Loading...
Note: File does not exist in v4.17.
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 <pzijlstr@redhat.com>
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/module.h>
21#include <linux/kdebug.h>
22#include <linux/sched.h>
23#include <linux/uaccess.h>
24#include <linux/slab.h>
25#include <linux/cpu.h>
26#include <linux/bitops.h>
27#include <linux/device.h>
28
29#include <asm/apic.h>
30#include <asm/stacktrace.h>
31#include <asm/nmi.h>
32#include <asm/smp.h>
33#include <asm/alternative.h>
34#include <asm/timer.h>
35#include <asm/desc.h>
36#include <asm/ldt.h>
37
38#include "perf_event.h"
39
40struct x86_pmu x86_pmu __read_mostly;
41
42DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events) = {
43 .enabled = 1,
44};
45
46u64 __read_mostly hw_cache_event_ids
47 [PERF_COUNT_HW_CACHE_MAX]
48 [PERF_COUNT_HW_CACHE_OP_MAX]
49 [PERF_COUNT_HW_CACHE_RESULT_MAX];
50u64 __read_mostly hw_cache_extra_regs
51 [PERF_COUNT_HW_CACHE_MAX]
52 [PERF_COUNT_HW_CACHE_OP_MAX]
53 [PERF_COUNT_HW_CACHE_RESULT_MAX];
54
55/*
56 * Propagate event elapsed time into the generic event.
57 * Can only be executed on the CPU where the event is active.
58 * Returns the delta events processed.
59 */
60u64 x86_perf_event_update(struct perf_event *event)
61{
62 struct hw_perf_event *hwc = &event->hw;
63 int shift = 64 - x86_pmu.cntval_bits;
64 u64 prev_raw_count, new_raw_count;
65 int idx = hwc->idx;
66 s64 delta;
67
68 if (idx == INTEL_PMC_IDX_FIXED_BTS)
69 return 0;
70
71 /*
72 * Careful: an NMI might modify the previous event value.
73 *
74 * Our tactic to handle this is to first atomically read and
75 * exchange a new raw count - then add that new-prev delta
76 * count to the generic event atomically:
77 */
78again:
79 prev_raw_count = local64_read(&hwc->prev_count);
80 rdpmcl(hwc->event_base_rdpmc, new_raw_count);
81
82 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
83 new_raw_count) != prev_raw_count)
84 goto again;
85
86 /*
87 * Now we have the new raw value and have updated the prev
88 * timestamp already. We can now calculate the elapsed delta
89 * (event-)time and add that to the generic event.
90 *
91 * Careful, not all hw sign-extends above the physical width
92 * of the count.
93 */
94 delta = (new_raw_count << shift) - (prev_raw_count << shift);
95 delta >>= shift;
96
97 local64_add(delta, &event->count);
98 local64_sub(delta, &hwc->period_left);
99
100 return new_raw_count;
101}
102
103/*
104 * Find and validate any extra registers to set up.
105 */
106static int x86_pmu_extra_regs(u64 config, struct perf_event *event)
107{
108 struct hw_perf_event_extra *reg;
109 struct extra_reg *er;
110
111 reg = &event->hw.extra_reg;
112
113 if (!x86_pmu.extra_regs)
114 return 0;
115
116 for (er = x86_pmu.extra_regs; er->msr; er++) {
117 if (er->event != (config & er->config_mask))
118 continue;
119 if (event->attr.config1 & ~er->valid_mask)
120 return -EINVAL;
121
122 reg->idx = er->idx;
123 reg->config = event->attr.config1;
124 reg->reg = er->msr;
125 break;
126 }
127 return 0;
128}
129
130static atomic_t active_events;
131static DEFINE_MUTEX(pmc_reserve_mutex);
132
133#ifdef CONFIG_X86_LOCAL_APIC
134
135static bool reserve_pmc_hardware(void)
136{
137 int i;
138
139 for (i = 0; i < x86_pmu.num_counters; i++) {
140 if (!reserve_perfctr_nmi(x86_pmu_event_addr(i)))
141 goto perfctr_fail;
142 }
143
144 for (i = 0; i < x86_pmu.num_counters; i++) {
145 if (!reserve_evntsel_nmi(x86_pmu_config_addr(i)))
146 goto eventsel_fail;
147 }
148
149 return true;
150
151eventsel_fail:
152 for (i--; i >= 0; i--)
153 release_evntsel_nmi(x86_pmu_config_addr(i));
154
155 i = x86_pmu.num_counters;
156
157perfctr_fail:
158 for (i--; i >= 0; i--)
159 release_perfctr_nmi(x86_pmu_event_addr(i));
160
161 return false;
162}
163
164static void release_pmc_hardware(void)
165{
166 int i;
167
168 for (i = 0; i < x86_pmu.num_counters; i++) {
169 release_perfctr_nmi(x86_pmu_event_addr(i));
170 release_evntsel_nmi(x86_pmu_config_addr(i));
171 }
172}
173
174#else
175
176static bool reserve_pmc_hardware(void) { return true; }
177static void release_pmc_hardware(void) {}
178
179#endif
180
181static bool check_hw_exists(void)
182{
183 u64 val, val_fail, val_new= ~0;
184 int i, reg, reg_fail, ret = 0;
185 int bios_fail = 0;
186
187 /*
188 * Check to see if the BIOS enabled any of the counters, if so
189 * complain and bail.
190 */
191 for (i = 0; i < x86_pmu.num_counters; i++) {
192 reg = x86_pmu_config_addr(i);
193 ret = rdmsrl_safe(reg, &val);
194 if (ret)
195 goto msr_fail;
196 if (val & ARCH_PERFMON_EVENTSEL_ENABLE) {
197 bios_fail = 1;
198 val_fail = val;
199 reg_fail = reg;
200 }
201 }
202
203 if (x86_pmu.num_counters_fixed) {
204 reg = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
205 ret = rdmsrl_safe(reg, &val);
206 if (ret)
207 goto msr_fail;
208 for (i = 0; i < x86_pmu.num_counters_fixed; i++) {
209 if (val & (0x03 << i*4)) {
210 bios_fail = 1;
211 val_fail = val;
212 reg_fail = reg;
213 }
214 }
215 }
216
217 /*
218 * Read the current value, change it and read it back to see if it
219 * matches, this is needed to detect certain hardware emulators
220 * (qemu/kvm) that don't trap on the MSR access and always return 0s.
