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/sysfs.h>
22 #include <linux/dcache.h>
23 #include <linux/percpu.h>
24 #include <linux/ptrace.h>
25 #include <linux/reboot.h>
26 #include <linux/vmstat.h>
27 #include <linux/device.h>
28 #include <linux/export.h>
29 #include <linux/vmalloc.h>
30 #include <linux/hardirq.h>
31 #include <linux/rculist.h>
32 #include <linux/uaccess.h>
33 #include <linux/syscalls.h>
34 #include <linux/anon_inodes.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/perf_event.h>
37 #include <linux/ftrace_event.h>
38 #include <linux/hw_breakpoint.h>
39 #include <linux/mm_types.h>
43 #include <asm/irq_regs.h>
45 struct remote_function_call {
46 struct task_struct *p;
47 int (*func)(void *info);
52 static void remote_function(void *data)
54 struct remote_function_call *tfc = data;
55 struct task_struct *p = tfc->p;
59 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
63 tfc->ret = tfc->func(tfc->info);
67 * task_function_call - call a function on the cpu on which a task runs
68 * @p: the task to evaluate
69 * @func: the function to be called
70 * @info: the function call argument
72 * Calls the function @func when the task is currently running. This might
73 * be on the current CPU, which just calls the function directly
75 * returns: @func return value, or
76 * -ESRCH - when the process isn't running
77 * -EAGAIN - when the process moved away
80 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
82 struct remote_function_call data = {
86 .ret = -ESRCH, /* No such (running) process */
90 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
96 * cpu_function_call - call a function on the cpu
97 * @func: the function to be called
98 * @info: the function call argument
100 * Calls the function @func on the remote cpu.
102 * returns: @func return value or -ENXIO when the cpu is offline
104 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
106 struct remote_function_call data = {
110 .ret = -ENXIO, /* No such CPU */
113 smp_call_function_single(cpu, remote_function, &data, 1);
118 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
119 PERF_FLAG_FD_OUTPUT |\
120 PERF_FLAG_PID_CGROUP)
123 * branch priv levels that need permission checks
125 #define PERF_SAMPLE_BRANCH_PERM_PLM \
126 (PERF_SAMPLE_BRANCH_KERNEL |\
127 PERF_SAMPLE_BRANCH_HV)
130 EVENT_FLEXIBLE = 0x1,
132 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
136 * perf_sched_events : >0 events exist
137 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
139 struct static_key_deferred perf_sched_events __read_mostly;
140 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
141 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
143 static atomic_t nr_mmap_events __read_mostly;
144 static atomic_t nr_comm_events __read_mostly;
145 static atomic_t nr_task_events __read_mostly;
147 static LIST_HEAD(pmus);
148 static DEFINE_MUTEX(pmus_lock);
149 static struct srcu_struct pmus_srcu;
152 * perf event paranoia level:
153 * -1 - not paranoid at all
154 * 0 - disallow raw tracepoint access for unpriv
155 * 1 - disallow cpu events for unpriv
156 * 2 - disallow kernel profiling for unpriv
158 int sysctl_perf_event_paranoid __read_mostly = 1;
160 /* Minimum for 512 kiB + 1 user control page */
161 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
164 * max perf event sample rate
166 #define DEFAULT_MAX_SAMPLE_RATE 100000
167 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
168 static int max_samples_per_tick __read_mostly =
169 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
171 int perf_proc_update_handler(struct ctl_table *table, int write,
172 void __user *buffer, size_t *lenp,
175 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
180 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
185 static atomic64_t perf_event_id;
187 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
188 enum event_type_t event_type);
190 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
191 enum event_type_t event_type,
192 struct task_struct *task);
194 static void update_context_time(struct perf_event_context *ctx);
195 static u64 perf_event_time(struct perf_event *event);
197 static void ring_buffer_attach(struct perf_event *event,
198 struct ring_buffer *rb);
200 void __weak perf_event_print_debug(void) { }
202 extern __weak const char *perf_pmu_name(void)
207 static inline u64 perf_clock(void)
209 return local_clock();
212 static inline struct perf_cpu_context *
213 __get_cpu_context(struct perf_event_context *ctx)
215 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
218 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
219 struct perf_event_context *ctx)
221 raw_spin_lock(&cpuctx->ctx.lock);
223 raw_spin_lock(&ctx->lock);
226 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
227 struct perf_event_context *ctx)
230 raw_spin_unlock(&ctx->lock);
231 raw_spin_unlock(&cpuctx->ctx.lock);
234 #ifdef CONFIG_CGROUP_PERF
237 * Must ensure cgroup is pinned (css_get) before calling
238 * this function. In other words, we cannot call this function
239 * if there is no cgroup event for the current CPU context.
241 static inline struct perf_cgroup *
242 perf_cgroup_from_task(struct task_struct *task)
244 return container_of(task_subsys_state(task, perf_subsys_id),
245 struct perf_cgroup, css);
249 perf_cgroup_match(struct perf_event *event)
251 struct perf_event_context *ctx = event->ctx;
252 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
254 /* @event doesn't care about cgroup */
258 /* wants specific cgroup scope but @cpuctx isn't associated with any */
263 * Cgroup scoping is recursive. An event enabled for a cgroup is
264 * also enabled for all its descendant cgroups. If @cpuctx's
265 * cgroup is a descendant of @event's (the test covers identity
266 * case), it's a match.
268 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
269 event->cgrp->css.cgroup);
272 static inline bool perf_tryget_cgroup(struct perf_event *event)
274 return css_tryget(&event->cgrp->css);
277 static inline void perf_put_cgroup(struct perf_event *event)
279 css_put(&event->cgrp->css);
282 static inline void perf_detach_cgroup(struct perf_event *event)
284 perf_put_cgroup(event);
288 static inline int is_cgroup_event(struct perf_event *event)
290 return event->cgrp != NULL;
293 static inline u64 perf_cgroup_event_time(struct perf_event *event)
295 struct perf_cgroup_info *t;
297 t = per_cpu_ptr(event->cgrp->info, event->cpu);
301 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
303 struct perf_cgroup_info *info;
308 info = this_cpu_ptr(cgrp->info);
310 info->time += now - info->timestamp;
311 info->timestamp = now;
314 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
316 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
318 __update_cgrp_time(cgrp_out);
321 static inline void update_cgrp_time_from_event(struct perf_event *event)
323 struct perf_cgroup *cgrp;
326 * ensure we access cgroup data only when needed and
327 * when we know the cgroup is pinned (css_get)
329 if (!is_cgroup_event(event))
332 cgrp = perf_cgroup_from_task(current);
334 * Do not update time when cgroup is not active
336 if (cgrp == event->cgrp)
337 __update_cgrp_time(event->cgrp);
341 perf_cgroup_set_timestamp(struct task_struct *task,
342 struct perf_event_context *ctx)
344 struct perf_cgroup *cgrp;
345 struct perf_cgroup_info *info;
348 * ctx->lock held by caller
349 * ensure we do not access cgroup data
350 * unless we have the cgroup pinned (css_get)
352 if (!task || !ctx->nr_cgroups)
355 cgrp = perf_cgroup_from_task(task);
356 info = this_cpu_ptr(cgrp->info);
357 info->timestamp = ctx->timestamp;
360 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
361 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
364 * reschedule events based on the cgroup constraint of task.
366 * mode SWOUT : schedule out everything
367 * mode SWIN : schedule in based on cgroup for next
369 void perf_cgroup_switch(struct task_struct *task, int mode)
371 struct perf_cpu_context *cpuctx;
376 * disable interrupts to avoid geting nr_cgroup
377 * changes via __perf_event_disable(). Also
380 local_irq_save(flags);
383 * we reschedule only in the presence of cgroup
384 * constrained events.
388 list_for_each_entry_rcu(pmu, &pmus, entry) {
389 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
390 if (cpuctx->unique_pmu != pmu)
391 continue; /* ensure we process each cpuctx once */
394 * perf_cgroup_events says at least one
395 * context on this CPU has cgroup events.
397 * ctx->nr_cgroups reports the number of cgroup
398 * events for a context.
400 if (cpuctx->ctx.nr_cgroups > 0) {
401 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
402 perf_pmu_disable(cpuctx->ctx.pmu);
404 if (mode & PERF_CGROUP_SWOUT) {
405 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
407 * must not be done before ctxswout due
408 * to event_filter_match() in event_sched_out()
413 if (mode & PERF_CGROUP_SWIN) {
414 WARN_ON_ONCE(cpuctx->cgrp);
416 * set cgrp before ctxsw in to allow
417 * event_filter_match() to not have to pass
420 cpuctx->cgrp = perf_cgroup_from_task(task);
421 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
423 perf_pmu_enable(cpuctx->ctx.pmu);
424 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
430 local_irq_restore(flags);
433 static inline void perf_cgroup_sched_out(struct task_struct *task,
434 struct task_struct *next)
436 struct perf_cgroup *cgrp1;
437 struct perf_cgroup *cgrp2 = NULL;
440 * we come here when we know perf_cgroup_events > 0
442 cgrp1 = perf_cgroup_from_task(task);
445 * next is NULL when called from perf_event_enable_on_exec()
446 * that will systematically cause a cgroup_switch()
449 cgrp2 = perf_cgroup_from_task(next);
452 * only schedule out current cgroup events if we know
453 * that we are switching to a different cgroup. Otherwise,
454 * do no touch the cgroup events.
457 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
460 static inline void perf_cgroup_sched_in(struct task_struct *prev,
461 struct task_struct *task)
463 struct perf_cgroup *cgrp1;
464 struct perf_cgroup *cgrp2 = NULL;
467 * we come here when we know perf_cgroup_events > 0
469 cgrp1 = perf_cgroup_from_task(task);
471 /* prev can never be NULL */
472 cgrp2 = perf_cgroup_from_task(prev);
475 * only need to schedule in cgroup events if we are changing
476 * cgroup during ctxsw. Cgroup events were not scheduled
477 * out of ctxsw out if that was not the case.
480 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
483 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
484 struct perf_event_attr *attr,
485 struct perf_event *group_leader)
487 struct perf_cgroup *cgrp;
488 struct cgroup_subsys_state *css;
489 struct fd f = fdget(fd);
495 css = cgroup_css_from_dir(f.file, perf_subsys_id);
501 cgrp = container_of(css, struct perf_cgroup, css);
504 /* must be done before we fput() the file */
505 if (!perf_tryget_cgroup(event)) {
512 * all events in a group must monitor
513 * the same cgroup because a task belongs
514 * to only one perf cgroup at a time
516 if (group_leader && group_leader->cgrp != cgrp) {
517 perf_detach_cgroup(event);
526 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
528 struct perf_cgroup_info *t;
529 t = per_cpu_ptr(event->cgrp->info, event->cpu);
530 event->shadow_ctx_time = now - t->timestamp;
534 perf_cgroup_defer_enabled(struct perf_event *event)
537 * when the current task's perf cgroup does not match
538 * the event's, we need to remember to call the
539 * perf_mark_enable() function the first time a task with
540 * a matching perf cgroup is scheduled in.
542 if (is_cgroup_event(event) && !perf_cgroup_match(event))
543 event->cgrp_defer_enabled = 1;
547 perf_cgroup_mark_enabled(struct perf_event *event,
548 struct perf_event_context *ctx)
550 struct perf_event *sub;
551 u64 tstamp = perf_event_time(event);
553 if (!event->cgrp_defer_enabled)
556 event->cgrp_defer_enabled = 0;
558 event->tstamp_enabled = tstamp - event->total_time_enabled;
559 list_for_each_entry(sub, &event->sibling_list, group_entry) {
560 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
561 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
562 sub->cgrp_defer_enabled = 0;
566 #else /* !CONFIG_CGROUP_PERF */
569 perf_cgroup_match(struct perf_event *event)
574 static inline void perf_detach_cgroup(struct perf_event *event)
577 static inline int is_cgroup_event(struct perf_event *event)
582 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
587 static inline void update_cgrp_time_from_event(struct perf_event *event)
591 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
595 static inline void perf_cgroup_sched_out(struct task_struct *task,
596 struct task_struct *next)
600 static inline void perf_cgroup_sched_in(struct task_struct *prev,
601 struct task_struct *task)
605 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
606 struct perf_event_attr *attr,
607 struct perf_event *group_leader)
613 perf_cgroup_set_timestamp(struct task_struct *task,
614 struct perf_event_context *ctx)
619 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
624 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
628 static inline u64 perf_cgroup_event_time(struct perf_event *event)
634 perf_cgroup_defer_enabled(struct perf_event *event)
639 perf_cgroup_mark_enabled(struct perf_event *event,
640 struct perf_event_context *ctx)
645 void perf_pmu_disable(struct pmu *pmu)
647 int *count = this_cpu_ptr(pmu->pmu_disable_count);
649 pmu->pmu_disable(pmu);
652 void perf_pmu_enable(struct pmu *pmu)
654 int *count = this_cpu_ptr(pmu->pmu_disable_count);
656 pmu->pmu_enable(pmu);
659 static DEFINE_PER_CPU(struct list_head, rotation_list);
662 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
663 * because they're strictly cpu affine and rotate_start is called with IRQs
664 * disabled, while rotate_context is called from IRQ context.
666 static void perf_pmu_rotate_start(struct pmu *pmu)
668 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
669 struct list_head *head = &__get_cpu_var(rotation_list);
671 WARN_ON(!irqs_disabled());
673 if (list_empty(&cpuctx->rotation_list))
674 list_add(&cpuctx->rotation_list, head);
677 static void get_ctx(struct perf_event_context *ctx)
679 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
682 static void put_ctx(struct perf_event_context *ctx)
684 if (atomic_dec_and_test(&ctx->refcount)) {
686 put_ctx(ctx->parent_ctx);
688 put_task_struct(ctx->task);
689 kfree_rcu(ctx, rcu_head);
693 static void unclone_ctx(struct perf_event_context *ctx)
695 if (ctx->parent_ctx) {
696 put_ctx(ctx->parent_ctx);
697 ctx->parent_ctx = NULL;
701 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
704 * only top level events have the pid namespace they were created in
707 event = event->parent;
709 return task_tgid_nr_ns(p, event->ns);
712 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
715 * only top level events have the pid namespace they were created in
718 event = event->parent;
720 return task_pid_nr_ns(p, event->ns);
724 * If we inherit events we want to return the parent event id
727 static u64 primary_event_id(struct perf_event *event)
732 id = event->parent->id;
738 * Get the perf_event_context for a task and lock it.
739 * This has to cope with with the fact that until it is locked,
740 * the context could get moved to another task.
742 static struct perf_event_context *
743 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
745 struct perf_event_context *ctx;
749 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
752 * If this context is a clone of another, it might
753 * get swapped for another underneath us by
754 * perf_event_task_sched_out, though the
755 * rcu_read_lock() protects us from any context
756 * getting freed. Lock the context and check if it
757 * got swapped before we could get the lock, and retry
758 * if so. If we locked the right context, then it
759 * can't get swapped on us any more.
761 raw_spin_lock_irqsave(&ctx->lock, *flags);
762 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
763 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
767 if (!atomic_inc_not_zero(&ctx->refcount)) {
768 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
777 * Get the context for a task and increment its pin_count so it
778 * can't get swapped to another task. This also increments its
779 * reference count so that the context can't get freed.
781 static struct perf_event_context *
782 perf_pin_task_context(struct task_struct *task, int ctxn)
784 struct perf_event_context *ctx;
787 ctx = perf_lock_task_context(task, ctxn, &flags);
790 raw_spin_unlock_irqrestore(&ctx->lock, flags);
795 static void perf_unpin_context(struct perf_event_context *ctx)
799 raw_spin_lock_irqsave(&ctx->lock, flags);
801 raw_spin_unlock_irqrestore(&ctx->lock, flags);
805 * Update the record of the current time in a context.
807 static void update_context_time(struct perf_event_context *ctx)
809 u64 now = perf_clock();
811 ctx->time += now - ctx->timestamp;
812 ctx->timestamp = now;
815 static u64 perf_event_time(struct perf_event *event)
817 struct perf_event_context *ctx = event->ctx;
819 if (is_cgroup_event(event))
820 return perf_cgroup_event_time(event);
822 return ctx ? ctx->time : 0;
826 * Update the total_time_enabled and total_time_running fields for a event.
827 * The caller of this function needs to hold the ctx->lock.
829 static void update_event_times(struct perf_event *event)
831 struct perf_event_context *ctx = event->ctx;
834 if (event->state < PERF_EVENT_STATE_INACTIVE ||
835 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
838 * in cgroup mode, time_enabled represents
839 * the time the event was enabled AND active
840 * tasks were in the monitored cgroup. This is
841 * independent of the activity of the context as
842 * there may be a mix of cgroup and non-cgroup events.
844 * That is why we treat cgroup events differently
847 if (is_cgroup_event(event))
848 run_end = perf_cgroup_event_time(event);
849 else if (ctx->is_active)
852 run_end = event->tstamp_stopped;
854 event->total_time_enabled = run_end - event->tstamp_enabled;
856 if (event->state == PERF_EVENT_STATE_INACTIVE)
857 run_end = event->tstamp_stopped;
859 run_end = perf_event_time(event);
861 event->total_time_running = run_end - event->tstamp_running;
866 * Update total_time_enabled and total_time_running for all events in a group.
868 static void update_group_times(struct perf_event *leader)
870 struct perf_event *event;
872 update_event_times(leader);
873 list_for_each_entry(event, &leader->sibling_list, group_entry)
874 update_event_times(event);
877 static struct list_head *
878 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
880 if (event->attr.pinned)
881 return &ctx->pinned_groups;
883 return &ctx->flexible_groups;
887 * Add a event from the lists for its context.
888 * Must be called with ctx->mutex and ctx->lock held.
891 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
893 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
894 event->attach_state |= PERF_ATTACH_CONTEXT;
897 * If we're a stand alone event or group leader, we go to the context
898 * list, group events are kept attached to the group so that
899 * perf_group_detach can, at all times, locate all siblings.
901 if (event->group_leader == event) {
902 struct list_head *list;
904 if (is_software_event(event))
905 event->group_flags |= PERF_GROUP_SOFTWARE;
907 list = ctx_group_list(event, ctx);
908 list_add_tail(&event->group_entry, list);
911 if (is_cgroup_event(event))
914 if (has_branch_stack(event))
915 ctx->nr_branch_stack++;
917 list_add_rcu(&event->event_entry, &ctx->event_list);
919 perf_pmu_rotate_start(ctx->pmu);
921 if (event->attr.inherit_stat)
926 * Initialize event state based on the perf_event_attr::disabled.
928 static inline void perf_event__state_init(struct perf_event *event)
930 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
931 PERF_EVENT_STATE_INACTIVE;
935 * Called at perf_event creation and when events are attached/detached from a
938 static void perf_event__read_size(struct perf_event *event)
940 int entry = sizeof(u64); /* value */
944 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
947 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
950 if (event->attr.read_format & PERF_FORMAT_ID)
951 entry += sizeof(u64);
953 if (event->attr.read_format & PERF_FORMAT_GROUP) {
954 nr += event->group_leader->nr_siblings;
959 event->read_size = size;
962 static void perf_event__header_size(struct perf_event *event)
964 struct perf_sample_data *data;
965 u64 sample_type = event->attr.sample_type;
968 perf_event__read_size(event);
970 if (sample_type & PERF_SAMPLE_IP)
971 size += sizeof(data->ip);
973 if (sample_type & PERF_SAMPLE_ADDR)
974 size += sizeof(data->addr);
976 if (sample_type & PERF_SAMPLE_PERIOD)
977 size += sizeof(data->period);
979 if (sample_type & PERF_SAMPLE_READ)
980 size += event->read_size;
982 event->header_size = size;
985 static void perf_event__id_header_size(struct perf_event *event)
987 struct perf_sample_data *data;
988 u64 sample_type = event->attr.sample_type;
991 if (sample_type & PERF_SAMPLE_TID)
992 size += sizeof(data->tid_entry);
994 if (sample_type & PERF_SAMPLE_TIME)
995 size += sizeof(data->time);
997 if (sample_type & PERF_SAMPLE_ID)
998 size += sizeof(data->id);
1000 if (sample_type & PERF_SAMPLE_STREAM_ID)
1001 size += sizeof(data->stream_id);
1003 if (sample_type & PERF_SAMPLE_CPU)
1004 size += sizeof(data->cpu_entry);
1006 event->id_header_size = size;
1009 static void perf_group_attach(struct perf_event *event)
1011 struct perf_event *group_leader = event->group_leader, *pos;
1014 * We can have double attach due to group movement in perf_event_open.
