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_css(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);
596 css = css_from_dir(f.file->f_dentry, &perf_subsys);
602 cgrp = container_of(css, struct perf_cgroup, css);
605 /* must be done before we fput() the file */
606 if (!perf_tryget_cgroup(event)) {
613 * all events in a group must monitor
614 * the same cgroup because a task belongs
615 * to only one perf cgroup at a time
617 if (group_leader && group_leader->cgrp != cgrp) {
618 perf_detach_cgroup(event);
628 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
630 struct perf_cgroup_info *t;
631 t = per_cpu_ptr(event->cgrp->info, event->cpu);
632 event->shadow_ctx_time = now - t->timestamp;
636 perf_cgroup_defer_enabled(struct perf_event *event)
639 * when the current task's perf cgroup does not match
640 * the event's, we need to remember to call the
641 * perf_mark_enable() function the first time a task with
642 * a matching perf cgroup is scheduled in.
644 if (is_cgroup_event(event) && !perf_cgroup_match(event))
645 event->cgrp_defer_enabled = 1;
649 perf_cgroup_mark_enabled(struct perf_event *event,
650 struct perf_event_context *ctx)
652 struct perf_event *sub;
653 u64 tstamp = perf_event_time(event);
655 if (!event->cgrp_defer_enabled)
658 event->cgrp_defer_enabled = 0;
660 event->tstamp_enabled = tstamp - event->total_time_enabled;
661 list_for_each_entry(sub, &event->sibling_list, group_entry) {
662 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
663 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
664 sub->cgrp_defer_enabled = 0;
668 #else /* !CONFIG_CGROUP_PERF */
671 perf_cgroup_match(struct perf_event *event)
676 static inline void perf_detach_cgroup(struct perf_event *event)
679 static inline int is_cgroup_event(struct perf_event *event)
684 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
689 static inline void update_cgrp_time_from_event(struct perf_event *event)
693 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
697 static inline void perf_cgroup_sched_out(struct task_struct *task,
698 struct task_struct *next)
702 static inline void perf_cgroup_sched_in(struct task_struct *prev,
703 struct task_struct *task)
707 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
708 struct perf_event_attr *attr,
709 struct perf_event *group_leader)
715 perf_cgroup_set_timestamp(struct task_struct *task,
716 struct perf_event_context *ctx)
721 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
726 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
730 static inline u64 perf_cgroup_event_time(struct perf_event *event)
736 perf_cgroup_defer_enabled(struct perf_event *event)
741 perf_cgroup_mark_enabled(struct perf_event *event,
742 struct perf_event_context *ctx)
748 * set default to be dependent on timer tick just
751 #define PERF_CPU_HRTIMER (1000 / HZ)
753 * function must be called with interrupts disbled
755 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
757 struct perf_cpu_context *cpuctx;
758 enum hrtimer_restart ret = HRTIMER_NORESTART;
761 WARN_ON(!irqs_disabled());
763 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
765 rotations = perf_rotate_context(cpuctx);
768 * arm timer if needed
771 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
772 ret = HRTIMER_RESTART;
778 /* CPU is going down */
779 void perf_cpu_hrtimer_cancel(int cpu)
781 struct perf_cpu_context *cpuctx;
785 if (WARN_ON(cpu != smp_processor_id()))
788 local_irq_save(flags);
792 list_for_each_entry_rcu(pmu, &pmus, entry) {
793 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
795 if (pmu->task_ctx_nr == perf_sw_context)
798 hrtimer_cancel(&cpuctx->hrtimer);
803 local_irq_restore(flags);
806 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
808 struct hrtimer *hr = &cpuctx->hrtimer;
809 struct pmu *pmu = cpuctx->ctx.pmu;
812 /* no multiplexing needed for SW PMU */
813 if (pmu->task_ctx_nr == perf_sw_context)
817 * check default is sane, if not set then force to
818 * default interval (1/tick)
820 timer = pmu->hrtimer_interval_ms;
822 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
824 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
826 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
827 hr->function = perf_cpu_hrtimer_handler;
830 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
832 struct hrtimer *hr = &cpuctx->hrtimer;
833 struct pmu *pmu = cpuctx->ctx.pmu;
836 if (pmu->task_ctx_nr == perf_sw_context)
839 if (hrtimer_active(hr))
842 if (!hrtimer_callback_running(hr))
843 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
844 0, HRTIMER_MODE_REL_PINNED, 0);
847 void perf_pmu_disable(struct pmu *pmu)
849 int *count = this_cpu_ptr(pmu->pmu_disable_count);
851 pmu->pmu_disable(pmu);
854 void perf_pmu_enable(struct pmu *pmu)
856 int *count = this_cpu_ptr(pmu->pmu_disable_count);
858 pmu->pmu_enable(pmu);
861 static DEFINE_PER_CPU(struct list_head, rotation_list);
864 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
865 * because they're strictly cpu affine and rotate_start is called with IRQs
866 * disabled, while rotate_context is called from IRQ context.
868 static void perf_pmu_rotate_start(struct pmu *pmu)
870 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
871 struct list_head *head = &__get_cpu_var(rotation_list);
873 WARN_ON(!irqs_disabled());
875 if (list_empty(&cpuctx->rotation_list)) {
876 int was_empty = list_empty(head);
877 list_add(&cpuctx->rotation_list, head);
879 tick_nohz_full_kick();
883 static void get_ctx(struct perf_event_context *ctx)
885 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
888 static void put_ctx(struct perf_event_context *ctx)
890 if (atomic_dec_and_test(&ctx->refcount)) {
892 put_ctx(ctx->parent_ctx);
894 put_task_struct(ctx->task);
895 kfree_rcu(ctx, rcu_head);
899 static void unclone_ctx(struct perf_event_context *ctx)
901 if (ctx->parent_ctx) {
902 put_ctx(ctx->parent_ctx);
903 ctx->parent_ctx = NULL;
907 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
910 * only top level events have the pid namespace they were created in
913 event = event->parent;
915 return task_tgid_nr_ns(p, event->ns);
918 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
921 * only top level events have the pid namespace they were created in
924 event = event->parent;
926 return task_pid_nr_ns(p, event->ns);
930 * If we inherit events we want to return the parent event id
933 static u64 primary_event_id(struct perf_event *event)
938 id = event->parent->id;
944 * Get the perf_event_context for a task and lock it.
945 * This has to cope with with the fact that until it is locked,
946 * the context could get moved to another task.
948 static struct perf_event_context *
949 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
951 struct perf_event_context *ctx;
955 * One of the few rules of preemptible RCU is that one cannot do
956 * rcu_read_unlock() while holding a scheduler (or nested) lock when
957 * part of the read side critical section was preemptible -- see
958 * rcu_read_unlock_special().
960 * Since ctx->lock nests under rq->lock we must ensure the entire read
961 * side critical section is non-preemptible.
965 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
968 * If this context is a clone of another, it might
969 * get swapped for another underneath us by
970 * perf_event_task_sched_out, though the
971 * rcu_read_lock() protects us from any context
972 * getting freed. Lock the context and check if it
973 * got swapped before we could get the lock, and retry
974 * if so. If we locked the right context, then it
975 * can't get swapped on us any more.
977 raw_spin_lock_irqsave(&ctx->lock, *flags);
978 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
979 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
985 if (!atomic_inc_not_zero(&ctx->refcount)) {
986 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
996 * Get the context for a task and increment its pin_count so it
997 * can't get swapped to another task. This also increments its
998 * reference count so that the context can't get freed.
1000 static struct perf_event_context *
1001 perf_pin_task_context(struct task_struct *task, int ctxn)
1003 struct perf_event_context *ctx;
1004 unsigned long flags;
1006 ctx = perf_lock_task_context(task, ctxn, &flags);
1009 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1014 static void perf_unpin_context(struct perf_event_context *ctx)
1016 unsigned long flags;
1018 raw_spin_lock_irqsave(&ctx->lock, flags);
1020 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1024 * Update the record of the current time in a context.
1026 static void update_context_time(struct perf_event_context *ctx)
1028 u64 now = perf_clock();
1030 ctx->time += now - ctx->timestamp;
1031 ctx->timestamp = now;
1034 static u64 perf_event_time(struct perf_event *event)
1036 struct perf_event_context *ctx = event->ctx;
1038 if (is_cgroup_event(event))
1039 return perf_cgroup_event_time(event);
1041 return ctx ? ctx->time : 0;
1045 * Update the total_time_enabled and total_time_running fields for a event.
1046 * The caller of this function needs to hold the ctx->lock.
1048 static void update_event_times(struct perf_event *event)
1050 struct perf_event_context *ctx = event->ctx;
1053 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1054 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1057 * in cgroup mode, time_enabled represents
1058 * the time the event was enabled AND active
1059 * tasks were in the monitored cgroup. This is
1060 * independent of the activity of the context as
1061 * there may be a mix of cgroup and non-cgroup events.
1063 * That is why we treat cgroup events differently
1066 if (is_cgroup_event(event))
1067 run_end = perf_cgroup_event_time(event);
1068 else if (ctx->is_active)
1069 run_end = ctx->time;
1071 run_end = event->tstamp_stopped;
1073 event->total_time_enabled = run_end - event->tstamp_enabled;
1075 if (event->state == PERF_EVENT_STATE_INACTIVE)
1076 run_end = event->tstamp_stopped;
1078 run_end = perf_event_time(event);
1080 event->total_time_running = run_end - event->tstamp_running;
1085 * Update total_time_enabled and total_time_running for all events in a group.
1087 static void update_group_times(struct perf_event *leader)
1089 struct perf_event *event;
1091 update_event_times(leader);
1092 list_for_each_entry(event, &leader->sibling_list, group_entry)
1093 update_event_times(event);
1096 static struct list_head *
1097 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1099 if (event->attr.pinned)
1100 return &ctx->pinned_groups;
1102 return &ctx->flexible_groups;
1106 * Add a event from the lists for its context.
1107 * Must be called with ctx->mutex and ctx->lock held.
1110 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1112 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1113 event->attach_state |= PERF_ATTACH_CONTEXT;
1116 * If we're a stand alone event or group leader, we go to the context
1117 * list, group events are kept attached to the group so that
1118 * perf_group_detach can, at all times, locate all siblings.
1120 if (event->group_leader == event) {
1121 struct list_head *list;
1123 if (is_software_event(event))
1124 event->group_flags |= PERF_GROUP_SOFTWARE;
1126 list = ctx_group_list(event, ctx);
1127 list_add_tail(&event->group_entry, list);
1130 if (is_cgroup_event(event))
1133 if (has_branch_stack(event))
1134 ctx->nr_branch_stack++;
1136 list_add_rcu(&event->event_entry, &ctx->event_list);
1137 if (!ctx->nr_events)
1138 perf_pmu_rotate_start(ctx->pmu);
1140 if (event->attr.inherit_stat)
1145 * Initialize event state based on the perf_event_attr::disabled.
1147 static inline void perf_event__state_init(struct perf_event *event)
1149 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1150 PERF_EVENT_STATE_INACTIVE;
1154 * Called at perf_event creation and when events are attached/detached from a
1157 static void perf_event__read_size(struct perf_event *event)
1159 int entry = sizeof(u64); /* value */
1163 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1164 size += sizeof(u64);
1166 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1167 size += sizeof(u64);
1169 if (event->attr.read_format & PERF_FORMAT_ID)
1170 entry += sizeof(u64);
1172 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1173 nr += event->group_leader->nr_siblings;
1174 size += sizeof(u64);
1178 event->read_size = size;
1181 static void perf_event__header_size(struct perf_event *event)
1183 struct perf_sample_data *data;
1184 u64 sample_type = event->attr.sample_type;
1187 perf_event__read_size(event);
1189 if (sample_type & PERF_SAMPLE_IP)
1190 size += sizeof(data->ip);
1192 if (sample_type & PERF_SAMPLE_ADDR)
1193 size += sizeof(data->addr);
1195 if (sample_type & PERF_SAMPLE_PERIOD)
1196 size += sizeof(data->period);
1198 if (sample_type & PERF_SAMPLE_WEIGHT)
1199 size += sizeof(data->weight);
1201 if (sample_type & PERF_SAMPLE_READ)
1202 size += event->read_size;
1204 if (sample_type & PERF_SAMPLE_DATA_SRC)
1205 size += sizeof(data->data_src.val);
1207 event->header_size = size;
1210 static void perf_event__id_header_size(struct perf_event *event)
1212 struct perf_sample_data *data;
1213 u64 sample_type = event->attr.sample_type;
1216 if (sample_type & PERF_SAMPLE_TID)
1217 size += sizeof(data->tid_entry);
1219 if (sample_type & PERF_SAMPLE_TIME)
1220 size += sizeof(data->time);
1222 if (sample_type & PERF_SAMPLE_ID)
1223 size += sizeof(data->id);
1225 if (sample_type & PERF_SAMPLE_STREAM_ID)
1226 size += sizeof(data->stream_id);
1228 if (sample_type & PERF_SAMPLE_CPU)
1229 size += sizeof(data->cpu_entry);
1231 event->id_header_size = size;
1234 static void perf_group_attach(struct perf_event *event)
1236 struct perf_event *group_leader = event->group_leader, *pos;
1239 * We can have double attach due to group movement in perf_event_open.
1241 if (event->attach_state & PERF_ATTACH_GROUP)
1244 event->attach_state |= PERF_ATTACH_GROUP;
1246 if (group_leader == event)
1249 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1250 !is_software_event(event))
1251 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1253 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1254 group_leader->nr_siblings++;
1256 perf_event__header_size(group_leader);
1258 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1259 perf_event__header_size(pos);
1263 * Remove a event from the lists for its context.
1264 * Must be called with ctx->mutex and ctx->lock held.
1267 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1269 struct perf_cpu_context *cpuctx;
1271 * We can have double detach due to exit/hot-unplug + close.
1273 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1276 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1278 if (is_cgroup_event(event)) {
1280 cpuctx = __get_cpu_context(ctx);
1282 * if there are no more cgroup events
1283 * then cler cgrp to avoid stale pointer
1284 * in update_cgrp_time_from_cpuctx()
1286 if (!ctx->nr_cgroups)
1287 cpuctx->cgrp = NULL;
1290 if (has_branch_stack(event))
1291 ctx->nr_branch_stack--;
1294 if (event->attr.inherit_stat)
1297 list_del_rcu(&event->event_entry);
1299 if (event->group_leader == event)
1300 list_del_init(&event->group_entry);
1302 update_group_times(event);
1305 * If event was in error state, then keep it
1306 * that way, otherwise bogus counts will be
1307 * returned on read(). The only way to get out
1308 * of error state is by explicit re-enabling
1311 if (event->state > PERF_EVENT_STATE_OFF)
1312 event->state = PERF_EVENT_STATE_OFF;
1315 static void perf_group_detach(struct perf_event *event)
1317 struct perf_event *sibling, *tmp;
1318 struct list_head *list = NULL;
1321 * We can have double detach due to exit/hot-unplug + close.
1323 if (!(event->attach_state & PERF_ATTACH_GROUP))
1326 event->attach_state &= ~PERF_ATTACH_GROUP;
1329 * If this is a sibling, remove it from its group.
1331 if (event->group_leader != event) {
1332 list_del_init(&event->group_entry);
1333 event->group_leader->nr_siblings--;
1337 if (!list_empty(&event->group_entry))
1338 list = &event->group_entry;
1341 * If this was a group event with sibling events then
1342 * upgrade the siblings to singleton events by adding them
1343 * to whatever list we are on.
1345 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1347 list_move_tail(&sibling->group_entry, list);
1348 sibling->group_leader = sibling;
1350 /* Inherit group flags from the previous leader */
1351 sibling->group_flags = event->group_flags;
1355 perf_event__header_size(event->group_leader);
1357 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1358 perf_event__header_size(tmp);
1362 event_filter_match(struct perf_event *event)
1364 return (event->cpu == -1 || event->cpu == smp_processor_id())
1365 && perf_cgroup_match(event);
1369 event_sched_out(struct perf_event *event,
1370 struct perf_cpu_context *cpuctx,
1371 struct perf_event_context *ctx)
1373 u64 tstamp = perf_event_time(event);
1376 * An event which could not be activated because of
1377 * filter mismatch still needs to have its timings
1378 * maintained, otherwise bogus information is return
1379 * via read() for time_enabled, time_running:
1381 if (event->state == PERF_EVENT_STATE_INACTIVE
1382 && !event_filter_match(event)) {
1383 delta = tstamp - event->tstamp_stopped;
1384 event->tstamp_running += delta;
1385 event->tstamp_stopped = tstamp;
1388 if (event->state != PERF_EVENT_STATE_ACTIVE)
1391 event->state = PERF_EVENT_STATE_INACTIVE;
1392 if (event->pending_disable) {
1393 event->pending_disable = 0;
1394 event->state = PERF_EVENT_STATE_OFF;
1396 event->tstamp_stopped = tstamp;
1397 event->pmu->del(event, 0);
1400 if (!is_software_event(event))
1401 cpuctx->active_oncpu--;
1403 if (event->attr.freq && event->attr.sample_freq)
1405 if (event->attr.exclusive || !cpuctx->active_oncpu)
1406 cpuctx->exclusive = 0;
1410 group_sched_out(struct perf_event *group_event,
1411 struct perf_cpu_context *cpuctx,
1412 struct perf_event_context *ctx)
1414 struct perf_event *event;
1415 int state = group_event->state;
1417 event_sched_out(group_event, cpuctx, ctx);
1420 * Schedule out siblings (if any):
1422 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1423 event_sched_out(event, cpuctx, ctx);
1425 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1426 cpuctx->exclusive = 0;
1430 * Cross CPU call to remove a performance event
1432 * We disable the event on the hardware level first. After that we
1433 * remove it from the context list.
1435 static int __perf_remove_from_context(void *info)
1437 struct perf_event *event = info;
1438 struct perf_event_context *ctx = event->ctx;
1439 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1441 raw_spin_lock(&ctx->lock);
1442 event_sched_out(event, cpuctx, ctx);
1443 list_del_event(event, ctx);
1444 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1446 cpuctx->task_ctx = NULL;
1448 raw_spin_unlock(&ctx->lock);
1455 * Remove the event from a task's (or a CPU's) list of events.
1457 * CPU events are removed with a smp call. For task events we only
1458 * call when the task is on a CPU.
1460 * If event->ctx is a cloned context, callers must make sure that
1461 * every task struct that event->ctx->task could possibly point to
1462 * remains valid. This is OK when called from perf_release since
1463 * that only calls us on the top-level context, which can't be a clone.
1464 * When called from perf_event_exit_task, it's OK because the
1465 * context has been detached from its task.
1467 static void perf_remove_from_context(struct perf_event *event)
1469 struct perf_event_context *ctx = event->ctx;
1470 struct task_struct *task = ctx->task;
1472 lockdep_assert_held(&ctx->mutex);
1476 * Per cpu events are removed via an smp call and
1477 * the removal is always successful.
1479 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1484 if (!task_function_call(task, __perf_remove_from_context, event))
1487 raw_spin_lock_irq(&ctx->lock);
1489 * If we failed to find a running task, but find the context active now
1490 * that we've acquired the ctx->lock, retry.
1492 if (ctx->is_active) {
1493 raw_spin_unlock_irq(&ctx->lock);
1498 * Since the task isn't running, its safe to remove the event, us
1499 * holding the ctx->lock ensures the task won't get scheduled in.
1501 list_del_event(event, ctx);
1502 raw_spin_unlock_irq(&ctx->lock);
1506 * Cross CPU call to disable a performance event
1508 int __perf_event_disable(void *info)
1510 struct perf_event *event = info;
1511 struct perf_event_context *ctx = event->ctx;
1512 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1515 * If this is a per-task event, need to check whether this
1516 * event's task is the current task on this cpu.
