2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP)
125 * branch priv levels that need permission checks
127 #define PERF_SAMPLE_BRANCH_PERM_PLM \
128 (PERF_SAMPLE_BRANCH_KERNEL |\
129 PERF_SAMPLE_BRANCH_HV)
132 EVENT_FLEXIBLE = 0x1,
134 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
138 * perf_sched_events : >0 events exist
139 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 struct static_key_deferred perf_sched_events __read_mostly;
142 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
143 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
145 static atomic_t nr_mmap_events __read_mostly;
146 static atomic_t nr_comm_events __read_mostly;
147 static atomic_t nr_task_events __read_mostly;
149 static LIST_HEAD(pmus);
150 static DEFINE_MUTEX(pmus_lock);
151 static struct srcu_struct pmus_srcu;
154 * perf event paranoia level:
155 * -1 - not paranoid at all
156 * 0 - disallow raw tracepoint access for unpriv
157 * 1 - disallow cpu events for unpriv
158 * 2 - disallow kernel profiling for unpriv
160 int sysctl_perf_event_paranoid __read_mostly = 1;
162 /* Minimum for 512 kiB + 1 user control page */
163 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
166 * max perf event sample rate
168 #define DEFAULT_MAX_SAMPLE_RATE 100000
169 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
170 static int max_samples_per_tick __read_mostly =
171 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
173 int perf_proc_update_handler(struct ctl_table *table, int write,
174 void __user *buffer, size_t *lenp,
177 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
182 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
187 static atomic64_t perf_event_id;
189 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
190 enum event_type_t event_type);
192 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
193 enum event_type_t event_type,
194 struct task_struct *task);
196 static void update_context_time(struct perf_event_context *ctx);
197 static u64 perf_event_time(struct perf_event *event);
199 static void ring_buffer_attach(struct perf_event *event,
200 struct ring_buffer *rb);
202 void __weak perf_event_print_debug(void) { }
204 extern __weak const char *perf_pmu_name(void)
209 static inline u64 perf_clock(void)
211 return local_clock();
214 static inline struct perf_cpu_context *
215 __get_cpu_context(struct perf_event_context *ctx)
217 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
220 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
221 struct perf_event_context *ctx)
223 raw_spin_lock(&cpuctx->ctx.lock);
225 raw_spin_lock(&ctx->lock);
228 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
229 struct perf_event_context *ctx)
232 raw_spin_unlock(&ctx->lock);
233 raw_spin_unlock(&cpuctx->ctx.lock);
236 #ifdef CONFIG_CGROUP_PERF
239 * perf_cgroup_info keeps track of time_enabled for a cgroup.
240 * This is a per-cpu dynamically allocated data structure.
242 struct perf_cgroup_info {
248 struct cgroup_subsys_state css;
249 struct perf_cgroup_info __percpu *info;
253 * Must ensure cgroup is pinned (css_get) before calling
254 * this function. In other words, we cannot call this function
255 * if there is no cgroup event for the current CPU context.
257 static inline struct perf_cgroup *
258 perf_cgroup_from_task(struct task_struct *task)
260 return container_of(task_subsys_state(task, perf_subsys_id),
261 struct perf_cgroup, css);
265 perf_cgroup_match(struct perf_event *event)
267 struct perf_event_context *ctx = event->ctx;
268 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
270 /* @event doesn't care about cgroup */
274 /* wants specific cgroup scope but @cpuctx isn't associated with any */
279 * Cgroup scoping is recursive. An event enabled for a cgroup is
280 * also enabled for all its descendant cgroups. If @cpuctx's
281 * cgroup is a descendant of @event's (the test covers identity
282 * case), it's a match.
284 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
285 event->cgrp->css.cgroup);
288 static inline bool perf_tryget_cgroup(struct perf_event *event)
290 return css_tryget(&event->cgrp->css);
293 static inline void perf_put_cgroup(struct perf_event *event)
295 css_put(&event->cgrp->css);
298 static inline void perf_detach_cgroup(struct perf_event *event)
300 perf_put_cgroup(event);
304 static inline int is_cgroup_event(struct perf_event *event)
306 return event->cgrp != NULL;
309 static inline u64 perf_cgroup_event_time(struct perf_event *event)
311 struct perf_cgroup_info *t;
313 t = per_cpu_ptr(event->cgrp->info, event->cpu);
317 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
319 struct perf_cgroup_info *info;
324 info = this_cpu_ptr(cgrp->info);
326 info->time += now - info->timestamp;
327 info->timestamp = now;
330 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
332 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
334 __update_cgrp_time(cgrp_out);
337 static inline void update_cgrp_time_from_event(struct perf_event *event)
339 struct perf_cgroup *cgrp;
342 * ensure we access cgroup data only when needed and
343 * when we know the cgroup is pinned (css_get)
345 if (!is_cgroup_event(event))
348 cgrp = perf_cgroup_from_task(current);
350 * Do not update time when cgroup is not active
352 if (cgrp == event->cgrp)
353 __update_cgrp_time(event->cgrp);
357 perf_cgroup_set_timestamp(struct task_struct *task,
358 struct perf_event_context *ctx)
360 struct perf_cgroup *cgrp;
361 struct perf_cgroup_info *info;
364 * ctx->lock held by caller
365 * ensure we do not access cgroup data
366 * unless we have the cgroup pinned (css_get)
368 if (!task || !ctx->nr_cgroups)
371 cgrp = perf_cgroup_from_task(task);
372 info = this_cpu_ptr(cgrp->info);
373 info->timestamp = ctx->timestamp;
376 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
377 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
380 * reschedule events based on the cgroup constraint of task.
382 * mode SWOUT : schedule out everything
383 * mode SWIN : schedule in based on cgroup for next
385 void perf_cgroup_switch(struct task_struct *task, int mode)
387 struct perf_cpu_context *cpuctx;
392 * disable interrupts to avoid geting nr_cgroup
393 * changes via __perf_event_disable(). Also
396 local_irq_save(flags);
399 * we reschedule only in the presence of cgroup
400 * constrained events.
404 list_for_each_entry_rcu(pmu, &pmus, entry) {
405 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
406 if (cpuctx->unique_pmu != pmu)
407 continue; /* ensure we process each cpuctx once */
410 * perf_cgroup_events says at least one
411 * context on this CPU has cgroup events.
413 * ctx->nr_cgroups reports the number of cgroup
414 * events for a context.
416 if (cpuctx->ctx.nr_cgroups > 0) {
417 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
418 perf_pmu_disable(cpuctx->ctx.pmu);
420 if (mode & PERF_CGROUP_SWOUT) {
421 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
423 * must not be done before ctxswout due
424 * to event_filter_match() in event_sched_out()
429 if (mode & PERF_CGROUP_SWIN) {
430 WARN_ON_ONCE(cpuctx->cgrp);
432 * set cgrp before ctxsw in to allow
433 * event_filter_match() to not have to pass
436 cpuctx->cgrp = perf_cgroup_from_task(task);
437 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
439 perf_pmu_enable(cpuctx->ctx.pmu);
440 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
446 local_irq_restore(flags);
449 static inline void perf_cgroup_sched_out(struct task_struct *task,
450 struct task_struct *next)
452 struct perf_cgroup *cgrp1;
453 struct perf_cgroup *cgrp2 = NULL;
456 * we come here when we know perf_cgroup_events > 0
458 cgrp1 = perf_cgroup_from_task(task);
461 * next is NULL when called from perf_event_enable_on_exec()
462 * that will systematically cause a cgroup_switch()
465 cgrp2 = perf_cgroup_from_task(next);
468 * only schedule out current cgroup events if we know
469 * that we are switching to a different cgroup. Otherwise,
470 * do no touch the cgroup events.
473 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
476 static inline void perf_cgroup_sched_in(struct task_struct *prev,
477 struct task_struct *task)
479 struct perf_cgroup *cgrp1;
480 struct perf_cgroup *cgrp2 = NULL;
483 * we come here when we know perf_cgroup_events > 0
485 cgrp1 = perf_cgroup_from_task(task);
487 /* prev can never be NULL */
488 cgrp2 = perf_cgroup_from_task(prev);
491 * only need to schedule in cgroup events if we are changing
492 * cgroup during ctxsw. Cgroup events were not scheduled
493 * out of ctxsw out if that was not the case.
496 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
499 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
500 struct perf_event_attr *attr,
501 struct perf_event *group_leader)
503 struct perf_cgroup *cgrp;
504 struct cgroup_subsys_state *css;
505 struct fd f = fdget(fd);
511 css = cgroup_css_from_dir(f.file, perf_subsys_id);
517 cgrp = container_of(css, struct perf_cgroup, css);
520 /* must be done before we fput() the file */
521 if (!perf_tryget_cgroup(event)) {
528 * all events in a group must monitor
529 * the same cgroup because a task belongs
530 * to only one perf cgroup at a time
532 if (group_leader && group_leader->cgrp != cgrp) {
533 perf_detach_cgroup(event);
542 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
544 struct perf_cgroup_info *t;
545 t = per_cpu_ptr(event->cgrp->info, event->cpu);
546 event->shadow_ctx_time = now - t->timestamp;
550 perf_cgroup_defer_enabled(struct perf_event *event)
553 * when the current task's perf cgroup does not match
554 * the event's, we need to remember to call the
555 * perf_mark_enable() function the first time a task with
556 * a matching perf cgroup is scheduled in.
558 if (is_cgroup_event(event) && !perf_cgroup_match(event))
559 event->cgrp_defer_enabled = 1;
563 perf_cgroup_mark_enabled(struct perf_event *event,
564 struct perf_event_context *ctx)
566 struct perf_event *sub;
567 u64 tstamp = perf_event_time(event);
569 if (!event->cgrp_defer_enabled)
572 event->cgrp_defer_enabled = 0;
574 event->tstamp_enabled = tstamp - event->total_time_enabled;
575 list_for_each_entry(sub, &event->sibling_list, group_entry) {
576 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
577 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
578 sub->cgrp_defer_enabled = 0;
582 #else /* !CONFIG_CGROUP_PERF */
585 perf_cgroup_match(struct perf_event *event)
590 static inline void perf_detach_cgroup(struct perf_event *event)
593 static inline int is_cgroup_event(struct perf_event *event)
598 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
603 static inline void update_cgrp_time_from_event(struct perf_event *event)
607 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
611 static inline void perf_cgroup_sched_out(struct task_struct *task,
612 struct task_struct *next)
616 static inline void perf_cgroup_sched_in(struct task_struct *prev,
617 struct task_struct *task)
621 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
622 struct perf_event_attr *attr,
623 struct perf_event *group_leader)
629 perf_cgroup_set_timestamp(struct task_struct *task,
630 struct perf_event_context *ctx)
635 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
640 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
644 static inline u64 perf_cgroup_event_time(struct perf_event *event)
650 perf_cgroup_defer_enabled(struct perf_event *event)
655 perf_cgroup_mark_enabled(struct perf_event *event,
656 struct perf_event_context *ctx)
661 void perf_pmu_disable(struct pmu *pmu)
663 int *count = this_cpu_ptr(pmu->pmu_disable_count);
665 pmu->pmu_disable(pmu);
668 void perf_pmu_enable(struct pmu *pmu)
670 int *count = this_cpu_ptr(pmu->pmu_disable_count);
672 pmu->pmu_enable(pmu);
675 static DEFINE_PER_CPU(struct list_head, rotation_list);
678 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
679 * because they're strictly cpu affine and rotate_start is called with IRQs
680 * disabled, while rotate_context is called from IRQ context.
682 static void perf_pmu_rotate_start(struct pmu *pmu)
684 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
685 struct list_head *head = &__get_cpu_var(rotation_list);
687 WARN_ON(!irqs_disabled());
689 if (list_empty(&cpuctx->rotation_list)) {
690 int was_empty = list_empty(head);
691 list_add(&cpuctx->rotation_list, head);
693 tick_nohz_full_kick();
697 static void get_ctx(struct perf_event_context *ctx)
699 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
702 static void put_ctx(struct perf_event_context *ctx)
704 if (atomic_dec_and_test(&ctx->refcount)) {
706 put_ctx(ctx->parent_ctx);
708 put_task_struct(ctx->task);
709 kfree_rcu(ctx, rcu_head);
713 static void unclone_ctx(struct perf_event_context *ctx)
715 if (ctx->parent_ctx) {
716 put_ctx(ctx->parent_ctx);
717 ctx->parent_ctx = NULL;
721 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
724 * only top level events have the pid namespace they were created in
727 event = event->parent;
729 return task_tgid_nr_ns(p, event->ns);
732 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
735 * only top level events have the pid namespace they were created in
738 event = event->parent;
740 return task_pid_nr_ns(p, event->ns);
744 * If we inherit events we want to return the parent event id
747 static u64 primary_event_id(struct perf_event *event)
752 id = event->parent->id;
758 * Get the perf_event_context for a task and lock it.
759 * This has to cope with with the fact that until it is locked,
760 * the context could get moved to another task.
762 static struct perf_event_context *
763 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
765 struct perf_event_context *ctx;
769 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
772 * If this context is a clone of another, it might
773 * get swapped for another underneath us by
774 * perf_event_task_sched_out, though the
775 * rcu_read_lock() protects us from any context
776 * getting freed. Lock the context and check if it
777 * got swapped before we could get the lock, and retry
778 * if so. If we locked the right context, then it
779 * can't get swapped on us any more.
781 raw_spin_lock_irqsave(&ctx->lock, *flags);
782 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
783 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
787 if (!atomic_inc_not_zero(&ctx->refcount)) {
788 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
797 * Get the context for a task and increment its pin_count so it
798 * can't get swapped to another task. This also increments its
799 * reference count so that the context can't get freed.
801 static struct perf_event_context *
802 perf_pin_task_context(struct task_struct *task, int ctxn)
804 struct perf_event_context *ctx;
807 ctx = perf_lock_task_context(task, ctxn, &flags);
810 raw_spin_unlock_irqrestore(&ctx->lock, flags);
815 static void perf_unpin_context(struct perf_event_context *ctx)
819 raw_spin_lock_irqsave(&ctx->lock, flags);
821 raw_spin_unlock_irqrestore(&ctx->lock, flags);
825 * Update the record of the current time in a context.
827 static void update_context_time(struct perf_event_context *ctx)
829 u64 now = perf_clock();
831 ctx->time += now - ctx->timestamp;
832 ctx->timestamp = now;
835 static u64 perf_event_time(struct perf_event *event)
837 struct perf_event_context *ctx = event->ctx;
839 if (is_cgroup_event(event))
840 return perf_cgroup_event_time(event);
842 return ctx ? ctx->time : 0;
846 * Update the total_time_enabled and total_time_running fields for a event.
847 * The caller of this function needs to hold the ctx->lock.
849 static void update_event_times(struct perf_event *event)
851 struct perf_event_context *ctx = event->ctx;
854 if (event->state < PERF_EVENT_STATE_INACTIVE ||
855 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
858 * in cgroup mode, time_enabled represents
859 * the time the event was enabled AND active
860 * tasks were in the monitored cgroup. This is
861 * independent of the activity of the context as
862 * there may be a mix of cgroup and non-cgroup events.
864 * That is why we treat cgroup events differently
867 if (is_cgroup_event(event))
868 run_end = perf_cgroup_event_time(event);
869 else if (ctx->is_active)
872 run_end = event->tstamp_stopped;
874 event->total_time_enabled = run_end - event->tstamp_enabled;
876 if (event->state == PERF_EVENT_STATE_INACTIVE)
877 run_end = event->tstamp_stopped;
879 run_end = perf_event_time(event);
881 event->total_time_running = run_end - event->tstamp_running;
886 * Update total_time_enabled and total_time_running for all events in a group.
888 static void update_group_times(struct perf_event *leader)
890 struct perf_event *event;
892 update_event_times(leader);
893 list_for_each_entry(event, &leader->sibling_list, group_entry)
894 update_event_times(event);
897 static struct list_head *
898 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
900 if (event->attr.pinned)
901 return &ctx->pinned_groups;
903 return &ctx->flexible_groups;
907 * Add a event from the lists for its context.
908 * Must be called with ctx->mutex and ctx->lock held.
911 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
913 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
914 event->attach_state |= PERF_ATTACH_CONTEXT;
917 * If we're a stand alone event or group leader, we go to the context
918 * list, group events are kept attached to the group so that
919 * perf_group_detach can, at all times, locate all siblings.
921 if (event->group_leader == event) {
922 struct list_head *list;
924 if (is_software_event(event))
925 event->group_flags |= PERF_GROUP_SOFTWARE;
927 list = ctx_group_list(event, ctx);
928 list_add_tail(&event->group_entry, list);
931 if (is_cgroup_event(event))
934 if (has_branch_stack(event))
935 ctx->nr_branch_stack++;
937 list_add_rcu(&event->event_entry, &ctx->event_list);
939 perf_pmu_rotate_start(ctx->pmu);
941 if (event->attr.inherit_stat)
946 * Initialize event state based on the perf_event_attr::disabled.
948 static inline void perf_event__state_init(struct perf_event *event)
950 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
951 PERF_EVENT_STATE_INACTIVE;
955 * Called at perf_event creation and when events are attached/detached from a
958 static void perf_event__read_size(struct perf_event *event)
960 int entry = sizeof(u64); /* value */
964 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
967 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
970 if (event->attr.read_format & PERF_FORMAT_ID)
971 entry += sizeof(u64);
973 if (event->attr.read_format & PERF_FORMAT_GROUP) {
974 nr += event->group_leader->nr_siblings;
979 event->read_size = size;
982 static void perf_event__header_size(struct perf_event *event)
984 struct perf_sample_data *data;
985 u64 sample_type = event->attr.sample_type;
988 perf_event__read_size(event);
990 if (sample_type & PERF_SAMPLE_IP)
991 size += sizeof(data->ip);
993 if (sample_type & PERF_SAMPLE_ADDR)
994 size += sizeof(data->addr);
996 if (sample_type & PERF_SAMPLE_PERIOD)
997 size += sizeof(data->period);
999 if (sample_type & PERF_SAMPLE_WEIGHT)
1000 size += sizeof(data->weight);
1002 if (sample_type & PERF_SAMPLE_READ)
1003 size += event->read_size;
1005 if (sample_type & PERF_SAMPLE_DATA_SRC)
1006 size += sizeof(data->data_src.val);
1008 event->header_size = size;
1011 static void perf_event__id_header_size(struct perf_event *event)
1013 struct perf_sample_data *data;
1014 u64 sample_type = event->attr.sample_type;
1017 if (sample_type & PERF_SAMPLE_TID)
1018 size += sizeof(data->tid_entry);
1020 if (sample_type & PERF_SAMPLE_TIME)
1021 size += sizeof(data->time);
1023 if (sample_type & PERF_SAMPLE_ID)
1024 size += sizeof(data->id);
1026 if (sample_type & PERF_SAMPLE_STREAM_ID)
1027 size += sizeof(data->stream_id);
1029 if (sample_type & PERF_SAMPLE_CPU)
1030 size += sizeof(data->cpu_entry);
1032 event->id_header_size = size;
1035 static void perf_group_attach(struct perf_event *event)
1037 struct perf_event *group_leader = event->group_leader, *pos;
1040 * We can have double attach due to group movement in perf_event_open.
1042 if (event->attach_state & PERF_ATTACH_GROUP)
1045 event->attach_state |= PERF_ATTACH_GROUP;
1047 if (group_leader == event)
1050 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1051 !is_software_event(event))
1052 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1054 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1055 group_leader->nr_siblings++;
1057 perf_event__header_size(group_leader);
1059 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1060 perf_event__header_size(pos);
1064 * Remove a event from the lists for its context.
1065 * Must be called with ctx->mutex and ctx->lock held.
1068 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1070 struct perf_cpu_context *cpuctx;
1072 * We can have double detach due to exit/hot-unplug + close.
