046 			e = cpuc->event_list[i];
1047 
1048 			/*
1049 			 * release events that failed scheduling
1050 			 */
1051 			static_call_cond(x86_pmu_put_event_constraints)(cpuc, e);
1052 
1053 			cpuc->event_constraint[i] = NULL;
1054 		}
1055 	}
1056 
1057 	static_call_cond(x86_pmu_stop_scheduling)(cpuc);
1058 
1059 	return unsched ? -EINVAL : 0;
1060 }
1061 
1062 static int add_nr_metric_event(struct cpu_hw_events *cpuc,
1063 			       struct perf_event *event)
1064 {
1065 	if (is_metric_event(event)) {
1066 		if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
1067 			return -EINVAL;
1068 		cpuc->n_metric++;
1069 		cpuc->n_txn_metric++;
1070 	}
1071 
1072 	return 0;
1073 }
1074 
1075 static void del_nr_metric_event(struct cpu_hw_events *cpuc,
1076 				struct perf_event *event)
1077 {
1078 	if (is_metric_event(event))
1079 		cpuc->n_metric--;
1080 }
1081 
1082 static int collect_event(struct cpu_hw_events *cpuc, struct perf_event *event,
1083 			 int max_count, int n)
1084 {
1085 
1086 	if (x86_pmu.intel_cap.perf_metrics && add_nr_metric_event(cpuc, event))
1087 		return -EINVAL;
1088 
1089 	if (n >= max_count + cpuc->n_metric)
1090 		return -EINVAL;
1091 
1092 	cpuc->event_list[n] = event;
1093 	if (is_counter_pair(&event->hw)) {
1094 		cpuc->n_pair++;
1095 		cpuc->n_txn_pair++;
1096 	}
1097 
1098 	return 0;
1099 }
1100 
1101 /*
1102  * dogrp: true if must collect siblings events (group)
1103  * returns total number of events and error code
1104  */
1105 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1106 {
1107 	struct perf_event *event;
1108 	int n, max_count;
1109 
1110 	max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
1111 
1112 	/* current number of events already accepted */
1113 	n = cpuc->n_events;
1114 	if (!cpuc->n_events)
1115 		cpuc->pebs_output = 0;
1116 
1117 	if (!cpuc->is_fake && leader->attr.precise_ip) {
1118 		/*
1119 		 * For PEBS->PT, if !aux_event, the group leader (PT) went
1120 		 * away, the group was broken down and this singleton event
1121 		 * can't schedule any more.
1122 		 */
1123 		if (is_pebs_pt(leader) && !leader->aux_event)
1124 			return -EINVAL;
1125 
1126 		/*
1127 		 * pebs_output: 0: no PEBS so far, 1: PT, 2: DS
1128 		 */
1129 		if (cpuc->pebs_output &&
1130 		    cpuc->pebs_output != is_pebs_pt(leader) + 1)
1131 			return -EINVAL;
1132 
1133 		cpuc->pebs_output = is_pebs_pt(leader) + 1;
1134 	}
1135 
1136 	if (is_x86_event(leader)) {
1137 		if (collect_event(cpuc, leader, max_count, n))
1138 			return -EINVAL;
1139 		n++;
1140 	}
1141 
1142 	if (!dogrp)
1143 		return n;
1144 
1145 	for_each_sibling_event(event, leader) {
1146 		if (!is_x86_event(event) || event->state <= PERF_EVENT_STATE_OFF)
1147 			continue;
1148 
1149 		if (collect_event(cpuc, event, max_count, n))
1150 			return -EINVAL;
1151 
1152 		n++;
1153 	}
1154 	return n;
1155 }
1156 
1157 static inline void x86_assign_hw_event(struct perf_event *event,
1158 				struct cpu_hw_events *cpuc, int i)
1159 {
1160 	struct hw_perf_event *hwc = &event->hw;
1161 	int idx;
1162 
1163 	idx = hwc->idx = cpuc->assign[i];
1164 	hwc->last_cpu = smp_processor_id();
1165 	hwc->last_tag = ++cpuc->tags[i];
1166 
1167 	switch (hwc->idx) {
1168 	case INTEL_PMC_IDX_FIXED_BTS:
1169 	case INTEL_PMC_IDX_FIXED_VLBR:
1170 		hwc->config_base = 0;
1171 		hwc->event_base	= 0;
1172 		break;
1173 
1174 	case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
1175 		/* All the metric events are mapped onto the fixed counter 3. */
1176 		idx = INTEL_PMC_IDX_FIXED_SLOTS;
1177 		fallthrough;
1178 	case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS-1:
1179 		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1180 		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 +
1181 				(idx - INTEL_PMC_IDX_FIXED);
1182 		hwc->event_base_rdpmc = (idx - INTEL_PMC_IDX_FIXED) |
1183 					INTEL_PMC_FIXED_RDPMC_BASE;
1184 		break;
1185 
1186 	default:
1187 		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1188 		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1189 		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1190 		break;
1191 	}
1192 }
1193 
1194 /**
1195  * x86_perf_rdpmc_index - Return PMC counter used for event
1196  * @event: the perf_event to which the PMC counter was assigned
1197  *
1198  * The counter assigned to this performance event may change if interrupts
1199  * are enabled. This counter should thus never be used while interrupts are
1200  * enabled. Before this function is used to obtain the assigned counter the
1201  * event should be checked for validity using, for example,
1202  * perf_event_read_local(), within the same interrupt disabled section in
1203  * which this counter is planned to be used.
1204  *
1205  * Return: The index of the performance monitoring counter assigned to
1206  * @perf_event.
1207  */
1208 int x86_perf_rdpmc_index(struct perf_event *event)
1209 {
1210 	lockdep_assert_irqs_disabled();
1211 
1212 	return event->hw.event_base_rdpmc;
1213 }
1214 
1215 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1216 					struct cpu_hw_events *cpuc,
1217 					int i)
1218 {
1219 	return hwc->idx == cpuc->assign[i] &&
1220 		hwc->last_cpu == smp_processor_id() &&
1221 		hwc->last_tag == cpuc->tags[i];
1222 }
1223 
1224 static void x86_pmu_start(struct perf_event *event, int flags);
1225 
1226 static void x86_pmu_enable(struct pmu *pmu)
1227 {
1228 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1229 	struct perf_event *event;
1230 	struct hw_perf_event *hwc;
1231 	int i, added = cpuc->n_added;
1232 
1233 	if (!x86_pmu_initialized())
1234 		return;
1235 
1236 	if (cpuc->enabled)
1237 		return;
1238 
1239 	if (cpuc->n_added) {
1240 		int n_running = cpuc->n_events - cpuc->n_added;
1241 		/*
1242 		 * apply assignment obtained either from
1243 		 * hw_perf_group_sched_in() or x86_pmu_enable()
1244 		 *
1245 		 * step1: save events moving to new counters
1246 		 */
1247 		for (i = 0; i < n_running; i++) {
1248 			event = cpuc->event_list[i];
1249 			hwc = &event->hw;
1250 
1251 			/*
1252 			 * we can avoid reprogramming counter if:
1253 			 * - assigned same counter as last time
1254 			 * - running on same CPU as last time
1255 			 * - no other event has used the counter since
1256 			 */
1257 			if (hwc->idx == -1 ||
1258 			    match_prev_assignment(hwc, cpuc, i))
1259 				continue;
1260 
1261 			/*
1262 			 * Ensure we don't accidentally enable a stopped
1263 			 * counter simply because we rescheduled.
