xref: /openbmc/linux/arch/x86/events/core.c (revision 6a143a7c)
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 DEFINE_STATIC_CALL_NULL(x86_pmu_guest_get_msrs,  *x86_pmu.guest_get_msrs);
85 
86 u64 __read_mostly hw_cache_event_ids
87 				[PERF_COUNT_HW_CACHE_MAX]
88 				[PERF_COUNT_HW_CACHE_OP_MAX]
89 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
90 u64 __read_mostly hw_cache_extra_regs
91 				[PERF_COUNT_HW_CACHE_MAX]
92 				[PERF_COUNT_HW_CACHE_OP_MAX]
93 				[PERF_COUNT_HW_CACHE_RESULT_MAX];
94 
95 /*
96  * Propagate event elapsed time into the generic event.
97  * Can only be executed on the CPU where the event is active.
98  * Returns the delta events processed.
99  */
100 u64 x86_perf_event_update(struct perf_event *event)
101 {
102 	struct hw_perf_event *hwc = &event->hw;
103 	int shift = 64 - x86_pmu.cntval_bits;
104 	u64 prev_raw_count, new_raw_count;
105 	u64 delta;
106 
107 	if (unlikely(!hwc->event_base))
108 		return 0;
109 
110 	if (unlikely(is_topdown_count(event)) && x86_pmu.update_topdown_event)
111 		return x86_pmu.update_topdown_event(event);
112 
113 	/*
114 	 * Careful: an NMI might modify the previous event value.
115 	 *
116 	 * Our tactic to handle this is to first atomically read and
117 	 * exchange a new raw count - then add that new-prev delta
118 	 * count to the generic event atomically:
119 	 */
120 again:
121 	prev_raw_count = local64_read(&hwc->prev_count);
122 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
123 
124 	if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
125 					new_raw_count) != prev_raw_count)
126 		goto again;
127 
128 	/*
129 	 * Now we have the new raw value and have updated the prev
130 	 * timestamp already. We can now calculate the elapsed delta
131 	 * (event-)time and add that to the generic event.
132 	 *
133 	 * Careful, not all hw sign-extends above the physical width
134 	 * of the count.
135 	 */
136 	delta = (new_raw_count << shift) - (prev_raw_count << shift);
137 	delta >>= shift;
138 
139 	local64_add(delta, &event->count);
140 	local64_sub(delta, &hwc->period_left);
141 
142 	return new_raw_count;
143 }
144 
145 /*
146  * Find and validate any extra registers to set up.
147  */
148 static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
149 {
150 	struct hw_perf_event_extra *reg;
151 	struct extra_reg *er;
152 
153 	reg = &event->hw.extra_reg;
154 
155 	if (!x86_pmu.extra_regs)
156 		return 0;
157 
158 	for (er = x86_pmu.extra_regs; er->msr; er++) {
159 		if (er->event != (config & er->config_mask))
160 			continue;
161 		if (event->attr.config1 & ~er->valid_mask)
162 			return -EINVAL;
163 		/* Check if the extra msrs can be safely accessed*/
164 		if (!er->extra_msr_access)
165 			return -ENXIO;
166 
167 		reg->idx = er->idx;
168 		reg->config = event->attr.config1;
169 		reg->reg = er->msr;
170 		break;
171 	}
172 	return 0;
173 }
174 
175 static atomic_t active_events;
176 static atomic_t pmc_refcount;
177 static DEFINE_MUTEX(pmc_reserve_mutex);
178 
179 #ifdef CONFIG_X86_LOCAL_APIC
180 
181 static bool reserve_pmc_hardware(void)
182 {
183 	int i;
184 
185 	for (i = 0; i < x86_pmu.num_counters; i++) {
186 		if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
187 			goto perfctr_fail;
188 	}
189 
190 	for (i = 0; i < x86_pmu.num_counters; i++) {
191 		if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
192 			goto eventsel_fail;
193 	}
194 
195 	return true;
196 
197 eventsel_fail:
198 	for (i--; i >= 0; i--)
199 		release_evntsel_nmi(x86_pmu_config_addr(i));
200 
201 	i = x86_pmu.num_counters;
202 
203 perfctr_fail:
204 	for (i--; i >= 0; i--)
205 		release_perfctr_nmi(x86_pmu_event_addr(i));
206 
207 	return false;
208 }
209 
210 static void release_pmc_hardware(void)
211 {
212 	int i;
213 
214 	for (i = 0; i < x86_pmu.num_counters; i++) {
215 		release_perfctr_nmi(x86_pmu_event_addr(i));
216 		release_evntsel_nmi(x86_pmu_config_addr(i));
217 	}
218 }
219 
220 #else
221 
222 static bool reserve_pmc_hardware(void) { return true; }
223 static void release_pmc_hardware(void) {}
224 
225 #endif
226 
227 static bool check_hw_exists(void)
228 {
229 	u64 val, val_fail = -1, val_new= ~0;
230 	int i, reg, reg_fail = -1, ret = 0;
231 	int bios_fail = 0;
232 	int reg_safe = -1;
233 
234 	/*
235 	 * Check to see if the BIOS enabled any of the counters, if so
236 	 * complain and bail.
237 	 */
238 	for (i = 0; i < x86_pmu.num_counters; i++) {
239 		reg = x86_pmu_config_addr(i);
240 		ret = rdmsrl_safe(reg, &val);
241 		if (ret)
242 			goto msr_fail;
243 		if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
244 			bios_fail = 1;
245 			val_fail = val;
246 			reg_fail = reg;
247 		} else {
248 			reg_safe = i;
249 		}
250 	}
251 
252 	if (x86_pmu.num_counters_fixed) {
253 		reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
254 		ret = rdmsrl_safe(reg, &val);
255 		if (ret)
256 			goto msr_fail;
257 		for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
258 			if (fixed_counter_disabled(i))
259 				continue;
260 			if (val & (0x03 << i*4)) {
261 				bios_fail = 1;
262 				val_fail = val;
263 				reg_fail = reg;
264 			}
265 		}
266 	}
267 
268 	/*
269 	 * If all the counters are enabled, the below test will always
270 	 * fail.  The tools will also become useless in this scenario.
271 	 * Just fail and disable the hardware counters.
272 	 */
273 
274 	if (reg_safe == -1) {
275 		reg = reg_safe;
276 		goto msr_fail;
277 	}
278 
279 	/*
280 	 * Read the current value, change it and read it back to see if it
281 	 * matches, this is needed to detect certain hardware emulators
282 	 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
283 	 */
284 	reg = x86_pmu_event_addr(reg_safe);
285 	if (rdmsrl_safe(reg, &val))
286 		goto msr_fail;
287 	val ^= 0xffffUL;
288 	ret = wrmsrl_safe(reg, val);
289 	ret |= rdmsrl_safe(reg, &val_new);
290 	if (ret || val != val_new)
291 		goto msr_fail;
292 
293 	/*
294 	 * We still allow the PMU driver to operate:
295 	 */
296 	if (bios_fail) {
297 		pr_cont("Broken BIOS detected, complain to your hardware vendor.\n");
298 		pr_err(FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n",
299 			      reg_fail, val_fail);
300 	}
301 
302 	return true;
303 
304 msr_fail:
305 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
306 		pr_cont("PMU not available due to virtualization, using software events only.\n");
307 	} else {
308 		pr_cont("Broken PMU hardware detected, using software events only.\n");
309 		pr_err("Failed to access perfctr msr (MSR %x is %Lx)\n",
310 		       reg, val_new);
311 	}
312 
313 	return false;
314 }
315 
316 static void hw_perf_event_destroy(struct perf_event *event)
317 {
318 	x86_release_hardware();
319 	atomic_dec(&active_events);
320 }
321 
322 void hw_perf_lbr_event_destroy(struct perf_event *event)
323 {
324 	hw_perf_event_destroy(event);
325 
326 	/* undo the lbr/bts event accounting */
327 	x86_del_exclusive(x86_lbr_exclusive_lbr);
328 }
329 
330 static inline int x86_pmu_initialized(void)
331 {
332 	return x86_pmu.handle_irq != NULL;
333 }
334 
335 static inline int
336 set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
337 {
338 	struct perf_event_attr *attr = &event->attr;
339 	unsigned int cache_type, cache_op, cache_result;
340 	u64 config, val;
341 
342 	config = attr->config;
343 
344 	cache_type = (config >> 0) & 0xff;
345 	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
346 		return -EINVAL;
347 	cache_type = array_index_nospec(cache_type, PERF_COUNT_HW_CACHE_MAX);
348 
349 	cache_op = (config >>  8) & 0xff;
350 	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
351 		return -EINVAL;
352 	cache_op = array_index_nospec(cache_op, PERF_COUNT_HW_CACHE_OP_MAX);
353 
354 	cache_result = (config >> 16) & 0xff;
355 	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
356 		return -EINVAL;
357 	cache_result = array_index_nospec(cache_result, PERF_COUNT_HW_CACHE_RESULT_MAX);
358 
359 	val = hw_cache_event_ids[cache_type][cache_op][cache_result];
360 
361 	if (val == 0)
362 		return -ENOENT;
363 
364 	if (val == -1)
365 		return -EINVAL;
366 
367 	hwc->config |= val;
368 	attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
369 	return x86_pmu_extra_regs(val, event);
370 }
371 
372 int x86_reserve_hardware(void)
373 {
374 	int err = 0;
375 
376 	if (!atomic_inc_not_zero(&pmc_refcount)) {
377 		mutex_lock(&pmc_reserve_mutex);
378 		if (atomic_read(&pmc_refcount) == 0) {
379 			if (!reserve_pmc_hardware())
380 				err = -EBUSY;
381 			else
382 				reserve_ds_buffers();
383 		}
384 		if (!err)
385 			atomic_inc(&pmc_refcount);
386 		mutex_unlock(&pmc_reserve_mutex);
387 	}
388 
389 	return err;
390 }
391 
392 void x86_release_hardware(void)
393 {
394 	if (atomic_dec_and_mutex_lock(&pmc_refcount, &pmc_reserve_mutex)) {
395 		release_pmc_hardware();
396 		release_ds_buffers();
397 		release_lbr_buffers();
398 		mutex_unlock(&pmc_reserve_mutex);
399 	}
400 }
401 
402 /*
403  * Check if we can create event of a certain type (that no conflicting events
404  * are present).
405  */
406 int x86_add_exclusive(unsigned int what)
407 {
408 	int i;
409 
410 	/*
411 	 * When lbr_pt_coexist we allow PT to coexist with either LBR or BTS.
412 	 * LBR and BTS are still mutually exclusive.
413 	 */
414 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
415 		goto out;
416 
417 	if (!atomic_inc_not_zero(&x86_pmu.lbr_exclusive[what])) {
418 		mutex_lock(&pmc_reserve_mutex);
419 		for (i = 0; i < ARRAY_SIZE(x86_pmu.lbr_exclusive); i++) {
420 			if (i != what && atomic_read(&x86_pmu.lbr_exclusive[i]))
421 				goto fail_unlock;
422 		}
423 		atomic_inc(&x86_pmu.lbr_exclusive[what]);
424 		mutex_unlock(&pmc_reserve_mutex);
425 	}
426 
427 out:
428 	atomic_inc(&active_events);
429 	return 0;
430 
431 fail_unlock:
432 	mutex_unlock(&pmc_reserve_mutex);
433 	return -EBUSY;
434 }
435 
436 void x86_del_exclusive(unsigned int what)
437 {
438 	atomic_dec(&active_events);
439 
440 	/*
441 	 * See the comment in x86_add_exclusive().