1074 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1077 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1079 if (is_cgroup_event(event)) {
1081 cpuctx = __get_cpu_context(ctx);
1083 * if there are no more cgroup events
1084 * then cler cgrp to avoid stale pointer
1085 * in update_cgrp_time_from_cpuctx()
1087 if (!ctx->nr_cgroups)
1088 cpuctx->cgrp = NULL;
1091 if (has_branch_stack(event))
1092 ctx->nr_branch_stack--;
1095 if (event->attr.inherit_stat)
1098 list_del_rcu(&event->event_entry);
1100 if (event->group_leader == event)
1101 list_del_init(&event->group_entry);
1103 update_group_times(event);
1106 * If event was in error state, then keep it
1107 * that way, otherwise bogus counts will be
1108 * returned on read(). The only way to get out
1109 * of error state is by explicit re-enabling
1112 if (event->state > PERF_EVENT_STATE_OFF)
1113 event->state = PERF_EVENT_STATE_OFF;
1116 static void perf_group_detach(struct perf_event *event)
1118 struct perf_event *sibling, *tmp;
1119 struct list_head *list = NULL;
1122 * We can have double detach due to exit/hot-unplug + close.
1124 if (!(event->attach_state & PERF_ATTACH_GROUP))
1127 event->attach_state &= ~PERF_ATTACH_GROUP;
1130 * If this is a sibling, remove it from its group.
1132 if (event->group_leader != event) {
1133 list_del_init(&event->group_entry);
1134 event->group_leader->nr_siblings--;
1138 if (!list_empty(&event->group_entry))
1139 list = &event->group_entry;
1142 * If this was a group event with sibling events then
1143 * upgrade the siblings to singleton events by adding them
1144 * to whatever list we are on.
1146 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1148 list_move_tail(&sibling->group_entry, list);
1149 sibling->group_leader = sibling;
1151 /* Inherit group flags from the previous leader */
1152 sibling->group_flags = event->group_flags;
1156 perf_event__header_size(event->group_leader);
1158 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1159 perf_event__header_size(tmp);
1163 event_filter_match(struct perf_event *event)
1165 return (event->cpu == -1 || event->cpu == smp_processor_id())
1166 && perf_cgroup_match(event);
1170 event_sched_out(struct perf_event *event,
1171 struct perf_cpu_context *cpuctx,
1172 struct perf_event_context *ctx)
1174 u64 tstamp = perf_event_time(event);
1177 * An event which could not be activated because of
1178 * filter mismatch still needs to have its timings
1179 * maintained, otherwise bogus information is return
1180 * via read() for time_enabled, time_running:
1182 if (event->state == PERF_EVENT_STATE_INACTIVE
1183 && !event_filter_match(event)) {
1184 delta = tstamp - event->tstamp_stopped;
1185 event->tstamp_running += delta;
1186 event->tstamp_stopped = tstamp;
1189 if (event->state != PERF_EVENT_STATE_ACTIVE)
1192 event->state = PERF_EVENT_STATE_INACTIVE;
1193 if (event->pending_disable) {
1194 event->pending_disable = 0;
1195 event->state = PERF_EVENT_STATE_OFF;
1197 event->tstamp_stopped = tstamp;
1198 event->pmu->del(event, 0);
1201 if (!is_software_event(event))
1202 cpuctx->active_oncpu--;
1204 if (event->attr.freq && event->attr.sample_freq)
1206 if (event->attr.exclusive || !cpuctx->active_oncpu)
1207 cpuctx->exclusive = 0;
1211 group_sched_out(struct perf_event *group_event,
1212 struct perf_cpu_context *cpuctx,
1213 struct perf_event_context *ctx)
1215 struct perf_event *event;
1216 int state = group_event->state;
1218 event_sched_out(group_event, cpuctx, ctx);
1221 * Schedule out siblings (if any):
1223 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1224 event_sched_out(event, cpuctx, ctx);
1226 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1227 cpuctx->exclusive = 0;
1231 * Cross CPU call to remove a performance event
1233 * We disable the event on the hardware level first. After that we
1234 * remove it from the context list.
1236 static int __perf_remove_from_context(void *info)
1238 struct perf_event *event = info;
1239 struct perf_event_context *ctx = event->ctx;
1240 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1242 raw_spin_lock(&ctx->lock);
1243 event_sched_out(event, cpuctx, ctx);
1244 list_del_event(event, ctx);
1245 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1247 cpuctx->task_ctx = NULL;
1249 raw_spin_unlock(&ctx->lock);
1256 * Remove the event from a task's (or a CPU's) list of events.
1258 * CPU events are removed with a smp call. For task events we only
1259 * call when the task is on a CPU.
1261 * If event->ctx is a cloned context, callers must make sure that
1262 * every task struct that event->ctx->task could possibly point to
1263 * remains valid. This is OK when called from perf_release since
1264 * that only calls us on the top-level context, which can't be a clone.
1265 * When called from perf_event_exit_task, it's OK because the
1266 * context has been detached from its task.
1268 static void perf_remove_from_context(struct perf_event *event)
1270 struct perf_event_context *ctx = event->ctx;
1271 struct task_struct *task = ctx->task;
1273 lockdep_assert_held(&ctx->mutex);
1277 * Per cpu events are removed via an smp call and
1278 * the removal is always successful.
1280 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1285 if (!task_function_call(task, __perf_remove_from_context, event))
1288 raw_spin_lock_irq(&ctx->lock);
1290 * If we failed to find a running task, but find the context active now
1291 * that we've acquired the ctx->lock, retry.
1293 if (ctx->is_active) {
1294 raw_spin_unlock_irq(&ctx->lock);
1299 * Since the task isn't running, its safe to remove the event, us
1300 * holding the ctx->lock ensures the task won't get scheduled in.
1302 list_del_event(event, ctx);
1303 raw_spin_unlock_irq(&ctx->lock);
1307 * Cross CPU call to disable a performance event
1309 int __perf_event_disable(void *info)
1311 struct perf_event *event = info;
1312 struct perf_event_context *ctx = event->ctx;
1313 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1316 * If this is a per-task event, need to check whether this
1317 * event's task is the current task on this cpu.
1319 * Can trigger due to concurrent perf_event_context_sched_out()
1320 * flipping contexts around.
1322 if (ctx->task && cpuctx->task_ctx != ctx)
1325 raw_spin_lock(&ctx->lock);
1328 * If the event is on, turn it off.
1329 * If it is in error state, leave it in error state.
1331 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1332 update_context_time(ctx);
1333 update_cgrp_time_from_event(event);
1334 update_group_times(event);
1335 if (event == event->group_leader)
1336 group_sched_out(event, cpuctx, ctx);
1338 event_sched_out(event, cpuctx, ctx);
1339 event->state = PERF_EVENT_STATE_OFF;
1342 raw_spin_unlock(&ctx->lock);
1350 * If event->ctx is a cloned context, callers must make sure that
1351 * every task struct that event->ctx->task could possibly point to
1352 * remains valid. This condition is satisifed when called through
1353 * perf_event_for_each_child or perf_event_for_each because they
1354 * hold the top-level event's child_mutex, so any descendant that
1355 * goes to exit will block in sync_child_event.
1356 * When called from perf_pending_event it's OK because event->ctx
1357 * is the current context on this CPU and preemption is disabled,
1358 * hence we can't get into perf_event_task_sched_out for this context.
1360 void perf_event_disable(struct perf_event *event)
1362 struct perf_event_context *ctx = event->ctx;
1363 struct task_struct *task = ctx->task;
1367 * Disable the event on the cpu that it's on
1369 cpu_function_call(event->cpu, __perf_event_disable, event);
1374 if (!task_function_call(task, __perf_event_disable, event))
1377 raw_spin_lock_irq(&ctx->lock);
1379 * If the event is still active, we need to retry the cross-call.
1381 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1382 raw_spin_unlock_irq(&ctx->lock);
1384 * Reload the task pointer, it might have been changed by
1385 * a concurrent perf_event_context_sched_out().
1392 * Since we have the lock this context can't be scheduled
1393 * in, so we can change the state safely.
1395 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1396 update_group_times(event);
1397 event->state = PERF_EVENT_STATE_OFF;
1399 raw_spin_unlock_irq(&ctx->lock);
1401 EXPORT_SYMBOL_GPL(perf_event_disable);
1403 static void perf_set_shadow_time(struct perf_event *event,
1404 struct perf_event_context *ctx,
1408 * use the correct time source for the time snapshot
1410 * We could get by without this by leveraging the
1411 * fact that to get to this function, the caller
1412 * has most likely already called update_context_time()
1413 * and update_cgrp_time_xx() and thus both timestamp
1414 * are identical (or very close). Given that tstamp is,
1415 * already adjusted for cgroup, we could say that:
1416 * tstamp - ctx->timestamp
1418 * tstamp - cgrp->timestamp.
1420 * Then, in perf_output_read(), the calculation would
1421 * work with no changes because:
1422 * - event is guaranteed scheduled in
1423 * - no scheduled out in between
1424 * - thus the timestamp would be the same
1426 * But this is a bit hairy.
1428 * So instead, we have an explicit cgroup call to remain
1429 * within the time time source all along. We believe it
1430 * is cleaner and simpler to understand.
1432 if (is_cgroup_event(event))
1433 perf_cgroup_set_shadow_time(event, tstamp);
1435 event->shadow_ctx_time = tstamp - ctx->timestamp;
1438 #define MAX_INTERRUPTS (~0ULL)
1440 static void perf_log_throttle(struct perf_event *event, int enable);
1443 event_sched_in(struct perf_event *event,
1444 struct perf_cpu_context *cpuctx,
1445 struct perf_event_context *ctx)
1447 u64 tstamp = perf_event_time(event);
1449 if (event->state <= PERF_EVENT_STATE_OFF)
1452 event->state = PERF_EVENT_STATE_ACTIVE;
1453 event->oncpu = smp_processor_id();
1456 * Unthrottle events, since we scheduled we might have missed several
1457 * ticks already, also for a heavily scheduling task there is little
1458 * guarantee it'll get a tick in a timely manner.
1460 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1461 perf_log_throttle(event, 1);
1462 event->hw.interrupts = 0;
1466 * The new state must be visible before we turn it on in the hardware:
1470 if (event->pmu->add(event, PERF_EF_START)) {
1471 event->state = PERF_EVENT_STATE_INACTIVE;
1476 event->tstamp_running += tstamp - event->tstamp_stopped;
1478 perf_set_shadow_time(event, ctx, tstamp);
1480 if (!is_software_event(event))
1481 cpuctx->active_oncpu++;
1483 if (event->attr.freq && event->attr.sample_freq)
1486 if (event->attr.exclusive)
1487 cpuctx->exclusive = 1;
1493 group_sched_in(struct perf_event *group_event,
1494 struct perf_cpu_context *cpuctx,
1495 struct perf_event_context *ctx)
1497 struct perf_event *event, *partial_group = NULL;
1498 struct pmu *pmu = group_event->pmu;
1499 u64 now = ctx->time;
1500 bool simulate = false;
1502 if (group_event->state == PERF_EVENT_STATE_OFF)
1505 pmu->start_txn(pmu);
1507 if (event_sched_in(group_event, cpuctx, ctx)) {
1508 pmu->cancel_txn(pmu);
1513 * Schedule in siblings as one group (if any):
1515 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1516 if (event_sched_in(event, cpuctx, ctx)) {
1517 partial_group = event;
1522 if (!pmu->commit_txn(pmu))
1527 * Groups can be scheduled in as one unit only, so undo any
1528 * partial group before returning:
1529 * The events up to the failed event are scheduled out normally,
1530 * tstamp_stopped will be updated.
1532 * The failed events and the remaining siblings need to have
1533 * their timings updated as if they had gone thru event_sched_in()
1534 * and event_sched_out(). This is required to get consistent timings
1535 * across the group. This also takes care of the case where the group
1536 * could never be scheduled by ensuring tstamp_stopped is set to mark
1537 * the time the event was actually stopped, such that time delta
1538 * calculation in update_event_times() is correct.
1540 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1541 if (event == partial_group)
1545 event->tstamp_running += now - event->tstamp_stopped;
1546 event->tstamp_stopped = now;
1548 event_sched_out(event, cpuctx, ctx);
1551 event_sched_out(group_event, cpuctx, ctx);
1553 pmu->cancel_txn(pmu);
1559 * Work out whether we can put this event group on the CPU now.
1561 static int group_can_go_on(struct perf_event *event,
1562 struct perf_cpu_context *cpuctx,
1566 * Groups consisting entirely of software events can always go on.
1568 if (event->group_flags & PERF_GROUP_SOFTWARE)
1571 * If an exclusive group is already on, no other hardware
1574 if (cpuctx->exclusive)
1577 * If this group is exclusive and there are already
1578 * events on the CPU, it can't go on.
1580 if (event->attr.exclusive && cpuctx->active_oncpu)
1583 * Otherwise, try to add it if all previous groups were able
1589 static void add_event_to_ctx(struct perf_event *event,
1590 struct perf_event_context *ctx)
1592 u64 tstamp = perf_event_time(event);
1594 list_add_event(event, ctx);
1595 perf_group_attach(event);
1596 event->tstamp_enabled = tstamp;
1597 event->tstamp_running = tstamp;
1598 event->tstamp_stopped = tstamp;
1601 static void task_ctx_sched_out(struct perf_event_context *ctx);
1603 ctx_sched_in(struct perf_event_context *ctx,
1604 struct perf_cpu_context *cpuctx,
1605 enum event_type_t event_type,
1606 struct task_struct *task);
1608 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1609 struct perf_event_context *ctx,
1610 struct task_struct *task)
1612 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1614 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1615 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1617 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1621 * Cross CPU call to install and enable a performance event
1623 * Must be called with ctx->mutex held
1625 static int __perf_install_in_context(void *info)
1627 struct perf_event *event = info;
1628 struct perf_event_context *ctx = event->ctx;
1629 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1630 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1631 struct task_struct *task = current;
1633 perf_ctx_lock(cpuctx, task_ctx);
1634 perf_pmu_disable(cpuctx->ctx.pmu);
1637 * If there was an active task_ctx schedule it out.
1640 task_ctx_sched_out(task_ctx);
1643 * If the context we're installing events in is not the
1644 * active task_ctx, flip them.
1646 if (ctx->task && task_ctx != ctx) {
1648 raw_spin_unlock(&task_ctx->lock);
1649 raw_spin_lock(&ctx->lock);
1654 cpuctx->task_ctx = task_ctx;
1655 task = task_ctx->task;
1658 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1660 update_context_time(ctx);
1662 * update cgrp time only if current cgrp
1663 * matches event->cgrp. Must be done before
1664 * calling add_event_to_ctx()
1666 update_cgrp_time_from_event(event);
1668 add_event_to_ctx(event, ctx);
1671 * Schedule everything back in
1673 perf_event_sched_in(cpuctx, task_ctx, task);
1675 perf_pmu_enable(cpuctx->ctx.pmu);
1676 perf_ctx_unlock(cpuctx, task_ctx);
1682 * Attach a performance event to a context
1684 * First we add the event to the list with the hardware enable bit
1685 * in event->hw_config cleared.
1687 * If the event is attached to a task which is on a CPU we use a smp
1688 * call to enable it in the task context. The task might have been
1689 * scheduled away, but we check this in the smp call again.
1692 perf_install_in_context(struct perf_event_context *ctx,
1693 struct perf_event *event,
1696 struct task_struct *task = ctx->task;
1698 lockdep_assert_held(&ctx->mutex);
1701 if (event->cpu != -1)
1706 * Per cpu events are installed via an smp call and
1707 * the install is always successful.
1709 cpu_function_call(cpu, __perf_install_in_context, event);
1714 if (!task_function_call(task, __perf_install_in_context, event))
1717 raw_spin_lock_irq(&ctx->lock);
1719 * If we failed to find a running task, but find the context active now
1720 * that we've acquired the ctx->lock, retry.
1722 if (ctx->is_active) {
1723 raw_spin_unlock_irq(&ctx->lock);
1728 * Since the task isn't running, its safe to add the event, us holding
1729 * the ctx->lock ensures the task won't get scheduled in.
1731 add_event_to_ctx(event, ctx);
1732 raw_spin_unlock_irq(&ctx->lock);
1736 * Put a event into inactive state and update time fields.
1737 * Enabling the leader of a group effectively enables all
1738 * the group members that aren't explicitly disabled, so we
1739 * have to update their ->tstamp_enabled also.
1740 * Note: this works for group members as well as group leaders
1741 * since the non-leader members' sibling_lists will be empty.
1743 static void __perf_event_mark_enabled(struct perf_event *event)
1745 struct perf_event *sub;
1746 u64 tstamp = perf_event_time(event);
1748 event->state = PERF_EVENT_STATE_INACTIVE;
1749 event->tstamp_enabled = tstamp - event->total_time_enabled;
1750 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1751 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1752 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1757 * Cross CPU call to enable a performance event
1759 static int __perf_event_enable(void *info)
1761 struct perf_event *event = info;
1762 struct perf_event_context *ctx = event->ctx;
1763 struct perf_event *leader = event->group_leader;
1764 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1767 if (WARN_ON_ONCE(!ctx->is_active))
1770 raw_spin_lock(&ctx->lock);
1771 update_context_time(ctx);
1773 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1777 * set current task's cgroup time reference point
1779 perf_cgroup_set_timestamp(current, ctx);
1781 __perf_event_mark_enabled(event);
1783 if (!event_filter_match(event)) {
1784 if (is_cgroup_event(event))
1785 perf_cgroup_defer_enabled(event);
1790 * If the event is in a group and isn't the group leader,
1791 * then don't put it on unless the group is on.
1793 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1796 if (!group_can_go_on(event, cpuctx, 1)) {
1799 if (event == leader)
1800 err = group_sched_in(event, cpuctx, ctx);
1802 err = event_sched_in(event, cpuctx, ctx);
1807 * If this event can't go on and it's part of a
1808 * group, then the whole group has to come off.
1810 if (leader != event)
1811 group_sched_out(leader, cpuctx, ctx);
1812 if (leader->attr.pinned) {
1813 update_group_times(leader);
1814 leader->state = PERF_EVENT_STATE_ERROR;
1819 raw_spin_unlock(&ctx->lock);
1827 * If event->ctx is a cloned context, callers must make sure that
1828 * every task struct that event->ctx->task could possibly point to
1829 * remains valid. This condition is satisfied when called through
1830 * perf_event_for_each_child or perf_event_for_each as described
1831 * for perf_event_disable.
1833 void perf_event_enable(struct perf_event *event)
1835 struct perf_event_context *ctx = event->ctx;
1836 struct task_struct *task = ctx->task;
1840 * Enable the event on the cpu that it's on
1842 cpu_function_call(event->cpu, __perf_event_enable, event);
1846 raw_spin_lock_irq(&ctx->lock);
1847 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1851 * If the event is in error state, clear that first.
1852 * That way, if we see the event in error state below, we
1853 * know that it has gone back into error state, as distinct
1854 * from the task having been scheduled away before the
1855 * cross-call arrived.
1857 if (event->state == PERF_EVENT_STATE_ERROR)
1858 event->state = PERF_EVENT_STATE_OFF;
1861 if (!ctx->is_active) {
1862 __perf_event_mark_enabled(event);
1866 raw_spin_unlock_irq(&ctx->lock);
1868 if (!task_function_call(task, __perf_event_enable, event))
1871 raw_spin_lock_irq(&ctx->lock);
1874 * If the context is active and the event is still off,
1875 * we need to retry the cross-call.