1016 if (event->attach_state & PERF_ATTACH_GROUP)
1019 event->attach_state |= PERF_ATTACH_GROUP;
1021 if (group_leader == event)
1024 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1025 !is_software_event(event))
1026 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1028 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1029 group_leader->nr_siblings++;
1031 perf_event__header_size(group_leader);
1033 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1034 perf_event__header_size(pos);
1038 * Remove a event from the lists for its context.
1039 * Must be called with ctx->mutex and ctx->lock held.
1042 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1044 struct perf_cpu_context *cpuctx;
1046 * We can have double detach due to exit/hot-unplug + close.
1048 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1051 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1053 if (is_cgroup_event(event)) {
1055 cpuctx = __get_cpu_context(ctx);
1057 * if there are no more cgroup events
1058 * then cler cgrp to avoid stale pointer
1059 * in update_cgrp_time_from_cpuctx()
1061 if (!ctx->nr_cgroups)
1062 cpuctx->cgrp = NULL;
1065 if (has_branch_stack(event))
1066 ctx->nr_branch_stack--;
1069 if (event->attr.inherit_stat)
1072 list_del_rcu(&event->event_entry);
1074 if (event->group_leader == event)
1075 list_del_init(&event->group_entry);
1077 update_group_times(event);
1080 * If event was in error state, then keep it
1081 * that way, otherwise bogus counts will be
1082 * returned on read(). The only way to get out
1083 * of error state is by explicit re-enabling
1086 if (event->state > PERF_EVENT_STATE_OFF)
1087 event->state = PERF_EVENT_STATE_OFF;
1090 static void perf_group_detach(struct perf_event *event)
1092 struct perf_event *sibling, *tmp;
1093 struct list_head *list = NULL;
1096 * We can have double detach due to exit/hot-unplug + close.
1098 if (!(event->attach_state & PERF_ATTACH_GROUP))
1101 event->attach_state &= ~PERF_ATTACH_GROUP;
1104 * If this is a sibling, remove it from its group.
1106 if (event->group_leader != event) {
1107 list_del_init(&event->group_entry);
1108 event->group_leader->nr_siblings--;
1112 if (!list_empty(&event->group_entry))
1113 list = &event->group_entry;
1116 * If this was a group event with sibling events then
1117 * upgrade the siblings to singleton events by adding them
1118 * to whatever list we are on.
1120 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1122 list_move_tail(&sibling->group_entry, list);
1123 sibling->group_leader = sibling;
1125 /* Inherit group flags from the previous leader */
1126 sibling->group_flags = event->group_flags;
1130 perf_event__header_size(event->group_leader);
1132 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1133 perf_event__header_size(tmp);
1137 event_filter_match(struct perf_event *event)
1139 return (event->cpu == -1 || event->cpu == smp_processor_id())
1140 && perf_cgroup_match(event);
1144 event_sched_out(struct perf_event *event,
1145 struct perf_cpu_context *cpuctx,
1146 struct perf_event_context *ctx)
1148 u64 tstamp = perf_event_time(event);
1151 * An event which could not be activated because of
1152 * filter mismatch still needs to have its timings
1153 * maintained, otherwise bogus information is return
1154 * via read() for time_enabled, time_running:
1156 if (event->state == PERF_EVENT_STATE_INACTIVE
1157 && !event_filter_match(event)) {
1158 delta = tstamp - event->tstamp_stopped;
1159 event->tstamp_running += delta;
1160 event->tstamp_stopped = tstamp;
1163 if (event->state != PERF_EVENT_STATE_ACTIVE)
1166 event->state = PERF_EVENT_STATE_INACTIVE;
1167 if (event->pending_disable) {
1168 event->pending_disable = 0;
1169 event->state = PERF_EVENT_STATE_OFF;
1171 event->tstamp_stopped = tstamp;
1172 event->pmu->del(event, 0);
1175 if (!is_software_event(event))
1176 cpuctx->active_oncpu--;
1178 if (event->attr.freq && event->attr.sample_freq)
1180 if (event->attr.exclusive || !cpuctx->active_oncpu)
1181 cpuctx->exclusive = 0;
1185 group_sched_out(struct perf_event *group_event,
1186 struct perf_cpu_context *cpuctx,
1187 struct perf_event_context *ctx)
1189 struct perf_event *event;
1190 int state = group_event->state;
1192 event_sched_out(group_event, cpuctx, ctx);
1195 * Schedule out siblings (if any):
1197 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1198 event_sched_out(event, cpuctx, ctx);
1200 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1201 cpuctx->exclusive = 0;
1205 * Cross CPU call to remove a performance event
1207 * We disable the event on the hardware level first. After that we
1208 * remove it from the context list.
1210 static int __perf_remove_from_context(void *info)
1212 struct perf_event *event = info;
1213 struct perf_event_context *ctx = event->ctx;
1214 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1216 raw_spin_lock(&ctx->lock);
1217 event_sched_out(event, cpuctx, ctx);
1218 list_del_event(event, ctx);
1219 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1221 cpuctx->task_ctx = NULL;
1223 raw_spin_unlock(&ctx->lock);
1230 * Remove the event from a task's (or a CPU's) list of events.
1232 * CPU events are removed with a smp call. For task events we only
1233 * call when the task is on a CPU.
1235 * If event->ctx is a cloned context, callers must make sure that
1236 * every task struct that event->ctx->task could possibly point to
1237 * remains valid. This is OK when called from perf_release since
1238 * that only calls us on the top-level context, which can't be a clone.
1239 * When called from perf_event_exit_task, it's OK because the
1240 * context has been detached from its task.
1242 static void perf_remove_from_context(struct perf_event *event)
1244 struct perf_event_context *ctx = event->ctx;
1245 struct task_struct *task = ctx->task;
1247 lockdep_assert_held(&ctx->mutex);
1251 * Per cpu events are removed via an smp call and
1252 * the removal is always successful.
1254 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1259 if (!task_function_call(task, __perf_remove_from_context, event))
1262 raw_spin_lock_irq(&ctx->lock);
1264 * If we failed to find a running task, but find the context active now
1265 * that we've acquired the ctx->lock, retry.
1267 if (ctx->is_active) {
1268 raw_spin_unlock_irq(&ctx->lock);
1273 * Since the task isn't running, its safe to remove the event, us
1274 * holding the ctx->lock ensures the task won't get scheduled in.
1276 list_del_event(event, ctx);
1277 raw_spin_unlock_irq(&ctx->lock);
1281 * Cross CPU call to disable a performance event
1283 int __perf_event_disable(void *info)
1285 struct perf_event *event = info;
1286 struct perf_event_context *ctx = event->ctx;
1287 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1290 * If this is a per-task event, need to check whether this
1291 * event's task is the current task on this cpu.
1293 * Can trigger due to concurrent perf_event_context_sched_out()
1294 * flipping contexts around.
1296 if (ctx->task && cpuctx->task_ctx != ctx)
1299 raw_spin_lock(&ctx->lock);
1302 * If the event is on, turn it off.
1303 * If it is in error state, leave it in error state.
1305 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1306 update_context_time(ctx);
1307 update_cgrp_time_from_event(event);
1308 update_group_times(event);
1309 if (event == event->group_leader)
1310 group_sched_out(event, cpuctx, ctx);
1312 event_sched_out(event, cpuctx, ctx);
1313 event->state = PERF_EVENT_STATE_OFF;
1316 raw_spin_unlock(&ctx->lock);
1324 * If event->ctx is a cloned context, callers must make sure that
1325 * every task struct that event->ctx->task could possibly point to
1326 * remains valid. This condition is satisifed when called through
1327 * perf_event_for_each_child or perf_event_for_each because they
1328 * hold the top-level event's child_mutex, so any descendant that
1329 * goes to exit will block in sync_child_event.
1330 * When called from perf_pending_event it's OK because event->ctx
1331 * is the current context on this CPU and preemption is disabled,
1332 * hence we can't get into perf_event_task_sched_out for this context.
1334 void perf_event_disable(struct perf_event *event)
1336 struct perf_event_context *ctx = event->ctx;
1337 struct task_struct *task = ctx->task;
1341 * Disable the event on the cpu that it's on
1343 cpu_function_call(event->cpu, __perf_event_disable, event);
1348 if (!task_function_call(task, __perf_event_disable, event))
1351 raw_spin_lock_irq(&ctx->lock);
1353 * If the event is still active, we need to retry the cross-call.
1355 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1356 raw_spin_unlock_irq(&ctx->lock);
1358 * Reload the task pointer, it might have been changed by
1359 * a concurrent perf_event_context_sched_out().
1366 * Since we have the lock this context can't be scheduled
1367 * in, so we can change the state safely.
1369 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1370 update_group_times(event);
1371 event->state = PERF_EVENT_STATE_OFF;
1373 raw_spin_unlock_irq(&ctx->lock);
1375 EXPORT_SYMBOL_GPL(perf_event_disable);
1377 static void perf_set_shadow_time(struct perf_event *event,
1378 struct perf_event_context *ctx,
1382 * use the correct time source for the time snapshot
1384 * We could get by without this by leveraging the
1385 * fact that to get to this function, the caller
1386 * has most likely already called update_context_time()
1387 * and update_cgrp_time_xx() and thus both timestamp
1388 * are identical (or very close). Given that tstamp is,
1389 * already adjusted for cgroup, we could say that:
1390 * tstamp - ctx->timestamp
1392 * tstamp - cgrp->timestamp.
1394 * Then, in perf_output_read(), the calculation would
1395 * work with no changes because:
1396 * - event is guaranteed scheduled in
1397 * - no scheduled out in between
1398 * - thus the timestamp would be the same
1400 * But this is a bit hairy.
1402 * So instead, we have an explicit cgroup call to remain
1403 * within the time time source all along. We believe it
1404 * is cleaner and simpler to understand.
1406 if (is_cgroup_event(event))
1407 perf_cgroup_set_shadow_time(event, tstamp);
1409 event->shadow_ctx_time = tstamp - ctx->timestamp;
1412 #define MAX_INTERRUPTS (~0ULL)
1414 static void perf_log_throttle(struct perf_event *event, int enable);
1417 event_sched_in(struct perf_event *event,
1418 struct perf_cpu_context *cpuctx,
1419 struct perf_event_context *ctx)
1421 u64 tstamp = perf_event_time(event);
1423 if (event->state <= PERF_EVENT_STATE_OFF)
1426 event->state = PERF_EVENT_STATE_ACTIVE;
1427 event->oncpu = smp_processor_id();
1430 * Unthrottle events, since we scheduled we might have missed several
1431 * ticks already, also for a heavily scheduling task there is little
1432 * guarantee it'll get a tick in a timely manner.
1434 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1435 perf_log_throttle(event, 1);
1436 event->hw.interrupts = 0;
1440 * The new state must be visible before we turn it on in the hardware:
1444 if (event->pmu->add(event, PERF_EF_START)) {
1445 event->state = PERF_EVENT_STATE_INACTIVE;
1450 event->tstamp_running += tstamp - event->tstamp_stopped;
1452 perf_set_shadow_time(event, ctx, tstamp);
1454 if (!is_software_event(event))
1455 cpuctx->active_oncpu++;
1457 if (event->attr.freq && event->attr.sample_freq)
1460 if (event->attr.exclusive)
1461 cpuctx->exclusive = 1;
1467 group_sched_in(struct perf_event *group_event,
1468 struct perf_cpu_context *cpuctx,
1469 struct perf_event_context *ctx)
1471 struct perf_event *event, *partial_group = NULL;
1472 struct pmu *pmu = group_event->pmu;
1473 u64 now = ctx->time;
1474 bool simulate = false;
1476 if (group_event->state == PERF_EVENT_STATE_OFF)
1479 pmu->start_txn(pmu);
1481 if (event_sched_in(group_event, cpuctx, ctx)) {
1482 pmu->cancel_txn(pmu);
1487 * Schedule in siblings as one group (if any):
1489 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1490 if (event_sched_in(event, cpuctx, ctx)) {
1491 partial_group = event;
1496 if (!pmu->commit_txn(pmu))
1501 * Groups can be scheduled in as one unit only, so undo any
1502 * partial group before returning:
1503 * The events up to the failed event are scheduled out normally,
1504 * tstamp_stopped will be updated.
1506 * The failed events and the remaining siblings need to have
1507 * their timings updated as if they had gone thru event_sched_in()
1508 * and event_sched_out(). This is required to get consistent timings
1509 * across the group. This also takes care of the case where the group
1510 * could never be scheduled by ensuring tstamp_stopped is set to mark
1511 * the time the event was actually stopped, such that time delta
1512 * calculation in update_event_times() is correct.
1514 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1515 if (event == partial_group)
1519 event->tstamp_running += now - event->tstamp_stopped;
1520 event->tstamp_stopped = now;
1522 event_sched_out(event, cpuctx, ctx);
1525 event_sched_out(group_event, cpuctx, ctx);
1527 pmu->cancel_txn(pmu);
1533 * Work out whether we can put this event group on the CPU now.
1535 static int group_can_go_on(struct perf_event *event,
1536 struct perf_cpu_context *cpuctx,
1540 * Groups consisting entirely of software events can always go on.
1542 if (event->group_flags & PERF_GROUP_SOFTWARE)
1545 * If an exclusive group is already on, no other hardware
1548 if (cpuctx->exclusive)
1551 * If this group is exclusive and there are already
1552 * events on the CPU, it can't go on.
1554 if (event->attr.exclusive && cpuctx->active_oncpu)
1557 * Otherwise, try to add it if all previous groups were able
1563 static void add_event_to_ctx(struct perf_event *event,
1564 struct perf_event_context *ctx)
1566 u64 tstamp = perf_event_time(event);
1568 list_add_event(event, ctx);
1569 perf_group_attach(event);
1570 event->tstamp_enabled = tstamp;
1571 event->tstamp_running = tstamp;
1572 event->tstamp_stopped = tstamp;
1575 static void task_ctx_sched_out(struct perf_event_context *ctx);
1577 ctx_sched_in(struct perf_event_context *ctx,
1578 struct perf_cpu_context *cpuctx,
1579 enum event_type_t event_type,
1580 struct task_struct *task);
1582 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1583 struct perf_event_context *ctx,
1584 struct task_struct *task)
1586 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1588 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1589 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1591 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1595 * Cross CPU call to install and enable a performance event
1597 * Must be called with ctx->mutex held
1599 static int __perf_install_in_context(void *info)
1601 struct perf_event *event = info;
1602 struct perf_event_context *ctx = event->ctx;
1603 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1604 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1605 struct task_struct *task = current;
1607 perf_ctx_lock(cpuctx, task_ctx);
1608 perf_pmu_disable(cpuctx->ctx.pmu);
1611 * If there was an active task_ctx schedule it out.
1614 task_ctx_sched_out(task_ctx);
1617 * If the context we're installing events in is not the
1618 * active task_ctx, flip them.
1620 if (ctx->task && task_ctx != ctx) {
1622 raw_spin_unlock(&task_ctx->lock);
1623 raw_spin_lock(&ctx->lock);
1628 cpuctx->task_ctx = task_ctx;
1629 task = task_ctx->task;
1632 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1634 update_context_time(ctx);
1636 * update cgrp time only if current cgrp
1637 * matches event->cgrp. Must be done before
1638 * calling add_event_to_ctx()
1640 update_cgrp_time_from_event(event);
1642 add_event_to_ctx(event, ctx);
1645 * Schedule everything back in
1647 perf_event_sched_in(cpuctx, task_ctx, task);
1649 perf_pmu_enable(cpuctx->ctx.pmu);
1650 perf_ctx_unlock(cpuctx, task_ctx);
1656 * Attach a performance event to a context
1658 * First we add the event to the list with the hardware enable bit
1659 * in event->hw_config cleared.
1661 * If the event is attached to a task which is on a CPU we use a smp
1662 * call to enable it in the task context. The task might have been
1663 * scheduled away, but we check this in the smp call again.
1666 perf_install_in_context(struct perf_event_context *ctx,
1667 struct perf_event *event,
1670 struct task_struct *task = ctx->task;
1672 lockdep_assert_held(&ctx->mutex);
1675 if (event->cpu != -1)
1680 * Per cpu events are installed via an smp call and
1681 * the install is always successful.
1683 cpu_function_call(cpu, __perf_install_in_context, event);
1688 if (!task_function_call(task, __perf_install_in_context, event))
1691 raw_spin_lock_irq(&ctx->lock);
1693 * If we failed to find a running task, but find the context active now
1694 * that we've acquired the ctx->lock, retry.
1696 if (ctx->is_active) {
1697 raw_spin_unlock_irq(&ctx->lock);
1702 * Since the task isn't running, its safe to add the event, us holding
1703 * the ctx->lock ensures the task won't get scheduled in.
1705 add_event_to_ctx(event, ctx);
1706 raw_spin_unlock_irq(&ctx->lock);
1710 * Put a event into inactive state and update time fields.
1711 * Enabling the leader of a group effectively enables all
1712 * the group members that aren't explicitly disabled, so we
1713 * have to update their ->tstamp_enabled also.
1714 * Note: this works for group members as well as group leaders
1715 * since the non-leader members' sibling_lists will be empty.
1717 static void __perf_event_mark_enabled(struct perf_event *event)
1719 struct perf_event *sub;
1720 u64 tstamp = perf_event_time(event);
1722 event->state = PERF_EVENT_STATE_INACTIVE;
1723 event->tstamp_enabled = tstamp - event->total_time_enabled;
1724 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1725 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1726 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1731 * Cross CPU call to enable a performance event
1733 static int __perf_event_enable(void *info)
1735 struct perf_event *event = info;
1736 struct perf_event_context *ctx = event->ctx;
1737 struct perf_event *leader = event->group_leader;
1738 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1741 if (WARN_ON_ONCE(!ctx->is_active))
1744 raw_spin_lock(&ctx->lock);
1745 update_context_time(ctx);
1747 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1751 * set current task's cgroup time reference point
1753 perf_cgroup_set_timestamp(current, ctx);
1755 __perf_event_mark_enabled(event);
1757 if (!event_filter_match(event)) {
1758 if (is_cgroup_event(event))
1759 perf_cgroup_defer_enabled(event);
1764 * If the event is in a group and isn't the group leader,
1765 * then don't put it on unless the group is on.
1767 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1770 if (!group_can_go_on(event, cpuctx, 1)) {
1773 if (event == leader)
1774 err = group_sched_in(event, cpuctx, ctx);
1776 err = event_sched_in(event, cpuctx, ctx);
1781 * If this event can't go on and it's part of a
1782 * group, then the whole group has to come off.
1784 if (leader != event)
1785 group_sched_out(leader, cpuctx, ctx);
1786 if (leader->attr.pinned) {
1787 update_group_times(leader);
1788 leader->state = PERF_EVENT_STATE_ERROR;
1793 raw_spin_unlock(&ctx->lock);
1801 * If event->ctx is a cloned context, callers must make sure that
1802 * every task struct that event->ctx->task could possibly point to
1803 * remains valid. This condition is satisfied when called through
1804 * perf_event_for_each_child or perf_event_for_each as described
1805 * for perf_event_disable.
1807 void perf_event_enable(struct perf_event *event)
1809 struct perf_event_context *ctx = event->ctx;
1810 struct task_struct *task = ctx->task;
1814 * Enable the event on the cpu that it's on
1816 cpu_function_call(event->cpu, __perf_event_enable, event);
1820 raw_spin_lock_irq(&ctx->lock);
1821 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1825 * If the event is in error state, clear that first.
1826 * That way, if we see the event in error state below, we
1827 * know that it has gone back into error state, as distinct
1828 * from the task having been scheduled away before the
1829 * cross-call arrived.
1831 if (event->state == PERF_EVENT_STATE_ERROR)
1832 event->state = PERF_EVENT_STATE_OFF;
1835 if (!ctx->is_active) {
1836 __perf_event_mark_enabled(event);
1840 raw_spin_unlock_irq(&ctx->lock);
1842 if (!task_function_call(task, __perf_event_enable, event))
1845 raw_spin_lock_irq(&ctx->lock);
1848 * If the context is active and the event is still off,
1849 * we need to retry the cross-call.