221 */
222 reg = x86_pmu_event_addr(0);
223 if (rdmsrl_safe(reg, &val))
224 goto msr_fail;
225 val ^= 0xffffUL;
226 ret = wrmsrl_safe(reg, val);
227 ret |= rdmsrl_safe(reg, &val_new);
228 if (ret || val != val_new)
229 goto msr_fail;
230
231 /*
232 * We still allow the PMU driver to operate:
233 */
234 if (bios_fail) {
235 printk(KERN_CONT "Broken BIOS detected, complain to your hardware vendor.\n");
236 printk(KERN_ERR FW_BUG "the BIOS has corrupted hw-PMU resources (MSR %x is %Lx)\n", reg_fail, val_fail);
237 }
238
239 return true;
240
241msr_fail:
242 printk(KERN_CONT "Broken PMU hardware detected, using software events only.\n");
243 printk(KERN_ERR "Failed to access perfctr msr (MSR %x is %Lx)\n", reg, val_new);
244
245 return false;
246}
247
248static void hw_perf_event_destroy(struct perf_event *event)
249{
250 if (atomic_dec_and_mutex_lock(&active_events, &pmc_reserve_mutex)) {
251 release_pmc_hardware();
252 release_ds_buffers();
253 mutex_unlock(&pmc_reserve_mutex);
254 }
255}
256
257static inline int x86_pmu_initialized(void)
258{
259 return x86_pmu.handle_irq != NULL;
260}
261
262static inline int
263set_ext_hw_attr(struct hw_perf_event *hwc, struct perf_event *event)
264{
265 struct perf_event_attr *attr = &event->attr;
266 unsigned int cache_type, cache_op, cache_result;
267 u64 config, val;
268
269 config = attr->config;
270
271 cache_type = (config >> 0) & 0xff;
272 if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
273 return -EINVAL;
274
275 cache_op = (config >> 8) & 0xff;
276 if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
277 return -EINVAL;
278
279 cache_result = (config >> 16) & 0xff;
280 if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
281 return -EINVAL;
282
283 val = hw_cache_event_ids[cache_type][cache_op][cache_result];
284
285 if (val == 0)
286 return -ENOENT;
287
288 if (val == -1)
289 return -EINVAL;
290
291 hwc->config |= val;
292 attr->config1 = hw_cache_extra_regs[cache_type][cache_op][cache_result];
293 return x86_pmu_extra_regs(val, event);
294}
295
296int x86_setup_perfctr(struct perf_event *event)
297{
298 struct perf_event_attr *attr = &event->attr;
299 struct hw_perf_event *hwc = &event->hw;
300 u64 config;
301
302 if (!is_sampling_event(event)) {
303 hwc->sample_period = x86_pmu.max_period;
304 hwc->last_period = hwc->sample_period;
305 local64_set(&hwc->period_left, hwc->sample_period);
306 } else {
307 /*
308 * If we have a PMU initialized but no APIC
309 * interrupts, we cannot sample hardware
310 * events (user-space has to fall back and
311 * sample via a hrtimer based software event):
312 */
313 if (!x86_pmu.apic)
314 return -EOPNOTSUPP;
315 }
316
317 if (attr->type == PERF_TYPE_RAW)
318 return x86_pmu_extra_regs(event->attr.config, event);
319
320 if (attr->type == PERF_TYPE_HW_CACHE)
321 return set_ext_hw_attr(hwc, event);
322
323 if (attr->config >= x86_pmu.max_events)
324 return -EINVAL;
325
326 /*
327 * The generic map:
328 */
329 config = x86_pmu.event_map(attr->config);
330
331 if (config == 0)
332 return -ENOENT;
333
334 if (config == -1LL)
335 return -EINVAL;
336
337 /*
338 * Branch tracing:
339 */
340 if (attr->config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS &&
341 !attr->freq && hwc->sample_period == 1) {
342 /* BTS is not supported by this architecture. */
343 if (!x86_pmu.bts_active)
344 return -EOPNOTSUPP;
345
346 /* BTS is currently only allowed for user-mode. */
347 if (!attr->exclude_kernel)
348 return -EOPNOTSUPP;
349 }
350
351 hwc->config |= config;
352
353 return 0;
354}
355
356/*
357 * check that branch_sample_type is compatible with
358 * settings needed for precise_ip > 1 which implies
359 * using the LBR to capture ALL taken branches at the
360 * priv levels of the measurement
361 */
362static inline int precise_br_compat(struct perf_event *event)
363{
364 u64 m = event->attr.branch_sample_type;
365 u64 b = 0;
366
367 /* must capture all branches */
368 if (!(m & PERF_SAMPLE_BRANCH_ANY))
369 return 0;
370
371 m &= PERF_SAMPLE_BRANCH_KERNEL | PERF_SAMPLE_BRANCH_USER;
372
373 if (!event->attr.exclude_user)
374 b |= PERF_SAMPLE_BRANCH_USER;
375
376 if (!event->attr.exclude_kernel)
377 b |= PERF_SAMPLE_BRANCH_KERNEL;
378
379 /*
380 * ignore PERF_SAMPLE_BRANCH_HV, not supported on x86
381 */
382
383 return m == b;
384}
385
386int x86_pmu_hw_config(struct perf_event *event)
387{
388 if (event->attr.precise_ip) {
389 int precise = 0;
390
391 /* Support for constant skid */
392 if (x86_pmu.pebs_active && !x86_pmu.pebs_broken) {
393 precise++;
394
395 /* Support for IP fixup */
396 if (x86_pmu.lbr_nr)
397 precise++;
398 }
399
400 if (event->attr.precise_ip > precise)
401 return -EOPNOTSUPP;
402 /*
403 * check that PEBS LBR correction does not conflict with
404 * whatever the user is asking with attr->branch_sample_type
405 */
406 if (event->attr.precise_ip > 1 &&
407 x86_pmu.intel_cap.pebs_format < 2) {
408 u64 *br_type = &event->attr.branch_sample_type;
409
410 if (has_branch_stack(event)) {
411 if (!precise_br_compat(event))
412 return -EOPNOTSUPP;
413
414 /* branch_sample_type is compatible */
415
416 } else {
417 /*
418 * user did not specify branch_sample_type
419 *
420 * For PEBS fixups, we capture all
421 * the branches at the priv level of the
422 * event.
423 */
424 *br_type = PERF_SAMPLE_BRANCH_ANY;
425
426 if (!event->attr.exclude_user)
427 *br_type |= PERF_SAMPLE_BRANCH_USER;
428
429 if (!event->attr.exclude_kernel)
430 *br_type |= PERF_SAMPLE_BRANCH_KERNEL;
431 }
432 }
433 }
434
435 /*
436 * Generate PMC IRQs:
437 * (keep 'enabled' bit clear for now)
438 */
439 event->hw.config = ARCH_PERFMON_EVENTSEL_INT;
440
441 /*
442 * Count user and OS events unless requested not to
443 */
444 if (!event->attr.exclude_user)
445 event->hw.config |= ARCH_PERFMON_EVENTSEL_USR;
446 if (!event->attr.exclude_kernel)
447 event->hw.config |= ARCH_PERFMON_EVENTSEL_OS;
448
449 if (event->attr.type == PERF_TYPE_RAW)
450 event->hw.config |= event->attr.config & X86_RAW_EVENT_MASK;
451
452 return x86_setup_perfctr(event);
453}
454
455/*
456 * Setup the hardware configuration for a given attr_type
457 */
458static int __x86_pmu_event_init(struct perf_event *event)
459{
460 int err;
461
462 if (!x86_pmu_initialized())
463 return -ENODEV;
464
465 err = 0;
466 if (!atomic_inc_not_zero(&active_events)) {
467 mutex_lock(&pmc_reserve_mutex);
468 if (atomic_read(&active_events) == 0) {
469 if (!reserve_pmc_hardware())
470 err = -EBUSY;
471 else
472 reserve_ds_buffers();
473 }
474 if (!err)
475 atomic_inc(&active_events);
476 mutex_unlock(&pmc_reserve_mutex);
477 }
478 if (err)
479 return err;
480
481 event->destroy = hw_perf_event_destroy;
482
483 event->hw.idx = -1;
484 event->hw.last_cpu = -1;
485 event->hw.last_tag = ~0ULL;
486
487 /* mark unused */
488 event->hw.extra_reg.idx = EXTRA_REG_NONE;
489 event->hw.branch_reg.idx = EXTRA_REG_NONE;
490
491 return x86_pmu.hw_config(event);
492}
493
494void x86_pmu_disable_all(void)
495{
496 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
497 int idx;
498
499 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
500 u64 val;
501
502 if (!test_bit(idx, cpuc->active_mask))
503 continue;
504 rdmsrl(x86_pmu_config_addr(idx), val);
505 if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
506 continue;
507 val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
508 wrmsrl(x86_pmu_config_addr(idx), val);
509 }
510}
511
512static void x86_pmu_disable(struct pmu *pmu)
513{
514 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
515
516 if (!x86_pmu_initialized())
517 return;
518
519 if (!cpuc->enabled)
520 return;
521
522 cpuc->n_added = 0;
523 cpuc->enabled = 0;
524 barrier();
525
526 x86_pmu.disable_all();
527}
528
529void x86_pmu_enable_all(int added)
530{
531 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
532 int idx;
533
534 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
535 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
536
537 if (!test_bit(idx, cpuc->active_mask))
538 continue;
539
540 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
541 }
542}
543
544static struct pmu pmu;
545
546static inline int is_x86_event(struct perf_event *event)
547{
548 return event->pmu == &pmu;
549}
550
551/*
552 * Event scheduler state:
553 *
554 * Assign events iterating over all events and counters, beginning
555 * with events with least weights first. Keep the current iterator
556 * state in struct sched_state.
557 */
558struct sched_state {
559 int weight;
560 int event; /* event index */
561 int counter; /* counter index */
562 int unassigned; /* number of events to be assigned left */
563 unsigned long used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
564};
565
566/* Total max is X86_PMC_IDX_MAX, but we are O(n!) limited */
567#define SCHED_STATES_MAX 2
568
569struct perf_sched {
570 int max_weight;
571 int max_events;
572 struct perf_event **events;
573 struct sched_state state;
574 int saved_states;
575 struct sched_state saved[SCHED_STATES_MAX];
576};
577
578/*
579 * Initialize interator that runs through all events and counters.