1877 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1879 * task could have been flipped by a concurrent
1880 * perf_event_context_sched_out()
1887 raw_spin_unlock_irq(&ctx->lock);
1889 EXPORT_SYMBOL_GPL(perf_event_enable);
1891 int perf_event_refresh(struct perf_event *event, int refresh)
1894 * not supported on inherited events
1896 if (event->attr.inherit || !is_sampling_event(event))
1899 atomic_add(refresh, &event->event_limit);
1900 perf_event_enable(event);
1904 EXPORT_SYMBOL_GPL(perf_event_refresh);
1906 static void ctx_sched_out(struct perf_event_context *ctx,
1907 struct perf_cpu_context *cpuctx,
1908 enum event_type_t event_type)
1910 struct perf_event *event;
1911 int is_active = ctx->is_active;
1913 ctx->is_active &= ~event_type;
1914 if (likely(!ctx->nr_events))
1917 update_context_time(ctx);
1918 update_cgrp_time_from_cpuctx(cpuctx);
1919 if (!ctx->nr_active)
1922 perf_pmu_disable(ctx->pmu);
1923 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1924 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1925 group_sched_out(event, cpuctx, ctx);
1928 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1929 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1930 group_sched_out(event, cpuctx, ctx);
1932 perf_pmu_enable(ctx->pmu);
1936 * Test whether two contexts are equivalent, i.e. whether they
1937 * have both been cloned from the same version of the same context
1938 * and they both have the same number of enabled events.
1939 * If the number of enabled events is the same, then the set
1940 * of enabled events should be the same, because these are both
1941 * inherited contexts, therefore we can't access individual events
1942 * in them directly with an fd; we can only enable/disable all
1943 * events via prctl, or enable/disable all events in a family
1944 * via ioctl, which will have the same effect on both contexts.
1946 static int context_equiv(struct perf_event_context *ctx1,
1947 struct perf_event_context *ctx2)
1949 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1950 && ctx1->parent_gen == ctx2->parent_gen
1951 && !ctx1->pin_count && !ctx2->pin_count;
1954 static void __perf_event_sync_stat(struct perf_event *event,
1955 struct perf_event *next_event)
1959 if (!event->attr.inherit_stat)
1963 * Update the event value, we cannot use perf_event_read()
1964 * because we're in the middle of a context switch and have IRQs
1965 * disabled, which upsets smp_call_function_single(), however
1966 * we know the event must be on the current CPU, therefore we
1967 * don't need to use it.
1969 switch (event->state) {
1970 case PERF_EVENT_STATE_ACTIVE:
1971 event->pmu->read(event);
1974 case PERF_EVENT_STATE_INACTIVE:
1975 update_event_times(event);
1983 * In order to keep per-task stats reliable we need to flip the event
1984 * values when we flip the contexts.
1986 value = local64_read(&next_event->count);
1987 value = local64_xchg(&event->count, value);
1988 local64_set(&next_event->count, value);
1990 swap(event->total_time_enabled, next_event->total_time_enabled);
1991 swap(event->total_time_running, next_event->total_time_running);
1994 * Since we swizzled the values, update the user visible data too.
1996 perf_event_update_userpage(event);
1997 perf_event_update_userpage(next_event);
2000 #define list_next_entry(pos, member) \
2001 list_entry(pos->member.next, typeof(*pos), member)
2003 static void perf_event_sync_stat(struct perf_event_context *ctx,
2004 struct perf_event_context *next_ctx)
2006 struct perf_event *event, *next_event;
2011 update_context_time(ctx);
2013 event = list_first_entry(&ctx->event_list,
2014 struct perf_event, event_entry);
2016 next_event = list_first_entry(&next_ctx->event_list,
2017 struct perf_event, event_entry);
2019 while (&event->event_entry != &ctx->event_list &&
2020 &next_event->event_entry != &next_ctx->event_list) {
2022 __perf_event_sync_stat(event, next_event);
2024 event = list_next_entry(event, event_entry);
2025 next_event = list_next_entry(next_event, event_entry);
2029 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2030 struct task_struct *next)
2032 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2033 struct perf_event_context *next_ctx;
2034 struct perf_event_context *parent;
2035 struct perf_cpu_context *cpuctx;
2041 cpuctx = __get_cpu_context(ctx);
2042 if (!cpuctx->task_ctx)
2046 parent = rcu_dereference(ctx->parent_ctx);
2047 next_ctx = next->perf_event_ctxp[ctxn];
2048 if (parent && next_ctx &&
2049 rcu_dereference(next_ctx->parent_ctx) == parent) {
2051 * Looks like the two contexts are clones, so we might be
2052 * able to optimize the context switch. We lock both
2053 * contexts and check that they are clones under the
2054 * lock (including re-checking that neither has been
2055 * uncloned in the meantime). It doesn't matter which
2056 * order we take the locks because no other cpu could
2057 * be trying to lock both of these tasks.
2059 raw_spin_lock(&ctx->lock);
2060 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2061 if (context_equiv(ctx, next_ctx)) {
2063 * XXX do we need a memory barrier of sorts
2064 * wrt to rcu_dereference() of perf_event_ctxp
2066 task->perf_event_ctxp[ctxn] = next_ctx;
2067 next->perf_event_ctxp[ctxn] = ctx;
2069 next_ctx->task = task;
2072 perf_event_sync_stat(ctx, next_ctx);
2074 raw_spin_unlock(&next_ctx->lock);
2075 raw_spin_unlock(&ctx->lock);
2080 raw_spin_lock(&ctx->lock);
2081 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2082 cpuctx->task_ctx = NULL;
2083 raw_spin_unlock(&ctx->lock);
2087 #define for_each_task_context_nr(ctxn) \
2088 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2091 * Called from scheduler to remove the events of the current task,
2092 * with interrupts disabled.
2094 * We stop each event and update the event value in event->count.
2096 * This does not protect us against NMI, but disable()
2097 * sets the disabled bit in the control field of event _before_
2098 * accessing the event control register. If a NMI hits, then it will
2099 * not restart the event.
2101 void __perf_event_task_sched_out(struct task_struct *task,
2102 struct task_struct *next)
2106 for_each_task_context_nr(ctxn)
2107 perf_event_context_sched_out(task, ctxn, next);
2110 * if cgroup events exist on this CPU, then we need
2111 * to check if we have to switch out PMU state.
2112 * cgroup event are system-wide mode only
2114 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2115 perf_cgroup_sched_out(task, next);
2118 static void task_ctx_sched_out(struct perf_event_context *ctx)
2120 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2122 if (!cpuctx->task_ctx)
2125 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2128 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2129 cpuctx->task_ctx = NULL;
2133 * Called with IRQs disabled
2135 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2136 enum event_type_t event_type)
2138 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2142 ctx_pinned_sched_in(struct perf_event_context *ctx,
2143 struct perf_cpu_context *cpuctx)
2145 struct perf_event *event;
2147 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2148 if (event->state <= PERF_EVENT_STATE_OFF)
2150 if (!event_filter_match(event))
2153 /* may need to reset tstamp_enabled */
2154 if (is_cgroup_event(event))
2155 perf_cgroup_mark_enabled(event, ctx);
2157 if (group_can_go_on(event, cpuctx, 1))
2158 group_sched_in(event, cpuctx, ctx);
2161 * If this pinned group hasn't been scheduled,
2162 * put it in error state.
2164 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2165 update_group_times(event);
2166 event->state = PERF_EVENT_STATE_ERROR;
2172 ctx_flexible_sched_in(struct perf_event_context *ctx,
2173 struct perf_cpu_context *cpuctx)
2175 struct perf_event *event;
2178 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2179 /* Ignore events in OFF or ERROR state */
2180 if (event->state <= PERF_EVENT_STATE_OFF)
2183 * Listen to the 'cpu' scheduling filter constraint
2186 if (!event_filter_match(event))
2189 /* may need to reset tstamp_enabled */
2190 if (is_cgroup_event(event))
2191 perf_cgroup_mark_enabled(event, ctx);
2193 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2194 if (group_sched_in(event, cpuctx, ctx))
2201 ctx_sched_in(struct perf_event_context *ctx,
2202 struct perf_cpu_context *cpuctx,
2203 enum event_type_t event_type,
2204 struct task_struct *task)
2207 int is_active = ctx->is_active;
2209 ctx->is_active |= event_type;
2210 if (likely(!ctx->nr_events))
2214 ctx->timestamp = now;
2215 perf_cgroup_set_timestamp(task, ctx);
2217 * First go through the list and put on any pinned groups
2218 * in order to give them the best chance of going on.
2220 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2221 ctx_pinned_sched_in(ctx, cpuctx);
2223 /* Then walk through the lower prio flexible groups */
2224 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2225 ctx_flexible_sched_in(ctx, cpuctx);
2228 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2229 enum event_type_t event_type,
2230 struct task_struct *task)
2232 struct perf_event_context *ctx = &cpuctx->ctx;
2234 ctx_sched_in(ctx, cpuctx, event_type, task);
2237 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2238 struct task_struct *task)
2240 struct perf_cpu_context *cpuctx;
2242 cpuctx = __get_cpu_context(ctx);
2243 if (cpuctx->task_ctx == ctx)
2246 perf_ctx_lock(cpuctx, ctx);
2247 perf_pmu_disable(ctx->pmu);
2249 * We want to keep the following priority order:
2250 * cpu pinned (that don't need to move), task pinned,
2251 * cpu flexible, task flexible.
2253 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2256 cpuctx->task_ctx = ctx;
2258 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2260 perf_pmu_enable(ctx->pmu);
2261 perf_ctx_unlock(cpuctx, ctx);
2264 * Since these rotations are per-cpu, we need to ensure the
2265 * cpu-context we got scheduled on is actually rotating.
2267 perf_pmu_rotate_start(ctx->pmu);
2271 * When sampling the branck stack in system-wide, it may be necessary
2272 * to flush the stack on context switch. This happens when the branch
2273 * stack does not tag its entries with the pid of the current task.
2274 * Otherwise it becomes impossible to associate a branch entry with a
2275 * task. This ambiguity is more likely to appear when the branch stack
2276 * supports priv level filtering and the user sets it to monitor only
2277 * at the user level (which could be a useful measurement in system-wide
2278 * mode). In that case, the risk is high of having a branch stack with
2279 * branch from multiple tasks. Flushing may mean dropping the existing
2280 * entries or stashing them somewhere in the PMU specific code layer.
2282 * This function provides the context switch callback to the lower code
2283 * layer. It is invoked ONLY when there is at least one system-wide context
2284 * with at least one active event using taken branch sampling.
2286 static void perf_branch_stack_sched_in(struct task_struct *prev,
2287 struct task_struct *task)
2289 struct perf_cpu_context *cpuctx;
2291 unsigned long flags;
2293 /* no need to flush branch stack if not changing task */
2297 local_irq_save(flags);
2301 list_for_each_entry_rcu(pmu, &pmus, entry) {
2302 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2305 * check if the context has at least one
2306 * event using PERF_SAMPLE_BRANCH_STACK
2308 if (cpuctx->ctx.nr_branch_stack > 0
2309 && pmu->flush_branch_stack) {
2311 pmu = cpuctx->ctx.pmu;
2313 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2315 perf_pmu_disable(pmu);
2317 pmu->flush_branch_stack();
2319 perf_pmu_enable(pmu);
2321 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2327 local_irq_restore(flags);
2331 * Called from scheduler to add the events of the current task
2332 * with interrupts disabled.
2334 * We restore the event value and then enable it.
2336 * This does not protect us against NMI, but enable()
2337 * sets the enabled bit in the control field of event _before_
2338 * accessing the event control register. If a NMI hits, then it will
2339 * keep the event running.
2341 void __perf_event_task_sched_in(struct task_struct *prev,
2342 struct task_struct *task)
2344 struct perf_event_context *ctx;
2347 for_each_task_context_nr(ctxn) {
2348 ctx = task->perf_event_ctxp[ctxn];
2352 perf_event_context_sched_in(ctx, task);
2355 * if cgroup events exist on this CPU, then we need
2356 * to check if we have to switch in PMU state.
2357 * cgroup event are system-wide mode only
2359 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2360 perf_cgroup_sched_in(prev, task);
2362 /* check for system-wide branch_stack events */
2363 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2364 perf_branch_stack_sched_in(prev, task);
2367 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2369 u64 frequency = event->attr.sample_freq;
2370 u64 sec = NSEC_PER_SEC;
2371 u64 divisor, dividend;
2373 int count_fls, nsec_fls, frequency_fls, sec_fls;
2375 count_fls = fls64(count);
2376 nsec_fls = fls64(nsec);
2377 frequency_fls = fls64(frequency);
2381 * We got @count in @nsec, with a target of sample_freq HZ
2382 * the target period becomes:
2385 * period = -------------------
2386 * @nsec * sample_freq
2391 * Reduce accuracy by one bit such that @a and @b converge
2392 * to a similar magnitude.
2394 #define REDUCE_FLS(a, b) \
2396 if (a##_fls > b##_fls) { \
2406 * Reduce accuracy until either term fits in a u64, then proceed with
2407 * the other, so that finally we can do a u64/u64 division.
2409 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2410 REDUCE_FLS(nsec, frequency);
2411 REDUCE_FLS(sec, count);
2414 if (count_fls + sec_fls > 64) {
2415 divisor = nsec * frequency;
2417 while (count_fls + sec_fls > 64) {
2418 REDUCE_FLS(count, sec);
2422 dividend = count * sec;
2424 dividend = count * sec;
2426 while (nsec_fls + frequency_fls > 64) {
2427 REDUCE_FLS(nsec, frequency);
2431 divisor = nsec * frequency;
2437 return div64_u64(dividend, divisor);
2440 static DEFINE_PER_CPU(int, perf_throttled_count);
2441 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2443 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2445 struct hw_perf_event *hwc = &event->hw;
2446 s64 period, sample_period;
2449 period = perf_calculate_period(event, nsec, count);
2451 delta = (s64)(period - hwc->sample_period);
2452 delta = (delta + 7) / 8; /* low pass filter */
2454 sample_period = hwc->sample_period + delta;
2459 hwc->sample_period = sample_period;
2461 if (local64_read(&hwc->period_left) > 8*sample_period) {
2463 event->pmu->stop(event, PERF_EF_UPDATE);
2465 local64_set(&hwc->period_left, 0);
2468 event->pmu->start(event, PERF_EF_RELOAD);
2473 * combine freq adjustment with unthrottling to avoid two passes over the
2474 * events. At the same time, make sure, having freq events does not change
2475 * the rate of unthrottling as that would introduce bias.
2477 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2480 struct perf_event *event;
2481 struct hw_perf_event *hwc;
2482 u64 now, period = TICK_NSEC;
2486 * only need to iterate over all events iff:
2487 * - context have events in frequency mode (needs freq adjust)
2488 * - there are events to unthrottle on this cpu
2490 if (!(ctx->nr_freq || needs_unthr))
2493 raw_spin_lock(&ctx->lock);
2494 perf_pmu_disable(ctx->pmu);
2496 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2497 if (event->state != PERF_EVENT_STATE_ACTIVE)
2500 if (!event_filter_match(event))
2505 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2506 hwc->interrupts = 0;
2507 perf_log_throttle(event, 1);
2508 event->pmu->start(event, 0);
2511 if (!event->attr.freq || !event->attr.sample_freq)
2515 * stop the event and update event->count
2517 event->pmu->stop(event, PERF_EF_UPDATE);
2519 now = local64_read(&event->count);
2520 delta = now - hwc->freq_count_stamp;
2521 hwc->freq_count_stamp = now;
2525 * reload only if value has changed
2526 * we have stopped the event so tell that
2527 * to perf_adjust_period() to avoid stopping it
2531 perf_adjust_period(event, period, delta, false);
2533 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2536 perf_pmu_enable(ctx->pmu);
2537 raw_spin_unlock(&ctx->lock);
2541 * Round-robin a context's events:
2543 static void rotate_ctx(struct perf_event_context *ctx)
2546 * Rotate the first entry last of non-pinned groups. Rotation might be
2547 * disabled by the inheritance code.
2549 if (!ctx->rotate_disable)
2550 list_rotate_left(&ctx->flexible_groups);
2554 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2555 * because they're strictly cpu affine and rotate_start is called with IRQs
2556 * disabled, while rotate_context is called from IRQ context.
2558 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2560 struct perf_event_context *ctx = NULL;
2561 int rotate = 0, remove = 1;
2563 if (cpuctx->ctx.nr_events) {
2565 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2569 ctx = cpuctx->task_ctx;
2570 if (ctx && ctx->nr_events) {
2572 if (ctx->nr_events != ctx->nr_active)
2579 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2580 perf_pmu_disable(cpuctx->ctx.pmu);
2582 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2584 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2586 rotate_ctx(&cpuctx->ctx);
2590 perf_event_sched_in(cpuctx, ctx, current);
2592 perf_pmu_enable(cpuctx->ctx.pmu);
2593 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2596 list_del_init(&cpuctx->rotation_list);
2599 #ifdef CONFIG_NO_HZ_FULL
2600 bool perf_event_can_stop_tick(void)
2602 if (list_empty(&__get_cpu_var(rotation_list)))
2609 void perf_event_task_tick(void)
2611 struct list_head *head = &__get_cpu_var(rotation_list);
2612 struct perf_cpu_context *cpuctx, *tmp;
2613 struct perf_event_context *ctx;
2616 WARN_ON(!irqs_disabled());
2618 __this_cpu_inc(perf_throttled_seq);
2619 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2621 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2623 perf_adjust_freq_unthr_context(ctx, throttled);
2625 ctx = cpuctx->task_ctx;
2627 perf_adjust_freq_unthr_context(ctx, throttled);
2629 if (cpuctx->jiffies_interval == 1 ||
2630 !(jiffies % cpuctx->jiffies_interval))
2631 perf_rotate_context(cpuctx);
2635 static int event_enable_on_exec(struct perf_event *event,
2636 struct perf_event_context *ctx)
2638 if (!event->attr.enable_on_exec)
2641 event->attr.enable_on_exec = 0;
2642 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2645 __perf_event_mark_enabled(event);
2651 * Enable all of a task's events that have been marked enable-on-exec.
2652 * This expects task == current.
2654 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2656 struct perf_event *event;
2657 unsigned long flags;
2661 local_irq_save(flags);
2662 if (!ctx || !ctx->nr_events)
2666 * We must ctxsw out cgroup events to avoid conflict
2667 * when invoking perf_task_event_sched_in() later on
2668 * in this function. Otherwise we end up trying to
2669 * ctxswin cgroup events which are already scheduled
2672 perf_cgroup_sched_out(current, NULL);
2674 raw_spin_lock(&ctx->lock);
2675 task_ctx_sched_out(ctx);
2677 list_for_each_entry(event, &ctx->event_list, event_entry) {
2678 ret = event_enable_on_exec(event, ctx);
2684 * Unclone this context if we enabled any event.
2689 raw_spin_unlock(&ctx->lock);
2692 * Also calls ctxswin for cgroup events, if any:
2694 perf_event_context_sched_in(ctx, ctx->task);
2696 local_irq_restore(flags);
2700 * Cross CPU call to read the hardware event
2702 static void __perf_event_read(void *info)
2704 struct perf_event *event = info;
2705 struct perf_event_context *ctx = event->ctx;
2706 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2709 * If this is a task context, we need to check whether it is
2710 * the current task context of this cpu. If not it has been
2711 * scheduled out before the smp call arrived. In that case
2712 * event->count would have been updated to a recent sample
2713 * when the event was scheduled out.
