2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP)
125 * branch priv levels that need permission checks
127 #define PERF_SAMPLE_BRANCH_PERM_PLM \
128 (PERF_SAMPLE_BRANCH_KERNEL |\
129 PERF_SAMPLE_BRANCH_HV)
132 EVENT_FLEXIBLE = 0x1,
134 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
138 * perf_sched_events : >0 events exist
139 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 struct static_key_deferred perf_sched_events __read_mostly;
142 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
143 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
145 static atomic_t nr_mmap_events __read_mostly;
146 static atomic_t nr_comm_events __read_mostly;
147 static atomic_t nr_task_events __read_mostly;
149 static LIST_HEAD(pmus);
150 static DEFINE_MUTEX(pmus_lock);
151 static struct srcu_struct pmus_srcu;
154 * perf event paranoia level:
155 * -1 - not paranoid at all
156 * 0 - disallow raw tracepoint access for unpriv
157 * 1 - disallow cpu events for unpriv
158 * 2 - disallow kernel profiling for unpriv
160 int sysctl_perf_event_paranoid __read_mostly = 1;
162 /* Minimum for 512 kiB + 1 user control page */
163 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
166 * max perf event sample rate
168 #define DEFAULT_MAX_SAMPLE_RATE 100000
169 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
170 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
172 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
174 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
175 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
177 static atomic_t perf_sample_allowed_ns __read_mostly =
178 ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100);
180 void update_perf_cpu_limits(void)
182 u64 tmp = perf_sample_period_ns;
184 tmp *= sysctl_perf_cpu_time_max_percent;
186 atomic_set(&perf_sample_allowed_ns, tmp);
189 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
191 int perf_proc_update_handler(struct ctl_table *table, int write,
192 void __user *buffer, size_t *lenp,
195 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
200 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
201 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
202 update_perf_cpu_limits();
207 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
209 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
210 void __user *buffer, size_t *lenp,
213 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
218 update_perf_cpu_limits();
224 * perf samples are done in some very critical code paths (NMIs).
225 * If they take too much CPU time, the system can lock up and not
226 * get any real work done. This will drop the sample rate when
227 * we detect that events are taking too long.
229 #define NR_ACCUMULATED_SAMPLES 128
230 DEFINE_PER_CPU(u64, running_sample_length);
232 void perf_sample_event_took(u64 sample_len_ns)
234 u64 avg_local_sample_len;
235 u64 local_samples_len;
237 if (atomic_read(&perf_sample_allowed_ns) == 0)
240 /* decay the counter by 1 average sample */
241 local_samples_len = __get_cpu_var(running_sample_length);
242 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
243 local_samples_len += sample_len_ns;
244 __get_cpu_var(running_sample_length) = local_samples_len;
247 * note: this will be biased artifically low until we have
248 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
249 * from having to maintain a count.
251 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
253 if (avg_local_sample_len <= atomic_read(&perf_sample_allowed_ns))
256 if (max_samples_per_tick <= 1)
259 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
260 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
261 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
263 printk_ratelimited(KERN_WARNING
264 "perf samples too long (%lld > %d), lowering "
265 "kernel.perf_event_max_sample_rate to %d\n",
266 avg_local_sample_len,
267 atomic_read(&perf_sample_allowed_ns),
268 sysctl_perf_event_sample_rate);
270 update_perf_cpu_limits();
273 static atomic64_t perf_event_id;
275 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
276 enum event_type_t event_type);
278 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
279 enum event_type_t event_type,
280 struct task_struct *task);
282 static void update_context_time(struct perf_event_context *ctx);
283 static u64 perf_event_time(struct perf_event *event);
285 void __weak perf_event_print_debug(void) { }
287 extern __weak const char *perf_pmu_name(void)
292 static inline u64 perf_clock(void)
294 return local_clock();
297 static inline struct perf_cpu_context *
298 __get_cpu_context(struct perf_event_context *ctx)
300 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
303 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
304 struct perf_event_context *ctx)
306 raw_spin_lock(&cpuctx->ctx.lock);
308 raw_spin_lock(&ctx->lock);
311 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
312 struct perf_event_context *ctx)
315 raw_spin_unlock(&ctx->lock);
316 raw_spin_unlock(&cpuctx->ctx.lock);
319 #ifdef CONFIG_CGROUP_PERF
322 * perf_cgroup_info keeps track of time_enabled for a cgroup.
323 * This is a per-cpu dynamically allocated data structure.
325 struct perf_cgroup_info {
331 struct cgroup_subsys_state css;
332 struct perf_cgroup_info __percpu *info;
336 * Must ensure cgroup is pinned (css_get) before calling
337 * this function. In other words, we cannot call this function
338 * if there is no cgroup event for the current CPU context.
340 static inline struct perf_cgroup *
341 perf_cgroup_from_task(struct task_struct *task)
343 return container_of(task_subsys_state(task, perf_subsys_id),
344 struct perf_cgroup, css);
348 perf_cgroup_match(struct perf_event *event)
350 struct perf_event_context *ctx = event->ctx;
351 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
353 /* @event doesn't care about cgroup */
357 /* wants specific cgroup scope but @cpuctx isn't associated with any */
362 * Cgroup scoping is recursive. An event enabled for a cgroup is
363 * also enabled for all its descendant cgroups. If @cpuctx's
364 * cgroup is a descendant of @event's (the test covers identity
365 * case), it's a match.
367 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
368 event->cgrp->css.cgroup);
371 static inline bool perf_tryget_cgroup(struct perf_event *event)
373 return css_tryget(&event->cgrp->css);
376 static inline void perf_put_cgroup(struct perf_event *event)
378 css_put(&event->cgrp->css);
381 static inline void perf_detach_cgroup(struct perf_event *event)
383 perf_put_cgroup(event);
387 static inline int is_cgroup_event(struct perf_event *event)
389 return event->cgrp != NULL;
392 static inline u64 perf_cgroup_event_time(struct perf_event *event)
394 struct perf_cgroup_info *t;
396 t = per_cpu_ptr(event->cgrp->info, event->cpu);
400 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
402 struct perf_cgroup_info *info;
407 info = this_cpu_ptr(cgrp->info);
409 info->time += now - info->timestamp;
410 info->timestamp = now;
413 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
415 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
417 __update_cgrp_time(cgrp_out);
420 static inline void update_cgrp_time_from_event(struct perf_event *event)
422 struct perf_cgroup *cgrp;
425 * ensure we access cgroup data only when needed and
426 * when we know the cgroup is pinned (css_get)
428 if (!is_cgroup_event(event))
431 cgrp = perf_cgroup_from_task(current);
433 * Do not update time when cgroup is not active
435 if (cgrp == event->cgrp)
436 __update_cgrp_time(event->cgrp);
440 perf_cgroup_set_timestamp(struct task_struct *task,
441 struct perf_event_context *ctx)
443 struct perf_cgroup *cgrp;
444 struct perf_cgroup_info *info;
447 * ctx->lock held by caller
448 * ensure we do not access cgroup data
449 * unless we have the cgroup pinned (css_get)
451 if (!task || !ctx->nr_cgroups)
454 cgrp = perf_cgroup_from_task(task);
455 info = this_cpu_ptr(cgrp->info);
456 info->timestamp = ctx->timestamp;
459 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
460 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
463 * reschedule events based on the cgroup constraint of task.
465 * mode SWOUT : schedule out everything
466 * mode SWIN : schedule in based on cgroup for next
468 void perf_cgroup_switch(struct task_struct *task, int mode)
470 struct perf_cpu_context *cpuctx;
475 * disable interrupts to avoid geting nr_cgroup
476 * changes via __perf_event_disable(). Also
479 local_irq_save(flags);
482 * we reschedule only in the presence of cgroup
483 * constrained events.
487 list_for_each_entry_rcu(pmu, &pmus, entry) {
488 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
489 if (cpuctx->unique_pmu != pmu)
490 continue; /* ensure we process each cpuctx once */
493 * perf_cgroup_events says at least one
494 * context on this CPU has cgroup events.
496 * ctx->nr_cgroups reports the number of cgroup
497 * events for a context.
499 if (cpuctx->ctx.nr_cgroups > 0) {
500 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
501 perf_pmu_disable(cpuctx->ctx.pmu);
503 if (mode & PERF_CGROUP_SWOUT) {
504 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
506 * must not be done before ctxswout due
507 * to event_filter_match() in event_sched_out()
512 if (mode & PERF_CGROUP_SWIN) {
513 WARN_ON_ONCE(cpuctx->cgrp);
515 * set cgrp before ctxsw in to allow
516 * event_filter_match() to not have to pass
519 cpuctx->cgrp = perf_cgroup_from_task(task);
520 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
522 perf_pmu_enable(cpuctx->ctx.pmu);
523 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
529 local_irq_restore(flags);
532 static inline void perf_cgroup_sched_out(struct task_struct *task,
533 struct task_struct *next)
535 struct perf_cgroup *cgrp1;
536 struct perf_cgroup *cgrp2 = NULL;
539 * we come here when we know perf_cgroup_events > 0
541 cgrp1 = perf_cgroup_from_task(task);
544 * next is NULL when called from perf_event_enable_on_exec()
545 * that will systematically cause a cgroup_switch()
548 cgrp2 = perf_cgroup_from_task(next);
551 * only schedule out current cgroup events if we know
552 * that we are switching to a different cgroup. Otherwise,
553 * do no touch the cgroup events.
556 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
559 static inline void perf_cgroup_sched_in(struct task_struct *prev,
560 struct task_struct *task)
562 struct perf_cgroup *cgrp1;
563 struct perf_cgroup *cgrp2 = NULL;
566 * we come here when we know perf_cgroup_events > 0
568 cgrp1 = perf_cgroup_from_task(task);
570 /* prev can never be NULL */
571 cgrp2 = perf_cgroup_from_task(prev);
574 * only need to schedule in cgroup events if we are changing
575 * cgroup during ctxsw. Cgroup events were not scheduled
576 * out of ctxsw out if that was not the case.
579 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
582 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
583 struct perf_event_attr *attr,
584 struct perf_event *group_leader)
586 struct perf_cgroup *cgrp;
587 struct cgroup_subsys_state *css;
588 struct fd f = fdget(fd);
594 css = cgroup_css_from_dir(f.file, perf_subsys_id);
600 cgrp = container_of(css, struct perf_cgroup, css);
603 /* must be done before we fput() the file */
604 if (!perf_tryget_cgroup(event)) {
611 * all events in a group must monitor
612 * the same cgroup because a task belongs
613 * to only one perf cgroup at a time
615 if (group_leader && group_leader->cgrp != cgrp) {
616 perf_detach_cgroup(event);
625 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
627 struct perf_cgroup_info *t;
628 t = per_cpu_ptr(event->cgrp->info, event->cpu);
629 event->shadow_ctx_time = now - t->timestamp;
633 perf_cgroup_defer_enabled(struct perf_event *event)
636 * when the current task's perf cgroup does not match
637 * the event's, we need to remember to call the
638 * perf_mark_enable() function the first time a task with
639 * a matching perf cgroup is scheduled in.
641 if (is_cgroup_event(event) && !perf_cgroup_match(event))
642 event->cgrp_defer_enabled = 1;
646 perf_cgroup_mark_enabled(struct perf_event *event,
647 struct perf_event_context *ctx)
649 struct perf_event *sub;
650 u64 tstamp = perf_event_time(event);
652 if (!event->cgrp_defer_enabled)
655 event->cgrp_defer_enabled = 0;
657 event->tstamp_enabled = tstamp - event->total_time_enabled;
658 list_for_each_entry(sub, &event->sibling_list, group_entry) {
659 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
660 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
661 sub->cgrp_defer_enabled = 0;
665 #else /* !CONFIG_CGROUP_PERF */
668 perf_cgroup_match(struct perf_event *event)
673 static inline void perf_detach_cgroup(struct perf_event *event)
676 static inline int is_cgroup_event(struct perf_event *event)
681 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
686 static inline void update_cgrp_time_from_event(struct perf_event *event)
690 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
694 static inline void perf_cgroup_sched_out(struct task_struct *task,
695 struct task_struct *next)
699 static inline void perf_cgroup_sched_in(struct task_struct *prev,
700 struct task_struct *task)
704 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
705 struct perf_event_attr *attr,
706 struct perf_event *group_leader)
712 perf_cgroup_set_timestamp(struct task_struct *task,
713 struct perf_event_context *ctx)
718 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
723 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
727 static inline u64 perf_cgroup_event_time(struct perf_event *event)
733 perf_cgroup_defer_enabled(struct perf_event *event)
738 perf_cgroup_mark_enabled(struct perf_event *event,
739 struct perf_event_context *ctx)
745 * set default to be dependent on timer tick just
748 #define PERF_CPU_HRTIMER (1000 / HZ)
750 * function must be called with interrupts disbled
752 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
754 struct perf_cpu_context *cpuctx;
755 enum hrtimer_restart ret = HRTIMER_NORESTART;
758 WARN_ON(!irqs_disabled());
760 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
762 rotations = perf_rotate_context(cpuctx);
765 * arm timer if needed
768 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
769 ret = HRTIMER_RESTART;
775 /* CPU is going down */
776 void perf_cpu_hrtimer_cancel(int cpu)
778 struct perf_cpu_context *cpuctx;
782 if (WARN_ON(cpu != smp_processor_id()))
785 local_irq_save(flags);
789 list_for_each_entry_rcu(pmu, &pmus, entry) {
790 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
792 if (pmu->task_ctx_nr == perf_sw_context)
795 hrtimer_cancel(&cpuctx->hrtimer);
800 local_irq_restore(flags);
803 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
805 struct hrtimer *hr = &cpuctx->hrtimer;
806 struct pmu *pmu = cpuctx->ctx.pmu;
809 /* no multiplexing needed for SW PMU */
810 if (pmu->task_ctx_nr == perf_sw_context)
814 * check default is sane, if not set then force to
815 * default interval (1/tick)
817 timer = pmu->hrtimer_interval_ms;
819 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
821 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
823 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
824 hr->function = perf_cpu_hrtimer_handler;
827 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
829 struct hrtimer *hr = &cpuctx->hrtimer;
830 struct pmu *pmu = cpuctx->ctx.pmu;
833 if (pmu->task_ctx_nr == perf_sw_context)
836 if (hrtimer_active(hr))
839 if (!hrtimer_callback_running(hr))
840 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
841 0, HRTIMER_MODE_REL_PINNED, 0);
844 void perf_pmu_disable(struct pmu *pmu)
846 int *count = this_cpu_ptr(pmu->pmu_disable_count);
848 pmu->pmu_disable(pmu);
851 void perf_pmu_enable(struct pmu *pmu)
853 int *count = this_cpu_ptr(pmu->pmu_disable_count);
855 pmu->pmu_enable(pmu);
858 static DEFINE_PER_CPU(struct list_head, rotation_list);
861 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
862 * because they're strictly cpu affine and rotate_start is called with IRQs
863 * disabled, while rotate_context is called from IRQ context.
865 static void perf_pmu_rotate_start(struct pmu *pmu)
867 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
868 struct list_head *head = &__get_cpu_var(rotation_list);
870 WARN_ON(!irqs_disabled());
872 if (list_empty(&cpuctx->rotation_list)) {
873 int was_empty = list_empty(head);
874 list_add(&cpuctx->rotation_list, head);
876 tick_nohz_full_kick();
880 static void get_ctx(struct perf_event_context *ctx)
882 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
885 static void put_ctx(struct perf_event_context *ctx)
887 if (atomic_dec_and_test(&ctx->refcount)) {
889 put_ctx(ctx->parent_ctx);
891 put_task_struct(ctx->task);
892 kfree_rcu(ctx, rcu_head);
896 static void unclone_ctx(struct perf_event_context *ctx)
898 if (ctx->parent_ctx) {
899 put_ctx(ctx->parent_ctx);
900 ctx->parent_ctx = NULL;
904 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
907 * only top level events have the pid namespace they were created in
910 event = event->parent;
912 return task_tgid_nr_ns(p, event->ns);
915 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
918 * only top level events have the pid namespace they were created in
921 event = event->parent;
923 return task_pid_nr_ns(p, event->ns);
927 * If we inherit events we want to return the parent event id
930 static u64 primary_event_id(struct perf_event *event)
935 id = event->parent->id;
941 * Get the perf_event_context for a task and lock it.
942 * This has to cope with with the fact that until it is locked,
943 * the context could get moved to another task.
945 static struct perf_event_context *
946 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
948 struct perf_event_context *ctx;
952 * One of the few rules of preemptible RCU is that one cannot do
953 * rcu_read_unlock() while holding a scheduler (or nested) lock when
954 * part of the read side critical section was preemptible -- see
955 * rcu_read_unlock_special().
957 * Since ctx->lock nests under rq->lock we must ensure the entire read
958 * side critical section is non-preemptible.
962 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
965 * If this context is a clone of another, it might
966 * get swapped for another underneath us by
967 * perf_event_task_sched_out, though the
968 * rcu_read_lock() protects us from any context
969 * getting freed. Lock the context and check if it
970 * got swapped before we could get the lock, and retry
971 * if so. If we locked the right context, then it
972 * can't get swapped on us any more.
974 raw_spin_lock_irqsave(&ctx->lock, *flags);
975 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
976 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
982 if (!atomic_inc_not_zero(&ctx->refcount)) {
983 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
993 * Get the context for a task and increment its pin_count so it
994 * can't get swapped to another task. This also increments its
995 * reference count so that the context can't get freed.
997 static struct perf_event_context *
998 perf_pin_task_context(struct task_struct *task, int ctxn)
1000 struct perf_event_context *ctx;
1001 unsigned long flags;
1003 ctx = perf_lock_task_context(task, ctxn, &flags);
1006 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1011 static void perf_unpin_context(struct perf_event_context *ctx)
1013 unsigned long flags;
1015 raw_spin_lock_irqsave(&ctx->lock, flags);
1017 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1021 * Update the record of the current time in a context.
1023 static void update_context_time(struct perf_event_context *ctx)
1025 u64 now = perf_clock();
1027 ctx->time += now - ctx->timestamp;
1028 ctx->timestamp = now;
1031 static u64 perf_event_time(struct perf_event *event)
1033 struct perf_event_context *ctx = event->ctx;
1035 if (is_cgroup_event(event))
1036 return perf_cgroup_event_time(event);
1038 return ctx ? ctx->time : 0;
1042 * Update the total_time_enabled and total_time_running fields for a event.
1043 * The caller of this function needs to hold the ctx->lock.
1045 static void update_event_times(struct perf_event *event)
1047 struct perf_event_context *ctx = event->ctx;
1050 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1051 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1054 * in cgroup mode, time_enabled represents
1055 * the time the event was enabled AND active
1056 * tasks were in the monitored cgroup. This is
1057 * independent of the activity of the context as
1058 * there may be a mix of cgroup and non-cgroup events.
1060 * That is why we treat cgroup events differently
1063 if (is_cgroup_event(event))
1064 run_end = perf_cgroup_event_time(event);
1065 else if (ctx->is_active)
1066 run_end = ctx->time;
1068 run_end = event->tstamp_stopped;
1070 event->total_time_enabled = run_end - event->tstamp_enabled;
1072 if (event->state == PERF_EVENT_STATE_INACTIVE)
1073 run_end = event->tstamp_stopped;
1075 run_end = perf_event_time(event);
1077 event->total_time_running = run_end - event->tstamp_running;
1082 * Update total_time_enabled and total_time_running for all events in a group.
1084 static void update_group_times(struct perf_event *leader)
1086 struct perf_event *event;
1088 update_event_times(leader);
1089 list_for_each_entry(event, &leader->sibling_list, group_entry)
1090 update_event_times(event);
1093 static struct list_head *
1094 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1096 if (event->attr.pinned)
1097 return &ctx->pinned_groups;
1099 return &ctx->flexible_groups;
1103 * Add a event from the lists for its context.
1104 * Must be called with ctx->mutex and ctx->lock held.
1107 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1109 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1110 event->attach_state |= PERF_ATTACH_CONTEXT;
1113 * If we're a stand alone event or group leader, we go to the context
1114 * list, group events are kept attached to the group so that
1115 * perf_group_detach can, at all times, locate all siblings.
1117 if (event->group_leader == event) {
1118 struct list_head *list;
1120 if (is_software_event(event))
1121 event->group_flags |= PERF_GROUP_SOFTWARE;
1123 list = ctx_group_list(event, ctx);
1124 list_add_tail(&event->group_entry, list);
1127 if (is_cgroup_event(event))
1130 if (has_branch_stack(event))
1131 ctx->nr_branch_stack++;
1133 list_add_rcu(&event->event_entry, &ctx->event_list);
1134 if (!ctx->nr_events)
1135 perf_pmu_rotate_start(ctx->pmu);
1137 if (event->attr.inherit_stat)
1142 * Initialize event state based on the perf_event_attr::disabled.