442 	 */
443 	if (x86_pmu.lbr_pt_coexist && what == x86_lbr_exclusive_pt)
444 		return;
445 
446 	atomic_dec(&x86_pmu.lbr_exclusive[what]);
447 }
448 
449 int x86_setup_perfctr(struct perf_event *event)
450 {
451 	struct perf_event_attr *attr = &event->attr;
452 	struct hw_perf_event *hwc = &event->hw;
453 	u64 config;
454 
455 	if (!is_sampling_event(event)) {
456 		hwc->sample_period = x86_pmu.max_period;
457 		hwc->last_period = hwc->sample_period;
458 		local64_set(&hwc->period_left, hwc->sample_period);
459 	}
460 
461 	if (attr->type == PERF_TYPE_RAW)
462 		return x86_pmu_extra_regs(event->attr.config, event);
463 
464 	if (attr->type == PERF_TYPE_HW_CACHE)
465 		return set_ext_hw_attr(hwc, event);
466 
467 	if (attr->config >= x86_pmu.max_events)
468 		return -EINVAL;
469 
470 	attr->config = array_index_nospec((unsigned long)attr->config, x86_pmu.max_events);
471 
472 	/*
473 	 * The generic map:
474 	 */
475 	config = x86_pmu.event_map(attr->config);
476 
477 	if (config == 0)
478 		return -ENOENT;
479 
480 	if (config == -1LL)
481 		return -EINVAL;
482 
483 	hwc->config |= config;
484 
485 	return 0;
486 }
487 
488 /*
489  * check that branch_sample_type is compatible with
490  * settings needed for precise_ip > 1 which implies
491  * using the LBR to capture ALL taken branches at the
492  * priv levels of the measurement
493  */
494 static inline int precise_br_compat(struct perf_event *event)
495 {
496 	u64 m = event->attr.branch_sample_type;
497 	u64 b = 0;
498 
499 	/* must capture all branches */
500 	if (!(m & PERF_SAMPLE_BRANCH_ANY))
501 		return 0;
502 
503 	m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
504 
505 	if (!event->attr.exclude_user)
506 		b |= PERF_SAMPLE_BRANCH_USER;
507 
508 	if (!event->attr.exclude_kernel)
509 		b |= PERF_SAMPLE_BRANCH_KERNEL;
510 
511 	/*
512 	 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
513 	 */
514 
515 	return m == b;
516 }
517 
518 int x86_pmu_max_precise(void)
519 {
520 	int precise = 0;
521 
522 	/* Support for constant skid */
523 	if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
524 		precise++;
525 
526 		/* Support for IP fixup */
527 		if (x86_pmu.lbr_nr || x86_pmu.intel_cap.pebs_format >= 2)
528 			precise++;
529 
530 		if (x86_pmu.pebs_prec_dist)
531 			precise++;
532 	}
533 	return precise;
534 }
535 
536 int x86_pmu_hw_config(struct perf_event *event)
537 {
538 	if (event->attr.precise_ip) {
539 		int precise = x86_pmu_max_precise();
540 
541 		if (event->attr.precise_ip > precise)
542 			return -EOPNOTSUPP;
543 
544 		/* There's no sense in having PEBS for non sampling events: */
545 		if (!is_sampling_event(event))
546 			return -EINVAL;
547 	}
548 	/*
549 	 * check that PEBS LBR correction does not conflict with
550 	 * whatever the user is asking with attr->branch_sample_type
551 	 */
552 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format < 2) {
553 		u64 *br_type = &event->attr.branch_sample_type;
554 
555 		if (has_branch_stack(event)) {
556 			if (!precise_br_compat(event))
557 				return -EOPNOTSUPP;
558 
559 			/* branch_sample_type is compatible */
560 
561 		} else {
562 			/*
563 			 * user did not specify  branch_sample_type
564 			 *
565 			 * For PEBS fixups, we capture all
566 			 * the branches at the priv level of the
567 			 * event.
568 			 */
569 			*br_type = PERF_SAMPLE_BRANCH_ANY;
570 
571 			if (!event->attr.exclude_user)
572 				*br_type |= PERF_SAMPLE_BRANCH_USER;
573 
574 			if (!event->attr.exclude_kernel)
575 				*br_type |= PERF_SAMPLE_BRANCH_KERNEL;
576 		}
577 	}
578 
579 	if (event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK)
580 		event->attach_state |= PERF_ATTACH_TASK_DATA;
581 
582 	/*
583 	 * Generate PMC IRQs:
584 	 * (keep 'enabled' bit clear for now)
585 	 */
586 	event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
587 
588 	/*
589 	 * Count user and OS events unless requested not to
590 	 */
591 	if (!event->attr.exclude_user)
592 		event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
593 	if (!event->attr.exclude_kernel)
594 		event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
595 
596 	if (event->attr.type == PERF_TYPE_RAW)
597 		event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
598 
599 	if (event->attr.sample_period && x86_pmu.limit_period) {
600 		if (x86_pmu.limit_period(event, event->attr.sample_period) >
601 				event->attr.sample_period)
602 			return -EINVAL;
603 	}
604 
605 	/* sample_regs_user never support XMM registers */
606 	if (unlikely(event->attr.sample_regs_user & PERF_REG_EXTENDED_MASK))
607 		return -EINVAL;
608 	/*
609 	 * Besides the general purpose registers, XMM registers may
610 	 * be collected in PEBS on some platforms, e.g. Icelake
611 	 */
612 	if (unlikely(event->attr.sample_regs_intr & PERF_REG_EXTENDED_MASK)) {
613 		if (!(event->pmu->capabilities & PERF_PMU_CAP_EXTENDED_REGS))
614 			return -EINVAL;
615 
616 		if (!event->attr.precise_ip)
617 			return -EINVAL;
618 	}
619 
620 	return x86_setup_perfctr(event);
621 }
622 
623 /*
624  * Setup the hardware configuration for a given attr_type
625  */
626 static int __x86_pmu_event_init(struct perf_event *event)
627 {
628 	int err;
629 
630 	if (!x86_pmu_initialized())
631 		return -ENODEV;
632 
633 	err = x86_reserve_hardware();
634 	if (err)
635 		return err;
636 
637 	atomic_inc(&active_events);
638 	event->destroy = hw_perf_event_destroy;
639 
640 	event->hw.idx = -1;
641 	event->hw.last_cpu = -1;
642 	event->hw.last_tag = ~0ULL;
643 
644 	/* mark unused */
645 	event->hw.extra_reg.idx = EXTRA_REG_NONE;
646 	event->hw.branch_reg.idx = EXTRA_REG_NONE;
647 
648 	return x86_pmu.hw_config(event);
649 }
650 
651 void x86_pmu_disable_all(void)
652 {
653 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
654 	int idx;
655 
656 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
657 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
658 		u64 val;
659 
660 		if (!test_bit(idx, cpuc->active_mask))
661 			continue;
662 		rdmsrl(x86_pmu_config_addr(idx), val);
663 		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
664 			continue;
665 		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
666 		wrmsrl(x86_pmu_config_addr(idx), val);
667 		if (is_counter_pair(hwc))
668 			wrmsrl(x86_pmu_config_addr(idx + 1), 0);
669 	}
670 }
671 
672 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr)
673 {
674 	return static_call(x86_pmu_guest_get_msrs)(nr);
675 }
676 EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
677 
678 /*
679  * There may be PMI landing after enabled=0. The PMI hitting could be before or
680  * after disable_all.
681  *
682  * If PMI hits before disable_all, the PMU will be disabled in the NMI handler.
683  * It will not be re-enabled in the NMI handler again, because enabled=0. After
684  * handling the NMI, disable_all will be called, which will not change the
685  * state either. If PMI hits after disable_all, the PMU is already disabled
686  * before entering NMI handler. The NMI handler will not change the state
687  * either.
688  *
689  * So either situation is harmless.
690  */
691 static void x86_pmu_disable(struct pmu *pmu)
692 {
693 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
694 
695 	if (!x86_pmu_initialized())
696 		return;
697 
698 	if (!cpuc->enabled)
699 		return;
700 
701 	cpuc->n_added = 0;
702 	cpuc->enabled = 0;
703 	barrier();
704 
705 	static_call(x86_pmu_disable_all)();
706 }
707 
708 void x86_pmu_enable_all(int added)
709 {
710 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
711 	int idx;
712 
713 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
714 		struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
715 
716 		if (!test_bit(idx, cpuc->active_mask))
717 			continue;
718 
719 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
720 	}
721 }
722 
723 static struct pmu pmu;
724 
725 static inline int is_x86_event(struct perf_event *event)
726 {
727 	return event->pmu == &pmu;
728 }
729 
730 struct pmu *x86_get_pmu(void)
731 {
732 	return &pmu;
733 }
734 /*
735  * Event scheduler state:
736  *
737  * Assign events iterating over all events and counters, beginning
738  * with events with least weights first. Keep the current iterator
739  * state in struct sched_state.
740  */
741 struct sched_state {
742 	int	weight;
743 	int	event;		/* event index */
744 	int	counter;	/* counter index */
745 	int	unassigned;	/* number of events to be assigned left */
746 	int	nr_gp;		/* number of GP counters used */
747 	u64	used;
748 };
749 
750 /* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
751 #define	SCHED_STATES_MAX	2
752 
753 struct perf_sched {
754 	int			max_weight;
755 	int			max_events;
756 	int			max_gp;
757 	int			saved_states;
758 	struct event_constraint	**constraints;
759 	struct sched_state	state;
760 	struct sched_state	saved[SCHED_STATES_MAX];
761 };
762 
763 /*
764  * Initialize interator that runs through all events and counters.
765  */
766 static void perf_sched_init(struct perf_sched *sched, struct event_constraint **constraints,
767 			    int num, int wmin, int wmax, int gpmax)
768 {
769 	int idx;
770 
771 	memset(sched, 0, sizeof(*sched));
772 	sched->max_events	= num;
773 	sched->max_weight	= wmax;
774 	sched->max_gp		= gpmax;
775 	sched->constraints	= constraints;
776 
777 	for (idx = 0; idx < num; idx++) {
778 		if (constraints[idx]->weight == wmin)
779 			break;
780 	}
781 
782 	sched->state.event	= idx;		/* start with min weight */
783 	sched->state.weight	= wmin;
784 	sched->state.unassigned	= num;
785 }
786 
787 static void perf_sched_save_state(struct perf_sched *sched)
788 {
789 	if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
790 		return;
791 
792 	sched->saved[sched->saved_states] = sched->state;
793 	sched->saved_states++;
794 }
795 
796 static bool perf_sched_restore_state(struct perf_sched *sched)
797 {
798 	if (!sched->saved_states)
799 		return false;
800 
801 	sched->saved_states--;
802 	sched->state = sched->saved[sched->saved_states];
803 
804 	/* this assignment didn't work out */
805 	/* XXX broken vs EVENT_PAIR */
806 	sched->state.used &= ~BIT_ULL(sched->state.counter);
807 
808 	/* try the next one */
809 	sched->state.counter++;
810 
811 	return true;
812 }
813 
814 /*
815  * Select a counter for the current event to schedule. Return true on
816  * success.