2715 if (ctx->task && cpuctx->task_ctx != ctx)
2718 raw_spin_lock(&ctx->lock);
2719 if (ctx->is_active) {
2720 update_context_time(ctx);
2721 update_cgrp_time_from_event(event);
2723 update_event_times(event);
2724 if (event->state == PERF_EVENT_STATE_ACTIVE)
2725 event->pmu->read(event);
2726 raw_spin_unlock(&ctx->lock);
2729 static inline u64 perf_event_count(struct perf_event *event)
2731 return local64_read(&event->count) + atomic64_read(&event->child_count);
2734 static u64 perf_event_read(struct perf_event *event)
2737 * If event is enabled and currently active on a CPU, update the
2738 * value in the event structure:
2740 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2741 smp_call_function_single(event->oncpu,
2742 __perf_event_read, event, 1);
2743 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2744 struct perf_event_context *ctx = event->ctx;
2745 unsigned long flags;
2747 raw_spin_lock_irqsave(&ctx->lock, flags);
2749 * may read while context is not active
2750 * (e.g., thread is blocked), in that case
2751 * we cannot update context time
2753 if (ctx->is_active) {
2754 update_context_time(ctx);
2755 update_cgrp_time_from_event(event);
2757 update_event_times(event);
2758 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2761 return perf_event_count(event);
2765 * Initialize the perf_event context in a task_struct:
2767 static void __perf_event_init_context(struct perf_event_context *ctx)
2769 raw_spin_lock_init(&ctx->lock);
2770 mutex_init(&ctx->mutex);
2771 INIT_LIST_HEAD(&ctx->pinned_groups);
2772 INIT_LIST_HEAD(&ctx->flexible_groups);
2773 INIT_LIST_HEAD(&ctx->event_list);
2774 atomic_set(&ctx->refcount, 1);
2777 static struct perf_event_context *
2778 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2780 struct perf_event_context *ctx;
2782 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2786 __perf_event_init_context(ctx);
2789 get_task_struct(task);
2796 static struct task_struct *
2797 find_lively_task_by_vpid(pid_t vpid)
2799 struct task_struct *task;
2806 task = find_task_by_vpid(vpid);
2808 get_task_struct(task);
2812 return ERR_PTR(-ESRCH);
2814 /* Reuse ptrace permission checks for now. */
2816 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2821 put_task_struct(task);
2822 return ERR_PTR(err);
2827 * Returns a matching context with refcount and pincount.
2829 static struct perf_event_context *
2830 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2832 struct perf_event_context *ctx;
2833 struct perf_cpu_context *cpuctx;
2834 unsigned long flags;
2838 /* Must be root to operate on a CPU event: */
2839 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2840 return ERR_PTR(-EACCES);
2843 * We could be clever and allow to attach a event to an
2844 * offline CPU and activate it when the CPU comes up, but
2847 if (!cpu_online(cpu))
2848 return ERR_PTR(-ENODEV);
2850 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2859 ctxn = pmu->task_ctx_nr;
2864 ctx = perf_lock_task_context(task, ctxn, &flags);
2868 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2870 ctx = alloc_perf_context(pmu, task);
2876 mutex_lock(&task->perf_event_mutex);
2878 * If it has already passed perf_event_exit_task().
2879 * we must see PF_EXITING, it takes this mutex too.
2881 if (task->flags & PF_EXITING)
2883 else if (task->perf_event_ctxp[ctxn])
2888 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2890 mutex_unlock(&task->perf_event_mutex);
2892 if (unlikely(err)) {
2904 return ERR_PTR(err);
2907 static void perf_event_free_filter(struct perf_event *event);
2909 static void free_event_rcu(struct rcu_head *head)
2911 struct perf_event *event;
2913 event = container_of(head, struct perf_event, rcu_head);
2915 put_pid_ns(event->ns);
2916 perf_event_free_filter(event);
2920 static void ring_buffer_put(struct ring_buffer *rb);
2922 static void free_event(struct perf_event *event)
2924 irq_work_sync(&event->pending);
2926 if (!event->parent) {
2927 if (event->attach_state & PERF_ATTACH_TASK)
2928 static_key_slow_dec_deferred(&perf_sched_events);
2929 if (event->attr.mmap || event->attr.mmap_data)
2930 atomic_dec(&nr_mmap_events);
2931 if (event->attr.comm)
2932 atomic_dec(&nr_comm_events);
2933 if (event->attr.task)
2934 atomic_dec(&nr_task_events);
2935 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2936 put_callchain_buffers();
2937 if (is_cgroup_event(event)) {
2938 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2939 static_key_slow_dec_deferred(&perf_sched_events);
2942 if (has_branch_stack(event)) {
2943 static_key_slow_dec_deferred(&perf_sched_events);
2944 /* is system-wide event */
2945 if (!(event->attach_state & PERF_ATTACH_TASK))
2946 atomic_dec(&per_cpu(perf_branch_stack_events,
2952 ring_buffer_put(event->rb);
2956 if (is_cgroup_event(event))
2957 perf_detach_cgroup(event);
2960 event->destroy(event);
2963 put_ctx(event->ctx);
2965 call_rcu(&event->rcu_head, free_event_rcu);
2968 int perf_event_release_kernel(struct perf_event *event)
2970 struct perf_event_context *ctx = event->ctx;
2972 WARN_ON_ONCE(ctx->parent_ctx);
2974 * There are two ways this annotation is useful:
2976 * 1) there is a lock recursion from perf_event_exit_task
2977 * see the comment there.
2979 * 2) there is a lock-inversion with mmap_sem through
2980 * perf_event_read_group(), which takes faults while
2981 * holding ctx->mutex, however this is called after
2982 * the last filedesc died, so there is no possibility
2983 * to trigger the AB-BA case.
2985 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2986 raw_spin_lock_irq(&ctx->lock);
2987 perf_group_detach(event);
2988 raw_spin_unlock_irq(&ctx->lock);
2989 perf_remove_from_context(event);
2990 mutex_unlock(&ctx->mutex);
2996 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2999 * Called when the last reference to the file is gone.
3001 static void put_event(struct perf_event *event)
3003 struct task_struct *owner;
3005 if (!atomic_long_dec_and_test(&event->refcount))
3009 owner = ACCESS_ONCE(event->owner);
3011 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3012 * !owner it means the list deletion is complete and we can indeed
3013 * free this event, otherwise we need to serialize on
3014 * owner->perf_event_mutex.
3016 smp_read_barrier_depends();
3019 * Since delayed_put_task_struct() also drops the last
3020 * task reference we can safely take a new reference
3021 * while holding the rcu_read_lock().
3023 get_task_struct(owner);
3028 mutex_lock(&owner->perf_event_mutex);
3030 * We have to re-check the event->owner field, if it is cleared
3031 * we raced with perf_event_exit_task(), acquiring the mutex
3032 * ensured they're done, and we can proceed with freeing the
3036 list_del_init(&event->owner_entry);
3037 mutex_unlock(&owner->perf_event_mutex);
3038 put_task_struct(owner);
3041 perf_event_release_kernel(event);
3044 static int perf_release(struct inode *inode, struct file *file)
3046 put_event(file->private_data);
3050 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3052 struct perf_event *child;
3058 mutex_lock(&event->child_mutex);
3059 total += perf_event_read(event);
3060 *enabled += event->total_time_enabled +
3061 atomic64_read(&event->child_total_time_enabled);
3062 *running += event->total_time_running +
3063 atomic64_read(&event->child_total_time_running);
3065 list_for_each_entry(child, &event->child_list, child_list) {
3066 total += perf_event_read(child);
3067 *enabled += child->total_time_enabled;
3068 *running += child->total_time_running;
3070 mutex_unlock(&event->child_mutex);
3074 EXPORT_SYMBOL_GPL(perf_event_read_value);
3076 static int perf_event_read_group(struct perf_event *event,
3077 u64 read_format, char __user *buf)
3079 struct perf_event *leader = event->group_leader, *sub;
3080 int n = 0, size = 0, ret = -EFAULT;
3081 struct perf_event_context *ctx = leader->ctx;
3083 u64 count, enabled, running;
3085 mutex_lock(&ctx->mutex);
3086 count = perf_event_read_value(leader, &enabled, &running);
3088 values[n++] = 1 + leader->nr_siblings;
3089 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3090 values[n++] = enabled;
3091 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3092 values[n++] = running;
3093 values[n++] = count;
3094 if (read_format & PERF_FORMAT_ID)
3095 values[n++] = primary_event_id(leader);
3097 size = n * sizeof(u64);
3099 if (copy_to_user(buf, values, size))
3104 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3107 values[n++] = perf_event_read_value(sub, &enabled, &running);
3108 if (read_format & PERF_FORMAT_ID)
3109 values[n++] = primary_event_id(sub);
3111 size = n * sizeof(u64);
3113 if (copy_to_user(buf + ret, values, size)) {
3121 mutex_unlock(&ctx->mutex);
3126 static int perf_event_read_one(struct perf_event *event,
3127 u64 read_format, char __user *buf)
3129 u64 enabled, running;
3133 values[n++] = perf_event_read_value(event, &enabled, &running);
3134 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3135 values[n++] = enabled;
3136 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3137 values[n++] = running;
3138 if (read_format & PERF_FORMAT_ID)
3139 values[n++] = primary_event_id(event);
3141 if (copy_to_user(buf, values, n * sizeof(u64)))
3144 return n * sizeof(u64);
3148 * Read the performance event - simple non blocking version for now
3151 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3153 u64 read_format = event->attr.read_format;
3157 * Return end-of-file for a read on a event that is in
3158 * error state (i.e. because it was pinned but it couldn't be
3159 * scheduled on to the CPU at some point).
3161 if (event->state == PERF_EVENT_STATE_ERROR)
3164 if (count < event->read_size)
3167 WARN_ON_ONCE(event->ctx->parent_ctx);
3168 if (read_format & PERF_FORMAT_GROUP)
3169 ret = perf_event_read_group(event, read_format, buf);
3171 ret = perf_event_read_one(event, read_format, buf);
3177 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3179 struct perf_event *event = file->private_data;
3181 return perf_read_hw(event, buf, count);
3184 static unsigned int perf_poll(struct file *file, poll_table *wait)
3186 struct perf_event *event = file->private_data;
3187 struct ring_buffer *rb;
3188 unsigned int events = POLL_HUP;
3191 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3192 * grabs the rb reference but perf_event_set_output() overrides it.
3193 * Here is the timeline for two threads T1, T2:
3194 * t0: T1, rb = rcu_dereference(event->rb)
3195 * t1: T2, old_rb = event->rb
3196 * t2: T2, event->rb = new rb
3197 * t3: T2, ring_buffer_detach(old_rb)
3198 * t4: T1, ring_buffer_attach(rb1)
3199 * t5: T1, poll_wait(event->waitq)
3201 * To avoid this problem, we grab mmap_mutex in perf_poll()
3202 * thereby ensuring that the assignment of the new ring buffer
3203 * and the detachment of the old buffer appear atomic to perf_poll()
3205 mutex_lock(&event->mmap_mutex);
3208 rb = rcu_dereference(event->rb);
3210 ring_buffer_attach(event, rb);
3211 events = atomic_xchg(&rb->poll, 0);
3215 mutex_unlock(&event->mmap_mutex);
3217 poll_wait(file, &event->waitq, wait);
3222 static void perf_event_reset(struct perf_event *event)
3224 (void)perf_event_read(event);
3225 local64_set(&event->count, 0);
3226 perf_event_update_userpage(event);
3230 * Holding the top-level event's child_mutex means that any
3231 * descendant process that has inherited this event will block
3232 * in sync_child_event if it goes to exit, thus satisfying the
3233 * task existence requirements of perf_event_enable/disable.
3235 static void perf_event_for_each_child(struct perf_event *event,
3236 void (*func)(struct perf_event *))
3238 struct perf_event *child;
3240 WARN_ON_ONCE(event->ctx->parent_ctx);
3241 mutex_lock(&event->child_mutex);
3243 list_for_each_entry(child, &event->child_list, child_list)
3245 mutex_unlock(&event->child_mutex);
3248 static void perf_event_for_each(struct perf_event *event,
3249 void (*func)(struct perf_event *))
3251 struct perf_event_context *ctx = event->ctx;
3252 struct perf_event *sibling;
3254 WARN_ON_ONCE(ctx->parent_ctx);
3255 mutex_lock(&ctx->mutex);
3256 event = event->group_leader;
3258 perf_event_for_each_child(event, func);
3259 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3260 perf_event_for_each_child(sibling, func);
3261 mutex_unlock(&ctx->mutex);
3264 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3266 struct perf_event_context *ctx = event->ctx;
3270 if (!is_sampling_event(event))
3273 if (copy_from_user(&value, arg, sizeof(value)))
3279 raw_spin_lock_irq(&ctx->lock);
3280 if (event->attr.freq) {
3281 if (value > sysctl_perf_event_sample_rate) {
3286 event->attr.sample_freq = value;
3288 event->attr.sample_period = value;
3289 event->hw.sample_period = value;
3292 raw_spin_unlock_irq(&ctx->lock);
3297 static const struct file_operations perf_fops;
3299 static inline int perf_fget_light(int fd, struct fd *p)
3301 struct fd f = fdget(fd);
3305 if (f.file->f_op != &perf_fops) {
3313 static int perf_event_set_output(struct perf_event *event,
3314 struct perf_event *output_event);
3315 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3317 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3319 struct perf_event *event = file->private_data;
3320 void (*func)(struct perf_event *);
3324 case PERF_EVENT_IOC_ENABLE:
3325 func = perf_event_enable;
3327 case PERF_EVENT_IOC_DISABLE:
3328 func = perf_event_disable;
3330 case PERF_EVENT_IOC_RESET:
3331 func = perf_event_reset;
3334 case PERF_EVENT_IOC_REFRESH:
3335 return perf_event_refresh(event, arg);
3337 case PERF_EVENT_IOC_PERIOD:
3338 return perf_event_period(event, (u64 __user *)arg);
3340 case PERF_EVENT_IOC_SET_OUTPUT:
3344 struct perf_event *output_event;
3346 ret = perf_fget_light(arg, &output);
3349 output_event = output.file->private_data;
3350 ret = perf_event_set_output(event, output_event);
3353 ret = perf_event_set_output(event, NULL);
3358 case PERF_EVENT_IOC_SET_FILTER:
3359 return perf_event_set_filter(event, (void __user *)arg);
3365 if (flags & PERF_IOC_FLAG_GROUP)
3366 perf_event_for_each(event, func);
3368 perf_event_for_each_child(event, func);
3373 int perf_event_task_enable(void)
3375 struct perf_event *event;
3377 mutex_lock(¤t->perf_event_mutex);
3378 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3379 perf_event_for_each_child(event, perf_event_enable);
3380 mutex_unlock(¤t->perf_event_mutex);
3385 int perf_event_task_disable(void)
3387 struct perf_event *event;
3389 mutex_lock(¤t->perf_event_mutex);
3390 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3391 perf_event_for_each_child(event, perf_event_disable);
3392 mutex_unlock(¤t->perf_event_mutex);
3397 static int perf_event_index(struct perf_event *event)
3399 if (event->hw.state & PERF_HES_STOPPED)
3402 if (event->state != PERF_EVENT_STATE_ACTIVE)
3405 return event->pmu->event_idx(event);
3408 static void calc_timer_values(struct perf_event *event,
3415 *now = perf_clock();
3416 ctx_time = event->shadow_ctx_time + *now;
3417 *enabled = ctx_time - event->tstamp_enabled;
3418 *running = ctx_time - event->tstamp_running;
3421 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3426 * Callers need to ensure there can be no nesting of this function, otherwise
3427 * the seqlock logic goes bad. We can not serialize this because the arch
3428 * code calls this from NMI context.
3430 void perf_event_update_userpage(struct perf_event *event)
3432 struct perf_event_mmap_page *userpg;
3433 struct ring_buffer *rb;
3434 u64 enabled, running, now;
3438 * compute total_time_enabled, total_time_running
3439 * based on snapshot values taken when the event
3440 * was last scheduled in.
3442 * we cannot simply called update_context_time()
3443 * because of locking issue as we can be called in
3446 calc_timer_values(event, &now, &enabled, &running);
3447 rb = rcu_dereference(event->rb);
3451 userpg = rb->user_page;
3454 * Disable preemption so as to not let the corresponding user-space
3455 * spin too long if we get preempted.
3460 userpg->index = perf_event_index(event);
3461 userpg->offset = perf_event_count(event);
3463 userpg->offset -= local64_read(&event->hw.prev_count);
3465 userpg->time_enabled = enabled +
3466 atomic64_read(&event->child_total_time_enabled);
3468 userpg->time_running = running +
3469 atomic64_read(&event->child_total_time_running);
3471 arch_perf_update_userpage(userpg, now);
3480 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3482 struct perf_event *event = vma->vm_file->private_data;
3483 struct ring_buffer *rb;
3484 int ret = VM_FAULT_SIGBUS;
3486 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3487 if (vmf->pgoff == 0)
3493 rb = rcu_dereference(event->rb);
3497 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3500 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3504 get_page(vmf->page);
3505 vmf->page->mapping = vma->vm_file->f_mapping;
3506 vmf->page->index = vmf->pgoff;
3515 static void ring_buffer_attach(struct perf_event *event,
3516 struct ring_buffer *rb)
3518 unsigned long flags;
3520 if (!list_empty(&event->rb_entry))
3523 spin_lock_irqsave(&rb->event_lock, flags);
3524 if (!list_empty(&event->rb_entry))
3527 list_add(&event->rb_entry, &rb->event_list);
3529 spin_unlock_irqrestore(&rb->event_lock, flags);
3532 static void ring_buffer_detach(struct perf_event *event,
3533 struct ring_buffer *rb)
3535 unsigned long flags;
3537 if (list_empty(&event->rb_entry))
3540 spin_lock_irqsave(&rb->event_lock, flags);
3541 list_del_init(&event->rb_entry);
3542 wake_up_all(&event->waitq);
3543 spin_unlock_irqrestore(&rb->event_lock, flags);
3546 static void ring_buffer_wakeup(struct perf_event *event)
3548 struct ring_buffer *rb;
3551 rb = rcu_dereference(event->rb);
3555 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3556 wake_up_all(&event->waitq);
3562 static void rb_free_rcu(struct rcu_head *rcu_head)
3564 struct ring_buffer *rb;
3566 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3570 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3572 struct ring_buffer *rb;
3575 rb = rcu_dereference(event->rb);
3577 if (!atomic_inc_not_zero(&rb->refcount))
3585 static void ring_buffer_put(struct ring_buffer *rb)
3587 struct perf_event *event, *n;
3588 unsigned long flags;
3590 if (!atomic_dec_and_test(&rb->refcount))
3593 spin_lock_irqsave(&rb->event_lock, flags);
3594 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3595 list_del_init(&event->rb_entry);
3596 wake_up_all(&event->waitq);
3598 spin_unlock_irqrestore(&rb->event_lock, flags);
3600 call_rcu(&rb->rcu_head, rb_free_rcu);
3603 static void perf_mmap_open(struct vm_area_struct *vma)
3605 struct perf_event *event = vma->vm_file->private_data;
3607 atomic_inc(&event->mmap_count);
3610 static void perf_mmap_close(struct vm_area_struct *vma)
3612 struct perf_event *event = vma->vm_file->private_data;
3614 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3615 unsigned long size = perf_data_size(event->rb);
3616 struct user_struct *user = event->mmap_user;
3617 struct ring_buffer *rb = event->rb;
3619 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3620 vma->vm_mm->pinned_vm -= event->mmap_locked;
3621 rcu_assign_pointer(event->rb, NULL);
3622 ring_buffer_detach(event, rb);
3623 mutex_unlock(&event->mmap_mutex);
3625 ring_buffer_put(rb);
3630 static const struct vm_operations_struct perf_mmap_vmops = {
3631 .open = perf_mmap_open,
3632 .close = perf_mmap_close,
3633 .fault = perf_mmap_fault,
3634 .page_mkwrite = perf_mmap_fault,
3637 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3639 struct perf_event *event = file->private_data;
3640 unsigned long user_locked, user_lock_limit;
3641 struct user_struct *user = current_user();
3642 unsigned long locked, lock_limit;
3643 struct ring_buffer *rb;
3644 unsigned long vma_size;
3645 unsigned long nr_pages;
3646 long user_extra, extra;
3647 int ret = 0, flags = 0;
3650 * Don't allow mmap() of inherited per-task counters. This would
3651 * create a performance issue due to all children writing to the
3654 if (event->cpu == -1 && event->attr.inherit)
3657 if (!(vma->vm_flags & VM_SHARED))
3660 vma_size = vma->vm_end - vma->vm_start;
3661 nr_pages = (vma_size / PAGE_SIZE) - 1;
3664 * If we have rb pages ensure they're a power-of-two number, so we
3665 * can do bitmasks instead of modulo.
