aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/perf_event.c
diff options
context:
space:
mode:
Diffstat (limited to 'kernel/perf_event.c')
-rw-r--r--kernel/perf_event.c2724
1 files changed, 1630 insertions, 1094 deletions
diff --git a/kernel/perf_event.c b/kernel/perf_event.c
index 403d1804b198..2870feee81dd 100644
--- a/kernel/perf_event.c
+++ b/kernel/perf_event.c
@@ -35,20 +35,15 @@
35 35
36#include <asm/irq_regs.h> 36#include <asm/irq_regs.h>
37 37
38/* 38atomic_t perf_task_events __read_mostly;
39 * Each CPU has a list of per CPU events:
40 */
41static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
42
43int perf_max_events __read_mostly = 1;
44static int perf_reserved_percpu __read_mostly;
45static int perf_overcommit __read_mostly = 1;
46
47static atomic_t nr_events __read_mostly;
48static atomic_t nr_mmap_events __read_mostly; 39static atomic_t nr_mmap_events __read_mostly;
49static atomic_t nr_comm_events __read_mostly; 40static atomic_t nr_comm_events __read_mostly;
50static atomic_t nr_task_events __read_mostly; 41static atomic_t nr_task_events __read_mostly;
51 42
43static LIST_HEAD(pmus);
44static DEFINE_MUTEX(pmus_lock);
45static struct srcu_struct pmus_srcu;
46
52/* 47/*
53 * perf event paranoia level: 48 * perf event paranoia level:
54 * -1 - not paranoid at all 49 * -1 - not paranoid at all
@@ -67,36 +62,43 @@ int sysctl_perf_event_sample_rate __read_mostly = 100000;
67 62
68static atomic64_t perf_event_id; 63static atomic64_t perf_event_id;
69 64
70/* 65void __weak perf_event_print_debug(void) { }
71 * Lock for (sysadmin-configurable) event reservations:
72 */
73static DEFINE_SPINLOCK(perf_resource_lock);
74 66
75/* 67extern __weak const char *perf_pmu_name(void)
76 * Architecture provided APIs - weak aliases:
77 */
78extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
79{ 68{
80 return NULL; 69 return "pmu";
81} 70}
82 71
83void __weak hw_perf_disable(void) { barrier(); } 72void perf_pmu_disable(struct pmu *pmu)
84void __weak hw_perf_enable(void) { barrier(); } 73{
85 74 int *count = this_cpu_ptr(pmu->pmu_disable_count);
86void __weak perf_event_print_debug(void) { } 75 if (!(*count)++)
87 76 pmu->pmu_disable(pmu);
88static DEFINE_PER_CPU(int, perf_disable_count); 77}
89 78
90void perf_disable(void) 79void perf_pmu_enable(struct pmu *pmu)
91{ 80{
92 if (!__get_cpu_var(perf_disable_count)++) 81 int *count = this_cpu_ptr(pmu->pmu_disable_count);
93 hw_perf_disable(); 82 if (!--(*count))
83 pmu->pmu_enable(pmu);
94} 84}
95 85
96void perf_enable(void) 86static DEFINE_PER_CPU(struct list_head, rotation_list);
87
88/*
89 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
90 * because they're strictly cpu affine and rotate_start is called with IRQs
91 * disabled, while rotate_context is called from IRQ context.
92 */
93static void perf_pmu_rotate_start(struct pmu *pmu)
97{ 94{
98 if (!--__get_cpu_var(perf_disable_count)) 95 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
99 hw_perf_enable(); 96 struct list_head *head = &__get_cpu_var(rotation_list);
97
98 WARN_ON(!irqs_disabled());
99
100 if (list_empty(&cpuctx->rotation_list))
101 list_add(&cpuctx->rotation_list, head);
100} 102}
101 103
102static void get_ctx(struct perf_event_context *ctx) 104static void get_ctx(struct perf_event_context *ctx)
@@ -151,13 +153,13 @@ static u64 primary_event_id(struct perf_event *event)
151 * the context could get moved to another task. 153 * the context could get moved to another task.
152 */ 154 */
153static struct perf_event_context * 155static struct perf_event_context *
154perf_lock_task_context(struct task_struct *task, unsigned long *flags) 156perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
155{ 157{
156 struct perf_event_context *ctx; 158 struct perf_event_context *ctx;
157 159
158 rcu_read_lock(); 160 rcu_read_lock();
159 retry: 161retry:
160 ctx = rcu_dereference(task->perf_event_ctxp); 162 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
161 if (ctx) { 163 if (ctx) {
162 /* 164 /*
163 * If this context is a clone of another, it might 165 * If this context is a clone of another, it might
@@ -170,7 +172,7 @@ perf_lock_task_context(struct task_struct *task, unsigned long *flags)
170 * can't get swapped on us any more. 172 * can't get swapped on us any more.
171 */ 173 */
172 raw_spin_lock_irqsave(&ctx->lock, *flags); 174 raw_spin_lock_irqsave(&ctx->lock, *flags);
173 if (ctx != rcu_dereference(task->perf_event_ctxp)) { 175 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
174 raw_spin_unlock_irqrestore(&ctx->lock, *flags); 176 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
175 goto retry; 177 goto retry;
176 } 178 }
@@ -189,12 +191,13 @@ perf_lock_task_context(struct task_struct *task, unsigned long *flags)
189 * can't get swapped to another task. This also increments its 191 * can't get swapped to another task. This also increments its
190 * reference count so that the context can't get freed. 192 * reference count so that the context can't get freed.
191 */ 193 */
192static struct perf_event_context *perf_pin_task_context(struct task_struct *task) 194static struct perf_event_context *
195perf_pin_task_context(struct task_struct *task, int ctxn)
193{ 196{
194 struct perf_event_context *ctx; 197 struct perf_event_context *ctx;
195 unsigned long flags; 198 unsigned long flags;
196 199
197 ctx = perf_lock_task_context(task, &flags); 200 ctx = perf_lock_task_context(task, ctxn, &flags);
198 if (ctx) { 201 if (ctx) {
199 ++ctx->pin_count; 202 ++ctx->pin_count;
200 raw_spin_unlock_irqrestore(&ctx->lock, flags); 203 raw_spin_unlock_irqrestore(&ctx->lock, flags);
@@ -302,6 +305,8 @@ list_add_event(struct perf_event *event, struct perf_event_context *ctx)
302 } 305 }
303 306
304 list_add_rcu(&event->event_entry, &ctx->event_list); 307 list_add_rcu(&event->event_entry, &ctx->event_list);
308 if (!ctx->nr_events)
309 perf_pmu_rotate_start(ctx->pmu);
305 ctx->nr_events++; 310 ctx->nr_events++;
306 if (event->attr.inherit_stat) 311 if (event->attr.inherit_stat)
307 ctx->nr_stat++; 312 ctx->nr_stat++;
@@ -311,7 +316,12 @@ static void perf_group_attach(struct perf_event *event)
311{ 316{
312 struct perf_event *group_leader = event->group_leader; 317 struct perf_event *group_leader = event->group_leader;
313 318
314 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_GROUP); 319 /*
320 * We can have double attach due to group movement in perf_event_open.
321 */
322 if (event->attach_state & PERF_ATTACH_GROUP)
323 return;
324
315 event->attach_state |= PERF_ATTACH_GROUP; 325 event->attach_state |= PERF_ATTACH_GROUP;
316 326
317 if (group_leader == event) 327 if (group_leader == event)
@@ -402,11 +412,31 @@ static void perf_group_detach(struct perf_event *event)
402 } 412 }
403} 413}
404 414
415static inline int
416event_filter_match(struct perf_event *event)
417{
418 return event->cpu == -1 || event->cpu == smp_processor_id();
419}
420
405static void 421static void
406event_sched_out(struct perf_event *event, 422event_sched_out(struct perf_event *event,
407 struct perf_cpu_context *cpuctx, 423 struct perf_cpu_context *cpuctx,
408 struct perf_event_context *ctx) 424 struct perf_event_context *ctx)
409{ 425{
426 u64 delta;
427 /*
428 * An event which could not be activated because of
429 * filter mismatch still needs to have its timings
430 * maintained, otherwise bogus information is return
431 * via read() for time_enabled, time_running:
432 */
433 if (event->state == PERF_EVENT_STATE_INACTIVE
434 && !event_filter_match(event)) {
435 delta = ctx->time - event->tstamp_stopped;
436 event->tstamp_running += delta;
437 event->tstamp_stopped = ctx->time;
438 }
439
410 if (event->state != PERF_EVENT_STATE_ACTIVE) 440 if (event->state != PERF_EVENT_STATE_ACTIVE)
411 return; 441 return;
412 442
@@ -416,7 +446,7 @@ event_sched_out(struct perf_event *event,
416 event->state = PERF_EVENT_STATE_OFF; 446 event->state = PERF_EVENT_STATE_OFF;
417 } 447 }
418 event->tstamp_stopped = ctx->time; 448 event->tstamp_stopped = ctx->time;
419 event->pmu->disable(event); 449 event->pmu->del(event, 0);
420 event->oncpu = -1; 450 event->oncpu = -1;
421 451
422 if (!is_software_event(event)) 452 if (!is_software_event(event))
@@ -432,9 +462,7 @@ group_sched_out(struct perf_event *group_event,
432 struct perf_event_context *ctx) 462 struct perf_event_context *ctx)
433{ 463{
434 struct perf_event *event; 464 struct perf_event *event;
435 465 int state = group_event->state;
436 if (group_event->state != PERF_EVENT_STATE_ACTIVE)
437 return;
438 466
439 event_sched_out(group_event, cpuctx, ctx); 467 event_sched_out(group_event, cpuctx, ctx);
440 468
@@ -444,10 +472,16 @@ group_sched_out(struct perf_event *group_event,
444 list_for_each_entry(event, &group_event->sibling_list, group_entry) 472 list_for_each_entry(event, &group_event->sibling_list, group_entry)
445 event_sched_out(event, cpuctx, ctx); 473 event_sched_out(event, cpuctx, ctx);
446 474
447 if (group_event->attr.exclusive) 475 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
448 cpuctx->exclusive = 0; 476 cpuctx->exclusive = 0;
449} 477}
450 478
479static inline struct perf_cpu_context *
480__get_cpu_context(struct perf_event_context *ctx)
481{
482 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
483}
484
451/* 485/*
452 * Cross CPU call to remove a performance event 486 * Cross CPU call to remove a performance event
453 * 487 *
@@ -456,9 +490,9 @@ group_sched_out(struct perf_event *group_event,
456 */ 490 */
457static void __perf_event_remove_from_context(void *info) 491static void __perf_event_remove_from_context(void *info)
458{ 492{
459 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
460 struct perf_event *event = info; 493 struct perf_event *event = info;
461 struct perf_event_context *ctx = event->ctx; 494 struct perf_event_context *ctx = event->ctx;
495 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
462 496
463 /* 497 /*
464 * If this is a task context, we need to check whether it is 498 * If this is a task context, we need to check whether it is
@@ -469,27 +503,11 @@ static void __perf_event_remove_from_context(void *info)
469 return; 503 return;
470 504
471 raw_spin_lock(&ctx->lock); 505 raw_spin_lock(&ctx->lock);
472 /*
473 * Protect the list operation against NMI by disabling the
474 * events on a global level.
475 */
476 perf_disable();
477 506
478 event_sched_out(event, cpuctx, ctx); 507 event_sched_out(event, cpuctx, ctx);
479 508
480 list_del_event(event, ctx); 509 list_del_event(event, ctx);
481 510
482 if (!ctx->task) {
483 /*
484 * Allow more per task events with respect to the
485 * reservation:
486 */
487 cpuctx->max_pertask =
488 min(perf_max_events - ctx->nr_events,
489 perf_max_events - perf_reserved_percpu);
490 }
491
492 perf_enable();
493 raw_spin_unlock(&ctx->lock); 511 raw_spin_unlock(&ctx->lock);
494} 512}
495 513
@@ -554,8 +572,8 @@ retry:
554static void __perf_event_disable(void *info) 572static void __perf_event_disable(void *info)
555{ 573{
556 struct perf_event *event = info; 574 struct perf_event *event = info;
557 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
558 struct perf_event_context *ctx = event->ctx; 575 struct perf_event_context *ctx = event->ctx;
576 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
559 577
560 /* 578 /*
561 * If this is a per-task event, need to check whether this 579 * If this is a per-task event, need to check whether this
@@ -610,7 +628,7 @@ void perf_event_disable(struct perf_event *event)
610 return; 628 return;
611 } 629 }
612 630
613 retry: 631retry:
614 task_oncpu_function_call(task, __perf_event_disable, event); 632 task_oncpu_function_call(task, __perf_event_disable, event);
615 633
616 raw_spin_lock_irq(&ctx->lock); 634 raw_spin_lock_irq(&ctx->lock);
@@ -649,7 +667,7 @@ event_sched_in(struct perf_event *event,
649 */ 667 */
650 smp_wmb(); 668 smp_wmb();
651 669
652 if (event->pmu->enable(event)) { 670 if (event->pmu->add(event, PERF_EF_START)) {
653 event->state = PERF_EVENT_STATE_INACTIVE; 671 event->state = PERF_EVENT_STATE_INACTIVE;
654 event->oncpu = -1; 672 event->oncpu = -1;
655 return -EAGAIN; 673 return -EAGAIN;
@@ -657,6 +675,8 @@ event_sched_in(struct perf_event *event,
657 675
658 event->tstamp_running += ctx->time - event->tstamp_stopped; 676 event->tstamp_running += ctx->time - event->tstamp_stopped;
659 677
678 event->shadow_ctx_time = ctx->time - ctx->timestamp;
679
660 if (!is_software_event(event)) 680 if (!is_software_event(event))
661 cpuctx->active_oncpu++; 681 cpuctx->active_oncpu++;
662 ctx->nr_active++; 682 ctx->nr_active++;
@@ -673,22 +693,17 @@ group_sched_in(struct perf_event *group_event,
673 struct perf_event_context *ctx) 693 struct perf_event_context *ctx)
674{ 694{
675 struct perf_event *event, *partial_group = NULL; 695 struct perf_event *event, *partial_group = NULL;
676 const struct pmu *pmu = group_event->pmu; 696 struct pmu *pmu = group_event->pmu;
677 bool txn = false; 697 u64 now = ctx->time;
698 bool simulate = false;
678 699
679 if (group_event->state == PERF_EVENT_STATE_OFF) 700 if (group_event->state == PERF_EVENT_STATE_OFF)
680 return 0; 701 return 0;
681 702
682 /* Check if group transaction availabe */ 703 pmu->start_txn(pmu);
683 if (pmu->start_txn)
684 txn = true;
685
686 if (txn)
687 pmu->start_txn(pmu);
688 704
689 if (event_sched_in(group_event, cpuctx, ctx)) { 705 if (event_sched_in(group_event, cpuctx, ctx)) {
690 if (txn) 706 pmu->cancel_txn(pmu);
691 pmu->cancel_txn(pmu);
692 return -EAGAIN; 707 return -EAGAIN;
693 } 708 }
694 709
@@ -702,23 +717,38 @@ group_sched_in(struct perf_event *group_event,
702 } 717 }
703 } 718 }
704 719
705 if (!txn || !pmu->commit_txn(pmu)) 720 if (!pmu->commit_txn(pmu))
706 return 0; 721 return 0;
707 722
708group_error: 723group_error:
709 /* 724 /*
710 * Groups can be scheduled in as one unit only, so undo any 725 * Groups can be scheduled in as one unit only, so undo any
711 * partial group before returning: 726 * partial group before returning:
727 * The events up to the failed event are scheduled out normally,
728 * tstamp_stopped will be updated.
729 *
730 * The failed events and the remaining siblings need to have
731 * their timings updated as if they had gone thru event_sched_in()
732 * and event_sched_out(). This is required to get consistent timings
733 * across the group. This also takes care of the case where the group
734 * could never be scheduled by ensuring tstamp_stopped is set to mark
735 * the time the event was actually stopped, such that time delta
736 * calculation in update_event_times() is correct.
712 */ 737 */
713 list_for_each_entry(event, &group_event->sibling_list, group_entry) { 738 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
714 if (event == partial_group) 739 if (event == partial_group)
715 break; 740 simulate = true;
716 event_sched_out(event, cpuctx, ctx); 741
742 if (simulate) {
743 event->tstamp_running += now - event->tstamp_stopped;
744 event->tstamp_stopped = now;
745 } else {
746 event_sched_out(event, cpuctx, ctx);
747 }
717 } 748 }
718 event_sched_out(group_event, cpuctx, ctx); 749 event_sched_out(group_event, cpuctx, ctx);
719 750
720 if (txn) 751 pmu->cancel_txn(pmu);
721 pmu->cancel_txn(pmu);
722 752
723 return -EAGAIN; 753 return -EAGAIN;
724} 754}
@@ -771,10 +801,10 @@ static void add_event_to_ctx(struct perf_event *event,
771 */ 801 */
772static void __perf_install_in_context(void *info) 802static void __perf_install_in_context(void *info)
773{ 803{
774 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
775 struct perf_event *event = info; 804 struct perf_event *event = info;
776 struct perf_event_context *ctx = event->ctx; 805 struct perf_event_context *ctx = event->ctx;
777 struct perf_event *leader = event->group_leader; 806 struct perf_event *leader = event->group_leader;
807 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
778 int err; 808 int err;
779 809
780 /* 810 /*
@@ -794,12 +824,6 @@ static void __perf_install_in_context(void *info)
794 ctx->is_active = 1; 824 ctx->is_active = 1;
795 update_context_time(ctx); 825 update_context_time(ctx);
796 826
797 /*
798 * Protect the list operation against NMI by disabling the
799 * events on a global level. NOP for non NMI based events.