580 */
581static void perf_sched_init(struct perf_sched *sched, struct perf_event **events,
582 int num, int wmin, int wmax)
583{
584 int idx;
585
586 memset(sched, 0, sizeof(*sched));
587 sched->max_events = num;
588 sched->max_weight = wmax;
589 sched->events = events;
590
591 for (idx = 0; idx < num; idx++) {
592 if (events[idx]->hw.constraint->weight == wmin)
593 break;
594 }
595
596 sched->state.event = idx; /* start with min weight */
597 sched->state.weight = wmin;
598 sched->state.unassigned = num;
599}
600
601static void perf_sched_save_state(struct perf_sched *sched)
602{
603 if (WARN_ON_ONCE(sched->saved_states >= SCHED_STATES_MAX))
604 return;
605
606 sched->saved[sched->saved_states] = sched->state;
607 sched->saved_states++;
608}
609
610static bool perf_sched_restore_state(struct perf_sched *sched)
611{
612 if (!sched->saved_states)
613 return false;
614
615 sched->saved_states--;
616 sched->state = sched->saved[sched->saved_states];
617
618 /* continue with next counter: */
619 clear_bit(sched->state.counter++, sched->state.used);
620
621 return true;
622}
623
624/*
625 * Select a counter for the current event to schedule. Return true on
626 * success.
627 */
628static bool __perf_sched_find_counter(struct perf_sched *sched)
629{
630 struct event_constraint *c;
631 int idx;
632
633 if (!sched->state.unassigned)
634 return false;
635
636 if (sched->state.event >= sched->max_events)
637 return false;
638
639 c = sched->events[sched->state.event]->hw.constraint;
640 /* Prefer fixed purpose counters */
641 if (c->idxmsk64 & (~0ULL << INTEL_PMC_IDX_FIXED)) {
642 idx = INTEL_PMC_IDX_FIXED;
643 for_each_set_bit_from(idx, c->idxmsk, X86_PMC_IDX_MAX) {
644 if (!__test_and_set_bit(idx, sched->state.used))
645 goto done;
646 }
647 }
648 /* Grab the first unused counter starting with idx */
649 idx = sched->state.counter;
650 for_each_set_bit_from(idx, c->idxmsk, INTEL_PMC_IDX_FIXED) {
651 if (!__test_and_set_bit(idx, sched->state.used))
652 goto done;
653 }
654
655 return false;
656
657done:
658 sched->state.counter = idx;
659
660 if (c->overlap)
661 perf_sched_save_state(sched);
662
663 return true;
664}
665
666static bool perf_sched_find_counter(struct perf_sched *sched)
667{
668 while (!__perf_sched_find_counter(sched)) {
669 if (!perf_sched_restore_state(sched))
670 return false;
671 }
672
673 return true;
674}
675
676/*
677 * Go through all unassigned events and find the next one to schedule.
678 * Take events with the least weight first. Return true on success.
679 */
680static bool perf_sched_next_event(struct perf_sched *sched)
681{
682 struct event_constraint *c;
683
684 if (!sched->state.unassigned || !--sched->state.unassigned)
685 return false;
686
687 do {
688 /* next event */
689 sched->state.event++;
690 if (sched->state.event >= sched->max_events) {
691 /* next weight */
692 sched->state.event = 0;
693 sched->state.weight++;
694 if (sched->state.weight > sched->max_weight)
695 return false;
696 }
697 c = sched->events[sched->state.event]->hw.constraint;
698 } while (c->weight != sched->state.weight);
699
700 sched->state.counter = 0; /* start with first counter */
701
702 return true;
703}
704
705/*
706 * Assign a counter for each event.
707 */
708int perf_assign_events(struct perf_event **events, int n,
709 int wmin, int wmax, int *assign)
710{
711 struct perf_sched sched;
712
713 perf_sched_init(&sched, events, n, wmin, wmax);
714
715 do {
716 if (!perf_sched_find_counter(&sched))
717 break; /* failed */
718 if (assign)
719 assign[sched.state.event] = sched.state.counter;
720 } while (perf_sched_next_event(&sched));
721
722 return sched.state.unassigned;
723}
724
725int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
726{
727 struct event_constraint *c;
728 unsigned long used_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
729 struct perf_event *e;
730 int i, wmin, wmax, num = 0;
731 struct hw_perf_event *hwc;
732
733 bitmap_zero(used_mask, X86_PMC_IDX_MAX);
734
735 for (i = 0, wmin = X86_PMC_IDX_MAX, wmax = 0; i < n; i++) {
736 hwc = &cpuc->event_list[i]->hw;
737 c = x86_pmu.get_event_constraints(cpuc, cpuc->event_list[i]);
738 hwc->constraint = c;
739
740 wmin = min(wmin, c->weight);
741 wmax = max(wmax, c->weight);
742 }
743
744 /*
745 * fastpath, try to reuse previous register
746 */
747 for (i = 0; i < n; i++) {
748 hwc = &cpuc->event_list[i]->hw;
749 c = hwc->constraint;
750
751 /* never assigned */
752 if (hwc->idx == -1)
753 break;
754
755 /* constraint still honored */
756 if (!test_bit(hwc->idx, c->idxmsk))
757 break;
758
759 /* not already used */
760 if (test_bit(hwc->idx, used_mask))
761 break;
762
763 __set_bit(hwc->idx, used_mask);
764 if (assign)
765 assign[i] = hwc->idx;
766 }
767
768 /* slow path */
769 if (i != n)
770 num = perf_assign_events(cpuc->event_list, n, wmin,
771 wmax, assign);
772
773 /*
774 * Mark the event as committed, so we do not put_constraint()
775 * in case new events are added and fail scheduling.
776 */
777 if (!num && assign) {
778 for (i = 0; i < n; i++) {
779 e = cpuc->event_list[i];
780 e->hw.flags |= PERF_X86_EVENT_COMMITTED;
781 }
782 }
783 /*
784 * scheduling failed or is just a simulation,
785 * free resources if necessary
786 */
787 if (!assign || num) {
788 for (i = 0; i < n; i++) {
789 e = cpuc->event_list[i];
790 /*
791 * do not put_constraint() on comitted events,
792 * because they are good to go
793 */
794 if ((e->hw.flags & PERF_X86_EVENT_COMMITTED))
795 continue;
796
797 if (x86_pmu.put_event_constraints)
798 x86_pmu.put_event_constraints(cpuc, e);
799 }
800 }
801 return num ? -EINVAL : 0;
802}
803
804/*
805 * dogrp: true if must collect siblings events (group)
806 * returns total number of events and error code
807 */
808static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader, bool dogrp)
809{
810 struct perf_event *event;
811 int n, max_count;
812
813 max_count = x86_pmu.num_counters + x86_pmu.num_counters_fixed;
814
815 /* current number of events already accepted */
816 n = cpuc->n_events;
817
818 if (is_x86_event(leader)) {
819 if (n >= max_count)
820 return -EINVAL;
821 cpuc->event_list[n] = leader;
822 n++;
823 }
824 if (!dogrp)
825 return n;
826
827 list_for_each_entry(event, &leader->sibling_list, group_entry) {
828 if (!is_x86_event(event) ||
829 event->state <= PERF_EVENT_STATE_OFF)
830 continue;
831
832 if (n >= max_count)
833 return -EINVAL;
834
835 cpuc->event_list[n] = event;
836 n++;
837 }
838 return n;
839}
840
841static inline void x86_assign_hw_event(struct perf_event *event,
842 struct cpu_hw_events *cpuc, int i)
843{
844 struct hw_perf_event *hwc = &event->hw;
845
846 hwc->idx = cpuc->assign[i];
847 hwc->last_cpu = smp_processor_id();
848 hwc->last_tag = ++cpuc->tags[i];
849
850 if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
851 hwc->config_base = 0;
852 hwc->event_base = 0;
853 } else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
854 hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
855 hwc->event_base = MSR_ARCH_PERFMON_FIXED_CTR0 + (hwc->idx - INTEL_PMC_IDX_FIXED);
856 hwc->event_base_rdpmc = (hwc->idx - INTEL_PMC_IDX_FIXED) | 1<<30;
857 } else {
858 hwc->config_base = x86_pmu_config_addr(hwc->idx);
859 hwc->event_base = x86_pmu_event_addr(hwc->idx);
860 hwc->event_base_rdpmc = x86_pmu_rdpmc_index(hwc->idx);
861 }
862}
863
864static inline int match_prev_assignment(struct hw_perf_event *hwc,
865 struct cpu_hw_events *cpuc,
866 int i)
867{
868 return hwc->idx == cpuc->assign[i] &&
869 hwc->last_cpu == smp_processor_id() &&
870 hwc->last_tag == cpuc->tags[i];
871}
872
873static void x86_pmu_start(struct perf_event *event, int flags);
874
875static void x86_pmu_enable(struct pmu *pmu)
876{
877 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
878 struct perf_event *event;
879 struct hw_perf_event *hwc;
880 int i, added = cpuc->n_added;
881
882 if (!x86_pmu_initialized())
883 return;
884
885 if (cpuc->enabled)
886 return;
887
888 if (cpuc->n_added) {
889 int n_running = cpuc->n_events - cpuc->n_added;
890 /*
891 * apply assignment obtained either from
892 * hw_perf_group_sched_in() or x86_pmu_enable()
893 *
894 * step1: save events moving to new counters
895 */
896 for (i = 0; i < n_running; i++) {
897 event = cpuc->event_list[i];
898 hwc = &event->hw;
899
900 /*
901 * we can avoid reprogramming counter if:
902 * - assigned same counter as last time
903 * - running on same CPU as last time
904 * - no other event has used the counter since
905 */
906 if (hwc->idx == -1 ||
907 match_prev_assignment(hwc, cpuc, i))
908 continue;
909
910 /*
911 * Ensure we don't accidentally enable a stopped
912 * counter simply because we rescheduled.