1264 			 */
1265 			if (hwc->state & PERF_HES_STOPPED)
1266 				hwc->state |= PERF_HES_ARCH;
1267 
1268 			x86_pmu_stop(event, PERF_EF_UPDATE);
1269 		}
1270 
1271 		/*
1272 		 * step2: reprogram moved events into new counters
1273 		 */
1274 		for (i = 0; i < cpuc->n_events; i++) {
1275 			event = cpuc->event_list[i];
1276 			hwc = &event->hw;
1277 
1278 			if (!match_prev_assignment(hwc, cpuc, i))
1279 				x86_assign_hw_event(event, cpuc, i);
1280 			else if (i < n_running)
1281 				continue;
1282 
1283 			if (hwc->state & PERF_HES_ARCH)
1284 				continue;
1285 
1286 			x86_pmu_start(event, PERF_EF_RELOAD);
1287 		}
1288 		cpuc->n_added = 0;
1289 		perf_events_lapic_init();
1290 	}
1291 
1292 	cpuc->enabled = 1;
1293 	barrier();
1294 
1295 	static_call(x86_pmu_enable_all)(added);
1296 }
1297 
1298 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1299 
1300 /*
1301  * Set the next IRQ period, based on the hwc->period_left value.
1302  * To be called with the event disabled in hw:
1303  */
1304 int x86_perf_event_set_period(struct perf_event *event)
1305 {
1306 	struct hw_perf_event *hwc = &event->hw;
1307 	s64 left = local64_read(&hwc->period_left);
1308 	s64 period = hwc->sample_period;
1309 	int ret = 0, idx = hwc->idx;
1310 
1311 	if (unlikely(!hwc->event_base))
1312 		return 0;
1313 
1314 	if (unlikely(is_topdown_count(event)) &&
1315 	    x86_pmu.set_topdown_event_period)
1316 		return x86_pmu.set_topdown_event_period(event);
1317 
1318 	/*
1319 	 * If we are way outside a reasonable range then just skip forward:
1320 	 */
1321 	if (unlikely(left <= -period)) {
1322 		left = period;
1323 		local64_set(&hwc->period_left, left);
1324 		hwc->last_period = period;
1325 		ret = 1;
1326 	}
1327 
1328 	if (unlikely(left <= 0)) {
1329 		left += period;
1330 		local64_set(&hwc->period_left, left);
1331 		hwc->last_period = period;
1332 		ret = 1;
1333 	}
1334 	/*
1335 	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1336 	 */
1337 	if (unlikely(left < 2))
1338 		left = 2;
1339 
1340 	if (left > x86_pmu.max_period)
1341 		left = x86_pmu.max_period;
1342 
1343 	if (x86_pmu.limit_period)
1344 		left = x86_pmu.limit_period(event, left);
1345 
1346 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1347 
1348 	/*
1349 	 * The hw event starts counting from this event offset,
1350 	 * mark it to be able to extra future deltas:
1351 	 */
1352 	local64_set(&hwc->prev_count, (u64)-left);
1353 
1354 	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1355 
1356 	/*
1357 	 * Sign extend the Merge event counter's upper 16 bits since
1358 	 * we currently declare a 48-bit counter width
1359 	 */
1360 	if (is_counter_pair(hwc))
1361 		wrmsrl(x86_pmu_event_addr(idx + 1), 0xffff);
1362 
1363 	/*
1364 	 * Due to erratum on certan cpu we need
1365 	 * a second write to be sure the register
1366 	 * is updated properly
1367 	 */
1368 	if (x86_pmu.perfctr_second_write) {
1369 		wrmsrl(hwc->event_base,
1370 			(u64)(-left) & x86_pmu.cntval_mask);
1371 	}
1372 
1373 	perf_event_update_userpage(event);
1374 
1375 	return ret;
1376 }
1377 
1378 void x86_pmu_enable_event(struct perf_event *event)
1379 {
1380 	if (__this_cpu_read(cpu_hw_events.enabled))
1381 		__x86_pmu_enable_event(&event->hw,
1382 				       ARCH_PERFMON_EVENTSEL_ENABLE);
1383 }
1384 
1385 /*
1386  * Add a single event to the PMU.
1387  *
1388  * The event is added to the group of enabled events
1389  * but only if it can be scheduled with existing events.
1390  */
1391 static int x86_pmu_add(struct perf_event *event, int flags)
1392 {
1393 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1394 	struct hw_perf_event *hwc;
1395 	int assign[X86_PMC_IDX_MAX];
1396 	int n, n0, ret;
1397 
1398 	hwc = &event->hw;
1399 
1400 	n0 = cpuc->n_events;
1401 	ret = n = collect_events(cpuc, event, false);
1402 	if (ret < 0)
1403 		goto out;
1404 
1405 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1406 	if (!(flags & PERF_EF_START))
1407 		hwc->state |= PERF_HES_ARCH;
1408 
1409 	/*
1410 	 * If group events scheduling transaction was started,
1411 	 * skip the schedulability test here, it will be performed
1412 	 * at commit time (->commit_txn) as a whole.
1413 	 *
1414 	 * If commit fails, we'll call ->del() on all events
1415 	 * for which ->add() was called.
1416 	 */
1417 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1418 		goto done_collect;
1419 
1420 	ret = static_call(x86_pmu_schedule_events)(cpuc, n, assign);
1421 	if (ret)
1422 		goto out;
1423 	/*
1424 	 * copy new assignment, now we know it is possible
1425 	 * will be used by hw_perf_enable()
1426 	 */
1427 	memcpy(cpuc->assign, assign, n*sizeof(int));
1428 
1429 done_collect:
1430 	/*
1431 	 * Commit the collect_events() state. See x86_pmu_del() and
1432 	 * x86_pmu_*_txn().
1433 	 */
1434 	cpuc->n_events = n;
1435 	cpuc->n_added += n - n0;
1436 	cpuc->n_txn += n - n0;
1437 
1438 	/*
1439 	 * This is before x86_pmu_enable() will call x86_pmu_start(),
1440 	 * so we enable LBRs before an event needs them etc..