800 */
801 perf_disable();
802
803 add_event_to_ctx(event, ctx); 827 add_event_to_ctx(event, ctx);
804 828
805 if (event->cpu != -1 && event->cpu != smp_processor_id()) 829 if (event->cpu != -1 && event->cpu != smp_processor_id())
@@ -837,12 +861,7 @@ static void __perf_install_in_context(void *info)
837 } 861 }
838 } 862 }
839 863
840 if (!err && !ctx->task && cpuctx->max_pertask) 864unlock:
841 cpuctx->max_pertask--;
842
843 unlock:
844 perf_enable();
845
846 raw_spin_unlock(&ctx->lock); 865 raw_spin_unlock(&ctx->lock);
847} 866}
848 867
@@ -865,6 +884,8 @@ perf_install_in_context(struct perf_event_context *ctx,
865{ 884{
866 struct task_struct *task = ctx->task; 885 struct task_struct *task = ctx->task;
867 886
887 event->ctx = ctx;
888
868 if (!task) { 889 if (!task) {
869 /* 890 /*
870 * Per cpu events are installed via an smp call and 891 * Per cpu events are installed via an smp call and
@@ -913,10 +934,12 @@ static void __perf_event_mark_enabled(struct perf_event *event,
913 934
914 event->state = PERF_EVENT_STATE_INACTIVE; 935 event->state = PERF_EVENT_STATE_INACTIVE;
915 event->tstamp_enabled = ctx->time - event->total_time_enabled; 936 event->tstamp_enabled = ctx->time - event->total_time_enabled;
916 list_for_each_entry(sub, &event->sibling_list, group_entry) 937 list_for_each_entry(sub, &event->sibling_list, group_entry) {
917 if (sub->state >= PERF_EVENT_STATE_INACTIVE) 938 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
918 sub->tstamp_enabled = 939 sub->tstamp_enabled =
919 ctx->time - sub->total_time_enabled; 940 ctx->time - sub->total_time_enabled;
941 }
942 }
920} 943}
921 944
922/* 945/*
@@ -925,9 +948,9 @@ static void __perf_event_mark_enabled(struct perf_event *event,
925static void __perf_event_enable(void *info) 948static void __perf_event_enable(void *info)
926{ 949{
927 struct perf_event *event = info; 950 struct perf_event *event = info;
928 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
929 struct perf_event_context *ctx = event->ctx; 951 struct perf_event_context *ctx = event->ctx;
930 struct perf_event *leader = event->group_leader; 952 struct perf_event *leader = event->group_leader;
953 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
931 int err; 954 int err;
932 955
933 /* 956 /*
@@ -961,12 +984,10 @@ static void __perf_event_enable(void *info)
961 if (!group_can_go_on(event, cpuctx, 1)) { 984 if (!group_can_go_on(event, cpuctx, 1)) {
962 err = -EEXIST; 985 err = -EEXIST;
963 } else { 986 } else {
964 perf_disable();
965 if (event == leader) 987 if (event == leader)
966 err = group_sched_in(event, cpuctx, ctx); 988 err = group_sched_in(event, cpuctx, ctx);
967 else 989 else
968 err = event_sched_in(event, cpuctx, ctx); 990 err = event_sched_in(event, cpuctx, ctx);
969 perf_enable();
970 } 991 }
971 992
972 if (err) { 993 if (err) {
@@ -982,7 +1003,7 @@ static void __perf_event_enable(void *info)
982 } 1003 }
983 } 1004 }
984 1005
985 unlock: 1006unlock:
986 raw_spin_unlock(&ctx->lock); 1007 raw_spin_unlock(&ctx->lock);
987} 1008}
988 1009
@@ -1023,7 +1044,7 @@ void perf_event_enable(struct perf_event *event)
1023 if (event->state == PERF_EVENT_STATE_ERROR) 1044 if (event->state == PERF_EVENT_STATE_ERROR)
1024 event->state = PERF_EVENT_STATE_OFF; 1045 event->state = PERF_EVENT_STATE_OFF;
1025 1046
1026 retry: 1047retry:
1027 raw_spin_unlock_irq(&ctx->lock); 1048 raw_spin_unlock_irq(&ctx->lock);
1028 task_oncpu_function_call(task, __perf_event_enable, event); 1049 task_oncpu_function_call(task, __perf_event_enable, event);
1029 1050
@@ -1043,7 +1064,7 @@ void perf_event_enable(struct perf_event *event)
1043 if (event->state == PERF_EVENT_STATE_OFF) 1064 if (event->state == PERF_EVENT_STATE_OFF)
1044 __perf_event_mark_enabled(event, ctx); 1065 __perf_event_mark_enabled(event, ctx);
1045 1066
1046 out: 1067out:
1047 raw_spin_unlock_irq(&ctx->lock); 1068 raw_spin_unlock_irq(&ctx->lock);
1048} 1069}
1049 1070
@@ -1074,26 +1095,26 @@ static void ctx_sched_out(struct perf_event_context *ctx,
1074 struct perf_event *event; 1095 struct perf_event *event;
1075 1096
1076 raw_spin_lock(&ctx->lock); 1097 raw_spin_lock(&ctx->lock);
1098 perf_pmu_disable(ctx->pmu);
1077 ctx->is_active = 0; 1099 ctx->is_active = 0;
1078 if (likely(!ctx->nr_events)) 1100 if (likely(!ctx->nr_events))
1079 goto out; 1101 goto out;
1080 update_context_time(ctx); 1102 update_context_time(ctx);
1081 1103
1082 perf_disable();
1083 if (!ctx->nr_active) 1104 if (!ctx->nr_active)
1084 goto out_enable; 1105 goto out;
1085 1106
1086 if (event_type & EVENT_PINNED) 1107 if (event_type & EVENT_PINNED) {
1087 list_for_each_entry(event, &ctx->pinned_groups, group_entry) 1108 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1088 group_sched_out(event, cpuctx, ctx); 1109 group_sched_out(event, cpuctx, ctx);
1110 }
1089 1111
1090 if (event_type & EVENT_FLEXIBLE) 1112 if (event_type & EVENT_FLEXIBLE) {
1091 list_for_each_entry(event, &ctx->flexible_groups, group_entry) 1113 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1092 group_sched_out(event, cpuctx, ctx); 1114 group_sched_out(event, cpuctx, ctx);
1093 1115 }
1094 out_enable: 1116out:
1095 perf_enable(); 1117 perf_pmu_enable(ctx->pmu);
1096 out:
1097 raw_spin_unlock(&ctx->lock); 1118 raw_spin_unlock(&ctx->lock);
1098} 1119}
1099 1120
@@ -1191,34 +1212,25 @@ static void perf_event_sync_stat(struct perf_event_context *ctx,
1191 } 1212 }
1192} 1213}
1193 1214
1194/* 1215void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1195 * Called from scheduler to remove the events of the current task, 1216 struct task_struct *next)
1196 * with interrupts disabled.
1197 *
1198 * We stop each event and update the event value in event->count.
1199 *
1200 * This does not protect us against NMI, but disable()
1201 * sets the disabled bit in the control field of event _before_
1202 * accessing the event control register. If a NMI hits, then it will
1203 * not restart the event.
1204 */
1205void perf_event_task_sched_out(struct task_struct *task,
1206 struct task_struct *next)
1207{ 1217{
1208 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); 1218 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1209 struct perf_event_context *ctx = task->perf_event_ctxp;
1210 struct perf_event_context *next_ctx; 1219 struct perf_event_context *next_ctx;
1211 struct perf_event_context *parent; 1220 struct perf_event_context *parent;
1221 struct perf_cpu_context *cpuctx;
1212 int do_switch = 1; 1222 int do_switch = 1;
1213 1223
1214 perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0); 1224 if (likely(!ctx))
1225 return;
1215 1226
1216 if (likely(!ctx || !cpuctx->task_ctx)) 1227 cpuctx = __get_cpu_context(ctx);
1228 if (!cpuctx->task_ctx)
1217 return; 1229 return;
1218 1230
1219 rcu_read_lock(); 1231 rcu_read_lock();
1220 parent = rcu_dereference(ctx->parent_ctx); 1232 parent = rcu_dereference(ctx->parent_ctx);
1221 next_ctx = next->perf_event_ctxp; 1233 next_ctx = next->perf_event_ctxp[ctxn];
1222 if (parent && next_ctx && 1234 if (parent && next_ctx &&
1223 rcu_dereference(next_ctx->parent_ctx) == parent) { 1235 rcu_dereference(next_ctx->parent_ctx) == parent) {
1224 /* 1236 /*
@@ -1237,8 +1249,8 @@ void perf_event_task_sched_out(struct task_struct *task,
1237 * XXX do we need a memory barrier of sorts 1249 * XXX do we need a memory barrier of sorts
1238 * wrt to rcu_dereference() of perf_event_ctxp 1250 * wrt to rcu_dereference() of perf_event_ctxp
1239 */ 1251 */
1240 task->perf_event_ctxp = next_ctx; 1252 task->perf_event_ctxp[ctxn] = next_ctx;
1241 next->perf_event_ctxp = ctx; 1253 next->perf_event_ctxp[ctxn] = ctx;
1242 ctx->task = next; 1254 ctx->task = next;
1243 next_ctx->task = task; 1255 next_ctx->task = task;
1244 do_switch = 0; 1256 do_switch = 0;
@@ -1256,10 +1268,33 @@ void perf_event_task_sched_out(struct task_struct *task,
1256 } 1268 }
1257} 1269}
1258 1270
1271#define for_each_task_context_nr(ctxn) \
1272 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
1273
1274/*
1275 * Called from scheduler to remove the events of the current task,
1276 * with interrupts disabled.
1277 *
1278 * We stop each event and update the event value in event->count.
1279 *
1280 * This does not protect us against NMI, but disable()
1281 * sets the disabled bit in the control field of event _before_
1282 * accessing the event control register. If a NMI hits, then it will
1283 * not restart the event.
1284 */
1285void __perf_event_task_sched_out(struct task_struct *task,
1286 struct task_struct *next)
1287{
1288 int ctxn;
1289
1290 for_each_task_context_nr(ctxn)
1291 perf_event_context_sched_out(task, ctxn, next);
1292}
1293
1259static void task_ctx_sched_out(struct perf_event_context *ctx, 1294static void task_ctx_sched_out(struct perf_event_context *ctx,
1260 enum event_type_t event_type) 1295 enum event_type_t event_type)
1261{ 1296{
1262 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); 1297 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1263 1298
1264 if (!cpuctx->task_ctx) 1299 if (!cpuctx->task_ctx)
1265 return; 1300 return;
@@ -1274,14 +1309,6 @@ static void task_ctx_sched_out(struct perf_event_context *ctx,
1274/* 1309/*
1275 * Called with IRQs disabled 1310 * Called with IRQs disabled
1276 */ 1311 */
1277static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1278{
1279 task_ctx_sched_out(ctx, EVENT_ALL);
1280}
1281
1282/*
1283 * Called with IRQs disabled
1284 */
1285static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, 1312static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1286 enum event_type_t event_type) 1313 enum event_type_t event_type)
1287{ 1314{
@@ -1332,9 +1359,10 @@ ctx_flexible_sched_in(struct perf_event_context *ctx,
1332 if (event->cpu != -1 && event->cpu != smp_processor_id()) 1359 if (event->cpu != -1 && event->cpu != smp_processor_id())
1333 continue; 1360 continue;
1334 1361
1335 if (group_can_go_on(event, cpuctx, can_add_hw)) 1362 if (group_can_go_on(event, cpuctx, can_add_hw)) {
1336 if (group_sched_in(event, cpuctx, ctx)) 1363 if (group_sched_in(event, cpuctx, ctx))
1337 can_add_hw = 0; 1364 can_add_hw = 0;
1365 }
1338 } 1366 }
1339} 1367}
1340 1368
@@ -1350,8 +1378,6 @@ ctx_sched_in(struct perf_event_context *ctx,
1350 1378
1351 ctx->timestamp = perf_clock(); 1379 ctx->timestamp = perf_clock();
1352 1380
1353 perf_disable();
1354
1355 /* 1381 /*
1356 * First go through the list and put on any pinned groups 1382 * First go through the list and put on any pinned groups
1357 * in order to give them the best chance of going on. 1383 * in order to give them the best chance of going on.
@@ -1363,8 +1389,7 @@ ctx_sched_in(struct perf_event_context *ctx,
1363 if (event_type & EVENT_FLEXIBLE) 1389 if (event_type & EVENT_FLEXIBLE)
1364 ctx_flexible_sched_in(ctx, cpuctx); 1390 ctx_flexible_sched_in(ctx, cpuctx);
1365 1391
1366 perf_enable(); 1392out:
1367 out:
1368 raw_spin_unlock(&ctx->lock); 1393 raw_spin_unlock(&ctx->lock);
1369} 1394}
1370 1395
@@ -1376,43 +1401,28 @@ static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1376 ctx_sched_in(ctx, cpuctx, event_type); 1401 ctx_sched_in(ctx, cpuctx, event_type);
1377} 1402}
1378 1403
1379static void task_ctx_sched_in(struct task_struct *task, 1404static void task_ctx_sched_in(struct perf_event_context *ctx,
1380 enum event_type_t event_type) 1405 enum event_type_t event_type)
1381{ 1406{
1382 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); 1407 struct perf_cpu_context *cpuctx;
1383 struct perf_event_context *ctx = task->perf_event_ctxp;
1384 1408
1385 if (likely(!ctx)) 1409 cpuctx = __get_cpu_context(ctx);
1386 return;
1387 if (cpuctx->task_ctx == ctx) 1410 if (cpuctx->task_ctx == ctx)
1388 return; 1411 return;
1412
1389 ctx_sched_in(ctx, cpuctx, event_type); 1413 ctx_sched_in(ctx, cpuctx, event_type);
1390 cpuctx->task_ctx = ctx; 1414 cpuctx->task_ctx = ctx;
1391} 1415}
1392/*
1393 * Called from scheduler to add the events of the current task
1394 * with interrupts disabled.
1395 *
1396 * We restore the event value and then enable it.
1397 *
1398 * This does not protect us against NMI, but enable()
1399 * sets the enabled bit in the control field of event _before_
1400 * accessing the event control register. If a NMI hits, then it will
1401 * keep the event running.
1402 */
1403void perf_event_task_sched_in(struct task_struct *task)
1404{
1405 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1406 struct perf_event_context *ctx = task->perf_event_ctxp;
1407 1416
1408 if (likely(!ctx)) 1417void perf_event_context_sched_in(struct perf_event_context *ctx)
1409 return; 1418{
1419 struct perf_cpu_context *cpuctx;
1410 1420
1421 cpuctx = __get_cpu_context(ctx);
1411 if (cpuctx->task_ctx == ctx) 1422 if (cpuctx->task_ctx == ctx)
1412 return; 1423 return;
1413 1424
1414 perf_disable(); 1425 perf_pmu_disable(ctx->pmu);
1415
1416 /* 1426 /*
1417 * We want to keep the following priority order: 1427 * We want to keep the following priority order:
1418 * cpu pinned (that don't need to move), task pinned, 1428 * cpu pinned (that don't need to move), task pinned,
@@ -1426,7 +1436,37 @@ void perf_event_task_sched_in(struct task_struct *task)
1426 1436
1427 cpuctx->task_ctx = ctx; 1437 cpuctx->task_ctx = ctx;
1428 1438
1429 perf_enable(); 1439 /*
1440 * Since these rotations are per-cpu, we need to ensure the
1441 * cpu-context we got scheduled on is actually rotating.
1442 */
1443 perf_pmu_rotate_start(ctx->pmu);
1444 perf_pmu_enable(ctx->pmu);
1445}
1446
1447/*
1448 * Called from scheduler to add the events of the current task
1449 * with interrupts disabled.
1450 *
1451 * We restore the event value and then enable it.
1452 *
1453 * This does not protect us against NMI, but enable()
1454 * sets the enabled bit in the control field of event _before_
1455 * accessing the event control register. If a NMI hits, then it will
1456 * keep the event running.
1457 */
1458void __perf_event_task_sched_in(struct task_struct *task)
1459{
1460 struct perf_event_context *ctx;
1461 int ctxn;
1462
1463 for_each_task_context_nr(ctxn) {
1464 ctx = task->perf_event_ctxp[ctxn];
1465 if (likely(!ctx))
1466 continue;
1467
1468 perf_event_context_sched_in(ctx);
1469 }
1430} 1470}
1431 1471
1432#define MAX_INTERRUPTS (~0ULL) 1472#define MAX_INTERRUPTS (~0ULL)
@@ -1506,22 +1546,6 @@ do { \
1506 return div64_u64(dividend, divisor); 1546 return div64_u64(dividend, divisor);
1507} 1547}
1508 1548
1509static void perf_event_stop(struct perf_event *event)
1510{
1511 if (!event->pmu->stop)
1512 return event->pmu->disable(event);
1513
1514 return event->pmu->stop(event);
1515}
1516
1517static int perf_event_start(struct perf_event *event)
1518{
1519 if (!event->pmu->start)
1520 return event->pmu->enable(event);
1521
1522 return event->pmu->start(event);
1523}
1524
1525static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count) 1549static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1526{ 1550{
1527 struct hw_perf_event *hwc = &event->hw; 1551 struct hw_perf_event *hwc = &event->hw;
@@ -1541,15 +1565,13 @@ static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1541 hwc->sample_period = sample_period; 1565 hwc->sample_period = sample_period;
1542 1566
1543 if (local64_read(&hwc->period_left) > 8*sample_period) { 1567 if (local64_read(&hwc->period_left) > 8*sample_period) {
1544 perf_disable(); 1568 event->pmu->stop(event, PERF_EF_UPDATE);
1545 perf_event_stop(event);
1546 local64_set(&hwc->period_left, 0); 1569 local64_set(&hwc->period_left, 0);
1547 perf_event_start(event); 1570 event->pmu->start(event, PERF_EF_RELOAD);
1548 perf_enable();
1549 } 1571 }
1550} 1572}
1551 1573
1552static void perf_ctx_adjust_freq(struct perf_event_context *ctx) 1574static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
1553{ 1575{
1554 struct perf_event *event; 1576 struct perf_event *event;
1555 struct hw_perf_event *hwc; 1577 struct hw_perf_event *hwc;
@@ -1574,23 +1596,19 @@ static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1574 */ 1596 */
1575 if (interrupts == MAX_INTERRUPTS) { 1597 if (interrupts == MAX_INTERRUPTS) {
1576 perf_log_throttle(event, 1); 1598 perf_log_throttle(event, 1);
1577 perf_disable(); 1599 event->pmu->start(event, 0);
1578 event->pmu->unthrottle(event);
1579 perf_enable();
1580 } 1600 }
1581 1601
1582 if (!event->attr.freq || !event->attr.sample_freq) 1602 if (!event->attr.freq || !event->attr.sample_freq)
1583 continue; 1603 continue;
1584 1604
1585 perf_disable();
1586 event->pmu->read(event); 1605 event->pmu->read(event);
1587 now = local64_read(&event->count); 1606 now = local64_read(&event->count);
1588 delta = now - hwc->freq_count_stamp; 1607 delta = now - hwc->freq_count_stamp;
1589 hwc->freq_count_stamp = now; 1608 hwc->freq_count_stamp = now;
1590 1609
1591 if (delta > 0) 1610 if (delta > 0)
1592 perf_adjust_period(event, TICK_NSEC, delta); 1611 perf_adjust_period(event, period, delta);
1593 perf_enable();
1594 } 1612 }
1595 raw_spin_unlock(&ctx->lock); 1613 raw_spin_unlock(&ctx->lock);
1596} 1614}
@@ -1602,38 +1620,48 @@ static void rotate_ctx(struct perf_event_context *ctx)
1602{ 1620{
1603 raw_spin_lock(&ctx->lock); 1621 raw_spin_lock(&ctx->lock);
1604 1622
1605 /* Rotate the first entry last of non-pinned groups */ 1623 /*
1606 list_rotate_left(&ctx->flexible_groups); 1624 * Rotate the first entry last of non-pinned groups. Rotation might be
1625 * disabled by the inheritance code.
1626 */
1627 if (!ctx->rotate_disable)
1628 list_rotate_left(&ctx->flexible_groups);
1607 1629
1608 raw_spin_unlock(&ctx->lock); 1630 raw_spin_unlock(&ctx->lock);
1609} 1631}
1610 1632
1611void perf_event_task_tick(struct task_struct *curr) 1633/*
1634 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
1635 * because they're strictly cpu affine and rotate_start is called with IRQs
1636 * disabled, while rotate_context is called from IRQ context.
1637 */
1638static void perf_rotate_context(struct perf_cpu_context *cpuctx)
1612{ 1639{
1613 struct perf_cpu_context *cpuctx; 1640 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
1614 struct perf_event_context *ctx; 1641 struct perf_event_context *ctx = NULL;
1615 int rotate = 0; 1642 int rotate = 0, remove = 1;
1616 1643
1617 if (!atomic_read(&nr_events)) 1644 if (cpuctx->ctx.nr_events) {
1618 return; 1645 remove = 0;
1619 1646 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1620 cpuctx = &__get_cpu_var(perf_cpu_context); 1647 rotate = 1;
1621 if (cpuctx->ctx.nr_events && 1648 }
1622 cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1623 rotate = 1;
1624 1649
1625 ctx = curr->perf_event_ctxp; 1650 ctx = cpuctx->task_ctx;
1626 if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active) 1651 if (ctx && ctx->nr_events) {
1627 rotate = 1; 1652 remove = 0;
1653 if (ctx->nr_events != ctx->nr_active)
1654 rotate = 1;
1655 }
1628 1656
1629 perf_ctx_adjust_freq(&cpuctx->ctx); 1657 perf_pmu_disable(cpuctx->ctx.pmu);
1658 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
1630 if (ctx) 1659 if (ctx)
1631 perf_ctx_adjust_freq(ctx); 1660 perf_ctx_adjust_freq(ctx, interval);
1632 1661
1633 if (!rotate) 1662 if (!rotate)
1634 return; 1663 goto done;
1635 1664
1636 perf_disable();
1637 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); 1665 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1638 if (ctx) 1666 if (ctx)
1639 task_ctx_sched_out(ctx, EVENT_FLEXIBLE); 1667 task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
@@ -1644,8 +1672,27 @@ void perf_event_task_tick(struct task_struct *curr)
1644 1672
1645 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE); 1673 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1646 if (ctx) 1674 if (ctx)
1647 task_ctx_sched_in(curr, EVENT_FLEXIBLE); 1675 task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
1648 perf_enable(); 1676
1677done:
1678 if (remove)
1679 list_del_init(&cpuctx->rotation_list);
1680
1681 perf_pmu_enable(cpuctx->ctx.pmu);
1682}
1683
1684void perf_event_task_tick(void)
1685{
1686 struct list_head *head = &__get_cpu_var(rotation_list);
1687 struct perf_cpu_context *cpuctx, *tmp;
1688
1689 WARN_ON(!irqs_disabled());
1690
1691 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
1692 if (cpuctx->jiffies_interval == 1 ||
1693 !(jiffies % cpuctx->jiffies_interval))
1694 perf_rotate_context(cpuctx);
1695 }
1649} 1696}
1650 1697
1651static int event_enable_on_exec(struct perf_event *event, 1698static int event_enable_on_exec(struct perf_event *event,
@@ -1667,20 +1714,18 @@ static int event_enable_on_exec(struct perf_event *event,
1667 * Enable all of a task's events that have been marked enable-on-exec. 1714 * Enable all of a task's events that have been marked enable-on-exec.