913 */
914 if (hwc->state & PERF_HES_STOPPED)
915 hwc->state |= PERF_HES_ARCH;
916
917 x86_pmu_stop(event, PERF_EF_UPDATE);
918 }
919
920 /*
921 * step2: reprogram moved events into new counters
922 */
923 for (i = 0; i < cpuc->n_events; i++) {
924 event = cpuc->event_list[i];
925 hwc = &event->hw;
926
927 if (!match_prev_assignment(hwc, cpuc, i))
928 x86_assign_hw_event(event, cpuc, i);
929 else if (i < n_running)
930 continue;
931
932 if (hwc->state & PERF_HES_ARCH)
933 continue;
934
935 x86_pmu_start(event, PERF_EF_RELOAD);
936 }
937 cpuc->n_added = 0;
938 perf_events_lapic_init();
939 }
940
941 cpuc->enabled = 1;
942 barrier();
943
944 x86_pmu.enable_all(added);
945}
946
947static DEFINE_PER_CPU(u64 [X86_PMC_IDX_MAX], pmc_prev_left);
948
949/*
950 * Set the next IRQ period, based on the hwc->period_left value.
951 * To be called with the event disabled in hw:
952 */
953int x86_perf_event_set_period(struct perf_event *event)
954{
955 struct hw_perf_event *hwc = &event->hw;
956 s64 left = local64_read(&hwc->period_left);
957 s64 period = hwc->sample_period;
958 int ret = 0, idx = hwc->idx;
959
960 if (idx == INTEL_PMC_IDX_FIXED_BTS)
961 return 0;
962
963 /*
964 * If we are way outside a reasonable range then just skip forward:
965 */
966 if (unlikely(left <= -period)) {
967 left = period;
968 local64_set(&hwc->period_left, left);
969 hwc->last_period = period;
970 ret = 1;
971 }
972
973 if (unlikely(left <= 0)) {
974 left += period;
975 local64_set(&hwc->period_left, left);
976 hwc->last_period = period;
977 ret = 1;
978 }
979 /*
980 * Quirk: certain CPUs dont like it if just 1 hw_event is left:
981 */
982 if (unlikely(left < 2))
983 left = 2;
984
985 if (left > x86_pmu.max_period)
986 left = x86_pmu.max_period;
987
988 per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
989
990 /*
991 * The hw event starts counting from this event offset,
992 * mark it to be able to extra future deltas:
993 */
994 local64_set(&hwc->prev_count, (u64)-left);
995
996 wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
997
998 /*
999 * Due to erratum on certan cpu we need
1000 * a second write to be sure the register
1001 * is updated properly
1002 */
1003 if (x86_pmu.perfctr_second_write) {
1004 wrmsrl(hwc->event_base,
1005 (u64)(-left) & x86_pmu.cntval_mask);
1006 }
1007
1008 perf_event_update_userpage(event);
1009
1010 return ret;
1011}
1012
1013void x86_pmu_enable_event(struct perf_event *event)
1014{
1015 if (__this_cpu_read(cpu_hw_events.enabled))
1016 __x86_pmu_enable_event(&event->hw,
1017 ARCH_PERFMON_EVENTSEL_ENABLE);
1018}
1019
1020/*
1021 * Add a single event to the PMU.
1022 *
1023 * The event is added to the group of enabled events
1024 * but only if it can be scehduled with existing events.
1025 */
1026static int x86_pmu_add(struct perf_event *event, int flags)
1027{
1028 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1029 struct hw_perf_event *hwc;
1030 int assign[X86_PMC_IDX_MAX];
1031 int n, n0, ret;
1032
1033 hwc = &event->hw;
1034
1035 perf_pmu_disable(event->pmu);
1036 n0 = cpuc->n_events;
1037 ret = n = collect_events(cpuc, event, false);
1038 if (ret < 0)
1039 goto out;
1040
1041 hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
1042 if (!(flags & PERF_EF_START))
1043 hwc->state |= PERF_HES_ARCH;
1044
1045 /*
1046 * If group events scheduling transaction was started,
1047 * skip the schedulability test here, it will be performed
1048 * at commit time (->commit_txn) as a whole.
1049 */
1050 if (cpuc->group_flag & PERF_EVENT_TXN)
1051 goto done_collect;
1052
1053 ret = x86_pmu.schedule_events(cpuc, n, assign);
1054 if (ret)
1055 goto out;
1056 /*
1057 * copy new assignment, now we know it is possible
1058 * will be used by hw_perf_enable()
1059 */
1060 memcpy(cpuc->assign, assign, n*sizeof(int));
1061
1062done_collect:
1063 /*
1064 * Commit the collect_events() state. See x86_pmu_del() and
1065 * x86_pmu_*_txn().
1066 */
1067 cpuc->n_events = n;
1068 cpuc->n_added += n - n0;
1069 cpuc->n_txn += n - n0;
1070
1071 ret = 0;
1072out:
1073 perf_pmu_enable(event->pmu);
1074 return ret;
1075}
1076
1077static void x86_pmu_start(struct perf_event *event, int flags)
1078{
1079 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1080 int idx = event->hw.idx;
1081
1082 if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
1083 return;
1084
1085 if (WARN_ON_ONCE(idx == -1))
1086 return;
1087
1088 if (flags & PERF_EF_RELOAD) {
1089 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
1090 x86_perf_event_set_period(event);
1091 }
1092
1093 event->hw.state = 0;
1094
1095 cpuc->events[idx] = event;
1096 __set_bit(idx, cpuc->active_mask);
1097 __set_bit(idx, cpuc->running);
1098 x86_pmu.enable(event);
1099 perf_event_update_userpage(event);
1100}
1101
1102void perf_event_print_debug(void)
1103{
1104 u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
1105 u64 pebs;
1106 struct cpu_hw_events *cpuc;
1107 unsigned long flags;
1108 int cpu, idx;
1109
1110 if (!x86_pmu.num_counters)
1111 return;
1112
1113 local_irq_save(flags);
1114
1115 cpu = smp_processor_id();
1116 cpuc = &per_cpu(cpu_hw_events, cpu);
1117
1118 if (x86_pmu.version >= 2) {
1119 rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
1120 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
1121 rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
1122 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
1123 rdmsrl(MSR_IA32_PEBS_ENABLE, pebs);
1124
1125 pr_info("\n");
1126 pr_info("CPU#%d: ctrl: %016llx\n", cpu, ctrl);
1127 pr_info("CPU#%d: status: %016llx\n", cpu, status);
1128 pr_info("CPU#%d: overflow: %016llx\n", cpu, overflow);
1129 pr_info("CPU#%d: fixed: %016llx\n", cpu, fixed);
1130 pr_info("CPU#%d: pebs: %016llx\n", cpu, pebs);
1131 }
1132 pr_info("CPU#%d: active: %016llx\n", cpu, *(u64 *)cpuc->active_mask);
1133
1134 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1135 rdmsrl(x86_pmu_config_addr(idx), pmc_ctrl);
1136 rdmsrl(x86_pmu_event_addr(idx), pmc_count);
1137
1138 prev_left = per_cpu(pmc_prev_left[idx], cpu);
1139
1140 pr_info("CPU#%d: gen-PMC%d ctrl: %016llx\n",
1141 cpu, idx, pmc_ctrl);
1142 pr_info("CPU#%d: gen-PMC%d count: %016llx\n",
1143 cpu, idx, pmc_count);
1144 pr_info("CPU#%d: gen-PMC%d left: %016llx\n",
1145 cpu, idx, prev_left);
1146 }
1147 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) {
1148 rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
1149
1150 pr_info("CPU#%d: fixed-PMC%d count: %016llx\n",
1151 cpu, idx, pmc_count);
1152 }
1153 local_irq_restore(flags);
1154}
1155
1156void x86_pmu_stop(struct perf_event *event, int flags)
1157{
1158 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1159 struct hw_perf_event *hwc = &event->hw;
1160
1161 if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {
1162 x86_pmu.disable(event);
1163 cpuc->events[hwc->idx] = NULL;
1164 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
1165 hwc->state |= PERF_HES_STOPPED;
1166 }
1167
1168 if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
1169 /*
1170 * Drain the remaining delta count out of a event
1171 * that we are disabling:
1172 */
1173 x86_perf_event_update(event);
1174 hwc->state |= PERF_HES_UPTODATE;
1175 }
1176}
1177
1178static void x86_pmu_del(struct perf_event *event, int flags)
1179{
1180 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1181 int i;
1182
1183 /*
1184 * event is descheduled
1185 */
1186 event->hw.flags &= ~PERF_X86_EVENT_COMMITTED;
1187
1188 /*
1189 * If we're called during a txn, we don't need to do anything.