1144 static inline void perf_event__state_init(struct perf_event *event)
1146 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1147 PERF_EVENT_STATE_INACTIVE;
1151 * Called at perf_event creation and when events are attached/detached from a
1154 static void perf_event__read_size(struct perf_event *event)
1156 int entry = sizeof(u64); /* value */
1160 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1161 size += sizeof(u64);
1163 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1164 size += sizeof(u64);
1166 if (event->attr.read_format & PERF_FORMAT_ID)
1167 entry += sizeof(u64);
1169 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1170 nr += event->group_leader->nr_siblings;
1171 size += sizeof(u64);
1175 event->read_size = size;
1178 static void perf_event__header_size(struct perf_event *event)
1180 struct perf_sample_data *data;
1181 u64 sample_type = event->attr.sample_type;
1184 perf_event__read_size(event);
1186 if (sample_type & PERF_SAMPLE_IP)
1187 size += sizeof(data->ip);
1189 if (sample_type & PERF_SAMPLE_ADDR)
1190 size += sizeof(data->addr);
1192 if (sample_type & PERF_SAMPLE_PERIOD)
1193 size += sizeof(data->period);
1195 if (sample_type & PERF_SAMPLE_WEIGHT)
1196 size += sizeof(data->weight);
1198 if (sample_type & PERF_SAMPLE_READ)
1199 size += event->read_size;
1201 if (sample_type & PERF_SAMPLE_DATA_SRC)
1202 size += sizeof(data->data_src.val);
1204 event->header_size = size;
1207 static void perf_event__id_header_size(struct perf_event *event)
1209 struct perf_sample_data *data;
1210 u64 sample_type = event->attr.sample_type;
1213 if (sample_type & PERF_SAMPLE_TID)
1214 size += sizeof(data->tid_entry);
1216 if (sample_type & PERF_SAMPLE_TIME)
1217 size += sizeof(data->time);
1219 if (sample_type & PERF_SAMPLE_ID)
1220 size += sizeof(data->id);
1222 if (sample_type & PERF_SAMPLE_STREAM_ID)
1223 size += sizeof(data->stream_id);
1225 if (sample_type & PERF_SAMPLE_CPU)
1226 size += sizeof(data->cpu_entry);
1228 event->id_header_size = size;
1231 static void perf_group_attach(struct perf_event *event)
1233 struct perf_event *group_leader = event->group_leader, *pos;
1236 * We can have double attach due to group movement in perf_event_open.
1238 if (event->attach_state & PERF_ATTACH_GROUP)
1241 event->attach_state |= PERF_ATTACH_GROUP;
1243 if (group_leader == event)
1246 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1247 !is_software_event(event))
1248 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1250 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1251 group_leader->nr_siblings++;
1253 perf_event__header_size(group_leader);
1255 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1256 perf_event__header_size(pos);
1260 * Remove a event from the lists for its context.
1261 * Must be called with ctx->mutex and ctx->lock held.
1264 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1266 struct perf_cpu_context *cpuctx;
1268 * We can have double detach due to exit/hot-unplug + close.
1270 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1273 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1275 if (is_cgroup_event(event)) {
1277 cpuctx = __get_cpu_context(ctx);
1279 * if there are no more cgroup events
1280 * then cler cgrp to avoid stale pointer
1281 * in update_cgrp_time_from_cpuctx()
1283 if (!ctx->nr_cgroups)
1284 cpuctx->cgrp = NULL;
1287 if (has_branch_stack(event))
1288 ctx->nr_branch_stack--;
1291 if (event->attr.inherit_stat)
1294 list_del_rcu(&event->event_entry);
1296 if (event->group_leader == event)
1297 list_del_init(&event->group_entry);
1299 update_group_times(event);
1302 * If event was in error state, then keep it
1303 * that way, otherwise bogus counts will be
1304 * returned on read(). The only way to get out
1305 * of error state is by explicit re-enabling
1308 if (event->state > PERF_EVENT_STATE_OFF)
1309 event->state = PERF_EVENT_STATE_OFF;
1312 static void perf_group_detach(struct perf_event *event)
1314 struct perf_event *sibling, *tmp;
1315 struct list_head *list = NULL;
1318 * We can have double detach due to exit/hot-unplug + close.
1320 if (!(event->attach_state & PERF_ATTACH_GROUP))
1323 event->attach_state &= ~PERF_ATTACH_GROUP;
1326 * If this is a sibling, remove it from its group.
1328 if (event->group_leader != event) {
1329 list_del_init(&event->group_entry);
1330 event->group_leader->nr_siblings--;
1334 if (!list_empty(&event->group_entry))
1335 list = &event->group_entry;
1338 * If this was a group event with sibling events then
1339 * upgrade the siblings to singleton events by adding them
1340 * to whatever list we are on.
1342 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1344 list_move_tail(&sibling->group_entry, list);
1345 sibling->group_leader = sibling;
1347 /* Inherit group flags from the previous leader */
1348 sibling->group_flags = event->group_flags;
1352 perf_event__header_size(event->group_leader);
1354 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1355 perf_event__header_size(tmp);
1359 event_filter_match(struct perf_event *event)
1361 return (event->cpu == -1 || event->cpu == smp_processor_id())
1362 && perf_cgroup_match(event);
1366 event_sched_out(struct perf_event *event,
1367 struct perf_cpu_context *cpuctx,
1368 struct perf_event_context *ctx)
1370 u64 tstamp = perf_event_time(event);
1373 * An event which could not be activated because of
1374 * filter mismatch still needs to have its timings
1375 * maintained, otherwise bogus information is return
1376 * via read() for time_enabled, time_running:
1378 if (event->state == PERF_EVENT_STATE_INACTIVE
1379 && !event_filter_match(event)) {
1380 delta = tstamp - event->tstamp_stopped;
1381 event->tstamp_running += delta;
1382 event->tstamp_stopped = tstamp;
1385 if (event->state != PERF_EVENT_STATE_ACTIVE)
1388 event->state = PERF_EVENT_STATE_INACTIVE;
1389 if (event->pending_disable) {
1390 event->pending_disable = 0;
1391 event->state = PERF_EVENT_STATE_OFF;
1393 event->tstamp_stopped = tstamp;
1394 event->pmu->del(event, 0);
1397 if (!is_software_event(event))
1398 cpuctx->active_oncpu--;
1400 if (event->attr.freq && event->attr.sample_freq)
1402 if (event->attr.exclusive || !cpuctx->active_oncpu)
1403 cpuctx->exclusive = 0;
1407 group_sched_out(struct perf_event *group_event,
1408 struct perf_cpu_context *cpuctx,
1409 struct perf_event_context *ctx)
1411 struct perf_event *event;
1412 int state = group_event->state;
1414 event_sched_out(group_event, cpuctx, ctx);
1417 * Schedule out siblings (if any):
1419 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1420 event_sched_out(event, cpuctx, ctx);
1422 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1423 cpuctx->exclusive = 0;
1427 * Cross CPU call to remove a performance event
1429 * We disable the event on the hardware level first. After that we
1430 * remove it from the context list.
1432 static int __perf_remove_from_context(void *info)
1434 struct perf_event *event = info;
1435 struct perf_event_context *ctx = event->ctx;
1436 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1438 raw_spin_lock(&ctx->lock);
1439 event_sched_out(event, cpuctx, ctx);
1440 list_del_event(event, ctx);
1441 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1443 cpuctx->task_ctx = NULL;
1445 raw_spin_unlock(&ctx->lock);
1452 * Remove the event from a task's (or a CPU's) list of events.
1454 * CPU events are removed with a smp call. For task events we only
1455 * call when the task is on a CPU.
1457 * If event->ctx is a cloned context, callers must make sure that
1458 * every task struct that event->ctx->task could possibly point to
1459 * remains valid. This is OK when called from perf_release since
1460 * that only calls us on the top-level context, which can't be a clone.
1461 * When called from perf_event_exit_task, it's OK because the
1462 * context has been detached from its task.
1464 static void perf_remove_from_context(struct perf_event *event)
1466 struct perf_event_context *ctx = event->ctx;
1467 struct task_struct *task = ctx->task;
1469 lockdep_assert_held(&ctx->mutex);
1473 * Per cpu events are removed via an smp call and
1474 * the removal is always successful.
1476 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1481 if (!task_function_call(task, __perf_remove_from_context, event))
1484 raw_spin_lock_irq(&ctx->lock);
1486 * If we failed to find a running task, but find the context active now
1487 * that we've acquired the ctx->lock, retry.
1489 if (ctx->is_active) {
1490 raw_spin_unlock_irq(&ctx->lock);
1495 * Since the task isn't running, its safe to remove the event, us
1496 * holding the ctx->lock ensures the task won't get scheduled in.
1498 list_del_event(event, ctx);
1499 raw_spin_unlock_irq(&ctx->lock);
1503 * Cross CPU call to disable a performance event
1505 int __perf_event_disable(void *info)
1507 struct perf_event *event = info;
1508 struct perf_event_context *ctx = event->ctx;
1509 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1512 * If this is a per-task event, need to check whether this
1513 * event's task is the current task on this cpu.
1515 * Can trigger due to concurrent perf_event_context_sched_out()
1516 * flipping contexts around.
1518 if (ctx->task && cpuctx->task_ctx != ctx)
1521 raw_spin_lock(&ctx->lock);
1524 * If the event is on, turn it off.
1525 * If it is in error state, leave it in error state.
1527 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1528 update_context_time(ctx);
1529 update_cgrp_time_from_event(event);
1530 update_group_times(event);
1531 if (event == event->group_leader)
1532 group_sched_out(event, cpuctx, ctx);
1534 event_sched_out(event, cpuctx, ctx);
1535 event->state = PERF_EVENT_STATE_OFF;
1538 raw_spin_unlock(&ctx->lock);
1546 * If event->ctx is a cloned context, callers must make sure that
1547 * every task struct that event->ctx->task could possibly point to
1548 * remains valid. This condition is satisifed when called through
1549 * perf_event_for_each_child or perf_event_for_each because they
1550 * hold the top-level event's child_mutex, so any descendant that
1551 * goes to exit will block in sync_child_event.
1552 * When called from perf_pending_event it's OK because event->ctx
1553 * is the current context on this CPU and preemption is disabled,
1554 * hence we can't get into perf_event_task_sched_out for this context.
1556 void perf_event_disable(struct perf_event *event)
1558 struct perf_event_context *ctx = event->ctx;
1559 struct task_struct *task = ctx->task;
1563 * Disable the event on the cpu that it's on
1565 cpu_function_call(event->cpu, __perf_event_disable, event);
1570 if (!task_function_call(task, __perf_event_disable, event))
1573 raw_spin_lock_irq(&ctx->lock);
1575 * If the event is still active, we need to retry the cross-call.
1577 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1578 raw_spin_unlock_irq(&ctx->lock);
1580 * Reload the task pointer, it might have been changed by
1581 * a concurrent perf_event_context_sched_out().
1588 * Since we have the lock this context can't be scheduled
1589 * in, so we can change the state safely.
1591 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1592 update_group_times(event);
1593 event->state = PERF_EVENT_STATE_OFF;
1595 raw_spin_unlock_irq(&ctx->lock);
1597 EXPORT_SYMBOL_GPL(perf_event_disable);
1599 static void perf_set_shadow_time(struct perf_event *event,
1600 struct perf_event_context *ctx,
1604 * use the correct time source for the time snapshot
1606 * We could get by without this by leveraging the
1607 * fact that to get to this function, the caller
1608 * has most likely already called update_context_time()
1609 * and update_cgrp_time_xx() and thus both timestamp
1610 * are identical (or very close). Given that tstamp is,
1611 * already adjusted for cgroup, we could say that:
1612 * tstamp - ctx->timestamp
1614 * tstamp - cgrp->timestamp.
1616 * Then, in perf_output_read(), the calculation would
1617 * work with no changes because:
1618 * - event is guaranteed scheduled in
1619 * - no scheduled out in between
1620 * - thus the timestamp would be the same
1622 * But this is a bit hairy.
1624 * So instead, we have an explicit cgroup call to remain
1625 * within the time time source all along. We believe it
1626 * is cleaner and simpler to understand.
1628 if (is_cgroup_event(event))
1629 perf_cgroup_set_shadow_time(event, tstamp);
1631 event->shadow_ctx_time = tstamp - ctx->timestamp;
1634 #define MAX_INTERRUPTS (~0ULL)
1636 static void perf_log_throttle(struct perf_event *event, int enable);
1639 event_sched_in(struct perf_event *event,
1640 struct perf_cpu_context *cpuctx,
1641 struct perf_event_context *ctx)
1643 u64 tstamp = perf_event_time(event);
1645 if (event->state <= PERF_EVENT_STATE_OFF)
1648 event->state = PERF_EVENT_STATE_ACTIVE;
1649 event->oncpu = smp_processor_id();
1652 * Unthrottle events, since we scheduled we might have missed several
1653 * ticks already, also for a heavily scheduling task there is little
1654 * guarantee it'll get a tick in a timely manner.
1656 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1657 perf_log_throttle(event, 1);
1658 event->hw.interrupts = 0;
1662 * The new state must be visible before we turn it on in the hardware:
1666 if (event->pmu->add(event, PERF_EF_START)) {
1667 event->state = PERF_EVENT_STATE_INACTIVE;
1672 event->tstamp_running += tstamp - event->tstamp_stopped;
1674 perf_set_shadow_time(event, ctx, tstamp);
1676 if (!is_software_event(event))
1677 cpuctx->active_oncpu++;
1679 if (event->attr.freq && event->attr.sample_freq)
1682 if (event->attr.exclusive)
1683 cpuctx->exclusive = 1;
1689 group_sched_in(struct perf_event *group_event,
1690 struct perf_cpu_context *cpuctx,
1691 struct perf_event_context *ctx)
1693 struct perf_event *event, *partial_group = NULL;
1694 struct pmu *pmu = group_event->pmu;
1695 u64 now = ctx->time;
1696 bool simulate = false;
1698 if (group_event->state == PERF_EVENT_STATE_OFF)
1701 pmu->start_txn(pmu);
1703 if (event_sched_in(group_event, cpuctx, ctx)) {
1704 pmu->cancel_txn(pmu);
1705 perf_cpu_hrtimer_restart(cpuctx);
1710 * Schedule in siblings as one group (if any):
1712 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1713 if (event_sched_in(event, cpuctx, ctx)) {
1714 partial_group = event;
1719 if (!pmu->commit_txn(pmu))
1724 * Groups can be scheduled in as one unit only, so undo any
1725 * partial group before returning:
1726 * The events up to the failed event are scheduled out normally,
1727 * tstamp_stopped will be updated.
1729 * The failed events and the remaining siblings need to have
1730 * their timings updated as if they had gone thru event_sched_in()
1731 * and event_sched_out(). This is required to get consistent timings
1732 * across the group. This also takes care of the case where the group
1733 * could never be scheduled by ensuring tstamp_stopped is set to mark
1734 * the time the event was actually stopped, such that time delta
1735 * calculation in update_event_times() is correct.
1737 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1738 if (event == partial_group)
1742 event->tstamp_running += now - event->tstamp_stopped;
1743 event->tstamp_stopped = now;
1745 event_sched_out(event, cpuctx, ctx);
1748 event_sched_out(group_event, cpuctx, ctx);
1750 pmu->cancel_txn(pmu);
1752 perf_cpu_hrtimer_restart(cpuctx);
1758 * Work out whether we can put this event group on the CPU now.
1760 static int group_can_go_on(struct perf_event *event,
1761 struct perf_cpu_context *cpuctx,
1765 * Groups consisting entirely of software events can always go on.
1767 if (event->group_flags & PERF_GROUP_SOFTWARE)
1770 * If an exclusive group is already on, no other hardware
1773 if (cpuctx->exclusive)
1776 * If this group is exclusive and there are already
1777 * events on the CPU, it can't go on.
1779 if (event->attr.exclusive && cpuctx->active_oncpu)
1782 * Otherwise, try to add it if all previous groups were able
1788 static void add_event_to_ctx(struct perf_event *event,
1789 struct perf_event_context *ctx)
1791 u64 tstamp = perf_event_time(event);
1793 list_add_event(event, ctx);
1794 perf_group_attach(event);
1795 event->tstamp_enabled = tstamp;
1796 event->tstamp_running = tstamp;
1797 event->tstamp_stopped = tstamp;
1800 static void task_ctx_sched_out(struct perf_event_context *ctx);
1802 ctx_sched_in(struct perf_event_context *ctx,
1803 struct perf_cpu_context *cpuctx,
1804 enum event_type_t event_type,
1805 struct task_struct *task);
1807 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1808 struct perf_event_context *ctx,
1809 struct task_struct *task)
1811 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1813 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1814 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1816 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1820 * Cross CPU call to install and enable a performance event
1822 * Must be called with ctx->mutex held
1824 static int __perf_install_in_context(void *info)
1826 struct perf_event *event = info;
1827 struct perf_event_context *ctx = event->ctx;
1828 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1829 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1830 struct task_struct *task = current;
1832 perf_ctx_lock(cpuctx, task_ctx);
1833 perf_pmu_disable(cpuctx->ctx.pmu);
1836 * If there was an active task_ctx schedule it out.
1839 task_ctx_sched_out(task_ctx);
1842 * If the context we're installing events in is not the
1843 * active task_ctx, flip them.
1845 if (ctx->task && task_ctx != ctx) {
1847 raw_spin_unlock(&task_ctx->lock);
1848 raw_spin_lock(&ctx->lock);
1853 cpuctx->task_ctx = task_ctx;
1854 task = task_ctx->task;
1857 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1859 update_context_time(ctx);
1861 * update cgrp time only if current cgrp
1862 * matches event->cgrp. Must be done before
1863 * calling add_event_to_ctx()
1865 update_cgrp_time_from_event(event);
1867 add_event_to_ctx(event, ctx);
1870 * Schedule everything back in
1872 perf_event_sched_in(cpuctx, task_ctx, task);
1874 perf_pmu_enable(cpuctx->ctx.pmu);
1875 perf_ctx_unlock(cpuctx, task_ctx);
1881 * Attach a performance event to a context
1883 * First we add the event to the list with the hardware enable bit
1884 * in event->hw_config cleared.
1886 * If the event is attached to a task which is on a CPU we use a smp
1887 * call to enable it in the task context. The task might have been
1888 * scheduled away, but we check this in the smp call again.
1891 perf_install_in_context(struct perf_event_context *ctx,
1892 struct perf_event *event,
1895 struct task_struct *task = ctx->task;
1897 lockdep_assert_held(&ctx->mutex);
1900 if (event->cpu != -1)
1905 * Per cpu events are installed via an smp call and
1906 * the install is always successful.
1908 cpu_function_call(cpu, __perf_install_in_context, event);
1913 if (!task_function_call(task, __perf_install_in_context, event))
1916 raw_spin_lock_irq(&ctx->lock);
1918 * If we failed to find a running task, but find the context active now
1919 * that we've acquired the ctx->lock, retry.
1921 if (ctx->is_active) {
1922 raw_spin_unlock_irq(&ctx->lock);
1927 * Since the task isn't running, its safe to add the event, us holding
1928 * the ctx->lock ensures the task won't get scheduled in.
1930 add_event_to_ctx(event, ctx);
1931 raw_spin_unlock_irq(&ctx->lock);
1935 * Put a event into inactive state and update time fields.
1936 * Enabling the leader of a group effectively enables all
1937 * the group members that aren't explicitly disabled, so we
1938 * have to update their ->tstamp_enabled also.
1939 * Note: this works for group members as well as group leaders
1940 * since the non-leader members' sibling_lists will be empty.
1942 static void __perf_event_mark_enabled(struct perf_event *event)
1944 struct perf_event *sub;
1945 u64 tstamp = perf_event_time(event);
1947 event->state = PERF_EVENT_STATE_INACTIVE;
1948 event->tstamp_enabled = tstamp - event->total_time_enabled;
1949 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1950 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1951 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1956 * Cross CPU call to enable a performance event
1958 static int __perf_event_enable(void *info)
1960 struct perf_event *event = info;
1961 struct perf_event_context *ctx = event->ctx;
1962 struct perf_event *leader = event->group_leader;
1963 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1967 * There's a time window between 'ctx->is_active' check
1968 * in perf_event_enable function and this place having:
1970 * - ctx->lock unlocked
1972 * where the task could be killed and 'ctx' deactivated
1973 * by perf_event_exit_task.
1975 if (!ctx->is_active)
1978 raw_spin_lock(&ctx->lock);
1979 update_context_time(ctx);
1981 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1985 * set current task's cgroup time reference point
1987 perf_cgroup_set_timestamp(current, ctx);
1989 __perf_event_mark_enabled(event);
1991 if (!event_filter_match(event)) {
1992 if (is_cgroup_event(event))
1993 perf_cgroup_defer_enabled(event);
1998 * If the event is in a group and isn't the group leader,
1999 * then don't put it on unless the group is on.