1668 * This expects task == current. 1715 * This expects task == current.
1669 */ 1716 */
1670static void perf_event_enable_on_exec(struct task_struct *task) 1717static void perf_event_enable_on_exec(struct perf_event_context *ctx)
1671{ 1718{
1672 struct perf_event_context *ctx;
1673 struct perf_event *event; 1719 struct perf_event *event;
1674 unsigned long flags; 1720 unsigned long flags;
1675 int enabled = 0; 1721 int enabled = 0;
1676 int ret; 1722 int ret;
1677 1723
1678 local_irq_save(flags); 1724 local_irq_save(flags);
1679 ctx = task->perf_event_ctxp;
1680 if (!ctx || !ctx->nr_events) 1725 if (!ctx || !ctx->nr_events)
1681 goto out; 1726 goto out;
1682 1727
1683 __perf_event_task_sched_out(ctx); 1728 task_ctx_sched_out(ctx, EVENT_ALL);
1684 1729
1685 raw_spin_lock(&ctx->lock); 1730 raw_spin_lock(&ctx->lock);
1686 1731
@@ -1704,8 +1749,8 @@ static void perf_event_enable_on_exec(struct task_struct *task)
1704 1749
1705 raw_spin_unlock(&ctx->lock); 1750 raw_spin_unlock(&ctx->lock);
1706 1751
1707 perf_event_task_sched_in(task); 1752 perf_event_context_sched_in(ctx);
1708 out: 1753out:
1709 local_irq_restore(flags); 1754 local_irq_restore(flags);
1710} 1755}
1711 1756
@@ -1714,9 +1759,9 @@ static void perf_event_enable_on_exec(struct task_struct *task)
1714 */ 1759 */
1715static void __perf_event_read(void *info) 1760static void __perf_event_read(void *info)
1716{ 1761{
1717 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1718 struct perf_event *event = info; 1762 struct perf_event *event = info;
1719 struct perf_event_context *ctx = event->ctx; 1763 struct perf_event_context *ctx = event->ctx;
1764 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1720 1765
1721 /* 1766 /*
1722 * If this is a task context, we need to check whether it is 1767 * If this is a task context, we need to check whether it is
@@ -1755,7 +1800,13 @@ static u64 perf_event_read(struct perf_event *event)
1755 unsigned long flags; 1800 unsigned long flags;
1756 1801
1757 raw_spin_lock_irqsave(&ctx->lock, flags); 1802 raw_spin_lock_irqsave(&ctx->lock, flags);
1758 update_context_time(ctx); 1803 /*
1804 * may read while context is not active
1805 * (e.g., thread is blocked), in that case
1806 * we cannot update context time
1807 */
1808 if (ctx->is_active)
1809 update_context_time(ctx);
1759 update_event_times(event); 1810 update_event_times(event);
1760 raw_spin_unlock_irqrestore(&ctx->lock, flags); 1811 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1761 } 1812 }
@@ -1764,11 +1815,219 @@ static u64 perf_event_read(struct perf_event *event)
1764} 1815}
1765 1816
1766/* 1817/*
1767 * Initialize the perf_event context in a task_struct: 1818 * Callchain support
1768 */ 1819 */
1820
1821struct callchain_cpus_entries {
1822 struct rcu_head rcu_head;
1823 struct perf_callchain_entry *cpu_entries[0];
1824};
1825
1826static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
1827static atomic_t nr_callchain_events;
1828static DEFINE_MUTEX(callchain_mutex);
1829struct callchain_cpus_entries *callchain_cpus_entries;
1830
1831
1832__weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
1833 struct pt_regs *regs)
1834{
1835}
1836
1837__weak void perf_callchain_user(struct perf_callchain_entry *entry,
1838 struct pt_regs *regs)
1839{
1840}
1841
1842static void release_callchain_buffers_rcu(struct rcu_head *head)
1843{
1844 struct callchain_cpus_entries *entries;
1845 int cpu;
1846
1847 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
1848
1849 for_each_possible_cpu(cpu)
1850 kfree(entries->cpu_entries[cpu]);
1851
1852 kfree(entries);
1853}
1854
1855static void release_callchain_buffers(void)
1856{
1857 struct callchain_cpus_entries *entries;
1858
1859 entries = callchain_cpus_entries;
1860 rcu_assign_pointer(callchain_cpus_entries, NULL);
1861 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
1862}
1863
1864static int alloc_callchain_buffers(void)
1865{
1866 int cpu;
1867 int size;
1868 struct callchain_cpus_entries *entries;
1869
1870 /*
1871 * We can't use the percpu allocation API for data that can be
1872 * accessed from NMI. Use a temporary manual per cpu allocation
1873 * until that gets sorted out.
1874 */
1875 size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
1876 num_possible_cpus();
1877
1878 entries = kzalloc(size, GFP_KERNEL);
1879 if (!entries)
1880 return -ENOMEM;
1881
1882 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
1883
1884 for_each_possible_cpu(cpu) {
1885 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
1886 cpu_to_node(cpu));
1887 if (!entries->cpu_entries[cpu])
1888 goto fail;
1889 }
1890
1891 rcu_assign_pointer(callchain_cpus_entries, entries);
1892
1893 return 0;
1894
1895fail:
1896 for_each_possible_cpu(cpu)
1897 kfree(entries->cpu_entries[cpu]);
1898 kfree(entries);
1899
1900 return -ENOMEM;
1901}
1902
1903static int get_callchain_buffers(void)
1904{
1905 int err = 0;
1906 int count;
1907
1908 mutex_lock(&callchain_mutex);
1909
1910 count = atomic_inc_return(&nr_callchain_events);
1911 if (WARN_ON_ONCE(count < 1)) {
1912 err = -EINVAL;
1913 goto exit;
1914 }
1915
1916 if (count > 1) {
1917 /* If the allocation failed, give up */
1918 if (!callchain_cpus_entries)
1919 err = -ENOMEM;
1920 goto exit;
1921 }
1922
1923 err = alloc_callchain_buffers();
1924 if (err)
1925 release_callchain_buffers();
1926exit:
1927 mutex_unlock(&callchain_mutex);
1928
1929 return err;
1930}
1931
1932static void put_callchain_buffers(void)
1933{
1934 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
1935 release_callchain_buffers();
1936 mutex_unlock(&callchain_mutex);
1937 }
1938}
1939
1940static int get_recursion_context(int *recursion)
1941{
1942 int rctx;
1943
1944 if (in_nmi())
1945 rctx = 3;
1946 else if (in_irq())
1947 rctx = 2;
1948 else if (in_softirq())
1949 rctx = 1;
1950 else
1951 rctx = 0;
1952
1953 if (recursion[rctx])
1954 return -1;
1955
1956 recursion[rctx]++;
1957 barrier();
1958
1959 return rctx;
1960}
1961
1962static inline void put_recursion_context(int *recursion, int rctx)
1963{
1964 barrier();
1965 recursion[rctx]--;
1966}
1967
1968static struct perf_callchain_entry *get_callchain_entry(int *rctx)
1969{
1970 int cpu;
1971 struct callchain_cpus_entries *entries;
1972
1973 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
1974 if (*rctx == -1)
1975 return NULL;
1976
1977 entries = rcu_dereference(callchain_cpus_entries);
1978 if (!entries)
1979 return NULL;
1980
1981 cpu = smp_processor_id();
1982
1983 return &entries->cpu_entries[cpu][*rctx];
1984}
1985
1769static void 1986static void
1770__perf_event_init_context(struct perf_event_context *ctx, 1987put_callchain_entry(int rctx)
1771 struct task_struct *task) 1988{
1989 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
1990}
1991
1992static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
1993{
1994 int rctx;
1995 struct perf_callchain_entry *entry;
1996
1997
1998 entry = get_callchain_entry(&rctx);
1999 if (rctx == -1)
2000 return NULL;
2001
2002 if (!entry)
2003 goto exit_put;
2004
2005 entry->nr = 0;
2006
2007 if (!user_mode(regs)) {
2008 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2009 perf_callchain_kernel(entry, regs);
2010 if (current->mm)
2011 regs = task_pt_regs(current);
2012 else
2013 regs = NULL;
2014 }
2015
2016 if (regs) {
2017 perf_callchain_store(entry, PERF_CONTEXT_USER);
2018 perf_callchain_user(entry, regs);
2019 }
2020
2021exit_put:
2022 put_callchain_entry(rctx);
2023
2024 return entry;
2025}
2026
2027/*
2028 * Initialize the perf_event context in a task_struct:
2029 */
2030static void __perf_event_init_context(struct perf_event_context *ctx)
1772{ 2031{
1773 raw_spin_lock_init(&ctx->lock); 2032 raw_spin_lock_init(&ctx->lock);
1774 mutex_init(&ctx->mutex); 2033 mutex_init(&ctx->mutex);
@@ -1776,45 +2035,38 @@ __perf_event_init_context(struct perf_event_context *ctx,
1776 INIT_LIST_HEAD(&ctx->flexible_groups); 2035 INIT_LIST_HEAD(&ctx->flexible_groups);
1777 INIT_LIST_HEAD(&ctx->event_list); 2036 INIT_LIST_HEAD(&ctx->event_list);
1778 atomic_set(&ctx->refcount, 1); 2037 atomic_set(&ctx->refcount, 1);
1779 ctx->task = task;
1780} 2038}
1781 2039
1782static struct perf_event_context *find_get_context(pid_t pid, int cpu) 2040static struct perf_event_context *
2041alloc_perf_context(struct pmu *pmu, struct task_struct *task)
1783{ 2042{
1784 struct perf_event_context *ctx; 2043 struct perf_event_context *ctx;
1785 struct perf_cpu_context *cpuctx;
1786 struct task_struct *task;
1787 unsigned long flags;
1788 int err;
1789 2044
1790 if (pid == -1 && cpu != -1) { 2045 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1791 /* Must be root to operate on a CPU event: */ 2046 if (!ctx)
1792 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) 2047 return NULL;
1793 return ERR_PTR(-EACCES);
1794
1795 if (cpu < 0 || cpu >= nr_cpumask_bits)
1796 return ERR_PTR(-EINVAL);
1797 2048
1798 /* 2049 __perf_event_init_context(ctx);
1799 * We could be clever and allow to attach a event to an 2050 if (task) {
1800 * offline CPU and activate it when the CPU comes up, but 2051 ctx->task = task;
1801 * that's for later. 2052 get_task_struct(task);
1802 */ 2053 }
1803 if (!cpu_online(cpu)) 2054 ctx->pmu = pmu;
1804 return ERR_PTR(-ENODEV);
1805 2055
1806 cpuctx = &per_cpu(perf_cpu_context, cpu); 2056 return ctx;
1807 ctx = &cpuctx->ctx; 2057}
1808 get_ctx(ctx);
1809 2058
1810 return ctx; 2059static struct task_struct *
1811 } 2060find_lively_task_by_vpid(pid_t vpid)
2061{
2062 struct task_struct *task;
2063 int err;
1812 2064
1813 rcu_read_lock(); 2065 rcu_read_lock();
1814 if (!pid) 2066 if (!vpid)
1815 task = current; 2067 task = current;
1816 else 2068 else
1817 task = find_task_by_vpid(pid); 2069 task = find_task_by_vpid(vpid);
1818 if (task) 2070 if (task)
1819 get_task_struct(task); 2071 get_task_struct(task);
1820 rcu_read_unlock(); 2072 rcu_read_unlock();
@@ -1834,36 +2086,78 @@ static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1834 if (!ptrace_may_access(task, PTRACE_MODE_READ)) 2086 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1835 goto errout; 2087 goto errout;
1836 2088
1837 retry: 2089 return task;
1838 ctx = perf_lock_task_context(task, &flags); 2090errout:
2091 put_task_struct(task);
2092 return ERR_PTR(err);
2093
2094}
2095
2096static struct perf_event_context *
2097find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2098{
2099 struct perf_event_context *ctx;
2100 struct perf_cpu_context *cpuctx;
2101 unsigned long flags;
2102 int ctxn, err;
2103
2104 if (!task && cpu != -1) {
2105 /* Must be root to operate on a CPU event: */
2106 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2107 return ERR_PTR(-EACCES);
2108
2109 if (cpu < 0 || cpu >= nr_cpumask_bits)
2110 return ERR_PTR(-EINVAL);
2111
2112 /*
2113 * We could be clever and allow to attach a event to an
2114 * offline CPU and activate it when the CPU comes up, but
2115 * that's for later.
2116 */
2117 if (!cpu_online(cpu))
2118 return ERR_PTR(-ENODEV);
2119
2120 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2121 ctx = &cpuctx->ctx;
2122 get_ctx(ctx);
2123
2124 return ctx;
2125 }
2126
2127 err = -EINVAL;
2128 ctxn = pmu->task_ctx_nr;
2129 if (ctxn < 0)
2130 goto errout;
2131
2132retry:
2133 ctx = perf_lock_task_context(task, ctxn, &flags);
1839 if (ctx) { 2134 if (ctx) {
1840 unclone_ctx(ctx); 2135 unclone_ctx(ctx);
1841 raw_spin_unlock_irqrestore(&ctx->lock, flags); 2136 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1842 } 2137 }
1843 2138
1844 if (!ctx) { 2139 if (!ctx) {
1845 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); 2140 ctx = alloc_perf_context(pmu, task);
1846 err = -ENOMEM; 2141 err = -ENOMEM;
1847 if (!ctx) 2142 if (!ctx)
1848 goto errout; 2143 goto errout;
1849 __perf_event_init_context(ctx, task); 2144
1850 get_ctx(ctx); 2145 get_ctx(ctx);
1851 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) { 2146
2147 if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
1852 /* 2148 /*
1853 * We raced with some other task; use 2149 * We raced with some other task; use
1854 * the context they set. 2150 * the context they set.
1855 */ 2151 */
2152 put_task_struct(task);
1856 kfree(ctx); 2153 kfree(ctx);
1857 goto retry; 2154 goto retry;
1858 } 2155 }
1859 get_task_struct(task);
1860 } 2156 }
1861 2157
1862 put_task_struct(task);
1863 return ctx; 2158 return ctx;
1864 2159
1865 errout: 2160errout:
1866 put_task_struct(task);
1867 return ERR_PTR(err); 2161 return ERR_PTR(err);
1868} 2162}
1869 2163
@@ -1880,21 +2174,23 @@ static void free_event_rcu(struct rcu_head *head)
1880 kfree(event); 2174 kfree(event);
1881} 2175}
1882 2176
1883static void perf_pending_sync(struct perf_event *event);
1884static void perf_buffer_put(struct perf_buffer *buffer); 2177static void perf_buffer_put(struct perf_buffer *buffer);
1885 2178
1886static void free_event(struct perf_event *event) 2179static void free_event(struct perf_event *event)
1887{ 2180{
1888 perf_pending_sync(event); 2181 irq_work_sync(&event->pending);
1889 2182
1890 if (!event->parent) { 2183 if (!event->parent) {
1891 atomic_dec(&nr_events); 2184 if (event->attach_state & PERF_ATTACH_TASK)
2185 jump_label_dec(&perf_task_events);
1892 if (event->attr.mmap || event->attr.mmap_data) 2186 if (event->attr.mmap || event->attr.mmap_data)
1893 atomic_dec(&nr_mmap_events); 2187 atomic_dec(&nr_mmap_events);
1894 if (event->attr.comm) 2188 if (event->attr.comm)
1895 atomic_dec(&nr_comm_events); 2189 atomic_dec(&nr_comm_events);
1896 if (event->attr.task) 2190 if (event->attr.task)
1897 atomic_dec(&nr_task_events); 2191 atomic_dec(&nr_task_events);
2192 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2193 put_callchain_buffers();
1898 } 2194 }
1899 2195
1900 if (event->buffer) { 2196 if (event->buffer) {
@@ -1905,7 +2201,9 @@ static void free_event(struct perf_event *event)
1905 if (event->destroy) 2201 if (event->destroy)
1906 event->destroy(event); 2202 event->destroy(event);
1907 2203
1908 put_ctx(event->ctx); 2204 if (event->ctx)
2205 put_ctx(event->ctx);
2206
1909 call_rcu(&event->rcu_head, free_event_rcu); 2207 call_rcu(&event->rcu_head, free_event_rcu);
1910} 2208}
1911 2209
@@ -1939,11 +2237,6 @@ int perf_event_release_kernel(struct perf_event *event)
1939 raw_spin_unlock_irq(&ctx->lock); 2237 raw_spin_unlock_irq(&ctx->lock);
1940 mutex_unlock(&ctx->mutex); 2238 mutex_unlock(&ctx->mutex);
1941 2239
1942 mutex_lock(&event->owner->perf_event_mutex);
1943 list_del_init(&event->owner_entry);
1944 mutex_unlock(&event->owner->perf_event_mutex);
1945 put_task_struct(event->owner);
1946
1947 free_event(event); 2240 free_event(event);
1948 2241
1949 return 0; 2242 return 0;
@@ -1956,9 +2249,43 @@ EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1956static int perf_release(struct inode *inode, struct file *file) 2249static int perf_release(struct inode *inode, struct file *file)
1957{ 2250{
1958 struct perf_event *event = file->private_data; 2251 struct perf_event *event = file->private_data;
2252 struct task_struct *owner;
1959 2253
1960 file->private_data = NULL; 2254 file->private_data = NULL;
1961 2255
2256 rcu_read_lock();
2257 owner = ACCESS_ONCE(event->owner);
2258 /*
2259 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2260 * !owner it means the list deletion is complete and we can indeed
2261 * free this event, otherwise we need to serialize on
2262 * owner->perf_event_mutex.
2263 */
2264 smp_read_barrier_depends();
2265 if (owner) {
2266 /*
2267 * Since delayed_put_task_struct() also drops the last
2268 * task reference we can safely take a new reference
2269 * while holding the rcu_read_lock().
2270 */
2271 get_task_struct(owner);
2272 }
2273 rcu_read_unlock();
2274
2275 if (owner) {
2276 mutex_lock(&owner->perf_event_mutex);
2277 /*
2278 * We have to re-check the event->owner field, if it is cleared
2279 * we raced with perf_event_exit_task(), acquiring the mutex
2280 * ensured they're done, and we can proceed with freeing the
2281 * event.
2282 */
2283 if (event->owner)
2284 list_del_init(&event->owner_entry);
2285 mutex_unlock(&owner->perf_event_mutex);
2286 put_task_struct(owner);
2287 }
2288
1962 return perf_event_release_kernel(event); 2289 return perf_event_release_kernel(event);
1963} 2290}
1964 2291
@@ -2184,15 +2511,13 @@ static void perf_event_for_each(struct perf_event *event,
2184static int perf_event_period(struct perf_event *event, u64 __user *arg) 2511static int perf_event_period(struct perf_event *event, u64 __user *arg)
2185{ 2512{
2186 struct perf_event_context *ctx = event->ctx; 2513 struct perf_event_context *ctx = event->ctx;
2187 unsigned long size;
2188 int ret = 0; 2514 int ret = 0;
2189 u64 value; 2515 u64 value;
2190 2516
2191 if (!event->attr.sample_period) 2517 if (!event->attr.sample_period)
2192 return -EINVAL; 2518 return -EINVAL;
2193 2519
2194 size = copy_from_user(&value, arg, sizeof(value)); 2520 if (copy_from_user(&value, arg, sizeof(value)))
2195 if (size != sizeof(value))
2196 return -EFAULT; 2521 return -EFAULT;
2197 2522
2198 if (!value) 2523 if (!value)
@@ -2326,6 +2651,9 @@ int perf_event_task_disable(void)
2326 2651
2327static int perf_event_index(struct perf_event *event) 2652static int perf_event_index(struct perf_event *event)
2328{ 2653{
2654 if (event->hw.state & PERF_HES_STOPPED)
2655 return 0;
2656
2329 if (event->state != PERF_EVENT_STATE_ACTIVE) 2657 if (event->state != PERF_EVENT_STATE_ACTIVE)
2330 return 0; 2658 return 0;
2331 2659
@@ -2829,16 +3157,7 @@ void perf_event_wakeup(struct perf_event *event)
2829 } 3157 }
2830} 3158}
2831 3159
2832/* 3160static void perf_pending_event(struct irq_work *entry)
2833 * Pending wakeups
2834 *
2835 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2836 *
2837 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2838 * single linked list and use cmpxchg() to add entries lockless.