1441 	 */
1442 	static_call_cond(x86_pmu_add)(event);
1443 
1444 	ret = 0;
1445 out:
1446 	return ret;
1447 }
1448 
1449 static void x86_pmu_start(struct perf_event *event, int flags)
1450 {
1451 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1452 	int idx = event->hw.idx;
1453 
1454 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1455 		return;
1456 
1457 	if (WARN_ON_ONCE(idx == -1))
1458 		return;
1459 
1460 	if (flags & PERF_EF_RELOAD) {
1461 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1462 		x86_perf_event_set_period(event);
1463 	}
1464 
1465 	event->hw.state = 0;
1466 
1467 	cpuc->events[idx] = event;
1468 	__set_bit(idx, cpuc->active_mask);
1469 	__set_bit(idx, cpuc->running);
1470 	static_call(x86_pmu_enable)(event);
1471 	perf_event_update_userpage(event);
1472 }
1473 
1474 void perf_event_print_debug(void)
1475 {
1476 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1477 	u64 pebs, debugctl;
1478 	struct cpu_hw_events *cpuc;
1479 	unsigned long flags;
1480 	int cpu, idx;
1481 
1482 	if (!x86_pmu.num_counters)
1483 		return;
1484 
1485 	local_irq_save(flags);
1486 
1487 	cpu = smp_processor_id();
1488 	cpuc = &per_cpu(cpu_hw_events, cpu);
1489 
1490 	if (x86_pmu.version >= 2) {
1491 		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1492 		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1493 		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1494 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1495 
1496 		pr_info("\n");
1497 		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1498 		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1499 		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1500 		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1501 		if (x86_pmu.pebs_constraints) {
1502 			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1503 			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1504 		}
1505 		if (x86_pmu.lbr_nr) {
1506 			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1507 			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1508 		}
1509 	}
1510 	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1511 
1512 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1513 		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1514 		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1515 
1516 		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1517 
1518 		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1519 			cpu, idx, pmc_ctrl);
1520 		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1521 			cpu, idx, pmc_count);
1522 		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1523 			cpu, idx, prev_left);
1524 	}
1525 	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1526 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1527 
1528 		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1529 			cpu, idx, pmc_count);
1530 	}
1531 	local_irq_restore(flags);
1532 }
1533 
1534 void x86_pmu_stop(struct perf_event *event, int flags)
1535 {
1536 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1537 	struct hw_perf_event *hwc = &event->hw;
1538 
1539 	if (test_bit(hwc->idx, cpuc->active_mask)) {
1540 		static_call(x86_pmu_disable)(event);
1541 		__clear_bit(hwc->idx, cpuc->active_mask);
1542 		cpuc->events[hwc->idx] = NULL;
1543 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1544 		hwc->state |= PERF_HES_STOPPED;
1545 	}
1546 
1547 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1548 		/*
1549 		 * Drain the remaining delta count out of a event
1550 		 * that we are disabling:
1551 		 */
1552 		x86_perf_event_update(event);
1553 		hwc->state |= PERF_HES_UPTODATE;
1554 	}
1555 }
1556 
1557 static void x86_pmu_del(struct perf_event *event, int flags)
1558 {
1559 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1560 	int i;
1561 
1562 	/*
1563 	 * If we're called during a txn, we only need to undo x86_pmu.add.
1564 	 * The events never got scheduled and ->cancel_txn will truncate
1565 	 * the event_list.
1566 	 *
1567 	 * XXX assumes any ->del() called during a TXN will only be on
1568 	 * an event added during that same TXN.
1569 	 */
1570 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1571 		goto do_del;
1572 
1573 	/*
1574 	 * Not a TXN, therefore cleanup properly.
1575 	 */
1576 	x86_pmu_stop(event, PERF_EF_UPDATE);
1577 
1578 	for (i = 0; i < cpuc->n_events; i++) {
1579 		if (event == cpuc->event_list[i])
1580 			break;
1581 	}
1582 
1583 	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1584 		return;
1585 
1586 	/* If we have a newly added event; make sure to decrease n_added. */
1587 	if (i >= cpuc->n_events - cpuc->n_added)
1588 		--cpuc->n_added;
1589 
1590 	static_call_cond(x86_pmu_put_event_constraints)(cpuc, event);
1591 
1592 	/* Delete the array entry. */
1593 	while (++i < cpuc->n_events) {
1594 		cpuc->event_list[i-1] = cpuc->event_list[i];
1595 		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1596 	}
1597 	cpuc->event_constraint[i-1] = NULL;
1598 	--cpuc->n_events;
1599 	if (x86_pmu.intel_cap.perf_metrics)
1600 		del_nr_metric_event(cpuc, event);
1601 
1602 	perf_event_update_userpage(event);
1603 
1604 do_del:
1605 
1606 	/*
1607 	 * This is after x86_pmu_stop(); so we disable LBRs after any
1608 	 * event can need them etc..
1609 	 */
1610 	static_call_cond(x86_pmu_del)(event);
1611 }
1612 
1613 int x86_pmu_handle_irq(struct pt_regs *regs)
1614 {
1615 	struct perf_sample_data data;
1616 	struct cpu_hw_events *cpuc;
1617 	struct perf_event *event;
1618 	int idx, handled = 0;
1619 	u64 val;
1620 
1621 	cpuc = this_cpu_ptr(&cpu_hw_events);
1622 
1623 	/*
1624 	 * Some chipsets need to unmask the LVTPC in a particular spot
1625 	 * inside the nmi handler.  As a result, the unmasking was pushed
1626 	 * into all the nmi handlers.
1627 	 *
1628 	 * This generic handler doesn't seem to have any issues where the
1629 	 * unmasking occurs so it was left at the top.
1630 	 */
1631 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1632 
1633 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1634 		if (!test_bit(idx, cpuc->active_mask))
1635 			continue;
1636 
1637 		event = cpuc->events[idx];
1638 
1639 		val = x86_perf_event_update(event);
1640 		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1641 			continue;
1642 
1643 		/*
1644 		 * event overflow
1645 		 */
1646 		handled++;
1647 		perf_sample_data_init(&data, 0, event->hw.last_period);
1648 
1649 		if (!x86_perf_event_set_period(event))
1650 			continue;
1651 
1652 		if (perf_event_overflow(event, &data, regs))
1653 			x86_pmu_stop(event, 0);
1654 	}
1655 
1656 	if (handled)
1657 		inc_irq_stat(apic_perf_irqs);
1658 
1659 	return handled;
1660 }
1661 
1662 void perf_events_lapic_init(void)
1663 {
1664 	if (!x86_pmu.apic || !x86_pmu_initialized())
1665 		return;
1666 
1667 	/*
1668 	 * Always use NMI for PMU
1669 	 */
1670 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1671 }
1672 
1673 static int
1674 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1675 {
1676 	u64 start_clock;
1677 	u64 finish_clock;
1678 	int ret;
1679 
1680 	/*
1681 	 * All PMUs/events that share this PMI handler should make sure to
1682 	 * increment active_events for their events.