3667 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3670 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3673 if (vma->vm_pgoff != 0)
3676 WARN_ON_ONCE(event->ctx->parent_ctx);
3677 mutex_lock(&event->mmap_mutex);
3679 if (event->rb->nr_pages == nr_pages)
3680 atomic_inc(&event->rb->refcount);
3686 user_extra = nr_pages + 1;
3687 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3690 * Increase the limit linearly with more CPUs:
3692 user_lock_limit *= num_online_cpus();
3694 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3697 if (user_locked > user_lock_limit)
3698 extra = user_locked - user_lock_limit;
3700 lock_limit = rlimit(RLIMIT_MEMLOCK);
3701 lock_limit >>= PAGE_SHIFT;
3702 locked = vma->vm_mm->pinned_vm + extra;
3704 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3705 !capable(CAP_IPC_LOCK)) {
3712 if (vma->vm_flags & VM_WRITE)
3713 flags |= RING_BUFFER_WRITABLE;
3715 rb = rb_alloc(nr_pages,
3716 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3723 rcu_assign_pointer(event->rb, rb);
3725 atomic_long_add(user_extra, &user->locked_vm);
3726 event->mmap_locked = extra;
3727 event->mmap_user = get_current_user();
3728 vma->vm_mm->pinned_vm += event->mmap_locked;
3730 perf_event_update_userpage(event);
3734 atomic_inc(&event->mmap_count);
3735 mutex_unlock(&event->mmap_mutex);
3737 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3738 vma->vm_ops = &perf_mmap_vmops;
3743 static int perf_fasync(int fd, struct file *filp, int on)
3745 struct inode *inode = file_inode(filp);
3746 struct perf_event *event = filp->private_data;
3749 mutex_lock(&inode->i_mutex);
3750 retval = fasync_helper(fd, filp, on, &event->fasync);
3751 mutex_unlock(&inode->i_mutex);
3759 static const struct file_operations perf_fops = {
3760 .llseek = no_llseek,
3761 .release = perf_release,
3764 .unlocked_ioctl = perf_ioctl,
3765 .compat_ioctl = perf_ioctl,
3767 .fasync = perf_fasync,
3773 * If there's data, ensure we set the poll() state and publish everything
3774 * to user-space before waking everybody up.
3777 void perf_event_wakeup(struct perf_event *event)
3779 ring_buffer_wakeup(event);
3781 if (event->pending_kill) {
3782 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3783 event->pending_kill = 0;
3787 static void perf_pending_event(struct irq_work *entry)
3789 struct perf_event *event = container_of(entry,
3790 struct perf_event, pending);
3792 if (event->pending_disable) {
3793 event->pending_disable = 0;
3794 __perf_event_disable(event);
3797 if (event->pending_wakeup) {
3798 event->pending_wakeup = 0;
3799 perf_event_wakeup(event);
3804 * We assume there is only KVM supporting the callbacks.
3805 * Later on, we might change it to a list if there is
3806 * another virtualization implementation supporting the callbacks.
3808 struct perf_guest_info_callbacks *perf_guest_cbs;
3810 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3812 perf_guest_cbs = cbs;
3815 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3817 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3819 perf_guest_cbs = NULL;
3822 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3825 perf_output_sample_regs(struct perf_output_handle *handle,
3826 struct pt_regs *regs, u64 mask)
3830 for_each_set_bit(bit, (const unsigned long *) &mask,
3831 sizeof(mask) * BITS_PER_BYTE) {
3834 val = perf_reg_value(regs, bit);
3835 perf_output_put(handle, val);
3839 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3840 struct pt_regs *regs)
3842 if (!user_mode(regs)) {
3844 regs = task_pt_regs(current);
3850 regs_user->regs = regs;
3851 regs_user->abi = perf_reg_abi(current);
3856 * Get remaining task size from user stack pointer.
3858 * It'd be better to take stack vma map and limit this more
3859 * precisly, but there's no way to get it safely under interrupt,
3860 * so using TASK_SIZE as limit.
3862 static u64 perf_ustack_task_size(struct pt_regs *regs)
3864 unsigned long addr = perf_user_stack_pointer(regs);
3866 if (!addr || addr >= TASK_SIZE)
3869 return TASK_SIZE - addr;
3873 perf_sample_ustack_size(u16 stack_size, u16 header_size,
3874 struct pt_regs *regs)
3878 /* No regs, no stack pointer, no dump. */
3883 * Check if we fit in with the requested stack size into the:
3885 * If we don't, we limit the size to the TASK_SIZE.
3887 * - remaining sample size
3888 * If we don't, we customize the stack size to
3889 * fit in to the remaining sample size.
3892 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3893 stack_size = min(stack_size, (u16) task_size);
3895 /* Current header size plus static size and dynamic size. */
3896 header_size += 2 * sizeof(u64);
3898 /* Do we fit in with the current stack dump size? */
3899 if ((u16) (header_size + stack_size) < header_size) {
3901 * If we overflow the maximum size for the sample,
3902 * we customize the stack dump size to fit in.
3904 stack_size = USHRT_MAX - header_size - sizeof(u64);
3905 stack_size = round_up(stack_size, sizeof(u64));
3912 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
3913 struct pt_regs *regs)
3915 /* Case of a kernel thread, nothing to dump */
3918 perf_output_put(handle, size);
3927 * - the size requested by user or the best one we can fit
3928 * in to the sample max size
3930 * - user stack dump data
3932 * - the actual dumped size
3936 perf_output_put(handle, dump_size);
3939 sp = perf_user_stack_pointer(regs);
3940 rem = __output_copy_user(handle, (void *) sp, dump_size);
3941 dyn_size = dump_size - rem;
3943 perf_output_skip(handle, rem);
3946 perf_output_put(handle, dyn_size);
3950 static void __perf_event_header__init_id(struct perf_event_header *header,
3951 struct perf_sample_data *data,
3952 struct perf_event *event)
3954 u64 sample_type = event->attr.sample_type;
3956 data->type = sample_type;
3957 header->size += event->id_header_size;
3959 if (sample_type & PERF_SAMPLE_TID) {
3960 /* namespace issues */
3961 data->tid_entry.pid = perf_event_pid(event, current);
3962 data->tid_entry.tid = perf_event_tid(event, current);
3965 if (sample_type & PERF_SAMPLE_TIME)
3966 data->time = perf_clock();
3968 if (sample_type & PERF_SAMPLE_ID)
3969 data->id = primary_event_id(event);
3971 if (sample_type & PERF_SAMPLE_STREAM_ID)
3972 data->stream_id = event->id;
3974 if (sample_type & PERF_SAMPLE_CPU) {
3975 data->cpu_entry.cpu = raw_smp_processor_id();
3976 data->cpu_entry.reserved = 0;
3980 void perf_event_header__init_id(struct perf_event_header *header,
3981 struct perf_sample_data *data,
3982 struct perf_event *event)
3984 if (event->attr.sample_id_all)
3985 __perf_event_header__init_id(header, data, event);
3988 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3989 struct perf_sample_data *data)
3991 u64 sample_type = data->type;
3993 if (sample_type & PERF_SAMPLE_TID)
3994 perf_output_put(handle, data->tid_entry);
3996 if (sample_type & PERF_SAMPLE_TIME)
3997 perf_output_put(handle, data->time);
3999 if (sample_type & PERF_SAMPLE_ID)
4000 perf_output_put(handle, data->id);
4002 if (sample_type & PERF_SAMPLE_STREAM_ID)
4003 perf_output_put(handle, data->stream_id);
4005 if (sample_type & PERF_SAMPLE_CPU)
4006 perf_output_put(handle, data->cpu_entry);
4009 void perf_event__output_id_sample(struct perf_event *event,
4010 struct perf_output_handle *handle,
4011 struct perf_sample_data *sample)
4013 if (event->attr.sample_id_all)
4014 __perf_event__output_id_sample(handle, sample);
4017 static void perf_output_read_one(struct perf_output_handle *handle,
4018 struct perf_event *event,
4019 u64 enabled, u64 running)
4021 u64 read_format = event->attr.read_format;
4025 values[n++] = perf_event_count(event);
4026 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4027 values[n++] = enabled +
4028 atomic64_read(&event->child_total_time_enabled);
4030 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4031 values[n++] = running +
4032 atomic64_read(&event->child_total_time_running);
4034 if (read_format & PERF_FORMAT_ID)
4035 values[n++] = primary_event_id(event);
4037 __output_copy(handle, values, n * sizeof(u64));
4041 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4043 static void perf_output_read_group(struct perf_output_handle *handle,
4044 struct perf_event *event,
4045 u64 enabled, u64 running)
4047 struct perf_event *leader = event->group_leader, *sub;
4048 u64 read_format = event->attr.read_format;
4052 values[n++] = 1 + leader->nr_siblings;
4054 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4055 values[n++] = enabled;
4057 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4058 values[n++] = running;
4060 if (leader != event)
4061 leader->pmu->read(leader);
4063 values[n++] = perf_event_count(leader);
4064 if (read_format & PERF_FORMAT_ID)
4065 values[n++] = primary_event_id(leader);
4067 __output_copy(handle, values, n * sizeof(u64));
4069 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4073 sub->pmu->read(sub);
4075 values[n++] = perf_event_count(sub);
4076 if (read_format & PERF_FORMAT_ID)
4077 values[n++] = primary_event_id(sub);
4079 __output_copy(handle, values, n * sizeof(u64));
4083 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4084 PERF_FORMAT_TOTAL_TIME_RUNNING)
4086 static void perf_output_read(struct perf_output_handle *handle,
4087 struct perf_event *event)
4089 u64 enabled = 0, running = 0, now;
4090 u64 read_format = event->attr.read_format;
4093 * compute total_time_enabled, total_time_running
4094 * based on snapshot values taken when the event
4095 * was last scheduled in.
4097 * we cannot simply called update_context_time()
4098 * because of locking issue as we are called in
4101 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4102 calc_timer_values(event, &now, &enabled, &running);
4104 if (event->attr.read_format & PERF_FORMAT_GROUP)
4105 perf_output_read_group(handle, event, enabled, running);
4107 perf_output_read_one(handle, event, enabled, running);
4110 void perf_output_sample(struct perf_output_handle *handle,
4111 struct perf_event_header *header,
4112 struct perf_sample_data *data,
4113 struct perf_event *event)
4115 u64 sample_type = data->type;
4117 perf_output_put(handle, *header);
4119 if (sample_type & PERF_SAMPLE_IP)
4120 perf_output_put(handle, data->ip);
4122 if (sample_type & PERF_SAMPLE_TID)
4123 perf_output_put(handle, data->tid_entry);
4125 if (sample_type & PERF_SAMPLE_TIME)
4126 perf_output_put(handle, data->time);
4128 if (sample_type & PERF_SAMPLE_ADDR)
4129 perf_output_put(handle, data->addr);
4131 if (sample_type & PERF_SAMPLE_ID)
4132 perf_output_put(handle, data->id);
4134 if (sample_type & PERF_SAMPLE_STREAM_ID)
4135 perf_output_put(handle, data->stream_id);
4137 if (sample_type & PERF_SAMPLE_CPU)
4138 perf_output_put(handle, data->cpu_entry);
4140 if (sample_type & PERF_SAMPLE_PERIOD)
4141 perf_output_put(handle, data->period);
4143 if (sample_type & PERF_SAMPLE_READ)
4144 perf_output_read(handle, event);
4146 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4147 if (data->callchain) {
4150 if (data->callchain)
4151 size += data->callchain->nr;
4153 size *= sizeof(u64);
4155 __output_copy(handle, data->callchain, size);
4158 perf_output_put(handle, nr);
4162 if (sample_type & PERF_SAMPLE_RAW) {
4164 perf_output_put(handle, data->raw->size);
4165 __output_copy(handle, data->raw->data,
4172 .size = sizeof(u32),
4175 perf_output_put(handle, raw);
4179 if (!event->attr.watermark) {
4180 int wakeup_events = event->attr.wakeup_events;
4182 if (wakeup_events) {
4183 struct ring_buffer *rb = handle->rb;
4184 int events = local_inc_return(&rb->events);
4186 if (events >= wakeup_events) {
4187 local_sub(wakeup_events, &rb->events);
4188 local_inc(&rb->wakeup);
4193 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4194 if (data->br_stack) {
4197 size = data->br_stack->nr
4198 * sizeof(struct perf_branch_entry);
4200 perf_output_put(handle, data->br_stack->nr);
4201 perf_output_copy(handle, data->br_stack->entries, size);
4204 * we always store at least the value of nr
4207 perf_output_put(handle, nr);
4211 if (sample_type & PERF_SAMPLE_REGS_USER) {
4212 u64 abi = data->regs_user.abi;
4215 * If there are no regs to dump, notice it through
4216 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4218 perf_output_put(handle, abi);
4221 u64 mask = event->attr.sample_regs_user;
4222 perf_output_sample_regs(handle,
4223 data->regs_user.regs,
4228 if (sample_type & PERF_SAMPLE_STACK_USER)
4229 perf_output_sample_ustack(handle,
4230 data->stack_user_size,
4231 data->regs_user.regs);
4233 if (sample_type & PERF_SAMPLE_WEIGHT)
4234 perf_output_put(handle, data->weight);
4236 if (sample_type & PERF_SAMPLE_DATA_SRC)
4237 perf_output_put(handle, data->data_src.val);
4240 void perf_prepare_sample(struct perf_event_header *header,
4241 struct perf_sample_data *data,
4242 struct perf_event *event,
4243 struct pt_regs *regs)
4245 u64 sample_type = event->attr.sample_type;
4247 header->type = PERF_RECORD_SAMPLE;
4248 header->size = sizeof(*header) + event->header_size;
4251 header->misc |= perf_misc_flags(regs);
4253 __perf_event_header__init_id(header, data, event);
4255 if (sample_type & PERF_SAMPLE_IP)
4256 data->ip = perf_instruction_pointer(regs);
4258 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4261 data->callchain = perf_callchain(event, regs);
4263 if (data->callchain)
4264 size += data->callchain->nr;
4266 header->size += size * sizeof(u64);
4269 if (sample_type & PERF_SAMPLE_RAW) {
4270 int size = sizeof(u32);
4273 size += data->raw->size;
4275 size += sizeof(u32);
4277 WARN_ON_ONCE(size & (sizeof(u64)-1));
4278 header->size += size;
4281 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4282 int size = sizeof(u64); /* nr */
4283 if (data->br_stack) {
4284 size += data->br_stack->nr
4285 * sizeof(struct perf_branch_entry);
4287 header->size += size;
4290 if (sample_type & PERF_SAMPLE_REGS_USER) {
4291 /* regs dump ABI info */
4292 int size = sizeof(u64);
4294 perf_sample_regs_user(&data->regs_user, regs);
4296 if (data->regs_user.regs) {
4297 u64 mask = event->attr.sample_regs_user;
4298 size += hweight64(mask) * sizeof(u64);
4301 header->size += size;
4304 if (sample_type & PERF_SAMPLE_STACK_USER) {
4306 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4307 * processed as the last one or have additional check added
4308 * in case new sample type is added, because we could eat
4309 * up the rest of the sample size.
4311 struct perf_regs_user *uregs = &data->regs_user;
4312 u16 stack_size = event->attr.sample_stack_user;
4313 u16 size = sizeof(u64);
4316 perf_sample_regs_user(uregs, regs);
4318 stack_size = perf_sample_ustack_size(stack_size, header->size,
4322 * If there is something to dump, add space for the dump
4323 * itself and for the field that tells the dynamic size,
4324 * which is how many have been actually dumped.
4327 size += sizeof(u64) + stack_size;
4329 data->stack_user_size = stack_size;
4330 header->size += size;
4334 static void perf_event_output(struct perf_event *event,
4335 struct perf_sample_data *data,
4336 struct pt_regs *regs)
4338 struct perf_output_handle handle;
4339 struct perf_event_header header;
4341 /* protect the callchain buffers */
4344 perf_prepare_sample(&header, data, event, regs);
4346 if (perf_output_begin(&handle, event, header.size))
4349 perf_output_sample(&handle, &header, data, event);
4351 perf_output_end(&handle);
4361 struct perf_read_event {
4362 struct perf_event_header header;
4369 perf_event_read_event(struct perf_event *event,
4370 struct task_struct *task)
4372 struct perf_output_handle handle;
4373 struct perf_sample_data sample;
4374 struct perf_read_event read_event = {
4376 .type = PERF_RECORD_READ,
4378 .size = sizeof(read_event) + event->read_size,
4380 .pid = perf_event_pid(event, task),
4381 .tid = perf_event_tid(event, task),
4385 perf_event_header__init_id(&read_event.header, &sample, event);
4386 ret = perf_output_begin(&handle, event, read_event.header.size);
4390 perf_output_put(&handle, read_event);
4391 perf_output_read(&handle, event);
4392 perf_event__output_id_sample(event, &handle, &sample);
4394 perf_output_end(&handle);
4397 typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data);
4398 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4401 perf_event_aux_ctx(struct perf_event_context *ctx,
4402 perf_event_aux_match_cb match,
4403 perf_event_aux_output_cb output,
4406 struct perf_event *event;
4408 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4409 if (event->state < PERF_EVENT_STATE_INACTIVE)
4411 if (!event_filter_match(event))
4413 if (match(event, data))
4414 output(event, data);
4419 perf_event_aux(perf_event_aux_match_cb match,
4420 perf_event_aux_output_cb output,
4422 struct perf_event_context *task_ctx)
4424 struct perf_cpu_context *cpuctx;
4425 struct perf_event_context *ctx;
4430 list_for_each_entry_rcu(pmu, &pmus, entry) {
4431 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4432 if (cpuctx->unique_pmu != pmu)
4434 perf_event_aux_ctx(&cpuctx->ctx, match, output, data);
4437 ctxn = pmu->task_ctx_nr;
4440 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4442 perf_event_aux_ctx(ctx, match, output, data);
4444 put_cpu_ptr(pmu->pmu_cpu_context);
4449 perf_event_aux_ctx(task_ctx, match, output, data);
4456 * task tracking -- fork/exit
4458 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4461 struct perf_task_event {
4462 struct task_struct *task;
4463 struct perf_event_context *task_ctx;
4466 struct perf_event_header header;
4476 static void perf_event_task_output(struct perf_event *event,
4479 struct perf_task_event *task_event = data;
4480 struct perf_output_handle handle;
4481 struct perf_sample_data sample;
4482 struct task_struct *task = task_event->task;
4483 int ret, size = task_event->event_id.header.size;
4485 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4487 ret = perf_output_begin(&handle, event,
4488 task_event->event_id.header.size);
4492 task_event->event_id.pid = perf_event_pid(event, task);
4493 task_event->event_id.ppid = perf_event_pid(event, current);
4495 task_event->event_id.tid = perf_event_tid(event, task);
4496 task_event->event_id.ptid = perf_event_tid(event, current);
4498 perf_output_put(&handle, task_event->event_id);
4500 perf_event__output_id_sample(event, &handle, &sample);
4502 perf_output_end(&handle);
4504 task_event->event_id.header.size = size;
4507 static int perf_event_task_match(struct perf_event *event,
4508 void *data __maybe_unused)
4510 return event->attr.comm || event->attr.mmap ||
4511 event->attr.mmap_data || event->attr.task;
4514 static void perf_event_task(struct task_struct *task,
4515 struct perf_event_context *task_ctx,
4518 struct perf_task_event task_event;
4520 if (!atomic_read(&nr_comm_events) &&
4521 !atomic_read(&nr_mmap_events) &&
4522 !atomic_read(&nr_task_events))
4525 task_event = (struct perf_task_event){
4527 .task_ctx = task_ctx,
4530 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4532 .size = sizeof(task_event.event_id),
4538 .time = perf_clock(),
4542 perf_event_aux(perf_event_task_match,
4543 perf_event_task_output,
4548 void perf_event_fork(struct task_struct *task)
4550 perf_event_task(task, NULL, 1);
4557 struct perf_comm_event {
4558 struct task_struct *task;
4563 struct perf_event_header header;
4570 static void perf_event_comm_output(struct perf_event *event,
4573 struct perf_comm_event *comm_event = data;
4574 struct perf_output_handle handle;
4575 struct perf_sample_data sample;
4576 int size = comm_event->event_id.header.size;
4579 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4580 ret = perf_output_begin(&handle, event,
4581 comm_event->event_id.header.size);
4586 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4587 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4589 perf_output_put(&handle, comm_event->event_id);
4590 __output_copy(&handle, comm_event->comm,
4591 comm_event->comm_size);
4593 perf_event__output_id_sample(event, &handle, &sample);
4595 perf_output_end(&handle);
4597 comm_event->event_id.header.size = size;
4600 static int perf_event_comm_match(struct perf_event *event,
4601 void *data __maybe_unused)
4603 return event->attr.comm;
4606 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4608 char comm[TASK_COMM_LEN];
4611 memset(comm, 0, sizeof(comm));
4612 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4613 size = ALIGN(strlen(comm)+1, sizeof(u64));
4615 comm_event->comm = comm;
4616 comm_event->comm_size = size;
4618 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4620 perf_event_aux(perf_event_comm_match,
4621 perf_event_comm_output,
4626 void perf_event_comm(struct task_struct *task)
4628 struct perf_comm_event comm_event;
4629 struct perf_event_context *ctx;
4633 for_each_task_context_nr(ctxn) {
4634 ctx = task->perf_event_ctxp[ctxn];
4638 perf_event_enable_on_exec(ctx);
4642 if (!atomic_read(&nr_comm_events))
4645 comm_event = (struct perf_comm_event){
4651 .type = PERF_RECORD_COMM,
4660 perf_event_comm_event(&comm_event);
4667 struct perf_mmap_event {
4668 struct vm_area_struct *vma;
4670 const char *file_name;
4674 struct perf_event_header header;
4684 static void perf_event_mmap_output(struct perf_event *event,
4687 struct perf_mmap_event *mmap_event = data;
4688 struct perf_output_handle handle;
4689 struct perf_sample_data sample;
4690 int size = mmap_event->event_id.header.size;
4693 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4694 ret = perf_output_begin(&handle, event,
4695 mmap_event->event_id.header.size);
4699 mmap_event->event_id.pid = perf_event_pid(event, current);
4700 mmap_event->event_id.tid = perf_event_tid(event, current);
4702 perf_output_put(&handle, mmap_event->event_id);
4703 __output_copy(&handle, mmap_event->file_name,
4704 mmap_event->file_size);
4706 perf_event__output_id_sample(event, &handle, &sample);
4708 perf_output_end(&handle);
4710 mmap_event->event_id.header.size = size;
4713 static int perf_event_mmap_match(struct perf_event *event,
4716 struct perf_mmap_event *mmap_event = data;
4717 struct vm_area_struct *vma = mmap_event->vma;
4718 int executable = vma->vm_flags & VM_EXEC;
4720 return (!executable && event->attr.mmap_data) ||
4721 (executable && event->attr.mmap);
4724 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4726 struct vm_area_struct *vma = mmap_event->vma;
4727 struct file *file = vma->vm_file;
4733 memset(tmp, 0, sizeof(tmp));
4737 * d_path works from the end of the rb backwards, so we
4738 * need to add enough zero bytes after the string to handle
4739 * the 64bit alignment we do later.