2839 */
2840
2841static void perf_pending_event(struct perf_pending_entry *entry)
2842{ 3161{
2843 struct perf_event *event = container_of(entry, 3162 struct perf_event *event = container_of(entry,
2844 struct perf_event, pending); 3163 struct perf_event, pending);
@@ -2854,99 +3173,6 @@ static void perf_pending_event(struct perf_pending_entry *entry)
2854 } 3173 }
2855} 3174}
2856 3175
2857#define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2858
2859static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2860 PENDING_TAIL,
2861};
2862
2863static void perf_pending_queue(struct perf_pending_entry *entry,
2864 void (*func)(struct perf_pending_entry *))
2865{
2866 struct perf_pending_entry **head;
2867
2868 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2869 return;
2870
2871 entry->func = func;
2872
2873 head = &get_cpu_var(perf_pending_head);
2874
2875 do {
2876 entry->next = *head;
2877 } while (cmpxchg(head, entry->next, entry) != entry->next);
2878
2879 set_perf_event_pending();
2880
2881 put_cpu_var(perf_pending_head);
2882}
2883
2884static int __perf_pending_run(void)
2885{
2886 struct perf_pending_entry *list;
2887 int nr = 0;
2888
2889 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2890 while (list != PENDING_TAIL) {
2891 void (*func)(struct perf_pending_entry *);
2892 struct perf_pending_entry *entry = list;
2893
2894 list = list->next;
2895
2896 func = entry->func;
2897 entry->next = NULL;
2898 /*
2899 * Ensure we observe the unqueue before we issue the wakeup,
2900 * so that we won't be waiting forever.
2901 * -- see perf_not_pending().
2902 */
2903 smp_wmb();
2904
2905 func(entry);
2906 nr++;
2907 }
2908
2909 return nr;
2910}
2911
2912static inline int perf_not_pending(struct perf_event *event)
2913{
2914 /*
2915 * If we flush on whatever cpu we run, there is a chance we don't
2916 * need to wait.
2917 */
2918 get_cpu();
2919 __perf_pending_run();
2920 put_cpu();
2921
2922 /*
2923 * Ensure we see the proper queue state before going to sleep
2924 * so that we do not miss the wakeup. -- see perf_pending_handle()
2925 */
2926 smp_rmb();
2927 return event->pending.next == NULL;
2928}
2929
2930static void perf_pending_sync(struct perf_event *event)
2931{
2932 wait_event(event->waitq, perf_not_pending(event));
2933}
2934
2935void perf_event_do_pending(void)
2936{
2937 __perf_pending_run();
2938}
2939
2940/*
2941 * Callchain support -- arch specific
2942 */
2943
2944__weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2945{
2946 return NULL;
2947}
2948
2949
2950/* 3176/*
2951 * We assume there is only KVM supporting the callbacks. 3177 * We assume there is only KVM supporting the callbacks.
2952 * Later on, we might change it to a list if there is 3178 * Later on, we might change it to a list if there is
@@ -2996,8 +3222,7 @@ static void perf_output_wakeup(struct perf_output_handle *handle)
2996 3222
2997 if (handle->nmi) { 3223 if (handle->nmi) {
2998 handle->event->pending_wakeup = 1; 3224 handle->event->pending_wakeup = 1;
2999 perf_pending_queue(&handle->event->pending, 3225 irq_work_queue(&handle->event->pending);
3000 perf_pending_event);
3001 } else 3226 } else
3002 perf_event_wakeup(handle->event); 3227 perf_event_wakeup(handle->event);
3003} 3228}
@@ -3053,7 +3278,7 @@ again:
3053 if (handle->wakeup != local_read(&buffer->wakeup)) 3278 if (handle->wakeup != local_read(&buffer->wakeup))
3054 perf_output_wakeup(handle); 3279 perf_output_wakeup(handle);
3055 3280
3056 out: 3281out:
3057 preempt_enable(); 3282 preempt_enable();
3058} 3283}
3059 3284
@@ -3205,7 +3430,8 @@ static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
3205} 3430}
3206 3431
3207static void perf_output_read_one(struct perf_output_handle *handle, 3432static void perf_output_read_one(struct perf_output_handle *handle,
3208 struct perf_event *event) 3433 struct perf_event *event,
3434 u64 enabled, u64 running)
3209{ 3435{
3210 u64 read_format = event->attr.read_format; 3436 u64 read_format = event->attr.read_format;
3211 u64 values[4]; 3437 u64 values[4];
@@ -3213,11 +3439,11 @@ static void perf_output_read_one(struct perf_output_handle *handle,
3213 3439
3214 values[n++] = perf_event_count(event); 3440 values[n++] = perf_event_count(event);
3215 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 3441 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3216 values[n++] = event->total_time_enabled + 3442 values[n++] = enabled +
3217 atomic64_read(&event->child_total_time_enabled); 3443 atomic64_read(&event->child_total_time_enabled);
3218 } 3444 }
3219 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 3445 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3220 values[n++] = event->total_time_running + 3446 values[n++] = running +
3221 atomic64_read(&event->child_total_time_running); 3447 atomic64_read(&event->child_total_time_running);
3222 } 3448 }
3223 if (read_format & PERF_FORMAT_ID) 3449 if (read_format & PERF_FORMAT_ID)
@@ -3230,7 +3456,8 @@ static void perf_output_read_one(struct perf_output_handle *handle,
3230 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. 3456 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3231 */ 3457 */
3232static void perf_output_read_group(struct perf_output_handle *handle, 3458static void perf_output_read_group(struct perf_output_handle *handle,
3233 struct perf_event *event) 3459 struct perf_event *event,
3460 u64 enabled, u64 running)
3234{ 3461{
3235 struct perf_event *leader = event->group_leader, *sub; 3462 struct perf_event *leader = event->group_leader, *sub;
3236 u64 read_format = event->attr.read_format; 3463 u64 read_format = event->attr.read_format;
@@ -3240,10 +3467,10 @@ static void perf_output_read_group(struct perf_output_handle *handle,
3240 values[n++] = 1 + leader->nr_siblings; 3467 values[n++] = 1 + leader->nr_siblings;
3241 3468
3242 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 3469 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3243 values[n++] = leader->total_time_enabled; 3470 values[n++] = enabled;
3244 3471
3245 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 3472 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3246 values[n++] = leader->total_time_running; 3473 values[n++] = running;
3247 3474
3248 if (leader != event) 3475 if (leader != event)
3249 leader->pmu->read(leader); 3476 leader->pmu->read(leader);
@@ -3268,13 +3495,35 @@ static void perf_output_read_group(struct perf_output_handle *handle,
3268 } 3495 }
3269} 3496}
3270 3497
3498#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3499 PERF_FORMAT_TOTAL_TIME_RUNNING)
3500
3271static void perf_output_read(struct perf_output_handle *handle, 3501static void perf_output_read(struct perf_output_handle *handle,
3272 struct perf_event *event) 3502 struct perf_event *event)
3273{ 3503{
3504 u64 enabled = 0, running = 0, now, ctx_time;
3505 u64 read_format = event->attr.read_format;
3506
3507 /*
3508 * compute total_time_enabled, total_time_running
3509 * based on snapshot values taken when the event
3510 * was last scheduled in.
3511 *
3512 * we cannot simply called update_context_time()
3513 * because of locking issue as we are called in
3514 * NMI context
3515 */
3516 if (read_format & PERF_FORMAT_TOTAL_TIMES) {
3517 now = perf_clock();
3518 ctx_time = event->shadow_ctx_time + now;
3519 enabled = ctx_time - event->tstamp_enabled;
3520 running = ctx_time - event->tstamp_running;
3521 }
3522
3274 if (event->attr.read_format & PERF_FORMAT_GROUP) 3523 if (event->attr.read_format & PERF_FORMAT_GROUP)
3275 perf_output_read_group(handle, event); 3524 perf_output_read_group(handle, event, enabled, running);
3276 else 3525 else
3277 perf_output_read_one(handle, event); 3526 perf_output_read_one(handle, event, enabled, running);
3278} 3527}
3279 3528
3280void perf_output_sample(struct perf_output_handle *handle, 3529void perf_output_sample(struct perf_output_handle *handle,
@@ -3441,14 +3690,20 @@ static void perf_event_output(struct perf_event *event, int nmi,
3441 struct perf_output_handle handle; 3690 struct perf_output_handle handle;
3442 struct perf_event_header header; 3691 struct perf_event_header header;
3443 3692
3693 /* protect the callchain buffers */
3694 rcu_read_lock();
3695
3444 perf_prepare_sample(&header, data, event, regs); 3696 perf_prepare_sample(&header, data, event, regs);
3445 3697
3446 if (perf_output_begin(&handle, event, header.size, nmi, 1)) 3698 if (perf_output_begin(&handle, event, header.size, nmi, 1))
3447 return; 3699 goto exit;
3448 3700
3449 perf_output_sample(&handle, &header, data, event); 3701 perf_output_sample(&handle, &header, data, event);
3450 3702
3451 perf_output_end(&handle); 3703 perf_output_end(&handle);
3704
3705exit:
3706 rcu_read_unlock();
3452} 3707}
3453 3708
3454/* 3709/*
@@ -3562,16 +3817,29 @@ static void perf_event_task_ctx(struct perf_event_context *ctx,
3562static void perf_event_task_event(struct perf_task_event *task_event) 3817static void perf_event_task_event(struct perf_task_event *task_event)
3563{ 3818{
3564 struct perf_cpu_context *cpuctx; 3819 struct perf_cpu_context *cpuctx;
3565 struct perf_event_context *ctx = task_event->task_ctx; 3820 struct perf_event_context *ctx;
3821 struct pmu *pmu;
3822 int ctxn;
3566 3823
3567 rcu_read_lock(); 3824 rcu_read_lock();
3568 cpuctx = &get_cpu_var(perf_cpu_context); 3825 list_for_each_entry_rcu(pmu, &pmus, entry) {
3569 perf_event_task_ctx(&cpuctx->ctx, task_event); 3826 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3570 if (!ctx) 3827 if (cpuctx->active_pmu != pmu)
3571 ctx = rcu_dereference(current->perf_event_ctxp); 3828 goto next;
3572 if (ctx) 3829 perf_event_task_ctx(&cpuctx->ctx, task_event);
3573 perf_event_task_ctx(ctx, task_event); 3830
3574 put_cpu_var(perf_cpu_context); 3831 ctx = task_event->task_ctx;
3832 if (!ctx) {
3833 ctxn = pmu->task_ctx_nr;
3834 if (ctxn < 0)
3835 goto next;
3836 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3837 }
3838 if (ctx)
3839 perf_event_task_ctx(ctx, task_event);
3840next:
3841 put_cpu_ptr(pmu->pmu_cpu_context);
3842 }
3575 rcu_read_unlock(); 3843 rcu_read_unlock();
3576} 3844}
3577 3845
@@ -3676,8 +3944,10 @@ static void perf_event_comm_event(struct perf_comm_event *comm_event)
3676{ 3944{
3677 struct perf_cpu_context *cpuctx; 3945 struct perf_cpu_context *cpuctx;
3678 struct perf_event_context *ctx; 3946 struct perf_event_context *ctx;
3679 unsigned int size;
3680 char comm[TASK_COMM_LEN]; 3947 char comm[TASK_COMM_LEN];
3948 unsigned int size;
3949 struct pmu *pmu;
3950 int ctxn;
3681 3951
3682 memset(comm, 0, sizeof(comm)); 3952 memset(comm, 0, sizeof(comm));
3683 strlcpy(comm, comm_event->task->comm, sizeof(comm)); 3953 strlcpy(comm, comm_event->task->comm, sizeof(comm));
@@ -3689,21 +3959,38 @@ static void perf_event_comm_event(struct perf_comm_event *comm_event)
3689 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; 3959 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3690 3960
3691 rcu_read_lock(); 3961 rcu_read_lock();
3692 cpuctx = &get_cpu_var(perf_cpu_context); 3962 list_for_each_entry_rcu(pmu, &pmus, entry) {
3693 perf_event_comm_ctx(&cpuctx->ctx, comm_event); 3963 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3694 ctx = rcu_dereference(current->perf_event_ctxp); 3964 if (cpuctx->active_pmu != pmu)
3695 if (ctx) 3965 goto next;
3696 perf_event_comm_ctx(ctx, comm_event); 3966 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3697 put_cpu_var(perf_cpu_context); 3967
3968 ctxn = pmu->task_ctx_nr;
3969 if (ctxn < 0)
3970 goto next;
3971
3972 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
3973 if (ctx)
3974 perf_event_comm_ctx(ctx, comm_event);
3975next:
3976 put_cpu_ptr(pmu->pmu_cpu_context);
3977 }
3698 rcu_read_unlock(); 3978 rcu_read_unlock();
3699} 3979}
3700 3980
3701void perf_event_comm(struct task_struct *task) 3981void perf_event_comm(struct task_struct *task)
3702{ 3982{
3703 struct perf_comm_event comm_event; 3983 struct perf_comm_event comm_event;
3984 struct perf_event_context *ctx;
3985 int ctxn;
3986
3987 for_each_task_context_nr(ctxn) {
3988 ctx = task->perf_event_ctxp[ctxn];
3989 if (!ctx)
3990 continue;
3704 3991
3705 if (task->perf_event_ctxp) 3992 perf_event_enable_on_exec(ctx);
3706 perf_event_enable_on_exec(task); 3993 }
3707 3994
3708 if (!atomic_read(&nr_comm_events)) 3995 if (!atomic_read(&nr_comm_events))
3709 return; 3996 return;
@@ -3805,6 +4092,8 @@ static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3805 char tmp[16]; 4092 char tmp[16];
3806 char *buf = NULL; 4093 char *buf = NULL;
3807 const char *name; 4094 const char *name;
4095 struct pmu *pmu;
4096 int ctxn;
3808 4097
3809 memset(tmp, 0, sizeof(tmp)); 4098 memset(tmp, 0, sizeof(tmp));
3810 4099
@@ -3857,12 +4146,25 @@ got_name:
3857 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; 4146 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3858 4147
3859 rcu_read_lock(); 4148 rcu_read_lock();
3860 cpuctx = &get_cpu_var(perf_cpu_context); 4149 list_for_each_entry_rcu(pmu, &pmus, entry) {
3861 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event, vma->vm_flags & VM_EXEC); 4150 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
3862 ctx = rcu_dereference(current->perf_event_ctxp); 4151 if (cpuctx->active_pmu != pmu)
3863 if (ctx) 4152 goto next;
3864 perf_event_mmap_ctx(ctx, mmap_event, vma->vm_flags & VM_EXEC); 4153 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
3865 put_cpu_var(perf_cpu_context); 4154 vma->vm_flags & VM_EXEC);
4155
4156 ctxn = pmu->task_ctx_nr;
4157 if (ctxn < 0)
4158 goto next;
4159
4160 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4161 if (ctx) {
4162 perf_event_mmap_ctx(ctx, mmap_event,
4163 vma->vm_flags & VM_EXEC);
4164 }
4165next:
4166 put_cpu_ptr(pmu->pmu_cpu_context);
4167 }
3866 rcu_read_unlock(); 4168 rcu_read_unlock();
3867 4169
3868 kfree(buf); 4170 kfree(buf);
@@ -3944,8 +4246,6 @@ static int __perf_event_overflow(struct perf_event *event, int nmi,
3944 struct hw_perf_event *hwc = &event->hw; 4246 struct hw_perf_event *hwc = &event->hw;
3945 int ret = 0; 4247 int ret = 0;
3946 4248
3947 throttle = (throttle && event->pmu->unthrottle != NULL);
3948
3949 if (!throttle) { 4249 if (!throttle) {
3950 hwc->interrupts++; 4250 hwc->interrupts++;
3951 } else { 4251 } else {
@@ -3988,8 +4288,7 @@ static int __perf_event_overflow(struct perf_event *event, int nmi,
3988 event->pending_kill = POLL_HUP; 4288 event->pending_kill = POLL_HUP;
3989 if (nmi) { 4289 if (nmi) {
3990 event->pending_disable = 1; 4290 event->pending_disable = 1;
3991 perf_pending_queue(&event->pending, 4291 irq_work_queue(&event->pending);
3992 perf_pending_event);
3993 } else 4292 } else
3994 perf_event_disable(event); 4293 perf_event_disable(event);
3995 } 4294 }
@@ -4013,6 +4312,17 @@ int perf_event_overflow(struct perf_event *event, int nmi,
4013 * Generic software event infrastructure 4312 * Generic software event infrastructure
4014 */ 4313 */
4015 4314
4315struct swevent_htable {
4316 struct swevent_hlist *swevent_hlist;
4317 struct mutex hlist_mutex;
4318 int hlist_refcount;
4319
4320 /* Recursion avoidance in each contexts */
4321 int recursion[PERF_NR_CONTEXTS];
4322};
4323
4324static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4325
4016/* 4326/*
4017 * We directly increment event->count and keep a second value in 4327 * We directly increment event->count and keep a second value in
4018 * event->hw.period_left to count intervals. This period event 4328 * event->hw.period_left to count intervals. This period event
@@ -4070,7 +4380,7 @@ static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4070 } 4380 }
4071} 4381}
4072 4382
4073static void perf_swevent_add(struct perf_event *event, u64 nr, 4383static void perf_swevent_event(struct perf_event *event, u64 nr,
4074 int nmi, struct perf_sample_data *data, 4384 int nmi, struct perf_sample_data *data,
4075 struct pt_regs *regs) 4385 struct pt_regs *regs)
4076{ 4386{
@@ -4096,6 +4406,9 @@ static void perf_swevent_add(struct perf_event *event, u64 nr,
4096static int perf_exclude_event(struct perf_event *event, 4406static int perf_exclude_event(struct perf_event *event,
4097 struct pt_regs *regs) 4407 struct pt_regs *regs)
4098{ 4408{
4409 if (event->hw.state & PERF_HES_STOPPED)
4410 return 0;
4411
4099 if (regs) { 4412 if (regs) {
4100 if (event->attr.exclude_user && user_mode(regs)) 4413 if (event->attr.exclude_user && user_mode(regs))
4101 return 1; 4414 return 1;
@@ -4142,11 +4455,11 @@ __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4142 4455
4143/* For the read side: events when they trigger */ 4456/* For the read side: events when they trigger */
4144static inline struct hlist_head * 4457static inline struct hlist_head *
4145find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id) 4458find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4146{ 4459{
4147 struct swevent_hlist *hlist; 4460 struct swevent_hlist *hlist;
4148 4461
4149 hlist = rcu_dereference(ctx->swevent_hlist); 4462 hlist = rcu_dereference(swhash->swevent_hlist);
4150 if (!hlist) 4463 if (!hlist)
4151 return NULL; 4464 return NULL;
4152 4465
@@ -4155,7 +4468,7 @@ find_swevent_head_rcu(struct perf_cpu_context *ctx, u64 type, u32 event_id)
4155 4468
4156/* For the event head insertion and removal in the hlist */ 4469/* For the event head insertion and removal in the hlist */
4157static inline struct hlist_head * 4470static inline struct hlist_head *
4158find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event) 4471find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4159{ 4472{
4160 struct swevent_hlist *hlist; 4473 struct swevent_hlist *hlist;
4161 u32 event_id = event->attr.config; 4474 u32 event_id = event->attr.config;
@@ -4166,7 +4479,7 @@ find_swevent_head(struct perf_cpu_context *ctx, struct perf_event *event)
4166 * and release. Which makes the protected version suitable here. 4479 * and release. Which makes the protected version suitable here.