2001 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2004 if (!group_can_go_on(event, cpuctx, 1)) {
2007 if (event == leader)
2008 err = group_sched_in(event, cpuctx, ctx);
2010 err = event_sched_in(event, cpuctx, ctx);
2015 * If this event can't go on and it's part of a
2016 * group, then the whole group has to come off.
2018 if (leader != event) {
2019 group_sched_out(leader, cpuctx, ctx);
2020 perf_cpu_hrtimer_restart(cpuctx);
2022 if (leader->attr.pinned) {
2023 update_group_times(leader);
2024 leader->state = PERF_EVENT_STATE_ERROR;
2029 raw_spin_unlock(&ctx->lock);
2037 * If event->ctx is a cloned context, callers must make sure that
2038 * every task struct that event->ctx->task could possibly point to
2039 * remains valid. This condition is satisfied when called through
2040 * perf_event_for_each_child or perf_event_for_each as described
2041 * for perf_event_disable.
2043 void perf_event_enable(struct perf_event *event)
2045 struct perf_event_context *ctx = event->ctx;
2046 struct task_struct *task = ctx->task;
2050 * Enable the event on the cpu that it's on
2052 cpu_function_call(event->cpu, __perf_event_enable, event);
2056 raw_spin_lock_irq(&ctx->lock);
2057 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2061 * If the event is in error state, clear that first.
2062 * That way, if we see the event in error state below, we
2063 * know that it has gone back into error state, as distinct
2064 * from the task having been scheduled away before the
2065 * cross-call arrived.
2067 if (event->state == PERF_EVENT_STATE_ERROR)
2068 event->state = PERF_EVENT_STATE_OFF;
2071 if (!ctx->is_active) {
2072 __perf_event_mark_enabled(event);
2076 raw_spin_unlock_irq(&ctx->lock);
2078 if (!task_function_call(task, __perf_event_enable, event))
2081 raw_spin_lock_irq(&ctx->lock);
2084 * If the context is active and the event is still off,
2085 * we need to retry the cross-call.
2087 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2089 * task could have been flipped by a concurrent
2090 * perf_event_context_sched_out()
2097 raw_spin_unlock_irq(&ctx->lock);
2099 EXPORT_SYMBOL_GPL(perf_event_enable);
2101 int perf_event_refresh(struct perf_event *event, int refresh)
2104 * not supported on inherited events
2106 if (event->attr.inherit || !is_sampling_event(event))
2109 atomic_add(refresh, &event->event_limit);
2110 perf_event_enable(event);
2114 EXPORT_SYMBOL_GPL(perf_event_refresh);
2116 static void ctx_sched_out(struct perf_event_context *ctx,
2117 struct perf_cpu_context *cpuctx,
2118 enum event_type_t event_type)
2120 struct perf_event *event;
2121 int is_active = ctx->is_active;
2123 ctx->is_active &= ~event_type;
2124 if (likely(!ctx->nr_events))
2127 update_context_time(ctx);
2128 update_cgrp_time_from_cpuctx(cpuctx);
2129 if (!ctx->nr_active)
2132 perf_pmu_disable(ctx->pmu);
2133 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2134 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2135 group_sched_out(event, cpuctx, ctx);
2138 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2139 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2140 group_sched_out(event, cpuctx, ctx);
2142 perf_pmu_enable(ctx->pmu);
2146 * Test whether two contexts are equivalent, i.e. whether they
2147 * have both been cloned from the same version of the same context
2148 * and they both have the same number of enabled events.
2149 * If the number of enabled events is the same, then the set
2150 * of enabled events should be the same, because these are both
2151 * inherited contexts, therefore we can't access individual events
2152 * in them directly with an fd; we can only enable/disable all
2153 * events via prctl, or enable/disable all events in a family
2154 * via ioctl, which will have the same effect on both contexts.
2156 static int context_equiv(struct perf_event_context *ctx1,
2157 struct perf_event_context *ctx2)
2159 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
2160 && ctx1->parent_gen == ctx2->parent_gen
2161 && !ctx1->pin_count && !ctx2->pin_count;
2164 static void __perf_event_sync_stat(struct perf_event *event,
2165 struct perf_event *next_event)
2169 if (!event->attr.inherit_stat)
2173 * Update the event value, we cannot use perf_event_read()
2174 * because we're in the middle of a context switch and have IRQs
2175 * disabled, which upsets smp_call_function_single(), however
2176 * we know the event must be on the current CPU, therefore we
2177 * don't need to use it.
2179 switch (event->state) {
2180 case PERF_EVENT_STATE_ACTIVE:
2181 event->pmu->read(event);
2184 case PERF_EVENT_STATE_INACTIVE:
2185 update_event_times(event);
2193 * In order to keep per-task stats reliable we need to flip the event
2194 * values when we flip the contexts.
2196 value = local64_read(&next_event->count);
2197 value = local64_xchg(&event->count, value);
2198 local64_set(&next_event->count, value);
2200 swap(event->total_time_enabled, next_event->total_time_enabled);
2201 swap(event->total_time_running, next_event->total_time_running);
2204 * Since we swizzled the values, update the user visible data too.
2206 perf_event_update_userpage(event);
2207 perf_event_update_userpage(next_event);
2210 #define list_next_entry(pos, member) \
2211 list_entry(pos->member.next, typeof(*pos), member)
2213 static void perf_event_sync_stat(struct perf_event_context *ctx,
2214 struct perf_event_context *next_ctx)
2216 struct perf_event *event, *next_event;
2221 update_context_time(ctx);
2223 event = list_first_entry(&ctx->event_list,
2224 struct perf_event, event_entry);
2226 next_event = list_first_entry(&next_ctx->event_list,
2227 struct perf_event, event_entry);
2229 while (&event->event_entry != &ctx->event_list &&
2230 &next_event->event_entry != &next_ctx->event_list) {
2232 __perf_event_sync_stat(event, next_event);
2234 event = list_next_entry(event, event_entry);
2235 next_event = list_next_entry(next_event, event_entry);
2239 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2240 struct task_struct *next)
2242 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2243 struct perf_event_context *next_ctx;
2244 struct perf_event_context *parent;
2245 struct perf_cpu_context *cpuctx;
2251 cpuctx = __get_cpu_context(ctx);
2252 if (!cpuctx->task_ctx)
2256 parent = rcu_dereference(ctx->parent_ctx);
2257 next_ctx = next->perf_event_ctxp[ctxn];
2258 if (parent && next_ctx &&
2259 rcu_dereference(next_ctx->parent_ctx) == parent) {
2261 * Looks like the two contexts are clones, so we might be
2262 * able to optimize the context switch. We lock both
2263 * contexts and check that they are clones under the
2264 * lock (including re-checking that neither has been
2265 * uncloned in the meantime). It doesn't matter which
2266 * order we take the locks because no other cpu could
2267 * be trying to lock both of these tasks.
2269 raw_spin_lock(&ctx->lock);
2270 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2271 if (context_equiv(ctx, next_ctx)) {
2273 * XXX do we need a memory barrier of sorts
2274 * wrt to rcu_dereference() of perf_event_ctxp
2276 task->perf_event_ctxp[ctxn] = next_ctx;
2277 next->perf_event_ctxp[ctxn] = ctx;
2279 next_ctx->task = task;
2282 perf_event_sync_stat(ctx, next_ctx);
2284 raw_spin_unlock(&next_ctx->lock);
2285 raw_spin_unlock(&ctx->lock);
2290 raw_spin_lock(&ctx->lock);
2291 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2292 cpuctx->task_ctx = NULL;
2293 raw_spin_unlock(&ctx->lock);
2297 #define for_each_task_context_nr(ctxn) \
2298 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2301 * Called from scheduler to remove the events of the current task,
2302 * with interrupts disabled.
2304 * We stop each event and update the event value in event->count.
2306 * This does not protect us against NMI, but disable()
2307 * sets the disabled bit in the control field of event _before_
2308 * accessing the event control register. If a NMI hits, then it will
2309 * not restart the event.
2311 void __perf_event_task_sched_out(struct task_struct *task,
2312 struct task_struct *next)
2316 for_each_task_context_nr(ctxn)
2317 perf_event_context_sched_out(task, ctxn, next);
2320 * if cgroup events exist on this CPU, then we need
2321 * to check if we have to switch out PMU state.
2322 * cgroup event are system-wide mode only
2324 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2325 perf_cgroup_sched_out(task, next);
2328 static void task_ctx_sched_out(struct perf_event_context *ctx)
2330 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2332 if (!cpuctx->task_ctx)
2335 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2338 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2339 cpuctx->task_ctx = NULL;
2343 * Called with IRQs disabled
2345 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2346 enum event_type_t event_type)
2348 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2352 ctx_pinned_sched_in(struct perf_event_context *ctx,
2353 struct perf_cpu_context *cpuctx)
2355 struct perf_event *event;
2357 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2358 if (event->state <= PERF_EVENT_STATE_OFF)
2360 if (!event_filter_match(event))
2363 /* may need to reset tstamp_enabled */
2364 if (is_cgroup_event(event))
2365 perf_cgroup_mark_enabled(event, ctx);
2367 if (group_can_go_on(event, cpuctx, 1))
2368 group_sched_in(event, cpuctx, ctx);
2371 * If this pinned group hasn't been scheduled,
2372 * put it in error state.
2374 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2375 update_group_times(event);
2376 event->state = PERF_EVENT_STATE_ERROR;
2382 ctx_flexible_sched_in(struct perf_event_context *ctx,
2383 struct perf_cpu_context *cpuctx)
2385 struct perf_event *event;
2388 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2389 /* Ignore events in OFF or ERROR state */
2390 if (event->state <= PERF_EVENT_STATE_OFF)
2393 * Listen to the 'cpu' scheduling filter constraint
2396 if (!event_filter_match(event))
2399 /* may need to reset tstamp_enabled */
2400 if (is_cgroup_event(event))
2401 perf_cgroup_mark_enabled(event, ctx);
2403 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2404 if (group_sched_in(event, cpuctx, ctx))
2411 ctx_sched_in(struct perf_event_context *ctx,
2412 struct perf_cpu_context *cpuctx,
2413 enum event_type_t event_type,
2414 struct task_struct *task)
2417 int is_active = ctx->is_active;
2419 ctx->is_active |= event_type;
2420 if (likely(!ctx->nr_events))
2424 ctx->timestamp = now;
2425 perf_cgroup_set_timestamp(task, ctx);
2427 * First go through the list and put on any pinned groups
2428 * in order to give them the best chance of going on.
2430 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2431 ctx_pinned_sched_in(ctx, cpuctx);
2433 /* Then walk through the lower prio flexible groups */
2434 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2435 ctx_flexible_sched_in(ctx, cpuctx);
2438 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2439 enum event_type_t event_type,
2440 struct task_struct *task)
2442 struct perf_event_context *ctx = &cpuctx->ctx;
2444 ctx_sched_in(ctx, cpuctx, event_type, task);
2447 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2448 struct task_struct *task)
2450 struct perf_cpu_context *cpuctx;
2452 cpuctx = __get_cpu_context(ctx);
2453 if (cpuctx->task_ctx == ctx)
2456 perf_ctx_lock(cpuctx, ctx);
2457 perf_pmu_disable(ctx->pmu);
2459 * We want to keep the following priority order:
2460 * cpu pinned (that don't need to move), task pinned,
2461 * cpu flexible, task flexible.
2463 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2466 cpuctx->task_ctx = ctx;
2468 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2470 perf_pmu_enable(ctx->pmu);
2471 perf_ctx_unlock(cpuctx, ctx);
2474 * Since these rotations are per-cpu, we need to ensure the
2475 * cpu-context we got scheduled on is actually rotating.
2477 perf_pmu_rotate_start(ctx->pmu);
2481 * When sampling the branck stack in system-wide, it may be necessary
2482 * to flush the stack on context switch. This happens when the branch
2483 * stack does not tag its entries with the pid of the current task.
2484 * Otherwise it becomes impossible to associate a branch entry with a
2485 * task. This ambiguity is more likely to appear when the branch stack
2486 * supports priv level filtering and the user sets it to monitor only
2487 * at the user level (which could be a useful measurement in system-wide
2488 * mode). In that case, the risk is high of having a branch stack with
2489 * branch from multiple tasks. Flushing may mean dropping the existing
2490 * entries or stashing them somewhere in the PMU specific code layer.
2492 * This function provides the context switch callback to the lower code
2493 * layer. It is invoked ONLY when there is at least one system-wide context
2494 * with at least one active event using taken branch sampling.
2496 static void perf_branch_stack_sched_in(struct task_struct *prev,
2497 struct task_struct *task)
2499 struct perf_cpu_context *cpuctx;
2501 unsigned long flags;
2503 /* no need to flush branch stack if not changing task */
2507 local_irq_save(flags);
2511 list_for_each_entry_rcu(pmu, &pmus, entry) {
2512 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2515 * check if the context has at least one
2516 * event using PERF_SAMPLE_BRANCH_STACK
2518 if (cpuctx->ctx.nr_branch_stack > 0
2519 && pmu->flush_branch_stack) {
2521 pmu = cpuctx->ctx.pmu;
2523 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2525 perf_pmu_disable(pmu);
2527 pmu->flush_branch_stack();
2529 perf_pmu_enable(pmu);
2531 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2537 local_irq_restore(flags);
2541 * Called from scheduler to add the events of the current task
2542 * with interrupts disabled.
2544 * We restore the event value and then enable it.
2546 * This does not protect us against NMI, but enable()
2547 * sets the enabled bit in the control field of event _before_
2548 * accessing the event control register. If a NMI hits, then it will
2549 * keep the event running.
2551 void __perf_event_task_sched_in(struct task_struct *prev,
2552 struct task_struct *task)
2554 struct perf_event_context *ctx;
2557 for_each_task_context_nr(ctxn) {
2558 ctx = task->perf_event_ctxp[ctxn];
2562 perf_event_context_sched_in(ctx, task);
2565 * if cgroup events exist on this CPU, then we need
2566 * to check if we have to switch in PMU state.
2567 * cgroup event are system-wide mode only
2569 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2570 perf_cgroup_sched_in(prev, task);
2572 /* check for system-wide branch_stack events */
2573 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2574 perf_branch_stack_sched_in(prev, task);
2577 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2579 u64 frequency = event->attr.sample_freq;
2580 u64 sec = NSEC_PER_SEC;
2581 u64 divisor, dividend;
2583 int count_fls, nsec_fls, frequency_fls, sec_fls;
2585 count_fls = fls64(count);
2586 nsec_fls = fls64(nsec);
2587 frequency_fls = fls64(frequency);
2591 * We got @count in @nsec, with a target of sample_freq HZ
2592 * the target period becomes:
2595 * period = -------------------
2596 * @nsec * sample_freq
2601 * Reduce accuracy by one bit such that @a and @b converge
2602 * to a similar magnitude.
2604 #define REDUCE_FLS(a, b) \
2606 if (a##_fls > b##_fls) { \
2616 * Reduce accuracy until either term fits in a u64, then proceed with
2617 * the other, so that finally we can do a u64/u64 division.
2619 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2620 REDUCE_FLS(nsec, frequency);
2621 REDUCE_FLS(sec, count);
2624 if (count_fls + sec_fls > 64) {
2625 divisor = nsec * frequency;
2627 while (count_fls + sec_fls > 64) {
2628 REDUCE_FLS(count, sec);
2632 dividend = count * sec;
2634 dividend = count * sec;
2636 while (nsec_fls + frequency_fls > 64) {
2637 REDUCE_FLS(nsec, frequency);
2641 divisor = nsec * frequency;
2647 return div64_u64(dividend, divisor);
2650 static DEFINE_PER_CPU(int, perf_throttled_count);
2651 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2653 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2655 struct hw_perf_event *hwc = &event->hw;
2656 s64 period, sample_period;
2659 period = perf_calculate_period(event, nsec, count);
2661 delta = (s64)(period - hwc->sample_period);
2662 delta = (delta + 7) / 8; /* low pass filter */
2664 sample_period = hwc->sample_period + delta;
2669 hwc->sample_period = sample_period;
2671 if (local64_read(&hwc->period_left) > 8*sample_period) {
2673 event->pmu->stop(event, PERF_EF_UPDATE);
2675 local64_set(&hwc->period_left, 0);
2678 event->pmu->start(event, PERF_EF_RELOAD);
2683 * combine freq adjustment with unthrottling to avoid two passes over the
2684 * events. At the same time, make sure, having freq events does not change
2685 * the rate of unthrottling as that would introduce bias.
2687 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2690 struct perf_event *event;
2691 struct hw_perf_event *hwc;
2692 u64 now, period = TICK_NSEC;
2696 * only need to iterate over all events iff:
2697 * - context have events in frequency mode (needs freq adjust)
2698 * - there are events to unthrottle on this cpu
2700 if (!(ctx->nr_freq || needs_unthr))
2703 raw_spin_lock(&ctx->lock);
2704 perf_pmu_disable(ctx->pmu);
2706 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2707 if (event->state != PERF_EVENT_STATE_ACTIVE)
2710 if (!event_filter_match(event))
2715 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2716 hwc->interrupts = 0;
2717 perf_log_throttle(event, 1);
2718 event->pmu->start(event, 0);
2721 if (!event->attr.freq || !event->attr.sample_freq)
2725 * stop the event and update event->count
2727 event->pmu->stop(event, PERF_EF_UPDATE);
2729 now = local64_read(&event->count);
2730 delta = now - hwc->freq_count_stamp;
2731 hwc->freq_count_stamp = now;
2735 * reload only if value has changed
2736 * we have stopped the event so tell that
2737 * to perf_adjust_period() to avoid stopping it
2741 perf_adjust_period(event, period, delta, false);
2743 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2746 perf_pmu_enable(ctx->pmu);
2747 raw_spin_unlock(&ctx->lock);
2751 * Round-robin a context's events:
2753 static void rotate_ctx(struct perf_event_context *ctx)
2756 * Rotate the first entry last of non-pinned groups. Rotation might be
2757 * disabled by the inheritance code.
2759 if (!ctx->rotate_disable)
2760 list_rotate_left(&ctx->flexible_groups);
2764 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2765 * because they're strictly cpu affine and rotate_start is called with IRQs
2766 * disabled, while rotate_context is called from IRQ context.
2768 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2770 struct perf_event_context *ctx = NULL;
2771 int rotate = 0, remove = 1;
2773 if (cpuctx->ctx.nr_events) {
2775 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2779 ctx = cpuctx->task_ctx;
2780 if (ctx && ctx->nr_events) {
2782 if (ctx->nr_events != ctx->nr_active)
2789 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2790 perf_pmu_disable(cpuctx->ctx.pmu);
2792 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2794 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2796 rotate_ctx(&cpuctx->ctx);
2800 perf_event_sched_in(cpuctx, ctx, current);
2802 perf_pmu_enable(cpuctx->ctx.pmu);
2803 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2806 list_del_init(&cpuctx->rotation_list);
2811 #ifdef CONFIG_NO_HZ_FULL
2812 bool perf_event_can_stop_tick(void)
2814 if (list_empty(&__get_cpu_var(rotation_list)))
2821 void perf_event_task_tick(void)
2823 struct list_head *head = &__get_cpu_var(rotation_list);
2824 struct perf_cpu_context *cpuctx, *tmp;
2825 struct perf_event_context *ctx;
2828 WARN_ON(!irqs_disabled());
2830 __this_cpu_inc(perf_throttled_seq);
2831 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2833 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2835 perf_adjust_freq_unthr_context(ctx, throttled);
2837 ctx = cpuctx->task_ctx;
2839 perf_adjust_freq_unthr_context(ctx, throttled);
2843 static int event_enable_on_exec(struct perf_event *event,
2844 struct perf_event_context *ctx)
2846 if (!event->attr.enable_on_exec)
2849 event->attr.enable_on_exec = 0;
2850 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2853 __perf_event_mark_enabled(event);
2859 * Enable all of a task's events that have been marked enable-on-exec.
2860 * This expects task == current.
2862 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2864 struct perf_event *event;
2865 unsigned long flags;
2869 local_irq_save(flags);
2870 if (!ctx || !ctx->nr_events)
2874 * We must ctxsw out cgroup events to avoid conflict
2875 * when invoking perf_task_event_sched_in() later on
2876 * in this function. Otherwise we end up trying to
2877 * ctxswin cgroup events which are already scheduled
2880 perf_cgroup_sched_out(current, NULL);
2882 raw_spin_lock(&ctx->lock);
2883 task_ctx_sched_out(ctx);
2885 list_for_each_entry(event, &ctx->event_list, event_entry) {
2886 ret = event_enable_on_exec(event, ctx);
2892 * Unclone this context if we enabled any event.