4741 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4743 name = strncpy(tmp, "//enomem", sizeof(tmp));
4746 name = d_path(&file->f_path, buf, PATH_MAX);
4748 name = strncpy(tmp, "//toolong", sizeof(tmp));
4752 if (arch_vma_name(mmap_event->vma)) {
4753 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4755 tmp[sizeof(tmp) - 1] = '\0';
4760 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4762 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4763 vma->vm_end >= vma->vm_mm->brk) {
4764 name = strncpy(tmp, "[heap]", sizeof(tmp));
4766 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4767 vma->vm_end >= vma->vm_mm->start_stack) {
4768 name = strncpy(tmp, "[stack]", sizeof(tmp));
4772 name = strncpy(tmp, "//anon", sizeof(tmp));
4777 size = ALIGN(strlen(name)+1, sizeof(u64));
4779 mmap_event->file_name = name;
4780 mmap_event->file_size = size;
4782 if (!(vma->vm_flags & VM_EXEC))
4783 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
4785 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4787 perf_event_aux(perf_event_mmap_match,
4788 perf_event_mmap_output,
4795 void perf_event_mmap(struct vm_area_struct *vma)
4797 struct perf_mmap_event mmap_event;
4799 if (!atomic_read(&nr_mmap_events))
4802 mmap_event = (struct perf_mmap_event){
4808 .type = PERF_RECORD_MMAP,
4809 .misc = PERF_RECORD_MISC_USER,
4814 .start = vma->vm_start,
4815 .len = vma->vm_end - vma->vm_start,
4816 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4820 perf_event_mmap_event(&mmap_event);
4824 * IRQ throttle logging
4827 static void perf_log_throttle(struct perf_event *event, int enable)
4829 struct perf_output_handle handle;
4830 struct perf_sample_data sample;
4834 struct perf_event_header header;
4838 } throttle_event = {
4840 .type = PERF_RECORD_THROTTLE,
4842 .size = sizeof(throttle_event),
4844 .time = perf_clock(),
4845 .id = primary_event_id(event),
4846 .stream_id = event->id,
4850 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4852 perf_event_header__init_id(&throttle_event.header, &sample, event);
4854 ret = perf_output_begin(&handle, event,
4855 throttle_event.header.size);
4859 perf_output_put(&handle, throttle_event);
4860 perf_event__output_id_sample(event, &handle, &sample);
4861 perf_output_end(&handle);
4865 * Generic event overflow handling, sampling.
4868 static int __perf_event_overflow(struct perf_event *event,
4869 int throttle, struct perf_sample_data *data,
4870 struct pt_regs *regs)
4872 int events = atomic_read(&event->event_limit);
4873 struct hw_perf_event *hwc = &event->hw;
4878 * Non-sampling counters might still use the PMI to fold short
4879 * hardware counters, ignore those.
4881 if (unlikely(!is_sampling_event(event)))
4884 seq = __this_cpu_read(perf_throttled_seq);
4885 if (seq != hwc->interrupts_seq) {
4886 hwc->interrupts_seq = seq;
4887 hwc->interrupts = 1;
4890 if (unlikely(throttle
4891 && hwc->interrupts >= max_samples_per_tick)) {
4892 __this_cpu_inc(perf_throttled_count);
4893 hwc->interrupts = MAX_INTERRUPTS;
4894 perf_log_throttle(event, 0);
4899 if (event->attr.freq) {
4900 u64 now = perf_clock();
4901 s64 delta = now - hwc->freq_time_stamp;
4903 hwc->freq_time_stamp = now;
4905 if (delta > 0 && delta < 2*TICK_NSEC)
4906 perf_adjust_period(event, delta, hwc->last_period, true);
4910 * XXX event_limit might not quite work as expected on inherited
4914 event->pending_kill = POLL_IN;
4915 if (events && atomic_dec_and_test(&event->event_limit)) {
4917 event->pending_kill = POLL_HUP;
4918 event->pending_disable = 1;
4919 irq_work_queue(&event->pending);
4922 if (event->overflow_handler)
4923 event->overflow_handler(event, data, regs);
4925 perf_event_output(event, data, regs);
4927 if (event->fasync && event->pending_kill) {
4928 event->pending_wakeup = 1;
4929 irq_work_queue(&event->pending);
4935 int perf_event_overflow(struct perf_event *event,
4936 struct perf_sample_data *data,
4937 struct pt_regs *regs)
4939 return __perf_event_overflow(event, 1, data, regs);
4943 * Generic software event infrastructure
4946 struct swevent_htable {
4947 struct swevent_hlist *swevent_hlist;
4948 struct mutex hlist_mutex;
4951 /* Recursion avoidance in each contexts */
4952 int recursion[PERF_NR_CONTEXTS];
4955 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4958 * We directly increment event->count and keep a second value in
4959 * event->hw.period_left to count intervals. This period event
4960 * is kept in the range [-sample_period, 0] so that we can use the
4964 u64 perf_swevent_set_period(struct perf_event *event)
4966 struct hw_perf_event *hwc = &event->hw;
4967 u64 period = hwc->last_period;
4971 hwc->last_period = hwc->sample_period;
4974 old = val = local64_read(&hwc->period_left);
4978 nr = div64_u64(period + val, period);
4979 offset = nr * period;
4981 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4987 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4988 struct perf_sample_data *data,
4989 struct pt_regs *regs)
4991 struct hw_perf_event *hwc = &event->hw;
4995 overflow = perf_swevent_set_period(event);
4997 if (hwc->interrupts == MAX_INTERRUPTS)
5000 for (; overflow; overflow--) {
5001 if (__perf_event_overflow(event, throttle,
5004 * We inhibit the overflow from happening when
5005 * hwc->interrupts == MAX_INTERRUPTS.
5013 static void perf_swevent_event(struct perf_event *event, u64 nr,
5014 struct perf_sample_data *data,
5015 struct pt_regs *regs)
5017 struct hw_perf_event *hwc = &event->hw;
5019 local64_add(nr, &event->count);
5024 if (!is_sampling_event(event))
5027 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5029 return perf_swevent_overflow(event, 1, data, regs);
5031 data->period = event->hw.last_period;
5033 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5034 return perf_swevent_overflow(event, 1, data, regs);
5036 if (local64_add_negative(nr, &hwc->period_left))
5039 perf_swevent_overflow(event, 0, data, regs);
5042 static int perf_exclude_event(struct perf_event *event,
5043 struct pt_regs *regs)
5045 if (event->hw.state & PERF_HES_STOPPED)
5049 if (event->attr.exclude_user && user_mode(regs))
5052 if (event->attr.exclude_kernel && !user_mode(regs))
5059 static int perf_swevent_match(struct perf_event *event,
5060 enum perf_type_id type,
5062 struct perf_sample_data *data,
5063 struct pt_regs *regs)
5065 if (event->attr.type != type)
5068 if (event->attr.config != event_id)
5071 if (perf_exclude_event(event, regs))
5077 static inline u64 swevent_hash(u64 type, u32 event_id)
5079 u64 val = event_id | (type << 32);
5081 return hash_64(val, SWEVENT_HLIST_BITS);
5084 static inline struct hlist_head *
5085 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5087 u64 hash = swevent_hash(type, event_id);
5089 return &hlist->heads[hash];
5092 /* For the read side: events when they trigger */
5093 static inline struct hlist_head *
5094 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5096 struct swevent_hlist *hlist;
5098 hlist = rcu_dereference(swhash->swevent_hlist);
5102 return __find_swevent_head(hlist, type, event_id);
5105 /* For the event head insertion and removal in the hlist */
5106 static inline struct hlist_head *
5107 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5109 struct swevent_hlist *hlist;
5110 u32 event_id = event->attr.config;
5111 u64 type = event->attr.type;
5114 * Event scheduling is always serialized against hlist allocation
5115 * and release. Which makes the protected version suitable here.
5116 * The context lock guarantees that.
5118 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5119 lockdep_is_held(&event->ctx->lock));
5123 return __find_swevent_head(hlist, type, event_id);
5126 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5128 struct perf_sample_data *data,
5129 struct pt_regs *regs)
5131 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5132 struct perf_event *event;
5133 struct hlist_head *head;
5136 head = find_swevent_head_rcu(swhash, type, event_id);
5140 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5141 if (perf_swevent_match(event, type, event_id, data, regs))
5142 perf_swevent_event(event, nr, data, regs);
5148 int perf_swevent_get_recursion_context(void)
5150 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5152 return get_recursion_context(swhash->recursion);
5154 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5156 inline void perf_swevent_put_recursion_context(int rctx)
5158 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5160 put_recursion_context(swhash->recursion, rctx);
5163 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5165 struct perf_sample_data data;
5168 preempt_disable_notrace();
5169 rctx = perf_swevent_get_recursion_context();
5173 perf_sample_data_init(&data, addr, 0);
5175 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5177 perf_swevent_put_recursion_context(rctx);
5178 preempt_enable_notrace();
5181 static void perf_swevent_read(struct perf_event *event)
5185 static int perf_swevent_add(struct perf_event *event, int flags)
5187 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5188 struct hw_perf_event *hwc = &event->hw;
5189 struct hlist_head *head;
5191 if (is_sampling_event(event)) {
5192 hwc->last_period = hwc->sample_period;
5193 perf_swevent_set_period(event);
5196 hwc->state = !(flags & PERF_EF_START);
5198 head = find_swevent_head(swhash, event);
5199 if (WARN_ON_ONCE(!head))
5202 hlist_add_head_rcu(&event->hlist_entry, head);
5207 static void perf_swevent_del(struct perf_event *event, int flags)
5209 hlist_del_rcu(&event->hlist_entry);
5212 static void perf_swevent_start(struct perf_event *event, int flags)
5214 event->hw.state = 0;
5217 static void perf_swevent_stop(struct perf_event *event, int flags)
5219 event->hw.state = PERF_HES_STOPPED;
5222 /* Deref the hlist from the update side */
5223 static inline struct swevent_hlist *
5224 swevent_hlist_deref(struct swevent_htable *swhash)
5226 return rcu_dereference_protected(swhash->swevent_hlist,
5227 lockdep_is_held(&swhash->hlist_mutex));
5230 static void swevent_hlist_release(struct swevent_htable *swhash)
5232 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5237 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5238 kfree_rcu(hlist, rcu_head);
5241 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5243 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5245 mutex_lock(&swhash->hlist_mutex);
5247 if (!--swhash->hlist_refcount)
5248 swevent_hlist_release(swhash);
5250 mutex_unlock(&swhash->hlist_mutex);
5253 static void swevent_hlist_put(struct perf_event *event)
5257 if (event->cpu != -1) {
5258 swevent_hlist_put_cpu(event, event->cpu);
5262 for_each_possible_cpu(cpu)
5263 swevent_hlist_put_cpu(event, cpu);
5266 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5268 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5271 mutex_lock(&swhash->hlist_mutex);
5273 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5274 struct swevent_hlist *hlist;
5276 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5281 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5283 swhash->hlist_refcount++;
5285 mutex_unlock(&swhash->hlist_mutex);
5290 static int swevent_hlist_get(struct perf_event *event)
5293 int cpu, failed_cpu;
5295 if (event->cpu != -1)
5296 return swevent_hlist_get_cpu(event, event->cpu);
5299 for_each_possible_cpu(cpu) {
5300 err = swevent_hlist_get_cpu(event, cpu);
5310 for_each_possible_cpu(cpu) {
5311 if (cpu == failed_cpu)
5313 swevent_hlist_put_cpu(event, cpu);
5320 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5322 static void sw_perf_event_destroy(struct perf_event *event)
5324 u64 event_id = event->attr.config;
5326 WARN_ON(event->parent);
5328 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5329 swevent_hlist_put(event);
5332 static int perf_swevent_init(struct perf_event *event)
5334 u64 event_id = event->attr.config;
5336 if (event->attr.type != PERF_TYPE_SOFTWARE)
5340 * no branch sampling for software events
5342 if (has_branch_stack(event))
5346 case PERF_COUNT_SW_CPU_CLOCK:
5347 case PERF_COUNT_SW_TASK_CLOCK:
5354 if (event_id >= PERF_COUNT_SW_MAX)
5357 if (!event->parent) {
5360 err = swevent_hlist_get(event);
5364 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5365 event->destroy = sw_perf_event_destroy;
5371 static int perf_swevent_event_idx(struct perf_event *event)
5376 static struct pmu perf_swevent = {
5377 .task_ctx_nr = perf_sw_context,
5379 .event_init = perf_swevent_init,
5380 .add = perf_swevent_add,
5381 .del = perf_swevent_del,
5382 .start = perf_swevent_start,
5383 .stop = perf_swevent_stop,
5384 .read = perf_swevent_read,
5386 .event_idx = perf_swevent_event_idx,
5389 #ifdef CONFIG_EVENT_TRACING
5391 static int perf_tp_filter_match(struct perf_event *event,
5392 struct perf_sample_data *data)
5394 void *record = data->raw->data;
5396 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5401 static int perf_tp_event_match(struct perf_event *event,
5402 struct perf_sample_data *data,
5403 struct pt_regs *regs)
5405 if (event->hw.state & PERF_HES_STOPPED)
5408 * All tracepoints are from kernel-space.
5410 if (event->attr.exclude_kernel)
5413 if (!perf_tp_filter_match(event, data))
5419 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5420 struct pt_regs *regs, struct hlist_head *head, int rctx,
5421 struct task_struct *task)
5423 struct perf_sample_data data;
5424 struct perf_event *event;
5426 struct perf_raw_record raw = {
5431 perf_sample_data_init(&data, addr, 0);
5434 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5435 if (perf_tp_event_match(event, &data, regs))
5436 perf_swevent_event(event, count, &data, regs);
5440 * If we got specified a target task, also iterate its context and
5441 * deliver this event there too.
5443 if (task && task != current) {
5444 struct perf_event_context *ctx;
5445 struct trace_entry *entry = record;
5448 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5452 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5453 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5455 if (event->attr.config != entry->type)
5457 if (perf_tp_event_match(event, &data, regs))
5458 perf_swevent_event(event, count, &data, regs);
5464 perf_swevent_put_recursion_context(rctx);
5466 EXPORT_SYMBOL_GPL(perf_tp_event);
5468 static void tp_perf_event_destroy(struct perf_event *event)
5470 perf_trace_destroy(event);
5473 static int perf_tp_event_init(struct perf_event *event)
5477 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5481 * no branch sampling for tracepoint events
5483 if (has_branch_stack(event))
5486 err = perf_trace_init(event);
5490 event->destroy = tp_perf_event_destroy;
5495 static struct pmu perf_tracepoint = {
5496 .task_ctx_nr = perf_sw_context,
5498 .event_init = perf_tp_event_init,
5499 .add = perf_trace_add,
5500 .del = perf_trace_del,
5501 .start = perf_swevent_start,
5502 .stop = perf_swevent_stop,
5503 .read = perf_swevent_read,
5505 .event_idx = perf_swevent_event_idx,
5508 static inline void perf_tp_register(void)
5510 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5513 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5518 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5521 filter_str = strndup_user(arg, PAGE_SIZE);
5522 if (IS_ERR(filter_str))
5523 return PTR_ERR(filter_str);
5525 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5531 static void perf_event_free_filter(struct perf_event *event)
5533 ftrace_profile_free_filter(event);
5538 static inline void perf_tp_register(void)
5542 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5547 static void perf_event_free_filter(struct perf_event *event)
5551 #endif /* CONFIG_EVENT_TRACING */
5553 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5554 void perf_bp_event(struct perf_event *bp, void *data)
5556 struct perf_sample_data sample;
5557 struct pt_regs *regs = data;
5559 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5561 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5562 perf_swevent_event(bp, 1, &sample, regs);
5567 * hrtimer based swevent callback
5570 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5572 enum hrtimer_restart ret = HRTIMER_RESTART;
5573 struct perf_sample_data data;
5574 struct pt_regs *regs;
5575 struct perf_event *event;
5578 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5580 if (event->state != PERF_EVENT_STATE_ACTIVE)
5581 return HRTIMER_NORESTART;
5583 event->pmu->read(event);
5585 perf_sample_data_init(&data, 0, event->hw.last_period);
5586 regs = get_irq_regs();
5588 if (regs && !perf_exclude_event(event, regs)) {
5589 if (!(event->attr.exclude_idle && is_idle_task(current)))
5590 if (__perf_event_overflow(event, 1, &data, regs))
5591 ret = HRTIMER_NORESTART;
5594 period = max_t(u64, 10000, event->hw.sample_period);
5595 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5600 static void perf_swevent_start_hrtimer(struct perf_event *event)
5602 struct hw_perf_event *hwc = &event->hw;
5605 if (!is_sampling_event(event))
5608 period = local64_read(&hwc->period_left);
5613 local64_set(&hwc->period_left, 0);
5615 period = max_t(u64, 10000, hwc->sample_period);
5617 __hrtimer_start_range_ns(&hwc->hrtimer,
5618 ns_to_ktime(period), 0,
5619 HRTIMER_MODE_REL_PINNED, 0);
5622 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5624 struct hw_perf_event *hwc = &event->hw;
5626 if (is_sampling_event(event)) {
5627 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5628 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5630 hrtimer_cancel(&hwc->hrtimer);
5634 static void perf_swevent_init_hrtimer(struct perf_event *event)
5636 struct hw_perf_event *hwc = &event->hw;
5638 if (!is_sampling_event(event))
5641 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5642 hwc->hrtimer.function = perf_swevent_hrtimer;
5645 * Since hrtimers have a fixed rate, we can do a static freq->period
5646 * mapping and avoid the whole period adjust feedback stuff.