4167 * The context lock guarantees that. 4480 * The context lock guarantees that.
4168 */ 4481 */
4169 hlist = rcu_dereference_protected(ctx->swevent_hlist, 4482 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4170 lockdep_is_held(&event->ctx->lock)); 4483 lockdep_is_held(&event->ctx->lock));
4171 if (!hlist) 4484 if (!hlist)
4172 return NULL; 4485 return NULL;
@@ -4179,23 +4492,19 @@ static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4179 struct perf_sample_data *data, 4492 struct perf_sample_data *data,
4180 struct pt_regs *regs) 4493 struct pt_regs *regs)
4181{ 4494{
4182 struct perf_cpu_context *cpuctx; 4495 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4183 struct perf_event *event; 4496 struct perf_event *event;
4184 struct hlist_node *node; 4497 struct hlist_node *node;
4185 struct hlist_head *head; 4498 struct hlist_head *head;
4186 4499
4187 cpuctx = &__get_cpu_var(perf_cpu_context);
4188
4189 rcu_read_lock(); 4500 rcu_read_lock();
4190 4501 head = find_swevent_head_rcu(swhash, type, event_id);
4191 head = find_swevent_head_rcu(cpuctx, type, event_id);
4192
4193 if (!head) 4502 if (!head)
4194 goto end; 4503 goto end;
4195 4504
4196 hlist_for_each_entry_rcu(event, node, head, hlist_entry) { 4505 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4197 if (perf_swevent_match(event, type, event_id, data, regs)) 4506 if (perf_swevent_match(event, type, event_id, data, regs))
4198 perf_swevent_add(event, nr, nmi, data, regs); 4507 perf_swevent_event(event, nr, nmi, data, regs);
4199 } 4508 }
4200end: 4509end:
4201 rcu_read_unlock(); 4510 rcu_read_unlock();
@@ -4203,33 +4512,17 @@ end:
4203 4512
4204int perf_swevent_get_recursion_context(void) 4513int perf_swevent_get_recursion_context(void)
4205{ 4514{
4206 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); 4515 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4207 int rctx;
4208
4209 if (in_nmi())
4210 rctx = 3;
4211 else if (in_irq())
4212 rctx = 2;
4213 else if (in_softirq())
4214 rctx = 1;
4215 else
4216 rctx = 0;
4217
4218 if (cpuctx->recursion[rctx])
4219 return -1;
4220 4516
4221 cpuctx->recursion[rctx]++; 4517 return get_recursion_context(swhash->recursion);
4222 barrier();
4223
4224 return rctx;
4225} 4518}
4226EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); 4519EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4227 4520
4228void inline perf_swevent_put_recursion_context(int rctx) 4521void inline perf_swevent_put_recursion_context(int rctx)
4229{ 4522{
4230 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); 4523 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4231 barrier(); 4524
4232 cpuctx->recursion[rctx]--; 4525 put_recursion_context(swhash->recursion, rctx);
4233} 4526}
4234 4527
4235void __perf_sw_event(u32 event_id, u64 nr, int nmi, 4528void __perf_sw_event(u32 event_id, u64 nr, int nmi,
@@ -4255,20 +4548,20 @@ static void perf_swevent_read(struct perf_event *event)
4255{ 4548{
4256} 4549}
4257 4550
4258static int perf_swevent_enable(struct perf_event *event) 4551static int perf_swevent_add(struct perf_event *event, int flags)
4259{ 4552{
4553 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4260 struct hw_perf_event *hwc = &event->hw; 4554 struct hw_perf_event *hwc = &event->hw;
4261 struct perf_cpu_context *cpuctx;
4262 struct hlist_head *head; 4555 struct hlist_head *head;
4263 4556
4264 cpuctx = &__get_cpu_var(perf_cpu_context);
4265
4266 if (hwc->sample_period) { 4557 if (hwc->sample_period) {
4267 hwc->last_period = hwc->sample_period; 4558 hwc->last_period = hwc->sample_period;
4268 perf_swevent_set_period(event); 4559 perf_swevent_set_period(event);
4269 } 4560 }
4270 4561
4271 head = find_swevent_head(cpuctx, event); 4562 hwc->state = !(flags & PERF_EF_START);
4563
4564 head = find_swevent_head(swhash, event);
4272 if (WARN_ON_ONCE(!head)) 4565 if (WARN_ON_ONCE(!head))
4273 return -EINVAL; 4566 return -EINVAL;
4274 4567
@@ -4277,202 +4570,27 @@ static int perf_swevent_enable(struct perf_event *event)
4277 return 0; 4570 return 0;
4278} 4571}
4279 4572
4280static void perf_swevent_disable(struct perf_event *event) 4573static void perf_swevent_del(struct perf_event *event, int flags)
4281{ 4574{
4282 hlist_del_rcu(&event->hlist_entry); 4575 hlist_del_rcu(&event->hlist_entry);
4283} 4576}
4284 4577
4285static void perf_swevent_void(struct perf_event *event) 4578static void perf_swevent_start(struct perf_event *event, int flags)
4286{ 4579{
4580 event->hw.state = 0;
4287} 4581}
4288 4582
4289static int perf_swevent_int(struct perf_event *event) 4583static void perf_swevent_stop(struct perf_event *event, int flags)
4290{ 4584{
4291 return 0; 4585 event->hw.state = PERF_HES_STOPPED;
4292} 4586}
4293 4587
4294static const struct pmu perf_ops_generic = {
4295 .enable = perf_swevent_enable,
4296 .disable = perf_swevent_disable,
4297 .start = perf_swevent_int,
4298 .stop = perf_swevent_void,
4299 .read = perf_swevent_read,
4300 .unthrottle = perf_swevent_void, /* hwc->interrupts already reset */
4301};
4302
4303/*
4304 * hrtimer based swevent callback
4305 */
4306
4307static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4308{
4309 enum hrtimer_restart ret = HRTIMER_RESTART;
4310 struct perf_sample_data data;
4311 struct pt_regs *regs;
4312 struct perf_event *event;
4313 u64 period;
4314
4315 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4316 event->pmu->read(event);
4317
4318 perf_sample_data_init(&data, 0);
4319 data.period = event->hw.last_period;
4320 regs = get_irq_regs();
4321
4322 if (regs && !perf_exclude_event(event, regs)) {
4323 if (!(event->attr.exclude_idle && current->pid == 0))
4324 if (perf_event_overflow(event, 0, &data, regs))
4325 ret = HRTIMER_NORESTART;
4326 }
4327
4328 period = max_t(u64, 10000, event->hw.sample_period);
4329 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4330
4331 return ret;
4332}
4333
4334static void perf_swevent_start_hrtimer(struct perf_event *event)
4335{
4336 struct hw_perf_event *hwc = &event->hw;
4337
4338 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4339 hwc->hrtimer.function = perf_swevent_hrtimer;
4340 if (hwc->sample_period) {
4341 u64 period;
4342
4343 if (hwc->remaining) {
4344 if (hwc->remaining < 0)
4345 period = 10000;
4346 else
4347 period = hwc->remaining;
4348 hwc->remaining = 0;
4349 } else {
4350 period = max_t(u64, 10000, hwc->sample_period);
4351 }
4352 __hrtimer_start_range_ns(&hwc->hrtimer,
4353 ns_to_ktime(period), 0,
4354 HRTIMER_MODE_REL, 0);
4355 }
4356}
4357
4358static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4359{
4360 struct hw_perf_event *hwc = &event->hw;
4361
4362 if (hwc->sample_period) {
4363 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4364 hwc->remaining = ktime_to_ns(remaining);
4365
4366 hrtimer_cancel(&hwc->hrtimer);
4367 }
4368}
4369
4370/*
4371 * Software event: cpu wall time clock
4372 */
4373
4374static void cpu_clock_perf_event_update(struct perf_event *event)
4375{
4376 int cpu = raw_smp_processor_id();
4377 s64 prev;
4378 u64 now;
4379
4380 now = cpu_clock(cpu);
4381 prev = local64_xchg(&event->hw.prev_count, now);
4382 local64_add(now - prev, &event->count);
4383}
4384
4385static int cpu_clock_perf_event_enable(struct perf_event *event)
4386{
4387 struct hw_perf_event *hwc = &event->hw;
4388 int cpu = raw_smp_processor_id();
4389
4390 local64_set(&hwc->prev_count, cpu_clock(cpu));
4391 perf_swevent_start_hrtimer(event);
4392
4393 return 0;
4394}
4395
4396static void cpu_clock_perf_event_disable(struct perf_event *event)
4397{
4398 perf_swevent_cancel_hrtimer(event);
4399 cpu_clock_perf_event_update(event);
4400}
4401
4402static void cpu_clock_perf_event_read(struct perf_event *event)
4403{
4404 cpu_clock_perf_event_update(event);
4405}
4406
4407static const struct pmu perf_ops_cpu_clock = {
4408 .enable = cpu_clock_perf_event_enable,
4409 .disable = cpu_clock_perf_event_disable,
4410 .read = cpu_clock_perf_event_read,
4411};
4412
4413/*
4414 * Software event: task time clock
4415 */
4416
4417static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4418{
4419 u64 prev;
4420 s64 delta;
4421
4422 prev = local64_xchg(&event->hw.prev_count, now);
4423 delta = now - prev;
4424 local64_add(delta, &event->count);
4425}
4426
4427static int task_clock_perf_event_enable(struct perf_event *event)
4428{
4429 struct hw_perf_event *hwc = &event->hw;
4430 u64 now;
4431
4432 now = event->ctx->time;
4433
4434 local64_set(&hwc->prev_count, now);
4435
4436 perf_swevent_start_hrtimer(event);
4437
4438 return 0;
4439}
4440
4441static void task_clock_perf_event_disable(struct perf_event *event)
4442{
4443 perf_swevent_cancel_hrtimer(event);
4444 task_clock_perf_event_update(event, event->ctx->time);
4445
4446}
4447
4448static void task_clock_perf_event_read(struct perf_event *event)
4449{
4450 u64 time;
4451
4452 if (!in_nmi()) {
4453 update_context_time(event->ctx);
4454 time = event->ctx->time;
4455 } else {
4456 u64 now = perf_clock();
4457 u64 delta = now - event->ctx->timestamp;
4458 time = event->ctx->time + delta;
4459 }
4460
4461 task_clock_perf_event_update(event, time);
4462}
4463
4464static const struct pmu perf_ops_task_clock = {
4465 .enable = task_clock_perf_event_enable,
4466 .disable = task_clock_perf_event_disable,
4467 .read = task_clock_perf_event_read,
4468};
4469
4470/* Deref the hlist from the update side */ 4588/* Deref the hlist from the update side */
4471static inline struct swevent_hlist * 4589static inline struct swevent_hlist *
4472swevent_hlist_deref(struct perf_cpu_context *cpuctx) 4590swevent_hlist_deref(struct swevent_htable *swhash)
4473{ 4591{
4474 return rcu_dereference_protected(cpuctx->swevent_hlist, 4592 return rcu_dereference_protected(swhash->swevent_hlist,
4475 lockdep_is_held(&cpuctx->hlist_mutex)); 4593 lockdep_is_held(&swhash->hlist_mutex));
4476} 4594}
4477 4595
4478static void swevent_hlist_release_rcu(struct rcu_head *rcu_head) 4596static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
@@ -4483,27 +4601,27 @@ static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
4483 kfree(hlist); 4601 kfree(hlist);
4484} 4602}
4485 4603
4486static void swevent_hlist_release(struct perf_cpu_context *cpuctx) 4604static void swevent_hlist_release(struct swevent_htable *swhash)
4487{ 4605{
4488 struct swevent_hlist *hlist = swevent_hlist_deref(cpuctx); 4606 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4489 4607
4490 if (!hlist) 4608 if (!hlist)
4491 return; 4609 return;
4492 4610
4493 rcu_assign_pointer(cpuctx->swevent_hlist, NULL); 4611 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4494 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu); 4612 call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
4495} 4613}
4496 4614
4497static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) 4615static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4498{ 4616{
4499 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); 4617 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4500 4618
4501 mutex_lock(&cpuctx->hlist_mutex); 4619 mutex_lock(&swhash->hlist_mutex);
4502 4620
4503 if (!--cpuctx->hlist_refcount) 4621 if (!--swhash->hlist_refcount)
4504 swevent_hlist_release(cpuctx); 4622 swevent_hlist_release(swhash);
4505 4623
4506 mutex_unlock(&cpuctx->hlist_mutex); 4624 mutex_unlock(&swhash->hlist_mutex);
4507} 4625}
4508 4626
4509static void swevent_hlist_put(struct perf_event *event) 4627static void swevent_hlist_put(struct perf_event *event)
@@ -4521,12 +4639,12 @@ static void swevent_hlist_put(struct perf_event *event)
4521 4639
4522static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) 4640static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4523{ 4641{
4524 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); 4642 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4525 int err = 0; 4643 int err = 0;
4526 4644
4527 mutex_lock(&cpuctx->hlist_mutex); 4645 mutex_lock(&swhash->hlist_mutex);
4528 4646
4529 if (!swevent_hlist_deref(cpuctx) && cpu_online(cpu)) { 4647 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
4530 struct swevent_hlist *hlist; 4648 struct swevent_hlist *hlist;
4531 4649
4532 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); 4650 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
@@ -4534,11 +4652,11 @@ static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
4534 err = -ENOMEM; 4652 err = -ENOMEM;
4535 goto exit; 4653 goto exit;
4536 } 4654 }
4537 rcu_assign_pointer(cpuctx->swevent_hlist, hlist); 4655 rcu_assign_pointer(swhash->swevent_hlist, hlist);
4538 } 4656 }
4539 cpuctx->hlist_refcount++; 4657 swhash->hlist_refcount++;
4540 exit: 4658exit:
4541 mutex_unlock(&cpuctx->hlist_mutex); 4659 mutex_unlock(&swhash->hlist_mutex);
4542 4660
4543 return err; 4661 return err;
4544} 4662}
@@ -4562,7 +4680,7 @@ static int swevent_hlist_get(struct perf_event *event)
4562 put_online_cpus(); 4680 put_online_cpus();
4563 4681
4564 return 0; 4682 return 0;
4565 fail: 4683fail:
4566 for_each_possible_cpu(cpu) { 4684 for_each_possible_cpu(cpu) {
4567 if (cpu == failed_cpu) 4685 if (cpu == failed_cpu)
4568 break; 4686 break;
@@ -4573,17 +4691,64 @@ static int swevent_hlist_get(struct perf_event *event)
4573 return err; 4691 return err;
4574} 4692}
4575 4693
4576#ifdef CONFIG_EVENT_TRACING 4694atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4577 4695
4578static const struct pmu perf_ops_tracepoint = { 4696static void sw_perf_event_destroy(struct perf_event *event)
4579 .enable = perf_trace_enable, 4697{
4580 .disable = perf_trace_disable, 4698 u64 event_id = event->attr.config;
4581 .start = perf_swevent_int, 4699
4582 .stop = perf_swevent_void, 4700 WARN_ON(event->parent);
4701
4702 jump_label_dec(&perf_swevent_enabled[event_id]);
4703 swevent_hlist_put(event);
4704}
4705
4706static int perf_swevent_init(struct perf_event *event)
4707{
4708 int event_id = event->attr.config;
4709
4710 if (event->attr.type != PERF_TYPE_SOFTWARE)
4711 return -ENOENT;
4712
4713 switch (event_id) {
4714 case PERF_COUNT_SW_CPU_CLOCK:
4715 case PERF_COUNT_SW_TASK_CLOCK:
4716 return -ENOENT;
4717
4718 default:
4719 break;
4720 }
4721
4722 if (event_id >= PERF_COUNT_SW_MAX)
4723 return -ENOENT;
4724
4725 if (!event->parent) {
4726 int err;
4727
4728 err = swevent_hlist_get(event);
4729 if (err)
4730 return err;
4731
4732 jump_label_inc(&perf_swevent_enabled[event_id]);
4733 event->destroy = sw_perf_event_destroy;
4734 }
4735
4736 return 0;
4737}
4738
4739static struct pmu perf_swevent = {
4740 .task_ctx_nr = perf_sw_context,
4741
4742 .event_init = perf_swevent_init,
4743 .add = perf_swevent_add,
4744 .del = perf_swevent_del,
4745 .start = perf_swevent_start,
4746 .stop = perf_swevent_stop,
4583 .read = perf_swevent_read, 4747 .read = perf_swevent_read,
4584 .unthrottle = perf_swevent_void,
4585}; 4748};
4586 4749
4750#ifdef CONFIG_EVENT_TRACING
4751
4587static int perf_tp_filter_match(struct perf_event *event, 4752static int perf_tp_filter_match(struct perf_event *event,
4588 struct perf_sample_data *data) 4753 struct perf_sample_data *data)
4589{ 4754{
@@ -4627,7 +4792,7 @@ void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
4627 4792
4628 hlist_for_each_entry_rcu(event, node, head, hlist_entry) { 4793 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4629 if (perf_tp_event_match(event, &data, regs)) 4794 if (perf_tp_event_match(event, &data, regs))
4630 perf_swevent_add(event, count, 1, &data, regs); 4795 perf_swevent_event(event, count, 1, &data, regs);
4631 } 4796 }
4632 4797
4633 perf_swevent_put_recursion_context(rctx); 4798 perf_swevent_put_recursion_context(rctx);
@@ -4639,10 +4804,13 @@ static void tp_perf_event_destroy(struct perf_event *event)
4639 perf_trace_destroy(event); 4804 perf_trace_destroy(event);
4640} 4805}
4641 4806
4642static const struct pmu *tp_perf_event_init(struct perf_event *event) 4807static int perf_tp_event_init(struct perf_event *event)
4643{ 4808{
4644 int err; 4809 int err;
4645 4810
4811 if (event->attr.type != PERF_TYPE_TRACEPOINT)
4812 return -ENOENT;
4813
4646 /* 4814 /*
4647 * Raw tracepoint data is a severe data leak, only allow root to 4815 * Raw tracepoint data is a severe data leak, only allow root to
4648 * have these. 4816 * have these.