1683 	 */
1684 	if (!atomic_read(&active_events))
1685 		return NMI_DONE;
1686 
1687 	start_clock = sched_clock();
1688 	ret = static_call(x86_pmu_handle_irq)(regs);
1689 	finish_clock = sched_clock();
1690 
1691 	perf_sample_event_took(finish_clock - start_clock);
1692 
1693 	return ret;
1694 }
1695 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1696 
1697 struct event_constraint emptyconstraint;
1698 struct event_constraint unconstrained;
1699 
1700 static int x86_pmu_prepare_cpu(unsigned int cpu)
1701 {
1702 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1703 	int i;
1704 
1705 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1706 		cpuc->kfree_on_online[i] = NULL;
1707 	if (x86_pmu.cpu_prepare)
1708 		return x86_pmu.cpu_prepare(cpu);
1709 	return 0;
1710 }
1711 
1712 static int x86_pmu_dead_cpu(unsigned int cpu)
1713 {
1714 	if (x86_pmu.cpu_dead)
1715 		x86_pmu.cpu_dead(cpu);
1716 	return 0;
1717 }
1718 
1719 static int x86_pmu_online_cpu(unsigned int cpu)
1720 {
1721 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1722 	int i;
1723 
1724 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1725 		kfree(cpuc->kfree_on_online[i]);
1726 		cpuc->kfree_on_online[i] = NULL;
1727 	}
1728 	return 0;
1729 }
1730 
1731 static int x86_pmu_starting_cpu(unsigned int cpu)
1732 {
1733 	if (x86_pmu.cpu_starting)
1734 		x86_pmu.cpu_starting(cpu);
1735 	return 0;
1736 }
1737 
1738 static int x86_pmu_dying_cpu(unsigned int cpu)
1739 {
1740 	if (x86_pmu.cpu_dying)
1741 		x86_pmu.cpu_dying(cpu);
1742 	return 0;
1743 }
1744 
1745 static void __init pmu_check_apic(void)
1746 {
1747 	if (boot_cpu_has(X86_FEATURE_APIC))
1748 		return;
1749 
1750 	x86_pmu.apic = 0;
1751 	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1752 	pr_info("no hardware sampling interrupt available.\n");
1753 
1754 	/*
1755 	 * If we have a PMU initialized but no APIC
1756 	 * interrupts, we cannot sample hardware
1757 	 * events (user-space has to fall back and
1758 	 * sample via a hrtimer based software event):
1759 	 */
1760 	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1761 
1762 }
1763 
1764 static struct attribute_group x86_pmu_format_group __ro_after_init = {
1765 	.name = "format",
1766 	.attrs = NULL,
1767 };
1768 
1769 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1770 {
1771 	struct perf_pmu_events_attr *pmu_attr =
1772 		container_of(attr, struct perf_pmu_events_attr, attr);
1773 	u64 config = 0;
1774 
1775 	if (pmu_attr->id < x86_pmu.max_events)
1776 		config = x86_pmu.event_map(pmu_attr->id);
1777 
1778 	/* string trumps id */
1779 	if (pmu_attr->event_str)
1780 		return sprintf(page, "%s", pmu_attr->event_str);
1781 
1782 	return x86_pmu.events_sysfs_show(page, config);
1783 }
1784 EXPORT_SYMBOL_GPL(events_sysfs_show);
1785 
1786 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1787 			  char *page)
1788 {
1789 	struct perf_pmu_events_ht_attr *pmu_attr =
1790 		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1791 
1792 	/*
1793 	 * Report conditional events depending on Hyper-Threading.
1794 	 *
1795 	 * This is overly conservative as usually the HT special
1796 	 * handling is not needed if the other CPU thread is idle.
1797 	 *
1798 	 * Note this does not (and cannot) handle the case when thread
1799 	 * siblings are invisible, for example with virtualization
1800 	 * if they are owned by some other guest.  The user tool
1801 	 * has to re-read when a thread sibling gets onlined later.
1802 	 */
1803 	return sprintf(page, "%s",
1804 			topology_max_smt_threads() > 1 ?
1805 			pmu_attr->event_str_ht :
1806 			pmu_attr->event_str_noht);
1807 }
1808 
1809 EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1810 EVENT_ATTR(instructions,		INSTRUCTIONS		);
1811 EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1812 EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1813 EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1814 EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1815 EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1816 EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1817 EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1818 EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1819 
1820 static struct attribute *empty_attrs;
1821 
1822 static struct attribute *events_attr[] = {
1823 	EVENT_PTR(CPU_CYCLES),
1824 	EVENT_PTR(INSTRUCTIONS),
1825 	EVENT_PTR(CACHE_REFERENCES),
1826 	EVENT_PTR(CACHE_MISSES),
1827 	EVENT_PTR(BRANCH_INSTRUCTIONS),
1828 	EVENT_PTR(BRANCH_MISSES),
1829 	EVENT_PTR(BUS_CYCLES),
1830 	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1831 	EVENT_PTR(STALLED_CYCLES_BACKEND),
1832 	EVENT_PTR(REF_CPU_CYCLES),
1833 	NULL,
1834 };
1835 
1836 /*
1837  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1838  * out of events_attr attributes.
1839  */
1840 static umode_t
1841 is_visible(struct kobject *kobj, struct attribute *attr, int idx)
1842 {
1843 	struct perf_pmu_events_attr *pmu_attr;
1844 
1845 	if (idx >= x86_pmu.max_events)
1846 		return 0;
1847 
1848 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
1849 	/* str trumps id */
1850 	return pmu_attr->event_str || x86_pmu.event_map(idx) ? attr->mode : 0;
1851 }
1852 
1853 static struct attribute_group x86_pmu_events_group __ro_after_init = {
1854 	.name = "events",
1855 	.attrs = events_attr,
1856 	.is_visible = is_visible,
1857 };
1858 
1859 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1860 {
1861 	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1862 	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1863 	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1864 	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1865 	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1866 	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1867 	ssize_t ret;
1868 
1869 	/*
1870 	* We have whole page size to spend and just little data
1871 	* to write, so we can safely use sprintf.