1851 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1853 * task could have been flipped by a concurrent
1854 * perf_event_context_sched_out()
1861 raw_spin_unlock_irq(&ctx->lock);
1863 EXPORT_SYMBOL_GPL(perf_event_enable);
1865 int perf_event_refresh(struct perf_event *event, int refresh)
1868 * not supported on inherited events
1870 if (event->attr.inherit || !is_sampling_event(event))
1873 atomic_add(refresh, &event->event_limit);
1874 perf_event_enable(event);
1878 EXPORT_SYMBOL_GPL(perf_event_refresh);
1880 static void ctx_sched_out(struct perf_event_context *ctx,
1881 struct perf_cpu_context *cpuctx,
1882 enum event_type_t event_type)
1884 struct perf_event *event;
1885 int is_active = ctx->is_active;
1887 ctx->is_active &= ~event_type;
1888 if (likely(!ctx->nr_events))
1891 update_context_time(ctx);
1892 update_cgrp_time_from_cpuctx(cpuctx);
1893 if (!ctx->nr_active)
1896 perf_pmu_disable(ctx->pmu);
1897 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1898 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1899 group_sched_out(event, cpuctx, ctx);
1902 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1903 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1904 group_sched_out(event, cpuctx, ctx);
1906 perf_pmu_enable(ctx->pmu);
1910 * Test whether two contexts are equivalent, i.e. whether they
1911 * have both been cloned from the same version of the same context
1912 * and they both have the same number of enabled events.
1913 * If the number of enabled events is the same, then the set
1914 * of enabled events should be the same, because these are both
1915 * inherited contexts, therefore we can't access individual events
1916 * in them directly with an fd; we can only enable/disable all
1917 * events via prctl, or enable/disable all events in a family
1918 * via ioctl, which will have the same effect on both contexts.
1920 static int context_equiv(struct perf_event_context *ctx1,
1921 struct perf_event_context *ctx2)
1923 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1924 && ctx1->parent_gen == ctx2->parent_gen
1925 && !ctx1->pin_count && !ctx2->pin_count;
1928 static void __perf_event_sync_stat(struct perf_event *event,
1929 struct perf_event *next_event)
1933 if (!event->attr.inherit_stat)
1937 * Update the event value, we cannot use perf_event_read()
1938 * because we're in the middle of a context switch and have IRQs
1939 * disabled, which upsets smp_call_function_single(), however
1940 * we know the event must be on the current CPU, therefore we
1941 * don't need to use it.
1943 switch (event->state) {
1944 case PERF_EVENT_STATE_ACTIVE:
1945 event->pmu->read(event);
1948 case PERF_EVENT_STATE_INACTIVE:
1949 update_event_times(event);
1957 * In order to keep per-task stats reliable we need to flip the event
1958 * values when we flip the contexts.
1960 value = local64_read(&next_event->count);
1961 value = local64_xchg(&event->count, value);
1962 local64_set(&next_event->count, value);
1964 swap(event->total_time_enabled, next_event->total_time_enabled);
1965 swap(event->total_time_running, next_event->total_time_running);
1968 * Since we swizzled the values, update the user visible data too.
1970 perf_event_update_userpage(event);
1971 perf_event_update_userpage(next_event);
1974 #define list_next_entry(pos, member) \
1975 list_entry(pos->member.next, typeof(*pos), member)
1977 static void perf_event_sync_stat(struct perf_event_context *ctx,
1978 struct perf_event_context *next_ctx)
1980 struct perf_event *event, *next_event;
1985 update_context_time(ctx);
1987 event = list_first_entry(&ctx->event_list,
1988 struct perf_event, event_entry);
1990 next_event = list_first_entry(&next_ctx->event_list,
1991 struct perf_event, event_entry);
1993 while (&event->event_entry != &ctx->event_list &&
1994 &next_event->event_entry != &next_ctx->event_list) {
1996 __perf_event_sync_stat(event, next_event);
1998 event = list_next_entry(event, event_entry);
1999 next_event = list_next_entry(next_event, event_entry);
2003 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2004 struct task_struct *next)
2006 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2007 struct perf_event_context *next_ctx;
2008 struct perf_event_context *parent;
2009 struct perf_cpu_context *cpuctx;
2015 cpuctx = __get_cpu_context(ctx);
2016 if (!cpuctx->task_ctx)
2020 parent = rcu_dereference(ctx->parent_ctx);
2021 next_ctx = next->perf_event_ctxp[ctxn];
2022 if (parent && next_ctx &&
2023 rcu_dereference(next_ctx->parent_ctx) == parent) {
2025 * Looks like the two contexts are clones, so we might be
2026 * able to optimize the context switch. We lock both
2027 * contexts and check that they are clones under the
2028 * lock (including re-checking that neither has been
2029 * uncloned in the meantime). It doesn't matter which
2030 * order we take the locks because no other cpu could
2031 * be trying to lock both of these tasks.
2033 raw_spin_lock(&ctx->lock);
2034 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2035 if (context_equiv(ctx, next_ctx)) {
2037 * XXX do we need a memory barrier of sorts
2038 * wrt to rcu_dereference() of perf_event_ctxp
2040 task->perf_event_ctxp[ctxn] = next_ctx;
2041 next->perf_event_ctxp[ctxn] = ctx;
2043 next_ctx->task = task;
2046 perf_event_sync_stat(ctx, next_ctx);
2048 raw_spin_unlock(&next_ctx->lock);
2049 raw_spin_unlock(&ctx->lock);
2054 raw_spin_lock(&ctx->lock);
2055 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2056 cpuctx->task_ctx = NULL;
2057 raw_spin_unlock(&ctx->lock);
2061 #define for_each_task_context_nr(ctxn) \
2062 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2065 * Called from scheduler to remove the events of the current task,
2066 * with interrupts disabled.
2068 * We stop each event and update the event value in event->count.
2070 * This does not protect us against NMI, but disable()
2071 * sets the disabled bit in the control field of event _before_
2072 * accessing the event control register. If a NMI hits, then it will
2073 * not restart the event.
2075 void __perf_event_task_sched_out(struct task_struct *task,
2076 struct task_struct *next)
2080 for_each_task_context_nr(ctxn)
2081 perf_event_context_sched_out(task, ctxn, next);
2084 * if cgroup events exist on this CPU, then we need
2085 * to check if we have to switch out PMU state.
2086 * cgroup event are system-wide mode only
2088 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2089 perf_cgroup_sched_out(task, next);
2092 static void task_ctx_sched_out(struct perf_event_context *ctx)
2094 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2096 if (!cpuctx->task_ctx)
2099 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2102 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2103 cpuctx->task_ctx = NULL;
2107 * Called with IRQs disabled
2109 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2110 enum event_type_t event_type)
2112 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2116 ctx_pinned_sched_in(struct perf_event_context *ctx,
2117 struct perf_cpu_context *cpuctx)
2119 struct perf_event *event;
2121 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2122 if (event->state <= PERF_EVENT_STATE_OFF)
2124 if (!event_filter_match(event))
2127 /* may need to reset tstamp_enabled */
2128 if (is_cgroup_event(event))
2129 perf_cgroup_mark_enabled(event, ctx);
2131 if (group_can_go_on(event, cpuctx, 1))
2132 group_sched_in(event, cpuctx, ctx);
2135 * If this pinned group hasn't been scheduled,
2136 * put it in error state.
2138 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2139 update_group_times(event);
2140 event->state = PERF_EVENT_STATE_ERROR;
2146 ctx_flexible_sched_in(struct perf_event_context *ctx,
2147 struct perf_cpu_context *cpuctx)
2149 struct perf_event *event;
2152 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2153 /* Ignore events in OFF or ERROR state */
2154 if (event->state <= PERF_EVENT_STATE_OFF)
2157 * Listen to the 'cpu' scheduling filter constraint
2160 if (!event_filter_match(event))
2163 /* may need to reset tstamp_enabled */
2164 if (is_cgroup_event(event))
2165 perf_cgroup_mark_enabled(event, ctx);
2167 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2168 if (group_sched_in(event, cpuctx, ctx))
2175 ctx_sched_in(struct perf_event_context *ctx,
2176 struct perf_cpu_context *cpuctx,
2177 enum event_type_t event_type,
2178 struct task_struct *task)
2181 int is_active = ctx->is_active;
2183 ctx->is_active |= event_type;
2184 if (likely(!ctx->nr_events))
2188 ctx->timestamp = now;
2189 perf_cgroup_set_timestamp(task, ctx);
2191 * First go through the list and put on any pinned groups
2192 * in order to give them the best chance of going on.
2194 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2195 ctx_pinned_sched_in(ctx, cpuctx);
2197 /* Then walk through the lower prio flexible groups */
2198 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2199 ctx_flexible_sched_in(ctx, cpuctx);
2202 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2203 enum event_type_t event_type,
2204 struct task_struct *task)
2206 struct perf_event_context *ctx = &cpuctx->ctx;
2208 ctx_sched_in(ctx, cpuctx, event_type, task);
2211 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2212 struct task_struct *task)
2214 struct perf_cpu_context *cpuctx;
2216 cpuctx = __get_cpu_context(ctx);
2217 if (cpuctx->task_ctx == ctx)
2220 perf_ctx_lock(cpuctx, ctx);
2221 perf_pmu_disable(ctx->pmu);
2223 * We want to keep the following priority order:
2224 * cpu pinned (that don't need to move), task pinned,
2225 * cpu flexible, task flexible.
2227 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2230 cpuctx->task_ctx = ctx;
2232 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2234 perf_pmu_enable(ctx->pmu);
2235 perf_ctx_unlock(cpuctx, ctx);
2238 * Since these rotations are per-cpu, we need to ensure the
2239 * cpu-context we got scheduled on is actually rotating.
2241 perf_pmu_rotate_start(ctx->pmu);
2245 * When sampling the branck stack in system-wide, it may be necessary
2246 * to flush the stack on context switch. This happens when the branch
2247 * stack does not tag its entries with the pid of the current task.
2248 * Otherwise it becomes impossible to associate a branch entry with a
2249 * task. This ambiguity is more likely to appear when the branch stack
2250 * supports priv level filtering and the user sets it to monitor only
2251 * at the user level (which could be a useful measurement in system-wide
2252 * mode). In that case, the risk is high of having a branch stack with
2253 * branch from multiple tasks. Flushing may mean dropping the existing
2254 * entries or stashing them somewhere in the PMU specific code layer.
2256 * This function provides the context switch callback to the lower code
2257 * layer. It is invoked ONLY when there is at least one system-wide context
2258 * with at least one active event using taken branch sampling.
2260 static void perf_branch_stack_sched_in(struct task_struct *prev,
2261 struct task_struct *task)
2263 struct perf_cpu_context *cpuctx;
2265 unsigned long flags;
2267 /* no need to flush branch stack if not changing task */
2271 local_irq_save(flags);
2275 list_for_each_entry_rcu(pmu, &pmus, entry) {
2276 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2279 * check if the context has at least one
2280 * event using PERF_SAMPLE_BRANCH_STACK
2282 if (cpuctx->ctx.nr_branch_stack > 0
2283 && pmu->flush_branch_stack) {
2285 pmu = cpuctx->ctx.pmu;
2287 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2289 perf_pmu_disable(pmu);
2291 pmu->flush_branch_stack();
2293 perf_pmu_enable(pmu);
2295 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2301 local_irq_restore(flags);
2305 * Called from scheduler to add the events of the current task
2306 * with interrupts disabled.
2308 * We restore the event value and then enable it.
2310 * This does not protect us against NMI, but enable()
2311 * sets the enabled bit in the control field of event _before_
2312 * accessing the event control register. If a NMI hits, then it will
2313 * keep the event running.
2315 void __perf_event_task_sched_in(struct task_struct *prev,
2316 struct task_struct *task)
2318 struct perf_event_context *ctx;
2321 for_each_task_context_nr(ctxn) {
2322 ctx = task->perf_event_ctxp[ctxn];
2326 perf_event_context_sched_in(ctx, task);
2329 * if cgroup events exist on this CPU, then we need
2330 * to check if we have to switch in PMU state.
2331 * cgroup event are system-wide mode only
2333 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2334 perf_cgroup_sched_in(prev, task);
2336 /* check for system-wide branch_stack events */
2337 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2338 perf_branch_stack_sched_in(prev, task);
2341 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2343 u64 frequency = event->attr.sample_freq;
2344 u64 sec = NSEC_PER_SEC;
2345 u64 divisor, dividend;
2347 int count_fls, nsec_fls, frequency_fls, sec_fls;
2349 count_fls = fls64(count);
2350 nsec_fls = fls64(nsec);
2351 frequency_fls = fls64(frequency);
2355 * We got @count in @nsec, with a target of sample_freq HZ
2356 * the target period becomes:
2359 * period = -------------------
2360 * @nsec * sample_freq
2365 * Reduce accuracy by one bit such that @a and @b converge
2366 * to a similar magnitude.
2368 #define REDUCE_FLS(a, b) \
2370 if (a##_fls > b##_fls) { \
2380 * Reduce accuracy until either term fits in a u64, then proceed with
2381 * the other, so that finally we can do a u64/u64 division.
2383 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2384 REDUCE_FLS(nsec, frequency);
2385 REDUCE_FLS(sec, count);
2388 if (count_fls + sec_fls > 64) {
2389 divisor = nsec * frequency;
2391 while (count_fls + sec_fls > 64) {
2392 REDUCE_FLS(count, sec);
2396 dividend = count * sec;
2398 dividend = count * sec;
2400 while (nsec_fls + frequency_fls > 64) {
2401 REDUCE_FLS(nsec, frequency);
2405 divisor = nsec * frequency;
2411 return div64_u64(dividend, divisor);
2414 static DEFINE_PER_CPU(int, perf_throttled_count);
2415 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2417 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2419 struct hw_perf_event *hwc = &event->hw;
2420 s64 period, sample_period;
2423 period = perf_calculate_period(event, nsec, count);
2425 delta = (s64)(period - hwc->sample_period);
2426 delta = (delta + 7) / 8; /* low pass filter */
2428 sample_period = hwc->sample_period + delta;
2433 hwc->sample_period = sample_period;
2435 if (local64_read(&hwc->period_left) > 8*sample_period) {
2437 event->pmu->stop(event, PERF_EF_UPDATE);
2439 local64_set(&hwc->period_left, 0);
2442 event->pmu->start(event, PERF_EF_RELOAD);
2447 * combine freq adjustment with unthrottling to avoid two passes over the
2448 * events. At the same time, make sure, having freq events does not change
2449 * the rate of unthrottling as that would introduce bias.
2451 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2454 struct perf_event *event;
2455 struct hw_perf_event *hwc;
2456 u64 now, period = TICK_NSEC;
2460 * only need to iterate over all events iff:
2461 * - context have events in frequency mode (needs freq adjust)
2462 * - there are events to unthrottle on this cpu
2464 if (!(ctx->nr_freq || needs_unthr))
2467 raw_spin_lock(&ctx->lock);
2468 perf_pmu_disable(ctx->pmu);
2470 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2471 if (event->state != PERF_EVENT_STATE_ACTIVE)
2474 if (!event_filter_match(event))
2479 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2480 hwc->interrupts = 0;
2481 perf_log_throttle(event, 1);
2482 event->pmu->start(event, 0);
2485 if (!event->attr.freq || !event->attr.sample_freq)
2489 * stop the event and update event->count
2491 event->pmu->stop(event, PERF_EF_UPDATE);
2493 now = local64_read(&event->count);
2494 delta = now - hwc->freq_count_stamp;
2495 hwc->freq_count_stamp = now;
2499 * reload only if value has changed
2500 * we have stopped the event so tell that
2501 * to perf_adjust_period() to avoid stopping it
2505 perf_adjust_period(event, period, delta, false);
2507 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2510 perf_pmu_enable(ctx->pmu);
2511 raw_spin_unlock(&ctx->lock);
2515 * Round-robin a context's events:
2517 static void rotate_ctx(struct perf_event_context *ctx)
2520 * Rotate the first entry last of non-pinned groups. Rotation might be
2521 * disabled by the inheritance code.
2523 if (!ctx->rotate_disable)
2524 list_rotate_left(&ctx->flexible_groups);
2528 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2529 * because they're strictly cpu affine and rotate_start is called with IRQs
2530 * disabled, while rotate_context is called from IRQ context.
2532 static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2534 struct perf_event_context *ctx = NULL;
2535 int rotate = 0, remove = 1;
2537 if (cpuctx->ctx.nr_events) {
2539 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2543 ctx = cpuctx->task_ctx;
2544 if (ctx && ctx->nr_events) {
2546 if (ctx->nr_events != ctx->nr_active)
2553 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2554 perf_pmu_disable(cpuctx->ctx.pmu);
2556 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2558 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2560 rotate_ctx(&cpuctx->ctx);
2564 perf_event_sched_in(cpuctx, ctx, current);
2566 perf_pmu_enable(cpuctx->ctx.pmu);
2567 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2570 list_del_init(&cpuctx->rotation_list);
2573 void perf_event_task_tick(void)
2575 struct list_head *head = &__get_cpu_var(rotation_list);
2576 struct perf_cpu_context *cpuctx, *tmp;
2577 struct perf_event_context *ctx;
2580 WARN_ON(!irqs_disabled());
2582 __this_cpu_inc(perf_throttled_seq);
2583 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2585 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2587 perf_adjust_freq_unthr_context(ctx, throttled);
2589 ctx = cpuctx->task_ctx;
2591 perf_adjust_freq_unthr_context(ctx, throttled);
2593 if (cpuctx->jiffies_interval == 1 ||
2594 !(jiffies % cpuctx->jiffies_interval))
2595 perf_rotate_context(cpuctx);
2599 static int event_enable_on_exec(struct perf_event *event,
2600 struct perf_event_context *ctx)
2602 if (!event->attr.enable_on_exec)
2605 event->attr.enable_on_exec = 0;
2606 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2609 __perf_event_mark_enabled(event);
2615 * Enable all of a task's events that have been marked enable-on-exec.
2616 * This expects task == current.
2618 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2620 struct perf_event *event;
2621 unsigned long flags;
2625 local_irq_save(flags);
2626 if (!ctx || !ctx->nr_events)
2630 * We must ctxsw out cgroup events to avoid conflict
2631 * when invoking perf_task_event_sched_in() later on
2632 * in this function. Otherwise we end up trying to
2633 * ctxswin cgroup events which are already scheduled
2636 perf_cgroup_sched_out(current, NULL);
2638 raw_spin_lock(&ctx->lock);
2639 task_ctx_sched_out(ctx);
2641 list_for_each_entry(event, &ctx->event_list, event_entry) {
2642 ret = event_enable_on_exec(event, ctx);
2648 * Unclone this context if we enabled any event.
2653 raw_spin_unlock(&ctx->lock);
2656 * Also calls ctxswin for cgroup events, if any:
2658 perf_event_context_sched_in(ctx, ctx->task);
2660 local_irq_restore(flags);
2664 * Cross CPU call to read the hardware event
2666 static void __perf_event_read(void *info)
2668 struct perf_event *event = info;
2669 struct perf_event_context *ctx = event->ctx;
2670 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2673 * If this is a task context, we need to check whether it is
2674 * the current task context of this cpu. If not it has been
2675 * scheduled out before the smp call arrived. In that case
2676 * event->count would have been updated to a recent sample
2677 * when the event was scheduled out.