@@ -4650,15 +4818,31 @@ static const struct pmu *tp_perf_event_init(struct perf_event *event)
4650 if ((event->attr.sample_type & PERF_SAMPLE_RAW) && 4818 if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4651 perf_paranoid_tracepoint_raw() && 4819 perf_paranoid_tracepoint_raw() &&
4652 !capable(CAP_SYS_ADMIN)) 4820 !capable(CAP_SYS_ADMIN))
4653 return ERR_PTR(-EPERM); 4821 return -EPERM;
4654 4822
4655 err = perf_trace_init(event); 4823 err = perf_trace_init(event);
4656 if (err) 4824 if (err)
4657 return NULL; 4825 return err;
4658 4826
4659 event->destroy = tp_perf_event_destroy; 4827 event->destroy = tp_perf_event_destroy;
4660 4828
4661 return &perf_ops_tracepoint; 4829 return 0;
4830}
4831
4832static struct pmu perf_tracepoint = {
4833 .task_ctx_nr = perf_sw_context,
4834
4835 .event_init = perf_tp_event_init,
4836 .add = perf_trace_add,
4837 .del = perf_trace_del,
4838 .start = perf_swevent_start,
4839 .stop = perf_swevent_stop,
4840 .read = perf_swevent_read,
4841};
4842
4843static inline void perf_tp_register(void)
4844{
4845 perf_pmu_register(&perf_tracepoint);
4662} 4846}
4663 4847
4664static int perf_event_set_filter(struct perf_event *event, void __user *arg) 4848static int perf_event_set_filter(struct perf_event *event, void __user *arg)
@@ -4686,9 +4870,8 @@ static void perf_event_free_filter(struct perf_event *event)
4686 4870
4687#else 4871#else
4688 4872
4689static const struct pmu *tp_perf_event_init(struct perf_event *event) 4873static inline void perf_tp_register(void)
4690{ 4874{
4691 return NULL;
4692} 4875}
4693 4876
4694static int perf_event_set_filter(struct perf_event *event, void __user *arg) 4877static int perf_event_set_filter(struct perf_event *event, void __user *arg)
@@ -4703,105 +4886,406 @@ static void perf_event_free_filter(struct perf_event *event)
4703#endif /* CONFIG_EVENT_TRACING */ 4886#endif /* CONFIG_EVENT_TRACING */
4704 4887
4705#ifdef CONFIG_HAVE_HW_BREAKPOINT 4888#ifdef CONFIG_HAVE_HW_BREAKPOINT
4706static void bp_perf_event_destroy(struct perf_event *event) 4889void perf_bp_event(struct perf_event *bp, void *data)
4707{ 4890{
4708 release_bp_slot(event); 4891 struct perf_sample_data sample;
4892 struct pt_regs *regs = data;
4893
4894 perf_sample_data_init(&sample, bp->attr.bp_addr);
4895
4896 if (!bp->hw.state && !perf_exclude_event(bp, regs))
4897 perf_swevent_event(bp, 1, 1, &sample, regs);
4709} 4898}
4899#endif
4710 4900
4711static const struct pmu *bp_perf_event_init(struct perf_event *bp) 4901/*
4902 * hrtimer based swevent callback
4903 */
4904
4905static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4712{ 4906{
4713 int err; 4907 enum hrtimer_restart ret = HRTIMER_RESTART;
4908 struct perf_sample_data data;
4909 struct pt_regs *regs;
4910 struct perf_event *event;
4911 u64 period;
4714 4912
4715 err = register_perf_hw_breakpoint(bp); 4913 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4716 if (err) 4914 event->pmu->read(event);
4717 return ERR_PTR(err);
4718 4915
4719 bp->destroy = bp_perf_event_destroy; 4916 perf_sample_data_init(&data, 0);
4917 data.period = event->hw.last_period;
4918 regs = get_irq_regs();
4919
4920 if (regs && !perf_exclude_event(event, regs)) {
4921 if (!(event->attr.exclude_idle && current->pid == 0))
4922 if (perf_event_overflow(event, 0, &data, regs))
4923 ret = HRTIMER_NORESTART;
4924 }
4925
4926 period = max_t(u64, 10000, event->hw.sample_period);
4927 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4720 4928
4721 return &perf_ops_bp; 4929 return ret;
4722} 4930}
4723 4931
4724void perf_bp_event(struct perf_event *bp, void *data) 4932static void perf_swevent_start_hrtimer(struct perf_event *event)
4725{ 4933{
4726 struct perf_sample_data sample; 4934 struct hw_perf_event *hwc = &event->hw;
4727 struct pt_regs *regs = data;
4728 4935
4729 perf_sample_data_init(&sample, bp->attr.bp_addr); 4936 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4937 hwc->hrtimer.function = perf_swevent_hrtimer;
4938 if (hwc->sample_period) {
4939 s64 period = local64_read(&hwc->period_left);
4730 4940
4731 if (!perf_exclude_event(bp, regs)) 4941 if (period) {
4732 perf_swevent_add(bp, 1, 1, &sample, regs); 4942 if (period < 0)
4943 period = 10000;
4944
4945 local64_set(&hwc->period_left, 0);
4946 } else {
4947 period = max_t(u64, 10000, hwc->sample_period);
4948 }
4949 __hrtimer_start_range_ns(&hwc->hrtimer,
4950 ns_to_ktime(period), 0,
4951 HRTIMER_MODE_REL_PINNED, 0);
4952 }
4733} 4953}
4734#else 4954
4735static const struct pmu *bp_perf_event_init(struct perf_event *bp) 4955static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4956{
4957 struct hw_perf_event *hwc = &event->hw;
4958
4959 if (hwc->sample_period) {
4960 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4961 local64_set(&hwc->period_left, ktime_to_ns(remaining));
4962
4963 hrtimer_cancel(&hwc->hrtimer);
4964 }
4965}
4966
4967/*
4968 * Software event: cpu wall time clock
4969 */
4970
4971static void cpu_clock_event_update(struct perf_event *event)
4972{
4973 s64 prev;
4974 u64 now;
4975
4976 now = local_clock();
4977 prev = local64_xchg(&event->hw.prev_count, now);
4978 local64_add(now - prev, &event->count);
4979}
4980
4981static void cpu_clock_event_start(struct perf_event *event, int flags)
4982{
4983 local64_set(&event->hw.prev_count, local_clock());
4984 perf_swevent_start_hrtimer(event);
4985}
4986
4987static void cpu_clock_event_stop(struct perf_event *event, int flags)
4988{
4989 perf_swevent_cancel_hrtimer(event);
4990 cpu_clock_event_update(event);
4991}
4992
4993static int cpu_clock_event_add(struct perf_event *event, int flags)
4994{
4995 if (flags & PERF_EF_START)
4996 cpu_clock_event_start(event, flags);
4997
4998 return 0;
4999}
5000
5001static void cpu_clock_event_del(struct perf_event *event, int flags)
5002{
5003 cpu_clock_event_stop(event, flags);
5004}
5005
5006static void cpu_clock_event_read(struct perf_event *event)
5007{
5008 cpu_clock_event_update(event);
5009}
5010
5011static int cpu_clock_event_init(struct perf_event *event)
5012{
5013 if (event->attr.type != PERF_TYPE_SOFTWARE)
5014 return -ENOENT;
5015
5016 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5017 return -ENOENT;
5018
5019 return 0;
5020}
5021
5022static struct pmu perf_cpu_clock = {
5023 .task_ctx_nr = perf_sw_context,
5024
5025 .event_init = cpu_clock_event_init,
5026 .add = cpu_clock_event_add,
5027 .del = cpu_clock_event_del,
5028 .start = cpu_clock_event_start,
5029 .stop = cpu_clock_event_stop,
5030 .read = cpu_clock_event_read,
5031};
5032
5033/*
5034 * Software event: task time clock
5035 */
5036
5037static void task_clock_event_update(struct perf_event *event, u64 now)
5038{
5039 u64 prev;
5040 s64 delta;
5041
5042 prev = local64_xchg(&event->hw.prev_count, now);
5043 delta = now - prev;
5044 local64_add(delta, &event->count);
5045}
5046
5047static void task_clock_event_start(struct perf_event *event, int flags)
5048{
5049 local64_set(&event->hw.prev_count, event->ctx->time);
5050 perf_swevent_start_hrtimer(event);
5051}
5052
5053static void task_clock_event_stop(struct perf_event *event, int flags)
5054{
5055 perf_swevent_cancel_hrtimer(event);
5056 task_clock_event_update(event, event->ctx->time);
5057}
5058
5059static int task_clock_event_add(struct perf_event *event, int flags)
5060{
5061 if (flags & PERF_EF_START)
5062 task_clock_event_start(event, flags);
5063
5064 return 0;
5065}
5066
5067static void task_clock_event_del(struct perf_event *event, int flags)
5068{
5069 task_clock_event_stop(event, PERF_EF_UPDATE);
5070}
5071
5072static void task_clock_event_read(struct perf_event *event)
5073{
5074 u64 time;
5075
5076 if (!in_nmi()) {
5077 update_context_time(event->ctx);
5078 time = event->ctx->time;
5079 } else {
5080 u64 now = perf_clock();
5081 u64 delta = now - event->ctx->timestamp;
5082 time = event->ctx->time + delta;
5083 }
5084
5085 task_clock_event_update(event, time);
5086}
5087
5088static int task_clock_event_init(struct perf_event *event)
5089{
5090 if (event->attr.type != PERF_TYPE_SOFTWARE)
5091 return -ENOENT;
5092
5093 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5094 return -ENOENT;
5095
5096 return 0;
5097}
5098
5099static struct pmu perf_task_clock = {
5100 .task_ctx_nr = perf_sw_context,
5101
5102 .event_init = task_clock_event_init,
5103 .add = task_clock_event_add,
5104 .del = task_clock_event_del,
5105 .start = task_clock_event_start,
5106 .stop = task_clock_event_stop,
5107 .read = task_clock_event_read,
5108};
5109
5110static void perf_pmu_nop_void(struct pmu *pmu)
5111{
5112}
5113
5114static int perf_pmu_nop_int(struct pmu *pmu)
4736{ 5115{
5116 return 0;
5117}
5118
5119static void perf_pmu_start_txn(struct pmu *pmu)
5120{
5121 perf_pmu_disable(pmu);
5122}
5123
5124static int perf_pmu_commit_txn(struct pmu *pmu)
5125{
5126 perf_pmu_enable(pmu);
5127 return 0;
5128}
5129
5130static void perf_pmu_cancel_txn(struct pmu *pmu)
5131{
5132 perf_pmu_enable(pmu);
5133}
5134
5135/*
5136 * Ensures all contexts with the same task_ctx_nr have the same
5137 * pmu_cpu_context too.
5138 */
5139static void *find_pmu_context(int ctxn)
5140{
5141 struct pmu *pmu;
5142
5143 if (ctxn < 0)
5144 return NULL;
5145
5146 list_for_each_entry(pmu, &pmus, entry) {
5147 if (pmu->task_ctx_nr == ctxn)
5148 return pmu->pmu_cpu_context;
5149 }
5150
4737 return NULL; 5151 return NULL;
4738} 5152}
4739 5153
4740void perf_bp_event(struct perf_event *bp, void *regs) 5154static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
4741{ 5155{
5156 int cpu;
5157
5158 for_each_possible_cpu(cpu) {
5159 struct perf_cpu_context *cpuctx;
5160
5161 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5162
5163 if (cpuctx->active_pmu == old_pmu)
5164 cpuctx->active_pmu = pmu;
5165 }
4742} 5166}
4743#endif
4744 5167
4745atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX]; 5168static void free_pmu_context(struct pmu *pmu)
5169{
5170 struct pmu *i;
4746 5171
4747static void sw_perf_event_destroy(struct perf_event *event) 5172 mutex_lock(&pmus_lock);
5173 /*
5174 * Like a real lame refcount.
5175 */
5176 list_for_each_entry(i, &pmus, entry) {
5177 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5178 update_pmu_context(i, pmu);
5179 goto out;
5180 }
5181 }
5182
5183 free_percpu(pmu->pmu_cpu_context);
5184out:
5185 mutex_unlock(&pmus_lock);
5186}
5187
5188int perf_pmu_register(struct pmu *pmu)
4748{ 5189{
4749 u64 event_id = event->attr.config; 5190 int cpu, ret;
4750 5191
4751 WARN_ON(event->parent); 5192 mutex_lock(&pmus_lock);
5193 ret = -ENOMEM;
5194 pmu->pmu_disable_count = alloc_percpu(int);
5195 if (!pmu->pmu_disable_count)
5196 goto unlock;
4752 5197
4753 atomic_dec(&perf_swevent_enabled[event_id]); 5198 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
4754 swevent_hlist_put(event); 5199 if (pmu->pmu_cpu_context)
5200 goto got_cpu_context;
5201
5202 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5203 if (!pmu->pmu_cpu_context)
5204 goto free_pdc;
5205
5206 for_each_possible_cpu(cpu) {
5207 struct perf_cpu_context *cpuctx;
5208
5209 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5210 __perf_event_init_context(&cpuctx->ctx);
5211 cpuctx->ctx.type = cpu_context;
5212 cpuctx->ctx.pmu = pmu;
5213 cpuctx->jiffies_interval = 1;
5214 INIT_LIST_HEAD(&cpuctx->rotation_list);
5215 cpuctx->active_pmu = pmu;
5216 }
5217
5218got_cpu_context:
5219 if (!pmu->start_txn) {
5220 if (pmu->pmu_enable) {
5221 /*
5222 * If we have pmu_enable/pmu_disable calls, install
5223 * transaction stubs that use that to try and batch
5224 * hardware accesses.
5225 */
5226 pmu->start_txn = perf_pmu_start_txn;
5227 pmu->commit_txn = perf_pmu_commit_txn;
5228 pmu->cancel_txn = perf_pmu_cancel_txn;
5229 } else {
5230 pmu->start_txn = perf_pmu_nop_void;
5231 pmu->commit_txn = perf_pmu_nop_int;
5232 pmu->cancel_txn = perf_pmu_nop_void;
5233 }
5234 }
5235
5236 if (!pmu->pmu_enable) {
5237 pmu->pmu_enable = perf_pmu_nop_void;
5238 pmu->pmu_disable = perf_pmu_nop_void;
5239 }
5240
5241 list_add_rcu(&pmu->entry, &pmus);
5242 ret = 0;
5243unlock:
5244 mutex_unlock(&pmus_lock);
5245
5246 return ret;
5247
5248free_pdc:
5249 free_percpu(pmu->pmu_disable_count);
5250 goto unlock;
4755} 5251}
4756 5252
4757static const struct pmu *sw_perf_event_init(struct perf_event *event) 5253void perf_pmu_unregister(struct pmu *pmu)
4758{ 5254{
4759 const struct pmu *pmu = NULL; 5255 mutex_lock(&pmus_lock);
4760 u64 event_id = event->attr.config; 5256 list_del_rcu(&pmu->entry);
5257 mutex_unlock(&pmus_lock);
4761 5258
4762 /* 5259 /*
4763 * Software events (currently) can't in general distinguish 5260 * We dereference the pmu list under both SRCU and regular RCU, so
4764 * between user, kernel and hypervisor events. 5261 * synchronize against both of those.
4765 * However, context switches and cpu migrations are considered
4766 * to be kernel events, and page faults are never hypervisor
4767 * events.
4768 */ 5262 */
4769 switch (event_id) { 5263 synchronize_srcu(&pmus_srcu);
4770 case PERF_COUNT_SW_CPU_CLOCK: 5264 synchronize_rcu();
4771 pmu = &perf_ops_cpu_clock;
4772 5265
4773 break; 5266 free_percpu(pmu->pmu_disable_count);
4774 case PERF_COUNT_SW_TASK_CLOCK: 5267 free_pmu_context(pmu);
4775 /* 5268}
4776 * If the user instantiates this as a per-cpu event,
4777 * use the cpu_clock event instead.
4778 */
4779 if (event->ctx->task)
4780 pmu = &perf_ops_task_clock;
4781 else
4782 pmu = &perf_ops_cpu_clock;
4783 5269
4784 break; 5270struct pmu *perf_init_event(struct perf_event *event)
4785 case PERF_COUNT_SW_PAGE_FAULTS: 5271{
4786 case PERF_COUNT_SW_PAGE_FAULTS_MIN: 5272 struct pmu *pmu = NULL;
4787 case PERF_COUNT_SW_PAGE_FAULTS_MAJ: 5273 int idx;
4788 case PERF_COUNT_SW_CONTEXT_SWITCHES:
4789 case PERF_COUNT_SW_CPU_MIGRATIONS:
4790 case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4791 case PERF_COUNT_SW_EMULATION_FAULTS:
4792 if (!event->parent) {
4793 int err;
4794
4795 err = swevent_hlist_get(event);
4796 if (err)
4797 return ERR_PTR(err);
4798 5274
4799 atomic_inc(&perf_swevent_enabled[event_id]); 5275 idx = srcu_read_lock(&pmus_srcu);
4800 event->destroy = sw_perf_event_destroy; 5276 list_for_each_entry_rcu(pmu, &pmus, entry) {
5277 int ret = pmu->event_init(event);
5278 if (!ret)
5279 goto unlock;
5280
5281 if (ret != -ENOENT) {
5282 pmu = ERR_PTR(ret);
5283 goto unlock;
4801 } 5284 }
4802 pmu = &perf_ops_generic;
4803 break;
4804 } 5285 }
5286 pmu = ERR_PTR(-ENOENT);
5287unlock:
5288 srcu_read_unlock(&pmus_srcu, idx);
4805 5289
4806 return pmu; 5290 return pmu;
4807} 5291}
@@ -4810,20 +5294,18 @@ static const struct pmu *sw_perf_event_init(struct perf_event *event)
4810 * Allocate and initialize a event structure 5294 * Allocate and initialize a event structure
4811 */ 5295 */
4812static struct perf_event * 5296static struct perf_event *
4813perf_event_alloc(struct perf_event_attr *attr, 5297perf_event_alloc(struct perf_event_attr *attr, int cpu,
4814 int cpu, 5298 struct task_struct *task,
4815 struct perf_event_context *ctx, 5299 struct perf_event *group_leader,
4816 struct perf_event *group_leader, 5300 struct perf_event *parent_event,
4817 struct perf_event *parent_event, 5301 perf_overflow_handler_t overflow_handler)
4818 perf_overflow_handler_t overflow_handler, 5302{
4819 gfp_t gfpflags) 5303 struct pmu *pmu;
4820{
4821 const struct pmu *pmu;
4822 struct perf_event *event; 5304 struct perf_event *event;
4823 struct hw_perf_event *hwc; 5305 struct hw_perf_event *hwc;
4824 long err; 5306 long err;
4825 5307
4826 event = kzalloc(sizeof(*event), gfpflags); 5308 event = kzalloc(sizeof(*event), GFP_KERNEL);
4827 if (!event) 5309 if (!event)
4828 return ERR_PTR(-ENOMEM); 5310 return ERR_PTR(-ENOMEM);
4829 5311
@@ -4841,6 +5323,7 @@ perf_event_alloc(struct perf_event_attr *attr,
4841 INIT_LIST_HEAD(&event->event_entry); 5323 INIT_LIST_HEAD(&event->event_entry);
4842 INIT_LIST_HEAD(&event->sibling_list); 5324 INIT_LIST_HEAD(&event->sibling_list);
4843 init_waitqueue_head(&event->waitq); 5325 init_waitqueue_head(&event->waitq);
5326 init_irq_work(&event->pending, perf_pending_event);
4844 5327
4845 mutex_init(&event->mmap_mutex); 5328 mutex_init(&event->mmap_mutex);
4846 5329
@@ -4848,7 +5331,6 @@ perf_event_alloc(struct perf_event_attr *attr,
4848 event->attr = *attr; 5331 event->attr = *attr;
4849 event->group_leader = group_leader; 5332 event->group_leader = group_leader;
4850 event->pmu = NULL; 5333 event->pmu = NULL;
4851 event->ctx = ctx;
4852 event->oncpu = -1; 5334 event->oncpu = -1;
4853 5335
4854 event->parent = parent_event; 5336 event->parent = parent_event;
@@ -4858,6 +5340,17 @@ perf_event_alloc(struct perf_event_attr *attr,
4858 5340
4859 event->state = PERF_EVENT_STATE_INACTIVE; 5341 event->state = PERF_EVENT_STATE_INACTIVE;
4860 5342
5343 if (task) {
5344 event->attach_state = PERF_ATTACH_TASK;
5345#ifdef CONFIG_HAVE_HW_BREAKPOINT
5346 /*
5347 * hw_breakpoint is a bit difficult here..