1190 * The events never got scheduled and ->cancel_txn will truncate
1191 * the event_list.
1192 *
1193 * XXX assumes any ->del() called during a TXN will only be on
1194 * an event added during that same TXN.
1195 */
1196 if (cpuc->group_flag & PERF_EVENT_TXN)
1197 return;
1198
1199 /*
1200 * Not a TXN, therefore cleanup properly.
1201 */
1202 x86_pmu_stop(event, PERF_EF_UPDATE);
1203
1204 for (i = 0; i < cpuc->n_events; i++) {
1205 if (event == cpuc->event_list[i])
1206 break;
1207 }
1208
1209 if (WARN_ON_ONCE(i == cpuc->n_events)) /* called ->del() without ->add() ? */
1210 return;
1211
1212 /* If we have a newly added event; make sure to decrease n_added. */
1213 if (i >= cpuc->n_events - cpuc->n_added)
1214 --cpuc->n_added;
1215
1216 if (x86_pmu.put_event_constraints)
1217 x86_pmu.put_event_constraints(cpuc, event);
1218
1219 /* Delete the array entry. */
1220 while (++i < cpuc->n_events)
1221 cpuc->event_list[i-1] = cpuc->event_list[i];
1222 --cpuc->n_events;
1223
1224 perf_event_update_userpage(event);
1225}
1226
1227int x86_pmu_handle_irq(struct pt_regs *regs)
1228{
1229 struct perf_sample_data data;
1230 struct cpu_hw_events *cpuc;
1231 struct perf_event *event;
1232 int idx, handled = 0;
1233 u64 val;
1234
1235 cpuc = &__get_cpu_var(cpu_hw_events);
1236
1237 /*
1238 * Some chipsets need to unmask the LVTPC in a particular spot
1239 * inside the nmi handler. As a result, the unmasking was pushed
1240 * into all the nmi handlers.
1241 *
1242 * This generic handler doesn't seem to have any issues where the
1243 * unmasking occurs so it was left at the top.
1244 */
1245 apic_write(APIC_LVTPC, APIC_DM_NMI);
1246
1247 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1248 if (!test_bit(idx, cpuc->active_mask)) {
1249 /*
1250 * Though we deactivated the counter some cpus
1251 * might still deliver spurious interrupts still
1252 * in flight. Catch them:
1253 */
1254 if (__test_and_clear_bit(idx, cpuc->running))
1255 handled++;
1256 continue;
1257 }
1258
1259 event = cpuc->events[idx];
1260
1261 val = x86_perf_event_update(event);
1262 if (val & (1ULL << (x86_pmu.cntval_bits - 1)))
1263 continue;
1264
1265 /*
1266 * event overflow
1267 */
1268 handled++;
1269 perf_sample_data_init(&data, 0, event->hw.last_period);
1270
1271 if (!x86_perf_event_set_period(event))
1272 continue;
1273
1274 if (perf_event_overflow(event, &data, regs))
1275 x86_pmu_stop(event, 0);
1276 }
1277
1278 if (handled)
1279 inc_irq_stat(apic_perf_irqs);
1280
1281 return handled;
1282}
1283
1284void perf_events_lapic_init(void)
1285{
1286 if (!x86_pmu.apic || !x86_pmu_initialized())
1287 return;
1288
1289 /*
1290 * Always use NMI for PMU
1291 */
1292 apic_write(APIC_LVTPC, APIC_DM_NMI);
1293}
1294
1295static int __kprobes
1296perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs)
1297{
1298 u64 start_clock;
1299 u64 finish_clock;
1300 int ret;
1301
1302 if (!atomic_read(&active_events))
1303 return NMI_DONE;
1304
1305 start_clock = sched_clock();
1306 ret = x86_pmu.handle_irq(regs);
1307 finish_clock = sched_clock();
1308
1309 perf_sample_event_took(finish_clock - start_clock);
1310
1311 return ret;
1312}
1313
1314struct event_constraint emptyconstraint;
1315struct event_constraint unconstrained;
1316
1317static int
1318x86_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
1319{
1320 unsigned int cpu = (long)hcpu;
1321 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1322 int ret = NOTIFY_OK;
1323
1324 switch (action & ~CPU_TASKS_FROZEN) {
1325 case CPU_UP_PREPARE:
1326 cpuc->kfree_on_online = NULL;
1327 if (x86_pmu.cpu_prepare)
1328 ret = x86_pmu.cpu_prepare(cpu);
1329 break;
1330
1331 case CPU_STARTING:
1332 if (x86_pmu.attr_rdpmc)
1333 set_in_cr4(X86_CR4_PCE);
1334 if (x86_pmu.cpu_starting)
1335 x86_pmu.cpu_starting(cpu);
1336 break;
1337
1338 case CPU_ONLINE:
1339 kfree(cpuc->kfree_on_online);
1340 break;
1341
1342 case CPU_DYING:
1343 if (x86_pmu.cpu_dying)
1344 x86_pmu.cpu_dying(cpu);
1345 break;
1346
1347 case CPU_UP_CANCELED:
1348 case CPU_DEAD:
1349 if (x86_pmu.cpu_dead)
1350 x86_pmu.cpu_dead(cpu);
1351 break;
1352
1353 default:
1354 break;
1355 }
1356
1357 return ret;
1358}
1359
1360static void __init pmu_check_apic(void)
1361{
1362 if (cpu_has_apic)
1363 return;
1364
1365 x86_pmu.apic = 0;
1366 pr_info("no APIC, boot with the \"lapic\" boot parameter to force-enable it.\n");
1367 pr_info("no hardware sampling interrupt available.\n");
1368}
1369
1370static struct attribute_group x86_pmu_format_group = {
1371 .name = "format",
1372 .attrs = NULL,
1373};
1374
1375/*
1376 * Remove all undefined events (x86_pmu.event_map(id) == 0)
1377 * out of events_attr attributes.