817  */
818 static bool __perf_sched_find_counter(struct perf_sched *sched)
819 {
820 	struct event_constraint *c;
821 	int idx;
822 
823 	if (!sched->state.unassigned)
824 		return false;
825 
826 	if (sched->state.event >= sched->max_events)
827 		return false;
828 
829 	c = sched->constraints[sched->state.event];
830 	/* Prefer fixed purpose counters */
831 	if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
832 		idx = INTEL_PMC_IDX_FIXED;
833 		for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
834 			u64 mask = BIT_ULL(idx);
835 
836 			if (sched->state.used & mask)
837 				continue;
838 
839 			sched->state.used |= mask;
840 			goto done;
841 		}
842 	}
843 
844 	/* Grab the first unused counter starting with idx */
845 	idx = sched->state.counter;
846 	for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
847 		u64 mask = BIT_ULL(idx);
848 
849 		if (c->flags & PERF_X86_EVENT_PAIR)
850 			mask |= mask << 1;
851 
852 		if (sched->state.used & mask)
853 			continue;
854 
855 		if (sched->state.nr_gp++ >= sched->max_gp)
856 			return false;
857 
858 		sched->state.used |= mask;
859 		goto done;
860 	}
861 
862 	return false;
863 
864 done:
865 	sched->state.counter = idx;
866 
867 	if (c->overlap)
868 		perf_sched_save_state(sched);
869 
870 	return true;
871 }
872 
873 static bool perf_sched_find_counter(struct perf_sched *sched)
874 {
875 	while (!__perf_sched_find_counter(sched)) {
876 		if (!perf_sched_restore_state(sched))
877 			return false;
878 	}
879 
880 	return true;
881 }
882 
883 /*
884  * Go through all unassigned events and find the next one to schedule.
885  * Take events with the least weight first. Return true on success.
886  */
887 static bool perf_sched_next_event(struct perf_sched *sched)
888 {
889 	struct event_constraint *c;
890 
891 	if (!sched->state.unassigned || !--sched->state.unassigned)
892 		return false;
893 
894 	do {
895 		/* next event */
896 		sched->state.event++;
897 		if (sched->state.event >= sched->max_events) {
898 			/* next weight */
899 			sched->state.event = 0;
900 			sched->state.weight++;
901 			if (sched->state.weight > sched->max_weight)
902 				return false;
903 		}
904 		c = sched->constraints[sched->state.event];
905 	} while (c->weight != sched->state.weight);
906 
907 	sched->state.counter = 0;	/* start with first counter */
908 
909 	return true;
910 }
911 
912 /*
913  * Assign a counter for each event.
914  */
915 int perf_assign_events(struct event_constraint **constraints, int n,
916 			int wmin, int wmax, int gpmax, int *assign)
917 {
918 	struct perf_sched sched;
919 
920 	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
921 
922 	do {
923 		if (!perf_sched_find_counter(&sched))
924 			break;	/* failed */
925 		if (assign)
926 			assign[sched.state.event] = sched.state.counter;
927 	} while (perf_sched_next_event(&sched));
928 
929 	return sched.state.unassigned;
930 }
931 EXPORT_SYMBOL_GPL(perf_assign_events);
932 
933 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
934 {
935 	struct event_constraint *c;
936 	struct perf_event *e;
937 	int n0, i, wmin, wmax, unsched = 0;
938 	struct hw_perf_event *hwc;
939 	u64 used_mask = 0;
940 
941 	/*
942 	 * Compute the number of events already present; see x86_pmu_add(),
943 	 * validate_group() and x86_pmu_commit_txn(). For the former two
944 	 * cpuc->n_events hasn't been updated yet, while for the latter
945 	 * cpuc->n_txn contains the number of events added in the current
946 	 * transaction.
947 	 */
948 	n0 = cpuc->n_events;
949 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
950 		n0 -= cpuc->n_txn;
951 
952 	static_call_cond(x86_pmu_start_scheduling)(cpuc);
953 
954 	for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
955 		c = cpuc->event_constraint[i];
956 
957 		/*
958 		 * Previously scheduled events should have a cached constraint,
959 		 * while new events should not have one.
960 		 */
961 		WARN_ON_ONCE((c && i >= n0) || (!c && i < n0));
962 
963 		/*
964 		 * Request constraints for new events; or for those events that
965 		 * have a dynamic constraint -- for those the constraint can
966 		 * change due to external factors (sibling state, allow_tfa).
967 		 */
968 		if (!c || (c->flags & PERF_X86_EVENT_DYNAMIC)) {
969 			c = static_call(x86_pmu_get_event_constraints)(cpuc, i, cpuc->event_list[i]);
970 			cpuc->event_constraint[i] = c;
971 		}
972 
973 		wmin = min(wmin, c->weight);
974 		wmax = max(wmax, c->weight);
975 	}
976 
977 	/*
978 	 * fastpath, try to reuse previous register
979 	 */
980 	for (i = 0; i < n; i++) {
981 		u64 mask;
982 
983 		hwc = &cpuc->event_list[i]->hw;
984 		c = cpuc->event_constraint[i];
985 
986 		/* never assigned */
987 		if (hwc->idx == -1)
988 			break;
989 
990 		/* constraint still honored */
991 		if (!test_bit(hwc->idx, c->idxmsk))
992 			break;
993 
994 		mask = BIT_ULL(hwc->idx);
995 		if (is_counter_pair(hwc))
996 			mask |= mask << 1;
997 
998 		/* not already used */
999 		if (used_mask & mask)
1000 			break;
1001 
1002 		used_mask |= mask;
1003 
1004 		if (assign)
1005 			assign[i] = hwc->idx;
1006 	}
1007 
1008 	/* slow path */
1009 	if (i != n) {
1010 		int gpmax = x86_pmu.num_counters;
1011 
1012 		/*
1013 		 * Do not allow scheduling of more than half the available
1014 		 * generic counters.
1015 		 *
1016 		 * This helps avoid counter starvation of sibling thread by
1017 		 * ensuring at most half the counters cannot be in exclusive
1018 		 * mode. There is no designated counters for the limits. Any
1019 		 * N/2 counters can be used. This helps with events with
1020 		 * specific counter constraints.
1021 		 */
1022 		if (is_ht_workaround_enabled() && !cpuc->is_fake &&
1023 		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
1024 			gpmax /= 2;
1025 
1026 		/*
1027 		 * Reduce the amount of available counters to allow fitting
1028 		 * the extra Merge events needed by large increment events.
1029 		 */
1030 		if (x86_pmu.flags & PMU_FL_PAIR) {
1031 			gpmax = x86_pmu.num_counters - cpuc->n_pair;
1032 			WARN_ON(gpmax <= 0);
1033 		}
1034 
1035 		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
1036 					     wmax, gpmax, assign);
1037 	}
1038 
1039 	/*
1040 	 * In case of success (unsched = 0), mark events as committed,
1041 	 * so we do not put_constraint() in case new events are added
1042 	 * and fail to be scheduled
1043 	 *
1044 	 * We invoke the lower level commit callback to lock the resource
1045 	 *
1046 	 * We do not need to do all of this in case we are called to
1047 	 * validate an event group (assign == NULL)
1048 	 */
1049 	if (!unsched && assign) {
1050 		for (i = 0; i < n; i++) {
1051 			e = cpuc->event_list[i];
1052 			static_call_cond(x86_pmu_commit_scheduling)(cpuc, i, assign[i]);
1053 		}
1054 	} else {
1055 		for (i = n0; i < n; i++) {
1056 			e = cpuc->event_list[i];
1057 
1058 			/*
1059 			 * release events that failed scheduling
1060 			 */
1061 			static_call_cond(x86_pmu_put_event_constraints)(cpuc, e);
1062 
1063 			cpuc->event_constraint[i] = NULL;
1064 		}
1065 	}
1066 
1067 	static_call_cond(x86_pmu_stop_scheduling)(cpuc);
1068 
1069 	return unsched ? -EINVAL : 0;
1070 }
1071 
1072 static int add_nr_metric_event(struct cpu_hw_events *cpuc,
1073 			       struct perf_event *event)
1074 {
1075 	if (is_metric_event(event)) {
1076 		if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
1077 			return -EINVAL;
1078 		cpuc->n_metric++;
1079 		cpuc->n_txn_metric++;
1080 	}
1081 
1082 	return 0;
1083 }
1084 
1085 static void del_nr_metric_event(struct cpu_hw_events *cpuc,
1086 				struct perf_event *event)
1087 {
1088 	if (is_metric_event(event))
1089 		cpuc->n_metric--;
1090 }
1091 
1092 static int collect_event(struct cpu_hw_events *cpuc, struct perf_event *event,
1093 			 int max_count, int n)
1094 {
1095 
1096 	if (x86_pmu.intel_cap.perf_metrics && add_nr_metric_event(cpuc, event))
1097 		return -EINVAL;
1098 
1099 	if (n >= max_count + cpuc->n_metric)
1100 		return -EINVAL;
1101 
1102 	cpuc->event_list[n] = event;
1103 	if (is_counter_pair(&event->hw)) {
1104 		cpuc->n_pair++;
1105 		cpuc->n_txn_pair++;
1106 	}
1107 
1108 	return 0;
1109 }
1110 
1111 /*
1112  * dogrp: true if must collect siblings events (group)
1113  * returns total number of events and error code
1114  */
1115 static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
1116 {
1117 	struct perf_event *event;
1118 	int n, max_count;
1119 
1120 	max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
1121 
1122 	/* current number of events already accepted */
1123 	n = cpuc->n_events;
1124 	if (!cpuc->n_events)
1125 		cpuc->pebs_output = 0;
1126 
1127 	if (!cpuc->is_fake && leader->attr.precise_ip) {
1128 		/*
1129 		 * For PEBS->PT, if !aux_event, the group leader (PT) went
1130 		 * away, the group was broken down and this singleton event
1131 		 * can't schedule any more.