1518 * Can trigger due to concurrent perf_event_context_sched_out()
1519 * flipping contexts around.
1521 if (ctx->task && cpuctx->task_ctx != ctx)
1524 raw_spin_lock(&ctx->lock);
1527 * If the event is on, turn it off.
1528 * If it is in error state, leave it in error state.
1530 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1531 update_context_time(ctx);
1532 update_cgrp_time_from_event(event);
1533 update_group_times(event);
1534 if (event == event->group_leader)
1535 group_sched_out(event, cpuctx, ctx);
1537 event_sched_out(event, cpuctx, ctx);
1538 event->state = PERF_EVENT_STATE_OFF;
1541 raw_spin_unlock(&ctx->lock);
1549 * If event->ctx is a cloned context, callers must make sure that
1550 * every task struct that event->ctx->task could possibly point to
1551 * remains valid. This condition is satisifed when called through
1552 * perf_event_for_each_child or perf_event_for_each because they
1553 * hold the top-level event's child_mutex, so any descendant that
1554 * goes to exit will block in sync_child_event.
1555 * When called from perf_pending_event it's OK because event->ctx
1556 * is the current context on this CPU and preemption is disabled,
1557 * hence we can't get into perf_event_task_sched_out for this context.
1559 void perf_event_disable(struct perf_event *event)
1561 struct perf_event_context *ctx = event->ctx;
1562 struct task_struct *task = ctx->task;
1566 * Disable the event on the cpu that it's on
1568 cpu_function_call(event->cpu, __perf_event_disable, event);
1573 if (!task_function_call(task, __perf_event_disable, event))
1576 raw_spin_lock_irq(&ctx->lock);
1578 * If the event is still active, we need to retry the cross-call.
1580 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1581 raw_spin_unlock_irq(&ctx->lock);
1583 * Reload the task pointer, it might have been changed by
1584 * a concurrent perf_event_context_sched_out().
1591 * Since we have the lock this context can't be scheduled
1592 * in, so we can change the state safely.
1594 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1595 update_group_times(event);
1596 event->state = PERF_EVENT_STATE_OFF;
1598 raw_spin_unlock_irq(&ctx->lock);
1600 EXPORT_SYMBOL_GPL(perf_event_disable);
1602 static void perf_set_shadow_time(struct perf_event *event,
1603 struct perf_event_context *ctx,
1607 * use the correct time source for the time snapshot
1609 * We could get by without this by leveraging the
1610 * fact that to get to this function, the caller
1611 * has most likely already called update_context_time()
1612 * and update_cgrp_time_xx() and thus both timestamp
1613 * are identical (or very close). Given that tstamp is,
1614 * already adjusted for cgroup, we could say that:
1615 * tstamp - ctx->timestamp
1617 * tstamp - cgrp->timestamp.
1619 * Then, in perf_output_read(), the calculation would
1620 * work with no changes because:
1621 * - event is guaranteed scheduled in
1622 * - no scheduled out in between
1623 * - thus the timestamp would be the same
1625 * But this is a bit hairy.
1627 * So instead, we have an explicit cgroup call to remain
1628 * within the time time source all along. We believe it
1629 * is cleaner and simpler to understand.
1631 if (is_cgroup_event(event))
1632 perf_cgroup_set_shadow_time(event, tstamp);
1634 event->shadow_ctx_time = tstamp - ctx->timestamp;
1637 #define MAX_INTERRUPTS (~0ULL)
1639 static void perf_log_throttle(struct perf_event *event, int enable);
1642 event_sched_in(struct perf_event *event,
1643 struct perf_cpu_context *cpuctx,
1644 struct perf_event_context *ctx)
1646 u64 tstamp = perf_event_time(event);
1648 if (event->state <= PERF_EVENT_STATE_OFF)
1651 event->state = PERF_EVENT_STATE_ACTIVE;
1652 event->oncpu = smp_processor_id();
1655 * Unthrottle events, since we scheduled we might have missed several
1656 * ticks already, also for a heavily scheduling task there is little
1657 * guarantee it'll get a tick in a timely manner.
1659 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1660 perf_log_throttle(event, 1);
1661 event->hw.interrupts = 0;
1665 * The new state must be visible before we turn it on in the hardware:
1669 if (event->pmu->add(event, PERF_EF_START)) {
1670 event->state = PERF_EVENT_STATE_INACTIVE;
1675 event->tstamp_running += tstamp - event->tstamp_stopped;
1677 perf_set_shadow_time(event, ctx, tstamp);
1679 if (!is_software_event(event))
1680 cpuctx->active_oncpu++;
1682 if (event->attr.freq && event->attr.sample_freq)
1685 if (event->attr.exclusive)
1686 cpuctx->exclusive = 1;
1692 group_sched_in(struct perf_event *group_event,
1693 struct perf_cpu_context *cpuctx,
1694 struct perf_event_context *ctx)
1696 struct perf_event *event, *partial_group = NULL;
1697 struct pmu *pmu = group_event->pmu;
1698 u64 now = ctx->time;
1699 bool simulate = false;
1701 if (group_event->state == PERF_EVENT_STATE_OFF)
1704 pmu->start_txn(pmu);
1706 if (event_sched_in(group_event, cpuctx, ctx)) {
1707 pmu->cancel_txn(pmu);
1708 perf_cpu_hrtimer_restart(cpuctx);
1713 * Schedule in siblings as one group (if any):
1715 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1716 if (event_sched_in(event, cpuctx, ctx)) {
1717 partial_group = event;
1722 if (!pmu->commit_txn(pmu))
1727 * Groups can be scheduled in as one unit only, so undo any
1728 * partial group before returning:
1729 * The events up to the failed event are scheduled out normally,
1730 * tstamp_stopped will be updated.
1732 * The failed events and the remaining siblings need to have
1733 * their timings updated as if they had gone thru event_sched_in()
1734 * and event_sched_out(). This is required to get consistent timings
1735 * across the group. This also takes care of the case where the group
1736 * could never be scheduled by ensuring tstamp_stopped is set to mark
1737 * the time the event was actually stopped, such that time delta
1738 * calculation in update_event_times() is correct.
1740 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1741 if (event == partial_group)
1745 event->tstamp_running += now - event->tstamp_stopped;
1746 event->tstamp_stopped = now;
1748 event_sched_out(event, cpuctx, ctx);
1751 event_sched_out(group_event, cpuctx, ctx);
1753 pmu->cancel_txn(pmu);
1755 perf_cpu_hrtimer_restart(cpuctx);
1761 * Work out whether we can put this event group on the CPU now.
1763 static int group_can_go_on(struct perf_event *event,
1764 struct perf_cpu_context *cpuctx,
1768 * Groups consisting entirely of software events can always go on.
1770 if (event->group_flags & PERF_GROUP_SOFTWARE)
1773 * If an exclusive group is already on, no other hardware
1776 if (cpuctx->exclusive)
1779 * If this group is exclusive and there are already
1780 * events on the CPU, it can't go on.
1782 if (event->attr.exclusive && cpuctx->active_oncpu)
1785 * Otherwise, try to add it if all previous groups were able
1791 static void add_event_to_ctx(struct perf_event *event,
1792 struct perf_event_context *ctx)
1794 u64 tstamp = perf_event_time(event);
1796 list_add_event(event, ctx);
1797 perf_group_attach(event);
1798 event->tstamp_enabled = tstamp;
1799 event->tstamp_running = tstamp;
1800 event->tstamp_stopped = tstamp;
1803 static void task_ctx_sched_out(struct perf_event_context *ctx);
1805 ctx_sched_in(struct perf_event_context *ctx,
1806 struct perf_cpu_context *cpuctx,
1807 enum event_type_t event_type,
1808 struct task_struct *task);
1810 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1811 struct perf_event_context *ctx,
1812 struct task_struct *task)
1814 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1816 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1817 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1819 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1823 * Cross CPU call to install and enable a performance event
1825 * Must be called with ctx->mutex held
1827 static int __perf_install_in_context(void *info)
1829 struct perf_event *event = info;
1830 struct perf_event_context *ctx = event->ctx;
1831 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1832 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1833 struct task_struct *task = current;
1835 perf_ctx_lock(cpuctx, task_ctx);
1836 perf_pmu_disable(cpuctx->ctx.pmu);
1839 * If there was an active task_ctx schedule it out.
1842 task_ctx_sched_out(task_ctx);
1845 * If the context we're installing events in is not the
1846 * active task_ctx, flip them.
1848 if (ctx->task && task_ctx != ctx) {
1850 raw_spin_unlock(&task_ctx->lock);
1851 raw_spin_lock(&ctx->lock);
1856 cpuctx->task_ctx = task_ctx;
1857 task = task_ctx->task;
1860 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1862 update_context_time(ctx);
1864 * update cgrp time only if current cgrp
1865 * matches event->cgrp. Must be done before
1866 * calling add_event_to_ctx()
1868 update_cgrp_time_from_event(event);
1870 add_event_to_ctx(event, ctx);
1873 * Schedule everything back in
1875 perf_event_sched_in(cpuctx, task_ctx, task);
1877 perf_pmu_enable(cpuctx->ctx.pmu);
1878 perf_ctx_unlock(cpuctx, task_ctx);
1884 * Attach a performance event to a context
1886 * First we add the event to the list with the hardware enable bit
1887 * in event->hw_config cleared.
1889 * If the event is attached to a task which is on a CPU we use a smp
1890 * call to enable it in the task context. The task might have been
1891 * scheduled away, but we check this in the smp call again.
1894 perf_install_in_context(struct perf_event_context *ctx,
1895 struct perf_event *event,
1898 struct task_struct *task = ctx->task;
1900 lockdep_assert_held(&ctx->mutex);
1903 if (event->cpu != -1)
1908 * Per cpu events are installed via an smp call and
1909 * the install is always successful.
1911 cpu_function_call(cpu, __perf_install_in_context, event);
1916 if (!task_function_call(task, __perf_install_in_context, event))
1919 raw_spin_lock_irq(&ctx->lock);
1921 * If we failed to find a running task, but find the context active now
1922 * that we've acquired the ctx->lock, retry.
1924 if (ctx->is_active) {
1925 raw_spin_unlock_irq(&ctx->lock);
1930 * Since the task isn't running, its safe to add the event, us holding
1931 * the ctx->lock ensures the task won't get scheduled in.
1933 add_event_to_ctx(event, ctx);
1934 raw_spin_unlock_irq(&ctx->lock);
1938 * Put a event into inactive state and update time fields.
1939 * Enabling the leader of a group effectively enables all
1940 * the group members that aren't explicitly disabled, so we
1941 * have to update their ->tstamp_enabled also.
1942 * Note: this works for group members as well as group leaders
1943 * since the non-leader members' sibling_lists will be empty.
1945 static void __perf_event_mark_enabled(struct perf_event *event)
1947 struct perf_event *sub;
1948 u64 tstamp = perf_event_time(event);
1950 event->state = PERF_EVENT_STATE_INACTIVE;
1951 event->tstamp_enabled = tstamp - event->total_time_enabled;
1952 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1953 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1954 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1959 * Cross CPU call to enable a performance event
1961 static int __perf_event_enable(void *info)
1963 struct perf_event *event = info;
1964 struct perf_event_context *ctx = event->ctx;
1965 struct perf_event *leader = event->group_leader;
1966 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1970 * There's a time window between 'ctx->is_active' check
1971 * in perf_event_enable function and this place having:
1973 * - ctx->lock unlocked
1975 * where the task could be killed and 'ctx' deactivated
1976 * by perf_event_exit_task.
1978 if (!ctx->is_active)
1981 raw_spin_lock(&ctx->lock);
1982 update_context_time(ctx);
1984 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1988 * set current task's cgroup time reference point
1990 perf_cgroup_set_timestamp(current, ctx);
1992 __perf_event_mark_enabled(event);
1994 if (!event_filter_match(event)) {
1995 if (is_cgroup_event(event))
1996 perf_cgroup_defer_enabled(event);
2001 * If the event is in a group and isn't the group leader,
2002 * then don't put it on unless the group is on.
2004 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2007 if (!group_can_go_on(event, cpuctx, 1)) {
2010 if (event == leader)
2011 err = group_sched_in(event, cpuctx, ctx);
2013 err = event_sched_in(event, cpuctx, ctx);
2018 * If this event can't go on and it's part of a
2019 * group, then the whole group has to come off.
2021 if (leader != event) {
2022 group_sched_out(leader, cpuctx, ctx);
2023 perf_cpu_hrtimer_restart(cpuctx);
2025 if (leader->attr.pinned) {
2026 update_group_times(leader);
2027 leader->state = PERF_EVENT_STATE_ERROR;
2032 raw_spin_unlock(&ctx->lock);
2040 * If event->ctx is a cloned context, callers must make sure that
2041 * every task struct that event->ctx->task could possibly point to
2042 * remains valid. This condition is satisfied when called through
2043 * perf_event_for_each_child or perf_event_for_each as described
2044 * for perf_event_disable.
2046 void perf_event_enable(struct perf_event *event)
2048 struct perf_event_context *ctx = event->ctx;
2049 struct task_struct *task = ctx->task;
2053 * Enable the event on the cpu that it's on
2055 cpu_function_call(event->cpu, __perf_event_enable, event);
2059 raw_spin_lock_irq(&ctx->lock);
2060 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2064 * If the event is in error state, clear that first.
2065 * That way, if we see the event in error state below, we
2066 * know that it has gone back into error state, as distinct
2067 * from the task having been scheduled away before the
2068 * cross-call arrived.
2070 if (event->state == PERF_EVENT_STATE_ERROR)
2071 event->state = PERF_EVENT_STATE_OFF;
2074 if (!ctx->is_active) {
2075 __perf_event_mark_enabled(event);
2079 raw_spin_unlock_irq(&ctx->lock);
2081 if (!task_function_call(task, __perf_event_enable, event))
2084 raw_spin_lock_irq(&ctx->lock);
2087 * If the context is active and the event is still off,
2088 * we need to retry the cross-call.
2090 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2092 * task could have been flipped by a concurrent
2093 * perf_event_context_sched_out()
2100 raw_spin_unlock_irq(&ctx->lock);
2102 EXPORT_SYMBOL_GPL(perf_event_enable);
2104 int perf_event_refresh(struct perf_event *event, int refresh)
2107 * not supported on inherited events
2109 if (event->attr.inherit || !is_sampling_event(event))
2112 atomic_add(refresh, &event->event_limit);
2113 perf_event_enable(event);
2117 EXPORT_SYMBOL_GPL(perf_event_refresh);
2119 static void ctx_sched_out(struct perf_event_context *ctx,
2120 struct perf_cpu_context *cpuctx,
2121 enum event_type_t event_type)
2123 struct perf_event *event;
2124 int is_active = ctx->is_active;
2126 ctx->is_active &= ~event_type;
2127 if (likely(!ctx->nr_events))
2130 update_context_time(ctx);
2131 update_cgrp_time_from_cpuctx(cpuctx);
2132 if (!ctx->nr_active)
2135 perf_pmu_disable(ctx->pmu);
2136 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2137 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2138 group_sched_out(event, cpuctx, ctx);
2141 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2142 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2143 group_sched_out(event, cpuctx, ctx);
2145 perf_pmu_enable(ctx->pmu);
2149 * Test whether two contexts are equivalent, i.e. whether they
2150 * have both been cloned from the same version of the same context
2151 * and they both have the same number of enabled events.
2152 * If the number of enabled events is the same, then the set
2153 * of enabled events should be the same, because these are both
2154 * inherited contexts, therefore we can't access individual events
2155 * in them directly with an fd; we can only enable/disable all
2156 * events via prctl, or enable/disable all events in a family
2157 * via ioctl, which will have the same effect on both contexts.
2159 static int context_equiv(struct perf_event_context *ctx1,
2160 struct perf_event_context *ctx2)
2162 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
2163 && ctx1->parent_gen == ctx2->parent_gen
2164 && !ctx1->pin_count && !ctx2->pin_count;
2167 static void __perf_event_sync_stat(struct perf_event *event,
2168 struct perf_event *next_event)
2172 if (!event->attr.inherit_stat)
2176 * Update the event value, we cannot use perf_event_read()
2177 * because we're in the middle of a context switch and have IRQs
2178 * disabled, which upsets smp_call_function_single(), however
2179 * we know the event must be on the current CPU, therefore we
2180 * don't need to use it.
2182 switch (event->state) {
2183 case PERF_EVENT_STATE_ACTIVE:
2184 event->pmu->read(event);
2187 case PERF_EVENT_STATE_INACTIVE:
2188 update_event_times(event);
2196 * In order to keep per-task stats reliable we need to flip the event
2197 * values when we flip the contexts.
2199 value = local64_read(&next_event->count);
2200 value = local64_xchg(&event->count, value);
2201 local64_set(&next_event->count, value);
2203 swap(event->total_time_enabled, next_event->total_time_enabled);
2204 swap(event->total_time_running, next_event->total_time_running);
2207 * Since we swizzled the values, update the user visible data too.
2209 perf_event_update_userpage(event);
2210 perf_event_update_userpage(next_event);
2213 #define list_next_entry(pos, member) \
2214 list_entry(pos->member.next, typeof(*pos), member)
2216 static void perf_event_sync_stat(struct perf_event_context *ctx,
2217 struct perf_event_context *next_ctx)
2219 struct perf_event *event, *next_event;
2224 update_context_time(ctx);
2226 event = list_first_entry(&ctx->event_list,
2227 struct perf_event, event_entry);
2229 next_event = list_first_entry(&next_ctx->event_list,
2230 struct perf_event, event_entry);
2232 while (&event->event_entry != &ctx->event_list &&
2233 &next_event->event_entry != &next_ctx->event_list) {
2235 __perf_event_sync_stat(event, next_event);
2237 event = list_next_entry(event, event_entry);
2238 next_event = list_next_entry(next_event, event_entry);
2242 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2243 struct task_struct *next)
2245 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2246 struct perf_event_context *next_ctx;
2247 struct perf_event_context *parent;
2248 struct perf_cpu_context *cpuctx;
2254 cpuctx = __get_cpu_context(ctx);
2255 if (!cpuctx->task_ctx)
2259 parent = rcu_dereference(ctx->parent_ctx);
2260 next_ctx = next->perf_event_ctxp[ctxn];
2261 if (parent && next_ctx &&
2262 rcu_dereference(next_ctx->parent_ctx) == parent) {
2264 * Looks like the two contexts are clones, so we might be
2265 * able to optimize the context switch. We lock both
2266 * contexts and check that they are clones under the
2267 * lock (including re-checking that neither has been
2268 * uncloned in the meantime). It doesn't matter which
2269 * order we take the locks because no other cpu could
2270 * be trying to lock both of these tasks.
2272 raw_spin_lock(&ctx->lock);
2273 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2274 if (context_equiv(ctx, next_ctx)) {
2276 * XXX do we need a memory barrier of sorts
2277 * wrt to rcu_dereference() of perf_event_ctxp
2279 task->perf_event_ctxp[ctxn] = next_ctx;
2280 next->perf_event_ctxp[ctxn] = ctx;
2282 next_ctx->task = task;
2285 perf_event_sync_stat(ctx, next_ctx);
2287 raw_spin_unlock(&next_ctx->lock);
2288 raw_spin_unlock(&ctx->lock);
2293 raw_spin_lock(&ctx->lock);
2294 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2295 cpuctx->task_ctx = NULL;
2296 raw_spin_unlock(&ctx->lock);
2300 #define for_each_task_context_nr(ctxn) \
2301 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2304 * Called from scheduler to remove the events of the current task,
2305 * with interrupts disabled.