5648 if (event->attr.freq) {
5649 long freq = event->attr.sample_freq;
5651 event->attr.sample_period = NSEC_PER_SEC / freq;
5652 hwc->sample_period = event->attr.sample_period;
5653 local64_set(&hwc->period_left, hwc->sample_period);
5654 hwc->last_period = hwc->sample_period;
5655 event->attr.freq = 0;
5660 * Software event: cpu wall time clock
5663 static void cpu_clock_event_update(struct perf_event *event)
5668 now = local_clock();
5669 prev = local64_xchg(&event->hw.prev_count, now);
5670 local64_add(now - prev, &event->count);
5673 static void cpu_clock_event_start(struct perf_event *event, int flags)
5675 local64_set(&event->hw.prev_count, local_clock());
5676 perf_swevent_start_hrtimer(event);
5679 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5681 perf_swevent_cancel_hrtimer(event);
5682 cpu_clock_event_update(event);
5685 static int cpu_clock_event_add(struct perf_event *event, int flags)
5687 if (flags & PERF_EF_START)
5688 cpu_clock_event_start(event, flags);
5693 static void cpu_clock_event_del(struct perf_event *event, int flags)
5695 cpu_clock_event_stop(event, flags);
5698 static void cpu_clock_event_read(struct perf_event *event)
5700 cpu_clock_event_update(event);
5703 static int cpu_clock_event_init(struct perf_event *event)
5705 if (event->attr.type != PERF_TYPE_SOFTWARE)
5708 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5712 * no branch sampling for software events
5714 if (has_branch_stack(event))
5717 perf_swevent_init_hrtimer(event);
5722 static struct pmu perf_cpu_clock = {
5723 .task_ctx_nr = perf_sw_context,
5725 .event_init = cpu_clock_event_init,
5726 .add = cpu_clock_event_add,
5727 .del = cpu_clock_event_del,
5728 .start = cpu_clock_event_start,
5729 .stop = cpu_clock_event_stop,
5730 .read = cpu_clock_event_read,
5732 .event_idx = perf_swevent_event_idx,
5736 * Software event: task time clock
5739 static void task_clock_event_update(struct perf_event *event, u64 now)
5744 prev = local64_xchg(&event->hw.prev_count, now);
5746 local64_add(delta, &event->count);
5749 static void task_clock_event_start(struct perf_event *event, int flags)
5751 local64_set(&event->hw.prev_count, event->ctx->time);
5752 perf_swevent_start_hrtimer(event);
5755 static void task_clock_event_stop(struct perf_event *event, int flags)
5757 perf_swevent_cancel_hrtimer(event);
5758 task_clock_event_update(event, event->ctx->time);
5761 static int task_clock_event_add(struct perf_event *event, int flags)
5763 if (flags & PERF_EF_START)
5764 task_clock_event_start(event, flags);
5769 static void task_clock_event_del(struct perf_event *event, int flags)
5771 task_clock_event_stop(event, PERF_EF_UPDATE);
5774 static void task_clock_event_read(struct perf_event *event)
5776 u64 now = perf_clock();
5777 u64 delta = now - event->ctx->timestamp;
5778 u64 time = event->ctx->time + delta;
5780 task_clock_event_update(event, time);
5783 static int task_clock_event_init(struct perf_event *event)
5785 if (event->attr.type != PERF_TYPE_SOFTWARE)
5788 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5792 * no branch sampling for software events
5794 if (has_branch_stack(event))
5797 perf_swevent_init_hrtimer(event);
5802 static struct pmu perf_task_clock = {
5803 .task_ctx_nr = perf_sw_context,
5805 .event_init = task_clock_event_init,
5806 .add = task_clock_event_add,
5807 .del = task_clock_event_del,
5808 .start = task_clock_event_start,
5809 .stop = task_clock_event_stop,
5810 .read = task_clock_event_read,
5812 .event_idx = perf_swevent_event_idx,
5815 static void perf_pmu_nop_void(struct pmu *pmu)
5819 static int perf_pmu_nop_int(struct pmu *pmu)
5824 static void perf_pmu_start_txn(struct pmu *pmu)
5826 perf_pmu_disable(pmu);
5829 static int perf_pmu_commit_txn(struct pmu *pmu)
5831 perf_pmu_enable(pmu);
5835 static void perf_pmu_cancel_txn(struct pmu *pmu)
5837 perf_pmu_enable(pmu);
5840 static int perf_event_idx_default(struct perf_event *event)
5842 return event->hw.idx + 1;
5846 * Ensures all contexts with the same task_ctx_nr have the same
5847 * pmu_cpu_context too.
5849 static void *find_pmu_context(int ctxn)
5856 list_for_each_entry(pmu, &pmus, entry) {
5857 if (pmu->task_ctx_nr == ctxn)
5858 return pmu->pmu_cpu_context;
5864 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5868 for_each_possible_cpu(cpu) {
5869 struct perf_cpu_context *cpuctx;
5871 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5873 if (cpuctx->unique_pmu == old_pmu)
5874 cpuctx->unique_pmu = pmu;
5878 static void free_pmu_context(struct pmu *pmu)
5882 mutex_lock(&pmus_lock);
5884 * Like a real lame refcount.
5886 list_for_each_entry(i, &pmus, entry) {
5887 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5888 update_pmu_context(i, pmu);
5893 free_percpu(pmu->pmu_cpu_context);
5895 mutex_unlock(&pmus_lock);
5897 static struct idr pmu_idr;
5900 type_show(struct device *dev, struct device_attribute *attr, char *page)
5902 struct pmu *pmu = dev_get_drvdata(dev);
5904 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5907 static struct device_attribute pmu_dev_attrs[] = {
5912 static int pmu_bus_running;
5913 static struct bus_type pmu_bus = {
5914 .name = "event_source",
5915 .dev_attrs = pmu_dev_attrs,
5918 static void pmu_dev_release(struct device *dev)
5923 static int pmu_dev_alloc(struct pmu *pmu)
5927 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5931 pmu->dev->groups = pmu->attr_groups;
5932 device_initialize(pmu->dev);
5933 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5937 dev_set_drvdata(pmu->dev, pmu);
5938 pmu->dev->bus = &pmu_bus;
5939 pmu->dev->release = pmu_dev_release;
5940 ret = device_add(pmu->dev);
5948 put_device(pmu->dev);
5952 static struct lock_class_key cpuctx_mutex;
5953 static struct lock_class_key cpuctx_lock;
5955 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5959 mutex_lock(&pmus_lock);
5961 pmu->pmu_disable_count = alloc_percpu(int);
5962 if (!pmu->pmu_disable_count)
5971 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
5979 if (pmu_bus_running) {
5980 ret = pmu_dev_alloc(pmu);
5986 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5987 if (pmu->pmu_cpu_context)
5988 goto got_cpu_context;
5991 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5992 if (!pmu->pmu_cpu_context)
5995 for_each_possible_cpu(cpu) {
5996 struct perf_cpu_context *cpuctx;
5998 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5999 __perf_event_init_context(&cpuctx->ctx);
6000 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6001 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6002 cpuctx->ctx.type = cpu_context;
6003 cpuctx->ctx.pmu = pmu;
6004 cpuctx->jiffies_interval = 1;
6005 INIT_LIST_HEAD(&cpuctx->rotation_list);
6006 cpuctx->unique_pmu = pmu;
6010 if (!pmu->start_txn) {
6011 if (pmu->pmu_enable) {
6013 * If we have pmu_enable/pmu_disable calls, install
6014 * transaction stubs that use that to try and batch
6015 * hardware accesses.
6017 pmu->start_txn = perf_pmu_start_txn;
6018 pmu->commit_txn = perf_pmu_commit_txn;
6019 pmu->cancel_txn = perf_pmu_cancel_txn;
6021 pmu->start_txn = perf_pmu_nop_void;
6022 pmu->commit_txn = perf_pmu_nop_int;
6023 pmu->cancel_txn = perf_pmu_nop_void;
6027 if (!pmu->pmu_enable) {
6028 pmu->pmu_enable = perf_pmu_nop_void;
6029 pmu->pmu_disable = perf_pmu_nop_void;
6032 if (!pmu->event_idx)
6033 pmu->event_idx = perf_event_idx_default;
6035 list_add_rcu(&pmu->entry, &pmus);
6038 mutex_unlock(&pmus_lock);
6043 device_del(pmu->dev);
6044 put_device(pmu->dev);
6047 if (pmu->type >= PERF_TYPE_MAX)
6048 idr_remove(&pmu_idr, pmu->type);
6051 free_percpu(pmu->pmu_disable_count);
6055 void perf_pmu_unregister(struct pmu *pmu)
6057 mutex_lock(&pmus_lock);
6058 list_del_rcu(&pmu->entry);
6059 mutex_unlock(&pmus_lock);
6062 * We dereference the pmu list under both SRCU and regular RCU, so
6063 * synchronize against both of those.
6065 synchronize_srcu(&pmus_srcu);
6068 free_percpu(pmu->pmu_disable_count);
6069 if (pmu->type >= PERF_TYPE_MAX)
6070 idr_remove(&pmu_idr, pmu->type);
6071 device_del(pmu->dev);
6072 put_device(pmu->dev);
6073 free_pmu_context(pmu);
6076 struct pmu *perf_init_event(struct perf_event *event)
6078 struct pmu *pmu = NULL;
6082 idx = srcu_read_lock(&pmus_srcu);
6085 pmu = idr_find(&pmu_idr, event->attr.type);
6089 ret = pmu->event_init(event);
6095 list_for_each_entry_rcu(pmu, &pmus, entry) {
6097 ret = pmu->event_init(event);
6101 if (ret != -ENOENT) {
6106 pmu = ERR_PTR(-ENOENT);
6108 srcu_read_unlock(&pmus_srcu, idx);
6114 * Allocate and initialize a event structure
6116 static struct perf_event *
6117 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6118 struct task_struct *task,
6119 struct perf_event *group_leader,
6120 struct perf_event *parent_event,
6121 perf_overflow_handler_t overflow_handler,
6125 struct perf_event *event;
6126 struct hw_perf_event *hwc;
6129 if ((unsigned)cpu >= nr_cpu_ids) {
6130 if (!task || cpu != -1)
6131 return ERR_PTR(-EINVAL);
6134 event = kzalloc(sizeof(*event), GFP_KERNEL);
6136 return ERR_PTR(-ENOMEM);
6139 * Single events are their own group leaders, with an
6140 * empty sibling list:
6143 group_leader = event;
6145 mutex_init(&event->child_mutex);
6146 INIT_LIST_HEAD(&event->child_list);
6148 INIT_LIST_HEAD(&event->group_entry);
6149 INIT_LIST_HEAD(&event->event_entry);
6150 INIT_LIST_HEAD(&event->sibling_list);
6151 INIT_LIST_HEAD(&event->rb_entry);
6153 init_waitqueue_head(&event->waitq);
6154 init_irq_work(&event->pending, perf_pending_event);
6156 mutex_init(&event->mmap_mutex);
6158 atomic_long_set(&event->refcount, 1);
6160 event->attr = *attr;
6161 event->group_leader = group_leader;
6165 event->parent = parent_event;
6167 event->ns = get_pid_ns(task_active_pid_ns(current));
6168 event->id = atomic64_inc_return(&perf_event_id);
6170 event->state = PERF_EVENT_STATE_INACTIVE;
6173 event->attach_state = PERF_ATTACH_TASK;
6175 if (attr->type == PERF_TYPE_TRACEPOINT)
6176 event->hw.tp_target = task;
6177 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6179 * hw_breakpoint is a bit difficult here..
6181 else if (attr->type == PERF_TYPE_BREAKPOINT)
6182 event->hw.bp_target = task;
6186 if (!overflow_handler && parent_event) {
6187 overflow_handler = parent_event->overflow_handler;
6188 context = parent_event->overflow_handler_context;
6191 event->overflow_handler = overflow_handler;
6192 event->overflow_handler_context = context;
6194 perf_event__state_init(event);
6199 hwc->sample_period = attr->sample_period;
6200 if (attr->freq && attr->sample_freq)
6201 hwc->sample_period = 1;
6202 hwc->last_period = hwc->sample_period;
6204 local64_set(&hwc->period_left, hwc->sample_period);
6207 * we currently do not support PERF_FORMAT_GROUP on inherited events
6209 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6212 pmu = perf_init_event(event);
6218 else if (IS_ERR(pmu))
6223 put_pid_ns(event->ns);
6225 return ERR_PTR(err);
6228 if (!event->parent) {
6229 if (event->attach_state & PERF_ATTACH_TASK)
6230 static_key_slow_inc(&perf_sched_events.key);
6231 if (event->attr.mmap || event->attr.mmap_data)
6232 atomic_inc(&nr_mmap_events);
6233 if (event->attr.comm)
6234 atomic_inc(&nr_comm_events);
6235 if (event->attr.task)
6236 atomic_inc(&nr_task_events);
6237 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6238 err = get_callchain_buffers();
6241 return ERR_PTR(err);
6244 if (has_branch_stack(event)) {
6245 static_key_slow_inc(&perf_sched_events.key);
6246 if (!(event->attach_state & PERF_ATTACH_TASK))
6247 atomic_inc(&per_cpu(perf_branch_stack_events,
6255 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6256 struct perf_event_attr *attr)
6261 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6265 * zero the full structure, so that a short copy will be nice.
6267 memset(attr, 0, sizeof(*attr));
6269 ret = get_user(size, &uattr->size);
6273 if (size > PAGE_SIZE) /* silly large */
6276 if (!size) /* abi compat */
6277 size = PERF_ATTR_SIZE_VER0;
6279 if (size < PERF_ATTR_SIZE_VER0)
6283 * If we're handed a bigger struct than we know of,
6284 * ensure all the unknown bits are 0 - i.e. new
6285 * user-space does not rely on any kernel feature
6286 * extensions we dont know about yet.
6288 if (size > sizeof(*attr)) {
6289 unsigned char __user *addr;
6290 unsigned char __user *end;
6293 addr = (void __user *)uattr + sizeof(*attr);
6294 end = (void __user *)uattr + size;
6296 for (; addr < end; addr++) {
6297 ret = get_user(val, addr);
6303 size = sizeof(*attr);
6306 ret = copy_from_user(attr, uattr, size);
6310 if (attr->__reserved_1)
6313 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6316 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6319 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6320 u64 mask = attr->branch_sample_type;
6322 /* only using defined bits */
6323 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6326 /* at least one branch bit must be set */
6327 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6330 /* kernel level capture: check permissions */
6331 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6332 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6335 /* propagate priv level, when not set for branch */
6336 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6338 /* exclude_kernel checked on syscall entry */
6339 if (!attr->exclude_kernel)
6340 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6342 if (!attr->exclude_user)
6343 mask |= PERF_SAMPLE_BRANCH_USER;
6345 if (!attr->exclude_hv)
6346 mask |= PERF_SAMPLE_BRANCH_HV;
6348 * adjust user setting (for HW filter setup)
6350 attr->branch_sample_type = mask;
6354 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6355 ret = perf_reg_validate(attr->sample_regs_user);
6360 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6361 if (!arch_perf_have_user_stack_dump())
6365 * We have __u32 type for the size, but so far
6366 * we can only use __u16 as maximum due to the
6367 * __u16 sample size limit.
6369 if (attr->sample_stack_user >= USHRT_MAX)
6371 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6379 put_user(sizeof(*attr), &uattr->size);
6385 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6387 struct ring_buffer *rb = NULL, *old_rb = NULL;
6393 /* don't allow circular references */
6394 if (event == output_event)
6398 * Don't allow cross-cpu buffers
6400 if (output_event->cpu != event->cpu)
6404 * If its not a per-cpu rb, it must be the same task.
6406 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6410 mutex_lock(&event->mmap_mutex);
6411 /* Can't redirect output if we've got an active mmap() */
6412 if (atomic_read(&event->mmap_count))
6416 /* get the rb we want to redirect to */
6417 rb = ring_buffer_get(output_event);
6423 rcu_assign_pointer(event->rb, rb);
6425 ring_buffer_detach(event, old_rb);
6428 mutex_unlock(&event->mmap_mutex);
6431 ring_buffer_put(old_rb);
6437 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6439 * @attr_uptr: event_id type attributes for monitoring/sampling
6442 * @group_fd: group leader event fd
6444 SYSCALL_DEFINE5(perf_event_open,
6445 struct perf_event_attr __user *, attr_uptr,
6446 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6448 struct perf_event *group_leader = NULL, *output_event = NULL;
6449 struct perf_event *event, *sibling;
6450 struct perf_event_attr attr;
6451 struct perf_event_context *ctx;
6452 struct file *event_file = NULL;
6453 struct fd group = {NULL, 0};
6454 struct task_struct *task = NULL;
6460 /* for future expandability... */
6461 if (flags & ~PERF_FLAG_ALL)
6464 err = perf_copy_attr(attr_uptr, &attr);
6468 if (!attr.exclude_kernel) {
6469 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6474 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6479 * In cgroup mode, the pid argument is used to pass the fd
6480 * opened to the cgroup directory in cgroupfs. The cpu argument
6481 * designates the cpu on which to monitor threads from that
6484 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6487 event_fd = get_unused_fd();
6491 if (group_fd != -1) {
6492 err = perf_fget_light(group_fd, &group);
6495 group_leader = group.file->private_data;
6496 if (flags & PERF_FLAG_FD_OUTPUT)
6497 output_event = group_leader;
6498 if (flags & PERF_FLAG_FD_NO_GROUP)
6499 group_leader = NULL;
6502 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6503 task = find_lively_task_by_vpid(pid);
6505 err = PTR_ERR(task);
6512 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6514 if (IS_ERR(event)) {
6515 err = PTR_ERR(event);
6519 if (flags & PERF_FLAG_PID_CGROUP) {
6520 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6525 * - that has cgroup constraint on event->cpu
6526 * - that may need work on context switch
6528 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6529 static_key_slow_inc(&perf_sched_events.key);
6533 * Special case software events and allow them to be part of
6534 * any hardware group.
6539 (is_software_event(event) != is_software_event(group_leader))) {
6540 if (is_software_event(event)) {
6542 * If event and group_leader are not both a software
6543 * event, and event is, then group leader is not.
6545 * Allow the addition of software events to !software
6546 * groups, this is safe because software events never
6549 pmu = group_leader->pmu;
6550 } else if (is_software_event(group_leader) &&
6551 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6553 * In case the group is a pure software group, and we
6554 * try to add a hardware event, move the whole group to
6555 * the hardware context.
6562 * Get the target context (task or percpu):
6564 ctx = find_get_context(pmu, task, event->cpu);
6571 put_task_struct(task);
6576 * Look up the group leader (we will attach this event to it):
6582 * Do not allow a recursive hierarchy (this new sibling
6583 * becoming part of another group-sibling):
6585 if (group_leader->group_leader != group_leader)
6588 * Do not allow to attach to a group in a different
6589 * task or CPU context:
6592 if (group_leader->ctx->type != ctx->type)
6595 if (group_leader->ctx != ctx)
6600 * Only a group leader can be exclusive or pinned
6602 if (attr.exclusive || attr.pinned)
6607 err = perf_event_set_output(event, output_event);
6612 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6613 if (IS_ERR(event_file)) {
6614 err = PTR_ERR(event_file);
6619 struct perf_event_context *gctx = group_leader->ctx;
6621 mutex_lock(&gctx->mutex);
6622 perf_remove_from_context(group_leader);
6625 * Removing from the context ends up with disabled
6626 * event. What we want here is event in the initial
6627 * startup state, ready to be add into new context.