1378 */
1379static void __init filter_events(struct attribute **attrs)
1380{
1381 struct device_attribute *d;
1382 struct perf_pmu_events_attr *pmu_attr;
1383 int i, j;
1384
1385 for (i = 0; attrs[i]; i++) {
1386 d = (struct device_attribute *)attrs[i];
1387 pmu_attr = container_of(d, struct perf_pmu_events_attr, attr);
1388 /* str trumps id */
1389 if (pmu_attr->event_str)
1390 continue;
1391 if (x86_pmu.event_map(i))
1392 continue;
1393
1394 for (j = i; attrs[j]; j++)
1395 attrs[j] = attrs[j + 1];
1396
1397 /* Check the shifted attr. */
1398 i--;
1399 }
1400}
1401
1402/* Merge two pointer arrays */
1403static __init struct attribute **merge_attr(struct attribute **a, struct attribute **b)
1404{
1405 struct attribute **new;
1406 int j, i;
1407
1408 for (j = 0; a[j]; j++)
1409 ;
1410 for (i = 0; b[i]; i++)
1411 j++;
1412 j++;
1413
1414 new = kmalloc(sizeof(struct attribute *) * j, GFP_KERNEL);
1415 if (!new)
1416 return NULL;
1417
1418 j = 0;
1419 for (i = 0; a[i]; i++)
1420 new[j++] = a[i];
1421 for (i = 0; b[i]; i++)
1422 new[j++] = b[i];
1423 new[j] = NULL;
1424
1425 return new;
1426}
1427
1428ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
1429 char *page)
1430{
1431 struct perf_pmu_events_attr *pmu_attr = \
1432 container_of(attr, struct perf_pmu_events_attr, attr);
1433 u64 config = x86_pmu.event_map(pmu_attr->id);
1434
1435 /* string trumps id */
1436 if (pmu_attr->event_str)
1437 return sprintf(page, "%s", pmu_attr->event_str);
1438
1439 return x86_pmu.events_sysfs_show(page, config);
1440}
1441
1442EVENT_ATTR(cpu-cycles, CPU_CYCLES );
1443EVENT_ATTR(instructions, INSTRUCTIONS );
1444EVENT_ATTR(cache-references, CACHE_REFERENCES );
1445EVENT_ATTR(cache-misses, CACHE_MISSES );
1446EVENT_ATTR(branch-instructions, BRANCH_INSTRUCTIONS );
1447EVENT_ATTR(branch-misses, BRANCH_MISSES );
1448EVENT_ATTR(bus-cycles, BUS_CYCLES );
1449EVENT_ATTR(stalled-cycles-frontend, STALLED_CYCLES_FRONTEND );
1450EVENT_ATTR(stalled-cycles-backend, STALLED_CYCLES_BACKEND );
1451EVENT_ATTR(ref-cycles, REF_CPU_CYCLES );
1452
1453static struct attribute *empty_attrs;
1454
1455static struct attribute *events_attr[] = {
1456 EVENT_PTR(CPU_CYCLES),
1457 EVENT_PTR(INSTRUCTIONS),
1458 EVENT_PTR(CACHE_REFERENCES),
1459 EVENT_PTR(CACHE_MISSES),
1460 EVENT_PTR(BRANCH_INSTRUCTIONS),
1461 EVENT_PTR(BRANCH_MISSES),
1462 EVENT_PTR(BUS_CYCLES),
1463 EVENT_PTR(STALLED_CYCLES_FRONTEND),
1464 EVENT_PTR(STALLED_CYCLES_BACKEND),
1465 EVENT_PTR(REF_CPU_CYCLES),
1466 NULL,
1467};
1468
1469static struct attribute_group x86_pmu_events_group = {
1470 .name = "events",
1471 .attrs = events_attr,
1472};
1473
1474ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event)
1475{
1476 u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
1477 u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
1478 bool edge = (config & ARCH_PERFMON_EVENTSEL_EDGE);
1479 bool pc = (config & ARCH_PERFMON_EVENTSEL_PIN_CONTROL);
1480 bool any = (config & ARCH_PERFMON_EVENTSEL_ANY);
1481 bool inv = (config & ARCH_PERFMON_EVENTSEL_INV);
1482 ssize_t ret;
1483
1484 /*
1485 * We have whole page size to spend and just little data
1486 * to write, so we can safely use sprintf.
1487 */
1488 ret = sprintf(page, "event=0x%02llx", event);
1489
1490 if (umask)
1491 ret += sprintf(page + ret, ",umask=0x%02llx", umask);
1492
1493 if (edge)
1494 ret += sprintf(page + ret, ",edge");
1495
1496 if (pc)
1497 ret += sprintf(page + ret, ",pc");
1498
1499 if (any)
1500 ret += sprintf(page + ret, ",any");
1501
1502 if (inv)
1503 ret += sprintf(page + ret, ",inv");
1504
1505 if (cmask)
1506 ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
1507
1508 ret += sprintf(page + ret, "\n");
1509
1510 return ret;
1511}
1512
1513static int __init init_hw_perf_events(void)
1514{
1515 struct x86_pmu_quirk *quirk;
1516 int err;
1517
1518 pr_info("Performance Events: ");
1519
1520 switch (boot_cpu_data.x86_vendor) {
1521 case X86_VENDOR_INTEL:
1522 err = intel_pmu_init();
1523 break;
1524 case X86_VENDOR_AMD:
1525 err = amd_pmu_init();
1526 break;
1527 default:
1528 err = -ENOTSUPP;
1529 }
1530 if (err != 0) {
1531 pr_cont("no PMU driver, software events only.\n");
1532 return 0;
1533 }
1534
1535 pmu_check_apic();
1536
1537 /* sanity check that the hardware exists or is emulated */
1538 if (!check_hw_exists())
1539 return 0;
1540
1541 pr_cont("%s PMU driver.\n", x86_pmu.name);
1542
1543 x86_pmu.attr_rdpmc = 1; /* enable userspace RDPMC usage by default */
1544
1545 for (quirk = x86_pmu.quirks; quirk; quirk = quirk->next)
1546 quirk->func();
1547
1548 if (!x86_pmu.intel_ctrl)
1549 x86_pmu.intel_ctrl = (1 << x86_pmu.num_counters) - 1;
1550
1551 perf_events_lapic_init();
1552 register_nmi_handler(NMI_LOCAL, perf_event_nmi_handler, 0, "PMI");
1553
1554 unconstrained = (struct event_constraint)
1555 __EVENT_CONSTRAINT(0, (1ULL << x86_pmu.num_counters) - 1,
1556 0, x86_pmu.num_counters, 0, 0);
1557
1558 x86_pmu_format_group.attrs = x86_pmu.format_attrs;
1559
1560 if (x86_pmu.event_attrs)
1561 x86_pmu_events_group.attrs = x86_pmu.event_attrs;
1562
1563 if (!x86_pmu.events_sysfs_show)
1564 x86_pmu_events_group.attrs = &empty_attrs;
1565 else
1566 filter_events(x86_pmu_events_group.attrs);
1567
1568 if (x86_pmu.cpu_events) {
1569 struct attribute **tmp;
1570
1571 tmp = merge_attr(x86_pmu_events_group.attrs, x86_pmu.cpu_events);
1572 if (!WARN_ON(!tmp))
1573 x86_pmu_events_group.attrs = tmp;
1574 }
1575
1576 pr_info("... version: %d\n", x86_pmu.version);
1577 pr_info("... bit width: %d\n", x86_pmu.cntval_bits);
1578 pr_info("... generic registers: %d\n", x86_pmu.num_counters);
1579 pr_info("... value mask: %016Lx\n", x86_pmu.cntval_mask);
1580 pr_info("... max period: %016Lx\n", x86_pmu.max_period);
1581 pr_info("... fixed-purpose events: %d\n", x86_pmu.num_counters_fixed);
1582 pr_info("... event mask: %016Lx\n", x86_pmu.intel_ctrl);
1583
1584 perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
1585 perf_cpu_notifier(x86_pmu_notifier);
1586
1587 return 0;
1588}
1589early_initcall(init_hw_perf_events);
1590
1591static inline void x86_pmu_read(struct perf_event *event)
1592{
1593 x86_perf_event_update(event);
1594}
1595
1596/*
1597 * Start group events scheduling transaction
1598 * Set the flag to make pmu::enable() not perform the
1599 * schedulability test, it will be performed at commit time
1600 */
1601static void x86_pmu_start_txn(struct pmu *pmu)
1602{
1603 perf_pmu_disable(pmu);
1604 __this_cpu_or(cpu_hw_events.group_flag, PERF_EVENT_TXN);
1605 __this_cpu_write(cpu_hw_events.n_txn, 0);
1606}
1607
1608/*
1609 * Stop group events scheduling transaction
1610 * Clear the flag and pmu::enable() will perform the
1611 * schedulability test.
1612 */
1613static void x86_pmu_cancel_txn(struct pmu *pmu)
1614{
1615 __this_cpu_and(cpu_hw_events.group_flag, ~PERF_EVENT_TXN);
1616 /*
1617 * Truncate collected array by the number of events added in this
1618 * transaction. See x86_pmu_add() and x86_pmu_*_txn().
1619 */
1620 __this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
1621 __this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
1622 perf_pmu_enable(pmu);
1623}
1624
1625/*
1626 * Commit group events scheduling transaction
1627 * Perform the group schedulability test as a whole
1628 * Return 0 if success
1629 *
1630 * Does not cancel the transaction on failure; expects the caller to do this.