1872 	*/
1873 	ret = sprintf(page, "event=0x%02llx", event);
1874 
1875 	if (umask)
1876 		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1877 
1878 	if (edge)
1879 		ret += sprintf(page + ret, ",edge");
1880 
1881 	if (pc)
1882 		ret += sprintf(page + ret, ",pc");
1883 
1884 	if (any)
1885 		ret += sprintf(page + ret, ",any");
1886 
1887 	if (inv)
1888 		ret += sprintf(page + ret, ",inv");
1889 
1890 	if (cmask)
1891 		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1892 
1893 	ret += sprintf(page + ret, "\n");
1894 
1895 	return ret;
1896 }
1897 
1898 static struct attribute_group x86_pmu_attr_group;
1899 static struct attribute_group x86_pmu_caps_group;
1900 
1901 static void x86_pmu_static_call_update(void)
1902 {
1903 	static_call_update(x86_pmu_handle_irq, x86_pmu.handle_irq);
1904 	static_call_update(x86_pmu_disable_all, x86_pmu.disable_all);
1905 	static_call_update(x86_pmu_enable_all, x86_pmu.enable_all);
1906 	static_call_update(x86_pmu_enable, x86_pmu.enable);
1907 	static_call_update(x86_pmu_disable, x86_pmu.disable);
1908 
1909 	static_call_update(x86_pmu_add, x86_pmu.add);
1910 	static_call_update(x86_pmu_del, x86_pmu.del);
1911 	static_call_update(x86_pmu_read, x86_pmu.read);
1912 
1913 	static_call_update(x86_pmu_schedule_events, x86_pmu.schedule_events);
1914 	static_call_update(x86_pmu_get_event_constraints, x86_pmu.get_event_constraints);
1915 	static_call_update(x86_pmu_put_event_constraints, x86_pmu.put_event_constraints);
1916 
1917 	static_call_update(x86_pmu_start_scheduling, x86_pmu.start_scheduling);
1918 	static_call_update(x86_pmu_commit_scheduling, x86_pmu.commit_scheduling);
1919 	static_call_update(x86_pmu_stop_scheduling, x86_pmu.stop_scheduling);
1920 
1921 	static_call_update(x86_pmu_sched_task, x86_pmu.sched_task);
1922 	static_call_update(x86_pmu_swap_task_ctx, x86_pmu.swap_task_ctx);
1923 
1924 	static_call_update(x86_pmu_drain_pebs, x86_pmu.drain_pebs);
1925 	static_call_update(x86_pmu_pebs_aliases, x86_pmu.pebs_aliases);
1926 }
1927 
1928 static void _x86_pmu_read(struct perf_event *event)
1929 {
1930 	x86_perf_event_update(event);
1931 }
1932 
1933 static int __init init_hw_perf_events(void)
1934 {
1935 	struct x86_pmu_quirk *quirk;
1936 	int err;
1937 
1938 	pr_info("Performance Events: ");
1939 
1940 	switch (boot_cpu_data.x86_vendor) {
1941 	case X86_VENDOR_INTEL:
1942 		err = intel_pmu_init();
1943 		break;
1944 	case X86_VENDOR_AMD:
1945 		err = amd_pmu_init();
1946 		break;
1947 	case X86_VENDOR_HYGON:
1948 		err = amd_pmu_init();
1949 		x86_pmu.name = "HYGON";
1950 		break;
1951 	case X86_VENDOR_ZHAOXIN:
1952 	case X86_VENDOR_CENTAUR:
1953 		err = zhaoxin_pmu_init();
1954 		break;
1955 	default:
1956 		err = -ENOTSUPP;
1957 	}
1958 	if (err != 0) {
1959 		pr_cont("no PMU driver, software events only.\n");
1960 		return 0;
1961 	}
1962 
1963 	pmu_check_apic();
1964 
1965 	/* sanity check that the hardware exists or is emulated */
1966 	if (!check_hw_exists())
1967 		return 0;
1968 
1969 	pr_cont("%s PMU driver.\n", x86_pmu.name);
1970 
1971 	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1972 
1973 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1974 		quirk->func();
1975 
1976 	if (!x86_pmu.intel_ctrl)
1977 		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1978 
1979 	perf_events_lapic_init();
1980 	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1981 
1982 	unconstrained = (struct event_constraint)
1983 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1984 				   0, x86_pmu.num_counters, 0, 0);
1985 
1986 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1987 
1988 	if (!x86_pmu.events_sysfs_show)
1989 		x86_pmu_events_group.attrs = &empty_attrs;
1990 
1991 	pmu.attr_update = x86_pmu.attr_update;
1992 
1993 	pr_info("... version:                %d\n",     x86_pmu.version);
1994 	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
1995 	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
1996 	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
1997 	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
1998 	pr_info("... fixed-purpose events:   %d\n",     x86_pmu.num_counters_fixed);
1999 	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
2000 
2001 	if (!x86_pmu.read)
2002 		x86_pmu.read = _x86_pmu_read;
2003 
2004 	x86_pmu_static_call_update();
2005 
2006 	/*
2007 	 * Install callbacks. Core will call them for each online
2008 	 * cpu.
2009 	 */
2010 	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
2011 				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
2012 	if (err)
2013 		return err;
2014 
2015 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
2016 				"perf/x86:starting", x86_pmu_starting_cpu,
2017 				x86_pmu_dying_cpu);
2018 	if (err)
2019 		goto out;
2020 
2021 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
2022 				x86_pmu_online_cpu, NULL);
2023 	if (err)
2024 		goto out1;
2025 
2026 	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
2027 	if (err)
2028 		goto out2;
2029 
2030 	return 0;
2031 
2032 out2:
2033 	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
2034 out1:
2035 	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
2036 out:
2037 	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
2038 	return err;
2039 }
2040 early_initcall(init_hw_perf_events);
2041 
2042 static void x86_pmu_read(struct perf_event *event)
2043 {
2044 	static_call(x86_pmu_read)(event);
2045 }
2046 
2047 /*
2048  * Start group events scheduling transaction
2049  * Set the flag to make pmu::enable() not perform the
2050  * schedulability test, it will be performed at commit time
2051  *
2052  * We only support PERF_PMU_TXN_ADD transactions. Save the
2053  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
2054  * transactions.
2055  */
2056 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
2057 {
2058 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2059 
2060 	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
2061 
2062 	cpuc->txn_flags = txn_flags;
2063 	if (txn_flags & ~PERF_PMU_TXN_ADD)
2064 		return;
2065 
2066 	perf_pmu_disable(pmu);
2067 	__this_cpu_write(cpu_hw_events.n_txn, 0);
2068 	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
2069 	__this_cpu_write(cpu_hw_events.n_txn_metric, 0);
2070 }
2071 
2072 /*
2073  * Stop group events scheduling transaction
2074  * Clear the flag and pmu::enable() will perform the
2075  * schedulability test.
2076  */
2077 static void x86_pmu_cancel_txn(struct pmu *pmu)
2078 {
2079 	unsigned int txn_flags;
2080 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2081 
2082 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
2083 
2084 	txn_flags = cpuc->txn_flags;
2085 	cpuc->txn_flags = 0;
2086 	if (txn_flags & ~PERF_PMU_TXN_ADD)
2087 		return;
2088 
2089 	/*
2090 	 * Truncate collected array by the number of events added in this
2091 	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
2092 	 */
2093 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
2094 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
2095 	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
2096 	__this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
2097 	perf_pmu_enable(pmu);
2098 }
2099 
2100 /*
2101  * Commit group events scheduling transaction
2102  * Perform the group schedulability test as a whole
2103  * Return 0 if success
2104  *
2105  * Does not cancel the transaction on failure; expects the caller to do this.