6629 perf_event__state_init(group_leader);
6630 list_for_each_entry(sibling, &group_leader->sibling_list,
6632 perf_remove_from_context(sibling);
6633 perf_event__state_init(sibling);
6636 mutex_unlock(&gctx->mutex);
6640 WARN_ON_ONCE(ctx->parent_ctx);
6641 mutex_lock(&ctx->mutex);
6645 perf_install_in_context(ctx, group_leader, event->cpu);
6647 list_for_each_entry(sibling, &group_leader->sibling_list,
6649 perf_install_in_context(ctx, sibling, event->cpu);
6654 perf_install_in_context(ctx, event, event->cpu);
6656 perf_unpin_context(ctx);
6657 mutex_unlock(&ctx->mutex);
6661 event->owner = current;
6663 mutex_lock(¤t->perf_event_mutex);
6664 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6665 mutex_unlock(¤t->perf_event_mutex);
6668 * Precalculate sample_data sizes
6670 perf_event__header_size(event);
6671 perf_event__id_header_size(event);
6674 * Drop the reference on the group_event after placing the
6675 * new event on the sibling_list. This ensures destruction
6676 * of the group leader will find the pointer to itself in
6677 * perf_group_detach().
6680 fd_install(event_fd, event_file);
6684 perf_unpin_context(ctx);
6691 put_task_struct(task);
6695 put_unused_fd(event_fd);
6700 * perf_event_create_kernel_counter
6702 * @attr: attributes of the counter to create
6703 * @cpu: cpu in which the counter is bound
6704 * @task: task to profile (NULL for percpu)
6707 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6708 struct task_struct *task,
6709 perf_overflow_handler_t overflow_handler,
6712 struct perf_event_context *ctx;
6713 struct perf_event *event;
6717 * Get the target context (task or percpu):
6720 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6721 overflow_handler, context);
6722 if (IS_ERR(event)) {
6723 err = PTR_ERR(event);
6727 ctx = find_get_context(event->pmu, task, cpu);
6733 WARN_ON_ONCE(ctx->parent_ctx);
6734 mutex_lock(&ctx->mutex);
6735 perf_install_in_context(ctx, event, cpu);
6737 perf_unpin_context(ctx);
6738 mutex_unlock(&ctx->mutex);
6745 return ERR_PTR(err);
6747 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6749 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6751 struct perf_event_context *src_ctx;
6752 struct perf_event_context *dst_ctx;
6753 struct perf_event *event, *tmp;
6756 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6757 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6759 mutex_lock(&src_ctx->mutex);
6760 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6762 perf_remove_from_context(event);
6764 list_add(&event->event_entry, &events);
6766 mutex_unlock(&src_ctx->mutex);
6770 mutex_lock(&dst_ctx->mutex);
6771 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6772 list_del(&event->event_entry);
6773 if (event->state >= PERF_EVENT_STATE_OFF)
6774 event->state = PERF_EVENT_STATE_INACTIVE;
6775 perf_install_in_context(dst_ctx, event, dst_cpu);
6778 mutex_unlock(&dst_ctx->mutex);
6780 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6782 static void sync_child_event(struct perf_event *child_event,
6783 struct task_struct *child)
6785 struct perf_event *parent_event = child_event->parent;
6788 if (child_event->attr.inherit_stat)
6789 perf_event_read_event(child_event, child);
6791 child_val = perf_event_count(child_event);
6794 * Add back the child's count to the parent's count:
6796 atomic64_add(child_val, &parent_event->child_count);
6797 atomic64_add(child_event->total_time_enabled,
6798 &parent_event->child_total_time_enabled);
6799 atomic64_add(child_event->total_time_running,
6800 &parent_event->child_total_time_running);
6803 * Remove this event from the parent's list
6805 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6806 mutex_lock(&parent_event->child_mutex);
6807 list_del_init(&child_event->child_list);
6808 mutex_unlock(&parent_event->child_mutex);
6811 * Release the parent event, if this was the last
6814 put_event(parent_event);
6818 __perf_event_exit_task(struct perf_event *child_event,
6819 struct perf_event_context *child_ctx,
6820 struct task_struct *child)
6822 if (child_event->parent) {
6823 raw_spin_lock_irq(&child_ctx->lock);
6824 perf_group_detach(child_event);
6825 raw_spin_unlock_irq(&child_ctx->lock);
6828 perf_remove_from_context(child_event);
6831 * It can happen that the parent exits first, and has events
6832 * that are still around due to the child reference. These
6833 * events need to be zapped.
6835 if (child_event->parent) {
6836 sync_child_event(child_event, child);
6837 free_event(child_event);
6841 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6843 struct perf_event *child_event, *tmp;
6844 struct perf_event_context *child_ctx;
6845 unsigned long flags;
6847 if (likely(!child->perf_event_ctxp[ctxn])) {
6848 perf_event_task(child, NULL, 0);
6852 local_irq_save(flags);
6854 * We can't reschedule here because interrupts are disabled,
6855 * and either child is current or it is a task that can't be
6856 * scheduled, so we are now safe from rescheduling changing
6859 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6862 * Take the context lock here so that if find_get_context is
6863 * reading child->perf_event_ctxp, we wait until it has
6864 * incremented the context's refcount before we do put_ctx below.
6866 raw_spin_lock(&child_ctx->lock);
6867 task_ctx_sched_out(child_ctx);
6868 child->perf_event_ctxp[ctxn] = NULL;
6870 * If this context is a clone; unclone it so it can't get
6871 * swapped to another process while we're removing all
6872 * the events from it.
6874 unclone_ctx(child_ctx);
6875 update_context_time(child_ctx);
6876 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6879 * Report the task dead after unscheduling the events so that we
6880 * won't get any samples after PERF_RECORD_EXIT. We can however still
6881 * get a few PERF_RECORD_READ events.
6883 perf_event_task(child, child_ctx, 0);
6886 * We can recurse on the same lock type through:
6888 * __perf_event_exit_task()
6889 * sync_child_event()
6891 * mutex_lock(&ctx->mutex)
6893 * But since its the parent context it won't be the same instance.
6895 mutex_lock(&child_ctx->mutex);
6898 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6900 __perf_event_exit_task(child_event, child_ctx, child);
6902 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6904 __perf_event_exit_task(child_event, child_ctx, child);
6907 * If the last event was a group event, it will have appended all
6908 * its siblings to the list, but we obtained 'tmp' before that which
6909 * will still point to the list head terminating the iteration.
6911 if (!list_empty(&child_ctx->pinned_groups) ||
6912 !list_empty(&child_ctx->flexible_groups))
6915 mutex_unlock(&child_ctx->mutex);
6921 * When a child task exits, feed back event values to parent events.
6923 void perf_event_exit_task(struct task_struct *child)
6925 struct perf_event *event, *tmp;
6928 mutex_lock(&child->perf_event_mutex);
6929 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6931 list_del_init(&event->owner_entry);
6934 * Ensure the list deletion is visible before we clear
6935 * the owner, closes a race against perf_release() where
6936 * we need to serialize on the owner->perf_event_mutex.
6939 event->owner = NULL;
6941 mutex_unlock(&child->perf_event_mutex);
6943 for_each_task_context_nr(ctxn)
6944 perf_event_exit_task_context(child, ctxn);
6947 static void perf_free_event(struct perf_event *event,
6948 struct perf_event_context *ctx)
6950 struct perf_event *parent = event->parent;
6952 if (WARN_ON_ONCE(!parent))
6955 mutex_lock(&parent->child_mutex);
6956 list_del_init(&event->child_list);
6957 mutex_unlock(&parent->child_mutex);
6961 perf_group_detach(event);
6962 list_del_event(event, ctx);
6967 * free an unexposed, unused context as created by inheritance by
6968 * perf_event_init_task below, used by fork() in case of fail.
6970 void perf_event_free_task(struct task_struct *task)
6972 struct perf_event_context *ctx;
6973 struct perf_event *event, *tmp;
6976 for_each_task_context_nr(ctxn) {
6977 ctx = task->perf_event_ctxp[ctxn];
6981 mutex_lock(&ctx->mutex);
6983 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6985 perf_free_event(event, ctx);
6987 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6989 perf_free_event(event, ctx);
6991 if (!list_empty(&ctx->pinned_groups) ||
6992 !list_empty(&ctx->flexible_groups))
6995 mutex_unlock(&ctx->mutex);
7001 void perf_event_delayed_put(struct task_struct *task)
7005 for_each_task_context_nr(ctxn)
7006 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7010 * inherit a event from parent task to child task:
7012 static struct perf_event *
7013 inherit_event(struct perf_event *parent_event,
7014 struct task_struct *parent,
7015 struct perf_event_context *parent_ctx,
7016 struct task_struct *child,
7017 struct perf_event *group_leader,
7018 struct perf_event_context *child_ctx)
7020 struct perf_event *child_event;
7021 unsigned long flags;
7024 * Instead of creating recursive hierarchies of events,
7025 * we link inherited events back to the original parent,
7026 * which has a filp for sure, which we use as the reference
7029 if (parent_event->parent)
7030 parent_event = parent_event->parent;
7032 child_event = perf_event_alloc(&parent_event->attr,
7035 group_leader, parent_event,
7037 if (IS_ERR(child_event))
7040 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7041 free_event(child_event);
7048 * Make the child state follow the state of the parent event,
7049 * not its attr.disabled bit. We hold the parent's mutex,
7050 * so we won't race with perf_event_{en, dis}able_family.
7052 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7053 child_event->state = PERF_EVENT_STATE_INACTIVE;
7055 child_event->state = PERF_EVENT_STATE_OFF;
7057 if (parent_event->attr.freq) {
7058 u64 sample_period = parent_event->hw.sample_period;
7059 struct hw_perf_event *hwc = &child_event->hw;
7061 hwc->sample_period = sample_period;
7062 hwc->last_period = sample_period;
7064 local64_set(&hwc->period_left, sample_period);
7067 child_event->ctx = child_ctx;
7068 child_event->overflow_handler = parent_event->overflow_handler;
7069 child_event->overflow_handler_context
7070 = parent_event->overflow_handler_context;
7073 * Precalculate sample_data sizes
7075 perf_event__header_size(child_event);
7076 perf_event__id_header_size(child_event);
7079 * Link it up in the child's context:
7081 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7082 add_event_to_ctx(child_event, child_ctx);
7083 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7086 * Link this into the parent event's child list
7088 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7089 mutex_lock(&parent_event->child_mutex);
7090 list_add_tail(&child_event->child_list, &parent_event->child_list);
7091 mutex_unlock(&parent_event->child_mutex);
7096 static int inherit_group(struct perf_event *parent_event,
7097 struct task_struct *parent,
7098 struct perf_event_context *parent_ctx,
7099 struct task_struct *child,
7100 struct perf_event_context *child_ctx)
7102 struct perf_event *leader;
7103 struct perf_event *sub;
7104 struct perf_event *child_ctr;
7106 leader = inherit_event(parent_event, parent, parent_ctx,
7107 child, NULL, child_ctx);
7109 return PTR_ERR(leader);
7110 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7111 child_ctr = inherit_event(sub, parent, parent_ctx,
7112 child, leader, child_ctx);
7113 if (IS_ERR(child_ctr))
7114 return PTR_ERR(child_ctr);
7120 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7121 struct perf_event_context *parent_ctx,
7122 struct task_struct *child, int ctxn,
7126 struct perf_event_context *child_ctx;
7128 if (!event->attr.inherit) {
7133 child_ctx = child->perf_event_ctxp[ctxn];
7136 * This is executed from the parent task context, so
7137 * inherit events that have been marked for cloning.
7138 * First allocate and initialize a context for the
7142 child_ctx = alloc_perf_context(event->pmu, child);
7146 child->perf_event_ctxp[ctxn] = child_ctx;
7149 ret = inherit_group(event, parent, parent_ctx,
7159 * Initialize the perf_event context in task_struct
7161 int perf_event_init_context(struct task_struct *child, int ctxn)
7163 struct perf_event_context *child_ctx, *parent_ctx;
7164 struct perf_event_context *cloned_ctx;
7165 struct perf_event *event;
7166 struct task_struct *parent = current;
7167 int inherited_all = 1;
7168 unsigned long flags;
7171 if (likely(!parent->perf_event_ctxp[ctxn]))
7175 * If the parent's context is a clone, pin it so it won't get
7178 parent_ctx = perf_pin_task_context(parent, ctxn);
7181 * No need to check if parent_ctx != NULL here; since we saw
7182 * it non-NULL earlier, the only reason for it to become NULL
7183 * is if we exit, and since we're currently in the middle of
7184 * a fork we can't be exiting at the same time.
7188 * Lock the parent list. No need to lock the child - not PID
7189 * hashed yet and not running, so nobody can access it.
7191 mutex_lock(&parent_ctx->mutex);
7194 * We dont have to disable NMIs - we are only looking at
7195 * the list, not manipulating it:
7197 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7198 ret = inherit_task_group(event, parent, parent_ctx,
7199 child, ctxn, &inherited_all);
7205 * We can't hold ctx->lock when iterating the ->flexible_group list due
7206 * to allocations, but we need to prevent rotation because
7207 * rotate_ctx() will change the list from interrupt context.
7209 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7210 parent_ctx->rotate_disable = 1;
7211 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7213 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7214 ret = inherit_task_group(event, parent, parent_ctx,
7215 child, ctxn, &inherited_all);
7220 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7221 parent_ctx->rotate_disable = 0;
7223 child_ctx = child->perf_event_ctxp[ctxn];
7225 if (child_ctx && inherited_all) {
7227 * Mark the child context as a clone of the parent
7228 * context, or of whatever the parent is a clone of.
7230 * Note that if the parent is a clone, the holding of
7231 * parent_ctx->lock avoids it from being uncloned.
7233 cloned_ctx = parent_ctx->parent_ctx;
7235 child_ctx->parent_ctx = cloned_ctx;
7236 child_ctx->parent_gen = parent_ctx->parent_gen;
7238 child_ctx->parent_ctx = parent_ctx;
7239 child_ctx->parent_gen = parent_ctx->generation;
7241 get_ctx(child_ctx->parent_ctx);
7244 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7245 mutex_unlock(&parent_ctx->mutex);
7247 perf_unpin_context(parent_ctx);
7248 put_ctx(parent_ctx);
7254 * Initialize the perf_event context in task_struct
7256 int perf_event_init_task(struct task_struct *child)
7260 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7261 mutex_init(&child->perf_event_mutex);
7262 INIT_LIST_HEAD(&child->perf_event_list);
7264 for_each_task_context_nr(ctxn) {
7265 ret = perf_event_init_context(child, ctxn);
7273 static void __init perf_event_init_all_cpus(void)
7275 struct swevent_htable *swhash;
7278 for_each_possible_cpu(cpu) {
7279 swhash = &per_cpu(swevent_htable, cpu);
7280 mutex_init(&swhash->hlist_mutex);
7281 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7285 static void __cpuinit perf_event_init_cpu(int cpu)
7287 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7289 mutex_lock(&swhash->hlist_mutex);
7290 if (swhash->hlist_refcount > 0) {
7291 struct swevent_hlist *hlist;
7293 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7295 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7297 mutex_unlock(&swhash->hlist_mutex);
7300 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7301 static void perf_pmu_rotate_stop(struct pmu *pmu)
7303 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7305 WARN_ON(!irqs_disabled());
7307 list_del_init(&cpuctx->rotation_list);
7310 static void __perf_event_exit_context(void *__info)
7312 struct perf_event_context *ctx = __info;
7313 struct perf_event *event, *tmp;
7315 perf_pmu_rotate_stop(ctx->pmu);
7317 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7318 __perf_remove_from_context(event);
7319 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7320 __perf_remove_from_context(event);
7323 static void perf_event_exit_cpu_context(int cpu)
7325 struct perf_event_context *ctx;
7329 idx = srcu_read_lock(&pmus_srcu);
7330 list_for_each_entry_rcu(pmu, &pmus, entry) {
7331 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7333 mutex_lock(&ctx->mutex);
7334 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7335 mutex_unlock(&ctx->mutex);
7337 srcu_read_unlock(&pmus_srcu, idx);
7340 static void perf_event_exit_cpu(int cpu)
7342 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7344 mutex_lock(&swhash->hlist_mutex);
7345 swevent_hlist_release(swhash);
7346 mutex_unlock(&swhash->hlist_mutex);
7348 perf_event_exit_cpu_context(cpu);
7351 static inline void perf_event_exit_cpu(int cpu) { }
7355 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7359 for_each_online_cpu(cpu)
7360 perf_event_exit_cpu(cpu);
7366 * Run the perf reboot notifier at the very last possible moment so that
7367 * the generic watchdog code runs as long as possible.
7369 static struct notifier_block perf_reboot_notifier = {
7370 .notifier_call = perf_reboot,
7371 .priority = INT_MIN,
7374 static int __cpuinit
7375 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7377 unsigned int cpu = (long)hcpu;
7379 switch (action & ~CPU_TASKS_FROZEN) {
7381 case CPU_UP_PREPARE:
7382 case CPU_DOWN_FAILED:
7383 perf_event_init_cpu(cpu);
7386 case CPU_UP_CANCELED:
7387 case CPU_DOWN_PREPARE:
7388 perf_event_exit_cpu(cpu);
7398 void __init perf_event_init(void)
7404 perf_event_init_all_cpus();
7405 init_srcu_struct(&pmus_srcu);
7406 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7407 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7408 perf_pmu_register(&perf_task_clock, NULL, -1);
7410 perf_cpu_notifier(perf_cpu_notify);
7411 register_reboot_notifier(&perf_reboot_notifier);
7413 ret = init_hw_breakpoint();
7414 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7416 /* do not patch jump label more than once per second */
7417 jump_label_rate_limit(&perf_sched_events, HZ);
7420 * Build time assertion that we keep the data_head at the intended
7421 * location. IOW, validation we got the __reserved[] size right.
7423 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7427 static int __init perf_event_sysfs_init(void)
7432 mutex_lock(&pmus_lock);
7434 ret = bus_register(&pmu_bus);
7438 list_for_each_entry(pmu, &pmus, entry) {
7439 if (!pmu->name || pmu->type < 0)
7442 ret = pmu_dev_alloc(pmu);
7443 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7445 pmu_bus_running = 1;
7449 mutex_unlock(&pmus_lock);
7453 device_initcall(perf_event_sysfs_init);
7455 #ifdef CONFIG_CGROUP_PERF
7456 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7458 struct perf_cgroup *jc;
7460 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7462 return ERR_PTR(-ENOMEM);
7464 jc->info = alloc_percpu(struct perf_cgroup_info);
7467 return ERR_PTR(-ENOMEM);
7473 static void perf_cgroup_css_free(struct cgroup *cont)
7475 struct perf_cgroup *jc;
7476 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7477 struct perf_cgroup, css);
7478 free_percpu(jc->info);
7482 static int __perf_cgroup_move(void *info)
7484 struct task_struct *task = info;
7485 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7489 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7491 struct task_struct *task;
7493 cgroup_taskset_for_each(task, cgrp, tset)
7494 task_function_call(task, __perf_cgroup_move, task);
7497 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7498 struct task_struct *task)
7501 * cgroup_exit() is called in the copy_process() failure path.
7502 * Ignore this case since the task hasn't ran yet, this avoids
7503 * trying to poke a half freed task state from generic code.
7505 if (!(task->flags & PF_EXITING))
7508 task_function_call(task, __perf_cgroup_move, task);
7511 struct cgroup_subsys perf_subsys = {
7512 .name = "perf_event",
7513 .subsys_id = perf_subsys_id,
7514 .css_alloc = perf_cgroup_css_alloc,
7515 .css_free = perf_cgroup_css_free,
7516 .exit = perf_cgroup_exit,
7517 .attach = perf_cgroup_attach,
7519 #endif /* CONFIG_CGROUP_PERF */