5348 */
5349 if (attr->type == PERF_TYPE_BREAKPOINT)
5350 event->hw.bp_target = task;
5351#endif
5352 }
5353
4861 if (!overflow_handler && parent_event) 5354 if (!overflow_handler && parent_event)
4862 overflow_handler = parent_event->overflow_handler; 5355 overflow_handler = parent_event->overflow_handler;
4863 5356
@@ -4882,29 +5375,8 @@ perf_event_alloc(struct perf_event_attr *attr,
4882 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) 5375 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4883 goto done; 5376 goto done;
4884 5377
4885 switch (attr->type) { 5378 pmu = perf_init_event(event);
4886 case PERF_TYPE_RAW:
4887 case PERF_TYPE_HARDWARE:
4888 case PERF_TYPE_HW_CACHE:
4889 pmu = hw_perf_event_init(event);
4890 break;
4891
4892 case PERF_TYPE_SOFTWARE:
4893 pmu = sw_perf_event_init(event);
4894 break;
4895 5379
4896 case PERF_TYPE_TRACEPOINT:
4897 pmu = tp_perf_event_init(event);
4898 break;
4899
4900 case PERF_TYPE_BREAKPOINT:
4901 pmu = bp_perf_event_init(event);
4902 break;
4903
4904
4905 default:
4906 break;
4907 }
4908done: 5380done:
4909 err = 0; 5381 err = 0;
4910 if (!pmu) 5382 if (!pmu)
@@ -4922,13 +5394,21 @@ done:
4922 event->pmu = pmu; 5394 event->pmu = pmu;
4923 5395
4924 if (!event->parent) { 5396 if (!event->parent) {
4925 atomic_inc(&nr_events); 5397 if (event->attach_state & PERF_ATTACH_TASK)
5398 jump_label_inc(&perf_task_events);
4926 if (event->attr.mmap || event->attr.mmap_data) 5399 if (event->attr.mmap || event->attr.mmap_data)
4927 atomic_inc(&nr_mmap_events); 5400 atomic_inc(&nr_mmap_events);
4928 if (event->attr.comm) 5401 if (event->attr.comm)
4929 atomic_inc(&nr_comm_events); 5402 atomic_inc(&nr_comm_events);
4930 if (event->attr.task) 5403 if (event->attr.task)
4931 atomic_inc(&nr_task_events); 5404 atomic_inc(&nr_task_events);
5405 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5406 err = get_callchain_buffers();
5407 if (err) {
5408 free_event(event);
5409 return ERR_PTR(err);
5410 }
5411 }
4932 } 5412 }
4933 5413
4934 return event; 5414 return event;
@@ -5076,12 +5556,16 @@ SYSCALL_DEFINE5(perf_event_open,
5076 struct perf_event_attr __user *, attr_uptr, 5556 struct perf_event_attr __user *, attr_uptr,
5077 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) 5557 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
5078{ 5558{
5079 struct perf_event *event, *group_leader = NULL, *output_event = NULL; 5559 struct perf_event *group_leader = NULL, *output_event = NULL;
5560 struct perf_event *event, *sibling;
5080 struct perf_event_attr attr; 5561 struct perf_event_attr attr;
5081 struct perf_event_context *ctx; 5562 struct perf_event_context *ctx;
5082 struct file *event_file = NULL; 5563 struct file *event_file = NULL;
5083 struct file *group_file = NULL; 5564 struct file *group_file = NULL;
5565 struct task_struct *task = NULL;
5566 struct pmu *pmu;
5084 int event_fd; 5567 int event_fd;
5568 int move_group = 0;
5085 int fput_needed = 0; 5569 int fput_needed = 0;
5086 int err; 5570 int err;
5087 5571
@@ -5107,20 +5591,11 @@ SYSCALL_DEFINE5(perf_event_open,
5107 if (event_fd < 0) 5591 if (event_fd < 0)
5108 return event_fd; 5592 return event_fd;
5109 5593
5110 /*
5111 * Get the target context (task or percpu):
5112 */
5113 ctx = find_get_context(pid, cpu);
5114 if (IS_ERR(ctx)) {
5115 err = PTR_ERR(ctx);
5116 goto err_fd;
5117 }
5118
5119 if (group_fd != -1) { 5594 if (group_fd != -1) {
5120 group_leader = perf_fget_light(group_fd, &fput_needed); 5595 group_leader = perf_fget_light(group_fd, &fput_needed);
5121 if (IS_ERR(group_leader)) { 5596 if (IS_ERR(group_leader)) {
5122 err = PTR_ERR(group_leader); 5597 err = PTR_ERR(group_leader);
5123 goto err_put_context; 5598 goto err_fd;
5124 } 5599 }
5125 group_file = group_leader->filp; 5600 group_file = group_leader->filp;
5126 if (flags & PERF_FLAG_FD_OUTPUT) 5601 if (flags & PERF_FLAG_FD_OUTPUT)
@@ -5129,6 +5604,58 @@ SYSCALL_DEFINE5(perf_event_open,
5129 group_leader = NULL; 5604 group_leader = NULL;
5130 } 5605 }
5131 5606
5607 if (pid != -1) {
5608 task = find_lively_task_by_vpid(pid);
5609 if (IS_ERR(task)) {
5610 err = PTR_ERR(task);
5611 goto err_group_fd;
5612 }
5613 }
5614
5615 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
5616 if (IS_ERR(event)) {
5617 err = PTR_ERR(event);
5618 goto err_task;
5619 }
5620
5621 /*
5622 * Special case software events and allow them to be part of
5623 * any hardware group.
5624 */
5625 pmu = event->pmu;
5626
5627 if (group_leader &&
5628 (is_software_event(event) != is_software_event(group_leader))) {
5629 if (is_software_event(event)) {
5630 /*
5631 * If event and group_leader are not both a software
5632 * event, and event is, then group leader is not.
5633 *
5634 * Allow the addition of software events to !software
5635 * groups, this is safe because software events never
5636 * fail to schedule.
5637 */
5638 pmu = group_leader->pmu;
5639 } else if (is_software_event(group_leader) &&
5640 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
5641 /*
5642 * In case the group is a pure software group, and we
5643 * try to add a hardware event, move the whole group to
5644 * the hardware context.
5645 */
5646 move_group = 1;
5647 }
5648 }
5649
5650 /*
5651 * Get the target context (task or percpu):
5652 */
5653 ctx = find_get_context(pmu, task, cpu);
5654 if (IS_ERR(ctx)) {
5655 err = PTR_ERR(ctx);
5656 goto err_alloc;
5657 }
5658
5132 /* 5659 /*
5133 * Look up the group leader (we will attach this event to it): 5660 * Look up the group leader (we will attach this event to it):
5134 */ 5661 */
@@ -5140,48 +5667,72 @@ SYSCALL_DEFINE5(perf_event_open,
5140 * becoming part of another group-sibling): 5667 * becoming part of another group-sibling):
5141 */ 5668 */
5142 if (group_leader->group_leader != group_leader) 5669 if (group_leader->group_leader != group_leader)
5143 goto err_put_context; 5670 goto err_context;
5144 /* 5671 /*
5145 * Do not allow to attach to a group in a different 5672 * Do not allow to attach to a group in a different
5146 * task or CPU context: 5673 * task or CPU context:
5147 */ 5674 */
5148 if (group_leader->ctx != ctx) 5675 if (move_group) {
5149 goto err_put_context; 5676 if (group_leader->ctx->type != ctx->type)
5677 goto err_context;
5678 } else {
5679 if (group_leader->ctx != ctx)
5680 goto err_context;
5681 }
5682
5150 /* 5683 /*
5151 * Only a group leader can be exclusive or pinned 5684 * Only a group leader can be exclusive or pinned
5152 */ 5685 */
5153 if (attr.exclusive || attr.pinned) 5686 if (attr.exclusive || attr.pinned)
5154 goto err_put_context; 5687 goto err_context;
5155 }
5156
5157 event = perf_event_alloc(&attr, cpu, ctx, group_leader,
5158 NULL, NULL, GFP_KERNEL);
5159 if (IS_ERR(event)) {
5160 err = PTR_ERR(event);
5161 goto err_put_context;
5162 } 5688 }
5163 5689
5164 if (output_event) { 5690 if (output_event) {
5165 err = perf_event_set_output(event, output_event); 5691 err = perf_event_set_output(event, output_event);
5166 if (err) 5692 if (err)
5167 goto err_free_put_context; 5693 goto err_context;
5168 } 5694 }
5169 5695
5170 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); 5696 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
5171 if (IS_ERR(event_file)) { 5697 if (IS_ERR(event_file)) {
5172 err = PTR_ERR(event_file); 5698 err = PTR_ERR(event_file);
5173 goto err_free_put_context; 5699 goto err_context;
5700 }
5701
5702 if (move_group) {
5703 struct perf_event_context *gctx = group_leader->ctx;
5704
5705 mutex_lock(&gctx->mutex);
5706 perf_event_remove_from_context(group_leader);
5707 list_for_each_entry(sibling, &group_leader->sibling_list,
5708 group_entry) {
5709 perf_event_remove_from_context(sibling);
5710 put_ctx(gctx);
5711 }
5712 mutex_unlock(&gctx->mutex);
5713 put_ctx(gctx);
5174 } 5714 }
5175 5715
5176 event->filp = event_file; 5716 event->filp = event_file;
5177 WARN_ON_ONCE(ctx->parent_ctx); 5717 WARN_ON_ONCE(ctx->parent_ctx);
5178 mutex_lock(&ctx->mutex); 5718 mutex_lock(&ctx->mutex);
5719
5720 if (move_group) {
5721 perf_install_in_context(ctx, group_leader, cpu);
5722 get_ctx(ctx);
5723 list_for_each_entry(sibling, &group_leader->sibling_list,
5724 group_entry) {
5725 perf_install_in_context(ctx, sibling, cpu);
5726 get_ctx(ctx);
5727 }
5728 }
5729
5179 perf_install_in_context(ctx, event, cpu); 5730 perf_install_in_context(ctx, event, cpu);
5180 ++ctx->generation; 5731 ++ctx->generation;
5181 mutex_unlock(&ctx->mutex); 5732 mutex_unlock(&ctx->mutex);
5182 5733
5183 event->owner = current; 5734 event->owner = current;
5184 get_task_struct(current); 5735
5185 mutex_lock(&current->perf_event_mutex); 5736 mutex_lock(&current->perf_event_mutex);
5186 list_add_tail(&event->owner_entry, &current->perf_event_list); 5737 list_add_tail(&event->owner_entry, &current->perf_event_list);
5187 mutex_unlock(&current->perf_event_mutex); 5738 mutex_unlock(&current->perf_event_mutex);
@@ -5196,11 +5747,15 @@ SYSCALL_DEFINE5(perf_event_open,
5196 fd_install(event_fd, event_file); 5747 fd_install(event_fd, event_file);
5197 return event_fd; 5748 return event_fd;
5198 5749
5199err_free_put_context: 5750err_context:
5751 put_ctx(ctx);
5752err_alloc:
5200 free_event(event); 5753 free_event(event);
5201err_put_context: 5754err_task:
5755 if (task)
5756 put_task_struct(task);
5757err_group_fd:
5202 fput_light(group_file, fput_needed); 5758 fput_light(group_file, fput_needed);
5203 put_ctx(ctx);
5204err_fd: 5759err_fd:
5205 put_unused_fd(event_fd); 5760 put_unused_fd(event_fd);
5206 return err; 5761 return err;
@@ -5211,32 +5766,31 @@ err_fd:
5211 * 5766 *
5212 * @attr: attributes of the counter to create 5767 * @attr: attributes of the counter to create
5213 * @cpu: cpu in which the counter is bound 5768 * @cpu: cpu in which the counter is bound
5214 * @pid: task to profile 5769 * @task: task to profile (NULL for percpu)
5215 */ 5770 */
5216struct perf_event * 5771struct perf_event *
5217perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, 5772perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5218 pid_t pid, 5773 struct task_struct *task,
5219 perf_overflow_handler_t overflow_handler) 5774 perf_overflow_handler_t overflow_handler)
5220{ 5775{
5221 struct perf_event *event;
5222 struct perf_event_context *ctx; 5776 struct perf_event_context *ctx;
5777 struct perf_event *event;
5223 int err; 5778 int err;
5224 5779
5225 /* 5780 /*
5226 * Get the target context (task or percpu): 5781 * Get the target context (task or percpu):
5227 */ 5782 */
5228 5783
5229 ctx = find_get_context(pid, cpu); 5784 event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
5230 if (IS_ERR(ctx)) {
5231 err = PTR_ERR(ctx);
5232 goto err_exit;
5233 }
5234
5235 event = perf_event_alloc(attr, cpu, ctx, NULL,
5236 NULL, overflow_handler, GFP_KERNEL);
5237 if (IS_ERR(event)) { 5785 if (IS_ERR(event)) {
5238 err = PTR_ERR(event); 5786 err = PTR_ERR(event);
5239 goto err_put_context; 5787 goto err;
5788 }
5789
5790 ctx = find_get_context(event->pmu, task, cpu);
5791 if (IS_ERR(ctx)) {
5792 err = PTR_ERR(ctx);
5793 goto err_free;
5240 } 5794 }
5241 5795
5242 event->filp = NULL; 5796 event->filp = NULL;
@@ -5246,120 +5800,15 @@ perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
5246 ++ctx->generation; 5800 ++ctx->generation;
5247 mutex_unlock(&ctx->mutex); 5801 mutex_unlock(&ctx->mutex);
5248 5802
5249 event->owner = current;
5250 get_task_struct(current);
5251 mutex_lock(&current->perf_event_mutex);
5252 list_add_tail(&event->owner_entry, &current->perf_event_list);
5253 mutex_unlock(&current->perf_event_mutex);
5254
5255 return event; 5803 return event;
5256 5804
5257 err_put_context: 5805err_free:
5258 put_ctx(ctx); 5806 free_event(event);
5259 err_exit: 5807err:
5260 return ERR_PTR(err); 5808 return ERR_PTR(err);
5261} 5809}
5262EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); 5810EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
5263 5811
5264/*
5265 * inherit a event from parent task to child task:
5266 */
5267static struct perf_event *
5268inherit_event(struct perf_event *parent_event,
5269 struct task_struct *parent,
5270 struct perf_event_context *parent_ctx,
5271 struct task_struct *child,
5272 struct perf_event *group_leader,
5273 struct perf_event_context *child_ctx)
5274{
5275 struct perf_event *child_event;
5276
5277 /*
5278 * Instead of creating recursive hierarchies of events,
5279 * we link inherited events back to the original parent,
5280 * which has a filp for sure, which we use as the reference
5281 * count:
5282 */
5283 if (parent_event->parent)
5284 parent_event = parent_event->parent;
5285
5286 child_event = perf_event_alloc(&parent_event->attr,
5287 parent_event->cpu, child_ctx,
5288 group_leader, parent_event,
5289 NULL, GFP_KERNEL);
5290 if (IS_ERR(child_event))
5291 return child_event;
5292 get_ctx(child_ctx);
5293
5294 /*
5295 * Make the child state follow the state of the parent event,
5296 * not its attr.disabled bit. We hold the parent's mutex,
5297 * so we won't race with perf_event_{en, dis}able_family.
5298 */
5299 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5300 child_event->state = PERF_EVENT_STATE_INACTIVE;
5301 else
5302 child_event->state = PERF_EVENT_STATE_OFF;
5303
5304 if (parent_event->attr.freq) {
5305 u64 sample_period = parent_event->hw.sample_period;
5306 struct hw_perf_event *hwc = &child_event->hw;
5307
5308 hwc->sample_period = sample_period;
5309 hwc->last_period = sample_period;
5310
5311 local64_set(&hwc->period_left, sample_period);
5312 }
5313
5314 child_event->overflow_handler = parent_event->overflow_handler;
5315
5316 /*
5317 * Link it up in the child's context:
5318 */
5319 add_event_to_ctx(child_event, child_ctx);
5320
5321 /*
5322 * Get a reference to the parent filp - we will fput it
5323 * when the child event exits. This is safe to do because
5324 * we are in the parent and we know that the filp still
5325 * exists and has a nonzero count:
5326 */
5327 atomic_long_inc(&parent_event->filp->f_count);
5328
5329 /*
5330 * Link this into the parent event's child list
5331 */
5332 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5333 mutex_lock(&parent_event->child_mutex);
5334 list_add_tail(&child_event->child_list, &parent_event->child_list);
5335 mutex_unlock(&parent_event->child_mutex);
5336
5337 return child_event;
5338}
5339
5340static int inherit_group(struct perf_event *parent_event,
5341 struct task_struct *parent,
5342 struct perf_event_context *parent_ctx,
5343 struct task_struct *child,
5344 struct perf_event_context *child_ctx)
5345{
5346 struct perf_event *leader;
5347 struct perf_event *sub;
5348 struct perf_event *child_ctr;
5349
5350 leader = inherit_event(parent_event, parent, parent_ctx,
5351 child, NULL, child_ctx);
5352 if (IS_ERR(leader))
5353 return PTR_ERR(leader);
5354 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5355 child_ctr = inherit_event(sub, parent, parent_ctx,
5356 child, leader, child_ctx);
5357 if (IS_ERR(child_ctr))
5358 return PTR_ERR(child_ctr);
5359 }
5360 return 0;
5361}
5362
5363static void sync_child_event(struct perf_event *child_event, 5812static void sync_child_event(struct perf_event *child_event,
5364 struct task_struct *child) 5813 struct task_struct *child)
5365{ 5814{
@@ -5416,16 +5865,13 @@ __perf_event_exit_task(struct perf_event *child_event,
5416 } 5865 }
5417} 5866}
5418 5867
5419/* 5868static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
5420 * When a child task exits, feed back event values to parent events.
5421 */
5422void perf_event_exit_task(struct task_struct *child)
5423{ 5869{
5424 struct perf_event *child_event, *tmp; 5870 struct perf_event *child_event, *tmp;
5425 struct perf_event_context *child_ctx; 5871 struct perf_event_context *child_ctx;
5426 unsigned long flags; 5872 unsigned long flags;
5427 5873
5428 if (likely(!child->perf_event_ctxp)) { 5874 if (likely(!child->perf_event_ctxp[ctxn])) {
5429 perf_event_task(child, NULL, 0); 5875 perf_event_task(child, NULL, 0);
5430 return; 5876 return;
5431 } 5877 }
@@ -5437,8 +5883,8 @@ void perf_event_exit_task(struct task_struct *child)
5437 * scheduled, so we are now safe from rescheduling changing 5883 * scheduled, so we are now safe from rescheduling changing
5438 * our context. 5884 * our context.
5439 */ 5885 */
5440 child_ctx = child->perf_event_ctxp; 5886 child_ctx = child->perf_event_ctxp[ctxn];
5441 __perf_event_task_sched_out(child_ctx); 5887 task_ctx_sched_out(child_ctx, EVENT_ALL);
5442 5888
5443 /* 5889 /*
5444 * Take the context lock here so that if find_get_context is 5890 * Take the context lock here so that if find_get_context is
@@ -5446,7 +5892,7 @@ void perf_event_exit_task(struct task_struct *child)
5446 * incremented the context's refcount before we do put_ctx below. 5892 * incremented the context's refcount before we do put_ctx below.
5447 */ 5893 */
5448 raw_spin_lock(&child_ctx->lock); 5894 raw_spin_lock(&child_ctx->lock);
5449 child->perf_event_ctxp = NULL; 5895 child->perf_event_ctxp[ctxn] = NULL;
5450 /* 5896 /*
5451 * If this context is a clone; unclone it so it can't get 5897 * If this context is a clone; unclone it so it can't get
5452 * swapped to another process while we're removing all 5898 * swapped to another process while we're removing all
@@ -5499,6 +5945,33 @@ again:
5499 put_ctx(child_ctx); 5945 put_ctx(child_ctx);
5500} 5946}
5501 5947
5948/*
5949 * When a child task exits, feed back event values to parent events.
5950 */
5951void perf_event_exit_task(struct task_struct *child)
5952{
5953 struct perf_event *event, *tmp;
5954 int ctxn;
5955
5956 mutex_lock(&child->perf_event_mutex);
5957 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
5958 owner_entry) {
5959 list_del_init(&event->owner_entry);
5960
5961 /*
5962 * Ensure the list deletion is visible before we clear
5963 * the owner, closes a race against perf_release() where
5964 * we need to serialize on the owner->perf_event_mutex.
5965 */
5966 smp_wmb();
5967 event->owner = NULL;
5968 }
5969 mutex_unlock(&child->perf_event_mutex);
5970
5971 for_each_task_context_nr(ctxn)
5972 perf_event_exit_task_context(child, ctxn);
5973}
5974
5502static void perf_free_event(struct perf_event *event, 5975static void perf_free_event(struct perf_event *event,
5503 struct perf_event_context *ctx) 5976 struct perf_event_context *ctx)
5504{ 5977{
@@ -5520,48 +5993,166 @@ static void perf_free_event(struct perf_event *event,
5520 5993
5521/* 5994/*
5522 * free an unexposed, unused context as created by inheritance by 5995 * free an unexposed, unused context as created by inheritance by
5523 * init_task below, used by fork() in case of fail. 5996 * perf_event_init_task below, used by fork() in case of fail.