2679 if (ctx->task && cpuctx->task_ctx != ctx)
2682 raw_spin_lock(&ctx->lock);
2683 if (ctx->is_active) {
2684 update_context_time(ctx);
2685 update_cgrp_time_from_event(event);
2687 update_event_times(event);
2688 if (event->state == PERF_EVENT_STATE_ACTIVE)
2689 event->pmu->read(event);
2690 raw_spin_unlock(&ctx->lock);
2693 static inline u64 perf_event_count(struct perf_event *event)
2695 return local64_read(&event->count) + atomic64_read(&event->child_count);
2698 static u64 perf_event_read(struct perf_event *event)
2701 * If event is enabled and currently active on a CPU, update the
2702 * value in the event structure:
2704 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2705 smp_call_function_single(event->oncpu,
2706 __perf_event_read, event, 1);
2707 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2708 struct perf_event_context *ctx = event->ctx;
2709 unsigned long flags;
2711 raw_spin_lock_irqsave(&ctx->lock, flags);
2713 * may read while context is not active
2714 * (e.g., thread is blocked), in that case
2715 * we cannot update context time
2717 if (ctx->is_active) {
2718 update_context_time(ctx);
2719 update_cgrp_time_from_event(event);
2721 update_event_times(event);
2722 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2725 return perf_event_count(event);
2729 * Initialize the perf_event context in a task_struct:
2731 static void __perf_event_init_context(struct perf_event_context *ctx)
2733 raw_spin_lock_init(&ctx->lock);
2734 mutex_init(&ctx->mutex);
2735 INIT_LIST_HEAD(&ctx->pinned_groups);
2736 INIT_LIST_HEAD(&ctx->flexible_groups);
2737 INIT_LIST_HEAD(&ctx->event_list);
2738 atomic_set(&ctx->refcount, 1);
2741 static struct perf_event_context *
2742 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2744 struct perf_event_context *ctx;
2746 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2750 __perf_event_init_context(ctx);
2753 get_task_struct(task);
2760 static struct task_struct *
2761 find_lively_task_by_vpid(pid_t vpid)
2763 struct task_struct *task;
2770 task = find_task_by_vpid(vpid);
2772 get_task_struct(task);
2776 return ERR_PTR(-ESRCH);
2778 /* Reuse ptrace permission checks for now. */
2780 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2785 put_task_struct(task);
2786 return ERR_PTR(err);
2791 * Returns a matching context with refcount and pincount.
2793 static struct perf_event_context *
2794 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2796 struct perf_event_context *ctx;
2797 struct perf_cpu_context *cpuctx;
2798 unsigned long flags;
2802 /* Must be root to operate on a CPU event: */
2803 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2804 return ERR_PTR(-EACCES);
2807 * We could be clever and allow to attach a event to an
2808 * offline CPU and activate it when the CPU comes up, but
2811 if (!cpu_online(cpu))
2812 return ERR_PTR(-ENODEV);
2814 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2823 ctxn = pmu->task_ctx_nr;
2828 ctx = perf_lock_task_context(task, ctxn, &flags);
2832 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2834 ctx = alloc_perf_context(pmu, task);
2840 mutex_lock(&task->perf_event_mutex);
2842 * If it has already passed perf_event_exit_task().
2843 * we must see PF_EXITING, it takes this mutex too.
2845 if (task->flags & PF_EXITING)
2847 else if (task->perf_event_ctxp[ctxn])
2852 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2854 mutex_unlock(&task->perf_event_mutex);
2856 if (unlikely(err)) {
2868 return ERR_PTR(err);
2871 static void perf_event_free_filter(struct perf_event *event);
2873 static void free_event_rcu(struct rcu_head *head)
2875 struct perf_event *event;
2877 event = container_of(head, struct perf_event, rcu_head);
2879 put_pid_ns(event->ns);
2880 perf_event_free_filter(event);
2884 static void ring_buffer_put(struct ring_buffer *rb);
2886 static void free_event(struct perf_event *event)
2888 irq_work_sync(&event->pending);
2890 if (!event->parent) {
2891 if (event->attach_state & PERF_ATTACH_TASK)
2892 static_key_slow_dec_deferred(&perf_sched_events);
2893 if (event->attr.mmap || event->attr.mmap_data)
2894 atomic_dec(&nr_mmap_events);
2895 if (event->attr.comm)
2896 atomic_dec(&nr_comm_events);
2897 if (event->attr.task)
2898 atomic_dec(&nr_task_events);
2899 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2900 put_callchain_buffers();
2901 if (is_cgroup_event(event)) {
2902 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2903 static_key_slow_dec_deferred(&perf_sched_events);
2906 if (has_branch_stack(event)) {
2907 static_key_slow_dec_deferred(&perf_sched_events);
2908 /* is system-wide event */
2909 if (!(event->attach_state & PERF_ATTACH_TASK))
2910 atomic_dec(&per_cpu(perf_branch_stack_events,
2916 ring_buffer_put(event->rb);
2920 if (is_cgroup_event(event))
2921 perf_detach_cgroup(event);
2924 event->destroy(event);
2927 put_ctx(event->ctx);
2929 call_rcu(&event->rcu_head, free_event_rcu);
2932 int perf_event_release_kernel(struct perf_event *event)
2934 struct perf_event_context *ctx = event->ctx;
2936 WARN_ON_ONCE(ctx->parent_ctx);
2938 * There are two ways this annotation is useful:
2940 * 1) there is a lock recursion from perf_event_exit_task
2941 * see the comment there.
2943 * 2) there is a lock-inversion with mmap_sem through
2944 * perf_event_read_group(), which takes faults while
2945 * holding ctx->mutex, however this is called after
2946 * the last filedesc died, so there is no possibility
2947 * to trigger the AB-BA case.
2949 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2950 raw_spin_lock_irq(&ctx->lock);
2951 perf_group_detach(event);
2952 raw_spin_unlock_irq(&ctx->lock);
2953 perf_remove_from_context(event);
2954 mutex_unlock(&ctx->mutex);
2960 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
2963 * Called when the last reference to the file is gone.
2965 static void put_event(struct perf_event *event)
2967 struct task_struct *owner;
2969 if (!atomic_long_dec_and_test(&event->refcount))
2973 owner = ACCESS_ONCE(event->owner);
2975 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
2976 * !owner it means the list deletion is complete and we can indeed
2977 * free this event, otherwise we need to serialize on
2978 * owner->perf_event_mutex.
2980 smp_read_barrier_depends();
2983 * Since delayed_put_task_struct() also drops the last
2984 * task reference we can safely take a new reference
2985 * while holding the rcu_read_lock().
2987 get_task_struct(owner);
2992 mutex_lock(&owner->perf_event_mutex);
2994 * We have to re-check the event->owner field, if it is cleared
2995 * we raced with perf_event_exit_task(), acquiring the mutex
2996 * ensured they're done, and we can proceed with freeing the
3000 list_del_init(&event->owner_entry);
3001 mutex_unlock(&owner->perf_event_mutex);
3002 put_task_struct(owner);
3005 perf_event_release_kernel(event);
3008 static int perf_release(struct inode *inode, struct file *file)
3010 put_event(file->private_data);
3014 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3016 struct perf_event *child;
3022 mutex_lock(&event->child_mutex);
3023 total += perf_event_read(event);
3024 *enabled += event->total_time_enabled +
3025 atomic64_read(&event->child_total_time_enabled);
3026 *running += event->total_time_running +
3027 atomic64_read(&event->child_total_time_running);
3029 list_for_each_entry(child, &event->child_list, child_list) {
3030 total += perf_event_read(child);
3031 *enabled += child->total_time_enabled;
3032 *running += child->total_time_running;
3034 mutex_unlock(&event->child_mutex);
3038 EXPORT_SYMBOL_GPL(perf_event_read_value);
3040 static int perf_event_read_group(struct perf_event *event,
3041 u64 read_format, char __user *buf)
3043 struct perf_event *leader = event->group_leader, *sub;
3044 int n = 0, size = 0, ret = -EFAULT;
3045 struct perf_event_context *ctx = leader->ctx;
3047 u64 count, enabled, running;
3049 mutex_lock(&ctx->mutex);
3050 count = perf_event_read_value(leader, &enabled, &running);
3052 values[n++] = 1 + leader->nr_siblings;
3053 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3054 values[n++] = enabled;
3055 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3056 values[n++] = running;
3057 values[n++] = count;
3058 if (read_format & PERF_FORMAT_ID)
3059 values[n++] = primary_event_id(leader);
3061 size = n * sizeof(u64);
3063 if (copy_to_user(buf, values, size))
3068 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3071 values[n++] = perf_event_read_value(sub, &enabled, &running);
3072 if (read_format & PERF_FORMAT_ID)
3073 values[n++] = primary_event_id(sub);
3075 size = n * sizeof(u64);
3077 if (copy_to_user(buf + ret, values, size)) {
3085 mutex_unlock(&ctx->mutex);
3090 static int perf_event_read_one(struct perf_event *event,
3091 u64 read_format, char __user *buf)
3093 u64 enabled, running;
3097 values[n++] = perf_event_read_value(event, &enabled, &running);
3098 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3099 values[n++] = enabled;
3100 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3101 values[n++] = running;
3102 if (read_format & PERF_FORMAT_ID)
3103 values[n++] = primary_event_id(event);
3105 if (copy_to_user(buf, values, n * sizeof(u64)))
3108 return n * sizeof(u64);
3112 * Read the performance event - simple non blocking version for now
3115 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3117 u64 read_format = event->attr.read_format;
3121 * Return end-of-file for a read on a event that is in
3122 * error state (i.e. because it was pinned but it couldn't be
3123 * scheduled on to the CPU at some point).
3125 if (event->state == PERF_EVENT_STATE_ERROR)
3128 if (count < event->read_size)
3131 WARN_ON_ONCE(event->ctx->parent_ctx);
3132 if (read_format & PERF_FORMAT_GROUP)
3133 ret = perf_event_read_group(event, read_format, buf);
3135 ret = perf_event_read_one(event, read_format, buf);
3141 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3143 struct perf_event *event = file->private_data;
3145 return perf_read_hw(event, buf, count);
3148 static unsigned int perf_poll(struct file *file, poll_table *wait)
3150 struct perf_event *event = file->private_data;
3151 struct ring_buffer *rb;
3152 unsigned int events = POLL_HUP;
3155 * Race between perf_event_set_output() and perf_poll(): perf_poll()
3156 * grabs the rb reference but perf_event_set_output() overrides it.
3157 * Here is the timeline for two threads T1, T2:
3158 * t0: T1, rb = rcu_dereference(event->rb)
3159 * t1: T2, old_rb = event->rb
3160 * t2: T2, event->rb = new rb
3161 * t3: T2, ring_buffer_detach(old_rb)
3162 * t4: T1, ring_buffer_attach(rb1)
3163 * t5: T1, poll_wait(event->waitq)
3165 * To avoid this problem, we grab mmap_mutex in perf_poll()
3166 * thereby ensuring that the assignment of the new ring buffer
3167 * and the detachment of the old buffer appear atomic to perf_poll()
3169 mutex_lock(&event->mmap_mutex);
3172 rb = rcu_dereference(event->rb);
3174 ring_buffer_attach(event, rb);
3175 events = atomic_xchg(&rb->poll, 0);
3179 mutex_unlock(&event->mmap_mutex);
3181 poll_wait(file, &event->waitq, wait);
3186 static void perf_event_reset(struct perf_event *event)
3188 (void)perf_event_read(event);
3189 local64_set(&event->count, 0);
3190 perf_event_update_userpage(event);
3194 * Holding the top-level event's child_mutex means that any
3195 * descendant process that has inherited this event will block
3196 * in sync_child_event if it goes to exit, thus satisfying the
3197 * task existence requirements of perf_event_enable/disable.
3199 static void perf_event_for_each_child(struct perf_event *event,
3200 void (*func)(struct perf_event *))
3202 struct perf_event *child;
3204 WARN_ON_ONCE(event->ctx->parent_ctx);
3205 mutex_lock(&event->child_mutex);
3207 list_for_each_entry(child, &event->child_list, child_list)
3209 mutex_unlock(&event->child_mutex);
3212 static void perf_event_for_each(struct perf_event *event,
3213 void (*func)(struct perf_event *))
3215 struct perf_event_context *ctx = event->ctx;
3216 struct perf_event *sibling;
3218 WARN_ON_ONCE(ctx->parent_ctx);
3219 mutex_lock(&ctx->mutex);
3220 event = event->group_leader;
3222 perf_event_for_each_child(event, func);
3223 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3224 perf_event_for_each_child(sibling, func);
3225 mutex_unlock(&ctx->mutex);
3228 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3230 struct perf_event_context *ctx = event->ctx;
3234 if (!is_sampling_event(event))
3237 if (copy_from_user(&value, arg, sizeof(value)))
3243 raw_spin_lock_irq(&ctx->lock);
3244 if (event->attr.freq) {
3245 if (value > sysctl_perf_event_sample_rate) {
3250 event->attr.sample_freq = value;
3252 event->attr.sample_period = value;
3253 event->hw.sample_period = value;
3256 raw_spin_unlock_irq(&ctx->lock);
3261 static const struct file_operations perf_fops;
3263 static inline int perf_fget_light(int fd, struct fd *p)
3265 struct fd f = fdget(fd);
3269 if (f.file->f_op != &perf_fops) {
3277 static int perf_event_set_output(struct perf_event *event,
3278 struct perf_event *output_event);
3279 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3281 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3283 struct perf_event *event = file->private_data;
3284 void (*func)(struct perf_event *);
3288 case PERF_EVENT_IOC_ENABLE:
3289 func = perf_event_enable;
3291 case PERF_EVENT_IOC_DISABLE:
3292 func = perf_event_disable;
3294 case PERF_EVENT_IOC_RESET:
3295 func = perf_event_reset;
3298 case PERF_EVENT_IOC_REFRESH:
3299 return perf_event_refresh(event, arg);
3301 case PERF_EVENT_IOC_PERIOD:
3302 return perf_event_period(event, (u64 __user *)arg);
3304 case PERF_EVENT_IOC_SET_OUTPUT:
3308 struct perf_event *output_event;
3310 ret = perf_fget_light(arg, &output);
3313 output_event = output.file->private_data;
3314 ret = perf_event_set_output(event, output_event);
3317 ret = perf_event_set_output(event, NULL);
3322 case PERF_EVENT_IOC_SET_FILTER:
3323 return perf_event_set_filter(event, (void __user *)arg);
3329 if (flags & PERF_IOC_FLAG_GROUP)
3330 perf_event_for_each(event, func);
3332 perf_event_for_each_child(event, func);
3337 int perf_event_task_enable(void)
3339 struct perf_event *event;
3341 mutex_lock(¤t->perf_event_mutex);
3342 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3343 perf_event_for_each_child(event, perf_event_enable);
3344 mutex_unlock(¤t->perf_event_mutex);
3349 int perf_event_task_disable(void)
3351 struct perf_event *event;
3353 mutex_lock(¤t->perf_event_mutex);
3354 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3355 perf_event_for_each_child(event, perf_event_disable);
3356 mutex_unlock(¤t->perf_event_mutex);
3361 static int perf_event_index(struct perf_event *event)
3363 if (event->hw.state & PERF_HES_STOPPED)
3366 if (event->state != PERF_EVENT_STATE_ACTIVE)
3369 return event->pmu->event_idx(event);
3372 static void calc_timer_values(struct perf_event *event,
3379 *now = perf_clock();
3380 ctx_time = event->shadow_ctx_time + *now;
3381 *enabled = ctx_time - event->tstamp_enabled;
3382 *running = ctx_time - event->tstamp_running;
3385 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3390 * Callers need to ensure there can be no nesting of this function, otherwise
3391 * the seqlock logic goes bad. We can not serialize this because the arch
3392 * code calls this from NMI context.
3394 void perf_event_update_userpage(struct perf_event *event)
3396 struct perf_event_mmap_page *userpg;
3397 struct ring_buffer *rb;
3398 u64 enabled, running, now;
3402 * compute total_time_enabled, total_time_running
3403 * based on snapshot values taken when the event
3404 * was last scheduled in.
3406 * we cannot simply called update_context_time()
3407 * because of locking issue as we can be called in
3410 calc_timer_values(event, &now, &enabled, &running);
3411 rb = rcu_dereference(event->rb);
3415 userpg = rb->user_page;
3418 * Disable preemption so as to not let the corresponding user-space
3419 * spin too long if we get preempted.
3424 userpg->index = perf_event_index(event);
3425 userpg->offset = perf_event_count(event);
3427 userpg->offset -= local64_read(&event->hw.prev_count);
3429 userpg->time_enabled = enabled +
3430 atomic64_read(&event->child_total_time_enabled);
3432 userpg->time_running = running +
3433 atomic64_read(&event->child_total_time_running);
3435 arch_perf_update_userpage(userpg, now);
3444 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3446 struct perf_event *event = vma->vm_file->private_data;
3447 struct ring_buffer *rb;
3448 int ret = VM_FAULT_SIGBUS;
3450 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3451 if (vmf->pgoff == 0)
3457 rb = rcu_dereference(event->rb);
3461 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3464 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3468 get_page(vmf->page);
3469 vmf->page->mapping = vma->vm_file->f_mapping;
3470 vmf->page->index = vmf->pgoff;
3479 static void ring_buffer_attach(struct perf_event *event,
3480 struct ring_buffer *rb)
3482 unsigned long flags;
3484 if (!list_empty(&event->rb_entry))
3487 spin_lock_irqsave(&rb->event_lock, flags);
3488 if (!list_empty(&event->rb_entry))
3491 list_add(&event->rb_entry, &rb->event_list);
3493 spin_unlock_irqrestore(&rb->event_lock, flags);
3496 static void ring_buffer_detach(struct perf_event *event,
3497 struct ring_buffer *rb)
3499 unsigned long flags;
3501 if (list_empty(&event->rb_entry))
3504 spin_lock_irqsave(&rb->event_lock, flags);
3505 list_del_init(&event->rb_entry);
3506 wake_up_all(&event->waitq);
3507 spin_unlock_irqrestore(&rb->event_lock, flags);
3510 static void ring_buffer_wakeup(struct perf_event *event)
3512 struct ring_buffer *rb;
3515 rb = rcu_dereference(event->rb);
3519 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3520 wake_up_all(&event->waitq);
3526 static void rb_free_rcu(struct rcu_head *rcu_head)
3528 struct ring_buffer *rb;
3530 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3534 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3536 struct ring_buffer *rb;
3539 rb = rcu_dereference(event->rb);
3541 if (!atomic_inc_not_zero(&rb->refcount))
3549 static void ring_buffer_put(struct ring_buffer *rb)
3551 struct perf_event *event, *n;
3552 unsigned long flags;
3554 if (!atomic_dec_and_test(&rb->refcount))
3557 spin_lock_irqsave(&rb->event_lock, flags);
3558 list_for_each_entry_safe(event, n, &rb->event_list, rb_entry) {
3559 list_del_init(&event->rb_entry);
3560 wake_up_all(&event->waitq);
3562 spin_unlock_irqrestore(&rb->event_lock, flags);
3564 call_rcu(&rb->rcu_head, rb_free_rcu);
3567 static void perf_mmap_open(struct vm_area_struct *vma)
3569 struct perf_event *event = vma->vm_file->private_data;
3571 atomic_inc(&event->mmap_count);
3574 static void perf_mmap_close(struct vm_area_struct *vma)
3576 struct perf_event *event = vma->vm_file->private_data;
3578 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3579 unsigned long size = perf_data_size(event->rb);
3580 struct user_struct *user = event->mmap_user;
3581 struct ring_buffer *rb = event->rb;
3583 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3584 vma->vm_mm->pinned_vm -= event->mmap_locked;
3585 rcu_assign_pointer(event->rb, NULL);
3586 ring_buffer_detach(event, rb);
3587 mutex_unlock(&event->mmap_mutex);
3589 ring_buffer_put(rb);
3594 static const struct vm_operations_struct perf_mmap_vmops = {
3595 .open = perf_mmap_open,
3596 .close = perf_mmap_close,
3597 .fault = perf_mmap_fault,
3598 .page_mkwrite = perf_mmap_fault,
3601 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3603 struct perf_event *event = file->private_data;
3604 unsigned long user_locked, user_lock_limit;
3605 struct user_struct *user = current_user();
3606 unsigned long locked, lock_limit;
3607 struct ring_buffer *rb;
3608 unsigned long vma_size;
3609 unsigned long nr_pages;
3610 long user_extra, extra;
3611 int ret = 0, flags = 0;
3614 * Don't allow mmap() of inherited per-task counters. This would
3615 * create a performance issue due to all children writing to the
3618 if (event->cpu == -1 && event->attr.inherit)
3621 if (!(vma->vm_flags & VM_SHARED))
3624 vma_size = vma->vm_end - vma->vm_start;
3625 nr_pages = (vma_size / PAGE_SIZE) - 1;
3628 * If we have rb pages ensure they're a power-of-two number, so we
3629 * can do bitmasks instead of modulo.