1132 		 */
1133 		if (is_pebs_pt(leader) && !leader->aux_event)
1134 			return -EINVAL;
1135 
1136 		/*
1137 		 * pebs_output: 0: no PEBS so far, 1: PT, 2: DS
1138 		 */
1139 		if (cpuc->pebs_output &&
1140 		    cpuc->pebs_output != is_pebs_pt(leader) + 1)
1141 			return -EINVAL;
1142 
1143 		cpuc->pebs_output = is_pebs_pt(leader) + 1;
1144 	}
1145 
1146 	if (is_x86_event(leader)) {
1147 		if (collect_event(cpuc, leader, max_count, n))
1148 			return -EINVAL;
1149 		n++;
1150 	}
1151 
1152 	if (!dogrp)
1153 		return n;
1154 
1155 	for_each_sibling_event(event, leader) {
1156 		if (!is_x86_event(event) || event->state <= PERF_EVENT_STATE_OFF)
1157 			continue;
1158 
1159 		if (collect_event(cpuc, event, max_count, n))
1160 			return -EINVAL;
1161 
1162 		n++;
1163 	}
1164 	return n;
1165 }
1166 
1167 static inline void x86_assign_hw_event(struct perf_event *event,
1168 				struct cpu_hw_events *cpuc, int i)
1169 {
1170 	struct hw_perf_event *hwc = &event->hw;
1171 	int idx;
1172 
1173 	idx = hwc->idx = cpuc->assign[i];
1174 	hwc->last_cpu = smp_processor_id();
1175 	hwc->last_tag = ++cpuc->tags[i];
1176 
1177 	switch (hwc->idx) {
1178 	case INTEL_PMC_IDX_FIXED_BTS:
1179 	case INTEL_PMC_IDX_FIXED_VLBR:
1180 		hwc->config_base = 0;
1181 		hwc->event_base	= 0;
1182 		break;
1183 
1184 	case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END:
1185 		/* All the metric events are mapped onto the fixed counter 3. */
1186 		idx = INTEL_PMC_IDX_FIXED_SLOTS;
1187 		fallthrough;
1188 	case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS-1:
1189 		hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
1190 		hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 +
1191 				(idx - INTEL_PMC_IDX_FIXED);
1192 		hwc->event_base_rdpmc = (idx - INTEL_PMC_IDX_FIXED) |
1193 					INTEL_PMC_FIXED_RDPMC_BASE;
1194 		break;
1195 
1196 	default:
1197 		hwc->config_base = x86_pmu_config_addr(hwc->idx);
1198 		hwc->event_base  = x86_pmu_event_addr(hwc->idx);
1199 		hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
1200 		break;
1201 	}
1202 }
1203 
1204 /**
1205  * x86_perf_rdpmc_index - Return PMC counter used for event
1206  * @event: the perf_event to which the PMC counter was assigned
1207  *
1208  * The counter assigned to this performance event may change if interrupts
1209  * are enabled. This counter should thus never be used while interrupts are
1210  * enabled. Before this function is used to obtain the assigned counter the
1211  * event should be checked for validity using, for example,
1212  * perf_event_read_local(), within the same interrupt disabled section in
1213  * which this counter is planned to be used.
1214  *
1215  * Return: The index of the performance monitoring counter assigned to
1216  * @perf_event.
1217  */
1218 int x86_perf_rdpmc_index(struct perf_event *event)
1219 {
1220 	lockdep_assert_irqs_disabled();
1221 
1222 	return event->hw.event_base_rdpmc;
1223 }
1224 
1225 static inline int match_prev_assignment(struct hw_perf_event *hwc,
1226 					struct cpu_hw_events *cpuc,
1227 					int i)
1228 {
1229 	return hwc->idx == cpuc->assign[i] &&
1230 		hwc->last_cpu == smp_processor_id() &&
1231 		hwc->last_tag == cpuc->tags[i];
1232 }
1233 
1234 static void x86_pmu_start(struct perf_event *event, int flags);
1235 
1236 static void x86_pmu_enable(struct pmu *pmu)
1237 {
1238 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1239 	struct perf_event *event;
1240 	struct hw_perf_event *hwc;
1241 	int i, added = cpuc->n_added;
1242 
1243 	if (!x86_pmu_initialized())
1244 		return;
1245 
1246 	if (cpuc->enabled)
1247 		return;
1248 
1249 	if (cpuc->n_added) {
1250 		int n_running = cpuc->n_events - cpuc->n_added;
1251 		/*
1252 		 * apply assignment obtained either from
1253 		 * hw_perf_group_sched_in() or x86_pmu_enable()
1254 		 *
1255 		 * step1: save events moving to new counters
1256 		 */
1257 		for (i = 0; i < n_running; i++) {
1258 			event = cpuc->event_list[i];
1259 			hwc = &event->hw;
1260 
1261 			/*
1262 			 * we can avoid reprogramming counter if:
1263 			 * - assigned same counter as last time
1264 			 * - running on same CPU as last time
1265 			 * - no other event has used the counter since
1266 			 */
1267 			if (hwc->idx == -1 ||
1268 			    match_prev_assignment(hwc, cpuc, i))
1269 				continue;
1270 
1271 			/*
1272 			 * Ensure we don't accidentally enable a stopped
1273 			 * counter simply because we rescheduled.
1274 			 */
1275 			if (hwc->state & PERF_HES_STOPPED)
1276 				hwc->state |= PERF_HES_ARCH;
1277 
1278 			x86_pmu_stop(event, PERF_EF_UPDATE);
1279 		}
1280 
1281 		/*
1282 		 * step2: reprogram moved events into new counters
1283 		 */
1284 		for (i = 0; i < cpuc->n_events; i++) {
1285 			event = cpuc->event_list[i];
1286 			hwc = &event->hw;
1287 
1288 			if (!match_prev_assignment(hwc, cpuc, i))
1289 				x86_assign_hw_event(event, cpuc, i);
1290 			else if (i < n_running)
1291 				continue;
1292 
1293 			if (hwc->state & PERF_HES_ARCH)
1294 				continue;
1295 
1296 			x86_pmu_start(event, PERF_EF_RELOAD);
1297 		}
1298 		cpuc->n_added = 0;
1299 		perf_events_lapic_init();
1300 	}
1301 
1302 	cpuc->enabled = 1;
1303 	barrier();
1304 
1305 	static_call(x86_pmu_enable_all)(added);
1306 }
1307 
1308 static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
1309 
1310 /*
1311  * Set the next IRQ period, based on the hwc->period_left value.
1312  * To be called with the event disabled in hw:
1313  */
1314 int x86_perf_event_set_period(struct perf_event *event)
1315 {
1316 	struct hw_perf_event *hwc = &event->hw;
1317 	s64 left = local64_read(&hwc->period_left);
1318 	s64 period = hwc->sample_period;
1319 	int ret = 0, idx = hwc->idx;
1320 
1321 	if (unlikely(!hwc->event_base))
1322 		return 0;
1323 
1324 	if (unlikely(is_topdown_count(event)) &&
1325 	    x86_pmu.set_topdown_event_period)
1326 		return x86_pmu.set_topdown_event_period(event);
1327 
1328 	/*
1329 	 * If we are way outside a reasonable range then just skip forward:
1330 	 */
1331 	if (unlikely(left <= -period)) {
1332 		left = period;
1333 		local64_set(&hwc->period_left, left);
1334 		hwc->last_period = period;
1335 		ret = 1;
1336 	}
1337 
1338 	if (unlikely(left <= 0)) {
1339 		left += period;
1340 		local64_set(&hwc->period_left, left);
1341 		hwc->last_period = period;
1342 		ret = 1;
1343 	}
1344 	/*
1345 	 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
1346 	 */
1347 	if (unlikely(left < 2))
1348 		left = 2;
1349 
1350 	if (left > x86_pmu.max_period)
1351 		left = x86_pmu.max_period;
1352 
1353 	if (x86_pmu.limit_period)
1354 		left = x86_pmu.limit_period(event, left);
1355 
1356 	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
1357 
1358 	/*
1359 	 * The hw event starts counting from this event offset,
1360 	 * mark it to be able to extra future deltas:
1361 	 */
1362 	local64_set(&hwc->prev_count, (u64)-left);
1363 
1364 	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
1365 
1366 	/*
1367 	 * Sign extend the Merge event counter's upper 16 bits since
1368 	 * we currently declare a 48-bit counter width
1369 	 */
1370 	if (is_counter_pair(hwc))
1371 		wrmsrl(x86_pmu_event_addr(idx + 1), 0xffff);
1372 
1373 	/*
1374 	 * Due to erratum on certan cpu we need
1375 	 * a second write to be sure the register
1376 	 * is updated properly
1377 	 */
1378 	if (x86_pmu.perfctr_second_write) {
1379 		wrmsrl(hwc->event_base,
1380 			(u64)(-left) & x86_pmu.cntval_mask);
1381 	}
1382 
1383 	perf_event_update_userpage(event);
1384 
1385 	return ret;
1386 }
1387 
1388 void x86_pmu_enable_event(struct perf_event *event)
1389 {
1390 	if (__this_cpu_read(cpu_hw_events.enabled))
1391 		__x86_pmu_enable_event(&event->hw,
1392 				       ARCH_PERFMON_EVENTSEL_ENABLE);
1393 }
1394 
1395 /*
1396  * Add a single event to the PMU.
1397  *
1398  * The event is added to the group of enabled events
1399  * but only if it can be scheduled with existing events.
1400  */
1401 static int x86_pmu_add(struct perf_event *event, int flags)
1402 {
1403 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1404 	struct hw_perf_event *hwc;
1405 	int assign[X86_PMC_IDX_MAX];
1406 	int n, n0, ret;
1407 
1408 	hwc = &event->hw;
1409 
1410 	n0 = cpuc->n_events;
1411 	ret = n = collect_events(cpuc, event, false);
1412 	if (ret < 0)
1413 		goto out;
1414 
1415 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1416 	if (!(flags & PERF_EF_START))
1417 		hwc->state |= PERF_HES_ARCH;
1418 
1419 	/*
1420 	 * If group events scheduling transaction was started,
1421 	 * skip the schedulability test here, it will be performed
1422 	 * at commit time (->commit_txn) as a whole.
1423 	 *
1424 	 * If commit fails, we'll call ->del() on all events
1425 	 * for which ->add() was called.
1426 	 */
1427 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1428 		goto done_collect;
1429 
1430 	ret = static_call(x86_pmu_schedule_events)(cpuc, n, assign);
1431 	if (ret)
1432 		goto out;
1433 	/*
1434 	 * copy new assignment, now we know it is possible
1435 	 * will be used by hw_perf_enable()
1436 	 */
1437 	memcpy(cpuc->assign, assign, n*sizeof(int));
1438 
1439 done_collect:
1440 	/*
1441 	 * Commit the collect_events() state. See x86_pmu_del() and
1442 	 * x86_pmu_*_txn().
1443 	 */
1444 	cpuc->n_events = n;
1445 	cpuc->n_added += n - n0;
1446 	cpuc->n_txn += n - n0;
1447 
1448 	/*
1449 	 * This is before x86_pmu_enable() will call x86_pmu_start(),
1450 	 * so we enable LBRs before an event needs them etc..