2307 * We stop each event and update the event value in event->count.
2309 * This does not protect us against NMI, but disable()
2310 * sets the disabled bit in the control field of event _before_
2311 * accessing the event control register. If a NMI hits, then it will
2312 * not restart the event.
2314 void __perf_event_task_sched_out(struct task_struct *task,
2315 struct task_struct *next)
2319 for_each_task_context_nr(ctxn)
2320 perf_event_context_sched_out(task, ctxn, next);
2323 * if cgroup events exist on this CPU, then we need
2324 * to check if we have to switch out PMU state.
2325 * cgroup event are system-wide mode only
2327 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2328 perf_cgroup_sched_out(task, next);
2331 static void task_ctx_sched_out(struct perf_event_context *ctx)
2333 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2335 if (!cpuctx->task_ctx)
2338 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2341 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2342 cpuctx->task_ctx = NULL;
2346 * Called with IRQs disabled
2348 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2349 enum event_type_t event_type)
2351 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2355 ctx_pinned_sched_in(struct perf_event_context *ctx,
2356 struct perf_cpu_context *cpuctx)
2358 struct perf_event *event;
2360 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2361 if (event->state <= PERF_EVENT_STATE_OFF)
2363 if (!event_filter_match(event))
2366 /* may need to reset tstamp_enabled */
2367 if (is_cgroup_event(event))
2368 perf_cgroup_mark_enabled(event, ctx);
2370 if (group_can_go_on(event, cpuctx, 1))
2371 group_sched_in(event, cpuctx, ctx);
2374 * If this pinned group hasn't been scheduled,
2375 * put it in error state.
2377 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2378 update_group_times(event);
2379 event->state = PERF_EVENT_STATE_ERROR;
2385 ctx_flexible_sched_in(struct perf_event_context *ctx,
2386 struct perf_cpu_context *cpuctx)
2388 struct perf_event *event;
2391 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2392 /* Ignore events in OFF or ERROR state */
2393 if (event->state <= PERF_EVENT_STATE_OFF)
2396 * Listen to the 'cpu' scheduling filter constraint
2399 if (!event_filter_match(event))
2402 /* may need to reset tstamp_enabled */
2403 if (is_cgroup_event(event))
2404 perf_cgroup_mark_enabled(event, ctx);
2406 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2407 if (group_sched_in(event, cpuctx, ctx))
2414 ctx_sched_in(struct perf_event_context *ctx,
2415 struct perf_cpu_context *cpuctx,
2416 enum event_type_t event_type,
2417 struct task_struct *task)
2420 int is_active = ctx->is_active;
2422 ctx->is_active |= event_type;
2423 if (likely(!ctx->nr_events))
2427 ctx->timestamp = now;
2428 perf_cgroup_set_timestamp(task, ctx);
2430 * First go through the list and put on any pinned groups
2431 * in order to give them the best chance of going on.
2433 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2434 ctx_pinned_sched_in(ctx, cpuctx);
2436 /* Then walk through the lower prio flexible groups */
2437 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2438 ctx_flexible_sched_in(ctx, cpuctx);
2441 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2442 enum event_type_t event_type,
2443 struct task_struct *task)
2445 struct perf_event_context *ctx = &cpuctx->ctx;
2447 ctx_sched_in(ctx, cpuctx, event_type, task);
2450 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2451 struct task_struct *task)
2453 struct perf_cpu_context *cpuctx;
2455 cpuctx = __get_cpu_context(ctx);
2456 if (cpuctx->task_ctx == ctx)
2459 perf_ctx_lock(cpuctx, ctx);
2460 perf_pmu_disable(ctx->pmu);
2462 * We want to keep the following priority order:
2463 * cpu pinned (that don't need to move), task pinned,
2464 * cpu flexible, task flexible.
2466 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2469 cpuctx->task_ctx = ctx;
2471 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2473 perf_pmu_enable(ctx->pmu);
2474 perf_ctx_unlock(cpuctx, ctx);
2477 * Since these rotations are per-cpu, we need to ensure the
2478 * cpu-context we got scheduled on is actually rotating.
2480 perf_pmu_rotate_start(ctx->pmu);
2484 * When sampling the branck stack in system-wide, it may be necessary
2485 * to flush the stack on context switch. This happens when the branch
2486 * stack does not tag its entries with the pid of the current task.
2487 * Otherwise it becomes impossible to associate a branch entry with a
2488 * task. This ambiguity is more likely to appear when the branch stack
2489 * supports priv level filtering and the user sets it to monitor only
2490 * at the user level (which could be a useful measurement in system-wide
2491 * mode). In that case, the risk is high of having a branch stack with
2492 * branch from multiple tasks. Flushing may mean dropping the existing
2493 * entries or stashing them somewhere in the PMU specific code layer.
2495 * This function provides the context switch callback to the lower code
2496 * layer. It is invoked ONLY when there is at least one system-wide context
2497 * with at least one active event using taken branch sampling.
2499 static void perf_branch_stack_sched_in(struct task_struct *prev,
2500 struct task_struct *task)
2502 struct perf_cpu_context *cpuctx;
2504 unsigned long flags;
2506 /* no need to flush branch stack if not changing task */
2510 local_irq_save(flags);
2514 list_for_each_entry_rcu(pmu, &pmus, entry) {
2515 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2518 * check if the context has at least one
2519 * event using PERF_SAMPLE_BRANCH_STACK
2521 if (cpuctx->ctx.nr_branch_stack > 0
2522 && pmu->flush_branch_stack) {
2524 pmu = cpuctx->ctx.pmu;
2526 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2528 perf_pmu_disable(pmu);
2530 pmu->flush_branch_stack();
2532 perf_pmu_enable(pmu);
2534 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2540 local_irq_restore(flags);
2544 * Called from scheduler to add the events of the current task
2545 * with interrupts disabled.
2547 * We restore the event value and then enable it.
2549 * This does not protect us against NMI, but enable()
2550 * sets the enabled bit in the control field of event _before_
2551 * accessing the event control register. If a NMI hits, then it will
2552 * keep the event running.
2554 void __perf_event_task_sched_in(struct task_struct *prev,
2555 struct task_struct *task)
2557 struct perf_event_context *ctx;
2560 for_each_task_context_nr(ctxn) {
2561 ctx = task->perf_event_ctxp[ctxn];
2565 perf_event_context_sched_in(ctx, task);
2568 * if cgroup events exist on this CPU, then we need
2569 * to check if we have to switch in PMU state.
2570 * cgroup event are system-wide mode only
2572 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2573 perf_cgroup_sched_in(prev, task);
2575 /* check for system-wide branch_stack events */
2576 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2577 perf_branch_stack_sched_in(prev, task);
2580 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2582 u64 frequency = event->attr.sample_freq;
2583 u64 sec = NSEC_PER_SEC;
2584 u64 divisor, dividend;
2586 int count_fls, nsec_fls, frequency_fls, sec_fls;
2588 count_fls = fls64(count);
2589 nsec_fls = fls64(nsec);
2590 frequency_fls = fls64(frequency);
2594 * We got @count in @nsec, with a target of sample_freq HZ
2595 * the target period becomes:
2598 * period = -------------------
2599 * @nsec * sample_freq
2604 * Reduce accuracy by one bit such that @a and @b converge
2605 * to a similar magnitude.
2607 #define REDUCE_FLS(a, b) \
2609 if (a##_fls > b##_fls) { \
2619 * Reduce accuracy until either term fits in a u64, then proceed with
2620 * the other, so that finally we can do a u64/u64 division.
2622 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2623 REDUCE_FLS(nsec, frequency);
2624 REDUCE_FLS(sec, count);
2627 if (count_fls + sec_fls > 64) {
2628 divisor = nsec * frequency;
2630 while (count_fls + sec_fls > 64) {
2631 REDUCE_FLS(count, sec);
2635 dividend = count * sec;
2637 dividend = count * sec;
2639 while (nsec_fls + frequency_fls > 64) {
2640 REDUCE_FLS(nsec, frequency);
2644 divisor = nsec * frequency;
2650 return div64_u64(dividend, divisor);
2653 static DEFINE_PER_CPU(int, perf_throttled_count);
2654 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2656 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2658 struct hw_perf_event *hwc = &event->hw;
2659 s64 period, sample_period;
2662 period = perf_calculate_period(event, nsec, count);
2664 delta = (s64)(period - hwc->sample_period);
2665 delta = (delta + 7) / 8; /* low pass filter */
2667 sample_period = hwc->sample_period + delta;
2672 hwc->sample_period = sample_period;
2674 if (local64_read(&hwc->period_left) > 8*sample_period) {
2676 event->pmu->stop(event, PERF_EF_UPDATE);
2678 local64_set(&hwc->period_left, 0);
2681 event->pmu->start(event, PERF_EF_RELOAD);
2686 * combine freq adjustment with unthrottling to avoid two passes over the
2687 * events. At the same time, make sure, having freq events does not change
2688 * the rate of unthrottling as that would introduce bias.
2690 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2693 struct perf_event *event;
2694 struct hw_perf_event *hwc;
2695 u64 now, period = TICK_NSEC;
2699 * only need to iterate over all events iff:
2700 * - context have events in frequency mode (needs freq adjust)
2701 * - there are events to unthrottle on this cpu
2703 if (!(ctx->nr_freq || needs_unthr))
2706 raw_spin_lock(&ctx->lock);
2707 perf_pmu_disable(ctx->pmu);
2709 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2710 if (event->state != PERF_EVENT_STATE_ACTIVE)
2713 if (!event_filter_match(event))
2718 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2719 hwc->interrupts = 0;
2720 perf_log_throttle(event, 1);
2721 event->pmu->start(event, 0);
2724 if (!event->attr.freq || !event->attr.sample_freq)
2728 * stop the event and update event->count
2730 event->pmu->stop(event, PERF_EF_UPDATE);
2732 now = local64_read(&event->count);
2733 delta = now - hwc->freq_count_stamp;
2734 hwc->freq_count_stamp = now;
2738 * reload only if value has changed
2739 * we have stopped the event so tell that
2740 * to perf_adjust_period() to avoid stopping it
2744 perf_adjust_period(event, period, delta, false);
2746 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2749 perf_pmu_enable(ctx->pmu);
2750 raw_spin_unlock(&ctx->lock);
2754 * Round-robin a context's events:
2756 static void rotate_ctx(struct perf_event_context *ctx)
2759 * Rotate the first entry last of non-pinned groups. Rotation might be
2760 * disabled by the inheritance code.
2762 if (!ctx->rotate_disable)
2763 list_rotate_left(&ctx->flexible_groups);
2767 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2768 * because they're strictly cpu affine and rotate_start is called with IRQs
2769 * disabled, while rotate_context is called from IRQ context.
2771 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2773 struct perf_event_context *ctx = NULL;
2774 int rotate = 0, remove = 1;
2776 if (cpuctx->ctx.nr_events) {
2778 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2782 ctx = cpuctx->task_ctx;
2783 if (ctx && ctx->nr_events) {
2785 if (ctx->nr_events != ctx->nr_active)
2792 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2793 perf_pmu_disable(cpuctx->ctx.pmu);
2795 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2797 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2799 rotate_ctx(&cpuctx->ctx);
2803 perf_event_sched_in(cpuctx, ctx, current);
2805 perf_pmu_enable(cpuctx->ctx.pmu);
2806 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2809 list_del_init(&cpuctx->rotation_list);
2814 #ifdef CONFIG_NO_HZ_FULL
2815 bool perf_event_can_stop_tick(void)
2817 if (list_empty(&__get_cpu_var(rotation_list)))
2824 void perf_event_task_tick(void)
2826 struct list_head *head = &__get_cpu_var(rotation_list);
2827 struct perf_cpu_context *cpuctx, *tmp;
2828 struct perf_event_context *ctx;
2831 WARN_ON(!irqs_disabled());
2833 __this_cpu_inc(perf_throttled_seq);
2834 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2836 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2838 perf_adjust_freq_unthr_context(ctx, throttled);
2840 ctx = cpuctx->task_ctx;
2842 perf_adjust_freq_unthr_context(ctx, throttled);
2846 static int event_enable_on_exec(struct perf_event *event,
2847 struct perf_event_context *ctx)
2849 if (!event->attr.enable_on_exec)
2852 event->attr.enable_on_exec = 0;
2853 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2856 __perf_event_mark_enabled(event);
2862 * Enable all of a task's events that have been marked enable-on-exec.
2863 * This expects task == current.
2865 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2867 struct perf_event *event;
2868 unsigned long flags;
2872 local_irq_save(flags);
2873 if (!ctx || !ctx->nr_events)
2877 * We must ctxsw out cgroup events to avoid conflict
2878 * when invoking perf_task_event_sched_in() later on
2879 * in this function. Otherwise we end up trying to
2880 * ctxswin cgroup events which are already scheduled
2883 perf_cgroup_sched_out(current, NULL);
2885 raw_spin_lock(&ctx->lock);
2886 task_ctx_sched_out(ctx);
2888 list_for_each_entry(event, &ctx->event_list, event_entry) {
2889 ret = event_enable_on_exec(event, ctx);
2895 * Unclone this context if we enabled any event.
2900 raw_spin_unlock(&ctx->lock);
2903 * Also calls ctxswin for cgroup events, if any:
2905 perf_event_context_sched_in(ctx, ctx->task);
2907 local_irq_restore(flags);
2911 * Cross CPU call to read the hardware event
2913 static void __perf_event_read(void *info)
2915 struct perf_event *event = info;
2916 struct perf_event_context *ctx = event->ctx;
2917 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2920 * If this is a task context, we need to check whether it is
2921 * the current task context of this cpu. If not it has been
2922 * scheduled out before the smp call arrived. In that case
2923 * event->count would have been updated to a recent sample
2924 * when the event was scheduled out.
2926 if (ctx->task && cpuctx->task_ctx != ctx)
2929 raw_spin_lock(&ctx->lock);
2930 if (ctx->is_active) {
2931 update_context_time(ctx);
2932 update_cgrp_time_from_event(event);
2934 update_event_times(event);
2935 if (event->state == PERF_EVENT_STATE_ACTIVE)
2936 event->pmu->read(event);
2937 raw_spin_unlock(&ctx->lock);
2940 static inline u64 perf_event_count(struct perf_event *event)
2942 return local64_read(&event->count) + atomic64_read(&event->child_count);
2945 static u64 perf_event_read(struct perf_event *event)
2948 * If event is enabled and currently active on a CPU, update the
2949 * value in the event structure:
2951 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2952 smp_call_function_single(event->oncpu,
2953 __perf_event_read, event, 1);
2954 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2955 struct perf_event_context *ctx = event->ctx;
2956 unsigned long flags;
2958 raw_spin_lock_irqsave(&ctx->lock, flags);
2960 * may read while context is not active
2961 * (e.g., thread is blocked), in that case
2962 * we cannot update context time
2964 if (ctx->is_active) {
2965 update_context_time(ctx);
2966 update_cgrp_time_from_event(event);
2968 update_event_times(event);
2969 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2972 return perf_event_count(event);
2976 * Initialize the perf_event context in a task_struct:
2978 static void __perf_event_init_context(struct perf_event_context *ctx)
2980 raw_spin_lock_init(&ctx->lock);
2981 mutex_init(&ctx->mutex);
2982 INIT_LIST_HEAD(&ctx->pinned_groups);
2983 INIT_LIST_HEAD(&ctx->flexible_groups);
2984 INIT_LIST_HEAD(&ctx->event_list);
2985 atomic_set(&ctx->refcount, 1);
2988 static struct perf_event_context *
2989 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2991 struct perf_event_context *ctx;
2993 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2997 __perf_event_init_context(ctx);
3000 get_task_struct(task);
3007 static struct task_struct *
3008 find_lively_task_by_vpid(pid_t vpid)
3010 struct task_struct *task;
3017 task = find_task_by_vpid(vpid);
3019 get_task_struct(task);
3023 return ERR_PTR(-ESRCH);
3025 /* Reuse ptrace permission checks for now. */
3027 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3032 put_task_struct(task);
3033 return ERR_PTR(err);
3038 * Returns a matching context with refcount and pincount.
3040 static struct perf_event_context *
3041 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3043 struct perf_event_context *ctx;
3044 struct perf_cpu_context *cpuctx;
3045 unsigned long flags;
3049 /* Must be root to operate on a CPU event: */
3050 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3051 return ERR_PTR(-EACCES);
3054 * We could be clever and allow to attach a event to an
3055 * offline CPU and activate it when the CPU comes up, but
3058 if (!cpu_online(cpu))
3059 return ERR_PTR(-ENODEV);
3061 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3070 ctxn = pmu->task_ctx_nr;
3075 ctx = perf_lock_task_context(task, ctxn, &flags);
3079 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3081 ctx = alloc_perf_context(pmu, task);
3087 mutex_lock(&task->perf_event_mutex);
3089 * If it has already passed perf_event_exit_task().
3090 * we must see PF_EXITING, it takes this mutex too.
3092 if (task->flags & PF_EXITING)
3094 else if (task->perf_event_ctxp[ctxn])
3099 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3101 mutex_unlock(&task->perf_event_mutex);
3103 if (unlikely(err)) {
3115 return ERR_PTR(err);
3118 static void perf_event_free_filter(struct perf_event *event);
3120 static void free_event_rcu(struct rcu_head *head)
3122 struct perf_event *event;
3124 event = container_of(head, struct perf_event, rcu_head);
3126 put_pid_ns(event->ns);
3127 perf_event_free_filter(event);
3131 static void ring_buffer_put(struct ring_buffer *rb);
3132 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3134 static void free_event(struct perf_event *event)
3136 irq_work_sync(&event->pending);
3138 if (!event->parent) {
3139 if (event->attach_state & PERF_ATTACH_TASK)
3140 static_key_slow_dec_deferred(&perf_sched_events);
3141 if (event->attr.mmap || event->attr.mmap_data)
3142 atomic_dec(&nr_mmap_events);
3143 if (event->attr.comm)
3144 atomic_dec(&nr_comm_events);
3145 if (event->attr.task)
3146 atomic_dec(&nr_task_events);
3147 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3148 put_callchain_buffers();
3149 if (is_cgroup_event(event)) {
3150 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
3151 static_key_slow_dec_deferred(&perf_sched_events);
3154 if (has_branch_stack(event)) {
3155 static_key_slow_dec_deferred(&perf_sched_events);
3156 /* is system-wide event */
3157 if (!(event->attach_state & PERF_ATTACH_TASK)) {
3158 atomic_dec(&per_cpu(perf_branch_stack_events,
3165 struct ring_buffer *rb;
3168 * Can happen when we close an event with re-directed output.
3170 * Since we have a 0 refcount, perf_mmap_close() will skip
3171 * over us; possibly making our ring_buffer_put() the last.
3173 mutex_lock(&event->mmap_mutex);
3176 rcu_assign_pointer(event->rb, NULL);
3177 ring_buffer_detach(event, rb);
3178 ring_buffer_put(rb); /* could be last */
3180 mutex_unlock(&event->mmap_mutex);
3183 if (is_cgroup_event(event))
3184 perf_detach_cgroup(event);
3187 event->destroy(event);
3190 put_ctx(event->ctx);
3192 call_rcu(&event->rcu_head, free_event_rcu);
3195 int perf_event_release_kernel(struct perf_event *event)
3197 struct perf_event_context *ctx = event->ctx;
3199 WARN_ON_ONCE(ctx->parent_ctx);
3201 * There are two ways this annotation is useful:
3203 * 1) there is a lock recursion from perf_event_exit_task
3204 * see the comment there.
