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;
148 static atomic_t nr_freq_events __read_mostly;
150 static LIST_HEAD(pmus);
151 static DEFINE_MUTEX(pmus_lock);
152 static struct srcu_struct pmus_srcu;
155 * perf event paranoia level:
156 * -1 - not paranoid at all
157 * 0 - disallow raw tracepoint access for unpriv
158 * 1 - disallow cpu events for unpriv
159 * 2 - disallow kernel profiling for unpriv
161 int sysctl_perf_event_paranoid __read_mostly = 1;
163 /* Minimum for 512 kiB + 1 user control page */
164 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
167 * max perf event sample rate
169 #define DEFAULT_MAX_SAMPLE_RATE 100000
170 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
171 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
173 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
175 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
176 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
178 static atomic_t perf_sample_allowed_ns __read_mostly =
179 ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100);
181 void update_perf_cpu_limits(void)
183 u64 tmp = perf_sample_period_ns;
185 tmp *= sysctl_perf_cpu_time_max_percent;
187 atomic_set(&perf_sample_allowed_ns, tmp);
190 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
192 int perf_proc_update_handler(struct ctl_table *table, int write,
193 void __user *buffer, size_t *lenp,
196 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
201 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
202 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
203 update_perf_cpu_limits();
208 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
210 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
211 void __user *buffer, size_t *lenp,
214 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
219 update_perf_cpu_limits();
225 * perf samples are done in some very critical code paths (NMIs).
226 * If they take too much CPU time, the system can lock up and not
227 * get any real work done. This will drop the sample rate when
228 * we detect that events are taking too long.
230 #define NR_ACCUMULATED_SAMPLES 128
231 DEFINE_PER_CPU(u64, running_sample_length);
233 void perf_sample_event_took(u64 sample_len_ns)
235 u64 avg_local_sample_len;
236 u64 local_samples_len;
238 if (atomic_read(&perf_sample_allowed_ns) == 0)
241 /* decay the counter by 1 average sample */
242 local_samples_len = __get_cpu_var(running_sample_length);
243 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
244 local_samples_len += sample_len_ns;
245 __get_cpu_var(running_sample_length) = local_samples_len;
248 * note: this will be biased artifically low until we have
249 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
250 * from having to maintain a count.
252 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
254 if (avg_local_sample_len <= atomic_read(&perf_sample_allowed_ns))
257 if (max_samples_per_tick <= 1)
260 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
261 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
262 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
264 printk_ratelimited(KERN_WARNING
265 "perf samples too long (%lld > %d), lowering "
266 "kernel.perf_event_max_sample_rate to %d\n",
267 avg_local_sample_len,
268 atomic_read(&perf_sample_allowed_ns),
269 sysctl_perf_event_sample_rate);
271 update_perf_cpu_limits();
274 static atomic64_t perf_event_id;
276 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
277 enum event_type_t event_type);
279 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
280 enum event_type_t event_type,
281 struct task_struct *task);
283 static void update_context_time(struct perf_event_context *ctx);
284 static u64 perf_event_time(struct perf_event *event);
286 void __weak perf_event_print_debug(void) { }
288 extern __weak const char *perf_pmu_name(void)
293 static inline u64 perf_clock(void)
295 return local_clock();
298 static inline struct perf_cpu_context *
299 __get_cpu_context(struct perf_event_context *ctx)
301 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
304 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
305 struct perf_event_context *ctx)
307 raw_spin_lock(&cpuctx->ctx.lock);
309 raw_spin_lock(&ctx->lock);
312 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
313 struct perf_event_context *ctx)
316 raw_spin_unlock(&ctx->lock);
317 raw_spin_unlock(&cpuctx->ctx.lock);
320 #ifdef CONFIG_CGROUP_PERF
323 * perf_cgroup_info keeps track of time_enabled for a cgroup.
324 * This is a per-cpu dynamically allocated data structure.
326 struct perf_cgroup_info {
332 struct cgroup_subsys_state css;
333 struct perf_cgroup_info __percpu *info;
337 * Must ensure cgroup is pinned (css_get) before calling
338 * this function. In other words, we cannot call this function
339 * if there is no cgroup event for the current CPU context.
341 static inline struct perf_cgroup *
342 perf_cgroup_from_task(struct task_struct *task)
344 return container_of(task_subsys_state(task, perf_subsys_id),
345 struct perf_cgroup, css);
349 perf_cgroup_match(struct perf_event *event)
351 struct perf_event_context *ctx = event->ctx;
352 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
354 /* @event doesn't care about cgroup */
358 /* wants specific cgroup scope but @cpuctx isn't associated with any */
363 * Cgroup scoping is recursive. An event enabled for a cgroup is
364 * also enabled for all its descendant cgroups. If @cpuctx's
365 * cgroup is a descendant of @event's (the test covers identity
366 * case), it's a match.