2106  */
2107 static int x86_pmu_commit_txn(struct pmu *pmu)
2108 {
2109 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2110 	int assign[X86_PMC_IDX_MAX];
2111 	int n, ret;
2112 
2113 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
2114 
2115 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
2116 		cpuc->txn_flags = 0;
2117 		return 0;
2118 	}
2119 
2120 	n = cpuc->n_events;
2121 
2122 	if (!x86_pmu_initialized())
2123 		return -EAGAIN;
2124 
2125 	ret = static_call(x86_pmu_schedule_events)(cpuc, n, assign);
2126 	if (ret)
2127 		return ret;
2128 
2129 	/*
2130 	 * copy new assignment, now we know it is possible
2131 	 * will be used by hw_perf_enable()
2132 	 */
2133 	memcpy(cpuc->assign, assign, n*sizeof(int));
2134 
2135 	cpuc->txn_flags = 0;
2136 	perf_pmu_enable(pmu);
2137 	return 0;
2138 }
2139 /*
2140  * a fake_cpuc is used to validate event groups. Due to
2141  * the extra reg logic, we need to also allocate a fake
2142  * per_core and per_cpu structure. Otherwise, group events
2143  * using extra reg may conflict without the kernel being
2144  * able to catch this when the last event gets added to
2145  * the group.
2146  */
2147 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
2148 {
2149 	intel_cpuc_finish(cpuc);
2150 	kfree(cpuc);
2151 }
2152 
2153 static struct cpu_hw_events *allocate_fake_cpuc(void)
2154 {
2155 	struct cpu_hw_events *cpuc;
2156 	int cpu = raw_smp_processor_id();
2157 
2158 	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
2159 	if (!cpuc)
2160 		return ERR_PTR(-ENOMEM);
2161 	cpuc->is_fake = 1;
2162 
2163 	if (intel_cpuc_prepare(cpuc, cpu))
2164 		goto error;
2165 
2166 	return cpuc;
2167 error:
2168 	free_fake_cpuc(cpuc);
2169 	return ERR_PTR(-ENOMEM);
2170 }
2171 
2172 /*
2173  * validate that we can schedule this event
2174  */
2175 static int validate_event(struct perf_event *event)
2176 {
2177 	struct cpu_hw_events *fake_cpuc;
2178 	struct event_constraint *c;
2179 	int ret = 0;
2180 
2181 	fake_cpuc = allocate_fake_cpuc();
2182 	if (IS_ERR(fake_cpuc))
2183 		return PTR_ERR(fake_cpuc);
2184 
2185 	c = x86_pmu.get_event_constraints(fake_cpuc, 0, event);
2186 
2187 	if (!c || !c->weight)
2188 		ret = -EINVAL;
2189 
2190 	if (x86_pmu.put_event_constraints)
2191 		x86_pmu.put_event_constraints(fake_cpuc, event);
2192 
2193 	free_fake_cpuc(fake_cpuc);
2194 
2195 	return ret;
2196 }
2197 
2198 /*
2199  * validate a single event group
2200  *
2201  * validation include:
2202  *	- check events are compatible which each other
2203  *	- events do not compete for the same counter
2204  *	- number of events <= number of counters
2205  *
2206  * validation ensures the group can be loaded onto the
2207  * PMU if it was the only group available.
2208  */
2209 static int validate_group(struct perf_event *event)
2210 {
2211 	struct perf_event *leader = event->group_leader;
2212 	struct cpu_hw_events *fake_cpuc;
2213 	int ret = -EINVAL, n;
2214 
2215 	fake_cpuc = allocate_fake_cpuc();
2216 	if (IS_ERR(fake_cpuc))
2217 		return PTR_ERR(fake_cpuc);
2218 	/*
2219 	 * the event is not yet connected with its
2220 	 * siblings therefore we must first collect
2221 	 * existing siblings, then add the new event
2222 	 * before we can simulate the scheduling
2223 	 */
2224 	n = collect_events(fake_cpuc, leader, true);
2225 	if (n < 0)
2226 		goto out;
2227 
2228 	fake_cpuc->n_events = n;
2229 	n = collect_events(fake_cpuc, event, false);
2230 	if (n < 0)
2231 		goto out;
2232 
2233 	fake_cpuc->n_events = 0;
2234 	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2235 
2236 out:
2237 	free_fake_cpuc(fake_cpuc);
2238 	return ret;
2239 }
2240 
2241 static int x86_pmu_event_init(struct perf_event *event)
2242 {
2243 	struct pmu *tmp;
2244 	int err;
2245 
2246 	switch (event->attr.type) {
2247 	case PERF_TYPE_RAW:
2248 	case PERF_TYPE_HARDWARE:
2249 	case PERF_TYPE_HW_CACHE:
2250 		break;
2251 
2252 	default:
2253 		return -ENOENT;
2254 	}
2255 
2256 	err = __x86_pmu_event_init(event);
2257 	if (!err) {
2258 		/*
2259 		 * we temporarily connect event to its pmu
2260 		 * such that validate_group() can classify
2261 		 * it as an x86 event using is_x86_event()
2262 		 */
2263 		tmp = event->pmu;
2264 		event->pmu = &pmu;
2265 
2266 		if (event->group_leader != event)
2267 			err = validate_group(event);
2268 		else
2269 			err = validate_event(event);
2270 
2271 		event->pmu = tmp;
2272 	}
2273 	if (err) {
2274 		if (event->destroy)
2275 			event->destroy(event);
2276 	}
2277 
2278 	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2279 	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2280 		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2281 
2282 	return err;
2283 }
2284 
2285 static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2286 {
2287 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2288 		return;
2289 
2290 	/*
2291 	 * This function relies on not being called concurrently in two
2292 	 * tasks in the same mm.  Otherwise one task could observe
2293 	 * perf_rdpmc_allowed > 1 and return all the way back to
2294 	 * userspace with CR4.PCE clear while another task is still
2295 	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2296 	 *
2297 	 * For now, this can't happen because all callers hold mmap_lock
2298 	 * for write.  If this changes, we'll need a different solution.
2299 	 */
2300 	mmap_assert_write_locked(mm);
2301 
2302 	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2303 		on_each_cpu_mask(mm_cpumask(mm), cr4_update_pce, NULL, 1);
2304 }
2305 
2306 static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2307 {
2308 
2309 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2310 		return;
2311 
2312 	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2313 		on_each_cpu_mask(mm_cpumask(mm), cr4_update_pce, NULL, 1);
2314 }
2315 
2316 static int x86_pmu_event_idx(struct perf_event *event)
2317 {
2318 	struct hw_perf_event *hwc = &event->hw;
2319 
2320 	if (!(hwc->flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2321 		return 0;
2322 
2323 	if (is_metric_idx(hwc->idx))
2324 		return INTEL_PMC_FIXED_RDPMC_METRICS + 1;
2325 	else
2326 		return hwc->event_base_rdpmc + 1;
2327 }
2328 
2329 static ssize_t get_attr_rdpmc(struct device *cdev,
2330 			      struct device_attribute *attr,
2331 			      char *buf)
2332 {
2333 	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2334 }
2335 
2336 static ssize_t set_attr_rdpmc(struct device *cdev,
2337 			      struct device_attribute *attr,
2338 			      const char *buf, size_t count)
2339 {
2340 	unsigned long val;
2341 	ssize_t ret;
2342 
2343 	ret = kstrtoul(buf, 0, &val);
2344 	if (ret)
2345 		return ret;
2346 
2347 	if (val > 2)
2348 		return -EINVAL;
2349 
2350 	if (x86_pmu.attr_rdpmc_broken)
2351 		return -ENOTSUPP;
2352 
2353 	if (val != x86_pmu.attr_rdpmc) {
2354 		/*
2355 		 * Changing into or out of never available or always available,
2356 		 * aka perf-event-bypassing mode. This path is extremely slow,
2357 		 * but only root can trigger it, so it's okay.