3206 * 2) there is a lock-inversion with mmap_sem through
3207 * perf_event_read_group(), which takes faults while
3208 * holding ctx->mutex, however this is called after
3209 * the last filedesc died, so there is no possibility
3210 * to trigger the AB-BA case.
3212 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3213 raw_spin_lock_irq(&ctx->lock);
3214 perf_group_detach(event);
3215 raw_spin_unlock_irq(&ctx->lock);
3216 perf_remove_from_context(event);
3217 mutex_unlock(&ctx->mutex);
3223 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3226 * Called when the last reference to the file is gone.
3228 static void put_event(struct perf_event *event)
3230 struct task_struct *owner;
3232 if (!atomic_long_dec_and_test(&event->refcount))
3236 owner = ACCESS_ONCE(event->owner);
3238 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3239 * !owner it means the list deletion is complete and we can indeed
3240 * free this event, otherwise we need to serialize on
3241 * owner->perf_event_mutex.
3243 smp_read_barrier_depends();
3246 * Since delayed_put_task_struct() also drops the last
3247 * task reference we can safely take a new reference
3248 * while holding the rcu_read_lock().
3250 get_task_struct(owner);
3255 mutex_lock(&owner->perf_event_mutex);
3257 * We have to re-check the event->owner field, if it is cleared
3258 * we raced with perf_event_exit_task(), acquiring the mutex
3259 * ensured they're done, and we can proceed with freeing the
3263 list_del_init(&event->owner_entry);
3264 mutex_unlock(&owner->perf_event_mutex);
3265 put_task_struct(owner);
3268 perf_event_release_kernel(event);
3271 static int perf_release(struct inode *inode, struct file *file)
3273 put_event(file->private_data);
3277 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3279 struct perf_event *child;
3285 mutex_lock(&event->child_mutex);
3286 total += perf_event_read(event);
3287 *enabled += event->total_time_enabled +
3288 atomic64_read(&event->child_total_time_enabled);
3289 *running += event->total_time_running +
3290 atomic64_read(&event->child_total_time_running);
3292 list_for_each_entry(child, &event->child_list, child_list) {
3293 total += perf_event_read(child);
3294 *enabled += child->total_time_enabled;
3295 *running += child->total_time_running;
3297 mutex_unlock(&event->child_mutex);
3301 EXPORT_SYMBOL_GPL(perf_event_read_value);
3303 static int perf_event_read_group(struct perf_event *event,
3304 u64 read_format, char __user *buf)
3306 struct perf_event *leader = event->group_leader, *sub;
3307 int n = 0, size = 0, ret = -EFAULT;
3308 struct perf_event_context *ctx = leader->ctx;
3310 u64 count, enabled, running;
3312 mutex_lock(&ctx->mutex);
3313 count = perf_event_read_value(leader, &enabled, &running);
3315 values[n++] = 1 + leader->nr_siblings;
3316 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3317 values[n++] = enabled;
3318 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3319 values[n++] = running;
3320 values[n++] = count;
3321 if (read_format & PERF_FORMAT_ID)
3322 values[n++] = primary_event_id(leader);
3324 size = n * sizeof(u64);
3326 if (copy_to_user(buf, values, size))
3331 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3334 values[n++] = perf_event_read_value(sub, &enabled, &running);
3335 if (read_format & PERF_FORMAT_ID)
3336 values[n++] = primary_event_id(sub);
3338 size = n * sizeof(u64);
3340 if (copy_to_user(buf + ret, values, size)) {
3348 mutex_unlock(&ctx->mutex);
3353 static int perf_event_read_one(struct perf_event *event,
3354 u64 read_format, char __user *buf)
3356 u64 enabled, running;
3360 values[n++] = perf_event_read_value(event, &enabled, &running);
3361 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3362 values[n++] = enabled;
3363 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3364 values[n++] = running;
3365 if (read_format & PERF_FORMAT_ID)
3366 values[n++] = primary_event_id(event);
3368 if (copy_to_user(buf, values, n * sizeof(u64)))
3371 return n * sizeof(u64);
3375 * Read the performance event - simple non blocking version for now
3378 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3380 u64 read_format = event->attr.read_format;
3384 * Return end-of-file for a read on a event that is in
3385 * error state (i.e. because it was pinned but it couldn't be
3386 * scheduled on to the CPU at some point).
3388 if (event->state == PERF_EVENT_STATE_ERROR)
3391 if (count < event->read_size)
3394 WARN_ON_ONCE(event->ctx->parent_ctx);
3395 if (read_format & PERF_FORMAT_GROUP)
3396 ret = perf_event_read_group(event, read_format, buf);
3398 ret = perf_event_read_one(event, read_format, buf);
3404 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3406 struct perf_event *event = file->private_data;
3408 return perf_read_hw(event, buf, count);
3411 static unsigned int perf_poll(struct file *file, poll_table *wait)
3413 struct perf_event *event = file->private_data;
3414 struct ring_buffer *rb;
3415 unsigned int events = POLL_HUP;
3418 * Pin the event->rb by taking event->mmap_mutex; otherwise
3419 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3421 mutex_lock(&event->mmap_mutex);
3424 events = atomic_xchg(&rb->poll, 0);
3425 mutex_unlock(&event->mmap_mutex);
3427 poll_wait(file, &event->waitq, wait);
3432 static void perf_event_reset(struct perf_event *event)
3434 (void)perf_event_read(event);
3435 local64_set(&event->count, 0);
3436 perf_event_update_userpage(event);
3440 * Holding the top-level event's child_mutex means that any
3441 * descendant process that has inherited this event will block
3442 * in sync_child_event if it goes to exit, thus satisfying the
3443 * task existence requirements of perf_event_enable/disable.
3445 static void perf_event_for_each_child(struct perf_event *event,
3446 void (*func)(struct perf_event *))
3448 struct perf_event *child;
3450 WARN_ON_ONCE(event->ctx->parent_ctx);
3451 mutex_lock(&event->child_mutex);
3453 list_for_each_entry(child, &event->child_list, child_list)
3455 mutex_unlock(&event->child_mutex);
3458 static void perf_event_for_each(struct perf_event *event,
3459 void (*func)(struct perf_event *))
3461 struct perf_event_context *ctx = event->ctx;
3462 struct perf_event *sibling;
3464 WARN_ON_ONCE(ctx->parent_ctx);
3465 mutex_lock(&ctx->mutex);
3466 event = event->group_leader;
3468 perf_event_for_each_child(event, func);
3469 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3470 perf_event_for_each_child(sibling, func);
3471 mutex_unlock(&ctx->mutex);
3474 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3476 struct perf_event_context *ctx = event->ctx;
3480 if (!is_sampling_event(event))
3483 if (copy_from_user(&value, arg, sizeof(value)))
3489 raw_spin_lock_irq(&ctx->lock);
3490 if (event->attr.freq) {
3491 if (value > sysctl_perf_event_sample_rate) {
3496 event->attr.sample_freq = value;
3498 event->attr.sample_period = value;
3499 event->hw.sample_period = value;
3502 raw_spin_unlock_irq(&ctx->lock);
3507 static const struct file_operations perf_fops;
3509 static inline int perf_fget_light(int fd, struct fd *p)
3511 struct fd f = fdget(fd);
3515 if (f.file->f_op != &perf_fops) {
3523 static int perf_event_set_output(struct perf_event *event,
3524 struct perf_event *output_event);
3525 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3527 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3529 struct perf_event *event = file->private_data;
3530 void (*func)(struct perf_event *);
3534 case PERF_EVENT_IOC_ENABLE:
3535 func = perf_event_enable;
3537 case PERF_EVENT_IOC_DISABLE:
3538 func = perf_event_disable;
3540 case PERF_EVENT_IOC_RESET:
3541 func = perf_event_reset;
3544 case PERF_EVENT_IOC_REFRESH:
3545 return perf_event_refresh(event, arg);
3547 case PERF_EVENT_IOC_PERIOD:
3548 return perf_event_period(event, (u64 __user *)arg);
3550 case PERF_EVENT_IOC_SET_OUTPUT:
3554 struct perf_event *output_event;
3556 ret = perf_fget_light(arg, &output);
3559 output_event = output.file->private_data;
3560 ret = perf_event_set_output(event, output_event);
3563 ret = perf_event_set_output(event, NULL);
3568 case PERF_EVENT_IOC_SET_FILTER:
3569 return perf_event_set_filter(event, (void __user *)arg);
3575 if (flags & PERF_IOC_FLAG_GROUP)
3576 perf_event_for_each(event, func);
3578 perf_event_for_each_child(event, func);
3583 int perf_event_task_enable(void)
3585 struct perf_event *event;
3587 mutex_lock(¤t->perf_event_mutex);
3588 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3589 perf_event_for_each_child(event, perf_event_enable);
3590 mutex_unlock(¤t->perf_event_mutex);
3595 int perf_event_task_disable(void)
3597 struct perf_event *event;
3599 mutex_lock(¤t->perf_event_mutex);
3600 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3601 perf_event_for_each_child(event, perf_event_disable);
3602 mutex_unlock(¤t->perf_event_mutex);
3607 static int perf_event_index(struct perf_event *event)
3609 if (event->hw.state & PERF_HES_STOPPED)
3612 if (event->state != PERF_EVENT_STATE_ACTIVE)
3615 return event->pmu->event_idx(event);
3618 static void calc_timer_values(struct perf_event *event,
3625 *now = perf_clock();
3626 ctx_time = event->shadow_ctx_time + *now;
3627 *enabled = ctx_time - event->tstamp_enabled;
3628 *running = ctx_time - event->tstamp_running;
3631 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3636 * Callers need to ensure there can be no nesting of this function, otherwise
3637 * the seqlock logic goes bad. We can not serialize this because the arch
3638 * code calls this from NMI context.
3640 void perf_event_update_userpage(struct perf_event *event)
3642 struct perf_event_mmap_page *userpg;
3643 struct ring_buffer *rb;
3644 u64 enabled, running, now;
3648 * compute total_time_enabled, total_time_running
3649 * based on snapshot values taken when the event
3650 * was last scheduled in.
3652 * we cannot simply called update_context_time()
3653 * because of locking issue as we can be called in
3656 calc_timer_values(event, &now, &enabled, &running);
3657 rb = rcu_dereference(event->rb);
3661 userpg = rb->user_page;
3664 * Disable preemption so as to not let the corresponding user-space
3665 * spin too long if we get preempted.
3670 userpg->index = perf_event_index(event);
3671 userpg->offset = perf_event_count(event);
3673 userpg->offset -= local64_read(&event->hw.prev_count);
3675 userpg->time_enabled = enabled +
3676 atomic64_read(&event->child_total_time_enabled);
3678 userpg->time_running = running +
3679 atomic64_read(&event->child_total_time_running);
3681 arch_perf_update_userpage(userpg, now);
3690 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3692 struct perf_event *event = vma->vm_file->private_data;
3693 struct ring_buffer *rb;
3694 int ret = VM_FAULT_SIGBUS;
3696 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3697 if (vmf->pgoff == 0)
3703 rb = rcu_dereference(event->rb);
3707 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3710 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3714 get_page(vmf->page);
3715 vmf->page->mapping = vma->vm_file->f_mapping;
3716 vmf->page->index = vmf->pgoff;
3725 static void ring_buffer_attach(struct perf_event *event,
3726 struct ring_buffer *rb)
3728 unsigned long flags;
3730 if (!list_empty(&event->rb_entry))
3733 spin_lock_irqsave(&rb->event_lock, flags);
3734 if (list_empty(&event->rb_entry))
3735 list_add(&event->rb_entry, &rb->event_list);
3736 spin_unlock_irqrestore(&rb->event_lock, flags);
3739 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3741 unsigned long flags;
3743 if (list_empty(&event->rb_entry))
3746 spin_lock_irqsave(&rb->event_lock, flags);
3747 list_del_init(&event->rb_entry);
3748 wake_up_all(&event->waitq);
3749 spin_unlock_irqrestore(&rb->event_lock, flags);
3752 static void ring_buffer_wakeup(struct perf_event *event)
3754 struct ring_buffer *rb;
3757 rb = rcu_dereference(event->rb);
3759 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3760 wake_up_all(&event->waitq);
3765 static void rb_free_rcu(struct rcu_head *rcu_head)
3767 struct ring_buffer *rb;
3769 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3773 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3775 struct ring_buffer *rb;
3778 rb = rcu_dereference(event->rb);
3780 if (!atomic_inc_not_zero(&rb->refcount))
3788 static void ring_buffer_put(struct ring_buffer *rb)
3790 if (!atomic_dec_and_test(&rb->refcount))
3793 WARN_ON_ONCE(!list_empty(&rb->event_list));
3795 call_rcu(&rb->rcu_head, rb_free_rcu);
3798 static void perf_mmap_open(struct vm_area_struct *vma)
3800 struct perf_event *event = vma->vm_file->private_data;
3802 atomic_inc(&event->mmap_count);
3803 atomic_inc(&event->rb->mmap_count);
3807 * A buffer can be mmap()ed multiple times; either directly through the same
3808 * event, or through other events by use of perf_event_set_output().
3810 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3811 * the buffer here, where we still have a VM context. This means we need
3812 * to detach all events redirecting to us.
3814 static void perf_mmap_close(struct vm_area_struct *vma)
3816 struct perf_event *event = vma->vm_file->private_data;
3818 struct ring_buffer *rb = event->rb;
3819 struct user_struct *mmap_user = rb->mmap_user;
3820 int mmap_locked = rb->mmap_locked;
3821 unsigned long size = perf_data_size(rb);
3823 atomic_dec(&rb->mmap_count);
3825 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3828 /* Detach current event from the buffer. */
3829 rcu_assign_pointer(event->rb, NULL);
3830 ring_buffer_detach(event, rb);
3831 mutex_unlock(&event->mmap_mutex);
3833 /* If there's still other mmap()s of this buffer, we're done. */
3834 if (atomic_read(&rb->mmap_count)) {
3835 ring_buffer_put(rb); /* can't be last */
3840 * No other mmap()s, detach from all other events that might redirect
3841 * into the now unreachable buffer. Somewhat complicated by the
3842 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3846 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3847 if (!atomic_long_inc_not_zero(&event->refcount)) {
3849 * This event is en-route to free_event() which will
3850 * detach it and remove it from the list.
3856 mutex_lock(&event->mmap_mutex);
3858 * Check we didn't race with perf_event_set_output() which can
3859 * swizzle the rb from under us while we were waiting to
3860 * acquire mmap_mutex.
3862 * If we find a different rb; ignore this event, a next
3863 * iteration will no longer find it on the list. We have to
3864 * still restart the iteration to make sure we're not now
3865 * iterating the wrong list.
3867 if (event->rb == rb) {
3868 rcu_assign_pointer(event->rb, NULL);
3869 ring_buffer_detach(event, rb);
3870 ring_buffer_put(rb); /* can't be last, we still have one */
3872 mutex_unlock(&event->mmap_mutex);
3876 * Restart the iteration; either we're on the wrong list or
3877 * destroyed its integrity by doing a deletion.
3884 * It could be there's still a few 0-ref events on the list; they'll
3885 * get cleaned up by free_event() -- they'll also still have their
3886 * ref on the rb and will free it whenever they are done with it.
3888 * Aside from that, this buffer is 'fully' detached and unmapped,
3889 * undo the VM accounting.
3892 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3893 vma->vm_mm->pinned_vm -= mmap_locked;
3894 free_uid(mmap_user);
3896 ring_buffer_put(rb); /* could be last */
3899 static const struct vm_operations_struct perf_mmap_vmops = {
3900 .open = perf_mmap_open,
3901 .close = perf_mmap_close,
3902 .fault = perf_mmap_fault,
3903 .page_mkwrite = perf_mmap_fault,
3906 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3908 struct perf_event *event = file->private_data;
3909 unsigned long user_locked, user_lock_limit;
3910 struct user_struct *user = current_user();
3911 unsigned long locked, lock_limit;
3912 struct ring_buffer *rb;
3913 unsigned long vma_size;
3914 unsigned long nr_pages;
3915 long user_extra, extra;
3916 int ret = 0, flags = 0;
3919 * Don't allow mmap() of inherited per-task counters. This would
3920 * create a performance issue due to all children writing to the
3923 if (event->cpu == -1 && event->attr.inherit)
3926 if (!(vma->vm_flags & VM_SHARED))
3929 vma_size = vma->vm_end - vma->vm_start;
3930 nr_pages = (vma_size / PAGE_SIZE) - 1;
3933 * If we have rb pages ensure they're a power-of-two number, so we
3934 * can do bitmasks instead of modulo.
3936 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3939 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3942 if (vma->vm_pgoff != 0)
3945 WARN_ON_ONCE(event->ctx->parent_ctx);
3947 mutex_lock(&event->mmap_mutex);
3949 if (event->rb->nr_pages != nr_pages) {
3954 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
3956 * Raced against perf_mmap_close() through
3957 * perf_event_set_output(). Try again, hope for better
3960 mutex_unlock(&event->mmap_mutex);
3967 user_extra = nr_pages + 1;
3968 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3971 * Increase the limit linearly with more CPUs:
3973 user_lock_limit *= num_online_cpus();
3975 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3978 if (user_locked > user_lock_limit)
3979 extra = user_locked - user_lock_limit;
3981 lock_limit = rlimit(RLIMIT_MEMLOCK);
3982 lock_limit >>= PAGE_SHIFT;
3983 locked = vma->vm_mm->pinned_vm + extra;
3985 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3986 !capable(CAP_IPC_LOCK)) {
3993 if (vma->vm_flags & VM_WRITE)
3994 flags |= RING_BUFFER_WRITABLE;
3996 rb = rb_alloc(nr_pages,
3997 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4005 atomic_set(&rb->mmap_count, 1);
4006 rb->mmap_locked = extra;
4007 rb->mmap_user = get_current_user();
4009 atomic_long_add(user_extra, &user->locked_vm);
4010 vma->vm_mm->pinned_vm += extra;
4012 ring_buffer_attach(event, rb);
4013 rcu_assign_pointer(event->rb, rb);
4015 perf_event_update_userpage(event);
4019 atomic_inc(&event->mmap_count);
4020 mutex_unlock(&event->mmap_mutex);
4023 * Since pinned accounting is per vm we cannot allow fork() to copy our
4026 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4027 vma->vm_ops = &perf_mmap_vmops;
4032 static int perf_fasync(int fd, struct file *filp, int on)
4034 struct inode *inode = file_inode(filp);
4035 struct perf_event *event = filp->private_data;
4038 mutex_lock(&inode->i_mutex);
4039 retval = fasync_helper(fd, filp, on, &event->fasync);
4040 mutex_unlock(&inode->i_mutex);
4048 static const struct file_operations perf_fops = {
4049 .llseek = no_llseek,
4050 .release = perf_release,
4053 .unlocked_ioctl = perf_ioctl,
4054 .compat_ioctl = perf_ioctl,
4056 .fasync = perf_fasync,
4062 * If there's data, ensure we set the poll() state and publish everything
4063 * to user-space before waking everybody up.
4066 void perf_event_wakeup(struct perf_event *event)
4068 ring_buffer_wakeup(event);
4070 if (event->pending_kill) {
4071 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4072 event->pending_kill = 0;
4076 static void perf_pending_event(struct irq_work *entry)
4078 struct perf_event *event = container_of(entry,
4079 struct perf_event, pending);
4081 if (event->pending_disable) {
4082 event->pending_disable = 0;
4083 __perf_event_disable(event);
4086 if (event->pending_wakeup) {
4087 event->pending_wakeup = 0;
4088 perf_event_wakeup(event);
4093 * We assume there is only KVM supporting the callbacks.