2897 raw_spin_unlock(&ctx->lock);
2900 * Also calls ctxswin for cgroup events, if any:
2902 perf_event_context_sched_in(ctx, ctx->task);
2904 local_irq_restore(flags);
2908 * Cross CPU call to read the hardware event
2910 static void __perf_event_read(void *info)
2912 struct perf_event *event = info;
2913 struct perf_event_context *ctx = event->ctx;
2914 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2917 * If this is a task context, we need to check whether it is
2918 * the current task context of this cpu. If not it has been
2919 * scheduled out before the smp call arrived. In that case
2920 * event->count would have been updated to a recent sample
2921 * when the event was scheduled out.
2923 if (ctx->task && cpuctx->task_ctx != ctx)
2926 raw_spin_lock(&ctx->lock);
2927 if (ctx->is_active) {
2928 update_context_time(ctx);
2929 update_cgrp_time_from_event(event);
2931 update_event_times(event);
2932 if (event->state == PERF_EVENT_STATE_ACTIVE)
2933 event->pmu->read(event);
2934 raw_spin_unlock(&ctx->lock);
2937 static inline u64 perf_event_count(struct perf_event *event)
2939 return local64_read(&event->count) + atomic64_read(&event->child_count);
2942 static u64 perf_event_read(struct perf_event *event)
2945 * If event is enabled and currently active on a CPU, update the
2946 * value in the event structure:
2948 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2949 smp_call_function_single(event->oncpu,
2950 __perf_event_read, event, 1);
2951 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2952 struct perf_event_context *ctx = event->ctx;
2953 unsigned long flags;
2955 raw_spin_lock_irqsave(&ctx->lock, flags);
2957 * may read while context is not active
2958 * (e.g., thread is blocked), in that case
2959 * we cannot update context time
2961 if (ctx->is_active) {
2962 update_context_time(ctx);
2963 update_cgrp_time_from_event(event);
2965 update_event_times(event);
2966 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2969 return perf_event_count(event);
2973 * Initialize the perf_event context in a task_struct:
2975 static void __perf_event_init_context(struct perf_event_context *ctx)
2977 raw_spin_lock_init(&ctx->lock);
2978 mutex_init(&ctx->mutex);
2979 INIT_LIST_HEAD(&ctx->pinned_groups);
2980 INIT_LIST_HEAD(&ctx->flexible_groups);
2981 INIT_LIST_HEAD(&ctx->event_list);
2982 atomic_set(&ctx->refcount, 1);
2985 static struct perf_event_context *
2986 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2988 struct perf_event_context *ctx;
2990 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2994 __perf_event_init_context(ctx);
2997 get_task_struct(task);
3004 static struct task_struct *
3005 find_lively_task_by_vpid(pid_t vpid)
3007 struct task_struct *task;
3014 task = find_task_by_vpid(vpid);
3016 get_task_struct(task);
3020 return ERR_PTR(-ESRCH);
3022 /* Reuse ptrace permission checks for now. */
3024 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3029 put_task_struct(task);
3030 return ERR_PTR(err);
3035 * Returns a matching context with refcount and pincount.
3037 static struct perf_event_context *
3038 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3040 struct perf_event_context *ctx;
3041 struct perf_cpu_context *cpuctx;
3042 unsigned long flags;
3046 /* Must be root to operate on a CPU event: */
3047 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3048 return ERR_PTR(-EACCES);
3051 * We could be clever and allow to attach a event to an
3052 * offline CPU and activate it when the CPU comes up, but
3055 if (!cpu_online(cpu))
3056 return ERR_PTR(-ENODEV);
3058 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3067 ctxn = pmu->task_ctx_nr;
3072 ctx = perf_lock_task_context(task, ctxn, &flags);
3076 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3078 ctx = alloc_perf_context(pmu, task);
3084 mutex_lock(&task->perf_event_mutex);
3086 * If it has already passed perf_event_exit_task().
3087 * we must see PF_EXITING, it takes this mutex too.
3089 if (task->flags & PF_EXITING)
3091 else if (task->perf_event_ctxp[ctxn])
3096 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3098 mutex_unlock(&task->perf_event_mutex);
3100 if (unlikely(err)) {
3112 return ERR_PTR(err);
3115 static void perf_event_free_filter(struct perf_event *event);
3117 static void free_event_rcu(struct rcu_head *head)
3119 struct perf_event *event;
3121 event = container_of(head, struct perf_event, rcu_head);
3123 put_pid_ns(event->ns);
3124 perf_event_free_filter(event);
3128 static void ring_buffer_put(struct ring_buffer *rb);
3129 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3131 static void free_event(struct perf_event *event)
3133 irq_work_sync(&event->pending);
3135 if (!event->parent) {
3136 if (event->attach_state & PERF_ATTACH_TASK)
3137 static_key_slow_dec_deferred(&perf_sched_events);
3138 if (event->attr.mmap || event->attr.mmap_data)
3139 atomic_dec(&nr_mmap_events);
3140 if (event->attr.comm)
3141 atomic_dec(&nr_comm_events);
3142 if (event->attr.task)
3143 atomic_dec(&nr_task_events);
3144 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3145 put_callchain_buffers();
3146 if (is_cgroup_event(event)) {
3147 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
3148 static_key_slow_dec_deferred(&perf_sched_events);
3151 if (has_branch_stack(event)) {
3152 static_key_slow_dec_deferred(&perf_sched_events);
3153 /* is system-wide event */
3154 if (!(event->attach_state & PERF_ATTACH_TASK)) {
3155 atomic_dec(&per_cpu(perf_branch_stack_events,
3162 struct ring_buffer *rb;
3165 * Can happen when we close an event with re-directed output.
3167 * Since we have a 0 refcount, perf_mmap_close() will skip
3168 * over us; possibly making our ring_buffer_put() the last.
3170 mutex_lock(&event->mmap_mutex);
3173 rcu_assign_pointer(event->rb, NULL);
3174 ring_buffer_detach(event, rb);
3175 ring_buffer_put(rb); /* could be last */
3177 mutex_unlock(&event->mmap_mutex);
3180 if (is_cgroup_event(event))
3181 perf_detach_cgroup(event);
3184 event->destroy(event);
3187 put_ctx(event->ctx);
3189 call_rcu(&event->rcu_head, free_event_rcu);
3192 int perf_event_release_kernel(struct perf_event *event)
3194 struct perf_event_context *ctx = event->ctx;
3196 WARN_ON_ONCE(ctx->parent_ctx);
3198 * There are two ways this annotation is useful:
3200 * 1) there is a lock recursion from perf_event_exit_task
3201 * see the comment there.
3203 * 2) there is a lock-inversion with mmap_sem through
3204 * perf_event_read_group(), which takes faults while
3205 * holding ctx->mutex, however this is called after
3206 * the last filedesc died, so there is no possibility
3207 * to trigger the AB-BA case.
3209 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3210 raw_spin_lock_irq(&ctx->lock);
3211 perf_group_detach(event);
3212 raw_spin_unlock_irq(&ctx->lock);
3213 perf_remove_from_context(event);
3214 mutex_unlock(&ctx->mutex);
3220 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3223 * Called when the last reference to the file is gone.
3225 static void put_event(struct perf_event *event)
3227 struct task_struct *owner;
3229 if (!atomic_long_dec_and_test(&event->refcount))
3233 owner = ACCESS_ONCE(event->owner);
3235 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3236 * !owner it means the list deletion is complete and we can indeed
3237 * free this event, otherwise we need to serialize on
3238 * owner->perf_event_mutex.
3240 smp_read_barrier_depends();
3243 * Since delayed_put_task_struct() also drops the last
3244 * task reference we can safely take a new reference
3245 * while holding the rcu_read_lock().
3247 get_task_struct(owner);
3252 mutex_lock(&owner->perf_event_mutex);
3254 * We have to re-check the event->owner field, if it is cleared
3255 * we raced with perf_event_exit_task(), acquiring the mutex
3256 * ensured they're done, and we can proceed with freeing the
3260 list_del_init(&event->owner_entry);
3261 mutex_unlock(&owner->perf_event_mutex);
3262 put_task_struct(owner);
3265 perf_event_release_kernel(event);
3268 static int perf_release(struct inode *inode, struct file *file)
3270 put_event(file->private_data);
3274 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3276 struct perf_event *child;
3282 mutex_lock(&event->child_mutex);
3283 total += perf_event_read(event);
3284 *enabled += event->total_time_enabled +
3285 atomic64_read(&event->child_total_time_enabled);
3286 *running += event->total_time_running +
3287 atomic64_read(&event->child_total_time_running);
3289 list_for_each_entry(child, &event->child_list, child_list) {
3290 total += perf_event_read(child);
3291 *enabled += child->total_time_enabled;
3292 *running += child->total_time_running;
3294 mutex_unlock(&event->child_mutex);
3298 EXPORT_SYMBOL_GPL(perf_event_read_value);
3300 static int perf_event_read_group(struct perf_event *event,
3301 u64 read_format, char __user *buf)
3303 struct perf_event *leader = event->group_leader, *sub;
3304 int n = 0, size = 0, ret = -EFAULT;
3305 struct perf_event_context *ctx = leader->ctx;
3307 u64 count, enabled, running;
3309 mutex_lock(&ctx->mutex);
3310 count = perf_event_read_value(leader, &enabled, &running);
3312 values[n++] = 1 + leader->nr_siblings;
3313 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3314 values[n++] = enabled;
3315 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3316 values[n++] = running;
3317 values[n++] = count;
3318 if (read_format & PERF_FORMAT_ID)
3319 values[n++] = primary_event_id(leader);
3321 size = n * sizeof(u64);
3323 if (copy_to_user(buf, values, size))
3328 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3331 values[n++] = perf_event_read_value(sub, &enabled, &running);
3332 if (read_format & PERF_FORMAT_ID)
3333 values[n++] = primary_event_id(sub);
3335 size = n * sizeof(u64);
3337 if (copy_to_user(buf + ret, values, size)) {
3345 mutex_unlock(&ctx->mutex);
3350 static int perf_event_read_one(struct perf_event *event,
3351 u64 read_format, char __user *buf)
3353 u64 enabled, running;
3357 values[n++] = perf_event_read_value(event, &enabled, &running);
3358 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3359 values[n++] = enabled;
3360 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3361 values[n++] = running;
3362 if (read_format & PERF_FORMAT_ID)
3363 values[n++] = primary_event_id(event);
3365 if (copy_to_user(buf, values, n * sizeof(u64)))
3368 return n * sizeof(u64);
3372 * Read the performance event - simple non blocking version for now
3375 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3377 u64 read_format = event->attr.read_format;
3381 * Return end-of-file for a read on a event that is in
3382 * error state (i.e. because it was pinned but it couldn't be
3383 * scheduled on to the CPU at some point).
3385 if (event->state == PERF_EVENT_STATE_ERROR)
3388 if (count < event->read_size)
3391 WARN_ON_ONCE(event->ctx->parent_ctx);
3392 if (read_format & PERF_FORMAT_GROUP)
3393 ret = perf_event_read_group(event, read_format, buf);
3395 ret = perf_event_read_one(event, read_format, buf);
3401 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3403 struct perf_event *event = file->private_data;
3405 return perf_read_hw(event, buf, count);
3408 static unsigned int perf_poll(struct file *file, poll_table *wait)
3410 struct perf_event *event = file->private_data;
3411 struct ring_buffer *rb;
3412 unsigned int events = POLL_HUP;
3415 * Pin the event->rb by taking event->mmap_mutex; otherwise
3416 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3418 mutex_lock(&event->mmap_mutex);
3421 events = atomic_xchg(&rb->poll, 0);
3422 mutex_unlock(&event->mmap_mutex);
3424 poll_wait(file, &event->waitq, wait);
3429 static void perf_event_reset(struct perf_event *event)
3431 (void)perf_event_read(event);
3432 local64_set(&event->count, 0);
3433 perf_event_update_userpage(event);
3437 * Holding the top-level event's child_mutex means that any
3438 * descendant process that has inherited this event will block
3439 * in sync_child_event if it goes to exit, thus satisfying the
3440 * task existence requirements of perf_event_enable/disable.
3442 static void perf_event_for_each_child(struct perf_event *event,
3443 void (*func)(struct perf_event *))
3445 struct perf_event *child;
3447 WARN_ON_ONCE(event->ctx->parent_ctx);
3448 mutex_lock(&event->child_mutex);
3450 list_for_each_entry(child, &event->child_list, child_list)
3452 mutex_unlock(&event->child_mutex);
3455 static void perf_event_for_each(struct perf_event *event,
3456 void (*func)(struct perf_event *))
3458 struct perf_event_context *ctx = event->ctx;
3459 struct perf_event *sibling;
3461 WARN_ON_ONCE(ctx->parent_ctx);
3462 mutex_lock(&ctx->mutex);
3463 event = event->group_leader;
3465 perf_event_for_each_child(event, func);
3466 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3467 perf_event_for_each_child(sibling, func);
3468 mutex_unlock(&ctx->mutex);
3471 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3473 struct perf_event_context *ctx = event->ctx;
3477 if (!is_sampling_event(event))
3480 if (copy_from_user(&value, arg, sizeof(value)))
3486 raw_spin_lock_irq(&ctx->lock);
3487 if (event->attr.freq) {
3488 if (value > sysctl_perf_event_sample_rate) {
3493 event->attr.sample_freq = value;
3495 event->attr.sample_period = value;
3496 event->hw.sample_period = value;
3499 raw_spin_unlock_irq(&ctx->lock);
3504 static const struct file_operations perf_fops;
3506 static inline int perf_fget_light(int fd, struct fd *p)
3508 struct fd f = fdget(fd);
3512 if (f.file->f_op != &perf_fops) {
3520 static int perf_event_set_output(struct perf_event *event,
3521 struct perf_event *output_event);
3522 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3524 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3526 struct perf_event *event = file->private_data;
3527 void (*func)(struct perf_event *);
3531 case PERF_EVENT_IOC_ENABLE:
3532 func = perf_event_enable;
3534 case PERF_EVENT_IOC_DISABLE:
3535 func = perf_event_disable;
3537 case PERF_EVENT_IOC_RESET:
3538 func = perf_event_reset;
3541 case PERF_EVENT_IOC_REFRESH:
3542 return perf_event_refresh(event, arg);
3544 case PERF_EVENT_IOC_PERIOD:
3545 return perf_event_period(event, (u64 __user *)arg);
3547 case PERF_EVENT_IOC_SET_OUTPUT:
3551 struct perf_event *output_event;
3553 ret = perf_fget_light(arg, &output);
3556 output_event = output.file->private_data;
3557 ret = perf_event_set_output(event, output_event);
3560 ret = perf_event_set_output(event, NULL);
3565 case PERF_EVENT_IOC_SET_FILTER:
3566 return perf_event_set_filter(event, (void __user *)arg);
3572 if (flags & PERF_IOC_FLAG_GROUP)
3573 perf_event_for_each(event, func);
3575 perf_event_for_each_child(event, func);
3580 int perf_event_task_enable(void)
3582 struct perf_event *event;
3584 mutex_lock(¤t->perf_event_mutex);
3585 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3586 perf_event_for_each_child(event, perf_event_enable);
3587 mutex_unlock(¤t->perf_event_mutex);
3592 int perf_event_task_disable(void)
3594 struct perf_event *event;
3596 mutex_lock(¤t->perf_event_mutex);
3597 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3598 perf_event_for_each_child(event, perf_event_disable);
3599 mutex_unlock(¤t->perf_event_mutex);
3604 static int perf_event_index(struct perf_event *event)
3606 if (event->hw.state & PERF_HES_STOPPED)
3609 if (event->state != PERF_EVENT_STATE_ACTIVE)
3612 return event->pmu->event_idx(event);
3615 static void calc_timer_values(struct perf_event *event,
3622 *now = perf_clock();
3623 ctx_time = event->shadow_ctx_time + *now;
3624 *enabled = ctx_time - event->tstamp_enabled;
3625 *running = ctx_time - event->tstamp_running;
3628 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3633 * Callers need to ensure there can be no nesting of this function, otherwise
3634 * the seqlock logic goes bad. We can not serialize this because the arch
3635 * code calls this from NMI context.
3637 void perf_event_update_userpage(struct perf_event *event)
3639 struct perf_event_mmap_page *userpg;
3640 struct ring_buffer *rb;
3641 u64 enabled, running, now;
3645 * compute total_time_enabled, total_time_running
3646 * based on snapshot values taken when the event
3647 * was last scheduled in.
3649 * we cannot simply called update_context_time()
3650 * because of locking issue as we can be called in
3653 calc_timer_values(event, &now, &enabled, &running);
3654 rb = rcu_dereference(event->rb);
3658 userpg = rb->user_page;
3661 * Disable preemption so as to not let the corresponding user-space
3662 * spin too long if we get preempted.
3667 userpg->index = perf_event_index(event);
3668 userpg->offset = perf_event_count(event);
3670 userpg->offset -= local64_read(&event->hw.prev_count);
3672 userpg->time_enabled = enabled +
3673 atomic64_read(&event->child_total_time_enabled);
3675 userpg->time_running = running +
3676 atomic64_read(&event->child_total_time_running);
3678 arch_perf_update_userpage(userpg, now);
3687 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3689 struct perf_event *event = vma->vm_file->private_data;
3690 struct ring_buffer *rb;
3691 int ret = VM_FAULT_SIGBUS;
3693 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3694 if (vmf->pgoff == 0)
3700 rb = rcu_dereference(event->rb);
3704 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3707 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3711 get_page(vmf->page);
3712 vmf->page->mapping = vma->vm_file->f_mapping;
3713 vmf->page->index = vmf->pgoff;
3722 static void ring_buffer_attach(struct perf_event *event,
3723 struct ring_buffer *rb)
3725 unsigned long flags;
3727 if (!list_empty(&event->rb_entry))
3730 spin_lock_irqsave(&rb->event_lock, flags);
3731 if (list_empty(&event->rb_entry))
3732 list_add(&event->rb_entry, &rb->event_list);
3733 spin_unlock_irqrestore(&rb->event_lock, flags);
3736 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3738 unsigned long flags;
3740 if (list_empty(&event->rb_entry))
3743 spin_lock_irqsave(&rb->event_lock, flags);
3744 list_del_init(&event->rb_entry);
3745 wake_up_all(&event->waitq);
3746 spin_unlock_irqrestore(&rb->event_lock, flags);
3749 static void ring_buffer_wakeup(struct perf_event *event)
3751 struct ring_buffer *rb;
3754 rb = rcu_dereference(event->rb);
3756 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3757 wake_up_all(&event->waitq);
3762 static void rb_free_rcu(struct rcu_head *rcu_head)
3764 struct ring_buffer *rb;
3766 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3770 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3772 struct ring_buffer *rb;
3775 rb = rcu_dereference(event->rb);
3777 if (!atomic_inc_not_zero(&rb->refcount))
3785 static void ring_buffer_put(struct ring_buffer *rb)
3787 if (!atomic_dec_and_test(&rb->refcount))
3790 WARN_ON_ONCE(!list_empty(&rb->event_list));
3792 call_rcu(&rb->rcu_head, rb_free_rcu);
3795 static void perf_mmap_open(struct vm_area_struct *vma)
3797 struct perf_event *event = vma->vm_file->private_data;
3799 atomic_inc(&event->mmap_count);
3800 atomic_inc(&event->rb->mmap_count);
3804 * A buffer can be mmap()ed multiple times; either directly through the same
3805 * event, or through other events by use of perf_event_set_output().
3807 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3808 * the buffer here, where we still have a VM context. This means we need
3809 * to detach all events redirecting to us.
3811 static void perf_mmap_close(struct vm_area_struct *vma)
3813 struct perf_event *event = vma->vm_file->private_data;
3815 struct ring_buffer *rb = event->rb;
3816 struct user_struct *mmap_user = rb->mmap_user;
3817 int mmap_locked = rb->mmap_locked;
3818 unsigned long size = perf_data_size(rb);
3820 atomic_dec(&rb->mmap_count);
3822 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3825 /* Detach current event from the buffer. */
3826 rcu_assign_pointer(event->rb, NULL);
3827 ring_buffer_detach(event, rb);
3828 mutex_unlock(&event->mmap_mutex);
3830 /* If there's still other mmap()s of this buffer, we're done. */
3831 if (atomic_read(&rb->mmap_count)) {
3832 ring_buffer_put(rb); /* can't be last */
3837 * No other mmap()s, detach from all other events that might redirect
3838 * into the now unreachable buffer. Somewhat complicated by the
3839 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3843 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3844 if (!atomic_long_inc_not_zero(&event->refcount)) {
3846 * This event is en-route to free_event() which will
3847 * detach it and remove it from the list.
3853 mutex_lock(&event->mmap_mutex);
3855 * Check we didn't race with perf_event_set_output() which can
3856 * swizzle the rb from under us while we were waiting to
3857 * acquire mmap_mutex.
3859 * If we find a different rb; ignore this event, a next
3860 * iteration will no longer find it on the list. We have to
3861 * still restart the iteration to make sure we're not now
3862 * iterating the wrong list.