3631 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3634 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3637 if (vma->vm_pgoff != 0)
3640 WARN_ON_ONCE(event->ctx->parent_ctx);
3641 mutex_lock(&event->mmap_mutex);
3643 if (event->rb->nr_pages == nr_pages)
3644 atomic_inc(&event->rb->refcount);
3650 user_extra = nr_pages + 1;
3651 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3654 * Increase the limit linearly with more CPUs:
3656 user_lock_limit *= num_online_cpus();
3658 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3661 if (user_locked > user_lock_limit)
3662 extra = user_locked - user_lock_limit;
3664 lock_limit = rlimit(RLIMIT_MEMLOCK);
3665 lock_limit >>= PAGE_SHIFT;
3666 locked = vma->vm_mm->pinned_vm + extra;
3668 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3669 !capable(CAP_IPC_LOCK)) {
3676 if (vma->vm_flags & VM_WRITE)
3677 flags |= RING_BUFFER_WRITABLE;
3679 rb = rb_alloc(nr_pages,
3680 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3687 rcu_assign_pointer(event->rb, rb);
3689 atomic_long_add(user_extra, &user->locked_vm);
3690 event->mmap_locked = extra;
3691 event->mmap_user = get_current_user();
3692 vma->vm_mm->pinned_vm += event->mmap_locked;
3694 perf_event_update_userpage(event);
3698 atomic_inc(&event->mmap_count);
3699 mutex_unlock(&event->mmap_mutex);
3701 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3702 vma->vm_ops = &perf_mmap_vmops;
3707 static int perf_fasync(int fd, struct file *filp, int on)
3709 struct inode *inode = file_inode(filp);
3710 struct perf_event *event = filp->private_data;
3713 mutex_lock(&inode->i_mutex);
3714 retval = fasync_helper(fd, filp, on, &event->fasync);
3715 mutex_unlock(&inode->i_mutex);
3723 static const struct file_operations perf_fops = {
3724 .llseek = no_llseek,
3725 .release = perf_release,
3728 .unlocked_ioctl = perf_ioctl,
3729 .compat_ioctl = perf_ioctl,
3731 .fasync = perf_fasync,
3737 * If there's data, ensure we set the poll() state and publish everything
3738 * to user-space before waking everybody up.
3741 void perf_event_wakeup(struct perf_event *event)
3743 ring_buffer_wakeup(event);
3745 if (event->pending_kill) {
3746 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3747 event->pending_kill = 0;
3751 static void perf_pending_event(struct irq_work *entry)
3753 struct perf_event *event = container_of(entry,
3754 struct perf_event, pending);
3756 if (event->pending_disable) {
3757 event->pending_disable = 0;
3758 __perf_event_disable(event);
3761 if (event->pending_wakeup) {
3762 event->pending_wakeup = 0;
3763 perf_event_wakeup(event);
3768 * We assume there is only KVM supporting the callbacks.
3769 * Later on, we might change it to a list if there is
3770 * another virtualization implementation supporting the callbacks.
3772 struct perf_guest_info_callbacks *perf_guest_cbs;
3774 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3776 perf_guest_cbs = cbs;
3779 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3781 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3783 perf_guest_cbs = NULL;
3786 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3789 perf_output_sample_regs(struct perf_output_handle *handle,
3790 struct pt_regs *regs, u64 mask)
3794 for_each_set_bit(bit, (const unsigned long *) &mask,
3795 sizeof(mask) * BITS_PER_BYTE) {
3798 val = perf_reg_value(regs, bit);
3799 perf_output_put(handle, val);
3803 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
3804 struct pt_regs *regs)
3806 if (!user_mode(regs)) {
3808 regs = task_pt_regs(current);
3814 regs_user->regs = regs;
3815 regs_user->abi = perf_reg_abi(current);
3820 * Get remaining task size from user stack pointer.
3822 * It'd be better to take stack vma map and limit this more
3823 * precisly, but there's no way to get it safely under interrupt,
3824 * so using TASK_SIZE as limit.
3826 static u64 perf_ustack_task_size(struct pt_regs *regs)
3828 unsigned long addr = perf_user_stack_pointer(regs);
3830 if (!addr || addr >= TASK_SIZE)
3833 return TASK_SIZE - addr;
3837 perf_sample_ustack_size(u16 stack_size, u16 header_size,
3838 struct pt_regs *regs)
3842 /* No regs, no stack pointer, no dump. */
3847 * Check if we fit in with the requested stack size into the:
3849 * If we don't, we limit the size to the TASK_SIZE.
3851 * - remaining sample size
3852 * If we don't, we customize the stack size to
3853 * fit in to the remaining sample size.
3856 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
3857 stack_size = min(stack_size, (u16) task_size);
3859 /* Current header size plus static size and dynamic size. */
3860 header_size += 2 * sizeof(u64);
3862 /* Do we fit in with the current stack dump size? */
3863 if ((u16) (header_size + stack_size) < header_size) {
3865 * If we overflow the maximum size for the sample,
3866 * we customize the stack dump size to fit in.
3868 stack_size = USHRT_MAX - header_size - sizeof(u64);
3869 stack_size = round_up(stack_size, sizeof(u64));
3876 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
3877 struct pt_regs *regs)
3879 /* Case of a kernel thread, nothing to dump */
3882 perf_output_put(handle, size);
3891 * - the size requested by user or the best one we can fit
3892 * in to the sample max size
3894 * - user stack dump data
3896 * - the actual dumped size
3900 perf_output_put(handle, dump_size);
3903 sp = perf_user_stack_pointer(regs);
3904 rem = __output_copy_user(handle, (void *) sp, dump_size);
3905 dyn_size = dump_size - rem;
3907 perf_output_skip(handle, rem);
3910 perf_output_put(handle, dyn_size);
3914 static void __perf_event_header__init_id(struct perf_event_header *header,
3915 struct perf_sample_data *data,
3916 struct perf_event *event)
3918 u64 sample_type = event->attr.sample_type;
3920 data->type = sample_type;
3921 header->size += event->id_header_size;
3923 if (sample_type & PERF_SAMPLE_TID) {
3924 /* namespace issues */
3925 data->tid_entry.pid = perf_event_pid(event, current);
3926 data->tid_entry.tid = perf_event_tid(event, current);
3929 if (sample_type & PERF_SAMPLE_TIME)
3930 data->time = perf_clock();
3932 if (sample_type & PERF_SAMPLE_ID)
3933 data->id = primary_event_id(event);
3935 if (sample_type & PERF_SAMPLE_STREAM_ID)
3936 data->stream_id = event->id;
3938 if (sample_type & PERF_SAMPLE_CPU) {
3939 data->cpu_entry.cpu = raw_smp_processor_id();
3940 data->cpu_entry.reserved = 0;
3944 void perf_event_header__init_id(struct perf_event_header *header,
3945 struct perf_sample_data *data,
3946 struct perf_event *event)
3948 if (event->attr.sample_id_all)
3949 __perf_event_header__init_id(header, data, event);
3952 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3953 struct perf_sample_data *data)
3955 u64 sample_type = data->type;
3957 if (sample_type & PERF_SAMPLE_TID)
3958 perf_output_put(handle, data->tid_entry);
3960 if (sample_type & PERF_SAMPLE_TIME)
3961 perf_output_put(handle, data->time);
3963 if (sample_type & PERF_SAMPLE_ID)
3964 perf_output_put(handle, data->id);
3966 if (sample_type & PERF_SAMPLE_STREAM_ID)
3967 perf_output_put(handle, data->stream_id);
3969 if (sample_type & PERF_SAMPLE_CPU)
3970 perf_output_put(handle, data->cpu_entry);
3973 void perf_event__output_id_sample(struct perf_event *event,
3974 struct perf_output_handle *handle,
3975 struct perf_sample_data *sample)
3977 if (event->attr.sample_id_all)
3978 __perf_event__output_id_sample(handle, sample);
3981 static void perf_output_read_one(struct perf_output_handle *handle,
3982 struct perf_event *event,
3983 u64 enabled, u64 running)
3985 u64 read_format = event->attr.read_format;
3989 values[n++] = perf_event_count(event);
3990 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3991 values[n++] = enabled +
3992 atomic64_read(&event->child_total_time_enabled);
3994 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3995 values[n++] = running +
3996 atomic64_read(&event->child_total_time_running);
3998 if (read_format & PERF_FORMAT_ID)
3999 values[n++] = primary_event_id(event);
4001 __output_copy(handle, values, n * sizeof(u64));
4005 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4007 static void perf_output_read_group(struct perf_output_handle *handle,
4008 struct perf_event *event,
4009 u64 enabled, u64 running)
4011 struct perf_event *leader = event->group_leader, *sub;
4012 u64 read_format = event->attr.read_format;
4016 values[n++] = 1 + leader->nr_siblings;
4018 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4019 values[n++] = enabled;
4021 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4022 values[n++] = running;
4024 if (leader != event)
4025 leader->pmu->read(leader);
4027 values[n++] = perf_event_count(leader);
4028 if (read_format & PERF_FORMAT_ID)
4029 values[n++] = primary_event_id(leader);
4031 __output_copy(handle, values, n * sizeof(u64));
4033 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4037 sub->pmu->read(sub);
4039 values[n++] = perf_event_count(sub);
4040 if (read_format & PERF_FORMAT_ID)
4041 values[n++] = primary_event_id(sub);
4043 __output_copy(handle, values, n * sizeof(u64));
4047 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4048 PERF_FORMAT_TOTAL_TIME_RUNNING)
4050 static void perf_output_read(struct perf_output_handle *handle,
4051 struct perf_event *event)
4053 u64 enabled = 0, running = 0, now;
4054 u64 read_format = event->attr.read_format;
4057 * compute total_time_enabled, total_time_running
4058 * based on snapshot values taken when the event
4059 * was last scheduled in.
4061 * we cannot simply called update_context_time()
4062 * because of locking issue as we are called in
4065 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4066 calc_timer_values(event, &now, &enabled, &running);
4068 if (event->attr.read_format & PERF_FORMAT_GROUP)
4069 perf_output_read_group(handle, event, enabled, running);
4071 perf_output_read_one(handle, event, enabled, running);
4074 void perf_output_sample(struct perf_output_handle *handle,
4075 struct perf_event_header *header,
4076 struct perf_sample_data *data,
4077 struct perf_event *event)
4079 u64 sample_type = data->type;
4081 perf_output_put(handle, *header);
4083 if (sample_type & PERF_SAMPLE_IP)
4084 perf_output_put(handle, data->ip);
4086 if (sample_type & PERF_SAMPLE_TID)
4087 perf_output_put(handle, data->tid_entry);
4089 if (sample_type & PERF_SAMPLE_TIME)
4090 perf_output_put(handle, data->time);
4092 if (sample_type & PERF_SAMPLE_ADDR)
4093 perf_output_put(handle, data->addr);
4095 if (sample_type & PERF_SAMPLE_ID)
4096 perf_output_put(handle, data->id);
4098 if (sample_type & PERF_SAMPLE_STREAM_ID)
4099 perf_output_put(handle, data->stream_id);
4101 if (sample_type & PERF_SAMPLE_CPU)
4102 perf_output_put(handle, data->cpu_entry);
4104 if (sample_type & PERF_SAMPLE_PERIOD)
4105 perf_output_put(handle, data->period);
4107 if (sample_type & PERF_SAMPLE_READ)
4108 perf_output_read(handle, event);
4110 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4111 if (data->callchain) {
4114 if (data->callchain)
4115 size += data->callchain->nr;
4117 size *= sizeof(u64);
4119 __output_copy(handle, data->callchain, size);
4122 perf_output_put(handle, nr);
4126 if (sample_type & PERF_SAMPLE_RAW) {
4128 perf_output_put(handle, data->raw->size);
4129 __output_copy(handle, data->raw->data,
4136 .size = sizeof(u32),
4139 perf_output_put(handle, raw);
4143 if (!event->attr.watermark) {
4144 int wakeup_events = event->attr.wakeup_events;
4146 if (wakeup_events) {
4147 struct ring_buffer *rb = handle->rb;
4148 int events = local_inc_return(&rb->events);
4150 if (events >= wakeup_events) {
4151 local_sub(wakeup_events, &rb->events);
4152 local_inc(&rb->wakeup);
4157 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4158 if (data->br_stack) {
4161 size = data->br_stack->nr
4162 * sizeof(struct perf_branch_entry);
4164 perf_output_put(handle, data->br_stack->nr);
4165 perf_output_copy(handle, data->br_stack->entries, size);
4168 * we always store at least the value of nr
4171 perf_output_put(handle, nr);
4175 if (sample_type & PERF_SAMPLE_REGS_USER) {
4176 u64 abi = data->regs_user.abi;
4179 * If there are no regs to dump, notice it through
4180 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4182 perf_output_put(handle, abi);
4185 u64 mask = event->attr.sample_regs_user;
4186 perf_output_sample_regs(handle,
4187 data->regs_user.regs,
4192 if (sample_type & PERF_SAMPLE_STACK_USER)
4193 perf_output_sample_ustack(handle,
4194 data->stack_user_size,
4195 data->regs_user.regs);
4198 void perf_prepare_sample(struct perf_event_header *header,
4199 struct perf_sample_data *data,
4200 struct perf_event *event,
4201 struct pt_regs *regs)
4203 u64 sample_type = event->attr.sample_type;
4205 header->type = PERF_RECORD_SAMPLE;
4206 header->size = sizeof(*header) + event->header_size;
4209 header->misc |= perf_misc_flags(regs);
4211 __perf_event_header__init_id(header, data, event);
4213 if (sample_type & PERF_SAMPLE_IP)
4214 data->ip = perf_instruction_pointer(regs);
4216 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4219 data->callchain = perf_callchain(event, regs);
4221 if (data->callchain)
4222 size += data->callchain->nr;
4224 header->size += size * sizeof(u64);
4227 if (sample_type & PERF_SAMPLE_RAW) {
4228 int size = sizeof(u32);
4231 size += data->raw->size;
4233 size += sizeof(u32);
4235 WARN_ON_ONCE(size & (sizeof(u64)-1));
4236 header->size += size;
4239 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4240 int size = sizeof(u64); /* nr */
4241 if (data->br_stack) {
4242 size += data->br_stack->nr
4243 * sizeof(struct perf_branch_entry);
4245 header->size += size;
4248 if (sample_type & PERF_SAMPLE_REGS_USER) {
4249 /* regs dump ABI info */
4250 int size = sizeof(u64);
4252 perf_sample_regs_user(&data->regs_user, regs);
4254 if (data->regs_user.regs) {
4255 u64 mask = event->attr.sample_regs_user;
4256 size += hweight64(mask) * sizeof(u64);
4259 header->size += size;
4262 if (sample_type & PERF_SAMPLE_STACK_USER) {
4264 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4265 * processed as the last one or have additional check added
4266 * in case new sample type is added, because we could eat
4267 * up the rest of the sample size.
4269 struct perf_regs_user *uregs = &data->regs_user;
4270 u16 stack_size = event->attr.sample_stack_user;
4271 u16 size = sizeof(u64);
4274 perf_sample_regs_user(uregs, regs);
4276 stack_size = perf_sample_ustack_size(stack_size, header->size,
4280 * If there is something to dump, add space for the dump
4281 * itself and for the field that tells the dynamic size,
4282 * which is how many have been actually dumped.
4285 size += sizeof(u64) + stack_size;
4287 data->stack_user_size = stack_size;
4288 header->size += size;
4292 static void perf_event_output(struct perf_event *event,
4293 struct perf_sample_data *data,
4294 struct pt_regs *regs)
4296 struct perf_output_handle handle;
4297 struct perf_event_header header;
4299 /* protect the callchain buffers */
4302 perf_prepare_sample(&header, data, event, regs);
4304 if (perf_output_begin(&handle, event, header.size))
4307 perf_output_sample(&handle, &header, data, event);
4309 perf_output_end(&handle);
4319 struct perf_read_event {
4320 struct perf_event_header header;
4327 perf_event_read_event(struct perf_event *event,
4328 struct task_struct *task)
4330 struct perf_output_handle handle;
4331 struct perf_sample_data sample;
4332 struct perf_read_event read_event = {
4334 .type = PERF_RECORD_READ,
4336 .size = sizeof(read_event) + event->read_size,
4338 .pid = perf_event_pid(event, task),
4339 .tid = perf_event_tid(event, task),
4343 perf_event_header__init_id(&read_event.header, &sample, event);
4344 ret = perf_output_begin(&handle, event, read_event.header.size);
4348 perf_output_put(&handle, read_event);
4349 perf_output_read(&handle, event);
4350 perf_event__output_id_sample(event, &handle, &sample);
4352 perf_output_end(&handle);
4356 * task tracking -- fork/exit
4358 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4361 struct perf_task_event {
4362 struct task_struct *task;
4363 struct perf_event_context *task_ctx;
4366 struct perf_event_header header;
4376 static void perf_event_task_output(struct perf_event *event,
4377 struct perf_task_event *task_event)
4379 struct perf_output_handle handle;
4380 struct perf_sample_data sample;
4381 struct task_struct *task = task_event->task;
4382 int ret, size = task_event->event_id.header.size;
4384 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4386 ret = perf_output_begin(&handle, event,
4387 task_event->event_id.header.size);
4391 task_event->event_id.pid = perf_event_pid(event, task);
4392 task_event->event_id.ppid = perf_event_pid(event, current);
4394 task_event->event_id.tid = perf_event_tid(event, task);
4395 task_event->event_id.ptid = perf_event_tid(event, current);
4397 perf_output_put(&handle, task_event->event_id);
4399 perf_event__output_id_sample(event, &handle, &sample);
4401 perf_output_end(&handle);
4403 task_event->event_id.header.size = size;
4406 static int perf_event_task_match(struct perf_event *event)
4408 if (event->state < PERF_EVENT_STATE_INACTIVE)
4411 if (!event_filter_match(event))
4414 if (event->attr.comm || event->attr.mmap ||
4415 event->attr.mmap_data || event->attr.task)
4421 static void perf_event_task_ctx(struct perf_event_context *ctx,
4422 struct perf_task_event *task_event)
4424 struct perf_event *event;
4426 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4427 if (perf_event_task_match(event))
4428 perf_event_task_output(event, task_event);
4432 static void perf_event_task_event(struct perf_task_event *task_event)
4434 struct perf_cpu_context *cpuctx;
4435 struct perf_event_context *ctx;
4440 list_for_each_entry_rcu(pmu, &pmus, entry) {
4441 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4442 if (cpuctx->unique_pmu != pmu)
4444 perf_event_task_ctx(&cpuctx->ctx, task_event);
4446 ctx = task_event->task_ctx;
4448 ctxn = pmu->task_ctx_nr;
4451 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4453 perf_event_task_ctx(ctx, task_event);
4456 put_cpu_ptr(pmu->pmu_cpu_context);
4458 if (task_event->task_ctx)
4459 perf_event_task_ctx(task_event->task_ctx, task_event);
4464 static void perf_event_task(struct task_struct *task,
4465 struct perf_event_context *task_ctx,
4468 struct perf_task_event task_event;
4470 if (!atomic_read(&nr_comm_events) &&
4471 !atomic_read(&nr_mmap_events) &&
4472 !atomic_read(&nr_task_events))
4475 task_event = (struct perf_task_event){
4477 .task_ctx = task_ctx,
4480 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4482 .size = sizeof(task_event.event_id),
4488 .time = perf_clock(),
4492 perf_event_task_event(&task_event);
4495 void perf_event_fork(struct task_struct *task)
4497 perf_event_task(task, NULL, 1);
4504 struct perf_comm_event {
4505 struct task_struct *task;
4510 struct perf_event_header header;
4517 static void perf_event_comm_output(struct perf_event *event,
4518 struct perf_comm_event *comm_event)
4520 struct perf_output_handle handle;
4521 struct perf_sample_data sample;
4522 int size = comm_event->event_id.header.size;
4525 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4526 ret = perf_output_begin(&handle, event,
4527 comm_event->event_id.header.size);
4532 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4533 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4535 perf_output_put(&handle, comm_event->event_id);
4536 __output_copy(&handle, comm_event->comm,
4537 comm_event->comm_size);
4539 perf_event__output_id_sample(event, &handle, &sample);
4541 perf_output_end(&handle);
4543 comm_event->event_id.header.size = size;
4546 static int perf_event_comm_match(struct perf_event *event)
4548 if (event->state < PERF_EVENT_STATE_INACTIVE)
4551 if (!event_filter_match(event))
4554 if (event->attr.comm)
4560 static void perf_event_comm_ctx(struct perf_event_context *ctx,
4561 struct perf_comm_event *comm_event)
4563 struct perf_event *event;
4565 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4566 if (perf_event_comm_match(event))
4567 perf_event_comm_output(event, comm_event);
4571 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4573 struct perf_cpu_context *cpuctx;
4574 struct perf_event_context *ctx;
4575 char comm[TASK_COMM_LEN];
4580 memset(comm, 0, sizeof(comm));
4581 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4582 size = ALIGN(strlen(comm)+1, sizeof(u64));
4584 comm_event->comm = comm;
4585 comm_event->comm_size = size;
4587 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4589 list_for_each_entry_rcu(pmu, &pmus, entry) {
4590 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4591 if (cpuctx->unique_pmu != pmu)
4593 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4595 ctxn = pmu->task_ctx_nr;
4599 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4601 perf_event_comm_ctx(ctx, comm_event);
4603 put_cpu_ptr(pmu->pmu_cpu_context);
4608 void perf_event_comm(struct task_struct *task)
4610 struct perf_comm_event comm_event;
4611 struct perf_event_context *ctx;
4615 for_each_task_context_nr(ctxn) {
4616 ctx = task->perf_event_ctxp[ctxn];
4620 perf_event_enable_on_exec(ctx);
4624 if (!atomic_read(&nr_comm_events))
4627 comm_event = (struct perf_comm_event){
4633 .type = PERF_RECORD_COMM,
4642 perf_event_comm_event(&comm_event);
4649 struct perf_mmap_event {
4650 struct vm_area_struct *vma;
4652 const char *file_name;
4656 struct perf_event_header header;
4666 static void perf_event_mmap_output(struct perf_event *event,
4667 struct perf_mmap_event *mmap_event)
4669 struct perf_output_handle handle;
4670 struct perf_sample_data sample;
4671 int size = mmap_event->event_id.header.size;
4674 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4675 ret = perf_output_begin(&handle, event,
4676 mmap_event->event_id.header.size);
4680 mmap_event->event_id.pid = perf_event_pid(event, current);
4681 mmap_event->event_id.tid = perf_event_tid(event, current);
4683 perf_output_put(&handle, mmap_event->event_id);
4684 __output_copy(&handle, mmap_event->file_name,
4685 mmap_event->file_size);
4687 perf_event__output_id_sample(event, &handle, &sample);
4689 perf_output_end(&handle);
4691 mmap_event->event_id.header.size = size;
4694 static int perf_event_mmap_match(struct perf_event *event,
4695 struct perf_mmap_event *mmap_event,
4698 if (event->state < PERF_EVENT_STATE_INACTIVE)
4701 if (!event_filter_match(event))
4704 if ((!executable && event->attr.mmap_data) ||
4705 (executable && event->attr.mmap))
4711 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4712 struct perf_mmap_event *mmap_event,
4715 struct perf_event *event;
4717 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4718 if (perf_event_mmap_match(event, mmap_event, executable))
4719 perf_event_mmap_output(event, mmap_event);
4723 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4725 struct perf_cpu_context *cpuctx;
4726 struct perf_event_context *ctx;
4727 struct vm_area_struct *vma = mmap_event->vma;
4728 struct file *file = vma->vm_file;
4736 memset(tmp, 0, sizeof(tmp));
4740 * d_path works from the end of the rb backwards, so we
4741 * need to add enough zero bytes after the string to handle
4742 * the 64bit alignment we do later.