4094 * Later on, we might change it to a list if there is
4095 * another virtualization implementation supporting the callbacks.
4097 struct perf_guest_info_callbacks *perf_guest_cbs;
4099 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4101 perf_guest_cbs = cbs;
4104 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4106 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4108 perf_guest_cbs = NULL;
4111 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4114 perf_output_sample_regs(struct perf_output_handle *handle,
4115 struct pt_regs *regs, u64 mask)
4119 for_each_set_bit(bit, (const unsigned long *) &mask,
4120 sizeof(mask) * BITS_PER_BYTE) {
4123 val = perf_reg_value(regs, bit);
4124 perf_output_put(handle, val);
4128 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4129 struct pt_regs *regs)
4131 if (!user_mode(regs)) {
4133 regs = task_pt_regs(current);
4139 regs_user->regs = regs;
4140 regs_user->abi = perf_reg_abi(current);
4145 * Get remaining task size from user stack pointer.
4147 * It'd be better to take stack vma map and limit this more
4148 * precisly, but there's no way to get it safely under interrupt,
4149 * so using TASK_SIZE as limit.
4151 static u64 perf_ustack_task_size(struct pt_regs *regs)
4153 unsigned long addr = perf_user_stack_pointer(regs);
4155 if (!addr || addr >= TASK_SIZE)
4158 return TASK_SIZE - addr;
4162 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4163 struct pt_regs *regs)
4167 /* No regs, no stack pointer, no dump. */
4172 * Check if we fit in with the requested stack size into the:
4174 * If we don't, we limit the size to the TASK_SIZE.
4176 * - remaining sample size
4177 * If we don't, we customize the stack size to
4178 * fit in to the remaining sample size.
4181 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4182 stack_size = min(stack_size, (u16) task_size);
4184 /* Current header size plus static size and dynamic size. */
4185 header_size += 2 * sizeof(u64);
4187 /* Do we fit in with the current stack dump size? */
4188 if ((u16) (header_size + stack_size) < header_size) {
4190 * If we overflow the maximum size for the sample,
4191 * we customize the stack dump size to fit in.
4193 stack_size = USHRT_MAX - header_size - sizeof(u64);
4194 stack_size = round_up(stack_size, sizeof(u64));
4201 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4202 struct pt_regs *regs)
4204 /* Case of a kernel thread, nothing to dump */
4207 perf_output_put(handle, size);
4216 * - the size requested by user or the best one we can fit
4217 * in to the sample max size
4219 * - user stack dump data
4221 * - the actual dumped size
4225 perf_output_put(handle, dump_size);
4228 sp = perf_user_stack_pointer(regs);
4229 rem = __output_copy_user(handle, (void *) sp, dump_size);
4230 dyn_size = dump_size - rem;
4232 perf_output_skip(handle, rem);
4235 perf_output_put(handle, dyn_size);
4239 static void __perf_event_header__init_id(struct perf_event_header *header,
4240 struct perf_sample_data *data,
4241 struct perf_event *event)
4243 u64 sample_type = event->attr.sample_type;
4245 data->type = sample_type;
4246 header->size += event->id_header_size;
4248 if (sample_type & PERF_SAMPLE_TID) {
4249 /* namespace issues */
4250 data->tid_entry.pid = perf_event_pid(event, current);
4251 data->tid_entry.tid = perf_event_tid(event, current);
4254 if (sample_type & PERF_SAMPLE_TIME)
4255 data->time = perf_clock();
4257 if (sample_type & PERF_SAMPLE_ID)
4258 data->id = primary_event_id(event);
4260 if (sample_type & PERF_SAMPLE_STREAM_ID)
4261 data->stream_id = event->id;
4263 if (sample_type & PERF_SAMPLE_CPU) {
4264 data->cpu_entry.cpu = raw_smp_processor_id();
4265 data->cpu_entry.reserved = 0;
4269 void perf_event_header__init_id(struct perf_event_header *header,
4270 struct perf_sample_data *data,
4271 struct perf_event *event)
4273 if (event->attr.sample_id_all)
4274 __perf_event_header__init_id(header, data, event);
4277 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4278 struct perf_sample_data *data)
4280 u64 sample_type = data->type;
4282 if (sample_type & PERF_SAMPLE_TID)
4283 perf_output_put(handle, data->tid_entry);
4285 if (sample_type & PERF_SAMPLE_TIME)
4286 perf_output_put(handle, data->time);
4288 if (sample_type & PERF_SAMPLE_ID)
4289 perf_output_put(handle, data->id);
4291 if (sample_type & PERF_SAMPLE_STREAM_ID)
4292 perf_output_put(handle, data->stream_id);
4294 if (sample_type & PERF_SAMPLE_CPU)
4295 perf_output_put(handle, data->cpu_entry);
4298 void perf_event__output_id_sample(struct perf_event *event,
4299 struct perf_output_handle *handle,
4300 struct perf_sample_data *sample)
4302 if (event->attr.sample_id_all)
4303 __perf_event__output_id_sample(handle, sample);
4306 static void perf_output_read_one(struct perf_output_handle *handle,
4307 struct perf_event *event,
4308 u64 enabled, u64 running)
4310 u64 read_format = event->attr.read_format;
4314 values[n++] = perf_event_count(event);
4315 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4316 values[n++] = enabled +
4317 atomic64_read(&event->child_total_time_enabled);
4319 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4320 values[n++] = running +
4321 atomic64_read(&event->child_total_time_running);
4323 if (read_format & PERF_FORMAT_ID)
4324 values[n++] = primary_event_id(event);
4326 __output_copy(handle, values, n * sizeof(u64));
4330 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4332 static void perf_output_read_group(struct perf_output_handle *handle,
4333 struct perf_event *event,
4334 u64 enabled, u64 running)
4336 struct perf_event *leader = event->group_leader, *sub;
4337 u64 read_format = event->attr.read_format;
4341 values[n++] = 1 + leader->nr_siblings;
4343 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4344 values[n++] = enabled;
4346 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4347 values[n++] = running;
4349 if (leader != event)
4350 leader->pmu->read(leader);
4352 values[n++] = perf_event_count(leader);
4353 if (read_format & PERF_FORMAT_ID)
4354 values[n++] = primary_event_id(leader);
4356 __output_copy(handle, values, n * sizeof(u64));
4358 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4362 sub->pmu->read(sub);
4364 values[n++] = perf_event_count(sub);
4365 if (read_format & PERF_FORMAT_ID)
4366 values[n++] = primary_event_id(sub);
4368 __output_copy(handle, values, n * sizeof(u64));
4372 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4373 PERF_FORMAT_TOTAL_TIME_RUNNING)
4375 static void perf_output_read(struct perf_output_handle *handle,
4376 struct perf_event *event)
4378 u64 enabled = 0, running = 0, now;
4379 u64 read_format = event->attr.read_format;
4382 * compute total_time_enabled, total_time_running
4383 * based on snapshot values taken when the event
4384 * was last scheduled in.
4386 * we cannot simply called update_context_time()
4387 * because of locking issue as we are called in
4390 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4391 calc_timer_values(event, &now, &enabled, &running);
4393 if (event->attr.read_format & PERF_FORMAT_GROUP)
4394 perf_output_read_group(handle, event, enabled, running);
4396 perf_output_read_one(handle, event, enabled, running);
4399 void perf_output_sample(struct perf_output_handle *handle,
4400 struct perf_event_header *header,
4401 struct perf_sample_data *data,
4402 struct perf_event *event)
4404 u64 sample_type = data->type;
4406 perf_output_put(handle, *header);
4408 if (sample_type & PERF_SAMPLE_IP)
4409 perf_output_put(handle, data->ip);
4411 if (sample_type & PERF_SAMPLE_TID)
4412 perf_output_put(handle, data->tid_entry);
4414 if (sample_type & PERF_SAMPLE_TIME)
4415 perf_output_put(handle, data->time);
4417 if (sample_type & PERF_SAMPLE_ADDR)
4418 perf_output_put(handle, data->addr);
4420 if (sample_type & PERF_SAMPLE_ID)
4421 perf_output_put(handle, data->id);
4423 if (sample_type & PERF_SAMPLE_STREAM_ID)
4424 perf_output_put(handle, data->stream_id);
4426 if (sample_type & PERF_SAMPLE_CPU)
4427 perf_output_put(handle, data->cpu_entry);
4429 if (sample_type & PERF_SAMPLE_PERIOD)
4430 perf_output_put(handle, data->period);
4432 if (sample_type & PERF_SAMPLE_READ)
4433 perf_output_read(handle, event);
4435 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4436 if (data->callchain) {
4439 if (data->callchain)
4440 size += data->callchain->nr;
4442 size *= sizeof(u64);
4444 __output_copy(handle, data->callchain, size);
4447 perf_output_put(handle, nr);
4451 if (sample_type & PERF_SAMPLE_RAW) {
4453 perf_output_put(handle, data->raw->size);
4454 __output_copy(handle, data->raw->data,
4461 .size = sizeof(u32),
4464 perf_output_put(handle, raw);
4468 if (!event->attr.watermark) {
4469 int wakeup_events = event->attr.wakeup_events;
4471 if (wakeup_events) {
4472 struct ring_buffer *rb = handle->rb;
4473 int events = local_inc_return(&rb->events);
4475 if (events >= wakeup_events) {
4476 local_sub(wakeup_events, &rb->events);
4477 local_inc(&rb->wakeup);
4482 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4483 if (data->br_stack) {
4486 size = data->br_stack->nr
4487 * sizeof(struct perf_branch_entry);
4489 perf_output_put(handle, data->br_stack->nr);
4490 perf_output_copy(handle, data->br_stack->entries, size);
4493 * we always store at least the value of nr
4496 perf_output_put(handle, nr);
4500 if (sample_type & PERF_SAMPLE_REGS_USER) {
4501 u64 abi = data->regs_user.abi;
4504 * If there are no regs to dump, notice it through
4505 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4507 perf_output_put(handle, abi);
4510 u64 mask = event->attr.sample_regs_user;
4511 perf_output_sample_regs(handle,
4512 data->regs_user.regs,
4517 if (sample_type & PERF_SAMPLE_STACK_USER)
4518 perf_output_sample_ustack(handle,
4519 data->stack_user_size,
4520 data->regs_user.regs);
4522 if (sample_type & PERF_SAMPLE_WEIGHT)
4523 perf_output_put(handle, data->weight);
4525 if (sample_type & PERF_SAMPLE_DATA_SRC)
4526 perf_output_put(handle, data->data_src.val);
4529 void perf_prepare_sample(struct perf_event_header *header,
4530 struct perf_sample_data *data,
4531 struct perf_event *event,
4532 struct pt_regs *regs)
4534 u64 sample_type = event->attr.sample_type;
4536 header->type = PERF_RECORD_SAMPLE;
4537 header->size = sizeof(*header) + event->header_size;
4540 header->misc |= perf_misc_flags(regs);
4542 __perf_event_header__init_id(header, data, event);
4544 if (sample_type & PERF_SAMPLE_IP)
4545 data->ip = perf_instruction_pointer(regs);
4547 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4550 data->callchain = perf_callchain(event, regs);
4552 if (data->callchain)
4553 size += data->callchain->nr;
4555 header->size += size * sizeof(u64);
4558 if (sample_type & PERF_SAMPLE_RAW) {
4559 int size = sizeof(u32);
4562 size += data->raw->size;
4564 size += sizeof(u32);
4566 WARN_ON_ONCE(size & (sizeof(u64)-1));
4567 header->size += size;
4570 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4571 int size = sizeof(u64); /* nr */
4572 if (data->br_stack) {
4573 size += data->br_stack->nr
4574 * sizeof(struct perf_branch_entry);
4576 header->size += size;
4579 if (sample_type & PERF_SAMPLE_REGS_USER) {
4580 /* regs dump ABI info */
4581 int size = sizeof(u64);
4583 perf_sample_regs_user(&data->regs_user, regs);
4585 if (data->regs_user.regs) {
4586 u64 mask = event->attr.sample_regs_user;
4587 size += hweight64(mask) * sizeof(u64);
4590 header->size += size;
4593 if (sample_type & PERF_SAMPLE_STACK_USER) {
4595 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4596 * processed as the last one or have additional check added
4597 * in case new sample type is added, because we could eat
4598 * up the rest of the sample size.
4600 struct perf_regs_user *uregs = &data->regs_user;
4601 u16 stack_size = event->attr.sample_stack_user;
4602 u16 size = sizeof(u64);
4605 perf_sample_regs_user(uregs, regs);
4607 stack_size = perf_sample_ustack_size(stack_size, header->size,
4611 * If there is something to dump, add space for the dump
4612 * itself and for the field that tells the dynamic size,
4613 * which is how many have been actually dumped.
4616 size += sizeof(u64) + stack_size;
4618 data->stack_user_size = stack_size;
4619 header->size += size;
4623 static void perf_event_output(struct perf_event *event,
4624 struct perf_sample_data *data,
4625 struct pt_regs *regs)
4627 struct perf_output_handle handle;
4628 struct perf_event_header header;
4630 /* protect the callchain buffers */
4633 perf_prepare_sample(&header, data, event, regs);
4635 if (perf_output_begin(&handle, event, header.size))
4638 perf_output_sample(&handle, &header, data, event);
4640 perf_output_end(&handle);
4650 struct perf_read_event {
4651 struct perf_event_header header;
4658 perf_event_read_event(struct perf_event *event,
4659 struct task_struct *task)
4661 struct perf_output_handle handle;
4662 struct perf_sample_data sample;
4663 struct perf_read_event read_event = {
4665 .type = PERF_RECORD_READ,
4667 .size = sizeof(read_event) + event->read_size,
4669 .pid = perf_event_pid(event, task),
4670 .tid = perf_event_tid(event, task),
4674 perf_event_header__init_id(&read_event.header, &sample, event);
4675 ret = perf_output_begin(&handle, event, read_event.header.size);
4679 perf_output_put(&handle, read_event);
4680 perf_output_read(&handle, event);
4681 perf_event__output_id_sample(event, &handle, &sample);
4683 perf_output_end(&handle);
4686 typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data);
4687 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4690 perf_event_aux_ctx(struct perf_event_context *ctx,
4691 perf_event_aux_match_cb match,
4692 perf_event_aux_output_cb output,
4695 struct perf_event *event;
4697 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4698 if (event->state < PERF_EVENT_STATE_INACTIVE)
4700 if (!event_filter_match(event))
4702 if (match(event, data))
4703 output(event, data);
4708 perf_event_aux(perf_event_aux_match_cb match,
4709 perf_event_aux_output_cb output,
4711 struct perf_event_context *task_ctx)
4713 struct perf_cpu_context *cpuctx;
4714 struct perf_event_context *ctx;
4719 list_for_each_entry_rcu(pmu, &pmus, entry) {
4720 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4721 if (cpuctx->unique_pmu != pmu)
4723 perf_event_aux_ctx(&cpuctx->ctx, match, output, data);
4726 ctxn = pmu->task_ctx_nr;
4729 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4731 perf_event_aux_ctx(ctx, match, output, data);
4733 put_cpu_ptr(pmu->pmu_cpu_context);
4738 perf_event_aux_ctx(task_ctx, match, output, data);
4745 * task tracking -- fork/exit
4747 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4750 struct perf_task_event {
4751 struct task_struct *task;
4752 struct perf_event_context *task_ctx;
4755 struct perf_event_header header;
4765 static void perf_event_task_output(struct perf_event *event,
4768 struct perf_task_event *task_event = data;
4769 struct perf_output_handle handle;
4770 struct perf_sample_data sample;
4771 struct task_struct *task = task_event->task;
4772 int ret, size = task_event->event_id.header.size;
4774 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4776 ret = perf_output_begin(&handle, event,
4777 task_event->event_id.header.size);
4781 task_event->event_id.pid = perf_event_pid(event, task);
4782 task_event->event_id.ppid = perf_event_pid(event, current);
4784 task_event->event_id.tid = perf_event_tid(event, task);
4785 task_event->event_id.ptid = perf_event_tid(event, current);
4787 perf_output_put(&handle, task_event->event_id);
4789 perf_event__output_id_sample(event, &handle, &sample);
4791 perf_output_end(&handle);
4793 task_event->event_id.header.size = size;
4796 static int perf_event_task_match(struct perf_event *event,
4797 void *data __maybe_unused)
4799 return event->attr.comm || event->attr.mmap ||
4800 event->attr.mmap_data || event->attr.task;
4803 static void perf_event_task(struct task_struct *task,
4804 struct perf_event_context *task_ctx,
4807 struct perf_task_event task_event;
4809 if (!atomic_read(&nr_comm_events) &&
4810 !atomic_read(&nr_mmap_events) &&
4811 !atomic_read(&nr_task_events))
4814 task_event = (struct perf_task_event){
4816 .task_ctx = task_ctx,
4819 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4821 .size = sizeof(task_event.event_id),
4827 .time = perf_clock(),
4831 perf_event_aux(perf_event_task_match,
4832 perf_event_task_output,
4837 void perf_event_fork(struct task_struct *task)
4839 perf_event_task(task, NULL, 1);
4846 struct perf_comm_event {
4847 struct task_struct *task;
4852 struct perf_event_header header;
4859 static void perf_event_comm_output(struct perf_event *event,
4862 struct perf_comm_event *comm_event = data;
4863 struct perf_output_handle handle;
4864 struct perf_sample_data sample;
4865 int size = comm_event->event_id.header.size;
4868 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4869 ret = perf_output_begin(&handle, event,
4870 comm_event->event_id.header.size);
4875 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4876 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4878 perf_output_put(&handle, comm_event->event_id);
4879 __output_copy(&handle, comm_event->comm,
4880 comm_event->comm_size);
4882 perf_event__output_id_sample(event, &handle, &sample);
4884 perf_output_end(&handle);
4886 comm_event->event_id.header.size = size;
4889 static int perf_event_comm_match(struct perf_event *event,
4890 void *data __maybe_unused)
4892 return event->attr.comm;
4895 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4897 char comm[TASK_COMM_LEN];
4900 memset(comm, 0, sizeof(comm));
4901 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4902 size = ALIGN(strlen(comm)+1, sizeof(u64));
4904 comm_event->comm = comm;
4905 comm_event->comm_size = size;
4907 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4909 perf_event_aux(perf_event_comm_match,
4910 perf_event_comm_output,
4915 void perf_event_comm(struct task_struct *task)
4917 struct perf_comm_event comm_event;
4918 struct perf_event_context *ctx;
4922 for_each_task_context_nr(ctxn) {
4923 ctx = task->perf_event_ctxp[ctxn];
4927 perf_event_enable_on_exec(ctx);
4931 if (!atomic_read(&nr_comm_events))
4934 comm_event = (struct perf_comm_event){
4940 .type = PERF_RECORD_COMM,
4949 perf_event_comm_event(&comm_event);
4956 struct perf_mmap_event {
4957 struct vm_area_struct *vma;
4959 const char *file_name;
4963 struct perf_event_header header;
4973 static void perf_event_mmap_output(struct perf_event *event,
4976 struct perf_mmap_event *mmap_event = data;
4977 struct perf_output_handle handle;
4978 struct perf_sample_data sample;
4979 int size = mmap_event->event_id.header.size;
4982 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4983 ret = perf_output_begin(&handle, event,
4984 mmap_event->event_id.header.size);
4988 mmap_event->event_id.pid = perf_event_pid(event, current);
4989 mmap_event->event_id.tid = perf_event_tid(event, current);
4991 perf_output_put(&handle, mmap_event->event_id);
4992 __output_copy(&handle, mmap_event->file_name,
4993 mmap_event->file_size);
4995 perf_event__output_id_sample(event, &handle, &sample);
4997 perf_output_end(&handle);
4999 mmap_event->event_id.header.size = size;
5002 static int perf_event_mmap_match(struct perf_event *event,
5005 struct perf_mmap_event *mmap_event = data;
5006 struct vm_area_struct *vma = mmap_event->vma;
5007 int executable = vma->vm_flags & VM_EXEC;
5009 return (!executable && event->attr.mmap_data) ||
5010 (executable && event->attr.mmap);
5013 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5015 struct vm_area_struct *vma = mmap_event->vma;
5016 struct file *file = vma->vm_file;
5022 memset(tmp, 0, sizeof(tmp));
5026 * d_path works from the end of the rb backwards, so we
5027 * need to add enough zero bytes after the string to handle
5028 * the 64bit alignment we do later.