1451 	 */
1452 	static_call_cond(x86_pmu_add)(event);
1453 
1454 	ret = 0;
1455 out:
1456 	return ret;
1457 }
1458 
1459 static void x86_pmu_start(struct perf_event *event, int flags)
1460 {
1461 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1462 	int idx = event->hw.idx;
1463 
1464 	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1465 		return;
1466 
1467 	if (WARN_ON_ONCE(idx == -1))
1468 		return;
1469 
1470 	if (flags & PERF_EF_RELOAD) {
1471 		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1472 		x86_perf_event_set_period(event);
1473 	}
1474 
1475 	event->hw.state = 0;
1476 
1477 	cpuc->events[idx] = event;
1478 	__set_bit(idx, cpuc->active_mask);
1479 	__set_bit(idx, cpuc->running);
1480 	static_call(x86_pmu_enable)(event);
1481 	perf_event_update_userpage(event);
1482 }
1483 
1484 void perf_event_print_debug(void)
1485 {
1486 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1487 	u64 pebs, debugctl;
1488 	struct cpu_hw_events *cpuc;
1489 	unsigned long flags;
1490 	int cpu, idx;
1491 
1492 	if (!x86_pmu.num_counters)
1493 		return;
1494 
1495 	local_irq_save(flags);
1496 
1497 	cpu = smp_processor_id();
1498 	cpuc = &per_cpu(cpu_hw_events, cpu);
1499 
1500 	if (x86_pmu.version >= 2) {
1501 		rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1502 		rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1503 		rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1504 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1505 
1506 		pr_info("\n");
1507 		pr_info("CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
1508 		pr_info("CPU#%d: status:     %016llx\n", cpu, status);
1509 		pr_info("CPU#%d: overflow:   %016llx\n", cpu, overflow);
1510 		pr_info("CPU#%d: fixed:      %016llx\n", cpu, fixed);
1511 		if (x86_pmu.pebs_constraints) {
1512 			rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1513 			pr_info("CPU#%d: pebs:       %016llx\n", cpu, pebs);
1514 		}
1515 		if (x86_pmu.lbr_nr) {
1516 			rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
1517 			pr_info("CPU#%d: debugctl:   %016llx\n", cpu, debugctl);
1518 		}
1519 	}
1520 	pr_info("CPU#%d: active:     %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1521 
1522 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1523 		rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1524 		rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1525 
1526 		prev_left = per_cpu(pmc_prev_left[idx], cpu);
1527 
1528 		pr_info("CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
1529 			cpu, idx, pmc_ctrl);
1530 		pr_info("CPU#%d:   gen-PMC%d count: %016llx\n",
1531 			cpu, idx, pmc_count);
1532 		pr_info("CPU#%d:   gen-PMC%d left:  %016llx\n",
1533 			cpu, idx, prev_left);
1534 	}
1535 	for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1536 		if (fixed_counter_disabled(idx))
1537 			continue;
1538 		rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1539 
1540 		pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1541 			cpu, idx, pmc_count);
1542 	}
1543 	local_irq_restore(flags);
1544 }
1545 
1546 void x86_pmu_stop(struct perf_event *event, int flags)
1547 {
1548 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1549 	struct hw_perf_event *hwc = &event->hw;
1550 
1551 	if (test_bit(hwc->idx, cpuc->active_mask)) {
1552 		static_call(x86_pmu_disable)(event);
1553 		__clear_bit(hwc->idx, cpuc->active_mask);
1554 		cpuc->events[hwc->idx] = NULL;
1555 		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1556 		hwc->state |= PERF_HES_STOPPED;
1557 	}
1558 
1559 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1560 		/*
1561 		 * Drain the remaining delta count out of a event
1562 		 * that we are disabling:
1563 		 */
1564 		x86_perf_event_update(event);
1565 		hwc->state |= PERF_HES_UPTODATE;
1566 	}
1567 }
1568 
1569 static void x86_pmu_del(struct perf_event *event, int flags)
1570 {
1571 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1572 	int i;
1573 
1574 	/*
1575 	 * If we're called during a txn, we only need to undo x86_pmu.add.
1576 	 * The events never got scheduled and ->cancel_txn will truncate
1577 	 * the event_list.
1578 	 *
1579 	 * XXX assumes any ->del() called during a TXN will only be on
1580 	 * an event added during that same TXN.
1581 	 */
1582 	if (cpuc->txn_flags & PERF_PMU_TXN_ADD)
1583 		goto do_del;
1584 
1585 	/*
1586 	 * Not a TXN, therefore cleanup properly.
1587 	 */
1588 	x86_pmu_stop(event, PERF_EF_UPDATE);
1589 
1590 	for (i = 0; i < cpuc->n_events; i++) {
1591 		if (event == cpuc->event_list[i])
1592 			break;
1593 	}
1594 
1595 	if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1596 		return;
1597 
1598 	/* If we have a newly added event; make sure to decrease n_added. */
1599 	if (i >= cpuc->n_events - cpuc->n_added)
1600 		--cpuc->n_added;
1601 
1602 	static_call_cond(x86_pmu_put_event_constraints)(cpuc, event);
1603 
1604 	/* Delete the array entry. */
1605 	while (++i < cpuc->n_events) {
1606 		cpuc->event_list[i-1] = cpuc->event_list[i];
1607 		cpuc->event_constraint[i-1] = cpuc->event_constraint[i];
1608 	}
1609 	cpuc->event_constraint[i-1] = NULL;
1610 	--cpuc->n_events;
1611 	if (x86_pmu.intel_cap.perf_metrics)
1612 		del_nr_metric_event(cpuc, event);
1613 
1614 	perf_event_update_userpage(event);
1615 
1616 do_del:
1617 
1618 	/*
1619 	 * This is after x86_pmu_stop(); so we disable LBRs after any
1620 	 * event can need them etc..
1621 	 */
1622 	static_call_cond(x86_pmu_del)(event);
1623 }
1624 
1625 int x86_pmu_handle_irq(struct pt_regs *regs)
1626 {
1627 	struct perf_sample_data data;
1628 	struct cpu_hw_events *cpuc;
1629 	struct perf_event *event;
1630 	int idx, handled = 0;
1631 	u64 val;
1632 
1633 	cpuc = this_cpu_ptr(&cpu_hw_events);
1634 
1635 	/*
1636 	 * Some chipsets need to unmask the LVTPC in a particular spot
1637 	 * inside the nmi handler.  As a result, the unmasking was pushed
1638 	 * into all the nmi handlers.
1639 	 *
1640 	 * This generic handler doesn't seem to have any issues where the
1641 	 * unmasking occurs so it was left at the top.
1642 	 */
1643 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1644 
1645 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1646 		if (!test_bit(idx, cpuc->active_mask))
1647 			continue;
1648 
1649 		event = cpuc->events[idx];
1650 
1651 		val = x86_perf_event_update(event);
1652 		if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1653 			continue;
1654 
1655 		/*
1656 		 * event overflow
1657 		 */
1658 		handled++;
1659 		perf_sample_data_init(&data, 0, event->hw.last_period);
1660 
1661 		if (!x86_perf_event_set_period(event))
1662 			continue;
1663 
1664 		if (perf_event_overflow(event, &data, regs))
1665 			x86_pmu_stop(event, 0);
1666 	}
1667 
1668 	if (handled)
1669 		inc_irq_stat(apic_perf_irqs);
1670 
1671 	return handled;
1672 }
1673 
1674 void perf_events_lapic_init(void)
1675 {
1676 	if (!x86_pmu.apic || !x86_pmu_initialized())
1677 		return;
1678 
1679 	/*
1680 	 * Always use NMI for PMU
1681 	 */
1682 	apic_write(APIC_LVTPC, APIC_DM_NMI);
1683 }
1684 
1685 static int
1686 perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1687 {
1688 	u64 start_clock;
1689 	u64 finish_clock;
1690 	int ret;
1691 
1692 	/*
1693 	 * All PMUs/events that share this PMI handler should make sure to
1694 	 * increment active_events for their events.
1695 	 */
1696 	if (!atomic_read(&active_events))
1697 		return NMI_DONE;
1698 
1699 	start_clock = sched_clock();
1700 	ret = static_call(x86_pmu_handle_irq)(regs);
1701 	finish_clock = sched_clock();
1702 
1703 	perf_sample_event_took(finish_clock - start_clock);
1704 
1705 	return ret;
1706 }
1707 NOKPROBE_SYMBOL(perf_event_nmi_handler);
1708 
1709 struct event_constraint emptyconstraint;
1710 struct event_constraint unconstrained;
1711 
1712 static int x86_pmu_prepare_cpu(unsigned int cpu)
1713 {
1714 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1715 	int i;
1716 
1717 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++)
1718 		cpuc->kfree_on_online[i] = NULL;
1719 	if (x86_pmu.cpu_prepare)
1720 		return x86_pmu.cpu_prepare(cpu);
1721 	return 0;
1722 }
1723 
1724 static int x86_pmu_dead_cpu(unsigned int cpu)
1725 {
1726 	if (x86_pmu.cpu_dead)
1727 		x86_pmu.cpu_dead(cpu);
1728 	return 0;
1729 }
1730 
1731 static int x86_pmu_online_cpu(unsigned int cpu)
1732 {
1733 	struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1734 	int i;
1735 
1736 	for (i = 0 ; i < X86_PERF_KFREE_MAX; i++) {
1737 		kfree(cpuc->kfree_on_online[i]);
1738 		cpuc->kfree_on_online[i] = NULL;
1739 	}
1740 	return 0;
1741 }
1742 
1743 static int x86_pmu_starting_cpu(unsigned int cpu)
1744 {
1745 	if (x86_pmu.cpu_starting)
1746 		x86_pmu.cpu_starting(cpu);
1747 	return 0;
1748 }
1749 
1750 static int x86_pmu_dying_cpu(unsigned int cpu)
1751 {
1752 	if (x86_pmu.cpu_dying)
1753 		x86_pmu.cpu_dying(cpu);
1754 	return 0;
1755 }
1756 
1757 static void __init pmu_check_apic(void)
1758 {
1759 	if (boot_cpu_has(X86_FEATURE_APIC))
1760 		return;
1761 
1762 	x86_pmu.apic = 0;
1763 	pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1764 	pr_info("no hardware sampling interrupt available.\n");
1765 
1766 	/*
1767 	 * If we have a PMU initialized but no APIC
1768 	 * interrupts, we cannot sample hardware
1769 	 * events (user-space has to fall back and
1770 	 * sample via a hrtimer based software event):
1771 	 */
1772 	pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
1773 
1774 }
1775 
1776 static struct attribute_group x86_pmu_format_group __ro_after_init = {
1777 	.name = "format",
1778 	.attrs = NULL,
1779 };
1780 
1781 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
1782 {
1783 	struct perf_pmu_events_attr *pmu_attr =
1784 		container_of(attr, struct perf_pmu_events_attr, attr);
1785 	u64 config = 0;
1786 
1787 	if (pmu_attr->id < x86_pmu.max_events)
1788 		config = x86_pmu.event_map(pmu_attr->id);
1789 
1790 	/* string trumps id */
1791 	if (pmu_attr->event_str)
1792 		return sprintf(page, "%s", pmu_attr->event_str);
1793 
1794 	return x86_pmu.events_sysfs_show(page, config);
1795 }
1796 EXPORT_SYMBOL_GPL(events_sysfs_show);
1797 
1798 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1799 			  char *page)
1800 {
1801 	struct perf_pmu_events_ht_attr *pmu_attr =
1802 		container_of(attr, struct perf_pmu_events_ht_attr, attr);
1803 
1804 	/*
1805 	 * Report conditional events depending on Hyper-Threading.
1806 	 *
1807 	 * This is overly conservative as usually the HT special
1808 	 * handling is not needed if the other CPU thread is idle.
1809 	 *
1810 	 * Note this does not (and cannot) handle the case when thread
1811 	 * siblings are invisible, for example with virtualization
1812 	 * if they are owned by some other guest.  The user tool
1813 	 * has to re-read when a thread sibling gets onlined later.
1814 	 */
1815 	return sprintf(page, "%s",
1816 			topology_max_smt_threads() > 1 ?