2358 		 */
2359 		if (val == 0)
2360 			static_branch_inc(&rdpmc_never_available_key);
2361 		else if (x86_pmu.attr_rdpmc == 0)
2362 			static_branch_dec(&rdpmc_never_available_key);
2363 
2364 		if (val == 2)
2365 			static_branch_inc(&rdpmc_always_available_key);
2366 		else if (x86_pmu.attr_rdpmc == 2)
2367 			static_branch_dec(&rdpmc_always_available_key);
2368 
2369 		on_each_cpu(cr4_update_pce, NULL, 1);
2370 		x86_pmu.attr_rdpmc = val;
2371 	}
2372 
2373 	return count;
2374 }
2375 
2376 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2377 
2378 static struct attribute *x86_pmu_attrs[] = {
2379 	&dev_attr_rdpmc.attr,
2380 	NULL,
2381 };
2382 
2383 static struct attribute_group x86_pmu_attr_group __ro_after_init = {
2384 	.attrs = x86_pmu_attrs,
2385 };
2386 
2387 static ssize_t max_precise_show(struct device *cdev,
2388 				  struct device_attribute *attr,
2389 				  char *buf)
2390 {
2391 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2392 }
2393 
2394 static DEVICE_ATTR_RO(max_precise);
2395 
2396 static struct attribute *x86_pmu_caps_attrs[] = {
2397 	&dev_attr_max_precise.attr,
2398 	NULL
2399 };
2400 
2401 static struct attribute_group x86_pmu_caps_group __ro_after_init = {
2402 	.name = "caps",
2403 	.attrs = x86_pmu_caps_attrs,
2404 };
2405 
2406 static const struct attribute_group *x86_pmu_attr_groups[] = {
2407 	&x86_pmu_attr_group,
2408 	&x86_pmu_format_group,
2409 	&x86_pmu_events_group,
2410 	&x86_pmu_caps_group,
2411 	NULL,
2412 };
2413 
2414 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2415 {
2416 	static_call_cond(x86_pmu_sched_task)(ctx, sched_in);
2417 }
2418 
2419 static void x86_pmu_swap_task_ctx(struct perf_event_context *prev,
2420 				  struct perf_event_context *next)
2421 {
2422 	static_call_cond(x86_pmu_swap_task_ctx)(prev, next);
2423 }
2424 
2425 void perf_check_microcode(void)
2426 {
2427 	if (x86_pmu.check_microcode)
2428 		x86_pmu.check_microcode();
2429 }
2430 
2431 static int x86_pmu_check_period(struct perf_event *event, u64 value)
2432 {
2433 	if (x86_pmu.check_period && x86_pmu.check_period(event, value))
2434 		return -EINVAL;
2435 
2436 	if (value && x86_pmu.limit_period) {
2437 		if (x86_pmu.limit_period(event, value) > value)
2438 			return -EINVAL;
2439 	}
2440 
2441 	return 0;
2442 }
2443 
2444 static int x86_pmu_aux_output_match(struct perf_event *event)
2445 {
2446 	if (!(pmu.capabilities & PERF_PMU_CAP_AUX_OUTPUT))
2447 		return 0;
2448 
2449 	if (x86_pmu.aux_output_match)
2450 		return x86_pmu.aux_output_match(event);
2451 
2452 	return 0;
2453 }
2454 
2455 static struct pmu pmu = {
2456 	.pmu_enable		= x86_pmu_enable,
2457 	.pmu_disable		= x86_pmu_disable,
2458 
2459 	.attr_groups		= x86_pmu_attr_groups,
2460 
2461 	.event_init		= x86_pmu_event_init,
2462 
2463 	.event_mapped		= x86_pmu_event_mapped,
2464 	.event_unmapped		= x86_pmu_event_unmapped,
2465 
2466 	.add			= x86_pmu_add,
2467 	.del			= x86_pmu_del,
2468 	.start			= x86_pmu_start,
2469 	.stop			= x86_pmu_stop,
2470 	.read			= x86_pmu_read,
2471 
2472 	.start_txn		= x86_pmu_start_txn,
2473 	.cancel_txn		= x86_pmu_cancel_txn,
2474 	.commit_txn		= x86_pmu_commit_txn,
2475 
2476 	.event_idx		= x86_pmu_event_idx,
2477 	.sched_task		= x86_pmu_sched_task,
2478 	.swap_task_ctx		= x86_pmu_swap_task_ctx,
2479 	.check_period		= x86_pmu_check_period,
2480 
2481 	.aux_output_match	= x86_pmu_aux_output_match,
2482 };
2483 
2484 void arch_perf_update_userpage(struct perf_event *event,
2485 			       struct perf_event_mmap_page *userpg, u64 now)
2486 {
2487 	struct cyc2ns_data data;
2488 	u64 offset;
2489 
2490 	userpg->cap_user_time = 0;
2491 	userpg->cap_user_time_zero = 0;
2492 	userpg->cap_user_rdpmc =
2493 		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2494 	userpg->pmc_width = x86_pmu.cntval_bits;
2495 
2496 	if (!using_native_sched_clock() || !sched_clock_stable())
2497 		return;
2498 
2499 	cyc2ns_read_begin(&data);
2500 
2501 	offset = data.cyc2ns_offset + __sched_clock_offset;
2502 
2503 	/*
2504 	 * Internal timekeeping for enabled/running/stopped times
2505 	 * is always in the local_clock domain.
2506 	 */
2507 	userpg->cap_user_time = 1;
2508 	userpg->time_mult = data.cyc2ns_mul;
2509 	userpg->time_shift = data.cyc2ns_shift;
2510 	userpg->time_offset = offset - now;
2511 
2512 	/*
2513 	 * cap_user_time_zero doesn't make sense when we're using a different
2514 	 * time base for the records.
2515 	 */
2516 	if (!event->attr.use_clockid) {
2517 		userpg->cap_user_time_zero = 1;
2518 		userpg->time_zero = offset;
2519 	}
2520 
2521 	cyc2ns_read_end();
2522 }
2523 
2524 /*
2525  * Determine whether the regs were taken from an irq/exception handler rather
2526  * than from perf_arch_fetch_caller_regs().