1631 */
1632static int x86_pmu_commit_txn(struct pmu *pmu)
1633{
1634 struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1635 int assign[X86_PMC_IDX_MAX];
1636 int n, ret;
1637
1638 n = cpuc->n_events;
1639
1640 if (!x86_pmu_initialized())
1641 return -EAGAIN;
1642
1643 ret = x86_pmu.schedule_events(cpuc, n, assign);
1644 if (ret)
1645 return ret;
1646
1647 /*
1648 * copy new assignment, now we know it is possible
1649 * will be used by hw_perf_enable()
1650 */
1651 memcpy(cpuc->assign, assign, n*sizeof(int));
1652
1653 cpuc->group_flag &= ~PERF_EVENT_TXN;
1654 perf_pmu_enable(pmu);
1655 return 0;
1656}
1657/*
1658 * a fake_cpuc is used to validate event groups. Due to
1659 * the extra reg logic, we need to also allocate a fake
1660 * per_core and per_cpu structure. Otherwise, group events
1661 * using extra reg may conflict without the kernel being
1662 * able to catch this when the last event gets added to
1663 * the group.
1664 */
1665static void free_fake_cpuc(struct cpu_hw_events *cpuc)
1666{
1667 kfree(cpuc->shared_regs);
1668 kfree(cpuc);
1669}
1670
1671static struct cpu_hw_events *allocate_fake_cpuc(void)
1672{
1673 struct cpu_hw_events *cpuc;
1674 int cpu = raw_smp_processor_id();
1675
1676 cpuc = kzalloc(sizeof(*cpuc), GFP_KERNEL);
1677 if (!cpuc)
1678 return ERR_PTR(-ENOMEM);
1679
1680 /* only needed, if we have extra_regs */
1681 if (x86_pmu.extra_regs) {
1682 cpuc->shared_regs = allocate_shared_regs(cpu);
1683 if (!cpuc->shared_regs)
1684 goto error;
1685 }
1686 cpuc->is_fake = 1;
1687 return cpuc;
1688error:
1689 free_fake_cpuc(cpuc);
1690 return ERR_PTR(-ENOMEM);
1691}
1692
1693/*
1694 * validate that we can schedule this event
1695 */
1696static int validate_event(struct perf_event *event)
1697{
1698 struct cpu_hw_events *fake_cpuc;
1699 struct event_constraint *c;
1700 int ret = 0;
1701
1702 fake_cpuc = allocate_fake_cpuc();
1703 if (IS_ERR(fake_cpuc))
1704 return PTR_ERR(fake_cpuc);
1705
1706 c = x86_pmu.get_event_constraints(fake_cpuc, event);
1707
1708 if (!c || !c->weight)
1709 ret = -EINVAL;
1710
1711 if (x86_pmu.put_event_constraints)
1712 x86_pmu.put_event_constraints(fake_cpuc, event);
1713
1714 free_fake_cpuc(fake_cpuc);
1715
1716 return ret;
1717}
1718
1719/*
1720 * validate a single event group
1721 *
1722 * validation include:
1723 * - check events are compatible which each other
1724 * - events do not compete for the same counter
1725 * - number of events <= number of counters
1726 *
1727 * validation ensures the group can be loaded onto the
1728 * PMU if it was the only group available.
1729 */
1730static int validate_group(struct perf_event *event)
1731{
1732 struct perf_event *leader = event->group_leader;
1733 struct cpu_hw_events *fake_cpuc;
1734 int ret = -EINVAL, n;
1735
1736 fake_cpuc = allocate_fake_cpuc();
1737 if (IS_ERR(fake_cpuc))
1738 return PTR_ERR(fake_cpuc);
1739 /*
1740 * the event is not yet connected with its
1741 * siblings therefore we must first collect
1742 * existing siblings, then add the new event
1743 * before we can simulate the scheduling
1744 */
1745 n = collect_events(fake_cpuc, leader, true);
1746 if (n < 0)
1747 goto out;
1748
1749 fake_cpuc->n_events = n;
1750 n = collect_events(fake_cpuc, event, false);
1751 if (n < 0)
1752 goto out;
1753
1754 fake_cpuc->n_events = n;
1755
1756 ret = x86_pmu.schedule_events(fake_cpuc, n, NULL);
1757
1758out:
1759 free_fake_cpuc(fake_cpuc);
1760 return ret;
1761}
1762
1763static int x86_pmu_event_init(struct perf_event *event)
1764{
1765 struct pmu *tmp;
1766 int err;
1767
1768 switch (event->attr.type) {
1769 case PERF_TYPE_RAW:
1770 case PERF_TYPE_HARDWARE:
1771 case PERF_TYPE_HW_CACHE:
1772 break;
1773
1774 default:
1775 return -ENOENT;
1776 }
1777
1778 err = __x86_pmu_event_init(event);
1779 if (!err) {
1780 /*
1781 * we temporarily connect event to its pmu
1782 * such that validate_group() can classify
1783 * it as an x86 event using is_x86_event()
1784 */
1785 tmp = event->pmu;
1786 event->pmu = &pmu;
1787
1788 if (event->group_leader != event)
1789 err = validate_group(event);
1790 else
1791 err = validate_event(event);
1792
1793 event->pmu = tmp;
1794 }
1795 if (err) {
1796 if (event->destroy)
1797 event->destroy(event);
1798 }
1799
1800 return err;
1801}
1802
1803static int x86_pmu_event_idx(struct perf_event *event)
1804{
1805 int idx = event->hw.idx;
1806
1807 if (!x86_pmu.attr_rdpmc)
1808 return 0;
1809
1810 if (x86_pmu.num_counters_fixed && idx >= INTEL_PMC_IDX_FIXED) {
1811 idx -= INTEL_PMC_IDX_FIXED;
1812 idx |= 1 << 30;
1813 }
1814
1815 return idx + 1;
1816}
1817
1818static ssize_t get_attr_rdpmc(struct device *cdev,
1819 struct device_attribute *attr,
1820 char *buf)
1821{
1822 return snprintf(buf, 40, "%d\n", x86_pmu.attr_rdpmc);
1823}
1824
1825static void change_rdpmc(void *info)
1826{
1827 bool enable = !!(unsigned long)info;
1828
1829 if (enable)
1830 set_in_cr4(X86_CR4_PCE);
1831 else
1832 clear_in_cr4(X86_CR4_PCE);
1833}
1834
1835static ssize_t set_attr_rdpmc(struct device *cdev,
1836 struct device_attribute *attr,
1837 const char *buf, size_t count)
1838{
1839 unsigned long val;
1840 ssize_t ret;
1841
1842 ret = kstrtoul(buf, 0, &val);
1843 if (ret)
1844 return ret;
1845
1846 if (x86_pmu.attr_rdpmc_broken)
1847 return -ENOTSUPP;
1848
1849 if (!!val != !!x86_pmu.attr_rdpmc) {
1850 x86_pmu.attr_rdpmc = !!val;
1851 on_each_cpu(change_rdpmc, (void *)val, 1);
1852 }
1853
1854 return count;
1855}
1856
1857static DEVICE_ATTR(rdpmc, S_IRUSR | S_IWUSR, get_attr_rdpmc, set_attr_rdpmc);
1858
1859static struct attribute *x86_pmu_attrs[] = {
1860 &dev_attr_rdpmc.attr,
1861 NULL,
1862};
1863
1864static struct attribute_group x86_pmu_attr_group = {
1865 .attrs = x86_pmu_attrs,
1866};
1867
1868static const struct attribute_group *x86_pmu_attr_groups[] = {
1869 &x86_pmu_attr_group,
1870 &x86_pmu_format_group,
1871 &x86_pmu_events_group,
1872 NULL,
1873};
1874
1875static void x86_pmu_flush_branch_stack(void)
1876{
1877 if (x86_pmu.flush_branch_stack)
1878 x86_pmu.flush_branch_stack();
1879}
1880
1881void perf_check_microcode(void)
1882{
1883 if (x86_pmu.check_microcode)
1884 x86_pmu.check_microcode();
1885}
1886EXPORT_SYMBOL_GPL(perf_check_microcode);
1887
1888static struct pmu pmu = {
1889 .pmu_enable = x86_pmu_enable,
1890 .pmu_disable = x86_pmu_disable,
1891
1892 .attr_groups = x86_pmu_attr_groups,
1893
1894 .event_init = x86_pmu_event_init,
1895
1896 .add = x86_pmu_add,
1897 .del = x86_pmu_del,
1898 .start = x86_pmu_start,
1899 .stop = x86_pmu_stop,
1900 .read = x86_pmu_read,
1901
1902 .start_txn = x86_pmu_start_txn,
1903 .cancel_txn = x86_pmu_cancel_txn,
1904 .commit_txn = x86_pmu_commit_txn,
1905
1906 .event_idx = x86_pmu_event_idx,
1907 .flush_branch_stack = x86_pmu_flush_branch_stack,
1908};
1909