4744 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4746 name = strncpy(tmp, "//enomem", sizeof(tmp));
4749 name = d_path(&file->f_path, buf, PATH_MAX);
4751 name = strncpy(tmp, "//toolong", sizeof(tmp));
4755 if (arch_vma_name(mmap_event->vma)) {
4756 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4758 tmp[sizeof(tmp) - 1] = '\0';
4763 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4765 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4766 vma->vm_end >= vma->vm_mm->brk) {
4767 name = strncpy(tmp, "[heap]", sizeof(tmp));
4769 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4770 vma->vm_end >= vma->vm_mm->start_stack) {
4771 name = strncpy(tmp, "[stack]", sizeof(tmp));
4775 name = strncpy(tmp, "//anon", sizeof(tmp));
4780 size = ALIGN(strlen(name)+1, sizeof(u64));
4782 mmap_event->file_name = name;
4783 mmap_event->file_size = size;
4785 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4788 list_for_each_entry_rcu(pmu, &pmus, entry) {
4789 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4790 if (cpuctx->unique_pmu != pmu)
4792 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4793 vma->vm_flags & VM_EXEC);
4795 ctxn = pmu->task_ctx_nr;
4799 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4801 perf_event_mmap_ctx(ctx, mmap_event,
4802 vma->vm_flags & VM_EXEC);
4805 put_cpu_ptr(pmu->pmu_cpu_context);
4812 void perf_event_mmap(struct vm_area_struct *vma)
4814 struct perf_mmap_event mmap_event;
4816 if (!atomic_read(&nr_mmap_events))
4819 mmap_event = (struct perf_mmap_event){
4825 .type = PERF_RECORD_MMAP,
4826 .misc = PERF_RECORD_MISC_USER,
4831 .start = vma->vm_start,
4832 .len = vma->vm_end - vma->vm_start,
4833 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4837 perf_event_mmap_event(&mmap_event);
4841 * IRQ throttle logging
4844 static void perf_log_throttle(struct perf_event *event, int enable)
4846 struct perf_output_handle handle;
4847 struct perf_sample_data sample;
4851 struct perf_event_header header;
4855 } throttle_event = {
4857 .type = PERF_RECORD_THROTTLE,
4859 .size = sizeof(throttle_event),
4861 .time = perf_clock(),
4862 .id = primary_event_id(event),
4863 .stream_id = event->id,
4867 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4869 perf_event_header__init_id(&throttle_event.header, &sample, event);
4871 ret = perf_output_begin(&handle, event,
4872 throttle_event.header.size);
4876 perf_output_put(&handle, throttle_event);
4877 perf_event__output_id_sample(event, &handle, &sample);
4878 perf_output_end(&handle);
4882 * Generic event overflow handling, sampling.
4885 static int __perf_event_overflow(struct perf_event *event,
4886 int throttle, struct perf_sample_data *data,
4887 struct pt_regs *regs)
4889 int events = atomic_read(&event->event_limit);
4890 struct hw_perf_event *hwc = &event->hw;
4895 * Non-sampling counters might still use the PMI to fold short
4896 * hardware counters, ignore those.
4898 if (unlikely(!is_sampling_event(event)))
4901 seq = __this_cpu_read(perf_throttled_seq);
4902 if (seq != hwc->interrupts_seq) {
4903 hwc->interrupts_seq = seq;
4904 hwc->interrupts = 1;
4907 if (unlikely(throttle
4908 && hwc->interrupts >= max_samples_per_tick)) {
4909 __this_cpu_inc(perf_throttled_count);
4910 hwc->interrupts = MAX_INTERRUPTS;
4911 perf_log_throttle(event, 0);
4916 if (event->attr.freq) {
4917 u64 now = perf_clock();
4918 s64 delta = now - hwc->freq_time_stamp;
4920 hwc->freq_time_stamp = now;
4922 if (delta > 0 && delta < 2*TICK_NSEC)
4923 perf_adjust_period(event, delta, hwc->last_period, true);
4927 * XXX event_limit might not quite work as expected on inherited
4931 event->pending_kill = POLL_IN;
4932 if (events && atomic_dec_and_test(&event->event_limit)) {
4934 event->pending_kill = POLL_HUP;
4935 event->pending_disable = 1;
4936 irq_work_queue(&event->pending);
4939 if (event->overflow_handler)
4940 event->overflow_handler(event, data, regs);
4942 perf_event_output(event, data, regs);
4944 if (event->fasync && event->pending_kill) {
4945 event->pending_wakeup = 1;
4946 irq_work_queue(&event->pending);
4952 int perf_event_overflow(struct perf_event *event,
4953 struct perf_sample_data *data,
4954 struct pt_regs *regs)
4956 return __perf_event_overflow(event, 1, data, regs);
4960 * Generic software event infrastructure
4963 struct swevent_htable {
4964 struct swevent_hlist *swevent_hlist;
4965 struct mutex hlist_mutex;
4968 /* Recursion avoidance in each contexts */
4969 int recursion[PERF_NR_CONTEXTS];
4972 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4975 * We directly increment event->count and keep a second value in
4976 * event->hw.period_left to count intervals. This period event
4977 * is kept in the range [-sample_period, 0] so that we can use the
4981 static u64 perf_swevent_set_period(struct perf_event *event)
4983 struct hw_perf_event *hwc = &event->hw;
4984 u64 period = hwc->last_period;
4988 hwc->last_period = hwc->sample_period;
4991 old = val = local64_read(&hwc->period_left);
4995 nr = div64_u64(period + val, period);
4996 offset = nr * period;
4998 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5004 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5005 struct perf_sample_data *data,
5006 struct pt_regs *regs)
5008 struct hw_perf_event *hwc = &event->hw;
5012 overflow = perf_swevent_set_period(event);
5014 if (hwc->interrupts == MAX_INTERRUPTS)
5017 for (; overflow; overflow--) {
5018 if (__perf_event_overflow(event, throttle,
5021 * We inhibit the overflow from happening when
5022 * hwc->interrupts == MAX_INTERRUPTS.
5030 static void perf_swevent_event(struct perf_event *event, u64 nr,
5031 struct perf_sample_data *data,
5032 struct pt_regs *regs)
5034 struct hw_perf_event *hwc = &event->hw;
5036 local64_add(nr, &event->count);
5041 if (!is_sampling_event(event))
5044 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5046 return perf_swevent_overflow(event, 1, data, regs);
5048 data->period = event->hw.last_period;
5050 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5051 return perf_swevent_overflow(event, 1, data, regs);
5053 if (local64_add_negative(nr, &hwc->period_left))
5056 perf_swevent_overflow(event, 0, data, regs);
5059 static int perf_exclude_event(struct perf_event *event,
5060 struct pt_regs *regs)
5062 if (event->hw.state & PERF_HES_STOPPED)
5066 if (event->attr.exclude_user && user_mode(regs))
5069 if (event->attr.exclude_kernel && !user_mode(regs))
5076 static int perf_swevent_match(struct perf_event *event,
5077 enum perf_type_id type,
5079 struct perf_sample_data *data,
5080 struct pt_regs *regs)
5082 if (event->attr.type != type)
5085 if (event->attr.config != event_id)
5088 if (perf_exclude_event(event, regs))
5094 static inline u64 swevent_hash(u64 type, u32 event_id)
5096 u64 val = event_id | (type << 32);
5098 return hash_64(val, SWEVENT_HLIST_BITS);
5101 static inline struct hlist_head *
5102 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5104 u64 hash = swevent_hash(type, event_id);
5106 return &hlist->heads[hash];
5109 /* For the read side: events when they trigger */
5110 static inline struct hlist_head *
5111 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5113 struct swevent_hlist *hlist;
5115 hlist = rcu_dereference(swhash->swevent_hlist);
5119 return __find_swevent_head(hlist, type, event_id);
5122 /* For the event head insertion and removal in the hlist */
5123 static inline struct hlist_head *
5124 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5126 struct swevent_hlist *hlist;
5127 u32 event_id = event->attr.config;
5128 u64 type = event->attr.type;
5131 * Event scheduling is always serialized against hlist allocation
5132 * and release. Which makes the protected version suitable here.
5133 * The context lock guarantees that.
5135 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5136 lockdep_is_held(&event->ctx->lock));
5140 return __find_swevent_head(hlist, type, event_id);
5143 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5145 struct perf_sample_data *data,
5146 struct pt_regs *regs)
5148 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5149 struct perf_event *event;
5150 struct hlist_head *head;
5153 head = find_swevent_head_rcu(swhash, type, event_id);
5157 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5158 if (perf_swevent_match(event, type, event_id, data, regs))
5159 perf_swevent_event(event, nr, data, regs);
5165 int perf_swevent_get_recursion_context(void)
5167 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5169 return get_recursion_context(swhash->recursion);
5171 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5173 inline void perf_swevent_put_recursion_context(int rctx)
5175 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5177 put_recursion_context(swhash->recursion, rctx);
5180 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5182 struct perf_sample_data data;
5185 preempt_disable_notrace();
5186 rctx = perf_swevent_get_recursion_context();
5190 perf_sample_data_init(&data, addr, 0);
5192 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5194 perf_swevent_put_recursion_context(rctx);
5195 preempt_enable_notrace();
5198 static void perf_swevent_read(struct perf_event *event)
5202 static int perf_swevent_add(struct perf_event *event, int flags)
5204 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5205 struct hw_perf_event *hwc = &event->hw;
5206 struct hlist_head *head;
5208 if (is_sampling_event(event)) {
5209 hwc->last_period = hwc->sample_period;
5210 perf_swevent_set_period(event);
5213 hwc->state = !(flags & PERF_EF_START);
5215 head = find_swevent_head(swhash, event);
5216 if (WARN_ON_ONCE(!head))
5219 hlist_add_head_rcu(&event->hlist_entry, head);
5224 static void perf_swevent_del(struct perf_event *event, int flags)
5226 hlist_del_rcu(&event->hlist_entry);
5229 static void perf_swevent_start(struct perf_event *event, int flags)
5231 event->hw.state = 0;
5234 static void perf_swevent_stop(struct perf_event *event, int flags)
5236 event->hw.state = PERF_HES_STOPPED;
5239 /* Deref the hlist from the update side */
5240 static inline struct swevent_hlist *
5241 swevent_hlist_deref(struct swevent_htable *swhash)
5243 return rcu_dereference_protected(swhash->swevent_hlist,
5244 lockdep_is_held(&swhash->hlist_mutex));
5247 static void swevent_hlist_release(struct swevent_htable *swhash)
5249 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5254 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5255 kfree_rcu(hlist, rcu_head);
5258 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5260 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5262 mutex_lock(&swhash->hlist_mutex);
5264 if (!--swhash->hlist_refcount)
5265 swevent_hlist_release(swhash);
5267 mutex_unlock(&swhash->hlist_mutex);
5270 static void swevent_hlist_put(struct perf_event *event)
5274 if (event->cpu != -1) {
5275 swevent_hlist_put_cpu(event, event->cpu);
5279 for_each_possible_cpu(cpu)
5280 swevent_hlist_put_cpu(event, cpu);
5283 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5285 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5288 mutex_lock(&swhash->hlist_mutex);
5290 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5291 struct swevent_hlist *hlist;
5293 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5298 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5300 swhash->hlist_refcount++;
5302 mutex_unlock(&swhash->hlist_mutex);
5307 static int swevent_hlist_get(struct perf_event *event)
5310 int cpu, failed_cpu;
5312 if (event->cpu != -1)
5313 return swevent_hlist_get_cpu(event, event->cpu);
5316 for_each_possible_cpu(cpu) {
5317 err = swevent_hlist_get_cpu(event, cpu);
5327 for_each_possible_cpu(cpu) {
5328 if (cpu == failed_cpu)
5330 swevent_hlist_put_cpu(event, cpu);
5337 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5339 static void sw_perf_event_destroy(struct perf_event *event)
5341 u64 event_id = event->attr.config;
5343 WARN_ON(event->parent);
5345 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5346 swevent_hlist_put(event);
5349 static int perf_swevent_init(struct perf_event *event)
5351 u64 event_id = event->attr.config;
5353 if (event->attr.type != PERF_TYPE_SOFTWARE)
5357 * no branch sampling for software events
5359 if (has_branch_stack(event))
5363 case PERF_COUNT_SW_CPU_CLOCK:
5364 case PERF_COUNT_SW_TASK_CLOCK:
5371 if (event_id >= PERF_COUNT_SW_MAX)
5374 if (!event->parent) {
5377 err = swevent_hlist_get(event);
5381 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5382 event->destroy = sw_perf_event_destroy;
5388 static int perf_swevent_event_idx(struct perf_event *event)
5393 static struct pmu perf_swevent = {
5394 .task_ctx_nr = perf_sw_context,
5396 .event_init = perf_swevent_init,
5397 .add = perf_swevent_add,
5398 .del = perf_swevent_del,
5399 .start = perf_swevent_start,
5400 .stop = perf_swevent_stop,
5401 .read = perf_swevent_read,
5403 .event_idx = perf_swevent_event_idx,
5406 #ifdef CONFIG_EVENT_TRACING
5408 static int perf_tp_filter_match(struct perf_event *event,
5409 struct perf_sample_data *data)
5411 void *record = data->raw->data;
5413 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5418 static int perf_tp_event_match(struct perf_event *event,
5419 struct perf_sample_data *data,
5420 struct pt_regs *regs)
5422 if (event->hw.state & PERF_HES_STOPPED)
5425 * All tracepoints are from kernel-space.
5427 if (event->attr.exclude_kernel)
5430 if (!perf_tp_filter_match(event, data))
5436 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5437 struct pt_regs *regs, struct hlist_head *head, int rctx,
5438 struct task_struct *task)
5440 struct perf_sample_data data;
5441 struct perf_event *event;
5443 struct perf_raw_record raw = {
5448 perf_sample_data_init(&data, addr, 0);
5451 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5452 if (perf_tp_event_match(event, &data, regs))
5453 perf_swevent_event(event, count, &data, regs);
5457 * If we got specified a target task, also iterate its context and
5458 * deliver this event there too.
5460 if (task && task != current) {
5461 struct perf_event_context *ctx;
5462 struct trace_entry *entry = record;
5465 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5469 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5470 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5472 if (event->attr.config != entry->type)
5474 if (perf_tp_event_match(event, &data, regs))
5475 perf_swevent_event(event, count, &data, regs);
5481 perf_swevent_put_recursion_context(rctx);
5483 EXPORT_SYMBOL_GPL(perf_tp_event);
5485 static void tp_perf_event_destroy(struct perf_event *event)
5487 perf_trace_destroy(event);
5490 static int perf_tp_event_init(struct perf_event *event)
5494 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5498 * no branch sampling for tracepoint events
5500 if (has_branch_stack(event))
5503 err = perf_trace_init(event);
5507 event->destroy = tp_perf_event_destroy;
5512 static struct pmu perf_tracepoint = {
5513 .task_ctx_nr = perf_sw_context,
5515 .event_init = perf_tp_event_init,
5516 .add = perf_trace_add,
5517 .del = perf_trace_del,
5518 .start = perf_swevent_start,
5519 .stop = perf_swevent_stop,
5520 .read = perf_swevent_read,
5522 .event_idx = perf_swevent_event_idx,
5525 static inline void perf_tp_register(void)
5527 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5530 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5535 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5538 filter_str = strndup_user(arg, PAGE_SIZE);
5539 if (IS_ERR(filter_str))
5540 return PTR_ERR(filter_str);
5542 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5548 static void perf_event_free_filter(struct perf_event *event)
5550 ftrace_profile_free_filter(event);
5555 static inline void perf_tp_register(void)
5559 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5564 static void perf_event_free_filter(struct perf_event *event)
5568 #endif /* CONFIG_EVENT_TRACING */
5570 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5571 void perf_bp_event(struct perf_event *bp, void *data)
5573 struct perf_sample_data sample;
5574 struct pt_regs *regs = data;
5576 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5578 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5579 perf_swevent_event(bp, 1, &sample, regs);
5584 * hrtimer based swevent callback
5587 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5589 enum hrtimer_restart ret = HRTIMER_RESTART;
5590 struct perf_sample_data data;
5591 struct pt_regs *regs;
5592 struct perf_event *event;
5595 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5597 if (event->state != PERF_EVENT_STATE_ACTIVE)
5598 return HRTIMER_NORESTART;
5600 event->pmu->read(event);
5602 perf_sample_data_init(&data, 0, event->hw.last_period);
5603 regs = get_irq_regs();
5605 if (regs && !perf_exclude_event(event, regs)) {
5606 if (!(event->attr.exclude_idle && is_idle_task(current)))
5607 if (__perf_event_overflow(event, 1, &data, regs))
5608 ret = HRTIMER_NORESTART;
5611 period = max_t(u64, 10000, event->hw.sample_period);
5612 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5617 static void perf_swevent_start_hrtimer(struct perf_event *event)
5619 struct hw_perf_event *hwc = &event->hw;
5622 if (!is_sampling_event(event))
5625 period = local64_read(&hwc->period_left);
5630 local64_set(&hwc->period_left, 0);
5632 period = max_t(u64, 10000, hwc->sample_period);
5634 __hrtimer_start_range_ns(&hwc->hrtimer,
5635 ns_to_ktime(period), 0,
5636 HRTIMER_MODE_REL_PINNED, 0);
5639 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5641 struct hw_perf_event *hwc = &event->hw;
5643 if (is_sampling_event(event)) {
5644 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5645 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5647 hrtimer_cancel(&hwc->hrtimer);
5651 static void perf_swevent_init_hrtimer(struct perf_event *event)
5653 struct hw_perf_event *hwc = &event->hw;
5655 if (!is_sampling_event(event))
5658 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5659 hwc->hrtimer.function = perf_swevent_hrtimer;
5662 * Since hrtimers have a fixed rate, we can do a static freq->period
5663 * mapping and avoid the whole period adjust feedback stuff.