5030 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
5032 name = strncpy(tmp, "//enomem", sizeof(tmp));
5035 name = d_path(&file->f_path, buf, PATH_MAX);
5037 name = strncpy(tmp, "//toolong", sizeof(tmp));
5041 if (arch_vma_name(mmap_event->vma)) {
5042 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
5044 tmp[sizeof(tmp) - 1] = '\0';
5049 name = strncpy(tmp, "[vdso]", sizeof(tmp));
5051 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
5052 vma->vm_end >= vma->vm_mm->brk) {
5053 name = strncpy(tmp, "[heap]", sizeof(tmp));
5055 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
5056 vma->vm_end >= vma->vm_mm->start_stack) {
5057 name = strncpy(tmp, "[stack]", sizeof(tmp));
5061 name = strncpy(tmp, "//anon", sizeof(tmp));
5066 size = ALIGN(strlen(name)+1, sizeof(u64));
5068 mmap_event->file_name = name;
5069 mmap_event->file_size = size;
5071 if (!(vma->vm_flags & VM_EXEC))
5072 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5074 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5076 perf_event_aux(perf_event_mmap_match,
5077 perf_event_mmap_output,
5084 void perf_event_mmap(struct vm_area_struct *vma)
5086 struct perf_mmap_event mmap_event;
5088 if (!atomic_read(&nr_mmap_events))
5091 mmap_event = (struct perf_mmap_event){
5097 .type = PERF_RECORD_MMAP,
5098 .misc = PERF_RECORD_MISC_USER,
5103 .start = vma->vm_start,
5104 .len = vma->vm_end - vma->vm_start,
5105 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5109 perf_event_mmap_event(&mmap_event);
5113 * IRQ throttle logging
5116 static void perf_log_throttle(struct perf_event *event, int enable)
5118 struct perf_output_handle handle;
5119 struct perf_sample_data sample;
5123 struct perf_event_header header;
5127 } throttle_event = {
5129 .type = PERF_RECORD_THROTTLE,
5131 .size = sizeof(throttle_event),
5133 .time = perf_clock(),
5134 .id = primary_event_id(event),
5135 .stream_id = event->id,
5139 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5141 perf_event_header__init_id(&throttle_event.header, &sample, event);
5143 ret = perf_output_begin(&handle, event,
5144 throttle_event.header.size);
5148 perf_output_put(&handle, throttle_event);
5149 perf_event__output_id_sample(event, &handle, &sample);
5150 perf_output_end(&handle);
5154 * Generic event overflow handling, sampling.
5157 static int __perf_event_overflow(struct perf_event *event,
5158 int throttle, struct perf_sample_data *data,
5159 struct pt_regs *regs)
5161 int events = atomic_read(&event->event_limit);
5162 struct hw_perf_event *hwc = &event->hw;
5167 * Non-sampling counters might still use the PMI to fold short
5168 * hardware counters, ignore those.
5170 if (unlikely(!is_sampling_event(event)))
5173 seq = __this_cpu_read(perf_throttled_seq);
5174 if (seq != hwc->interrupts_seq) {
5175 hwc->interrupts_seq = seq;
5176 hwc->interrupts = 1;
5179 if (unlikely(throttle
5180 && hwc->interrupts >= max_samples_per_tick)) {
5181 __this_cpu_inc(perf_throttled_count);
5182 hwc->interrupts = MAX_INTERRUPTS;
5183 perf_log_throttle(event, 0);
5188 if (event->attr.freq) {
5189 u64 now = perf_clock();
5190 s64 delta = now - hwc->freq_time_stamp;
5192 hwc->freq_time_stamp = now;
5194 if (delta > 0 && delta < 2*TICK_NSEC)
5195 perf_adjust_period(event, delta, hwc->last_period, true);
5199 * XXX event_limit might not quite work as expected on inherited
5203 event->pending_kill = POLL_IN;
5204 if (events && atomic_dec_and_test(&event->event_limit)) {
5206 event->pending_kill = POLL_HUP;
5207 event->pending_disable = 1;
5208 irq_work_queue(&event->pending);
5211 if (event->overflow_handler)
5212 event->overflow_handler(event, data, regs);
5214 perf_event_output(event, data, regs);
5216 if (event->fasync && event->pending_kill) {
5217 event->pending_wakeup = 1;
5218 irq_work_queue(&event->pending);
5224 int perf_event_overflow(struct perf_event *event,
5225 struct perf_sample_data *data,
5226 struct pt_regs *regs)
5228 return __perf_event_overflow(event, 1, data, regs);
5232 * Generic software event infrastructure
5235 struct swevent_htable {
5236 struct swevent_hlist *swevent_hlist;
5237 struct mutex hlist_mutex;
5240 /* Recursion avoidance in each contexts */
5241 int recursion[PERF_NR_CONTEXTS];
5244 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5247 * We directly increment event->count and keep a second value in
5248 * event->hw.period_left to count intervals. This period event
5249 * is kept in the range [-sample_period, 0] so that we can use the
5253 u64 perf_swevent_set_period(struct perf_event *event)
5255 struct hw_perf_event *hwc = &event->hw;
5256 u64 period = hwc->last_period;
5260 hwc->last_period = hwc->sample_period;
5263 old = val = local64_read(&hwc->period_left);
5267 nr = div64_u64(period + val, period);
5268 offset = nr * period;
5270 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5276 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5277 struct perf_sample_data *data,
5278 struct pt_regs *regs)
5280 struct hw_perf_event *hwc = &event->hw;
5284 overflow = perf_swevent_set_period(event);
5286 if (hwc->interrupts == MAX_INTERRUPTS)
5289 for (; overflow; overflow--) {
5290 if (__perf_event_overflow(event, throttle,
5293 * We inhibit the overflow from happening when
5294 * hwc->interrupts == MAX_INTERRUPTS.
5302 static void perf_swevent_event(struct perf_event *event, u64 nr,
5303 struct perf_sample_data *data,
5304 struct pt_regs *regs)
5306 struct hw_perf_event *hwc = &event->hw;
5308 local64_add(nr, &event->count);
5313 if (!is_sampling_event(event))
5316 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5318 return perf_swevent_overflow(event, 1, data, regs);
5320 data->period = event->hw.last_period;
5322 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5323 return perf_swevent_overflow(event, 1, data, regs);
5325 if (local64_add_negative(nr, &hwc->period_left))
5328 perf_swevent_overflow(event, 0, data, regs);
5331 static int perf_exclude_event(struct perf_event *event,
5332 struct pt_regs *regs)
5334 if (event->hw.state & PERF_HES_STOPPED)
5338 if (event->attr.exclude_user && user_mode(regs))
5341 if (event->attr.exclude_kernel && !user_mode(regs))
5348 static int perf_swevent_match(struct perf_event *event,
5349 enum perf_type_id type,
5351 struct perf_sample_data *data,
5352 struct pt_regs *regs)
5354 if (event->attr.type != type)
5357 if (event->attr.config != event_id)
5360 if (perf_exclude_event(event, regs))
5366 static inline u64 swevent_hash(u64 type, u32 event_id)
5368 u64 val = event_id | (type << 32);
5370 return hash_64(val, SWEVENT_HLIST_BITS);
5373 static inline struct hlist_head *
5374 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5376 u64 hash = swevent_hash(type, event_id);
5378 return &hlist->heads[hash];
5381 /* For the read side: events when they trigger */
5382 static inline struct hlist_head *
5383 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5385 struct swevent_hlist *hlist;
5387 hlist = rcu_dereference(swhash->swevent_hlist);
5391 return __find_swevent_head(hlist, type, event_id);
5394 /* For the event head insertion and removal in the hlist */
5395 static inline struct hlist_head *
5396 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5398 struct swevent_hlist *hlist;
5399 u32 event_id = event->attr.config;
5400 u64 type = event->attr.type;
5403 * Event scheduling is always serialized against hlist allocation
5404 * and release. Which makes the protected version suitable here.
5405 * The context lock guarantees that.
5407 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5408 lockdep_is_held(&event->ctx->lock));
5412 return __find_swevent_head(hlist, type, event_id);
5415 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5417 struct perf_sample_data *data,
5418 struct pt_regs *regs)
5420 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5421 struct perf_event *event;
5422 struct hlist_head *head;
5425 head = find_swevent_head_rcu(swhash, type, event_id);
5429 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5430 if (perf_swevent_match(event, type, event_id, data, regs))
5431 perf_swevent_event(event, nr, data, regs);
5437 int perf_swevent_get_recursion_context(void)
5439 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5441 return get_recursion_context(swhash->recursion);
5443 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5445 inline void perf_swevent_put_recursion_context(int rctx)
5447 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5449 put_recursion_context(swhash->recursion, rctx);
5452 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5454 struct perf_sample_data data;
5457 preempt_disable_notrace();
5458 rctx = perf_swevent_get_recursion_context();
5462 perf_sample_data_init(&data, addr, 0);
5464 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5466 perf_swevent_put_recursion_context(rctx);
5467 preempt_enable_notrace();
5470 static void perf_swevent_read(struct perf_event *event)
5474 static int perf_swevent_add(struct perf_event *event, int flags)
5476 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5477 struct hw_perf_event *hwc = &event->hw;
5478 struct hlist_head *head;
5480 if (is_sampling_event(event)) {
5481 hwc->last_period = hwc->sample_period;
5482 perf_swevent_set_period(event);
5485 hwc->state = !(flags & PERF_EF_START);
5487 head = find_swevent_head(swhash, event);
5488 if (WARN_ON_ONCE(!head))
5491 hlist_add_head_rcu(&event->hlist_entry, head);
5496 static void perf_swevent_del(struct perf_event *event, int flags)
5498 hlist_del_rcu(&event->hlist_entry);
5501 static void perf_swevent_start(struct perf_event *event, int flags)
5503 event->hw.state = 0;
5506 static void perf_swevent_stop(struct perf_event *event, int flags)
5508 event->hw.state = PERF_HES_STOPPED;
5511 /* Deref the hlist from the update side */
5512 static inline struct swevent_hlist *
5513 swevent_hlist_deref(struct swevent_htable *swhash)
5515 return rcu_dereference_protected(swhash->swevent_hlist,
5516 lockdep_is_held(&swhash->hlist_mutex));
5519 static void swevent_hlist_release(struct swevent_htable *swhash)
5521 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5526 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5527 kfree_rcu(hlist, rcu_head);
5530 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5532 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5534 mutex_lock(&swhash->hlist_mutex);
5536 if (!--swhash->hlist_refcount)
5537 swevent_hlist_release(swhash);
5539 mutex_unlock(&swhash->hlist_mutex);
5542 static void swevent_hlist_put(struct perf_event *event)
5546 if (event->cpu != -1) {
5547 swevent_hlist_put_cpu(event, event->cpu);
5551 for_each_possible_cpu(cpu)
5552 swevent_hlist_put_cpu(event, cpu);
5555 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5557 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5560 mutex_lock(&swhash->hlist_mutex);
5562 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5563 struct swevent_hlist *hlist;
5565 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5570 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5572 swhash->hlist_refcount++;
5574 mutex_unlock(&swhash->hlist_mutex);
5579 static int swevent_hlist_get(struct perf_event *event)
5582 int cpu, failed_cpu;
5584 if (event->cpu != -1)
5585 return swevent_hlist_get_cpu(event, event->cpu);
5588 for_each_possible_cpu(cpu) {
5589 err = swevent_hlist_get_cpu(event, cpu);
5599 for_each_possible_cpu(cpu) {
5600 if (cpu == failed_cpu)
5602 swevent_hlist_put_cpu(event, cpu);
5609 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5611 static void sw_perf_event_destroy(struct perf_event *event)
5613 u64 event_id = event->attr.config;
5615 WARN_ON(event->parent);
5617 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5618 swevent_hlist_put(event);
5621 static int perf_swevent_init(struct perf_event *event)
5623 u64 event_id = event->attr.config;
5625 if (event->attr.type != PERF_TYPE_SOFTWARE)
5629 * no branch sampling for software events
5631 if (has_branch_stack(event))
5635 case PERF_COUNT_SW_CPU_CLOCK:
5636 case PERF_COUNT_SW_TASK_CLOCK:
5643 if (event_id >= PERF_COUNT_SW_MAX)
5646 if (!event->parent) {
5649 err = swevent_hlist_get(event);
5653 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5654 event->destroy = sw_perf_event_destroy;
5660 static int perf_swevent_event_idx(struct perf_event *event)
5665 static struct pmu perf_swevent = {
5666 .task_ctx_nr = perf_sw_context,
5668 .event_init = perf_swevent_init,
5669 .add = perf_swevent_add,
5670 .del = perf_swevent_del,
5671 .start = perf_swevent_start,
5672 .stop = perf_swevent_stop,
5673 .read = perf_swevent_read,
5675 .event_idx = perf_swevent_event_idx,
5678 #ifdef CONFIG_EVENT_TRACING
5680 static int perf_tp_filter_match(struct perf_event *event,
5681 struct perf_sample_data *data)
5683 void *record = data->raw->data;
5685 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5690 static int perf_tp_event_match(struct perf_event *event,
5691 struct perf_sample_data *data,
5692 struct pt_regs *regs)
5694 if (event->hw.state & PERF_HES_STOPPED)
5697 * All tracepoints are from kernel-space.
5699 if (event->attr.exclude_kernel)
5702 if (!perf_tp_filter_match(event, data))
5708 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5709 struct pt_regs *regs, struct hlist_head *head, int rctx,
5710 struct task_struct *task)
5712 struct perf_sample_data data;
5713 struct perf_event *event;
5715 struct perf_raw_record raw = {
5720 perf_sample_data_init(&data, addr, 0);
5723 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5724 if (perf_tp_event_match(event, &data, regs))
5725 perf_swevent_event(event, count, &data, regs);
5729 * If we got specified a target task, also iterate its context and
5730 * deliver this event there too.
5732 if (task && task != current) {
5733 struct perf_event_context *ctx;
5734 struct trace_entry *entry = record;
5737 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5741 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5742 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5744 if (event->attr.config != entry->type)
5746 if (perf_tp_event_match(event, &data, regs))
5747 perf_swevent_event(event, count, &data, regs);
5753 perf_swevent_put_recursion_context(rctx);
5755 EXPORT_SYMBOL_GPL(perf_tp_event);
5757 static void tp_perf_event_destroy(struct perf_event *event)
5759 perf_trace_destroy(event);
5762 static int perf_tp_event_init(struct perf_event *event)
5766 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5770 * no branch sampling for tracepoint events
5772 if (has_branch_stack(event))
5775 err = perf_trace_init(event);
5779 event->destroy = tp_perf_event_destroy;
5784 static struct pmu perf_tracepoint = {
5785 .task_ctx_nr = perf_sw_context,
5787 .event_init = perf_tp_event_init,
5788 .add = perf_trace_add,
5789 .del = perf_trace_del,
5790 .start = perf_swevent_start,
5791 .stop = perf_swevent_stop,
5792 .read = perf_swevent_read,
5794 .event_idx = perf_swevent_event_idx,
5797 static inline void perf_tp_register(void)
5799 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5802 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5807 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5810 filter_str = strndup_user(arg, PAGE_SIZE);
5811 if (IS_ERR(filter_str))
5812 return PTR_ERR(filter_str);
5814 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5820 static void perf_event_free_filter(struct perf_event *event)
5822 ftrace_profile_free_filter(event);
5827 static inline void perf_tp_register(void)
5831 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5836 static void perf_event_free_filter(struct perf_event *event)
5840 #endif /* CONFIG_EVENT_TRACING */
5842 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5843 void perf_bp_event(struct perf_event *bp, void *data)
5845 struct perf_sample_data sample;
5846 struct pt_regs *regs = data;
5848 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5850 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5851 perf_swevent_event(bp, 1, &sample, regs);
5856 * hrtimer based swevent callback
5859 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5861 enum hrtimer_restart ret = HRTIMER_RESTART;
5862 struct perf_sample_data data;
5863 struct pt_regs *regs;
5864 struct perf_event *event;
5867 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5869 if (event->state != PERF_EVENT_STATE_ACTIVE)
5870 return HRTIMER_NORESTART;
5872 event->pmu->read(event);
5874 perf_sample_data_init(&data, 0, event->hw.last_period);
5875 regs = get_irq_regs();
5877 if (regs && !perf_exclude_event(event, regs)) {
5878 if (!(event->attr.exclude_idle && is_idle_task(current)))
5879 if (__perf_event_overflow(event, 1, &data, regs))
5880 ret = HRTIMER_NORESTART;
5883 period = max_t(u64, 10000, event->hw.sample_period);
5884 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5889 static void perf_swevent_start_hrtimer(struct perf_event *event)
5891 struct hw_perf_event *hwc = &event->hw;
5894 if (!is_sampling_event(event))
5897 period = local64_read(&hwc->period_left);
5902 local64_set(&hwc->period_left, 0);
5904 period = max_t(u64, 10000, hwc->sample_period);
5906 __hrtimer_start_range_ns(&hwc->hrtimer,
5907 ns_to_ktime(period), 0,
5908 HRTIMER_MODE_REL_PINNED, 0);
5911 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5913 struct hw_perf_event *hwc = &event->hw;
5915 if (is_sampling_event(event)) {
5916 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5917 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5919 hrtimer_cancel(&hwc->hrtimer);
5923 static void perf_swevent_init_hrtimer(struct perf_event *event)
5925 struct hw_perf_event *hwc = &event->hw;
5927 if (!is_sampling_event(event))
5930 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5931 hwc->hrtimer.function = perf_swevent_hrtimer;
5934 * Since hrtimers have a fixed rate, we can do a static freq->period
5935 * mapping and avoid the whole period adjust feedback stuff.