3864 if (event->rb == rb) {
3865 rcu_assign_pointer(event->rb, NULL);
3866 ring_buffer_detach(event, rb);
3867 ring_buffer_put(rb); /* can't be last, we still have one */
3869 mutex_unlock(&event->mmap_mutex);
3873 * Restart the iteration; either we're on the wrong list or
3874 * destroyed its integrity by doing a deletion.
3881 * It could be there's still a few 0-ref events on the list; they'll
3882 * get cleaned up by free_event() -- they'll also still have their
3883 * ref on the rb and will free it whenever they are done with it.
3885 * Aside from that, this buffer is 'fully' detached and unmapped,
3886 * undo the VM accounting.
3889 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3890 vma->vm_mm->pinned_vm -= mmap_locked;
3891 free_uid(mmap_user);
3893 ring_buffer_put(rb); /* could be last */
3896 static const struct vm_operations_struct perf_mmap_vmops = {
3897 .open = perf_mmap_open,
3898 .close = perf_mmap_close,
3899 .fault = perf_mmap_fault,
3900 .page_mkwrite = perf_mmap_fault,
3903 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3905 struct perf_event *event = file->private_data;
3906 unsigned long user_locked, user_lock_limit;
3907 struct user_struct *user = current_user();
3908 unsigned long locked, lock_limit;
3909 struct ring_buffer *rb;
3910 unsigned long vma_size;
3911 unsigned long nr_pages;
3912 long user_extra, extra;
3913 int ret = 0, flags = 0;
3916 * Don't allow mmap() of inherited per-task counters. This would
3917 * create a performance issue due to all children writing to the
3920 if (event->cpu == -1 && event->attr.inherit)
3923 if (!(vma->vm_flags & VM_SHARED))
3926 vma_size = vma->vm_end - vma->vm_start;
3927 nr_pages = (vma_size / PAGE_SIZE) - 1;
3930 * If we have rb pages ensure they're a power-of-two number, so we
3931 * can do bitmasks instead of modulo.
3933 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3936 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3939 if (vma->vm_pgoff != 0)
3942 WARN_ON_ONCE(event->ctx->parent_ctx);
3944 mutex_lock(&event->mmap_mutex);
3946 if (event->rb->nr_pages != nr_pages) {
3951 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
3953 * Raced against perf_mmap_close() through
3954 * perf_event_set_output(). Try again, hope for better
3957 mutex_unlock(&event->mmap_mutex);
3964 user_extra = nr_pages + 1;
3965 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3968 * Increase the limit linearly with more CPUs:
3970 user_lock_limit *= num_online_cpus();
3972 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3975 if (user_locked > user_lock_limit)
3976 extra = user_locked - user_lock_limit;
3978 lock_limit = rlimit(RLIMIT_MEMLOCK);
3979 lock_limit >>= PAGE_SHIFT;
3980 locked = vma->vm_mm->pinned_vm + extra;
3982 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3983 !capable(CAP_IPC_LOCK)) {
3990 if (vma->vm_flags & VM_WRITE)
3991 flags |= RING_BUFFER_WRITABLE;
3993 rb = rb_alloc(nr_pages,
3994 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4002 atomic_set(&rb->mmap_count, 1);
4003 rb->mmap_locked = extra;
4004 rb->mmap_user = get_current_user();
4006 atomic_long_add(user_extra, &user->locked_vm);
4007 vma->vm_mm->pinned_vm += extra;
4009 ring_buffer_attach(event, rb);
4010 rcu_assign_pointer(event->rb, rb);
4012 perf_event_update_userpage(event);
4016 atomic_inc(&event->mmap_count);
4017 mutex_unlock(&event->mmap_mutex);
4020 * Since pinned accounting is per vm we cannot allow fork() to copy our
4023 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4024 vma->vm_ops = &perf_mmap_vmops;
4029 static int perf_fasync(int fd, struct file *filp, int on)
4031 struct inode *inode = file_inode(filp);
4032 struct perf_event *event = filp->private_data;
4035 mutex_lock(&inode->i_mutex);
4036 retval = fasync_helper(fd, filp, on, &event->fasync);
4037 mutex_unlock(&inode->i_mutex);
4045 static const struct file_operations perf_fops = {
4046 .llseek = no_llseek,
4047 .release = perf_release,
4050 .unlocked_ioctl = perf_ioctl,
4051 .compat_ioctl = perf_ioctl,
4053 .fasync = perf_fasync,
4059 * If there's data, ensure we set the poll() state and publish everything
4060 * to user-space before waking everybody up.
4063 void perf_event_wakeup(struct perf_event *event)
4065 ring_buffer_wakeup(event);
4067 if (event->pending_kill) {
4068 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4069 event->pending_kill = 0;
4073 static void perf_pending_event(struct irq_work *entry)
4075 struct perf_event *event = container_of(entry,
4076 struct perf_event, pending);
4078 if (event->pending_disable) {
4079 event->pending_disable = 0;
4080 __perf_event_disable(event);
4083 if (event->pending_wakeup) {
4084 event->pending_wakeup = 0;
4085 perf_event_wakeup(event);
4090 * We assume there is only KVM supporting the callbacks.
4091 * Later on, we might change it to a list if there is
4092 * another virtualization implementation supporting the callbacks.
4094 struct perf_guest_info_callbacks *perf_guest_cbs;
4096 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4098 perf_guest_cbs = cbs;
4101 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4103 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4105 perf_guest_cbs = NULL;
4108 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4111 perf_output_sample_regs(struct perf_output_handle *handle,
4112 struct pt_regs *regs, u64 mask)
4116 for_each_set_bit(bit, (const unsigned long *) &mask,
4117 sizeof(mask) * BITS_PER_BYTE) {
4120 val = perf_reg_value(regs, bit);
4121 perf_output_put(handle, val);
4125 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4126 struct pt_regs *regs)
4128 if (!user_mode(regs)) {
4130 regs = task_pt_regs(current);
4136 regs_user->regs = regs;
4137 regs_user->abi = perf_reg_abi(current);
4142 * Get remaining task size from user stack pointer.
4144 * It'd be better to take stack vma map and limit this more
4145 * precisly, but there's no way to get it safely under interrupt,
4146 * so using TASK_SIZE as limit.
4148 static u64 perf_ustack_task_size(struct pt_regs *regs)
4150 unsigned long addr = perf_user_stack_pointer(regs);
4152 if (!addr || addr >= TASK_SIZE)
4155 return TASK_SIZE - addr;
4159 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4160 struct pt_regs *regs)
4164 /* No regs, no stack pointer, no dump. */
4169 * Check if we fit in with the requested stack size into the:
4171 * If we don't, we limit the size to the TASK_SIZE.
4173 * - remaining sample size
4174 * If we don't, we customize the stack size to
4175 * fit in to the remaining sample size.
4178 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4179 stack_size = min(stack_size, (u16) task_size);
4181 /* Current header size plus static size and dynamic size. */
4182 header_size += 2 * sizeof(u64);
4184 /* Do we fit in with the current stack dump size? */
4185 if ((u16) (header_size + stack_size) < header_size) {
4187 * If we overflow the maximum size for the sample,
4188 * we customize the stack dump size to fit in.
4190 stack_size = USHRT_MAX - header_size - sizeof(u64);
4191 stack_size = round_up(stack_size, sizeof(u64));
4198 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4199 struct pt_regs *regs)
4201 /* Case of a kernel thread, nothing to dump */
4204 perf_output_put(handle, size);
4213 * - the size requested by user or the best one we can fit
4214 * in to the sample max size
4216 * - user stack dump data
4218 * - the actual dumped size
4222 perf_output_put(handle, dump_size);
4225 sp = perf_user_stack_pointer(regs);
4226 rem = __output_copy_user(handle, (void *) sp, dump_size);
4227 dyn_size = dump_size - rem;
4229 perf_output_skip(handle, rem);
4232 perf_output_put(handle, dyn_size);
4236 static void __perf_event_header__init_id(struct perf_event_header *header,
4237 struct perf_sample_data *data,
4238 struct perf_event *event)
4240 u64 sample_type = event->attr.sample_type;
4242 data->type = sample_type;
4243 header->size += event->id_header_size;
4245 if (sample_type & PERF_SAMPLE_TID) {
4246 /* namespace issues */
4247 data->tid_entry.pid = perf_event_pid(event, current);
4248 data->tid_entry.tid = perf_event_tid(event, current);
4251 if (sample_type & PERF_SAMPLE_TIME)
4252 data->time = perf_clock();
4254 if (sample_type & PERF_SAMPLE_ID)
4255 data->id = primary_event_id(event);
4257 if (sample_type & PERF_SAMPLE_STREAM_ID)
4258 data->stream_id = event->id;
4260 if (sample_type & PERF_SAMPLE_CPU) {
4261 data->cpu_entry.cpu = raw_smp_processor_id();
4262 data->cpu_entry.reserved = 0;
4266 void perf_event_header__init_id(struct perf_event_header *header,
4267 struct perf_sample_data *data,
4268 struct perf_event *event)
4270 if (event->attr.sample_id_all)
4271 __perf_event_header__init_id(header, data, event);
4274 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4275 struct perf_sample_data *data)
4277 u64 sample_type = data->type;
4279 if (sample_type & PERF_SAMPLE_TID)
4280 perf_output_put(handle, data->tid_entry);
4282 if (sample_type & PERF_SAMPLE_TIME)
4283 perf_output_put(handle, data->time);
4285 if (sample_type & PERF_SAMPLE_ID)
4286 perf_output_put(handle, data->id);
4288 if (sample_type & PERF_SAMPLE_STREAM_ID)
4289 perf_output_put(handle, data->stream_id);
4291 if (sample_type & PERF_SAMPLE_CPU)
4292 perf_output_put(handle, data->cpu_entry);
4295 void perf_event__output_id_sample(struct perf_event *event,
4296 struct perf_output_handle *handle,
4297 struct perf_sample_data *sample)
4299 if (event->attr.sample_id_all)
4300 __perf_event__output_id_sample(handle, sample);
4303 static void perf_output_read_one(struct perf_output_handle *handle,
4304 struct perf_event *event,
4305 u64 enabled, u64 running)
4307 u64 read_format = event->attr.read_format;
4311 values[n++] = perf_event_count(event);
4312 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4313 values[n++] = enabled +
4314 atomic64_read(&event->child_total_time_enabled);
4316 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4317 values[n++] = running +
4318 atomic64_read(&event->child_total_time_running);
4320 if (read_format & PERF_FORMAT_ID)
4321 values[n++] = primary_event_id(event);
4323 __output_copy(handle, values, n * sizeof(u64));
4327 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4329 static void perf_output_read_group(struct perf_output_handle *handle,
4330 struct perf_event *event,
4331 u64 enabled, u64 running)
4333 struct perf_event *leader = event->group_leader, *sub;
4334 u64 read_format = event->attr.read_format;
4338 values[n++] = 1 + leader->nr_siblings;
4340 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4341 values[n++] = enabled;
4343 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4344 values[n++] = running;
4346 if (leader != event)
4347 leader->pmu->read(leader);
4349 values[n++] = perf_event_count(leader);
4350 if (read_format & PERF_FORMAT_ID)
4351 values[n++] = primary_event_id(leader);
4353 __output_copy(handle, values, n * sizeof(u64));
4355 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4359 sub->pmu->read(sub);
4361 values[n++] = perf_event_count(sub);
4362 if (read_format & PERF_FORMAT_ID)
4363 values[n++] = primary_event_id(sub);
4365 __output_copy(handle, values, n * sizeof(u64));
4369 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4370 PERF_FORMAT_TOTAL_TIME_RUNNING)
4372 static void perf_output_read(struct perf_output_handle *handle,
4373 struct perf_event *event)
4375 u64 enabled = 0, running = 0, now;
4376 u64 read_format = event->attr.read_format;
4379 * compute total_time_enabled, total_time_running
4380 * based on snapshot values taken when the event
4381 * was last scheduled in.
4383 * we cannot simply called update_context_time()
4384 * because of locking issue as we are called in
4387 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4388 calc_timer_values(event, &now, &enabled, &running);
4390 if (event->attr.read_format & PERF_FORMAT_GROUP)
4391 perf_output_read_group(handle, event, enabled, running);
4393 perf_output_read_one(handle, event, enabled, running);
4396 void perf_output_sample(struct perf_output_handle *handle,
4397 struct perf_event_header *header,
4398 struct perf_sample_data *data,
4399 struct perf_event *event)
4401 u64 sample_type = data->type;
4403 perf_output_put(handle, *header);
4405 if (sample_type & PERF_SAMPLE_IP)
4406 perf_output_put(handle, data->ip);
4408 if (sample_type & PERF_SAMPLE_TID)
4409 perf_output_put(handle, data->tid_entry);
4411 if (sample_type & PERF_SAMPLE_TIME)
4412 perf_output_put(handle, data->time);
4414 if (sample_type & PERF_SAMPLE_ADDR)
4415 perf_output_put(handle, data->addr);
4417 if (sample_type & PERF_SAMPLE_ID)
4418 perf_output_put(handle, data->id);
4420 if (sample_type & PERF_SAMPLE_STREAM_ID)
4421 perf_output_put(handle, data->stream_id);
4423 if (sample_type & PERF_SAMPLE_CPU)
4424 perf_output_put(handle, data->cpu_entry);
4426 if (sample_type & PERF_SAMPLE_PERIOD)
4427 perf_output_put(handle, data->period);
4429 if (sample_type & PERF_SAMPLE_READ)
4430 perf_output_read(handle, event);
4432 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4433 if (data->callchain) {
4436 if (data->callchain)
4437 size += data->callchain->nr;
4439 size *= sizeof(u64);
4441 __output_copy(handle, data->callchain, size);
4444 perf_output_put(handle, nr);
4448 if (sample_type & PERF_SAMPLE_RAW) {
4450 perf_output_put(handle, data->raw->size);
4451 __output_copy(handle, data->raw->data,
4458 .size = sizeof(u32),
4461 perf_output_put(handle, raw);
4465 if (!event->attr.watermark) {
4466 int wakeup_events = event->attr.wakeup_events;
4468 if (wakeup_events) {
4469 struct ring_buffer *rb = handle->rb;
4470 int events = local_inc_return(&rb->events);
4472 if (events >= wakeup_events) {
4473 local_sub(wakeup_events, &rb->events);
4474 local_inc(&rb->wakeup);
4479 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4480 if (data->br_stack) {
4483 size = data->br_stack->nr
4484 * sizeof(struct perf_branch_entry);
4486 perf_output_put(handle, data->br_stack->nr);
4487 perf_output_copy(handle, data->br_stack->entries, size);
4490 * we always store at least the value of nr
4493 perf_output_put(handle, nr);
4497 if (sample_type & PERF_SAMPLE_REGS_USER) {
4498 u64 abi = data->regs_user.abi;
4501 * If there are no regs to dump, notice it through
4502 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4504 perf_output_put(handle, abi);
4507 u64 mask = event->attr.sample_regs_user;
4508 perf_output_sample_regs(handle,
4509 data->regs_user.regs,
4514 if (sample_type & PERF_SAMPLE_STACK_USER)
4515 perf_output_sample_ustack(handle,
4516 data->stack_user_size,
4517 data->regs_user.regs);
4519 if (sample_type & PERF_SAMPLE_WEIGHT)
4520 perf_output_put(handle, data->weight);
4522 if (sample_type & PERF_SAMPLE_DATA_SRC)
4523 perf_output_put(handle, data->data_src.val);
4526 void perf_prepare_sample(struct perf_event_header *header,
4527 struct perf_sample_data *data,
4528 struct perf_event *event,
4529 struct pt_regs *regs)
4531 u64 sample_type = event->attr.sample_type;
4533 header->type = PERF_RECORD_SAMPLE;
4534 header->size = sizeof(*header) + event->header_size;
4537 header->misc |= perf_misc_flags(regs);
4539 __perf_event_header__init_id(header, data, event);
4541 if (sample_type & PERF_SAMPLE_IP)
4542 data->ip = perf_instruction_pointer(regs);
4544 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4547 data->callchain = perf_callchain(event, regs);
4549 if (data->callchain)
4550 size += data->callchain->nr;
4552 header->size += size * sizeof(u64);
4555 if (sample_type & PERF_SAMPLE_RAW) {
4556 int size = sizeof(u32);
4559 size += data->raw->size;
4561 size += sizeof(u32);
4563 WARN_ON_ONCE(size & (sizeof(u64)-1));
4564 header->size += size;
4567 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4568 int size = sizeof(u64); /* nr */
4569 if (data->br_stack) {
4570 size += data->br_stack->nr
4571 * sizeof(struct perf_branch_entry);
4573 header->size += size;
4576 if (sample_type & PERF_SAMPLE_REGS_USER) {
4577 /* regs dump ABI info */
4578 int size = sizeof(u64);
4580 perf_sample_regs_user(&data->regs_user, regs);
4582 if (data->regs_user.regs) {
4583 u64 mask = event->attr.sample_regs_user;
4584 size += hweight64(mask) * sizeof(u64);
4587 header->size += size;
4590 if (sample_type & PERF_SAMPLE_STACK_USER) {
4592 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4593 * processed as the last one or have additional check added
4594 * in case new sample type is added, because we could eat
4595 * up the rest of the sample size.
4597 struct perf_regs_user *uregs = &data->regs_user;
4598 u16 stack_size = event->attr.sample_stack_user;
4599 u16 size = sizeof(u64);
4602 perf_sample_regs_user(uregs, regs);
4604 stack_size = perf_sample_ustack_size(stack_size, header->size,
4608 * If there is something to dump, add space for the dump
4609 * itself and for the field that tells the dynamic size,
4610 * which is how many have been actually dumped.