5665 if (event->attr.freq) {
5666 long freq = event->attr.sample_freq;
5668 event->attr.sample_period = NSEC_PER_SEC / freq;
5669 hwc->sample_period = event->attr.sample_period;
5670 local64_set(&hwc->period_left, hwc->sample_period);
5671 hwc->last_period = hwc->sample_period;
5672 event->attr.freq = 0;
5677 * Software event: cpu wall time clock
5680 static void cpu_clock_event_update(struct perf_event *event)
5685 now = local_clock();
5686 prev = local64_xchg(&event->hw.prev_count, now);
5687 local64_add(now - prev, &event->count);
5690 static void cpu_clock_event_start(struct perf_event *event, int flags)
5692 local64_set(&event->hw.prev_count, local_clock());
5693 perf_swevent_start_hrtimer(event);
5696 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5698 perf_swevent_cancel_hrtimer(event);
5699 cpu_clock_event_update(event);
5702 static int cpu_clock_event_add(struct perf_event *event, int flags)
5704 if (flags & PERF_EF_START)
5705 cpu_clock_event_start(event, flags);
5710 static void cpu_clock_event_del(struct perf_event *event, int flags)
5712 cpu_clock_event_stop(event, flags);
5715 static void cpu_clock_event_read(struct perf_event *event)
5717 cpu_clock_event_update(event);
5720 static int cpu_clock_event_init(struct perf_event *event)
5722 if (event->attr.type != PERF_TYPE_SOFTWARE)
5725 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5729 * no branch sampling for software events
5731 if (has_branch_stack(event))
5734 perf_swevent_init_hrtimer(event);
5739 static struct pmu perf_cpu_clock = {
5740 .task_ctx_nr = perf_sw_context,
5742 .event_init = cpu_clock_event_init,
5743 .add = cpu_clock_event_add,
5744 .del = cpu_clock_event_del,
5745 .start = cpu_clock_event_start,
5746 .stop = cpu_clock_event_stop,
5747 .read = cpu_clock_event_read,
5749 .event_idx = perf_swevent_event_idx,
5753 * Software event: task time clock
5756 static void task_clock_event_update(struct perf_event *event, u64 now)
5761 prev = local64_xchg(&event->hw.prev_count, now);
5763 local64_add(delta, &event->count);
5766 static void task_clock_event_start(struct perf_event *event, int flags)
5768 local64_set(&event->hw.prev_count, event->ctx->time);
5769 perf_swevent_start_hrtimer(event);
5772 static void task_clock_event_stop(struct perf_event *event, int flags)
5774 perf_swevent_cancel_hrtimer(event);
5775 task_clock_event_update(event, event->ctx->time);
5778 static int task_clock_event_add(struct perf_event *event, int flags)
5780 if (flags & PERF_EF_START)
5781 task_clock_event_start(event, flags);
5786 static void task_clock_event_del(struct perf_event *event, int flags)
5788 task_clock_event_stop(event, PERF_EF_UPDATE);
5791 static void task_clock_event_read(struct perf_event *event)
5793 u64 now = perf_clock();
5794 u64 delta = now - event->ctx->timestamp;
5795 u64 time = event->ctx->time + delta;
5797 task_clock_event_update(event, time);
5800 static int task_clock_event_init(struct perf_event *event)
5802 if (event->attr.type != PERF_TYPE_SOFTWARE)
5805 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5809 * no branch sampling for software events
5811 if (has_branch_stack(event))
5814 perf_swevent_init_hrtimer(event);
5819 static struct pmu perf_task_clock = {
5820 .task_ctx_nr = perf_sw_context,
5822 .event_init = task_clock_event_init,
5823 .add = task_clock_event_add,
5824 .del = task_clock_event_del,
5825 .start = task_clock_event_start,
5826 .stop = task_clock_event_stop,
5827 .read = task_clock_event_read,
5829 .event_idx = perf_swevent_event_idx,
5832 static void perf_pmu_nop_void(struct pmu *pmu)
5836 static int perf_pmu_nop_int(struct pmu *pmu)
5841 static void perf_pmu_start_txn(struct pmu *pmu)
5843 perf_pmu_disable(pmu);
5846 static int perf_pmu_commit_txn(struct pmu *pmu)
5848 perf_pmu_enable(pmu);
5852 static void perf_pmu_cancel_txn(struct pmu *pmu)
5854 perf_pmu_enable(pmu);
5857 static int perf_event_idx_default(struct perf_event *event)
5859 return event->hw.idx + 1;
5863 * Ensures all contexts with the same task_ctx_nr have the same
5864 * pmu_cpu_context too.
5866 static void *find_pmu_context(int ctxn)
5873 list_for_each_entry(pmu, &pmus, entry) {
5874 if (pmu->task_ctx_nr == ctxn)
5875 return pmu->pmu_cpu_context;
5881 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5885 for_each_possible_cpu(cpu) {
5886 struct perf_cpu_context *cpuctx;
5888 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5890 if (cpuctx->unique_pmu == old_pmu)
5891 cpuctx->unique_pmu = pmu;
5895 static void free_pmu_context(struct pmu *pmu)
5899 mutex_lock(&pmus_lock);
5901 * Like a real lame refcount.
5903 list_for_each_entry(i, &pmus, entry) {
5904 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5905 update_pmu_context(i, pmu);
5910 free_percpu(pmu->pmu_cpu_context);
5912 mutex_unlock(&pmus_lock);
5914 static struct idr pmu_idr;
5917 type_show(struct device *dev, struct device_attribute *attr, char *page)
5919 struct pmu *pmu = dev_get_drvdata(dev);
5921 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5924 static struct device_attribute pmu_dev_attrs[] = {
5929 static int pmu_bus_running;
5930 static struct bus_type pmu_bus = {
5931 .name = "event_source",
5932 .dev_attrs = pmu_dev_attrs,
5935 static void pmu_dev_release(struct device *dev)
5940 static int pmu_dev_alloc(struct pmu *pmu)
5944 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5948 pmu->dev->groups = pmu->attr_groups;
5949 device_initialize(pmu->dev);
5950 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5954 dev_set_drvdata(pmu->dev, pmu);
5955 pmu->dev->bus = &pmu_bus;
5956 pmu->dev->release = pmu_dev_release;
5957 ret = device_add(pmu->dev);
5965 put_device(pmu->dev);
5969 static struct lock_class_key cpuctx_mutex;
5970 static struct lock_class_key cpuctx_lock;
5972 int perf_pmu_register(struct pmu *pmu, char *name, int type)
5976 mutex_lock(&pmus_lock);
5978 pmu->pmu_disable_count = alloc_percpu(int);
5979 if (!pmu->pmu_disable_count)
5988 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
5996 if (pmu_bus_running) {
5997 ret = pmu_dev_alloc(pmu);
6003 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6004 if (pmu->pmu_cpu_context)
6005 goto got_cpu_context;
6008 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6009 if (!pmu->pmu_cpu_context)
6012 for_each_possible_cpu(cpu) {
6013 struct perf_cpu_context *cpuctx;
6015 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6016 __perf_event_init_context(&cpuctx->ctx);
6017 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6018 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6019 cpuctx->ctx.type = cpu_context;
6020 cpuctx->ctx.pmu = pmu;
6021 cpuctx->jiffies_interval = 1;
6022 INIT_LIST_HEAD(&cpuctx->rotation_list);
6023 cpuctx->unique_pmu = pmu;
6027 if (!pmu->start_txn) {
6028 if (pmu->pmu_enable) {
6030 * If we have pmu_enable/pmu_disable calls, install
6031 * transaction stubs that use that to try and batch
6032 * hardware accesses.
6034 pmu->start_txn = perf_pmu_start_txn;
6035 pmu->commit_txn = perf_pmu_commit_txn;
6036 pmu->cancel_txn = perf_pmu_cancel_txn;
6038 pmu->start_txn = perf_pmu_nop_void;
6039 pmu->commit_txn = perf_pmu_nop_int;
6040 pmu->cancel_txn = perf_pmu_nop_void;
6044 if (!pmu->pmu_enable) {
6045 pmu->pmu_enable = perf_pmu_nop_void;
6046 pmu->pmu_disable = perf_pmu_nop_void;
6049 if (!pmu->event_idx)
6050 pmu->event_idx = perf_event_idx_default;
6052 list_add_rcu(&pmu->entry, &pmus);
6055 mutex_unlock(&pmus_lock);
6060 device_del(pmu->dev);
6061 put_device(pmu->dev);
6064 if (pmu->type >= PERF_TYPE_MAX)
6065 idr_remove(&pmu_idr, pmu->type);
6068 free_percpu(pmu->pmu_disable_count);
6072 void perf_pmu_unregister(struct pmu *pmu)
6074 mutex_lock(&pmus_lock);
6075 list_del_rcu(&pmu->entry);
6076 mutex_unlock(&pmus_lock);
6079 * We dereference the pmu list under both SRCU and regular RCU, so
6080 * synchronize against both of those.
6082 synchronize_srcu(&pmus_srcu);
6085 free_percpu(pmu->pmu_disable_count);
6086 if (pmu->type >= PERF_TYPE_MAX)
6087 idr_remove(&pmu_idr, pmu->type);
6088 device_del(pmu->dev);
6089 put_device(pmu->dev);
6090 free_pmu_context(pmu);
6093 struct pmu *perf_init_event(struct perf_event *event)
6095 struct pmu *pmu = NULL;
6099 idx = srcu_read_lock(&pmus_srcu);
6102 pmu = idr_find(&pmu_idr, event->attr.type);
6106 ret = pmu->event_init(event);
6112 list_for_each_entry_rcu(pmu, &pmus, entry) {
6114 ret = pmu->event_init(event);
6118 if (ret != -ENOENT) {
6123 pmu = ERR_PTR(-ENOENT);
6125 srcu_read_unlock(&pmus_srcu, idx);
6131 * Allocate and initialize a event structure
6133 static struct perf_event *
6134 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6135 struct task_struct *task,
6136 struct perf_event *group_leader,
6137 struct perf_event *parent_event,
6138 perf_overflow_handler_t overflow_handler,
6142 struct perf_event *event;
6143 struct hw_perf_event *hwc;
6146 if ((unsigned)cpu >= nr_cpu_ids) {
6147 if (!task || cpu != -1)
6148 return ERR_PTR(-EINVAL);
6151 event = kzalloc(sizeof(*event), GFP_KERNEL);
6153 return ERR_PTR(-ENOMEM);
6156 * Single events are their own group leaders, with an
6157 * empty sibling list:
6160 group_leader = event;
6162 mutex_init(&event->child_mutex);
6163 INIT_LIST_HEAD(&event->child_list);
6165 INIT_LIST_HEAD(&event->group_entry);
6166 INIT_LIST_HEAD(&event->event_entry);
6167 INIT_LIST_HEAD(&event->sibling_list);
6168 INIT_LIST_HEAD(&event->rb_entry);
6170 init_waitqueue_head(&event->waitq);
6171 init_irq_work(&event->pending, perf_pending_event);
6173 mutex_init(&event->mmap_mutex);
6175 atomic_long_set(&event->refcount, 1);
6177 event->attr = *attr;
6178 event->group_leader = group_leader;
6182 event->parent = parent_event;
6184 event->ns = get_pid_ns(task_active_pid_ns(current));
6185 event->id = atomic64_inc_return(&perf_event_id);
6187 event->state = PERF_EVENT_STATE_INACTIVE;
6190 event->attach_state = PERF_ATTACH_TASK;
6192 if (attr->type == PERF_TYPE_TRACEPOINT)
6193 event->hw.tp_target = task;
6194 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6196 * hw_breakpoint is a bit difficult here..
6198 else if (attr->type == PERF_TYPE_BREAKPOINT)
6199 event->hw.bp_target = task;
6203 if (!overflow_handler && parent_event) {
6204 overflow_handler = parent_event->overflow_handler;
6205 context = parent_event->overflow_handler_context;
6208 event->overflow_handler = overflow_handler;
6209 event->overflow_handler_context = context;
6211 perf_event__state_init(event);
6216 hwc->sample_period = attr->sample_period;
6217 if (attr->freq && attr->sample_freq)
6218 hwc->sample_period = 1;
6219 hwc->last_period = hwc->sample_period;
6221 local64_set(&hwc->period_left, hwc->sample_period);
6224 * we currently do not support PERF_FORMAT_GROUP on inherited events
6226 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6229 pmu = perf_init_event(event);
6235 else if (IS_ERR(pmu))
6240 put_pid_ns(event->ns);
6242 return ERR_PTR(err);
6245 if (!event->parent) {
6246 if (event->attach_state & PERF_ATTACH_TASK)
6247 static_key_slow_inc(&perf_sched_events.key);
6248 if (event->attr.mmap || event->attr.mmap_data)
6249 atomic_inc(&nr_mmap_events);
6250 if (event->attr.comm)
6251 atomic_inc(&nr_comm_events);
6252 if (event->attr.task)
6253 atomic_inc(&nr_task_events);
6254 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6255 err = get_callchain_buffers();
6258 return ERR_PTR(err);
6261 if (has_branch_stack(event)) {
6262 static_key_slow_inc(&perf_sched_events.key);
6263 if (!(event->attach_state & PERF_ATTACH_TASK))
6264 atomic_inc(&per_cpu(perf_branch_stack_events,
6272 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6273 struct perf_event_attr *attr)
6278 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6282 * zero the full structure, so that a short copy will be nice.
6284 memset(attr, 0, sizeof(*attr));
6286 ret = get_user(size, &uattr->size);
6290 if (size > PAGE_SIZE) /* silly large */
6293 if (!size) /* abi compat */
6294 size = PERF_ATTR_SIZE_VER0;
6296 if (size < PERF_ATTR_SIZE_VER0)
6300 * If we're handed a bigger struct than we know of,
6301 * ensure all the unknown bits are 0 - i.e. new
6302 * user-space does not rely on any kernel feature
6303 * extensions we dont know about yet.
6305 if (size > sizeof(*attr)) {
6306 unsigned char __user *addr;
6307 unsigned char __user *end;
6310 addr = (void __user *)uattr + sizeof(*attr);
6311 end = (void __user *)uattr + size;
6313 for (; addr < end; addr++) {
6314 ret = get_user(val, addr);
6320 size = sizeof(*attr);
6323 ret = copy_from_user(attr, uattr, size);
6327 if (attr->__reserved_1)
6330 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6333 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6336 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6337 u64 mask = attr->branch_sample_type;
6339 /* only using defined bits */
6340 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6343 /* at least one branch bit must be set */
6344 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6347 /* kernel level capture: check permissions */
6348 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6349 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6352 /* propagate priv level, when not set for branch */
6353 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6355 /* exclude_kernel checked on syscall entry */
6356 if (!attr->exclude_kernel)
6357 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6359 if (!attr->exclude_user)
6360 mask |= PERF_SAMPLE_BRANCH_USER;
6362 if (!attr->exclude_hv)
6363 mask |= PERF_SAMPLE_BRANCH_HV;
6365 * adjust user setting (for HW filter setup)
6367 attr->branch_sample_type = mask;
6371 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6372 ret = perf_reg_validate(attr->sample_regs_user);
6377 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6378 if (!arch_perf_have_user_stack_dump())
6382 * We have __u32 type for the size, but so far
6383 * we can only use __u16 as maximum due to the
6384 * __u16 sample size limit.
6386 if (attr->sample_stack_user >= USHRT_MAX)
6388 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6396 put_user(sizeof(*attr), &uattr->size);
6402 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6404 struct ring_buffer *rb = NULL, *old_rb = NULL;
6410 /* don't allow circular references */
6411 if (event == output_event)
6415 * Don't allow cross-cpu buffers
6417 if (output_event->cpu != event->cpu)
6421 * If its not a per-cpu rb, it must be the same task.
6423 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6427 mutex_lock(&event->mmap_mutex);
6428 /* Can't redirect output if we've got an active mmap() */
6429 if (atomic_read(&event->mmap_count))
6433 /* get the rb we want to redirect to */
6434 rb = ring_buffer_get(output_event);
6440 rcu_assign_pointer(event->rb, rb);
6442 ring_buffer_detach(event, old_rb);
6445 mutex_unlock(&event->mmap_mutex);
6448 ring_buffer_put(old_rb);
6454 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6456 * @attr_uptr: event_id type attributes for monitoring/sampling
6459 * @group_fd: group leader event fd
6461 SYSCALL_DEFINE5(perf_event_open,
6462 struct perf_event_attr __user *, attr_uptr,
6463 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6465 struct perf_event *group_leader = NULL, *output_event = NULL;
6466 struct perf_event *event, *sibling;
6467 struct perf_event_attr attr;
6468 struct perf_event_context *ctx;
6469 struct file *event_file = NULL;
6470 struct fd group = {NULL, 0};
6471 struct task_struct *task = NULL;
6477 /* for future expandability... */
6478 if (flags & ~PERF_FLAG_ALL)
6481 err = perf_copy_attr(attr_uptr, &attr);
6485 if (!attr.exclude_kernel) {
6486 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6491 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6496 * In cgroup mode, the pid argument is used to pass the fd
6497 * opened to the cgroup directory in cgroupfs. The cpu argument
6498 * designates the cpu on which to monitor threads from that
6501 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6504 event_fd = get_unused_fd();
6508 if (group_fd != -1) {
6509 err = perf_fget_light(group_fd, &group);
6512 group_leader = group.file->private_data;
6513 if (flags & PERF_FLAG_FD_OUTPUT)
6514 output_event = group_leader;
6515 if (flags & PERF_FLAG_FD_NO_GROUP)
6516 group_leader = NULL;
6519 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6520 task = find_lively_task_by_vpid(pid);
6522 err = PTR_ERR(task);
6529 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6531 if (IS_ERR(event)) {
6532 err = PTR_ERR(event);
6536 if (flags & PERF_FLAG_PID_CGROUP) {
6537 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6542 * - that has cgroup constraint on event->cpu
6543 * - that may need work on context switch
6545 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6546 static_key_slow_inc(&perf_sched_events.key);
6550 * Special case software events and allow them to be part of
6551 * any hardware group.
6556 (is_software_event(event) != is_software_event(group_leader))) {
6557 if (is_software_event(event)) {
6559 * If event and group_leader are not both a software
6560 * event, and event is, then group leader is not.
6562 * Allow the addition of software events to !software
6563 * groups, this is safe because software events never
6566 pmu = group_leader->pmu;
6567 } else if (is_software_event(group_leader) &&
6568 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6570 * In case the group is a pure software group, and we
6571 * try to add a hardware event, move the whole group to
6572 * the hardware context.
6579 * Get the target context (task or percpu):
6581 ctx = find_get_context(pmu, task, event->cpu);
6588 put_task_struct(task);
6593 * Look up the group leader (we will attach this event to it):
6599 * Do not allow a recursive hierarchy (this new sibling
6600 * becoming part of another group-sibling):
6602 if (group_leader->group_leader != group_leader)
6605 * Do not allow to attach to a group in a different
6606 * task or CPU context:
6609 if (group_leader->ctx->type != ctx->type)
6612 if (group_leader->ctx != ctx)
6617 * Only a group leader can be exclusive or pinned
6619 if (attr.exclusive || attr.pinned)
6624 err = perf_event_set_output(event, output_event);
6629 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6630 if (IS_ERR(event_file)) {
6631 err = PTR_ERR(event_file);
6636 struct perf_event_context *gctx = group_leader->ctx;
6638 mutex_lock(&gctx->mutex);
6639 perf_remove_from_context(group_leader);
6642 * Removing from the context ends up with disabled
6643 * event. What we want here is event in the initial
6644 * startup state, ready to be add into new context.