2527  */
2528 static bool perf_hw_regs(struct pt_regs *regs)
2529 {
2530 	return regs->flags & X86_EFLAGS_FIXED;
2531 }
2532 
2533 void
2534 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2535 {
2536 	struct unwind_state state;
2537 	unsigned long addr;
2538 
2539 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2540 		/* TODO: We don't support guest os callchain now */
2541 		return;
2542 	}
2543 
2544 	if (perf_callchain_store(entry, regs->ip))
2545 		return;
2546 
2547 	if (perf_hw_regs(regs))
2548 		unwind_start(&state, current, regs, NULL);
2549 	else
2550 		unwind_start(&state, current, NULL, (void *)regs->sp);
2551 
2552 	for (; !unwind_done(&state); unwind_next_frame(&state)) {
2553 		addr = unwind_get_return_address(&state);
2554 		if (!addr || perf_callchain_store(entry, addr))
2555 			return;
2556 	}
2557 }
2558 
2559 static inline int
2560 valid_user_frame(const void __user *fp, unsigned long size)
2561 {
2562 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2563 }
2564 
2565 static unsigned long get_segment_base(unsigned int segment)
2566 {
2567 	struct desc_struct *desc;
2568 	unsigned int idx = segment >> 3;
2569 
2570 	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2571 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2572 		struct ldt_struct *ldt;
2573 
2574 		/* IRQs are off, so this synchronizes with smp_store_release */
2575 		ldt = READ_ONCE(current->active_mm->context.ldt);
2576 		if (!ldt || idx >= ldt->nr_entries)
2577 			return 0;
2578 
2579 		desc = &ldt->entries[idx];
2580 #else
2581 		return 0;
2582 #endif
2583 	} else {
2584 		if (idx >= GDT_ENTRIES)
2585 			return 0;
2586 
2587 		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2588 	}
2589 
2590 	return get_desc_base(desc);
2591 }
2592 
2593 #ifdef CONFIG_IA32_EMULATION
2594 
2595 #include <linux/compat.h>
2596 
2597 static inline int
2598 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2599 {
2600 	/* 32-bit process in 64-bit kernel. */
2601 	unsigned long ss_base, cs_base;
2602 	struct stack_frame_ia32 frame;
2603 	const struct stack_frame_ia32 __user *fp;
2604 
2605 	if (user_64bit_mode(regs))
2606 		return 0;
2607 
2608 	cs_base = get_segment_base(regs->cs);
2609 	ss_base = get_segment_base(regs->ss);
2610 
2611 	fp = compat_ptr(ss_base + regs->bp);
2612 	pagefault_disable();
2613 	while (entry->nr < entry->max_stack) {
2614 		if (!valid_user_frame(fp, sizeof(frame)))
2615 			break;
2616 
2617 		if (__get_user(frame.next_frame, &fp->next_frame))
2618 			break;
2619 		if (__get_user(frame.return_address, &fp->return_address))
2620 			break;
2621 
2622 		perf_callchain_store(entry, cs_base + frame.return_address);
2623 		fp = compat_ptr(ss_base + frame.next_frame);
2624 	}
2625 	pagefault_enable();
2626 	return 1;
2627 }
2628 #else
2629 static inline int
2630 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2631 {
2632     return 0;
2633 }
2634 #endif
2635 
2636 void
2637 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2638 {
2639 	struct stack_frame frame;
2640 	const struct stack_frame __user *fp;
2641 
2642 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2643 		/* TODO: We don't support guest os callchain now */
2644 		return;
2645 	}
2646 
2647 	/*
2648 	 * We don't know what to do with VM86 stacks.. ignore them for now.
2649 	 */
2650 	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2651 		return;
2652 
2653 	fp = (void __user *)regs->bp;
2654 
2655 	perf_callchain_store(entry, regs->ip);
2656 
2657 	if (!nmi_uaccess_okay())
2658 		return;
2659 
2660 	if (perf_callchain_user32(regs, entry))
2661 		return;
2662 
2663 	pagefault_disable();
2664 	while (entry->nr < entry->max_stack) {
2665 		if (!valid_user_frame(fp, sizeof(frame)))
2666 			break;
2667 
2668 		if (__get_user(frame.next_frame, &fp->next_frame))
2669 			break;
2670 		if (__get_user(frame.return_address, &fp->return_address))
2671 			break;
2672 
2673 		perf_callchain_store(entry, frame.return_address);
2674 		fp = (void __user *)frame.next_frame;
2675 	}
2676 	pagefault_enable();
2677 }
2678 
2679 /*
2680  * Deal with code segment offsets for the various execution modes:
2681  *
2682  *   VM86 - the good olde 16 bit days, where the linear address is
2683  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2684  *
2685  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2686  *          to figure out what the 32bit base address is.
2687  *
2688  *    X32 - has TIF_X32 set, but is running in x86_64
2689  *
2690  * X86_64 - CS,DS,SS,ES are all zero based.
2691  */
2692 static unsigned long code_segment_base(struct pt_regs *regs)
2693 {
2694 	/*
2695 	 * For IA32 we look at the GDT/LDT segment base to convert the
2696 	 * effective IP to a linear address.
2697 	 */
2698 
2699 #ifdef CONFIG_X86_32
2700 	/*
2701 	 * If we are in VM86 mode, add the segment offset to convert to a
2702 	 * linear address.
2703 	 */
2704 	if (regs->flags & X86_VM_MASK)
2705 		return 0x10 * regs->cs;
2706 
2707 	if (user_mode(regs) && regs->cs != __USER_CS)
2708 		return get_segment_base(regs->cs);
2709 #else
2710 	if (user_mode(regs) && !user_64bit_mode(regs) &&
2711 	    regs->cs != __USER32_CS)
2712 		return get_segment_base(regs->cs);
2713 #endif
2714 	return 0;
2715 }
2716 
2717 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2718 {
2719 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2720 		return perf_guest_cbs->get_guest_ip();
2721 
2722 	return regs->ip + code_segment_base(regs);
2723 }
2724 
2725 unsigned long perf_misc_flags(struct pt_regs *regs)
2726 {
2727 	int misc = 0;
2728 
2729 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2730 		if (perf_guest_cbs->is_user_mode())
2731 			misc |= PERF_RECORD_MISC_GUEST_USER;
2732 		else
2733 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2734 	} else {
2735 		if (user_mode(regs))
2736 			misc |= PERF_RECORD_MISC_USER;
2737 		else
2738 			misc |= PERF_RECORD_MISC_KERNEL;
2739 	}
2740 
2741 	if (regs->flags & PERF_EFLAGS_EXACT)
2742 		misc |= PERF_RECORD_MISC_EXACT_IP;
2743 
2744 	return misc;
2745 }
2746 
2747 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2748 {
2749 	cap->version		= x86_pmu.version;
2750 	cap->num_counters_gp	= x86_pmu.num_counters;
2751 	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2752 	cap->bit_width_gp	= x86_pmu.cntval_bits;
2753 	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2754 	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2755 	cap->events_mask_len	= x86_pmu.events_mask_len;
2756 }
2757 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
2758