5937 if (event->attr.freq) {
5938 long freq = event->attr.sample_freq;
5940 event->attr.sample_period = NSEC_PER_SEC / freq;
5941 hwc->sample_period = event->attr.sample_period;
5942 local64_set(&hwc->period_left, hwc->sample_period);
5943 hwc->last_period = hwc->sample_period;
5944 event->attr.freq = 0;
5949 * Software event: cpu wall time clock
5952 static void cpu_clock_event_update(struct perf_event *event)
5957 now = local_clock();
5958 prev = local64_xchg(&event->hw.prev_count, now);
5959 local64_add(now - prev, &event->count);
5962 static void cpu_clock_event_start(struct perf_event *event, int flags)
5964 local64_set(&event->hw.prev_count, local_clock());
5965 perf_swevent_start_hrtimer(event);
5968 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5970 perf_swevent_cancel_hrtimer(event);
5971 cpu_clock_event_update(event);
5974 static int cpu_clock_event_add(struct perf_event *event, int flags)
5976 if (flags & PERF_EF_START)
5977 cpu_clock_event_start(event, flags);
5982 static void cpu_clock_event_del(struct perf_event *event, int flags)
5984 cpu_clock_event_stop(event, flags);
5987 static void cpu_clock_event_read(struct perf_event *event)
5989 cpu_clock_event_update(event);
5992 static int cpu_clock_event_init(struct perf_event *event)
5994 if (event->attr.type != PERF_TYPE_SOFTWARE)
5997 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6001 * no branch sampling for software events
6003 if (has_branch_stack(event))
6006 perf_swevent_init_hrtimer(event);
6011 static struct pmu perf_cpu_clock = {
6012 .task_ctx_nr = perf_sw_context,
6014 .event_init = cpu_clock_event_init,
6015 .add = cpu_clock_event_add,
6016 .del = cpu_clock_event_del,
6017 .start = cpu_clock_event_start,
6018 .stop = cpu_clock_event_stop,
6019 .read = cpu_clock_event_read,
6021 .event_idx = perf_swevent_event_idx,
6025 * Software event: task time clock
6028 static void task_clock_event_update(struct perf_event *event, u64 now)
6033 prev = local64_xchg(&event->hw.prev_count, now);
6035 local64_add(delta, &event->count);
6038 static void task_clock_event_start(struct perf_event *event, int flags)
6040 local64_set(&event->hw.prev_count, event->ctx->time);
6041 perf_swevent_start_hrtimer(event);
6044 static void task_clock_event_stop(struct perf_event *event, int flags)
6046 perf_swevent_cancel_hrtimer(event);
6047 task_clock_event_update(event, event->ctx->time);
6050 static int task_clock_event_add(struct perf_event *event, int flags)
6052 if (flags & PERF_EF_START)
6053 task_clock_event_start(event, flags);
6058 static void task_clock_event_del(struct perf_event *event, int flags)
6060 task_clock_event_stop(event, PERF_EF_UPDATE);
6063 static void task_clock_event_read(struct perf_event *event)
6065 u64 now = perf_clock();
6066 u64 delta = now - event->ctx->timestamp;
6067 u64 time = event->ctx->time + delta;
6069 task_clock_event_update(event, time);
6072 static int task_clock_event_init(struct perf_event *event)
6074 if (event->attr.type != PERF_TYPE_SOFTWARE)
6077 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6081 * no branch sampling for software events
6083 if (has_branch_stack(event))
6086 perf_swevent_init_hrtimer(event);
6091 static struct pmu perf_task_clock = {
6092 .task_ctx_nr = perf_sw_context,
6094 .event_init = task_clock_event_init,
6095 .add = task_clock_event_add,
6096 .del = task_clock_event_del,
6097 .start = task_clock_event_start,
6098 .stop = task_clock_event_stop,
6099 .read = task_clock_event_read,
6101 .event_idx = perf_swevent_event_idx,
6104 static void perf_pmu_nop_void(struct pmu *pmu)
6108 static int perf_pmu_nop_int(struct pmu *pmu)
6113 static void perf_pmu_start_txn(struct pmu *pmu)
6115 perf_pmu_disable(pmu);
6118 static int perf_pmu_commit_txn(struct pmu *pmu)
6120 perf_pmu_enable(pmu);
6124 static void perf_pmu_cancel_txn(struct pmu *pmu)
6126 perf_pmu_enable(pmu);
6129 static int perf_event_idx_default(struct perf_event *event)
6131 return event->hw.idx + 1;
6135 * Ensures all contexts with the same task_ctx_nr have the same
6136 * pmu_cpu_context too.
6138 static void *find_pmu_context(int ctxn)
6145 list_for_each_entry(pmu, &pmus, entry) {
6146 if (pmu->task_ctx_nr == ctxn)
6147 return pmu->pmu_cpu_context;
6153 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6157 for_each_possible_cpu(cpu) {
6158 struct perf_cpu_context *cpuctx;
6160 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6162 if (cpuctx->unique_pmu == old_pmu)
6163 cpuctx->unique_pmu = pmu;
6167 static void free_pmu_context(struct pmu *pmu)
6171 mutex_lock(&pmus_lock);
6173 * Like a real lame refcount.
6175 list_for_each_entry(i, &pmus, entry) {
6176 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6177 update_pmu_context(i, pmu);
6182 free_percpu(pmu->pmu_cpu_context);
6184 mutex_unlock(&pmus_lock);
6186 static struct idr pmu_idr;
6189 type_show(struct device *dev, struct device_attribute *attr, char *page)
6191 struct pmu *pmu = dev_get_drvdata(dev);
6193 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6197 perf_event_mux_interval_ms_show(struct device *dev,
6198 struct device_attribute *attr,
6201 struct pmu *pmu = dev_get_drvdata(dev);
6203 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6207 perf_event_mux_interval_ms_store(struct device *dev,
6208 struct device_attribute *attr,
6209 const char *buf, size_t count)
6211 struct pmu *pmu = dev_get_drvdata(dev);
6212 int timer, cpu, ret;
6214 ret = kstrtoint(buf, 0, &timer);
6221 /* same value, noting to do */
6222 if (timer == pmu->hrtimer_interval_ms)
6225 pmu->hrtimer_interval_ms = timer;
6227 /* update all cpuctx for this PMU */
6228 for_each_possible_cpu(cpu) {
6229 struct perf_cpu_context *cpuctx;
6230 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6231 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6233 if (hrtimer_active(&cpuctx->hrtimer))
6234 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6240 static struct device_attribute pmu_dev_attrs[] = {
6242 __ATTR_RW(perf_event_mux_interval_ms),
6246 static int pmu_bus_running;
6247 static struct bus_type pmu_bus = {
6248 .name = "event_source",
6249 .dev_attrs = pmu_dev_attrs,
6252 static void pmu_dev_release(struct device *dev)
6257 static int pmu_dev_alloc(struct pmu *pmu)
6261 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6265 pmu->dev->groups = pmu->attr_groups;
6266 device_initialize(pmu->dev);
6267 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6271 dev_set_drvdata(pmu->dev, pmu);
6272 pmu->dev->bus = &pmu_bus;
6273 pmu->dev->release = pmu_dev_release;
6274 ret = device_add(pmu->dev);
6282 put_device(pmu->dev);
6286 static struct lock_class_key cpuctx_mutex;
6287 static struct lock_class_key cpuctx_lock;
6289 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6293 mutex_lock(&pmus_lock);
6295 pmu->pmu_disable_count = alloc_percpu(int);
6296 if (!pmu->pmu_disable_count)
6305 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6313 if (pmu_bus_running) {
6314 ret = pmu_dev_alloc(pmu);
6320 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6321 if (pmu->pmu_cpu_context)
6322 goto got_cpu_context;
6325 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6326 if (!pmu->pmu_cpu_context)
6329 for_each_possible_cpu(cpu) {
6330 struct perf_cpu_context *cpuctx;
6332 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6333 __perf_event_init_context(&cpuctx->ctx);
6334 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6335 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6336 cpuctx->ctx.type = cpu_context;
6337 cpuctx->ctx.pmu = pmu;
6339 __perf_cpu_hrtimer_init(cpuctx, cpu);
6341 INIT_LIST_HEAD(&cpuctx->rotation_list);
6342 cpuctx->unique_pmu = pmu;
6346 if (!pmu->start_txn) {
6347 if (pmu->pmu_enable) {
6349 * If we have pmu_enable/pmu_disable calls, install
6350 * transaction stubs that use that to try and batch
6351 * hardware accesses.
6353 pmu->start_txn = perf_pmu_start_txn;
6354 pmu->commit_txn = perf_pmu_commit_txn;
6355 pmu->cancel_txn = perf_pmu_cancel_txn;
6357 pmu->start_txn = perf_pmu_nop_void;
6358 pmu->commit_txn = perf_pmu_nop_int;
6359 pmu->cancel_txn = perf_pmu_nop_void;
6363 if (!pmu->pmu_enable) {
6364 pmu->pmu_enable = perf_pmu_nop_void;
6365 pmu->pmu_disable = perf_pmu_nop_void;
6368 if (!pmu->event_idx)
6369 pmu->event_idx = perf_event_idx_default;
6371 list_add_rcu(&pmu->entry, &pmus);
6374 mutex_unlock(&pmus_lock);
6379 device_del(pmu->dev);
6380 put_device(pmu->dev);
6383 if (pmu->type >= PERF_TYPE_MAX)
6384 idr_remove(&pmu_idr, pmu->type);
6387 free_percpu(pmu->pmu_disable_count);
6391 void perf_pmu_unregister(struct pmu *pmu)
6393 mutex_lock(&pmus_lock);
6394 list_del_rcu(&pmu->entry);
6395 mutex_unlock(&pmus_lock);
6398 * We dereference the pmu list under both SRCU and regular RCU, so
6399 * synchronize against both of those.
6401 synchronize_srcu(&pmus_srcu);
6404 free_percpu(pmu->pmu_disable_count);
6405 if (pmu->type >= PERF_TYPE_MAX)
6406 idr_remove(&pmu_idr, pmu->type);
6407 device_del(pmu->dev);
6408 put_device(pmu->dev);
6409 free_pmu_context(pmu);
6412 struct pmu *perf_init_event(struct perf_event *event)
6414 struct pmu *pmu = NULL;
6418 idx = srcu_read_lock(&pmus_srcu);
6421 pmu = idr_find(&pmu_idr, event->attr.type);
6425 ret = pmu->event_init(event);
6431 list_for_each_entry_rcu(pmu, &pmus, entry) {
6433 ret = pmu->event_init(event);
6437 if (ret != -ENOENT) {
6442 pmu = ERR_PTR(-ENOENT);
6444 srcu_read_unlock(&pmus_srcu, idx);
6450 * Allocate and initialize a event structure
6452 static struct perf_event *
6453 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6454 struct task_struct *task,
6455 struct perf_event *group_leader,
6456 struct perf_event *parent_event,
6457 perf_overflow_handler_t overflow_handler,
6461 struct perf_event *event;
6462 struct hw_perf_event *hwc;
6465 if ((unsigned)cpu >= nr_cpu_ids) {
6466 if (!task || cpu != -1)
6467 return ERR_PTR(-EINVAL);
6470 event = kzalloc(sizeof(*event), GFP_KERNEL);
6472 return ERR_PTR(-ENOMEM);
6475 * Single events are their own group leaders, with an
6476 * empty sibling list:
6479 group_leader = event;
6481 mutex_init(&event->child_mutex);
6482 INIT_LIST_HEAD(&event->child_list);
6484 INIT_LIST_HEAD(&event->group_entry);
6485 INIT_LIST_HEAD(&event->event_entry);
6486 INIT_LIST_HEAD(&event->sibling_list);
6487 INIT_LIST_HEAD(&event->rb_entry);
6489 init_waitqueue_head(&event->waitq);
6490 init_irq_work(&event->pending, perf_pending_event);
6492 mutex_init(&event->mmap_mutex);
6494 atomic_long_set(&event->refcount, 1);
6496 event->attr = *attr;
6497 event->group_leader = group_leader;
6501 event->parent = parent_event;
6503 event->ns = get_pid_ns(task_active_pid_ns(current));
6504 event->id = atomic64_inc_return(&perf_event_id);
6506 event->state = PERF_EVENT_STATE_INACTIVE;
6509 event->attach_state = PERF_ATTACH_TASK;
6511 if (attr->type == PERF_TYPE_TRACEPOINT)
6512 event->hw.tp_target = task;
6513 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6515 * hw_breakpoint is a bit difficult here..
6517 else if (attr->type == PERF_TYPE_BREAKPOINT)
6518 event->hw.bp_target = task;
6522 if (!overflow_handler && parent_event) {
6523 overflow_handler = parent_event->overflow_handler;
6524 context = parent_event->overflow_handler_context;
6527 event->overflow_handler = overflow_handler;
6528 event->overflow_handler_context = context;
6530 perf_event__state_init(event);
6535 hwc->sample_period = attr->sample_period;
6536 if (attr->freq && attr->sample_freq)
6537 hwc->sample_period = 1;
6538 hwc->last_period = hwc->sample_period;
6540 local64_set(&hwc->period_left, hwc->sample_period);
6543 * we currently do not support PERF_FORMAT_GROUP on inherited events
6545 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6548 pmu = perf_init_event(event);
6554 else if (IS_ERR(pmu))
6559 put_pid_ns(event->ns);
6561 return ERR_PTR(err);
6564 if (!event->parent) {
6565 if (event->attach_state & PERF_ATTACH_TASK)
6566 static_key_slow_inc(&perf_sched_events.key);
6567 if (event->attr.mmap || event->attr.mmap_data)
6568 atomic_inc(&nr_mmap_events);
6569 if (event->attr.comm)
6570 atomic_inc(&nr_comm_events);
6571 if (event->attr.task)
6572 atomic_inc(&nr_task_events);
6573 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6574 err = get_callchain_buffers();
6577 return ERR_PTR(err);
6580 if (has_branch_stack(event)) {
6581 static_key_slow_inc(&perf_sched_events.key);
6582 if (!(event->attach_state & PERF_ATTACH_TASK))
6583 atomic_inc(&per_cpu(perf_branch_stack_events,
6591 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6592 struct perf_event_attr *attr)
6597 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6601 * zero the full structure, so that a short copy will be nice.
6603 memset(attr, 0, sizeof(*attr));
6605 ret = get_user(size, &uattr->size);
6609 if (size > PAGE_SIZE) /* silly large */
6612 if (!size) /* abi compat */
6613 size = PERF_ATTR_SIZE_VER0;
6615 if (size < PERF_ATTR_SIZE_VER0)
6619 * If we're handed a bigger struct than we know of,
6620 * ensure all the unknown bits are 0 - i.e. new
6621 * user-space does not rely on any kernel feature
6622 * extensions we dont know about yet.
6624 if (size > sizeof(*attr)) {
6625 unsigned char __user *addr;
6626 unsigned char __user *end;
6629 addr = (void __user *)uattr + sizeof(*attr);
6630 end = (void __user *)uattr + size;
6632 for (; addr < end; addr++) {
6633 ret = get_user(val, addr);
6639 size = sizeof(*attr);
6642 ret = copy_from_user(attr, uattr, size);
6646 if (attr->__reserved_1)
6649 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6652 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6655 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6656 u64 mask = attr->branch_sample_type;
6658 /* only using defined bits */
6659 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6662 /* at least one branch bit must be set */
6663 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6666 /* propagate priv level, when not set for branch */
6667 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6669 /* exclude_kernel checked on syscall entry */
6670 if (!attr->exclude_kernel)
6671 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6673 if (!attr->exclude_user)
6674 mask |= PERF_SAMPLE_BRANCH_USER;
6676 if (!attr->exclude_hv)
6677 mask |= PERF_SAMPLE_BRANCH_HV;
6679 * adjust user setting (for HW filter setup)
6681 attr->branch_sample_type = mask;
6683 /* privileged levels capture (kernel, hv): check permissions */
6684 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6685 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6689 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6690 ret = perf_reg_validate(attr->sample_regs_user);
6695 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6696 if (!arch_perf_have_user_stack_dump())
6700 * We have __u32 type for the size, but so far
6701 * we can only use __u16 as maximum due to the
6702 * __u16 sample size limit.
6704 if (attr->sample_stack_user >= USHRT_MAX)
6706 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6714 put_user(sizeof(*attr), &uattr->size);
6720 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6722 struct ring_buffer *rb = NULL, *old_rb = NULL;
6728 /* don't allow circular references */
6729 if (event == output_event)
6733 * Don't allow cross-cpu buffers
6735 if (output_event->cpu != event->cpu)
6739 * If its not a per-cpu rb, it must be the same task.
6741 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6745 mutex_lock(&event->mmap_mutex);
6746 /* Can't redirect output if we've got an active mmap() */
6747 if (atomic_read(&event->mmap_count))
6753 /* get the rb we want to redirect to */
6754 rb = ring_buffer_get(output_event);
6760 ring_buffer_detach(event, old_rb);
6763 ring_buffer_attach(event, rb);
6765 rcu_assign_pointer(event->rb, rb);
6768 ring_buffer_put(old_rb);
6770 * Since we detached before setting the new rb, so that we
6771 * could attach the new rb, we could have missed a wakeup.
6774 wake_up_all(&event->waitq);
6779 mutex_unlock(&event->mmap_mutex);
6786 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6788 * @attr_uptr: event_id type attributes for monitoring/sampling
6791 * @group_fd: group leader event fd
6793 SYSCALL_DEFINE5(perf_event_open,
6794 struct perf_event_attr __user *, attr_uptr,
6795 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6797 struct perf_event *group_leader = NULL, *output_event = NULL;
6798 struct perf_event *event, *sibling;
6799 struct perf_event_attr attr;
6800 struct perf_event_context *ctx;
6801 struct file *event_file = NULL;
6802 struct fd group = {NULL, 0};
6803 struct task_struct *task = NULL;
6809 /* for future expandability... */
6810 if (flags & ~PERF_FLAG_ALL)
6813 err = perf_copy_attr(attr_uptr, &attr);
6817 if (!attr.exclude_kernel) {
6818 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6823 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6828 * In cgroup mode, the pid argument is used to pass the fd
6829 * opened to the cgroup directory in cgroupfs. The cpu argument
6830 * designates the cpu on which to monitor threads from that
6833 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6836 event_fd = get_unused_fd();
6840 if (group_fd != -1) {
6841 err = perf_fget_light(group_fd, &group);
6844 group_leader = group.file->private_data;
6845 if (flags & PERF_FLAG_FD_OUTPUT)
6846 output_event = group_leader;
6847 if (flags & PERF_FLAG_FD_NO_GROUP)
6848 group_leader = NULL;
6851 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6852 task = find_lively_task_by_vpid(pid);
6854 err = PTR_ERR(task);
6861 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6863 if (IS_ERR(event)) {
6864 err = PTR_ERR(event);
6868 if (flags & PERF_FLAG_PID_CGROUP) {
6869 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6874 * - that has cgroup constraint on event->cpu
6875 * - that may need work on context switch
6877 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6878 static_key_slow_inc(&perf_sched_events.key);
6882 * Special case software events and allow them to be part of
6883 * any hardware group.
6888 (is_software_event(event) != is_software_event(group_leader))) {
6889 if (is_software_event(event)) {
6891 * If event and group_leader are not both a software
6892 * event, and event is, then group leader is not.
6894 * Allow the addition of software events to !software
6895 * groups, this is safe because software events never
6898 pmu = group_leader->pmu;
6899 } else if (is_software_event(group_leader) &&
6900 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6902 * In case the group is a pure software group, and we
6903 * try to add a hardware event, move the whole group to
6904 * the hardware context.
6911 * Get the target context (task or percpu):
6913 ctx = find_get_context(pmu, task, event->cpu);
6920 put_task_struct(task);
6925 * Look up the group leader (we will attach this event to it):
6931 * Do not allow a recursive hierarchy (this new sibling
6932 * becoming part of another group-sibling):
6934 if (group_leader->group_leader != group_leader)
6937 * Do not allow to attach to a group in a different
6938 * task or CPU context:
6941 if (group_leader->ctx->type != ctx->type)
6944 if (group_leader->ctx != ctx)
6949 * Only a group leader can be exclusive or pinned
6951 if (attr.exclusive || attr.pinned)
6956 err = perf_event_set_output(event, output_event);
6961 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6962 if (IS_ERR(event_file)) {
6963 err = PTR_ERR(event_file);
6968 struct perf_event_context *gctx = group_leader->ctx;
6970 mutex_lock(&gctx->mutex);
6971 perf_remove_from_context(group_leader);
6974 * Removing from the context ends up with disabled
6975 * event. What we want here is event in the initial
6976 * startup state, ready to be add into new context.