6646 perf_event__state_init(group_leader);
6647 list_for_each_entry(sibling, &group_leader->sibling_list,
6649 perf_remove_from_context(sibling);
6650 perf_event__state_init(sibling);
6653 mutex_unlock(&gctx->mutex);
6657 WARN_ON_ONCE(ctx->parent_ctx);
6658 mutex_lock(&ctx->mutex);
6662 perf_install_in_context(ctx, group_leader, event->cpu);
6664 list_for_each_entry(sibling, &group_leader->sibling_list,
6666 perf_install_in_context(ctx, sibling, event->cpu);
6671 perf_install_in_context(ctx, event, event->cpu);
6673 perf_unpin_context(ctx);
6674 mutex_unlock(&ctx->mutex);
6678 event->owner = current;
6680 mutex_lock(¤t->perf_event_mutex);
6681 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6682 mutex_unlock(¤t->perf_event_mutex);
6685 * Precalculate sample_data sizes
6687 perf_event__header_size(event);
6688 perf_event__id_header_size(event);
6691 * Drop the reference on the group_event after placing the
6692 * new event on the sibling_list. This ensures destruction
6693 * of the group leader will find the pointer to itself in
6694 * perf_group_detach().
6697 fd_install(event_fd, event_file);
6701 perf_unpin_context(ctx);
6708 put_task_struct(task);
6712 put_unused_fd(event_fd);
6717 * perf_event_create_kernel_counter
6719 * @attr: attributes of the counter to create
6720 * @cpu: cpu in which the counter is bound
6721 * @task: task to profile (NULL for percpu)
6724 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6725 struct task_struct *task,
6726 perf_overflow_handler_t overflow_handler,
6729 struct perf_event_context *ctx;
6730 struct perf_event *event;
6734 * Get the target context (task or percpu):
6737 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6738 overflow_handler, context);
6739 if (IS_ERR(event)) {
6740 err = PTR_ERR(event);
6744 ctx = find_get_context(event->pmu, task, cpu);
6750 WARN_ON_ONCE(ctx->parent_ctx);
6751 mutex_lock(&ctx->mutex);
6752 perf_install_in_context(ctx, event, cpu);
6754 perf_unpin_context(ctx);
6755 mutex_unlock(&ctx->mutex);
6762 return ERR_PTR(err);
6764 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6766 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
6768 struct perf_event_context *src_ctx;
6769 struct perf_event_context *dst_ctx;
6770 struct perf_event *event, *tmp;
6773 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
6774 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
6776 mutex_lock(&src_ctx->mutex);
6777 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
6779 perf_remove_from_context(event);
6781 list_add(&event->event_entry, &events);
6783 mutex_unlock(&src_ctx->mutex);
6787 mutex_lock(&dst_ctx->mutex);
6788 list_for_each_entry_safe(event, tmp, &events, event_entry) {
6789 list_del(&event->event_entry);
6790 if (event->state >= PERF_EVENT_STATE_OFF)
6791 event->state = PERF_EVENT_STATE_INACTIVE;
6792 perf_install_in_context(dst_ctx, event, dst_cpu);
6795 mutex_unlock(&dst_ctx->mutex);
6797 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
6799 static void sync_child_event(struct perf_event *child_event,
6800 struct task_struct *child)
6802 struct perf_event *parent_event = child_event->parent;
6805 if (child_event->attr.inherit_stat)
6806 perf_event_read_event(child_event, child);
6808 child_val = perf_event_count(child_event);
6811 * Add back the child's count to the parent's count:
6813 atomic64_add(child_val, &parent_event->child_count);
6814 atomic64_add(child_event->total_time_enabled,
6815 &parent_event->child_total_time_enabled);
6816 atomic64_add(child_event->total_time_running,
6817 &parent_event->child_total_time_running);
6820 * Remove this event from the parent's list
6822 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6823 mutex_lock(&parent_event->child_mutex);
6824 list_del_init(&child_event->child_list);
6825 mutex_unlock(&parent_event->child_mutex);
6828 * Release the parent event, if this was the last
6831 put_event(parent_event);
6835 __perf_event_exit_task(struct perf_event *child_event,
6836 struct perf_event_context *child_ctx,
6837 struct task_struct *child)
6839 if (child_event->parent) {
6840 raw_spin_lock_irq(&child_ctx->lock);
6841 perf_group_detach(child_event);
6842 raw_spin_unlock_irq(&child_ctx->lock);
6845 perf_remove_from_context(child_event);
6848 * It can happen that the parent exits first, and has events
6849 * that are still around due to the child reference. These
6850 * events need to be zapped.
6852 if (child_event->parent) {
6853 sync_child_event(child_event, child);
6854 free_event(child_event);
6858 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6860 struct perf_event *child_event, *tmp;
6861 struct perf_event_context *child_ctx;
6862 unsigned long flags;
6864 if (likely(!child->perf_event_ctxp[ctxn])) {
6865 perf_event_task(child, NULL, 0);
6869 local_irq_save(flags);
6871 * We can't reschedule here because interrupts are disabled,
6872 * and either child is current or it is a task that can't be
6873 * scheduled, so we are now safe from rescheduling changing
6876 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6879 * Take the context lock here so that if find_get_context is
6880 * reading child->perf_event_ctxp, we wait until it has
6881 * incremented the context's refcount before we do put_ctx below.
6883 raw_spin_lock(&child_ctx->lock);
6884 task_ctx_sched_out(child_ctx);
6885 child->perf_event_ctxp[ctxn] = NULL;
6887 * If this context is a clone; unclone it so it can't get
6888 * swapped to another process while we're removing all
6889 * the events from it.
6891 unclone_ctx(child_ctx);
6892 update_context_time(child_ctx);
6893 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6896 * Report the task dead after unscheduling the events so that we
6897 * won't get any samples after PERF_RECORD_EXIT. We can however still
6898 * get a few PERF_RECORD_READ events.
6900 perf_event_task(child, child_ctx, 0);
6903 * We can recurse on the same lock type through:
6905 * __perf_event_exit_task()
6906 * sync_child_event()
6908 * mutex_lock(&ctx->mutex)
6910 * But since its the parent context it won't be the same instance.
6912 mutex_lock(&child_ctx->mutex);
6915 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6917 __perf_event_exit_task(child_event, child_ctx, child);
6919 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6921 __perf_event_exit_task(child_event, child_ctx, child);
6924 * If the last event was a group event, it will have appended all
6925 * its siblings to the list, but we obtained 'tmp' before that which
6926 * will still point to the list head terminating the iteration.
6928 if (!list_empty(&child_ctx->pinned_groups) ||
6929 !list_empty(&child_ctx->flexible_groups))
6932 mutex_unlock(&child_ctx->mutex);
6938 * When a child task exits, feed back event values to parent events.
6940 void perf_event_exit_task(struct task_struct *child)
6942 struct perf_event *event, *tmp;
6945 mutex_lock(&child->perf_event_mutex);
6946 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6948 list_del_init(&event->owner_entry);
6951 * Ensure the list deletion is visible before we clear
6952 * the owner, closes a race against perf_release() where
6953 * we need to serialize on the owner->perf_event_mutex.
6956 event->owner = NULL;
6958 mutex_unlock(&child->perf_event_mutex);
6960 for_each_task_context_nr(ctxn)
6961 perf_event_exit_task_context(child, ctxn);
6964 static void perf_free_event(struct perf_event *event,
6965 struct perf_event_context *ctx)
6967 struct perf_event *parent = event->parent;
6969 if (WARN_ON_ONCE(!parent))
6972 mutex_lock(&parent->child_mutex);
6973 list_del_init(&event->child_list);
6974 mutex_unlock(&parent->child_mutex);
6978 perf_group_detach(event);
6979 list_del_event(event, ctx);
6984 * free an unexposed, unused context as created by inheritance by
6985 * perf_event_init_task below, used by fork() in case of fail.
6987 void perf_event_free_task(struct task_struct *task)
6989 struct perf_event_context *ctx;
6990 struct perf_event *event, *tmp;
6993 for_each_task_context_nr(ctxn) {
6994 ctx = task->perf_event_ctxp[ctxn];
6998 mutex_lock(&ctx->mutex);
7000 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7002 perf_free_event(event, ctx);
7004 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7006 perf_free_event(event, ctx);
7008 if (!list_empty(&ctx->pinned_groups) ||
7009 !list_empty(&ctx->flexible_groups))
7012 mutex_unlock(&ctx->mutex);
7018 void perf_event_delayed_put(struct task_struct *task)
7022 for_each_task_context_nr(ctxn)
7023 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7027 * inherit a event from parent task to child task:
7029 static struct perf_event *
7030 inherit_event(struct perf_event *parent_event,
7031 struct task_struct *parent,
7032 struct perf_event_context *parent_ctx,
7033 struct task_struct *child,
7034 struct perf_event *group_leader,
7035 struct perf_event_context *child_ctx)
7037 struct perf_event *child_event;
7038 unsigned long flags;
7041 * Instead of creating recursive hierarchies of events,
7042 * we link inherited events back to the original parent,
7043 * which has a filp for sure, which we use as the reference
7046 if (parent_event->parent)
7047 parent_event = parent_event->parent;
7049 child_event = perf_event_alloc(&parent_event->attr,
7052 group_leader, parent_event,
7054 if (IS_ERR(child_event))
7057 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7058 free_event(child_event);
7065 * Make the child state follow the state of the parent event,
7066 * not its attr.disabled bit. We hold the parent's mutex,
7067 * so we won't race with perf_event_{en, dis}able_family.
7069 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7070 child_event->state = PERF_EVENT_STATE_INACTIVE;
7072 child_event->state = PERF_EVENT_STATE_OFF;
7074 if (parent_event->attr.freq) {
7075 u64 sample_period = parent_event->hw.sample_period;
7076 struct hw_perf_event *hwc = &child_event->hw;
7078 hwc->sample_period = sample_period;
7079 hwc->last_period = sample_period;
7081 local64_set(&hwc->period_left, sample_period);
7084 child_event->ctx = child_ctx;
7085 child_event->overflow_handler = parent_event->overflow_handler;
7086 child_event->overflow_handler_context
7087 = parent_event->overflow_handler_context;
7090 * Precalculate sample_data sizes
7092 perf_event__header_size(child_event);
7093 perf_event__id_header_size(child_event);
7096 * Link it up in the child's context:
7098 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7099 add_event_to_ctx(child_event, child_ctx);
7100 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7103 * Link this into the parent event's child list
7105 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7106 mutex_lock(&parent_event->child_mutex);
7107 list_add_tail(&child_event->child_list, &parent_event->child_list);
7108 mutex_unlock(&parent_event->child_mutex);
7113 static int inherit_group(struct perf_event *parent_event,
7114 struct task_struct *parent,
7115 struct perf_event_context *parent_ctx,
7116 struct task_struct *child,
7117 struct perf_event_context *child_ctx)
7119 struct perf_event *leader;
7120 struct perf_event *sub;
7121 struct perf_event *child_ctr;
7123 leader = inherit_event(parent_event, parent, parent_ctx,
7124 child, NULL, child_ctx);
7126 return PTR_ERR(leader);
7127 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7128 child_ctr = inherit_event(sub, parent, parent_ctx,
7129 child, leader, child_ctx);
7130 if (IS_ERR(child_ctr))
7131 return PTR_ERR(child_ctr);
7137 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7138 struct perf_event_context *parent_ctx,
7139 struct task_struct *child, int ctxn,
7143 struct perf_event_context *child_ctx;
7145 if (!event->attr.inherit) {
7150 child_ctx = child->perf_event_ctxp[ctxn];
7153 * This is executed from the parent task context, so
7154 * inherit events that have been marked for cloning.
7155 * First allocate and initialize a context for the
7159 child_ctx = alloc_perf_context(event->pmu, child);
7163 child->perf_event_ctxp[ctxn] = child_ctx;
7166 ret = inherit_group(event, parent, parent_ctx,
7176 * Initialize the perf_event context in task_struct
7178 int perf_event_init_context(struct task_struct *child, int ctxn)
7180 struct perf_event_context *child_ctx, *parent_ctx;
7181 struct perf_event_context *cloned_ctx;
7182 struct perf_event *event;
7183 struct task_struct *parent = current;
7184 int inherited_all = 1;
7185 unsigned long flags;
7188 if (likely(!parent->perf_event_ctxp[ctxn]))
7192 * If the parent's context is a clone, pin it so it won't get
7195 parent_ctx = perf_pin_task_context(parent, ctxn);
7198 * No need to check if parent_ctx != NULL here; since we saw
7199 * it non-NULL earlier, the only reason for it to become NULL
7200 * is if we exit, and since we're currently in the middle of
7201 * a fork we can't be exiting at the same time.
7205 * Lock the parent list. No need to lock the child - not PID
7206 * hashed yet and not running, so nobody can access it.
7208 mutex_lock(&parent_ctx->mutex);
7211 * We dont have to disable NMIs - we are only looking at
7212 * the list, not manipulating it:
7214 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7215 ret = inherit_task_group(event, parent, parent_ctx,
7216 child, ctxn, &inherited_all);
7222 * We can't hold ctx->lock when iterating the ->flexible_group list due
7223 * to allocations, but we need to prevent rotation because
7224 * rotate_ctx() will change the list from interrupt context.
7226 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7227 parent_ctx->rotate_disable = 1;
7228 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7230 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7231 ret = inherit_task_group(event, parent, parent_ctx,
7232 child, ctxn, &inherited_all);
7237 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7238 parent_ctx->rotate_disable = 0;
7240 child_ctx = child->perf_event_ctxp[ctxn];
7242 if (child_ctx && inherited_all) {
7244 * Mark the child context as a clone of the parent
7245 * context, or of whatever the parent is a clone of.
7247 * Note that if the parent is a clone, the holding of
7248 * parent_ctx->lock avoids it from being uncloned.
7250 cloned_ctx = parent_ctx->parent_ctx;
7252 child_ctx->parent_ctx = cloned_ctx;
7253 child_ctx->parent_gen = parent_ctx->parent_gen;
7255 child_ctx->parent_ctx = parent_ctx;
7256 child_ctx->parent_gen = parent_ctx->generation;
7258 get_ctx(child_ctx->parent_ctx);
7261 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7262 mutex_unlock(&parent_ctx->mutex);
7264 perf_unpin_context(parent_ctx);
7265 put_ctx(parent_ctx);
7271 * Initialize the perf_event context in task_struct
7273 int perf_event_init_task(struct task_struct *child)
7277 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7278 mutex_init(&child->perf_event_mutex);
7279 INIT_LIST_HEAD(&child->perf_event_list);
7281 for_each_task_context_nr(ctxn) {
7282 ret = perf_event_init_context(child, ctxn);
7290 static void __init perf_event_init_all_cpus(void)
7292 struct swevent_htable *swhash;
7295 for_each_possible_cpu(cpu) {
7296 swhash = &per_cpu(swevent_htable, cpu);
7297 mutex_init(&swhash->hlist_mutex);
7298 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7302 static void __cpuinit perf_event_init_cpu(int cpu)
7304 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7306 mutex_lock(&swhash->hlist_mutex);
7307 if (swhash->hlist_refcount > 0) {
7308 struct swevent_hlist *hlist;
7310 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7312 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7314 mutex_unlock(&swhash->hlist_mutex);
7317 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7318 static void perf_pmu_rotate_stop(struct pmu *pmu)
7320 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7322 WARN_ON(!irqs_disabled());
7324 list_del_init(&cpuctx->rotation_list);
7327 static void __perf_event_exit_context(void *__info)
7329 struct perf_event_context *ctx = __info;
7330 struct perf_event *event, *tmp;
7332 perf_pmu_rotate_stop(ctx->pmu);
7334 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7335 __perf_remove_from_context(event);
7336 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7337 __perf_remove_from_context(event);
7340 static void perf_event_exit_cpu_context(int cpu)
7342 struct perf_event_context *ctx;
7346 idx = srcu_read_lock(&pmus_srcu);
7347 list_for_each_entry_rcu(pmu, &pmus, entry) {
7348 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7350 mutex_lock(&ctx->mutex);
7351 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7352 mutex_unlock(&ctx->mutex);
7354 srcu_read_unlock(&pmus_srcu, idx);
7357 static void perf_event_exit_cpu(int cpu)
7359 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7361 mutex_lock(&swhash->hlist_mutex);
7362 swevent_hlist_release(swhash);
7363 mutex_unlock(&swhash->hlist_mutex);
7365 perf_event_exit_cpu_context(cpu);
7368 static inline void perf_event_exit_cpu(int cpu) { }
7372 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7376 for_each_online_cpu(cpu)
7377 perf_event_exit_cpu(cpu);
7383 * Run the perf reboot notifier at the very last possible moment so that
7384 * the generic watchdog code runs as long as possible.
7386 static struct notifier_block perf_reboot_notifier = {
7387 .notifier_call = perf_reboot,
7388 .priority = INT_MIN,
7391 static int __cpuinit
7392 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7394 unsigned int cpu = (long)hcpu;
7396 switch (action & ~CPU_TASKS_FROZEN) {
7398 case CPU_UP_PREPARE:
7399 case CPU_DOWN_FAILED:
7400 perf_event_init_cpu(cpu);
7403 case CPU_UP_CANCELED:
7404 case CPU_DOWN_PREPARE:
7405 perf_event_exit_cpu(cpu);
7415 void __init perf_event_init(void)
7421 perf_event_init_all_cpus();
7422 init_srcu_struct(&pmus_srcu);
7423 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7424 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7425 perf_pmu_register(&perf_task_clock, NULL, -1);
7427 perf_cpu_notifier(perf_cpu_notify);
7428 register_reboot_notifier(&perf_reboot_notifier);
7430 ret = init_hw_breakpoint();
7431 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7433 /* do not patch jump label more than once per second */
7434 jump_label_rate_limit(&perf_sched_events, HZ);
7437 * Build time assertion that we keep the data_head at the intended
7438 * location. IOW, validation we got the __reserved[] size right.
7440 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7444 static int __init perf_event_sysfs_init(void)
7449 mutex_lock(&pmus_lock);
7451 ret = bus_register(&pmu_bus);
7455 list_for_each_entry(pmu, &pmus, entry) {
7456 if (!pmu->name || pmu->type < 0)
7459 ret = pmu_dev_alloc(pmu);
7460 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7462 pmu_bus_running = 1;
7466 mutex_unlock(&pmus_lock);
7470 device_initcall(perf_event_sysfs_init);
7472 #ifdef CONFIG_CGROUP_PERF
7473 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7475 struct perf_cgroup *jc;
7477 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7479 return ERR_PTR(-ENOMEM);
7481 jc->info = alloc_percpu(struct perf_cgroup_info);
7484 return ERR_PTR(-ENOMEM);
7490 static void perf_cgroup_css_free(struct cgroup *cont)
7492 struct perf_cgroup *jc;
7493 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7494 struct perf_cgroup, css);
7495 free_percpu(jc->info);
7499 static int __perf_cgroup_move(void *info)
7501 struct task_struct *task = info;
7502 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7506 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7508 struct task_struct *task;
7510 cgroup_taskset_for_each(task, cgrp, tset)
7511 task_function_call(task, __perf_cgroup_move, task);
7514 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7515 struct task_struct *task)
7518 * cgroup_exit() is called in the copy_process() failure path.
7519 * Ignore this case since the task hasn't ran yet, this avoids
7520 * trying to poke a half freed task state from generic code.
7522 if (!(task->flags & PF_EXITING))
7525 task_function_call(task, __perf_cgroup_move, task);
7528 struct cgroup_subsys perf_subsys = {
7529 .name = "perf_event",
7530 .subsys_id = perf_subsys_id,
7531 .css_alloc = perf_cgroup_css_alloc,
7532 .css_free = perf_cgroup_css_free,
7533 .exit = perf_cgroup_exit,
7534 .attach = perf_cgroup_attach,
7536 #endif /* CONFIG_CGROUP_PERF */