1817 			pmu_attr->event_str_ht :
1818 			pmu_attr->event_str_noht);
1819 }
1820 
1821 EVENT_ATTR(cpu-cycles,			CPU_CYCLES		);
1822 EVENT_ATTR(instructions,		INSTRUCTIONS		);
1823 EVENT_ATTR(cache-references,		CACHE_REFERENCES	);
1824 EVENT_ATTR(cache-misses, 		CACHE_MISSES		);
1825 EVENT_ATTR(branch-instructions,		BRANCH_INSTRUCTIONS	);
1826 EVENT_ATTR(branch-misses,		BRANCH_MISSES		);
1827 EVENT_ATTR(bus-cycles,			BUS_CYCLES		);
1828 EVENT_ATTR(stalled-cycles-frontend,	STALLED_CYCLES_FRONTEND	);
1829 EVENT_ATTR(stalled-cycles-backend,	STALLED_CYCLES_BACKEND	);
1830 EVENT_ATTR(ref-cycles,			REF_CPU_CYCLES		);
1831 
1832 static struct attribute *empty_attrs;
1833 
1834 static struct attribute *events_attr[] = {
1835 	EVENT_PTR(CPU_CYCLES),
1836 	EVENT_PTR(INSTRUCTIONS),
1837 	EVENT_PTR(CACHE_REFERENCES),
1838 	EVENT_PTR(CACHE_MISSES),
1839 	EVENT_PTR(BRANCH_INSTRUCTIONS),
1840 	EVENT_PTR(BRANCH_MISSES),
1841 	EVENT_PTR(BUS_CYCLES),
1842 	EVENT_PTR(STALLED_CYCLES_FRONTEND),
1843 	EVENT_PTR(STALLED_CYCLES_BACKEND),
1844 	EVENT_PTR(REF_CPU_CYCLES),
1845 	NULL,
1846 };
1847 
1848 /*
1849  * Remove all undefined events (x86_pmu.event_map(id) == 0)
1850  * out of events_attr attributes.
1851  */
1852 static umode_t
1853 is_visible(struct kobject *kobj, struct attribute *attr, int idx)
1854 {
1855 	struct perf_pmu_events_attr *pmu_attr;
1856 
1857 	if (idx >= x86_pmu.max_events)
1858 		return 0;
1859 
1860 	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
1861 	/* str trumps id */
1862 	return pmu_attr->event_str || x86_pmu.event_map(idx) ? attr->mode : 0;
1863 }
1864 
1865 static struct attribute_group x86_pmu_events_group __ro_after_init = {
1866 	.name = "events",
1867 	.attrs = events_attr,
1868 	.is_visible = is_visible,
1869 };
1870 
1871 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1872 {
1873 	u64 umask  = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1874 	u64 cmask  = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1875 	bool edge  = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1876 	bool pc    = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1877 	bool any   = (config & ARCH_PERFMON_EVENTSEL_ANY);
1878 	bool inv   = (config & ARCH_PERFMON_EVENTSEL_INV);
1879 	ssize_t ret;
1880 
1881 	/*
1882 	* We have whole page size to spend and just little data
1883 	* to write, so we can safely use sprintf.
1884 	*/
1885 	ret = sprintf(page, "event=0x%02llx", event);
1886 
1887 	if (umask)
1888 		ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1889 
1890 	if (edge)
1891 		ret += sprintf(page + ret, ",edge");
1892 
1893 	if (pc)
1894 		ret += sprintf(page + ret, ",pc");
1895 
1896 	if (any)
1897 		ret += sprintf(page + ret, ",any");
1898 
1899 	if (inv)
1900 		ret += sprintf(page + ret, ",inv");
1901 
1902 	if (cmask)
1903 		ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1904 
1905 	ret += sprintf(page + ret, "\n");
1906 
1907 	return ret;
1908 }
1909 
1910 static struct attribute_group x86_pmu_attr_group;
1911 static struct attribute_group x86_pmu_caps_group;
1912 
1913 static void x86_pmu_static_call_update(void)
1914 {
1915 	static_call_update(x86_pmu_handle_irq, x86_pmu.handle_irq);
1916 	static_call_update(x86_pmu_disable_all, x86_pmu.disable_all);
1917 	static_call_update(x86_pmu_enable_all, x86_pmu.enable_all);
1918 	static_call_update(x86_pmu_enable, x86_pmu.enable);
1919 	static_call_update(x86_pmu_disable, x86_pmu.disable);
1920 
1921 	static_call_update(x86_pmu_add, x86_pmu.add);
1922 	static_call_update(x86_pmu_del, x86_pmu.del);
1923 	static_call_update(x86_pmu_read, x86_pmu.read);
1924 
1925 	static_call_update(x86_pmu_schedule_events, x86_pmu.schedule_events);
1926 	static_call_update(x86_pmu_get_event_constraints, x86_pmu.get_event_constraints);
1927 	static_call_update(x86_pmu_put_event_constraints, x86_pmu.put_event_constraints);
1928 
1929 	static_call_update(x86_pmu_start_scheduling, x86_pmu.start_scheduling);
1930 	static_call_update(x86_pmu_commit_scheduling, x86_pmu.commit_scheduling);
1931 	static_call_update(x86_pmu_stop_scheduling, x86_pmu.stop_scheduling);
1932 
1933 	static_call_update(x86_pmu_sched_task, x86_pmu.sched_task);
1934 	static_call_update(x86_pmu_swap_task_ctx, x86_pmu.swap_task_ctx);
1935 
1936 	static_call_update(x86_pmu_drain_pebs, x86_pmu.drain_pebs);
1937 	static_call_update(x86_pmu_pebs_aliases, x86_pmu.pebs_aliases);
1938 
1939 	static_call_update(x86_pmu_guest_get_msrs, x86_pmu.guest_get_msrs);
1940 }
1941 
1942 static void _x86_pmu_read(struct perf_event *event)
1943 {
1944 	x86_perf_event_update(event);
1945 }
1946 
1947 static inline struct perf_guest_switch_msr *
1948 perf_guest_get_msrs_nop(int *nr)
1949 {
1950 	*nr = 0;
1951 	return NULL;
1952 }
1953 
1954 static int __init init_hw_perf_events(void)
1955 {
1956 	struct x86_pmu_quirk *quirk;
1957 	int err;
1958 
1959 	pr_info("Performance Events: ");
1960 
1961 	switch (boot_cpu_data.x86_vendor) {
1962 	case X86_VENDOR_INTEL:
1963 		err = intel_pmu_init();
1964 		break;
1965 	case X86_VENDOR_AMD:
1966 		err = amd_pmu_init();
1967 		break;
1968 	case X86_VENDOR_HYGON:
1969 		err = amd_pmu_init();
1970 		x86_pmu.name = "HYGON";
1971 		break;
1972 	case X86_VENDOR_ZHAOXIN:
1973 	case X86_VENDOR_CENTAUR:
1974 		err = zhaoxin_pmu_init();
1975 		break;
1976 	default:
1977 		err = -ENOTSUPP;
1978 	}
1979 	if (err != 0) {
1980 		pr_cont("no PMU driver, software events only.\n");
1981 		return 0;
1982 	}
1983 
1984 	pmu_check_apic();
1985 
1986 	/* sanity check that the hardware exists or is emulated */
1987 	if (!check_hw_exists())
1988 		return 0;
1989 
1990 	pr_cont("%s PMU driver.\n", x86_pmu.name);
1991 
1992 	x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1993 
1994 	for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1995 		quirk->func();
1996 
1997 	if (!x86_pmu.intel_ctrl)
1998 		x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1999 
2000 	perf_events_lapic_init();
2001 	register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
2002 
2003 	unconstrained = (struct event_constraint)
2004 		__EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
2005 				   0, x86_pmu.num_counters, 0, 0);
2006 
2007 	x86_pmu_format_group.attrs = x86_pmu.format_attrs;
2008 
2009 	if (!x86_pmu.events_sysfs_show)
2010 		x86_pmu_events_group.attrs = &empty_attrs;
2011 
2012 	pmu.attr_update = x86_pmu.attr_update;
2013 
2014 	pr_info("... version:                %d\n",     x86_pmu.version);
2015 	pr_info("... bit width:              %d\n",     x86_pmu.cntval_bits);
2016 	pr_info("... generic registers:      %d\n",     x86_pmu.num_counters);
2017 	pr_info("... value mask:             %016Lx\n", x86_pmu.cntval_mask);
2018 	pr_info("... max period:             %016Lx\n", x86_pmu.max_period);
2019 	pr_info("... fixed-purpose events:   %lu\n",
2020 			hweight64((((1ULL << x86_pmu.num_counters_fixed) - 1)
2021 					<< INTEL_PMC_IDX_FIXED) & x86_pmu.intel_ctrl));
2022 	pr_info("... event mask:             %016Lx\n", x86_pmu.intel_ctrl);
2023 
2024 	if (!x86_pmu.read)
2025 		x86_pmu.read = _x86_pmu_read;
2026 
2027 	if (!x86_pmu.guest_get_msrs)
2028 		x86_pmu.guest_get_msrs = perf_guest_get_msrs_nop;
2029 
2030 	x86_pmu_static_call_update();
2031 
2032 	/*
2033 	 * Install callbacks. Core will call them for each online
2034 	 * cpu.
2035 	 */
2036 	err = cpuhp_setup_state(CPUHP_PERF_X86_PREPARE, "perf/x86:prepare",
2037 				x86_pmu_prepare_cpu, x86_pmu_dead_cpu);
2038 	if (err)
2039 		return err;
2040 
2041 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_STARTING,
2042 				"perf/x86:starting", x86_pmu_starting_cpu,
2043 				x86_pmu_dying_cpu);
2044 	if (err)
2045 		goto out;
2046 
2047 	err = cpuhp_setup_state(CPUHP_AP_PERF_X86_ONLINE, "perf/x86:online",
2048 				x86_pmu_online_cpu, NULL);
2049 	if (err)
2050 		goto out1;
2051 
2052 	err = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
2053 	if (err)
2054 		goto out2;
2055 
2056 	return 0;
2057 
2058 out2:
2059 	cpuhp_remove_state(CPUHP_AP_PERF_X86_ONLINE);
2060 out1:
2061 	cpuhp_remove_state(CPUHP_AP_PERF_X86_STARTING);
2062 out:
2063 	cpuhp_remove_state(CPUHP_PERF_X86_PREPARE);
2064 	return err;
2065 }
2066 early_initcall(init_hw_perf_events);
2067 
2068 static void x86_pmu_read(struct perf_event *event)
2069 {
2070 	static_call(x86_pmu_read)(event);
2071 }
2072 
2073 /*
2074  * Start group events scheduling transaction
2075  * Set the flag to make pmu::enable() not perform the
2076  * schedulability test, it will be performed at commit time
2077  *
2078  * We only support PERF_PMU_TXN_ADD transactions. Save the
2079  * transaction flags but otherwise ignore non-PERF_PMU_TXN_ADD
2080  * transactions.
2081  */
2082 static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
2083 {
2084 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2085 
2086 	WARN_ON_ONCE(cpuc->txn_flags);		/* txn already in flight */
2087 
2088 	cpuc->txn_flags = txn_flags;
2089 	if (txn_flags & ~PERF_PMU_TXN_ADD)
2090 		return;
2091 
2092 	perf_pmu_disable(pmu);
2093 	__this_cpu_write(cpu_hw_events.n_txn, 0);
2094 	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
2095 	__this_cpu_write(cpu_hw_events.n_txn_metric, 0);
2096 }
2097 
2098 /*
2099  * Stop group events scheduling transaction
2100  * Clear the flag and pmu::enable() will perform the
2101  * schedulability test.
2102  */
2103 static void x86_pmu_cancel_txn(struct pmu *pmu)
2104 {
2105 	unsigned int txn_flags;
2106 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2107 
2108 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
2109 
2110 	txn_flags = cpuc->txn_flags;
2111 	cpuc->txn_flags = 0;
2112 	if (txn_flags & ~PERF_PMU_TXN_ADD)
2113 		return;
2114 
2115 	/*
2116 	 * Truncate collected array by the number of events added in this
2117 	 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
2118 	 */
2119 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
2120 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
2121 	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
2122 	__this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
2123 	perf_pmu_enable(pmu);
2124 }
2125 
2126 /*
2127  * Commit group events scheduling transaction
2128  * Perform the group schedulability test as a whole
2129  * Return 0 if success
2130  *
2131  * Does not cancel the transaction on failure; expects the caller to do this.