4613 size += sizeof(u64) + stack_size;
4615 data->stack_user_size = stack_size;
4616 header->size += size;
4620 static void perf_event_output(struct perf_event *event,
4621 struct perf_sample_data *data,
4622 struct pt_regs *regs)
4624 struct perf_output_handle handle;
4625 struct perf_event_header header;
4627 /* protect the callchain buffers */
4630 perf_prepare_sample(&header, data, event, regs);
4632 if (perf_output_begin(&handle, event, header.size))
4635 perf_output_sample(&handle, &header, data, event);
4637 perf_output_end(&handle);
4647 struct perf_read_event {
4648 struct perf_event_header header;
4655 perf_event_read_event(struct perf_event *event,
4656 struct task_struct *task)
4658 struct perf_output_handle handle;
4659 struct perf_sample_data sample;
4660 struct perf_read_event read_event = {
4662 .type = PERF_RECORD_READ,
4664 .size = sizeof(read_event) + event->read_size,
4666 .pid = perf_event_pid(event, task),
4667 .tid = perf_event_tid(event, task),
4671 perf_event_header__init_id(&read_event.header, &sample, event);
4672 ret = perf_output_begin(&handle, event, read_event.header.size);
4676 perf_output_put(&handle, read_event);
4677 perf_output_read(&handle, event);
4678 perf_event__output_id_sample(event, &handle, &sample);
4680 perf_output_end(&handle);
4683 typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data);
4684 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4687 perf_event_aux_ctx(struct perf_event_context *ctx,
4688 perf_event_aux_match_cb match,
4689 perf_event_aux_output_cb output,
4692 struct perf_event *event;
4694 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4695 if (event->state < PERF_EVENT_STATE_INACTIVE)
4697 if (!event_filter_match(event))
4699 if (match(event, data))
4700 output(event, data);
4705 perf_event_aux(perf_event_aux_match_cb match,
4706 perf_event_aux_output_cb output,
4708 struct perf_event_context *task_ctx)
4710 struct perf_cpu_context *cpuctx;
4711 struct perf_event_context *ctx;
4716 list_for_each_entry_rcu(pmu, &pmus, entry) {
4717 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4718 if (cpuctx->unique_pmu != pmu)
4720 perf_event_aux_ctx(&cpuctx->ctx, match, output, data);
4723 ctxn = pmu->task_ctx_nr;
4726 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4728 perf_event_aux_ctx(ctx, match, output, data);
4730 put_cpu_ptr(pmu->pmu_cpu_context);
4735 perf_event_aux_ctx(task_ctx, match, output, data);
4742 * task tracking -- fork/exit
4744 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4747 struct perf_task_event {
4748 struct task_struct *task;
4749 struct perf_event_context *task_ctx;
4752 struct perf_event_header header;
4762 static void perf_event_task_output(struct perf_event *event,
4765 struct perf_task_event *task_event = data;
4766 struct perf_output_handle handle;
4767 struct perf_sample_data sample;
4768 struct task_struct *task = task_event->task;
4769 int ret, size = task_event->event_id.header.size;
4771 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4773 ret = perf_output_begin(&handle, event,
4774 task_event->event_id.header.size);
4778 task_event->event_id.pid = perf_event_pid(event, task);
4779 task_event->event_id.ppid = perf_event_pid(event, current);
4781 task_event->event_id.tid = perf_event_tid(event, task);
4782 task_event->event_id.ptid = perf_event_tid(event, current);
4784 perf_output_put(&handle, task_event->event_id);
4786 perf_event__output_id_sample(event, &handle, &sample);
4788 perf_output_end(&handle);
4790 task_event->event_id.header.size = size;
4793 static int perf_event_task_match(struct perf_event *event,
4794 void *data __maybe_unused)
4796 return event->attr.comm || event->attr.mmap ||
4797 event->attr.mmap_data || event->attr.task;
4800 static void perf_event_task(struct task_struct *task,
4801 struct perf_event_context *task_ctx,
4804 struct perf_task_event task_event;
4806 if (!atomic_read(&nr_comm_events) &&
4807 !atomic_read(&nr_mmap_events) &&
4808 !atomic_read(&nr_task_events))
4811 task_event = (struct perf_task_event){
4813 .task_ctx = task_ctx,
4816 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4818 .size = sizeof(task_event.event_id),
4824 .time = perf_clock(),
4828 perf_event_aux(perf_event_task_match,
4829 perf_event_task_output,
4834 void perf_event_fork(struct task_struct *task)
4836 perf_event_task(task, NULL, 1);
4843 struct perf_comm_event {
4844 struct task_struct *task;
4849 struct perf_event_header header;
4856 static void perf_event_comm_output(struct perf_event *event,
4859 struct perf_comm_event *comm_event = data;
4860 struct perf_output_handle handle;
4861 struct perf_sample_data sample;
4862 int size = comm_event->event_id.header.size;
4865 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4866 ret = perf_output_begin(&handle, event,
4867 comm_event->event_id.header.size);
4872 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4873 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4875 perf_output_put(&handle, comm_event->event_id);
4876 __output_copy(&handle, comm_event->comm,
4877 comm_event->comm_size);
4879 perf_event__output_id_sample(event, &handle, &sample);
4881 perf_output_end(&handle);
4883 comm_event->event_id.header.size = size;
4886 static int perf_event_comm_match(struct perf_event *event,
4887 void *data __maybe_unused)
4889 return event->attr.comm;
4892 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4894 char comm[TASK_COMM_LEN];
4897 memset(comm, 0, sizeof(comm));
4898 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4899 size = ALIGN(strlen(comm)+1, sizeof(u64));
4901 comm_event->comm = comm;
4902 comm_event->comm_size = size;
4904 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4906 perf_event_aux(perf_event_comm_match,
4907 perf_event_comm_output,
4912 void perf_event_comm(struct task_struct *task)
4914 struct perf_comm_event comm_event;
4915 struct perf_event_context *ctx;
4919 for_each_task_context_nr(ctxn) {
4920 ctx = task->perf_event_ctxp[ctxn];
4924 perf_event_enable_on_exec(ctx);
4928 if (!atomic_read(&nr_comm_events))
4931 comm_event = (struct perf_comm_event){
4937 .type = PERF_RECORD_COMM,
4946 perf_event_comm_event(&comm_event);
4953 struct perf_mmap_event {
4954 struct vm_area_struct *vma;
4956 const char *file_name;
4960 struct perf_event_header header;
4970 static void perf_event_mmap_output(struct perf_event *event,
4973 struct perf_mmap_event *mmap_event = data;
4974 struct perf_output_handle handle;
4975 struct perf_sample_data sample;
4976 int size = mmap_event->event_id.header.size;
4979 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4980 ret = perf_output_begin(&handle, event,
4981 mmap_event->event_id.header.size);
4985 mmap_event->event_id.pid = perf_event_pid(event, current);
4986 mmap_event->event_id.tid = perf_event_tid(event, current);
4988 perf_output_put(&handle, mmap_event->event_id);
4989 __output_copy(&handle, mmap_event->file_name,
4990 mmap_event->file_size);
4992 perf_event__output_id_sample(event, &handle, &sample);
4994 perf_output_end(&handle);
4996 mmap_event->event_id.header.size = size;
4999 static int perf_event_mmap_match(struct perf_event *event,
5002 struct perf_mmap_event *mmap_event = data;
5003 struct vm_area_struct *vma = mmap_event->vma;
5004 int executable = vma->vm_flags & VM_EXEC;
5006 return (!executable && event->attr.mmap_data) ||
5007 (executable && event->attr.mmap);
5010 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5012 struct vm_area_struct *vma = mmap_event->vma;
5013 struct file *file = vma->vm_file;
5019 memset(tmp, 0, sizeof(tmp));
5023 * d_path works from the end of the rb backwards, so we
5024 * need to add enough zero bytes after the string to handle
5025 * the 64bit alignment we do later.
5027 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
5029 name = strncpy(tmp, "//enomem", sizeof(tmp));
5032 name = d_path(&file->f_path, buf, PATH_MAX);
5034 name = strncpy(tmp, "//toolong", sizeof(tmp));
5038 if (arch_vma_name(mmap_event->vma)) {
5039 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
5041 tmp[sizeof(tmp) - 1] = '\0';
5046 name = strncpy(tmp, "[vdso]", sizeof(tmp));
5048 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
5049 vma->vm_end >= vma->vm_mm->brk) {
5050 name = strncpy(tmp, "[heap]", sizeof(tmp));
5052 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
5053 vma->vm_end >= vma->vm_mm->start_stack) {
5054 name = strncpy(tmp, "[stack]", sizeof(tmp));
5058 name = strncpy(tmp, "//anon", sizeof(tmp));
5063 size = ALIGN(strlen(name)+1, sizeof(u64));
5065 mmap_event->file_name = name;
5066 mmap_event->file_size = size;
5068 if (!(vma->vm_flags & VM_EXEC))
5069 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5071 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5073 perf_event_aux(perf_event_mmap_match,
5074 perf_event_mmap_output,
5081 void perf_event_mmap(struct vm_area_struct *vma)
5083 struct perf_mmap_event mmap_event;
5085 if (!atomic_read(&nr_mmap_events))
5088 mmap_event = (struct perf_mmap_event){
5094 .type = PERF_RECORD_MMAP,
5095 .misc = PERF_RECORD_MISC_USER,
5100 .start = vma->vm_start,
5101 .len = vma->vm_end - vma->vm_start,
5102 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5106 perf_event_mmap_event(&mmap_event);
5110 * IRQ throttle logging
5113 static void perf_log_throttle(struct perf_event *event, int enable)
5115 struct perf_output_handle handle;
5116 struct perf_sample_data sample;
5120 struct perf_event_header header;
5124 } throttle_event = {
5126 .type = PERF_RECORD_THROTTLE,
5128 .size = sizeof(throttle_event),
5130 .time = perf_clock(),
5131 .id = primary_event_id(event),
5132 .stream_id = event->id,
5136 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5138 perf_event_header__init_id(&throttle_event.header, &sample, event);
5140 ret = perf_output_begin(&handle, event,
5141 throttle_event.header.size);
5145 perf_output_put(&handle, throttle_event);
5146 perf_event__output_id_sample(event, &handle, &sample);
5147 perf_output_end(&handle);
5151 * Generic event overflow handling, sampling.
5154 static int __perf_event_overflow(struct perf_event *event,
5155 int throttle, struct perf_sample_data *data,
5156 struct pt_regs *regs)
5158 int events = atomic_read(&event->event_limit);
5159 struct hw_perf_event *hwc = &event->hw;
5164 * Non-sampling counters might still use the PMI to fold short
5165 * hardware counters, ignore those.
5167 if (unlikely(!is_sampling_event(event)))
5170 seq = __this_cpu_read(perf_throttled_seq);
5171 if (seq != hwc->interrupts_seq) {
5172 hwc->interrupts_seq = seq;
5173 hwc->interrupts = 1;
5176 if (unlikely(throttle
5177 && hwc->interrupts >= max_samples_per_tick)) {
5178 __this_cpu_inc(perf_throttled_count);
5179 hwc->interrupts = MAX_INTERRUPTS;
5180 perf_log_throttle(event, 0);
5185 if (event->attr.freq) {
5186 u64 now = perf_clock();
5187 s64 delta = now - hwc->freq_time_stamp;
5189 hwc->freq_time_stamp = now;
5191 if (delta > 0 && delta < 2*TICK_NSEC)
5192 perf_adjust_period(event, delta, hwc->last_period, true);
5196 * XXX event_limit might not quite work as expected on inherited
5200 event->pending_kill = POLL_IN;
5201 if (events && atomic_dec_and_test(&event->event_limit)) {
5203 event->pending_kill = POLL_HUP;
5204 event->pending_disable = 1;
5205 irq_work_queue(&event->pending);
5208 if (event->overflow_handler)
5209 event->overflow_handler(event, data, regs);
5211 perf_event_output(event, data, regs);
5213 if (event->fasync && event->pending_kill) {
5214 event->pending_wakeup = 1;
5215 irq_work_queue(&event->pending);
5221 int perf_event_overflow(struct perf_event *event,
5222 struct perf_sample_data *data,
5223 struct pt_regs *regs)
5225 return __perf_event_overflow(event, 1, data, regs);
5229 * Generic software event infrastructure
5232 struct swevent_htable {
5233 struct swevent_hlist *swevent_hlist;
5234 struct mutex hlist_mutex;
5237 /* Recursion avoidance in each contexts */
5238 int recursion[PERF_NR_CONTEXTS];
5241 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5244 * We directly increment event->count and keep a second value in
5245 * event->hw.period_left to count intervals. This period event
5246 * is kept in the range [-sample_period, 0] so that we can use the
5250 u64 perf_swevent_set_period(struct perf_event *event)
5252 struct hw_perf_event *hwc = &event->hw;
5253 u64 period = hwc->last_period;
5257 hwc->last_period = hwc->sample_period;
5260 old = val = local64_read(&hwc->period_left);
5264 nr = div64_u64(period + val, period);
5265 offset = nr * period;
5267 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5273 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5274 struct perf_sample_data *data,
5275 struct pt_regs *regs)
5277 struct hw_perf_event *hwc = &event->hw;
5281 overflow = perf_swevent_set_period(event);
5283 if (hwc->interrupts == MAX_INTERRUPTS)
5286 for (; overflow; overflow--) {
5287 if (__perf_event_overflow(event, throttle,
5290 * We inhibit the overflow from happening when
5291 * hwc->interrupts == MAX_INTERRUPTS.
5299 static void perf_swevent_event(struct perf_event *event, u64 nr,
5300 struct perf_sample_data *data,
5301 struct pt_regs *regs)
5303 struct hw_perf_event *hwc = &event->hw;
5305 local64_add(nr, &event->count);
5310 if (!is_sampling_event(event))
5313 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5315 return perf_swevent_overflow(event, 1, data, regs);
5317 data->period = event->hw.last_period;
5319 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5320 return perf_swevent_overflow(event, 1, data, regs);
5322 if (local64_add_negative(nr, &hwc->period_left))
5325 perf_swevent_overflow(event, 0, data, regs);
5328 static int perf_exclude_event(struct perf_event *event,
5329 struct pt_regs *regs)
5331 if (event->hw.state & PERF_HES_STOPPED)
5335 if (event->attr.exclude_user && user_mode(regs))
5338 if (event->attr.exclude_kernel && !user_mode(regs))
5345 static int perf_swevent_match(struct perf_event *event,
5346 enum perf_type_id type,
5348 struct perf_sample_data *data,
5349 struct pt_regs *regs)
5351 if (event->attr.type != type)
5354 if (event->attr.config != event_id)
5357 if (perf_exclude_event(event, regs))
5363 static inline u64 swevent_hash(u64 type, u32 event_id)
5365 u64 val = event_id | (type << 32);
5367 return hash_64(val, SWEVENT_HLIST_BITS);
5370 static inline struct hlist_head *
5371 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5373 u64 hash = swevent_hash(type, event_id);
5375 return &hlist->heads[hash];
5378 /* For the read side: events when they trigger */
5379 static inline struct hlist_head *
5380 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5382 struct swevent_hlist *hlist;
5384 hlist = rcu_dereference(swhash->swevent_hlist);
5388 return __find_swevent_head(hlist, type, event_id);
5391 /* For the event head insertion and removal in the hlist */
5392 static inline struct hlist_head *
5393 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5395 struct swevent_hlist *hlist;
5396 u32 event_id = event->attr.config;
5397 u64 type = event->attr.type;
5400 * Event scheduling is always serialized against hlist allocation
5401 * and release. Which makes the protected version suitable here.
5402 * The context lock guarantees that.
5404 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5405 lockdep_is_held(&event->ctx->lock));
5409 return __find_swevent_head(hlist, type, event_id);
5412 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5414 struct perf_sample_data *data,
5415 struct pt_regs *regs)
5417 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5418 struct perf_event *event;
5419 struct hlist_head *head;
5422 head = find_swevent_head_rcu(swhash, type, event_id);
5426 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5427 if (perf_swevent_match(event, type, event_id, data, regs))
5428 perf_swevent_event(event, nr, data, regs);
5434 int perf_swevent_get_recursion_context(void)
5436 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5438 return get_recursion_context(swhash->recursion);
5440 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5442 inline void perf_swevent_put_recursion_context(int rctx)
5444 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5446 put_recursion_context(swhash->recursion, rctx);
5449 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5451 struct perf_sample_data data;
5454 preempt_disable_notrace();
5455 rctx = perf_swevent_get_recursion_context();
5459 perf_sample_data_init(&data, addr, 0);
5461 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5463 perf_swevent_put_recursion_context(rctx);
5464 preempt_enable_notrace();
5467 static void perf_swevent_read(struct perf_event *event)
5471 static int perf_swevent_add(struct perf_event *event, int flags)
5473 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5474 struct hw_perf_event *hwc = &event->hw;
5475 struct hlist_head *head;
5477 if (is_sampling_event(event)) {
5478 hwc->last_period = hwc->sample_period;
5479 perf_swevent_set_period(event);
5482 hwc->state = !(flags & PERF_EF_START);
5484 head = find_swevent_head(swhash, event);
5485 if (WARN_ON_ONCE(!head))
5488 hlist_add_head_rcu(&event->hlist_entry, head);
5493 static void perf_swevent_del(struct perf_event *event, int flags)
5495 hlist_del_rcu(&event->hlist_entry);
5498 static void perf_swevent_start(struct perf_event *event, int flags)
5500 event->hw.state = 0;
5503 static void perf_swevent_stop(struct perf_event *event, int flags)
5505 event->hw.state = PERF_HES_STOPPED;
5508 /* Deref the hlist from the update side */
5509 static inline struct swevent_hlist *
5510 swevent_hlist_deref(struct swevent_htable *swhash)
5512 return rcu_dereference_protected(swhash->swevent_hlist,
5513 lockdep_is_held(&swhash->hlist_mutex));
5516 static void swevent_hlist_release(struct swevent_htable *swhash)
5518 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5523 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5524 kfree_rcu(hlist, rcu_head);
5527 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5529 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5531 mutex_lock(&swhash->hlist_mutex);
5533 if (!--swhash->hlist_refcount)
5534 swevent_hlist_release(swhash);
5536 mutex_unlock(&swhash->hlist_mutex);
5539 static void swevent_hlist_put(struct perf_event *event)
5543 if (event->cpu != -1) {
5544 swevent_hlist_put_cpu(event, event->cpu);
5548 for_each_possible_cpu(cpu)
5549 swevent_hlist_put_cpu(event, cpu);
5552 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5554 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5557 mutex_lock(&swhash->hlist_mutex);
5559 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5560 struct swevent_hlist *hlist;
5562 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5567 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5569 swhash->hlist_refcount++;
5571 mutex_unlock(&swhash->hlist_mutex);
5576 static int swevent_hlist_get(struct perf_event *event)
5579 int cpu, failed_cpu;
5581 if (event->cpu != -1)
5582 return swevent_hlist_get_cpu(event, event->cpu);
5585 for_each_possible_cpu(cpu) {
5586 err = swevent_hlist_get_cpu(event, cpu);
5596 for_each_possible_cpu(cpu) {
5597 if (cpu == failed_cpu)
5599 swevent_hlist_put_cpu(event, cpu);
5606 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5608 static void sw_perf_event_destroy(struct perf_event *event)
5610 u64 event_id = event->attr.config;
5612 WARN_ON(event->parent);
5614 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5615 swevent_hlist_put(event);
5618 static int perf_swevent_init(struct perf_event *event)
5620 u64 event_id = event->attr.config;
5622 if (event->attr.type != PERF_TYPE_SOFTWARE)
5626 * no branch sampling for software events
5628 if (has_branch_stack(event))
5632 case PERF_COUNT_SW_CPU_CLOCK:
5633 case PERF_COUNT_SW_TASK_CLOCK:
5640 if (event_id >= PERF_COUNT_SW_MAX)
5643 if (!event->parent) {
5646 err = swevent_hlist_get(event);
5650 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5651 event->destroy = sw_perf_event_destroy;
5657 static int perf_swevent_event_idx(struct perf_event *event)
5662 static struct pmu perf_swevent = {
5663 .task_ctx_nr = perf_sw_context,
5665 .event_init = perf_swevent_init,
5666 .add = perf_swevent_add,
5667 .del = perf_swevent_del,
5668 .start = perf_swevent_start,
5669 .stop = perf_swevent_stop,
5670 .read = perf_swevent_read,
5672 .event_idx = perf_swevent_event_idx,
5675 #ifdef CONFIG_EVENT_TRACING
5677 static int perf_tp_filter_match(struct perf_event *event,
5678 struct perf_sample_data *data)
5680 void *record = data->raw->data;
5682 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5687 static int perf_tp_event_match(struct perf_event *event,
5688 struct perf_sample_data *data,
5689 struct pt_regs *regs)
5691 if (event->hw.state & PERF_HES_STOPPED)
5694 * All tracepoints are from kernel-space.
5696 if (event->attr.exclude_kernel)
5699 if (!perf_tp_filter_match(event, data))
5705 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5706 struct pt_regs *regs, struct hlist_head *head, int rctx,
5707 struct task_struct *task)
5709 struct perf_sample_data data;
5710 struct perf_event *event;
5712 struct perf_raw_record raw = {
5717 perf_sample_data_init(&data, addr, 0);
5720 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5721 if (perf_tp_event_match(event, &data, regs))
5722 perf_swevent_event(event, count, &data, regs);
5726 * If we got specified a target task, also iterate its context and
5727 * deliver this event there too.
5729 if (task && task != current) {
5730 struct perf_event_context *ctx;
5731 struct trace_entry *entry = record;
5734 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5738 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5739 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5741 if (event->attr.config != entry->type)
5743 if (perf_tp_event_match(event, &data, regs))
5744 perf_swevent_event(event, count, &data, regs);
5750 perf_swevent_put_recursion_context(rctx);
5752 EXPORT_SYMBOL_GPL(perf_tp_event);
5754 static void tp_perf_event_destroy(struct perf_event *event)
5756 perf_trace_destroy(event);
5759 static int perf_tp_event_init(struct perf_event *event)
5763 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5767 * no branch sampling for tracepoint events
5769 if (has_branch_stack(event))
5772 err = perf_trace_init(event);
5776 event->destroy = tp_perf_event_destroy;
5781 static struct pmu perf_tracepoint = {
5782 .task_ctx_nr = perf_sw_context,
5784 .event_init = perf_tp_event_init,
5785 .add = perf_trace_add,
5786 .del = perf_trace_del,
5787 .start = perf_swevent_start,
5788 .stop = perf_swevent_stop,
5789 .read = perf_swevent_read,
5791 .event_idx = perf_swevent_event_idx,
5794 static inline void perf_tp_register(void)
5796 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5799 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5804 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5807 filter_str = strndup_user(arg, PAGE_SIZE);
5808 if (IS_ERR(filter_str))
5809 return PTR_ERR(filter_str);
5811 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5817 static void perf_event_free_filter(struct perf_event *event)
5819 ftrace_profile_free_filter(event);
5824 static inline void perf_tp_register(void)
5828 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5833 static void perf_event_free_filter(struct perf_event *event)
5837 #endif /* CONFIG_EVENT_TRACING */
5839 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5840 void perf_bp_event(struct perf_event *bp, void *data)
5842 struct perf_sample_data sample;
5843 struct pt_regs *regs = data;
5845 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5847 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5848 perf_swevent_event(bp, 1, &sample, regs);
5853 * hrtimer based swevent callback
5856 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5858 enum hrtimer_restart ret = HRTIMER_RESTART;
5859 struct perf_sample_data data;
5860 struct pt_regs *regs;
5861 struct perf_event *event;
5864 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5866 if (event->state != PERF_EVENT_STATE_ACTIVE)
5867 return HRTIMER_NORESTART;
5869 event->pmu->read(event);
5871 perf_sample_data_init(&data, 0, event->hw.last_period);
5872 regs = get_irq_regs();
5874 if (regs && !perf_exclude_event(event, regs)) {
5875 if (!(event->attr.exclude_idle && is_idle_task(current)))
5876 if (__perf_event_overflow(event, 1, &data, regs))
5877 ret = HRTIMER_NORESTART;
5880 period = max_t(u64, 10000, event->hw.sample_period);
5881 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5886 static void perf_swevent_start_hrtimer(struct perf_event *event)
5888 struct hw_perf_event *hwc = &event->hw;
5891 if (!is_sampling_event(event))
5894 period = local64_read(&hwc->period_left);
5899 local64_set(&hwc->period_left, 0);
5901 period = max_t(u64, 10000, hwc->sample_period);
5903 __hrtimer_start_range_ns(&hwc->hrtimer,
5904 ns_to_ktime(period), 0,
5905 HRTIMER_MODE_REL_PINNED, 0);
5908 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5910 struct hw_perf_event *hwc = &event->hw;
5912 if (is_sampling_event(event)) {
5913 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5914 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5916 hrtimer_cancel(&hwc->hrtimer);
5920 static void perf_swevent_init_hrtimer(struct perf_event *event)
5922 struct hw_perf_event *hwc = &event->hw;
5924 if (!is_sampling_event(event))
5927 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5928 hwc->hrtimer.function = perf_swevent_hrtimer;
5931 * Since hrtimers have a fixed rate, we can do a static freq->period
5932 * mapping and avoid the whole period adjust feedback stuff.