1910void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
1911{
1912 struct cyc2ns_data *data;
1913
1914 userpg->cap_user_time = 0;
1915 userpg->cap_user_time_zero = 0;
1916 userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc;
1917 userpg->pmc_width = x86_pmu.cntval_bits;
1918
1919 if (!sched_clock_stable())
1920 return;
1921
1922 data = cyc2ns_read_begin();
1923
1924 userpg->cap_user_time = 1;
1925 userpg->time_mult = data->cyc2ns_mul;
1926 userpg->time_shift = data->cyc2ns_shift;
1927 userpg->time_offset = data->cyc2ns_offset - now;
1928
1929 userpg->cap_user_time_zero = 1;
1930 userpg->time_zero = data->cyc2ns_offset;
1931
1932 cyc2ns_read_end(data);
1933}
1934
1935/*
1936 * callchain support
1937 */
1938
1939static int backtrace_stack(void *data, char *name)
1940{
1941 return 0;
1942}
1943
1944static void backtrace_address(void *data, unsigned long addr, int reliable)
1945{
1946 struct perf_callchain_entry *entry = data;
1947
1948 perf_callchain_store(entry, addr);
1949}
1950
1951static const struct stacktrace_ops backtrace_ops = {
1952 .stack = backtrace_stack,
1953 .address = backtrace_address,
1954 .walk_stack = print_context_stack_bp,
1955};
1956
1957void
1958perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs)
1959{
1960 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
1961 /* TODO: We don't support guest os callchain now */
1962 return;
1963 }
1964
1965 perf_callchain_store(entry, regs->ip);
1966
1967 dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
1968}
1969
1970static inline int
1971valid_user_frame(const void __user *fp, unsigned long size)
1972{
1973 return (__range_not_ok(fp, size, TASK_SIZE) == 0);
1974}
1975
1976static unsigned long get_segment_base(unsigned int segment)
1977{
1978 struct desc_struct *desc;
1979 int idx = segment >> 3;
1980
1981 if ((segment & SEGMENT_TI_MASK) == SEGMENT_LDT) {
1982 if (idx > LDT_ENTRIES)
1983 return 0;
1984
1985 if (idx > current->active_mm->context.size)
1986 return 0;
1987
1988 desc = current->active_mm->context.ldt;
1989 } else {
1990 if (idx > GDT_ENTRIES)
1991 return 0;
1992
1993 desc = __this_cpu_ptr(&gdt_page.gdt[0]);
1994 }
1995
1996 return get_desc_base(desc + idx);
1997}
1998
1999#ifdef CONFIG_COMPAT
2000
2001#include <asm/compat.h>
2002
2003static inline int
2004perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2005{
2006 /* 32-bit process in 64-bit kernel. */
2007 unsigned long ss_base, cs_base;
2008 struct stack_frame_ia32 frame;
2009 const void __user *fp;
2010
2011 if (!test_thread_flag(TIF_IA32))
2012 return 0;
2013
2014 cs_base = get_segment_base(regs->cs);
2015 ss_base = get_segment_base(regs->ss);
2016
2017 fp = compat_ptr(ss_base + regs->bp);
2018 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2019 unsigned long bytes;
2020 frame.next_frame = 0;
2021 frame.return_address = 0;
2022
2023 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2024 if (bytes != 0)
2025 break;
2026
2027 if (!valid_user_frame(fp, sizeof(frame)))
2028 break;
2029
2030 perf_callchain_store(entry, cs_base + frame.return_address);
2031 fp = compat_ptr(ss_base + frame.next_frame);
2032 }
2033 return 1;
2034}
2035#else
2036static inline int
2037perf_callchain_user32(struct pt_regs *regs, struct perf_callchain_entry *entry)
2038{
2039 return 0;
2040}
2041#endif
2042
2043void
2044perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
2045{
2046 struct stack_frame frame;
2047 const void __user *fp;
2048
2049 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2050 /* TODO: We don't support guest os callchain now */
2051 return;
2052 }
2053
2054 /*
2055 * We don't know what to do with VM86 stacks.. ignore them for now.
2056 */
2057 if (regs->flags & (X86_VM_MASK | PERF_EFLAGS_VM))
2058 return;
2059
2060 fp = (void __user *)regs->bp;
2061
2062 perf_callchain_store(entry, regs->ip);
2063
2064 if (!current->mm)
2065 return;
2066
2067 if (perf_callchain_user32(regs, entry))
2068 return;
2069
2070 while (entry->nr < PERF_MAX_STACK_DEPTH) {
2071 unsigned long bytes;
2072 frame.next_frame = NULL;
2073 frame.return_address = 0;
2074
2075 bytes = copy_from_user_nmi(&frame, fp, sizeof(frame));
2076 if (bytes != 0)
2077 break;
2078
2079 if (!valid_user_frame(fp, sizeof(frame)))
2080 break;
2081
2082 perf_callchain_store(entry, frame.return_address);
2083 fp = frame.next_frame;
2084 }
2085}
2086
2087/*
2088 * Deal with code segment offsets for the various execution modes:
2089 *
2090 * VM86 - the good olde 16 bit days, where the linear address is
2091 * 20 bits and we use regs->ip + 0x10 * regs->cs.
2092 *
2093 * IA32 - Where we need to look at GDT/LDT segment descriptor tables
2094 * to figure out what the 32bit base address is.
2095 *
2096 * X32 - has TIF_X32 set, but is running in x86_64
2097 *
2098 * X86_64 - CS,DS,SS,ES are all zero based.
2099 */
2100static unsigned long code_segment_base(struct pt_regs *regs)
2101{
2102 /*
2103 * If we are in VM86 mode, add the segment offset to convert to a
2104 * linear address.
2105 */
2106 if (regs->flags & X86_VM_MASK)
2107 return 0x10 * regs->cs;
2108
2109 /*
2110 * For IA32 we look at the GDT/LDT segment base to convert the
2111 * effective IP to a linear address.
2112 */
2113#ifdef CONFIG_X86_32
2114 if (user_mode(regs) && regs->cs != __USER_CS)
2115 return get_segment_base(regs->cs);
2116#else
2117 if (test_thread_flag(TIF_IA32)) {
2118 if (user_mode(regs) && regs->cs != __USER32_CS)
2119 return get_segment_base(regs->cs);
2120 }
2121#endif
2122 return 0;
2123}
2124
2125unsigned long perf_instruction_pointer(struct pt_regs *regs)
2126{
2127 if (perf_guest_cbs && perf_guest_cbs->is_in_guest())
2128 return perf_guest_cbs->get_guest_ip();
2129
2130 return regs->ip + code_segment_base(regs);
2131}
2132
2133unsigned long perf_misc_flags(struct pt_regs *regs)
2134{
2135 int misc = 0;
2136
2137 if (perf_guest_cbs && perf_guest_cbs->is_in_guest()) {
2138 if (perf_guest_cbs->is_user_mode())
2139 misc |= PERF_RECORD_MISC_GUEST_USER;
2140 else
2141 misc |= PERF_RECORD_MISC_GUEST_KERNEL;
2142 } else {
2143 if (user_mode(regs))
2144 misc |= PERF_RECORD_MISC_USER;
2145 else
2146 misc |= PERF_RECORD_MISC_KERNEL;
2147 }
2148
2149 if (regs->flags & PERF_EFLAGS_EXACT)
2150 misc |= PERF_RECORD_MISC_EXACT_IP;
2151
2152 return misc;
2153}
2154
2155void perf_get_x86_pmu_capability(struct x86_pmu_capability *cap)
2156{
2157 cap->version = x86_pmu.version;
2158 cap->num_counters_gp = x86_pmu.num_counters;
2159 cap->num_counters_fixed = x86_pmu.num_counters_fixed;
2160 cap->bit_width_gp = x86_pmu.cntval_bits;
2161 cap->bit_width_fixed = x86_pmu.cntval_bits;
2162 cap->events_mask = (unsigned int)x86_pmu.events_maskl;
2163 cap->events_mask_len = x86_pmu.events_mask_len;
2164}
2165EXPORT_SYMBOL_GPL(perf_get_x86_pmu_capability);