2132  */
2133 static int x86_pmu_commit_txn(struct pmu *pmu)
2134 {
2135 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2136 	int assign[X86_PMC_IDX_MAX];
2137 	int n, ret;
2138 
2139 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
2140 
2141 	if (cpuc->txn_flags & ~PERF_PMU_TXN_ADD) {
2142 		cpuc->txn_flags = 0;
2143 		return 0;
2144 	}
2145 
2146 	n = cpuc->n_events;
2147 
2148 	if (!x86_pmu_initialized())
2149 		return -EAGAIN;
2150 
2151 	ret = static_call(x86_pmu_schedule_events)(cpuc, n, assign);
2152 	if (ret)
2153 		return ret;
2154 
2155 	/*
2156 	 * copy new assignment, now we know it is possible
2157 	 * will be used by hw_perf_enable()
2158 	 */
2159 	memcpy(cpuc->assign, assign, n*sizeof(int));
2160 
2161 	cpuc->txn_flags = 0;
2162 	perf_pmu_enable(pmu);
2163 	return 0;
2164 }
2165 /*
2166  * a fake_cpuc is used to validate event groups. Due to
2167  * the extra reg logic, we need to also allocate a fake
2168  * per_core and per_cpu structure. Otherwise, group events
2169  * using extra reg may conflict without the kernel being
2170  * able to catch this when the last event gets added to
2171  * the group.
2172  */
2173 static void free_fake_cpuc(struct cpu_hw_events *cpuc)
2174 {
2175 	intel_cpuc_finish(cpuc);
2176 	kfree(cpuc);
2177 }
2178 
2179 static struct cpu_hw_events *allocate_fake_cpuc(void)
2180 {
2181 	struct cpu_hw_events *cpuc;
2182 	int cpu = raw_smp_processor_id();
2183 
2184 	cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
2185 	if (!cpuc)
2186 		return ERR_PTR(-ENOMEM);
2187 	cpuc->is_fake = 1;
2188 
2189 	if (intel_cpuc_prepare(cpuc, cpu))
2190 		goto error;
2191 
2192 	return cpuc;
2193 error:
2194 	free_fake_cpuc(cpuc);
2195 	return ERR_PTR(-ENOMEM);
2196 }
2197 
2198 /*
2199  * validate that we can schedule this event
2200  */
2201 static int validate_event(struct perf_event *event)
2202 {
2203 	struct cpu_hw_events *fake_cpuc;
2204 	struct event_constraint *c;
2205 	int ret = 0;
2206 
2207 	fake_cpuc = allocate_fake_cpuc();
2208 	if (IS_ERR(fake_cpuc))
2209 		return PTR_ERR(fake_cpuc);
2210 
2211 	c = x86_pmu.get_event_constraints(fake_cpuc, 0, event);
2212 
2213 	if (!c || !c->weight)
2214 		ret = -EINVAL;
2215 
2216 	if (x86_pmu.put_event_constraints)
2217 		x86_pmu.put_event_constraints(fake_cpuc, event);
2218 
2219 	free_fake_cpuc(fake_cpuc);
2220 
2221 	return ret;
2222 }
2223 
2224 /*
2225  * validate a single event group
2226  *
2227  * validation include:
2228  *	- check events are compatible which each other
2229  *	- events do not compete for the same counter
2230  *	- number of events <= number of counters
2231  *
2232  * validation ensures the group can be loaded onto the
2233  * PMU if it was the only group available.
2234  */
2235 static int validate_group(struct perf_event *event)
2236 {
2237 	struct perf_event *leader = event->group_leader;
2238 	struct cpu_hw_events *fake_cpuc;
2239 	int ret = -EINVAL, n;
2240 
2241 	fake_cpuc = allocate_fake_cpuc();
2242 	if (IS_ERR(fake_cpuc))
2243 		return PTR_ERR(fake_cpuc);
2244 	/*
2245 	 * the event is not yet connected with its
2246 	 * siblings therefore we must first collect
2247 	 * existing siblings, then add the new event
2248 	 * before we can simulate the scheduling
2249 	 */
2250 	n = collect_events(fake_cpuc, leader, true);
2251 	if (n < 0)
2252 		goto out;
2253 
2254 	fake_cpuc->n_events = n;
2255 	n = collect_events(fake_cpuc, event, false);
2256 	if (n < 0)
2257 		goto out;
2258 
2259 	fake_cpuc->n_events = 0;
2260 	ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
2261 
2262 out:
2263 	free_fake_cpuc(fake_cpuc);
2264 	return ret;
2265 }
2266 
2267 static int x86_pmu_event_init(struct perf_event *event)
2268 {
2269 	struct pmu *tmp;
2270 	int err;
2271 
2272 	switch (event->attr.type) {
2273 	case PERF_TYPE_RAW:
2274 	case PERF_TYPE_HARDWARE:
2275 	case PERF_TYPE_HW_CACHE:
2276 		break;
2277 
2278 	default:
2279 		return -ENOENT;
2280 	}
2281 
2282 	err = __x86_pmu_event_init(event);
2283 	if (!err) {
2284 		/*
2285 		 * we temporarily connect event to its pmu
2286 		 * such that validate_group() can classify
2287 		 * it as an x86 event using is_x86_event()
2288 		 */
2289 		tmp = event->pmu;
2290 		event->pmu = &pmu;
2291 
2292 		if (event->group_leader != event)
2293 			err = validate_group(event);
2294 		else
2295 			err = validate_event(event);
2296 
2297 		event->pmu = tmp;
2298 	}
2299 	if (err) {
2300 		if (event->destroy)
2301 			event->destroy(event);
2302 	}
2303 
2304 	if (READ_ONCE(x86_pmu.attr_rdpmc) &&
2305 	    !(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
2306 		event->hw.flags |= PERF_X86_EVENT_RDPMC_ALLOWED;
2307 
2308 	return err;
2309 }
2310 
2311 static void x86_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
2312 {
2313 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2314 		return;
2315 
2316 	/*
2317 	 * This function relies on not being called concurrently in two
2318 	 * tasks in the same mm.  Otherwise one task could observe
2319 	 * perf_rdpmc_allowed > 1 and return all the way back to
2320 	 * userspace with CR4.PCE clear while another task is still
2321 	 * doing on_each_cpu_mask() to propagate CR4.PCE.
2322 	 *
2323 	 * For now, this can't happen because all callers hold mmap_lock
2324 	 * for write.  If this changes, we'll need a different solution.
2325 	 */
2326 	mmap_assert_write_locked(mm);
2327 
2328 	if (atomic_inc_return(&mm->context.perf_rdpmc_allowed) == 1)
2329 		on_each_cpu_mask(mm_cpumask(mm), cr4_update_pce, NULL, 1);
2330 }
2331 
2332 static void x86_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
2333 {
2334 
2335 	if (!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2336 		return;
2337 
2338 	if (atomic_dec_and_test(&mm->context.perf_rdpmc_allowed))
2339 		on_each_cpu_mask(mm_cpumask(mm), cr4_update_pce, NULL, 1);
2340 }
2341 
2342 static int x86_pmu_event_idx(struct perf_event *event)
2343 {
2344 	struct hw_perf_event *hwc = &event->hw;
2345 
2346 	if (!(hwc->flags & PERF_X86_EVENT_RDPMC_ALLOWED))
2347 		return 0;
2348 
2349 	if (is_metric_idx(hwc->idx))
2350 		return INTEL_PMC_FIXED_RDPMC_METRICS + 1;
2351 	else
2352 		return hwc->event_base_rdpmc + 1;
2353 }
2354 
2355 static ssize_t get_attr_rdpmc(struct device *cdev,
2356 			      struct device_attribute *attr,
2357 			      char *buf)
2358 {
2359 	return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
2360 }
2361 
2362 static ssize_t set_attr_rdpmc(struct device *cdev,
2363 			      struct device_attribute *attr,
2364 			      const char *buf, size_t count)
2365 {
2366 	unsigned long val;
2367 	ssize_t ret;
2368 
2369 	ret = kstrtoul(buf, 0, &val);
2370 	if (ret)
2371 		return ret;
2372 
2373 	if (val > 2)
2374 		return -EINVAL;
2375 
2376 	if (x86_pmu.attr_rdpmc_broken)
2377 		return -ENOTSUPP;
2378 
2379 	if (val != x86_pmu.attr_rdpmc) {
2380 		/*
2381 		 * Changing into or out of never available or always available,
2382 		 * aka perf-event-bypassing mode. This path is extremely slow,
2383 		 * but only root can trigger it, so it's okay.