5524 */ 5997 */
5525void perf_event_free_task(struct task_struct *task) 5998void perf_event_free_task(struct task_struct *task)
5526{ 5999{
5527 struct perf_event_context *ctx = task->perf_event_ctxp; 6000 struct perf_event_context *ctx;
5528 struct perf_event *event, *tmp; 6001 struct perf_event *event, *tmp;
6002 int ctxn;
5529 6003
5530 if (!ctx) 6004 for_each_task_context_nr(ctxn) {
5531 return; 6005 ctx = task->perf_event_ctxp[ctxn];
6006 if (!ctx)
6007 continue;
5532 6008
5533 mutex_lock(&ctx->mutex); 6009 mutex_lock(&ctx->mutex);
5534again: 6010again:
5535 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) 6011 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
5536 perf_free_event(event, ctx); 6012 group_entry)
6013 perf_free_event(event, ctx);
5537 6014
5538 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, 6015 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5539 group_entry) 6016 group_entry)
5540 perf_free_event(event, ctx); 6017 perf_free_event(event, ctx);
5541 6018
5542 if (!list_empty(&ctx->pinned_groups) || 6019 if (!list_empty(&ctx->pinned_groups) ||
5543 !list_empty(&ctx->flexible_groups)) 6020 !list_empty(&ctx->flexible_groups))
5544 goto again; 6021 goto again;
5545 6022
5546 mutex_unlock(&ctx->mutex); 6023 mutex_unlock(&ctx->mutex);
5547 6024
5548 put_ctx(ctx); 6025 put_ctx(ctx);
6026 }
6027}
6028
6029void perf_event_delayed_put(struct task_struct *task)
6030{
6031 int ctxn;
6032
6033 for_each_task_context_nr(ctxn)
6034 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6035}
6036
6037/*
6038 * inherit a event from parent task to child task:
6039 */
6040static struct perf_event *
6041inherit_event(struct perf_event *parent_event,
6042 struct task_struct *parent,
6043 struct perf_event_context *parent_ctx,
6044 struct task_struct *child,
6045 struct perf_event *group_leader,
6046 struct perf_event_context *child_ctx)
6047{
6048 struct perf_event *child_event;
6049 unsigned long flags;
6050
6051 /*
6052 * Instead of creating recursive hierarchies of events,
6053 * we link inherited events back to the original parent,
6054 * which has a filp for sure, which we use as the reference
6055 * count:
6056 */
6057 if (parent_event->parent)
6058 parent_event = parent_event->parent;
6059
6060 child_event = perf_event_alloc(&parent_event->attr,
6061 parent_event->cpu,
6062 child,
6063 group_leader, parent_event,
6064 NULL);
6065 if (IS_ERR(child_event))
6066 return child_event;
6067 get_ctx(child_ctx);
6068
6069 /*
6070 * Make the child state follow the state of the parent event,
6071 * not its attr.disabled bit. We hold the parent's mutex,
6072 * so we won't race with perf_event_{en, dis}able_family.
6073 */
6074 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6075 child_event->state = PERF_EVENT_STATE_INACTIVE;
6076 else
6077 child_event->state = PERF_EVENT_STATE_OFF;
6078
6079 if (parent_event->attr.freq) {
6080 u64 sample_period = parent_event->hw.sample_period;
6081 struct hw_perf_event *hwc = &child_event->hw;
6082
6083 hwc->sample_period = sample_period;
6084 hwc->last_period = sample_period;
6085
6086 local64_set(&hwc->period_left, sample_period);
6087 }
6088
6089 child_event->ctx = child_ctx;
6090 child_event->overflow_handler = parent_event->overflow_handler;
6091
6092 /*
6093 * Link it up in the child's context:
6094 */
6095 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6096 add_event_to_ctx(child_event, child_ctx);
6097 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6098
6099 /*
6100 * Get a reference to the parent filp - we will fput it
6101 * when the child event exits. This is safe to do because
6102 * we are in the parent and we know that the filp still
6103 * exists and has a nonzero count:
6104 */
6105 atomic_long_inc(&parent_event->filp->f_count);
6106
6107 /*
6108 * Link this into the parent event's child list
6109 */
6110 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6111 mutex_lock(&parent_event->child_mutex);
6112 list_add_tail(&child_event->child_list, &parent_event->child_list);
6113 mutex_unlock(&parent_event->child_mutex);
6114
6115 return child_event;
6116}
6117
6118static int inherit_group(struct perf_event *parent_event,
6119 struct task_struct *parent,
6120 struct perf_event_context *parent_ctx,
6121 struct task_struct *child,
6122 struct perf_event_context *child_ctx)
6123{
6124 struct perf_event *leader;
6125 struct perf_event *sub;
6126 struct perf_event *child_ctr;
6127
6128 leader = inherit_event(parent_event, parent, parent_ctx,
6129 child, NULL, child_ctx);
6130 if (IS_ERR(leader))
6131 return PTR_ERR(leader);
6132 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6133 child_ctr = inherit_event(sub, parent, parent_ctx,
6134 child, leader, child_ctx);
6135 if (IS_ERR(child_ctr))
6136 return PTR_ERR(child_ctr);
6137 }
6138 return 0;
5549} 6139}
5550 6140
5551static int 6141static int
5552inherit_task_group(struct perf_event *event, struct task_struct *parent, 6142inherit_task_group(struct perf_event *event, struct task_struct *parent,
5553 struct perf_event_context *parent_ctx, 6143 struct perf_event_context *parent_ctx,
5554 struct task_struct *child, 6144 struct task_struct *child, int ctxn,
5555 int *inherited_all) 6145 int *inherited_all)
5556{ 6146{
5557 int ret; 6147 int ret;
5558 struct perf_event_context *child_ctx = child->perf_event_ctxp; 6148 struct perf_event_context *child_ctx;
5559 6149
5560 if (!event->attr.inherit) { 6150 if (!event->attr.inherit) {
5561 *inherited_all = 0; 6151 *inherited_all = 0;
5562 return 0; 6152 return 0;
5563 } 6153 }
5564 6154
6155 child_ctx = child->perf_event_ctxp[ctxn];
5565 if (!child_ctx) { 6156 if (!child_ctx) {
5566 /* 6157 /*
5567 * This is executed from the parent task context, so 6158 * This is executed from the parent task context, so
@@ -5570,14 +6161,11 @@ inherit_task_group(struct perf_event *event, struct task_struct *parent,
5570 * child. 6161 * child.
5571 */ 6162 */
5572 6163
5573 child_ctx = kzalloc(sizeof(struct perf_event_context), 6164 child_ctx = alloc_perf_context(event->pmu, child);
5574 GFP_KERNEL);
5575 if (!child_ctx) 6165 if (!child_ctx)
5576 return -ENOMEM; 6166 return -ENOMEM;
5577 6167
5578 __perf_event_init_context(child_ctx, child); 6168 child->perf_event_ctxp[ctxn] = child_ctx;
5579 child->perf_event_ctxp = child_ctx;
5580 get_task_struct(child);
5581 } 6169 }
5582 6170
5583 ret = inherit_group(event, parent, parent_ctx, 6171 ret = inherit_group(event, parent, parent_ctx,
@@ -5589,32 +6177,32 @@ inherit_task_group(struct perf_event *event, struct task_struct *parent,
5589 return ret; 6177 return ret;
5590} 6178}
5591 6179
5592
5593/* 6180/*
5594 * Initialize the perf_event context in task_struct 6181 * Initialize the perf_event context in task_struct
5595 */ 6182 */
5596int perf_event_init_task(struct task_struct *child) 6183int perf_event_init_context(struct task_struct *child, int ctxn)
5597{ 6184{
5598 struct perf_event_context *child_ctx, *parent_ctx; 6185 struct perf_event_context *child_ctx, *parent_ctx;
5599 struct perf_event_context *cloned_ctx; 6186 struct perf_event_context *cloned_ctx;
5600 struct perf_event *event; 6187 struct perf_event *event;
5601 struct task_struct *parent = current; 6188 struct task_struct *parent = current;
5602 int inherited_all = 1; 6189 int inherited_all = 1;
6190 unsigned long flags;
5603 int ret = 0; 6191 int ret = 0;
5604 6192
5605 child->perf_event_ctxp = NULL; 6193 child->perf_event_ctxp[ctxn] = NULL;
5606 6194
5607 mutex_init(&child->perf_event_mutex); 6195 mutex_init(&child->perf_event_mutex);
5608 INIT_LIST_HEAD(&child->perf_event_list); 6196 INIT_LIST_HEAD(&child->perf_event_list);
5609 6197
5610 if (likely(!parent->perf_event_ctxp)) 6198 if (likely(!parent->perf_event_ctxp[ctxn]))
5611 return 0; 6199 return 0;
5612 6200
5613 /* 6201 /*
5614 * If the parent's context is a clone, pin it so it won't get 6202 * If the parent's context is a clone, pin it so it won't get
5615 * swapped under us. 6203 * swapped under us.
5616 */ 6204 */
5617 parent_ctx = perf_pin_task_context(parent); 6205 parent_ctx = perf_pin_task_context(parent, ctxn);
5618 6206
5619 /* 6207 /*
5620 * No need to check if parent_ctx != NULL here; since we saw 6208 * No need to check if parent_ctx != NULL here; since we saw
@@ -5634,20 +6222,33 @@ int perf_event_init_task(struct task_struct *child)
5634 * the list, not manipulating it: 6222 * the list, not manipulating it:
5635 */ 6223 */
5636 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { 6224 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5637 ret = inherit_task_group(event, parent, parent_ctx, child, 6225 ret = inherit_task_group(event, parent, parent_ctx,
5638 &inherited_all); 6226 child, ctxn, &inherited_all);
5639 if (ret) 6227 if (ret)
5640 break; 6228 break;
5641 } 6229 }
5642 6230
6231 /*
6232 * We can't hold ctx->lock when iterating the ->flexible_group list due
6233 * to allocations, but we need to prevent rotation because
6234 * rotate_ctx() will change the list from interrupt context.
6235 */
6236 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6237 parent_ctx->rotate_disable = 1;
6238 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6239
5643 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { 6240 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5644 ret = inherit_task_group(event, parent, parent_ctx, child, 6241 ret = inherit_task_group(event, parent, parent_ctx,
5645 &inherited_all); 6242 child, ctxn, &inherited_all);
5646 if (ret) 6243 if (ret)
5647 break; 6244 break;
5648 } 6245 }
5649 6246
5650 child_ctx = child->perf_event_ctxp; 6247 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6248 parent_ctx->rotate_disable = 0;
6249 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6250
6251 child_ctx = child->perf_event_ctxp[ctxn];
5651 6252
5652 if (child_ctx && inherited_all) { 6253 if (child_ctx && inherited_all) {
5653 /* 6254 /*
@@ -5676,63 +6277,98 @@ int perf_event_init_task(struct task_struct *child)
5676 return ret; 6277 return ret;
5677} 6278}
5678 6279
6280/*
6281 * Initialize the perf_event context in task_struct
6282 */
6283int perf_event_init_task(struct task_struct *child)
6284{
6285 int ctxn, ret;
6286
6287 for_each_task_context_nr(ctxn) {
6288 ret = perf_event_init_context(child, ctxn);
6289 if (ret)
6290 return ret;
6291 }
6292
6293 return 0;
6294}
6295
5679static void __init perf_event_init_all_cpus(void) 6296static void __init perf_event_init_all_cpus(void)
5680{ 6297{
6298 struct swevent_htable *swhash;
5681 int cpu; 6299 int cpu;
5682 struct perf_cpu_context *cpuctx;
5683 6300
5684 for_each_possible_cpu(cpu) { 6301 for_each_possible_cpu(cpu) {
5685 cpuctx = &per_cpu(perf_cpu_context, cpu); 6302 swhash = &per_cpu(swevent_htable, cpu);
5686 mutex_init(&cpuctx->hlist_mutex); 6303 mutex_init(&swhash->hlist_mutex);
5687 __perf_event_init_context(&cpuctx->ctx, NULL); 6304 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
5688 } 6305 }
5689} 6306}
5690 6307
5691static void __cpuinit perf_event_init_cpu(int cpu) 6308static void __cpuinit perf_event_init_cpu(int cpu)
5692{ 6309{
5693 struct perf_cpu_context *cpuctx; 6310 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5694
5695 cpuctx = &per_cpu(perf_cpu_context, cpu);
5696 6311
5697 spin_lock(&perf_resource_lock); 6312 mutex_lock(&swhash->hlist_mutex);
5698 cpuctx->max_pertask = perf_max_events - perf_reserved_percpu; 6313 if (swhash->hlist_refcount > 0) {
5699 spin_unlock(&perf_resource_lock);
5700
5701 mutex_lock(&cpuctx->hlist_mutex);
5702 if (cpuctx->hlist_refcount > 0) {
5703 struct swevent_hlist *hlist; 6314 struct swevent_hlist *hlist;
5704 6315
5705 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); 6316 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
5706 WARN_ON_ONCE(!hlist); 6317 WARN_ON(!hlist);
5707 rcu_assign_pointer(cpuctx->swevent_hlist, hlist); 6318 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5708 } 6319 }
5709 mutex_unlock(&cpuctx->hlist_mutex); 6320 mutex_unlock(&swhash->hlist_mutex);
5710} 6321}
5711 6322
5712#ifdef CONFIG_HOTPLUG_CPU 6323#ifdef CONFIG_HOTPLUG_CPU
5713static void __perf_event_exit_cpu(void *info) 6324static void perf_pmu_rotate_stop(struct pmu *pmu)
5714{ 6325{
5715 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context); 6326 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
5716 struct perf_event_context *ctx = &cpuctx->ctx; 6327
6328 WARN_ON(!irqs_disabled());
6329
6330 list_del_init(&cpuctx->rotation_list);
6331}
6332
6333static void __perf_event_exit_context(void *__info)
6334{
6335 struct perf_event_context *ctx = __info;
5717 struct perf_event *event, *tmp; 6336 struct perf_event *event, *tmp;
5718 6337
6338 perf_pmu_rotate_stop(ctx->pmu);
6339
5719 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) 6340 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5720 __perf_event_remove_from_context(event); 6341 __perf_event_remove_from_context(event);
5721 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) 6342 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5722 __perf_event_remove_from_context(event); 6343 __perf_event_remove_from_context(event);
5723} 6344}
6345
6346static void perf_event_exit_cpu_context(int cpu)
6347{
6348 struct perf_event_context *ctx;
6349 struct pmu *pmu;
6350 int idx;
6351
6352 idx = srcu_read_lock(&pmus_srcu);
6353 list_for_each_entry_rcu(pmu, &pmus, entry) {
6354 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6355
6356 mutex_lock(&ctx->mutex);
6357 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6358 mutex_unlock(&ctx->mutex);
6359 }
6360 srcu_read_unlock(&pmus_srcu, idx);
6361}
6362
5724static void perf_event_exit_cpu(int cpu) 6363static void perf_event_exit_cpu(int cpu)
5725{ 6364{
5726 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); 6365 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5727 struct perf_event_context *ctx = &cpuctx->ctx;
5728 6366
5729 mutex_lock(&cpuctx->hlist_mutex); 6367 mutex_lock(&swhash->hlist_mutex);
5730 swevent_hlist_release(cpuctx); 6368 swevent_hlist_release(swhash);
5731 mutex_unlock(&cpuctx->hlist_mutex); 6369 mutex_unlock(&swhash->hlist_mutex);
5732 6370
5733 mutex_lock(&ctx->mutex); 6371 perf_event_exit_cpu_context(cpu);
5734 smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5735 mutex_unlock(&ctx->mutex);
5736} 6372}
5737#else 6373#else
5738static inline void perf_event_exit_cpu(int cpu) { } 6374static inline void perf_event_exit_cpu(int cpu) { }
@@ -5743,15 +6379,15 @@ perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5743{ 6379{
5744 unsigned int cpu = (long)hcpu; 6380 unsigned int cpu = (long)hcpu;
5745 6381
5746 switch (action) { 6382 switch (action & ~CPU_TASKS_FROZEN) {
5747 6383
5748 case CPU_UP_PREPARE: 6384 case CPU_UP_PREPARE:
5749 case CPU_UP_PREPARE_FROZEN: 6385 case CPU_DOWN_FAILED:
5750 perf_event_init_cpu(cpu); 6386 perf_event_init_cpu(cpu);
5751 break; 6387 break;
5752 6388
6389 case CPU_UP_CANCELED:
5753 case CPU_DOWN_PREPARE: 6390 case CPU_DOWN_PREPARE:
5754 case CPU_DOWN_PREPARE_FROZEN:
5755 perf_event_exit_cpu(cpu); 6391 perf_event_exit_cpu(cpu);
5756 break; 6392 break;
5757 6393
@@ -5762,118 +6398,18 @@ perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5762 return NOTIFY_OK; 6398 return NOTIFY_OK;
5763} 6399}
5764 6400
5765/*
5766 * This has to have a higher priority than migration_notifier in sched.c.
5767 */
5768static struct notifier_block __cpuinitdata perf_cpu_nb = {
5769 .notifier_call = perf_cpu_notify,
5770 .priority = 20,
5771};
5772
5773void __init perf_event_init(void) 6401void __init perf_event_init(void)
5774{ 6402{
5775 perf_event_init_all_cpus(); 6403 int ret;
5776 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5777 (void *)(long)smp_processor_id());
5778 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5779 (void *)(long)smp_processor_id());
5780 register_cpu_notifier(&perf_cpu_nb);
5781}
5782
5783static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5784 struct sysdev_class_attribute *attr,
5785 char *buf)
5786{
5787 return sprintf(buf, "%d\n", perf_reserved_percpu);
5788}
5789
5790static ssize_t
5791perf_set_reserve_percpu(struct sysdev_class *class,
5792 struct sysdev_class_attribute *attr,
5793 const char *buf,
5794 size_t count)
5795{
5796 struct perf_cpu_context *cpuctx;
5797 unsigned long val;
5798 int err, cpu, mpt;
5799
5800 err = strict_strtoul(buf, 10, &val);
5801 if (err)
5802 return err;
5803 if (val > perf_max_events)
5804 return -EINVAL;
5805
5806 spin_lock(&perf_resource_lock);
5807 perf_reserved_percpu = val;
5808 for_each_online_cpu(cpu) {
5809 cpuctx = &per_cpu(perf_cpu_context, cpu);
5810 raw_spin_lock_irq(&cpuctx->ctx.lock);
5811 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5812 perf_max_events - perf_reserved_percpu);
5813 cpuctx->max_pertask = mpt;
5814 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5815 }
5816 spin_unlock(&perf_resource_lock);
5817
5818 return count;
5819}
5820
5821static ssize_t perf_show_overcommit(struct sysdev_class *class,
5822 struct sysdev_class_attribute *attr,
5823 char *buf)
5824{
5825 return sprintf(buf, "%d\n", perf_overcommit);
5826}
5827
5828static ssize_t
5829perf_set_overcommit(struct sysdev_class *class,
5830 struct sysdev_class_attribute *attr,
5831 const char *buf, size_t count)
5832{
5833 unsigned long val;
5834 int err;
5835
5836 err = strict_strtoul(buf, 10, &val);
5837 if (err)
5838 return err;
5839 if (val > 1)
5840 return -EINVAL;
5841
5842 spin_lock(&perf_resource_lock);
5843 perf_overcommit = val;
5844 spin_unlock(&perf_resource_lock);
5845
5846 return count;
5847}
5848
5849static SYSDEV_CLASS_ATTR(
5850 reserve_percpu,
5851 0644,
5852 perf_show_reserve_percpu,
5853 perf_set_reserve_percpu
5854 );
5855
5856static SYSDEV_CLASS_ATTR(
5857 overcommit,
5858 0644,
5859 perf_show_overcommit,
5860 perf_set_overcommit
5861 );
5862
5863static struct attribute *perfclass_attrs[] = {
5864 &attr_reserve_percpu.attr,
5865 &attr_overcommit.attr,
5866 NULL
5867};
5868 6404
5869static struct attribute_group perfclass_attr_group = { 6405 perf_event_init_all_cpus();
5870 .attrs = perfclass_attrs, 6406 init_srcu_struct(&pmus_srcu);
5871 .name = "perf_events", 6407 perf_pmu_register(&perf_swevent);
5872}; 6408 perf_pmu_register(&perf_cpu_clock);
6409 perf_pmu_register(&perf_task_clock);
6410 perf_tp_register();
6411 perf_cpu_notifier(perf_cpu_notify);
5873 6412
5874static int __init perf_event_sysfs_init(void) 6413 ret = init_hw_breakpoint();
5875{ 6414 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
5876 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5877 &perfclass_attr_group);
5878} 6415}
5879device_initcall(perf_event_sysfs_init);
generated by cgit v1.2.2 (git 2.25.0) at 2025-09-03 09:39:44 -0400