5934 if (event->attr.freq) {
5935 long freq = event->attr.sample_freq;
5937 event->attr.sample_period = NSEC_PER_SEC / freq;
5938 hwc->sample_period = event->attr.sample_period;
5939 local64_set(&hwc->period_left, hwc->sample_period);
5940 hwc->last_period = hwc->sample_period;
5941 event->attr.freq = 0;
5946 * Software event: cpu wall time clock
5949 static void cpu_clock_event_update(struct perf_event *event)
5954 now = local_clock();
5955 prev = local64_xchg(&event->hw.prev_count, now);
5956 local64_add(now - prev, &event->count);
5959 static void cpu_clock_event_start(struct perf_event *event, int flags)
5961 local64_set(&event->hw.prev_count, local_clock());
5962 perf_swevent_start_hrtimer(event);
5965 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5967 perf_swevent_cancel_hrtimer(event);
5968 cpu_clock_event_update(event);
5971 static int cpu_clock_event_add(struct perf_event *event, int flags)
5973 if (flags & PERF_EF_START)
5974 cpu_clock_event_start(event, flags);
5979 static void cpu_clock_event_del(struct perf_event *event, int flags)
5981 cpu_clock_event_stop(event, flags);
5984 static void cpu_clock_event_read(struct perf_event *event)
5986 cpu_clock_event_update(event);
5989 static int cpu_clock_event_init(struct perf_event *event)
5991 if (event->attr.type != PERF_TYPE_SOFTWARE)
5994 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5998 * no branch sampling for software events
6000 if (has_branch_stack(event))
6003 perf_swevent_init_hrtimer(event);
6008 static struct pmu perf_cpu_clock = {
6009 .task_ctx_nr = perf_sw_context,
6011 .event_init = cpu_clock_event_init,
6012 .add = cpu_clock_event_add,
6013 .del = cpu_clock_event_del,
6014 .start = cpu_clock_event_start,
6015 .stop = cpu_clock_event_stop,
6016 .read = cpu_clock_event_read,
6018 .event_idx = perf_swevent_event_idx,
6022 * Software event: task time clock
6025 static void task_clock_event_update(struct perf_event *event, u64 now)
6030 prev = local64_xchg(&event->hw.prev_count, now);
6032 local64_add(delta, &event->count);
6035 static void task_clock_event_start(struct perf_event *event, int flags)
6037 local64_set(&event->hw.prev_count, event->ctx->time);
6038 perf_swevent_start_hrtimer(event);
6041 static void task_clock_event_stop(struct perf_event *event, int flags)
6043 perf_swevent_cancel_hrtimer(event);
6044 task_clock_event_update(event, event->ctx->time);
6047 static int task_clock_event_add(struct perf_event *event, int flags)
6049 if (flags & PERF_EF_START)
6050 task_clock_event_start(event, flags);
6055 static void task_clock_event_del(struct perf_event *event, int flags)
6057 task_clock_event_stop(event, PERF_EF_UPDATE);
6060 static void task_clock_event_read(struct perf_event *event)
6062 u64 now = perf_clock();
6063 u64 delta = now - event->ctx->timestamp;
6064 u64 time = event->ctx->time + delta;
6066 task_clock_event_update(event, time);
6069 static int task_clock_event_init(struct perf_event *event)
6071 if (event->attr.type != PERF_TYPE_SOFTWARE)
6074 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6078 * no branch sampling for software events
6080 if (has_branch_stack(event))
6083 perf_swevent_init_hrtimer(event);
6088 static struct pmu perf_task_clock = {
6089 .task_ctx_nr = perf_sw_context,
6091 .event_init = task_clock_event_init,
6092 .add = task_clock_event_add,
6093 .del = task_clock_event_del,
6094 .start = task_clock_event_start,
6095 .stop = task_clock_event_stop,
6096 .read = task_clock_event_read,
6098 .event_idx = perf_swevent_event_idx,
6101 static void perf_pmu_nop_void(struct pmu *pmu)
6105 static int perf_pmu_nop_int(struct pmu *pmu)
6110 static void perf_pmu_start_txn(struct pmu *pmu)
6112 perf_pmu_disable(pmu);
6115 static int perf_pmu_commit_txn(struct pmu *pmu)
6117 perf_pmu_enable(pmu);
6121 static void perf_pmu_cancel_txn(struct pmu *pmu)
6123 perf_pmu_enable(pmu);
6126 static int perf_event_idx_default(struct perf_event *event)
6128 return event->hw.idx + 1;
6132 * Ensures all contexts with the same task_ctx_nr have the same
6133 * pmu_cpu_context too.
6135 static void *find_pmu_context(int ctxn)
6142 list_for_each_entry(pmu, &pmus, entry) {
6143 if (pmu->task_ctx_nr == ctxn)
6144 return pmu->pmu_cpu_context;
6150 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6154 for_each_possible_cpu(cpu) {
6155 struct perf_cpu_context *cpuctx;
6157 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6159 if (cpuctx->unique_pmu == old_pmu)
6160 cpuctx->unique_pmu = pmu;
6164 static void free_pmu_context(struct pmu *pmu)
6168 mutex_lock(&pmus_lock);
6170 * Like a real lame refcount.
6172 list_for_each_entry(i, &pmus, entry) {
6173 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6174 update_pmu_context(i, pmu);
6179 free_percpu(pmu->pmu_cpu_context);
6181 mutex_unlock(&pmus_lock);
6183 static struct idr pmu_idr;
6186 type_show(struct device *dev, struct device_attribute *attr, char *page)
6188 struct pmu *pmu = dev_get_drvdata(dev);
6190 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6194 perf_event_mux_interval_ms_show(struct device *dev,
6195 struct device_attribute *attr,
6198 struct pmu *pmu = dev_get_drvdata(dev);
6200 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6204 perf_event_mux_interval_ms_store(struct device *dev,
6205 struct device_attribute *attr,
6206 const char *buf, size_t count)
6208 struct pmu *pmu = dev_get_drvdata(dev);
6209 int timer, cpu, ret;
6211 ret = kstrtoint(buf, 0, &timer);
6218 /* same value, noting to do */
6219 if (timer == pmu->hrtimer_interval_ms)
6222 pmu->hrtimer_interval_ms = timer;
6224 /* update all cpuctx for this PMU */
6225 for_each_possible_cpu(cpu) {
6226 struct perf_cpu_context *cpuctx;
6227 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6228 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6230 if (hrtimer_active(&cpuctx->hrtimer))
6231 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6237 static struct device_attribute pmu_dev_attrs[] = {
6239 __ATTR_RW(perf_event_mux_interval_ms),
6243 static int pmu_bus_running;
6244 static struct bus_type pmu_bus = {
6245 .name = "event_source",
6246 .dev_attrs = pmu_dev_attrs,
6249 static void pmu_dev_release(struct device *dev)
6254 static int pmu_dev_alloc(struct pmu *pmu)
6258 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6262 pmu->dev->groups = pmu->attr_groups;
6263 device_initialize(pmu->dev);
6264 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6268 dev_set_drvdata(pmu->dev, pmu);
6269 pmu->dev->bus = &pmu_bus;
6270 pmu->dev->release = pmu_dev_release;
6271 ret = device_add(pmu->dev);
6279 put_device(pmu->dev);
6283 static struct lock_class_key cpuctx_mutex;
6284 static struct lock_class_key cpuctx_lock;
6286 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6290 mutex_lock(&pmus_lock);
6292 pmu->pmu_disable_count = alloc_percpu(int);
6293 if (!pmu->pmu_disable_count)
6302 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6310 if (pmu_bus_running) {
6311 ret = pmu_dev_alloc(pmu);
6317 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6318 if (pmu->pmu_cpu_context)
6319 goto got_cpu_context;
6322 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6323 if (!pmu->pmu_cpu_context)
6326 for_each_possible_cpu(cpu) {
6327 struct perf_cpu_context *cpuctx;
6329 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6330 __perf_event_init_context(&cpuctx->ctx);
6331 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6332 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6333 cpuctx->ctx.type = cpu_context;
6334 cpuctx->ctx.pmu = pmu;
6336 __perf_cpu_hrtimer_init(cpuctx, cpu);
6338 INIT_LIST_HEAD(&cpuctx->rotation_list);
6339 cpuctx->unique_pmu = pmu;
6343 if (!pmu->start_txn) {
6344 if (pmu->pmu_enable) {
6346 * If we have pmu_enable/pmu_disable calls, install
6347 * transaction stubs that use that to try and batch
6348 * hardware accesses.
6350 pmu->start_txn = perf_pmu_start_txn;
6351 pmu->commit_txn = perf_pmu_commit_txn;
6352 pmu->cancel_txn = perf_pmu_cancel_txn;
6354 pmu->start_txn = perf_pmu_nop_void;
6355 pmu->commit_txn = perf_pmu_nop_int;
6356 pmu->cancel_txn = perf_pmu_nop_void;
6360 if (!pmu->pmu_enable) {
6361 pmu->pmu_enable = perf_pmu_nop_void;
6362 pmu->pmu_disable = perf_pmu_nop_void;
6365 if (!pmu->event_idx)
6366 pmu->event_idx = perf_event_idx_default;
6368 list_add_rcu(&pmu->entry, &pmus);
6371 mutex_unlock(&pmus_lock);
6376 device_del(pmu->dev);
6377 put_device(pmu->dev);
6380 if (pmu->type >= PERF_TYPE_MAX)
6381 idr_remove(&pmu_idr, pmu->type);
6384 free_percpu(pmu->pmu_disable_count);
6388 void perf_pmu_unregister(struct pmu *pmu)
6390 mutex_lock(&pmus_lock);
6391 list_del_rcu(&pmu->entry);
6392 mutex_unlock(&pmus_lock);
6395 * We dereference the pmu list under both SRCU and regular RCU, so
6396 * synchronize against both of those.
6398 synchronize_srcu(&pmus_srcu);
6401 free_percpu(pmu->pmu_disable_count);
6402 if (pmu->type >= PERF_TYPE_MAX)
6403 idr_remove(&pmu_idr, pmu->type);
6404 device_del(pmu->dev);
6405 put_device(pmu->dev);
6406 free_pmu_context(pmu);
6409 struct pmu *perf_init_event(struct perf_event *event)
6411 struct pmu *pmu = NULL;
6415 idx = srcu_read_lock(&pmus_srcu);
6418 pmu = idr_find(&pmu_idr, event->attr.type);
6422 ret = pmu->event_init(event);
6428 list_for_each_entry_rcu(pmu, &pmus, entry) {
6430 ret = pmu->event_init(event);
6434 if (ret != -ENOENT) {
6439 pmu = ERR_PTR(-ENOENT);
6441 srcu_read_unlock(&pmus_srcu, idx);
6447 * Allocate and initialize a event structure
6449 static struct perf_event *
6450 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6451 struct task_struct *task,
6452 struct perf_event *group_leader,
6453 struct perf_event *parent_event,
6454 perf_overflow_handler_t overflow_handler,
6458 struct perf_event *event;
6459 struct hw_perf_event *hwc;
6462 if ((unsigned)cpu >= nr_cpu_ids) {
6463 if (!task || cpu != -1)
6464 return ERR_PTR(-EINVAL);
6467 event = kzalloc(sizeof(*event), GFP_KERNEL);
6469 return ERR_PTR(-ENOMEM);
6472 * Single events are their own group leaders, with an
6473 * empty sibling list:
6476 group_leader = event;
6478 mutex_init(&event->child_mutex);
6479 INIT_LIST_HEAD(&event->child_list);
6481 INIT_LIST_HEAD(&event->group_entry);
6482 INIT_LIST_HEAD(&event->event_entry);
6483 INIT_LIST_HEAD(&event->sibling_list);
6484 INIT_LIST_HEAD(&event->rb_entry);
6486 init_waitqueue_head(&event->waitq);
6487 init_irq_work(&event->pending, perf_pending_event);
6489 mutex_init(&event->mmap_mutex);
6491 atomic_long_set(&event->refcount, 1);
6493 event->attr = *attr;
6494 event->group_leader = group_leader;
6498 event->parent = parent_event;
6500 event->ns = get_pid_ns(task_active_pid_ns(current));
6501 event->id = atomic64_inc_return(&perf_event_id);
6503 event->state = PERF_EVENT_STATE_INACTIVE;
6506 event->attach_state = PERF_ATTACH_TASK;
6508 if (attr->type == PERF_TYPE_TRACEPOINT)
6509 event->hw.tp_target = task;
6510 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6512 * hw_breakpoint is a bit difficult here..
6514 else if (attr->type == PERF_TYPE_BREAKPOINT)
6515 event->hw.bp_target = task;
6519 if (!overflow_handler && parent_event) {
6520 overflow_handler = parent_event->overflow_handler;
6521 context = parent_event->overflow_handler_context;
6524 event->overflow_handler = overflow_handler;
6525 event->overflow_handler_context = context;
6527 perf_event__state_init(event);
6532 hwc->sample_period = attr->sample_period;
6533 if (attr->freq && attr->sample_freq)
6534 hwc->sample_period = 1;
6535 hwc->last_period = hwc->sample_period;
6537 local64_set(&hwc->period_left, hwc->sample_period);
6540 * we currently do not support PERF_FORMAT_GROUP on inherited events
6542 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6545 pmu = perf_init_event(event);
6551 else if (IS_ERR(pmu))
6556 put_pid_ns(event->ns);
6558 return ERR_PTR(err);
6561 if (!event->parent) {
6562 if (event->attach_state & PERF_ATTACH_TASK)
6563 static_key_slow_inc(&perf_sched_events.key);
6564 if (event->attr.mmap || event->attr.mmap_data)
6565 atomic_inc(&nr_mmap_events);
6566 if (event->attr.comm)
6567 atomic_inc(&nr_comm_events);
6568 if (event->attr.task)
6569 atomic_inc(&nr_task_events);
6570 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6571 err = get_callchain_buffers();
6574 return ERR_PTR(err);
6577 if (has_branch_stack(event)) {
6578 static_key_slow_inc(&perf_sched_events.key);
6579 if (!(event->attach_state & PERF_ATTACH_TASK))
6580 atomic_inc(&per_cpu(perf_branch_stack_events,
6588 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6589 struct perf_event_attr *attr)
6594 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6598 * zero the full structure, so that a short copy will be nice.
6600 memset(attr, 0, sizeof(*attr));
6602 ret = get_user(size, &uattr->size);
6606 if (size > PAGE_SIZE) /* silly large */
6609 if (!size) /* abi compat */
6610 size = PERF_ATTR_SIZE_VER0;
6612 if (size < PERF_ATTR_SIZE_VER0)
6616 * If we're handed a bigger struct than we know of,
6617 * ensure all the unknown bits are 0 - i.e. new
6618 * user-space does not rely on any kernel feature
6619 * extensions we dont know about yet.
6621 if (size > sizeof(*attr)) {
6622 unsigned char __user *addr;
6623 unsigned char __user *end;
6626 addr = (void __user *)uattr + sizeof(*attr);
6627 end = (void __user *)uattr + size;
6629 for (; addr < end; addr++) {
6630 ret = get_user(val, addr);
6636 size = sizeof(*attr);
6639 ret = copy_from_user(attr, uattr, size);
6643 if (attr->__reserved_1)
6646 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6649 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6652 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6653 u64 mask = attr->branch_sample_type;
6655 /* only using defined bits */
6656 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6659 /* at least one branch bit must be set */
6660 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6663 /* propagate priv level, when not set for branch */
6664 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6666 /* exclude_kernel checked on syscall entry */
6667 if (!attr->exclude_kernel)
6668 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6670 if (!attr->exclude_user)
6671 mask |= PERF_SAMPLE_BRANCH_USER;
6673 if (!attr->exclude_hv)
6674 mask |= PERF_SAMPLE_BRANCH_HV;
6676 * adjust user setting (for HW filter setup)
6678 attr->branch_sample_type = mask;
6680 /* privileged levels capture (kernel, hv): check permissions */
6681 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6682 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6686 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6687 ret = perf_reg_validate(attr->sample_regs_user);
6692 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6693 if (!arch_perf_have_user_stack_dump())
6697 * We have __u32 type for the size, but so far
6698 * we can only use __u16 as maximum due to the
6699 * __u16 sample size limit.
6701 if (attr->sample_stack_user >= USHRT_MAX)
6703 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6711 put_user(sizeof(*attr), &uattr->size);
6717 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6719 struct ring_buffer *rb = NULL, *old_rb = NULL;
6725 /* don't allow circular references */
6726 if (event == output_event)
6730 * Don't allow cross-cpu buffers
6732 if (output_event->cpu != event->cpu)
6736 * If its not a per-cpu rb, it must be the same task.
6738 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6742 mutex_lock(&event->mmap_mutex);
6743 /* Can't redirect output if we've got an active mmap() */
6744 if (atomic_read(&event->mmap_count))
6750 /* get the rb we want to redirect to */
6751 rb = ring_buffer_get(output_event);
6757 ring_buffer_detach(event, old_rb);
6760 ring_buffer_attach(event, rb);
6762 rcu_assign_pointer(event->rb, rb);
6765 ring_buffer_put(old_rb);
6767 * Since we detached before setting the new rb, so that we
6768 * could attach the new rb, we could have missed a wakeup.
6771 wake_up_all(&event->waitq);
6776 mutex_unlock(&event->mmap_mutex);
6783 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6785 * @attr_uptr: event_id type attributes for monitoring/sampling
6788 * @group_fd: group leader event fd
6790 SYSCALL_DEFINE5(perf_event_open,
6791 struct perf_event_attr __user *, attr_uptr,
6792 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6794 struct perf_event *group_leader = NULL, *output_event = NULL;
6795 struct perf_event *event, *sibling;
6796 struct perf_event_attr attr;
6797 struct perf_event_context *ctx;
6798 struct file *event_file = NULL;
6799 struct fd group = {NULL, 0};
6800 struct task_struct *task = NULL;
6806 /* for future expandability... */
6807 if (flags & ~PERF_FLAG_ALL)
6810 err = perf_copy_attr(attr_uptr, &attr);
6814 if (!attr.exclude_kernel) {
6815 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6820 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6825 * In cgroup mode, the pid argument is used to pass the fd
6826 * opened to the cgroup directory in cgroupfs. The cpu argument
6827 * designates the cpu on which to monitor threads from that
6830 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6833 event_fd = get_unused_fd();
6837 if (group_fd != -1) {
6838 err = perf_fget_light(group_fd, &group);
6841 group_leader = group.file->private_data;
6842 if (flags & PERF_FLAG_FD_OUTPUT)
6843 output_event = group_leader;
6844 if (flags & PERF_FLAG_FD_NO_GROUP)
6845 group_leader = NULL;
6848 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6849 task = find_lively_task_by_vpid(pid);
6851 err = PTR_ERR(task);
6858 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6860 if (IS_ERR(event)) {
6861 err = PTR_ERR(event);
6865 if (flags & PERF_FLAG_PID_CGROUP) {
6866 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6871 * - that has cgroup constraint on event->cpu
6872 * - that may need work on context switch
6874 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6875 static_key_slow_inc(&perf_sched_events.key);
6879 * Special case software events and allow them to be part of
6880 * any hardware group.
6885 (is_software_event(event) != is_software_event(group_leader))) {
6886 if (is_software_event(event)) {
6888 * If event and group_leader are not both a software
6889 * event, and event is, then group leader is not.
6891 * Allow the addition of software events to !software
6892 * groups, this is safe because software events never
6895 pmu = group_leader->pmu;
6896 } else if (is_software_event(group_leader) &&
6897 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6899 * In case the group is a pure software group, and we
6900 * try to add a hardware event, move the whole group to
6901 * the hardware context.