2384 		 */
2385 		if (val == 0)
2386 			static_branch_inc(&rdpmc_never_available_key);
2387 		else if (x86_pmu.attr_rdpmc == 0)
2388 			static_branch_dec(&rdpmc_never_available_key);
2389 
2390 		if (val == 2)
2391 			static_branch_inc(&rdpmc_always_available_key);
2392 		else if (x86_pmu.attr_rdpmc == 2)
2393 			static_branch_dec(&rdpmc_always_available_key);
2394 
2395 		on_each_cpu(cr4_update_pce, NULL, 1);
2396 		x86_pmu.attr_rdpmc = val;
2397 	}
2398 
2399 	return count;
2400 }
2401 
2402 static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
2403 
2404 static struct attribute *x86_pmu_attrs[] = {
2405 	&dev_attr_rdpmc.attr,
2406 	NULL,
2407 };
2408 
2409 static struct attribute_group x86_pmu_attr_group __ro_after_init = {
2410 	.attrs = x86_pmu_attrs,
2411 };
2412 
2413 static ssize_t max_precise_show(struct device *cdev,
2414 				  struct device_attribute *attr,
2415 				  char *buf)
2416 {
2417 	return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu_max_precise());
2418 }
2419 
2420 static DEVICE_ATTR_RO(max_precise);
2421 
2422 static struct attribute *x86_pmu_caps_attrs[] = {
2423 	&dev_attr_max_precise.attr,
2424 	NULL
2425 };
2426 
2427 static struct attribute_group x86_pmu_caps_group __ro_after_init = {
2428 	.name = "caps",
2429 	.attrs = x86_pmu_caps_attrs,
2430 };
2431 
2432 static const struct attribute_group *x86_pmu_attr_groups[] = {
2433 	&x86_pmu_attr_group,
2434 	&x86_pmu_format_group,
2435 	&x86_pmu_events_group,
2436 	&x86_pmu_caps_group,
2437 	NULL,
2438 };
2439 
2440 static void x86_pmu_sched_task(struct perf_event_context *ctx, bool sched_in)
2441 {
2442 	static_call_cond(x86_pmu_sched_task)(ctx, sched_in);
2443 }
2444 
2445 static void x86_pmu_swap_task_ctx(struct perf_event_context *prev,
2446 				  struct perf_event_context *next)
2447 {
2448 	static_call_cond(x86_pmu_swap_task_ctx)(prev, next);
2449 }
2450 
2451 void perf_check_microcode(void)
2452 {
2453 	if (x86_pmu.check_microcode)
2454 		x86_pmu.check_microcode();
2455 }
2456 
2457 static int x86_pmu_check_period(struct perf_event *event, u64 value)
2458 {
2459 	if (x86_pmu.check_period && x86_pmu.check_period(event, value))
2460 		return -EINVAL;
2461 
2462 	if (value && x86_pmu.limit_period) {
2463 		if (x86_pmu.limit_period(event, value) > value)
2464 			return -EINVAL;
2465 	}
2466 
2467 	return 0;
2468 }
2469 
2470 static int x86_pmu_aux_output_match(struct perf_event *event)
2471 {
2472 	if (!(pmu.capabilities & PERF_PMU_CAP_AUX_OUTPUT))
2473 		return 0;
2474 
2475 	if (x86_pmu.aux_output_match)
2476 		return x86_pmu.aux_output_match(event);
2477 
2478 	return 0;
2479 }
2480 
2481 static struct pmu pmu = {
2482 	.pmu_enable		= x86_pmu_enable,
2483 	.pmu_disable		= x86_pmu_disable,
2484 
2485 	.attr_groups		= x86_pmu_attr_groups,
2486 
2487 	.event_init		= x86_pmu_event_init,
2488 
2489 	.event_mapped		= x86_pmu_event_mapped,
2490 	.event_unmapped		= x86_pmu_event_unmapped,
2491 
2492 	.add			= x86_pmu_add,
2493 	.del			= x86_pmu_del,
2494 	.start			= x86_pmu_start,
2495 	.stop			= x86_pmu_stop,
2496 	.read			= x86_pmu_read,
2497 
2498 	.start_txn		= x86_pmu_start_txn,
2499 	.cancel_txn		= x86_pmu_cancel_txn,
2500 	.commit_txn		= x86_pmu_commit_txn,
2501 
2502 	.event_idx		= x86_pmu_event_idx,
2503 	.sched_task		= x86_pmu_sched_task,
2504 	.swap_task_ctx		= x86_pmu_swap_task_ctx,
2505 	.check_period		= x86_pmu_check_period,
2506 
2507 	.aux_output_match	= x86_pmu_aux_output_match,
2508 };
2509 
2510 void arch_perf_update_userpage(struct perf_event *event,
2511 			       struct perf_event_mmap_page *userpg, u64 now)
2512 {
2513 	struct cyc2ns_data data;
2514 	u64 offset;
2515 
2516 	userpg->cap_user_time = 0;
2517 	userpg->cap_user_time_zero = 0;
2518 	userpg->cap_user_rdpmc =
2519 		!!(event->hw.flags & PERF_X86_EVENT_RDPMC_ALLOWED);
2520 	userpg->pmc_width = x86_pmu.cntval_bits;
2521 
2522 	if (!using_native_sched_clock() || !sched_clock_stable())
2523 		return;
2524 
2525 	cyc2ns_read_begin(&data);
2526 
2527 	offset = data.cyc2ns_offset + __sched_clock_offset;
2528 
2529 	/*
2530 	 * Internal timekeeping for enabled/running/stopped times
2531 	 * is always in the local_clock domain.
2532 	 */
2533 	userpg->cap_user_time = 1;
2534 	userpg->time_mult = data.cyc2ns_mul;
2535 	userpg->time_shift = data.cyc2ns_shift;
2536 	userpg->time_offset = offset - now;
2537 
2538 	/*
2539 	 * cap_user_time_zero doesn't make sense when we're using a different
2540 	 * time base for the records.
2541 	 */
2542 	if (!event->attr.use_clockid) {
2543 		userpg->cap_user_time_zero = 1;
2544 		userpg->time_zero = offset;
2545 	}
2546 
2547 	cyc2ns_read_end();
2548 }
2549 
2550 /*
2551  * Determine whether the regs were taken from an irq/exception handler rather
2552  * than from perf_arch_fetch_caller_regs().
2553  */
2554 static bool perf_hw_regs(struct pt_regs *regs)
2555 {
2556 	return regs->flags & X86_EFLAGS_FIXED;
2557 }
2558 
2559 void
2560 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2561 {
2562 	struct unwind_state state;
2563 	unsigned long addr;
2564 
2565 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2566 		/* TODO: We don't support guest os callchain now */
2567 		return;
2568 	}
2569 
2570 	if (perf_callchain_store(entry, regs->ip))
2571 		return;
2572 
2573 	if (perf_hw_regs(regs))
2574 		unwind_start(&state, current, regs, NULL);
2575 	else
2576 		unwind_start(&state, current, NULL, (void *)regs->sp);
2577 
2578 	for (; !unwind_done(&state); unwind_next_frame(&state)) {
2579 		addr = unwind_get_return_address(&state);
2580 		if (!addr || perf_callchain_store(entry, addr))
2581 			return;
2582 	}
2583 }
2584 
2585 static inline int
2586 valid_user_frame(const void __user *fp, unsigned long size)
2587 {
2588 	return (__range_not_ok(fp, size, TASK_SIZE) == 0);
2589 }
2590 
2591 static unsigned long get_segment_base(unsigned int segment)
2592 {
2593 	struct desc_struct *desc;
2594 	unsigned int idx = segment >> 3;
2595 
2596 	if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
2597 #ifdef CONFIG_MODIFY_LDT_SYSCALL
2598 		struct ldt_struct *ldt;
2599 
2600 		/* IRQs are off, so this synchronizes with smp_store_release */
2601 		ldt = READ_ONCE(current->active_mm->context.ldt);
2602 		if (!ldt || idx >= ldt->nr_entries)
2603 			return 0;
2604 
2605 		desc = &ldt->entries[idx];
2606 #else
2607 		return 0;
2608 #endif
2609 	} else {
2610 		if (idx >= GDT_ENTRIES)
2611 			return 0;
2612 
2613 		desc = raw_cpu_ptr(gdt_page.gdt) + idx;
2614 	}
2615 
2616 	return get_desc_base(desc);
2617 }
2618 
2619 #ifdef CONFIG_IA32_EMULATION
2620 
2621 #include <linux/compat.h>
2622 
2623 static inline int
2624 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2625 {
2626 	/* 32-bit process in 64-bit kernel. */
2627 	unsigned long ss_base, cs_base;
2628 	struct stack_frame_ia32 frame;
2629 	const struct stack_frame_ia32 __user *fp;
2630 
2631 	if (user_64bit_mode(regs))
2632 		return 0;
2633 
2634 	cs_base = get_segment_base(regs->cs);
2635 	ss_base = get_segment_base(regs->ss);
2636 
2637 	fp = compat_ptr(ss_base + regs->bp);
2638 	pagefault_disable();
2639 	while (entry->nr < entry->max_stack) {
2640 		if (!valid_user_frame(fp, sizeof(frame)))
2641 			break;
2642 
2643 		if (__get_user(frame.next_frame, &fp->next_frame))
2644 			break;
2645 		if (__get_user(frame.return_address, &fp->return_address))
2646 			break;
2647 
2648 		perf_callchain_store(entry, cs_base + frame.return_address);
2649 		fp = compat_ptr(ss_base + frame.next_frame);
2650 	}
2651 	pagefault_enable();
2652 	return 1;
2653 }
2654 #else
2655 static inline int
2656 perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry_ctx *entry)
2657 {
2658     return 0;
2659 }
2660 #endif
2661 
2662 void
2663 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
2664 {
2665 	struct stack_frame frame;
2666 	const struct stack_frame __user *fp;
2667 
2668 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2669 		/* TODO: We don't support guest os callchain now */
2670 		return;
2671 	}
2672 
2673 	/*
2674 	 * We don't know what to do with VM86 stacks.. ignore them for now.
2675 	 */
2676 	if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2677 		return;
2678 
2679 	fp = (void __user *)regs->bp;
2680 
2681 	perf_callchain_store(entry, regs->ip);
2682 
2683 	if (!nmi_uaccess_okay())
2684 		return;
2685 
2686 	if (perf_callchain_user32(regs, entry))
2687 		return;
2688 
2689 	pagefault_disable();
2690 	while (entry->nr < entry->max_stack) {
2691 		if (!valid_user_frame(fp, sizeof(frame)))
2692 			break;
2693 
2694 		if (__get_user(frame.next_frame, &fp->next_frame))
2695 			break;
2696 		if (__get_user(frame.return_address, &fp->return_address))
2697 			break;
2698 
2699 		perf_callchain_store(entry, frame.return_address);
2700 		fp = (void __user *)frame.next_frame;
2701 	}
2702 	pagefault_enable();
2703 }
2704 
2705 /*
2706  * Deal with code segment offsets for the various execution modes:
2707  *
2708  *   VM86 - the good olde 16 bit days, where the linear address is
2709  *          20 bits and we use regs->ip + 0x10 * regs->cs.
2710  *
2711  *   IA32 - Where we need to look at GDT/LDT segment descriptor tables
2712  *          to figure out what the 32bit base address is.
2713  *
2714  *    X32 - has TIF_X32 set, but is running in x86_64
2715  *
2716  * X86_64 - CS,DS,SS,ES are all zero based.
2717  */
2718 static unsigned long code_segment_base(struct pt_regs *regs)
2719 {
2720 	/*
2721 	 * For IA32 we look at the GDT/LDT segment base to convert the
2722 	 * effective IP to a linear address.
2723 	 */
2724 
2725 #ifdef CONFIG_X86_32
2726 	/*
2727 	 * If we are in VM86 mode, add the segment offset to convert to a
2728 	 * linear address.
2729 	 */
2730 	if (regs->flags & X86_VM_MASK)
2731 		return 0x10 * regs->cs;
2732 
2733 	if (user_mode(regs) && regs->cs != __USER_CS)
2734 		return get_segment_base(regs->cs);
2735 #else
2736 	if (user_mode(regs) && !user_64bit_mode(regs) &&
2737 	    regs->cs != __USER32_CS)
2738 		return get_segment_base(regs->cs);
2739 #endif
2740 	return 0;
2741 }
2742 
2743 unsigned long perf_instruction_pointer(struct pt_regs *regs)
2744 {
2745 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2746 		return perf_guest_cbs->get_guest_ip();
2747 
2748 	return regs->ip + code_segment_base(regs);
2749 }
2750 
2751 unsigned long perf_misc_flags(struct pt_regs *regs)
2752 {
2753 	int misc = 0;
2754 
2755 	if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2756 		if (perf_guest_cbs->is_user_mode())
2757 			misc |= PERF_RECORD_MISC_GUEST_USER;
2758 		else
2759 			misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2760 	} else {
2761 		if (user_mode(regs))
2762 			misc |= PERF_RECORD_MISC_USER;
2763 		else
2764 			misc |= PERF_RECORD_MISC_KERNEL;
2765 	}
2766 
2767 	if (regs->flags & PERF_EFLAGS_EXACT)
2768 		misc |= PERF_RECORD_MISC_EXACT_IP;
2769 
2770 	return misc;
2771 }
2772 
2773 void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2774 {
2775 	cap->version		= x86_pmu.version;
2776 	cap->num_counters_gp	= x86_pmu.num_counters;
2777 	cap->num_counters_fixed	= x86_pmu.num_counters_fixed;
2778 	cap->bit_width_gp	= x86_pmu.cntval_bits;
2779 	cap->bit_width_fixed	= x86_pmu.cntval_bits;
2780 	cap->events_mask	= (unsigned int)x86_pmu.events_maskl;
2781 	cap->events_mask_len	= x86_pmu.events_mask_len;
2782 }
2783 EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);
2784