6978 perf_event__state_init(group_leader);
6979 list_for_each_entry(sibling, &group_leader->sibling_list,
6981 perf_remove_from_context(sibling);
6982 perf_event__state_init(sibling);
6985 mutex_unlock(&gctx->mutex);
6989 WARN_ON_ONCE(ctx->parent_ctx);
6990 mutex_lock(&ctx->mutex);
6994 perf_install_in_context(ctx, group_leader, event->cpu);
6996 list_for_each_entry(sibling, &group_leader->sibling_list,
6998 perf_install_in_context(ctx, sibling, event->cpu);
7003 perf_install_in_context(ctx, event, event->cpu);
7005 perf_unpin_context(ctx);
7006 mutex_unlock(&ctx->mutex);
7010 event->owner = current;
7012 mutex_lock(¤t->perf_event_mutex);
7013 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7014 mutex_unlock(¤t->perf_event_mutex);
7017 * Precalculate sample_data sizes
7019 perf_event__header_size(event);
7020 perf_event__id_header_size(event);
7023 * Drop the reference on the group_event after placing the
7024 * new event on the sibling_list. This ensures destruction
7025 * of the group leader will find the pointer to itself in
7026 * perf_group_detach().
7029 fd_install(event_fd, event_file);
7033 perf_unpin_context(ctx);
7040 put_task_struct(task);
7044 put_unused_fd(event_fd);
7049 * perf_event_create_kernel_counter
7051 * @attr: attributes of the counter to create
7052 * @cpu: cpu in which the counter is bound
7053 * @task: task to profile (NULL for percpu)
7056 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7057 struct task_struct *task,
7058 perf_overflow_handler_t overflow_handler,
7061 struct perf_event_context *ctx;
7062 struct perf_event *event;
7066 * Get the target context (task or percpu):
7069 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7070 overflow_handler, context);
7071 if (IS_ERR(event)) {
7072 err = PTR_ERR(event);
7076 ctx = find_get_context(event->pmu, task, cpu);
7082 WARN_ON_ONCE(ctx->parent_ctx);
7083 mutex_lock(&ctx->mutex);
7084 perf_install_in_context(ctx, event, cpu);
7086 perf_unpin_context(ctx);
7087 mutex_unlock(&ctx->mutex);
7094 return ERR_PTR(err);
7096 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7098 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7100 struct perf_event_context *src_ctx;
7101 struct perf_event_context *dst_ctx;
7102 struct perf_event *event, *tmp;
7105 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7106 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7108 mutex_lock(&src_ctx->mutex);
7109 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7111 perf_remove_from_context(event);
7113 list_add(&event->event_entry, &events);
7115 mutex_unlock(&src_ctx->mutex);
7119 mutex_lock(&dst_ctx->mutex);
7120 list_for_each_entry_safe(event, tmp, &events, event_entry) {
7121 list_del(&event->event_entry);
7122 if (event->state >= PERF_EVENT_STATE_OFF)
7123 event->state = PERF_EVENT_STATE_INACTIVE;
7124 perf_install_in_context(dst_ctx, event, dst_cpu);
7127 mutex_unlock(&dst_ctx->mutex);
7129 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7131 static void sync_child_event(struct perf_event *child_event,
7132 struct task_struct *child)
7134 struct perf_event *parent_event = child_event->parent;
7137 if (child_event->attr.inherit_stat)
7138 perf_event_read_event(child_event, child);
7140 child_val = perf_event_count(child_event);
7143 * Add back the child's count to the parent's count:
7145 atomic64_add(child_val, &parent_event->child_count);
7146 atomic64_add(child_event->total_time_enabled,
7147 &parent_event->child_total_time_enabled);
7148 atomic64_add(child_event->total_time_running,
7149 &parent_event->child_total_time_running);
7152 * Remove this event from the parent's list
7154 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7155 mutex_lock(&parent_event->child_mutex);
7156 list_del_init(&child_event->child_list);
7157 mutex_unlock(&parent_event->child_mutex);
7160 * Release the parent event, if this was the last
7163 put_event(parent_event);
7167 __perf_event_exit_task(struct perf_event *child_event,
7168 struct perf_event_context *child_ctx,
7169 struct task_struct *child)
7171 if (child_event->parent) {
7172 raw_spin_lock_irq(&child_ctx->lock);
7173 perf_group_detach(child_event);
7174 raw_spin_unlock_irq(&child_ctx->lock);
7177 perf_remove_from_context(child_event);
7180 * It can happen that the parent exits first, and has events
7181 * that are still around due to the child reference. These
7182 * events need to be zapped.
7184 if (child_event->parent) {
7185 sync_child_event(child_event, child);
7186 free_event(child_event);
7190 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7192 struct perf_event *child_event, *tmp;
7193 struct perf_event_context *child_ctx;
7194 unsigned long flags;
7196 if (likely(!child->perf_event_ctxp[ctxn])) {
7197 perf_event_task(child, NULL, 0);
7201 local_irq_save(flags);
7203 * We can't reschedule here because interrupts are disabled,
7204 * and either child is current or it is a task that can't be
7205 * scheduled, so we are now safe from rescheduling changing
7208 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7211 * Take the context lock here so that if find_get_context is
7212 * reading child->perf_event_ctxp, we wait until it has
7213 * incremented the context's refcount before we do put_ctx below.
7215 raw_spin_lock(&child_ctx->lock);
7216 task_ctx_sched_out(child_ctx);
7217 child->perf_event_ctxp[ctxn] = NULL;
7219 * If this context is a clone; unclone it so it can't get
7220 * swapped to another process while we're removing all
7221 * the events from it.
7223 unclone_ctx(child_ctx);
7224 update_context_time(child_ctx);
7225 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7228 * Report the task dead after unscheduling the events so that we
7229 * won't get any samples after PERF_RECORD_EXIT. We can however still
7230 * get a few PERF_RECORD_READ events.
7232 perf_event_task(child, child_ctx, 0);
7235 * We can recurse on the same lock type through:
7237 * __perf_event_exit_task()
7238 * sync_child_event()
7240 * mutex_lock(&ctx->mutex)
7242 * But since its the parent context it won't be the same instance.
7244 mutex_lock(&child_ctx->mutex);
7247 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7249 __perf_event_exit_task(child_event, child_ctx, child);
7251 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7253 __perf_event_exit_task(child_event, child_ctx, child);
7256 * If the last event was a group event, it will have appended all
7257 * its siblings to the list, but we obtained 'tmp' before that which
7258 * will still point to the list head terminating the iteration.
7260 if (!list_empty(&child_ctx->pinned_groups) ||
7261 !list_empty(&child_ctx->flexible_groups))
7264 mutex_unlock(&child_ctx->mutex);
7270 * When a child task exits, feed back event values to parent events.
7272 void perf_event_exit_task(struct task_struct *child)
7274 struct perf_event *event, *tmp;
7277 mutex_lock(&child->perf_event_mutex);
7278 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7280 list_del_init(&event->owner_entry);
7283 * Ensure the list deletion is visible before we clear
7284 * the owner, closes a race against perf_release() where
7285 * we need to serialize on the owner->perf_event_mutex.
7288 event->owner = NULL;
7290 mutex_unlock(&child->perf_event_mutex);
7292 for_each_task_context_nr(ctxn)
7293 perf_event_exit_task_context(child, ctxn);
7296 static void perf_free_event(struct perf_event *event,
7297 struct perf_event_context *ctx)
7299 struct perf_event *parent = event->parent;
7301 if (WARN_ON_ONCE(!parent))
7304 mutex_lock(&parent->child_mutex);
7305 list_del_init(&event->child_list);
7306 mutex_unlock(&parent->child_mutex);
7310 perf_group_detach(event);
7311 list_del_event(event, ctx);
7316 * free an unexposed, unused context as created by inheritance by
7317 * perf_event_init_task below, used by fork() in case of fail.
7319 void perf_event_free_task(struct task_struct *task)
7321 struct perf_event_context *ctx;
7322 struct perf_event *event, *tmp;
7325 for_each_task_context_nr(ctxn) {
7326 ctx = task->perf_event_ctxp[ctxn];
7330 mutex_lock(&ctx->mutex);
7332 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7334 perf_free_event(event, ctx);
7336 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7338 perf_free_event(event, ctx);
7340 if (!list_empty(&ctx->pinned_groups) ||
7341 !list_empty(&ctx->flexible_groups))
7344 mutex_unlock(&ctx->mutex);
7350 void perf_event_delayed_put(struct task_struct *task)
7354 for_each_task_context_nr(ctxn)
7355 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7359 * inherit a event from parent task to child task:
7361 static struct perf_event *
7362 inherit_event(struct perf_event *parent_event,
7363 struct task_struct *parent,
7364 struct perf_event_context *parent_ctx,
7365 struct task_struct *child,
7366 struct perf_event *group_leader,
7367 struct perf_event_context *child_ctx)
7369 struct perf_event *child_event;
7370 unsigned long flags;
7373 * Instead of creating recursive hierarchies of events,
7374 * we link inherited events back to the original parent,
7375 * which has a filp for sure, which we use as the reference
7378 if (parent_event->parent)
7379 parent_event = parent_event->parent;
7381 child_event = perf_event_alloc(&parent_event->attr,
7384 group_leader, parent_event,
7386 if (IS_ERR(child_event))
7389 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7390 free_event(child_event);
7397 * Make the child state follow the state of the parent event,
7398 * not its attr.disabled bit. We hold the parent's mutex,
7399 * so we won't race with perf_event_{en, dis}able_family.
7401 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7402 child_event->state = PERF_EVENT_STATE_INACTIVE;
7404 child_event->state = PERF_EVENT_STATE_OFF;
7406 if (parent_event->attr.freq) {
7407 u64 sample_period = parent_event->hw.sample_period;
7408 struct hw_perf_event *hwc = &child_event->hw;
7410 hwc->sample_period = sample_period;
7411 hwc->last_period = sample_period;
7413 local64_set(&hwc->period_left, sample_period);
7416 child_event->ctx = child_ctx;
7417 child_event->overflow_handler = parent_event->overflow_handler;
7418 child_event->overflow_handler_context
7419 = parent_event->overflow_handler_context;
7422 * Precalculate sample_data sizes
7424 perf_event__header_size(child_event);
7425 perf_event__id_header_size(child_event);
7428 * Link it up in the child's context:
7430 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7431 add_event_to_ctx(child_event, child_ctx);
7432 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7435 * Link this into the parent event's child list
7437 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7438 mutex_lock(&parent_event->child_mutex);
7439 list_add_tail(&child_event->child_list, &parent_event->child_list);
7440 mutex_unlock(&parent_event->child_mutex);
7445 static int inherit_group(struct perf_event *parent_event,
7446 struct task_struct *parent,
7447 struct perf_event_context *parent_ctx,
7448 struct task_struct *child,
7449 struct perf_event_context *child_ctx)
7451 struct perf_event *leader;
7452 struct perf_event *sub;
7453 struct perf_event *child_ctr;
7455 leader = inherit_event(parent_event, parent, parent_ctx,
7456 child, NULL, child_ctx);
7458 return PTR_ERR(leader);
7459 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7460 child_ctr = inherit_event(sub, parent, parent_ctx,
7461 child, leader, child_ctx);
7462 if (IS_ERR(child_ctr))
7463 return PTR_ERR(child_ctr);
7469 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7470 struct perf_event_context *parent_ctx,
7471 struct task_struct *child, int ctxn,
7475 struct perf_event_context *child_ctx;
7477 if (!event->attr.inherit) {
7482 child_ctx = child->perf_event_ctxp[ctxn];
7485 * This is executed from the parent task context, so
7486 * inherit events that have been marked for cloning.
7487 * First allocate and initialize a context for the
7491 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7495 child->perf_event_ctxp[ctxn] = child_ctx;
7498 ret = inherit_group(event, parent, parent_ctx,
7508 * Initialize the perf_event context in task_struct
7510 int perf_event_init_context(struct task_struct *child, int ctxn)
7512 struct perf_event_context *child_ctx, *parent_ctx;
7513 struct perf_event_context *cloned_ctx;
7514 struct perf_event *event;
7515 struct task_struct *parent = current;
7516 int inherited_all = 1;
7517 unsigned long flags;
7520 if (likely(!parent->perf_event_ctxp[ctxn]))
7524 * If the parent's context is a clone, pin it so it won't get
7527 parent_ctx = perf_pin_task_context(parent, ctxn);
7530 * No need to check if parent_ctx != NULL here; since we saw
7531 * it non-NULL earlier, the only reason for it to become NULL
7532 * is if we exit, and since we're currently in the middle of
7533 * a fork we can't be exiting at the same time.
7537 * Lock the parent list. No need to lock the child - not PID
7538 * hashed yet and not running, so nobody can access it.
7540 mutex_lock(&parent_ctx->mutex);
7543 * We dont have to disable NMIs - we are only looking at
7544 * the list, not manipulating it:
7546 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7547 ret = inherit_task_group(event, parent, parent_ctx,
7548 child, ctxn, &inherited_all);
7554 * We can't hold ctx->lock when iterating the ->flexible_group list due
7555 * to allocations, but we need to prevent rotation because
7556 * rotate_ctx() will change the list from interrupt context.
7558 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7559 parent_ctx->rotate_disable = 1;
7560 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7562 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7563 ret = inherit_task_group(event, parent, parent_ctx,
7564 child, ctxn, &inherited_all);
7569 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7570 parent_ctx->rotate_disable = 0;
7572 child_ctx = child->perf_event_ctxp[ctxn];
7574 if (child_ctx && inherited_all) {
7576 * Mark the child context as a clone of the parent
7577 * context, or of whatever the parent is a clone of.
7579 * Note that if the parent is a clone, the holding of
7580 * parent_ctx->lock avoids it from being uncloned.
7582 cloned_ctx = parent_ctx->parent_ctx;
7584 child_ctx->parent_ctx = cloned_ctx;
7585 child_ctx->parent_gen = parent_ctx->parent_gen;
7587 child_ctx->parent_ctx = parent_ctx;
7588 child_ctx->parent_gen = parent_ctx->generation;
7590 get_ctx(child_ctx->parent_ctx);
7593 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7594 mutex_unlock(&parent_ctx->mutex);
7596 perf_unpin_context(parent_ctx);
7597 put_ctx(parent_ctx);
7603 * Initialize the perf_event context in task_struct
7605 int perf_event_init_task(struct task_struct *child)
7609 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7610 mutex_init(&child->perf_event_mutex);
7611 INIT_LIST_HEAD(&child->perf_event_list);
7613 for_each_task_context_nr(ctxn) {
7614 ret = perf_event_init_context(child, ctxn);
7622 static void __init perf_event_init_all_cpus(void)
7624 struct swevent_htable *swhash;
7627 for_each_possible_cpu(cpu) {
7628 swhash = &per_cpu(swevent_htable, cpu);
7629 mutex_init(&swhash->hlist_mutex);
7630 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7634 static void perf_event_init_cpu(int cpu)
7636 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7638 mutex_lock(&swhash->hlist_mutex);
7639 if (swhash->hlist_refcount > 0) {
7640 struct swevent_hlist *hlist;
7642 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7644 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7646 mutex_unlock(&swhash->hlist_mutex);
7649 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7650 static void perf_pmu_rotate_stop(struct pmu *pmu)
7652 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7654 WARN_ON(!irqs_disabled());
7656 list_del_init(&cpuctx->rotation_list);
7659 static void __perf_event_exit_context(void *__info)
7661 struct perf_event_context *ctx = __info;
7662 struct perf_event *event, *tmp;
7664 perf_pmu_rotate_stop(ctx->pmu);
7666 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7667 __perf_remove_from_context(event);
7668 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7669 __perf_remove_from_context(event);
7672 static void perf_event_exit_cpu_context(int cpu)
7674 struct perf_event_context *ctx;
7678 idx = srcu_read_lock(&pmus_srcu);
7679 list_for_each_entry_rcu(pmu, &pmus, entry) {
7680 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7682 mutex_lock(&ctx->mutex);
7683 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7684 mutex_unlock(&ctx->mutex);
7686 srcu_read_unlock(&pmus_srcu, idx);
7689 static void perf_event_exit_cpu(int cpu)
7691 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7693 mutex_lock(&swhash->hlist_mutex);
7694 swevent_hlist_release(swhash);
7695 mutex_unlock(&swhash->hlist_mutex);
7697 perf_event_exit_cpu_context(cpu);
7700 static inline void perf_event_exit_cpu(int cpu) { }
7704 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7708 for_each_online_cpu(cpu)
7709 perf_event_exit_cpu(cpu);
7715 * Run the perf reboot notifier at the very last possible moment so that
7716 * the generic watchdog code runs as long as possible.
7718 static struct notifier_block perf_reboot_notifier = {
7719 .notifier_call = perf_reboot,
7720 .priority = INT_MIN,
7724 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7726 unsigned int cpu = (long)hcpu;
7728 switch (action & ~CPU_TASKS_FROZEN) {
7730 case CPU_UP_PREPARE:
7731 case CPU_DOWN_FAILED:
7732 perf_event_init_cpu(cpu);
7735 case CPU_UP_CANCELED:
7736 case CPU_DOWN_PREPARE:
7737 perf_event_exit_cpu(cpu);
7746 void __init perf_event_init(void)
7752 perf_event_init_all_cpus();
7753 init_srcu_struct(&pmus_srcu);
7754 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7755 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7756 perf_pmu_register(&perf_task_clock, NULL, -1);
7758 perf_cpu_notifier(perf_cpu_notify);
7759 register_reboot_notifier(&perf_reboot_notifier);
7761 ret = init_hw_breakpoint();
7762 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7764 /* do not patch jump label more than once per second */
7765 jump_label_rate_limit(&perf_sched_events, HZ);
7768 * Build time assertion that we keep the data_head at the intended
7769 * location. IOW, validation we got the __reserved[] size right.
7771 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7775 static int __init perf_event_sysfs_init(void)
7780 mutex_lock(&pmus_lock);
7782 ret = bus_register(&pmu_bus);
7786 list_for_each_entry(pmu, &pmus, entry) {
7787 if (!pmu->name || pmu->type < 0)
7790 ret = pmu_dev_alloc(pmu);
7791 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7793 pmu_bus_running = 1;
7797 mutex_unlock(&pmus_lock);
7801 device_initcall(perf_event_sysfs_init);
7803 #ifdef CONFIG_CGROUP_PERF
7804 static struct cgroup_subsys_state *
7805 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7807 struct perf_cgroup *jc;
7809 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7811 return ERR_PTR(-ENOMEM);
7813 jc->info = alloc_percpu(struct perf_cgroup_info);
7816 return ERR_PTR(-ENOMEM);
7822 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
7824 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
7826 free_percpu(jc->info);
7830 static int __perf_cgroup_move(void *info)
7832 struct task_struct *task = info;
7833 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7837 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
7838 struct cgroup_taskset *tset)
7840 struct task_struct *task;
7842 cgroup_taskset_for_each(task, css, tset)
7843 task_function_call(task, __perf_cgroup_move, task);
7846 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
7847 struct cgroup_subsys_state *old_css,
7848 struct task_struct *task)
7851 * cgroup_exit() is called in the copy_process() failure path.
7852 * Ignore this case since the task hasn't ran yet, this avoids
7853 * trying to poke a half freed task state from generic code.
7855 if (!(task->flags & PF_EXITING))
7858 task_function_call(task, __perf_cgroup_move, task);
7861 struct cgroup_subsys perf_subsys = {
7862 .name = "perf_event",
7863 .subsys_id = perf_subsys_id,
7864 .css_alloc = perf_cgroup_css_alloc,
7865 .css_free = perf_cgroup_css_free,
7866 .exit = perf_cgroup_exit,
7867 .attach = perf_cgroup_attach,
7869 #endif /* CONFIG_CGROUP_PERF */