6908 * Get the target context (task or percpu):
6910 ctx = find_get_context(pmu, task, event->cpu);
6917 put_task_struct(task);
6922 * Look up the group leader (we will attach this event to it):
6928 * Do not allow a recursive hierarchy (this new sibling
6929 * becoming part of another group-sibling):
6931 if (group_leader->group_leader != group_leader)
6934 * Do not allow to attach to a group in a different
6935 * task or CPU context:
6938 if (group_leader->ctx->type != ctx->type)
6941 if (group_leader->ctx != ctx)
6946 * Only a group leader can be exclusive or pinned
6948 if (attr.exclusive || attr.pinned)
6953 err = perf_event_set_output(event, output_event);
6958 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6959 if (IS_ERR(event_file)) {
6960 err = PTR_ERR(event_file);
6965 struct perf_event_context *gctx = group_leader->ctx;
6967 mutex_lock(&gctx->mutex);
6968 perf_remove_from_context(group_leader);
6971 * Removing from the context ends up with disabled
6972 * event. What we want here is event in the initial
6973 * startup state, ready to be add into new context.
6975 perf_event__state_init(group_leader);
6976 list_for_each_entry(sibling, &group_leader->sibling_list,
6978 perf_remove_from_context(sibling);
6979 perf_event__state_init(sibling);
6982 mutex_unlock(&gctx->mutex);
6986 WARN_ON_ONCE(ctx->parent_ctx);
6987 mutex_lock(&ctx->mutex);
6991 perf_install_in_context(ctx, group_leader, event->cpu);
6993 list_for_each_entry(sibling, &group_leader->sibling_list,
6995 perf_install_in_context(ctx, sibling, event->cpu);
7000 perf_install_in_context(ctx, event, event->cpu);
7002 perf_unpin_context(ctx);
7003 mutex_unlock(&ctx->mutex);
7007 event->owner = current;
7009 mutex_lock(¤t->perf_event_mutex);
7010 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7011 mutex_unlock(¤t->perf_event_mutex);
7014 * Precalculate sample_data sizes
7016 perf_event__header_size(event);
7017 perf_event__id_header_size(event);
7020 * Drop the reference on the group_event after placing the
7021 * new event on the sibling_list. This ensures destruction
7022 * of the group leader will find the pointer to itself in
7023 * perf_group_detach().
7026 fd_install(event_fd, event_file);
7030 perf_unpin_context(ctx);
7037 put_task_struct(task);
7041 put_unused_fd(event_fd);
7046 * perf_event_create_kernel_counter
7048 * @attr: attributes of the counter to create
7049 * @cpu: cpu in which the counter is bound
7050 * @task: task to profile (NULL for percpu)
7053 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7054 struct task_struct *task,
7055 perf_overflow_handler_t overflow_handler,
7058 struct perf_event_context *ctx;
7059 struct perf_event *event;
7063 * Get the target context (task or percpu):
7066 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7067 overflow_handler, context);
7068 if (IS_ERR(event)) {
7069 err = PTR_ERR(event);
7073 ctx = find_get_context(event->pmu, task, cpu);
7079 WARN_ON_ONCE(ctx->parent_ctx);
7080 mutex_lock(&ctx->mutex);
7081 perf_install_in_context(ctx, event, cpu);
7083 perf_unpin_context(ctx);
7084 mutex_unlock(&ctx->mutex);
7091 return ERR_PTR(err);
7093 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7095 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7097 struct perf_event_context *src_ctx;
7098 struct perf_event_context *dst_ctx;
7099 struct perf_event *event, *tmp;
7102 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7103 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7105 mutex_lock(&src_ctx->mutex);
7106 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7108 perf_remove_from_context(event);
7110 list_add(&event->event_entry, &events);
7112 mutex_unlock(&src_ctx->mutex);
7116 mutex_lock(&dst_ctx->mutex);
7117 list_for_each_entry_safe(event, tmp, &events, event_entry) {
7118 list_del(&event->event_entry);
7119 if (event->state >= PERF_EVENT_STATE_OFF)
7120 event->state = PERF_EVENT_STATE_INACTIVE;
7121 perf_install_in_context(dst_ctx, event, dst_cpu);
7124 mutex_unlock(&dst_ctx->mutex);
7126 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7128 static void sync_child_event(struct perf_event *child_event,
7129 struct task_struct *child)
7131 struct perf_event *parent_event = child_event->parent;
7134 if (child_event->attr.inherit_stat)
7135 perf_event_read_event(child_event, child);
7137 child_val = perf_event_count(child_event);
7140 * Add back the child's count to the parent's count:
7142 atomic64_add(child_val, &parent_event->child_count);
7143 atomic64_add(child_event->total_time_enabled,
7144 &parent_event->child_total_time_enabled);
7145 atomic64_add(child_event->total_time_running,
7146 &parent_event->child_total_time_running);
7149 * Remove this event from the parent's list
7151 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7152 mutex_lock(&parent_event->child_mutex);
7153 list_del_init(&child_event->child_list);
7154 mutex_unlock(&parent_event->child_mutex);
7157 * Release the parent event, if this was the last
7160 put_event(parent_event);
7164 __perf_event_exit_task(struct perf_event *child_event,
7165 struct perf_event_context *child_ctx,
7166 struct task_struct *child)
7168 if (child_event->parent) {
7169 raw_spin_lock_irq(&child_ctx->lock);
7170 perf_group_detach(child_event);
7171 raw_spin_unlock_irq(&child_ctx->lock);
7174 perf_remove_from_context(child_event);
7177 * It can happen that the parent exits first, and has events
7178 * that are still around due to the child reference. These
7179 * events need to be zapped.
7181 if (child_event->parent) {
7182 sync_child_event(child_event, child);
7183 free_event(child_event);
7187 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7189 struct perf_event *child_event, *tmp;
7190 struct perf_event_context *child_ctx;
7191 unsigned long flags;
7193 if (likely(!child->perf_event_ctxp[ctxn])) {
7194 perf_event_task(child, NULL, 0);
7198 local_irq_save(flags);
7200 * We can't reschedule here because interrupts are disabled,
7201 * and either child is current or it is a task that can't be
7202 * scheduled, so we are now safe from rescheduling changing
7205 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7208 * Take the context lock here so that if find_get_context is
7209 * reading child->perf_event_ctxp, we wait until it has
7210 * incremented the context's refcount before we do put_ctx below.
7212 raw_spin_lock(&child_ctx->lock);
7213 task_ctx_sched_out(child_ctx);
7214 child->perf_event_ctxp[ctxn] = NULL;
7216 * If this context is a clone; unclone it so it can't get
7217 * swapped to another process while we're removing all
7218 * the events from it.
7220 unclone_ctx(child_ctx);
7221 update_context_time(child_ctx);
7222 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7225 * Report the task dead after unscheduling the events so that we
7226 * won't get any samples after PERF_RECORD_EXIT. We can however still
7227 * get a few PERF_RECORD_READ events.
7229 perf_event_task(child, child_ctx, 0);
7232 * We can recurse on the same lock type through:
7234 * __perf_event_exit_task()
7235 * sync_child_event()
7237 * mutex_lock(&ctx->mutex)
7239 * But since its the parent context it won't be the same instance.
7241 mutex_lock(&child_ctx->mutex);
7244 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7246 __perf_event_exit_task(child_event, child_ctx, child);
7248 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7250 __perf_event_exit_task(child_event, child_ctx, child);
7253 * If the last event was a group event, it will have appended all
7254 * its siblings to the list, but we obtained 'tmp' before that which
7255 * will still point to the list head terminating the iteration.
7257 if (!list_empty(&child_ctx->pinned_groups) ||
7258 !list_empty(&child_ctx->flexible_groups))
7261 mutex_unlock(&child_ctx->mutex);
7267 * When a child task exits, feed back event values to parent events.
7269 void perf_event_exit_task(struct task_struct *child)
7271 struct perf_event *event, *tmp;
7274 mutex_lock(&child->perf_event_mutex);
7275 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7277 list_del_init(&event->owner_entry);
7280 * Ensure the list deletion is visible before we clear
7281 * the owner, closes a race against perf_release() where
7282 * we need to serialize on the owner->perf_event_mutex.
7285 event->owner = NULL;
7287 mutex_unlock(&child->perf_event_mutex);
7289 for_each_task_context_nr(ctxn)
7290 perf_event_exit_task_context(child, ctxn);
7293 static void perf_free_event(struct perf_event *event,
7294 struct perf_event_context *ctx)
7296 struct perf_event *parent = event->parent;
7298 if (WARN_ON_ONCE(!parent))
7301 mutex_lock(&parent->child_mutex);
7302 list_del_init(&event->child_list);
7303 mutex_unlock(&parent->child_mutex);
7307 perf_group_detach(event);
7308 list_del_event(event, ctx);
7313 * free an unexposed, unused context as created by inheritance by
7314 * perf_event_init_task below, used by fork() in case of fail.
7316 void perf_event_free_task(struct task_struct *task)
7318 struct perf_event_context *ctx;
7319 struct perf_event *event, *tmp;
7322 for_each_task_context_nr(ctxn) {
7323 ctx = task->perf_event_ctxp[ctxn];
7327 mutex_lock(&ctx->mutex);
7329 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7331 perf_free_event(event, ctx);
7333 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7335 perf_free_event(event, ctx);
7337 if (!list_empty(&ctx->pinned_groups) ||
7338 !list_empty(&ctx->flexible_groups))
7341 mutex_unlock(&ctx->mutex);
7347 void perf_event_delayed_put(struct task_struct *task)
7351 for_each_task_context_nr(ctxn)
7352 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7356 * inherit a event from parent task to child task:
7358 static struct perf_event *
7359 inherit_event(struct perf_event *parent_event,
7360 struct task_struct *parent,
7361 struct perf_event_context *parent_ctx,
7362 struct task_struct *child,
7363 struct perf_event *group_leader,
7364 struct perf_event_context *child_ctx)
7366 struct perf_event *child_event;
7367 unsigned long flags;
7370 * Instead of creating recursive hierarchies of events,
7371 * we link inherited events back to the original parent,
7372 * which has a filp for sure, which we use as the reference
7375 if (parent_event->parent)
7376 parent_event = parent_event->parent;
7378 child_event = perf_event_alloc(&parent_event->attr,
7381 group_leader, parent_event,
7383 if (IS_ERR(child_event))
7386 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7387 free_event(child_event);
7394 * Make the child state follow the state of the parent event,
7395 * not its attr.disabled bit. We hold the parent's mutex,
7396 * so we won't race with perf_event_{en, dis}able_family.
7398 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7399 child_event->state = PERF_EVENT_STATE_INACTIVE;
7401 child_event->state = PERF_EVENT_STATE_OFF;
7403 if (parent_event->attr.freq) {
7404 u64 sample_period = parent_event->hw.sample_period;
7405 struct hw_perf_event *hwc = &child_event->hw;
7407 hwc->sample_period = sample_period;
7408 hwc->last_period = sample_period;
7410 local64_set(&hwc->period_left, sample_period);
7413 child_event->ctx = child_ctx;
7414 child_event->overflow_handler = parent_event->overflow_handler;
7415 child_event->overflow_handler_context
7416 = parent_event->overflow_handler_context;
7419 * Precalculate sample_data sizes
7421 perf_event__header_size(child_event);
7422 perf_event__id_header_size(child_event);
7425 * Link it up in the child's context:
7427 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7428 add_event_to_ctx(child_event, child_ctx);
7429 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7432 * Link this into the parent event's child list
7434 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7435 mutex_lock(&parent_event->child_mutex);
7436 list_add_tail(&child_event->child_list, &parent_event->child_list);
7437 mutex_unlock(&parent_event->child_mutex);
7442 static int inherit_group(struct perf_event *parent_event,
7443 struct task_struct *parent,
7444 struct perf_event_context *parent_ctx,
7445 struct task_struct *child,
7446 struct perf_event_context *child_ctx)
7448 struct perf_event *leader;
7449 struct perf_event *sub;
7450 struct perf_event *child_ctr;
7452 leader = inherit_event(parent_event, parent, parent_ctx,
7453 child, NULL, child_ctx);
7455 return PTR_ERR(leader);
7456 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7457 child_ctr = inherit_event(sub, parent, parent_ctx,
7458 child, leader, child_ctx);
7459 if (IS_ERR(child_ctr))
7460 return PTR_ERR(child_ctr);
7466 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7467 struct perf_event_context *parent_ctx,
7468 struct task_struct *child, int ctxn,
7472 struct perf_event_context *child_ctx;
7474 if (!event->attr.inherit) {
7479 child_ctx = child->perf_event_ctxp[ctxn];
7482 * This is executed from the parent task context, so
7483 * inherit events that have been marked for cloning.
7484 * First allocate and initialize a context for the
7488 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7492 child->perf_event_ctxp[ctxn] = child_ctx;
7495 ret = inherit_group(event, parent, parent_ctx,
7505 * Initialize the perf_event context in task_struct
7507 int perf_event_init_context(struct task_struct *child, int ctxn)
7509 struct perf_event_context *child_ctx, *parent_ctx;
7510 struct perf_event_context *cloned_ctx;
7511 struct perf_event *event;
7512 struct task_struct *parent = current;
7513 int inherited_all = 1;
7514 unsigned long flags;
7517 if (likely(!parent->perf_event_ctxp[ctxn]))
7521 * If the parent's context is a clone, pin it so it won't get
7524 parent_ctx = perf_pin_task_context(parent, ctxn);
7527 * No need to check if parent_ctx != NULL here; since we saw
7528 * it non-NULL earlier, the only reason for it to become NULL
7529 * is if we exit, and since we're currently in the middle of
7530 * a fork we can't be exiting at the same time.
7534 * Lock the parent list. No need to lock the child - not PID
7535 * hashed yet and not running, so nobody can access it.
7537 mutex_lock(&parent_ctx->mutex);
7540 * We dont have to disable NMIs - we are only looking at
7541 * the list, not manipulating it:
7543 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7544 ret = inherit_task_group(event, parent, parent_ctx,
7545 child, ctxn, &inherited_all);
7551 * We can't hold ctx->lock when iterating the ->flexible_group list due
7552 * to allocations, but we need to prevent rotation because
7553 * rotate_ctx() will change the list from interrupt context.
7555 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7556 parent_ctx->rotate_disable = 1;
7557 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7559 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7560 ret = inherit_task_group(event, parent, parent_ctx,
7561 child, ctxn, &inherited_all);
7566 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7567 parent_ctx->rotate_disable = 0;
7569 child_ctx = child->perf_event_ctxp[ctxn];
7571 if (child_ctx && inherited_all) {
7573 * Mark the child context as a clone of the parent
7574 * context, or of whatever the parent is a clone of.
7576 * Note that if the parent is a clone, the holding of
7577 * parent_ctx->lock avoids it from being uncloned.
7579 cloned_ctx = parent_ctx->parent_ctx;
7581 child_ctx->parent_ctx = cloned_ctx;
7582 child_ctx->parent_gen = parent_ctx->parent_gen;
7584 child_ctx->parent_ctx = parent_ctx;
7585 child_ctx->parent_gen = parent_ctx->generation;
7587 get_ctx(child_ctx->parent_ctx);
7590 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7591 mutex_unlock(&parent_ctx->mutex);
7593 perf_unpin_context(parent_ctx);
7594 put_ctx(parent_ctx);
7600 * Initialize the perf_event context in task_struct
7602 int perf_event_init_task(struct task_struct *child)
7606 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7607 mutex_init(&child->perf_event_mutex);
7608 INIT_LIST_HEAD(&child->perf_event_list);
7610 for_each_task_context_nr(ctxn) {
7611 ret = perf_event_init_context(child, ctxn);
7619 static void __init perf_event_init_all_cpus(void)
7621 struct swevent_htable *swhash;
7624 for_each_possible_cpu(cpu) {
7625 swhash = &per_cpu(swevent_htable, cpu);
7626 mutex_init(&swhash->hlist_mutex);
7627 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7631 static void perf_event_init_cpu(int cpu)
7633 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7635 mutex_lock(&swhash->hlist_mutex);
7636 if (swhash->hlist_refcount > 0) {
7637 struct swevent_hlist *hlist;
7639 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7641 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7643 mutex_unlock(&swhash->hlist_mutex);
7646 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7647 static void perf_pmu_rotate_stop(struct pmu *pmu)
7649 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7651 WARN_ON(!irqs_disabled());
7653 list_del_init(&cpuctx->rotation_list);
7656 static void __perf_event_exit_context(void *__info)
7658 struct perf_event_context *ctx = __info;
7659 struct perf_event *event, *tmp;
7661 perf_pmu_rotate_stop(ctx->pmu);
7663 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7664 __perf_remove_from_context(event);
7665 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7666 __perf_remove_from_context(event);
7669 static void perf_event_exit_cpu_context(int cpu)
7671 struct perf_event_context *ctx;
7675 idx = srcu_read_lock(&pmus_srcu);
7676 list_for_each_entry_rcu(pmu, &pmus, entry) {
7677 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7679 mutex_lock(&ctx->mutex);
7680 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7681 mutex_unlock(&ctx->mutex);
7683 srcu_read_unlock(&pmus_srcu, idx);
7686 static void perf_event_exit_cpu(int cpu)
7688 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7690 mutex_lock(&swhash->hlist_mutex);
7691 swevent_hlist_release(swhash);
7692 mutex_unlock(&swhash->hlist_mutex);
7694 perf_event_exit_cpu_context(cpu);
7697 static inline void perf_event_exit_cpu(int cpu) { }
7701 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7705 for_each_online_cpu(cpu)
7706 perf_event_exit_cpu(cpu);
7712 * Run the perf reboot notifier at the very last possible moment so that
7713 * the generic watchdog code runs as long as possible.
7715 static struct notifier_block perf_reboot_notifier = {
7716 .notifier_call = perf_reboot,
7717 .priority = INT_MIN,
7721 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7723 unsigned int cpu = (long)hcpu;
7725 switch (action & ~CPU_TASKS_FROZEN) {
7727 case CPU_UP_PREPARE:
7728 case CPU_DOWN_FAILED:
7729 perf_event_init_cpu(cpu);
7732 case CPU_UP_CANCELED:
7733 case CPU_DOWN_PREPARE:
7734 perf_event_exit_cpu(cpu);
7743 void __init perf_event_init(void)
7749 perf_event_init_all_cpus();
7750 init_srcu_struct(&pmus_srcu);
7751 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7752 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7753 perf_pmu_register(&perf_task_clock, NULL, -1);
7755 perf_cpu_notifier(perf_cpu_notify);
7756 register_reboot_notifier(&perf_reboot_notifier);
7758 ret = init_hw_breakpoint();
7759 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7761 /* do not patch jump label more than once per second */
7762 jump_label_rate_limit(&perf_sched_events, HZ);
7765 * Build time assertion that we keep the data_head at the intended
7766 * location. IOW, validation we got the __reserved[] size right.
7768 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7772 static int __init perf_event_sysfs_init(void)
7777 mutex_lock(&pmus_lock);
7779 ret = bus_register(&pmu_bus);
7783 list_for_each_entry(pmu, &pmus, entry) {
7784 if (!pmu->name || pmu->type < 0)
7787 ret = pmu_dev_alloc(pmu);
7788 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7790 pmu_bus_running = 1;
7794 mutex_unlock(&pmus_lock);
7798 device_initcall(perf_event_sysfs_init);
7800 #ifdef CONFIG_CGROUP_PERF
7801 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7803 struct perf_cgroup *jc;
7805 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7807 return ERR_PTR(-ENOMEM);
7809 jc->info = alloc_percpu(struct perf_cgroup_info);
7812 return ERR_PTR(-ENOMEM);
7818 static void perf_cgroup_css_free(struct cgroup *cont)
7820 struct perf_cgroup *jc;
7821 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7822 struct perf_cgroup, css);
7823 free_percpu(jc->info);
7827 static int __perf_cgroup_move(void *info)
7829 struct task_struct *task = info;
7830 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7834 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7836 struct task_struct *task;
7838 cgroup_taskset_for_each(task, cgrp, tset)
7839 task_function_call(task, __perf_cgroup_move, task);
7842 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7843 struct task_struct *task)
7846 * cgroup_exit() is called in the copy_process() failure path.
7847 * Ignore this case since the task hasn't ran yet, this avoids
7848 * trying to poke a half freed task state from generic code.
7850 if (!(task->flags & PF_EXITING))
7853 task_function_call(task, __perf_cgroup_move, task);
7856 struct cgroup_subsys perf_subsys = {
7857 .name = "perf_event",
7858 .subsys_id = perf_subsys_id,
7859 .css_alloc = perf_cgroup_css_alloc,
7860 .css_free = perf_cgroup_css_free,
7861 .exit = perf_cgroup_exit,
7862 .attach = perf_cgroup_attach,
7864 #endif /* CONFIG_CGROUP_PERF */