368 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
369 event->cgrp->css.cgroup);
372 static inline bool perf_tryget_cgroup(struct perf_event *event)
374 return css_tryget(&event->cgrp->css);
377 static inline void perf_put_cgroup(struct perf_event *event)
379 css_put(&event->cgrp->css);
382 static inline void perf_detach_cgroup(struct perf_event *event)
384 perf_put_cgroup(event);
388 static inline int is_cgroup_event(struct perf_event *event)
390 return event->cgrp != NULL;
393 static inline u64 perf_cgroup_event_time(struct perf_event *event)
395 struct perf_cgroup_info *t;
397 t = per_cpu_ptr(event->cgrp->info, event->cpu);
401 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
403 struct perf_cgroup_info *info;
408 info = this_cpu_ptr(cgrp->info);
410 info->time += now - info->timestamp;
411 info->timestamp = now;
414 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
416 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
418 __update_cgrp_time(cgrp_out);
421 static inline void update_cgrp_time_from_event(struct perf_event *event)
423 struct perf_cgroup *cgrp;
426 * ensure we access cgroup data only when needed and
427 * when we know the cgroup is pinned (css_get)
429 if (!is_cgroup_event(event))
432 cgrp = perf_cgroup_from_task(current);
434 * Do not update time when cgroup is not active
436 if (cgrp == event->cgrp)
437 __update_cgrp_time(event->cgrp);
441 perf_cgroup_set_timestamp(struct task_struct *task,
442 struct perf_event_context *ctx)
444 struct perf_cgroup *cgrp;
445 struct perf_cgroup_info *info;
448 * ctx->lock held by caller
449 * ensure we do not access cgroup data
450 * unless we have the cgroup pinned (css_get)
452 if (!task || !ctx->nr_cgroups)
455 cgrp = perf_cgroup_from_task(task);
456 info = this_cpu_ptr(cgrp->info);
457 info->timestamp = ctx->timestamp;
460 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
461 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
464 * reschedule events based on the cgroup constraint of task.
466 * mode SWOUT : schedule out everything
467 * mode SWIN : schedule in based on cgroup for next
469 void perf_cgroup_switch(struct task_struct *task, int mode)
471 struct perf_cpu_context *cpuctx;
476 * disable interrupts to avoid geting nr_cgroup
477 * changes via __perf_event_disable(). Also
480 local_irq_save(flags);
483 * we reschedule only in the presence of cgroup
484 * constrained events.
488 list_for_each_entry_rcu(pmu, &pmus, entry) {
489 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
490 if (cpuctx->unique_pmu != pmu)
491 continue; /* ensure we process each cpuctx once */
494 * perf_cgroup_events says at least one
495 * context on this CPU has cgroup events.
497 * ctx->nr_cgroups reports the number of cgroup
498 * events for a context.
500 if (cpuctx->ctx.nr_cgroups > 0) {
501 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
502 perf_pmu_disable(cpuctx->ctx.pmu);
504 if (mode & PERF_CGROUP_SWOUT) {
505 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
507 * must not be done before ctxswout due
508 * to event_filter_match() in event_sched_out()
513 if (mode & PERF_CGROUP_SWIN) {
514 WARN_ON_ONCE(cpuctx->cgrp);
516 * set cgrp before ctxsw in to allow
517 * event_filter_match() to not have to pass
520 cpuctx->cgrp = perf_cgroup_from_task(task);
521 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
523 perf_pmu_enable(cpuctx->ctx.pmu);
524 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
530 local_irq_restore(flags);
533 static inline void perf_cgroup_sched_out(struct task_struct *task,
534 struct task_struct *next)
536 struct perf_cgroup *cgrp1;
537 struct perf_cgroup *cgrp2 = NULL;
540 * we come here when we know perf_cgroup_events > 0
542 cgrp1 = perf_cgroup_from_task(task);
545 * next is NULL when called from perf_event_enable_on_exec()
546 * that will systematically cause a cgroup_switch()
549 cgrp2 = perf_cgroup_from_task(next);
552 * only schedule out current cgroup events if we know
553 * that we are switching to a different cgroup. Otherwise,
554 * do no touch the cgroup events.
557 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
560 static inline void perf_cgroup_sched_in(struct task_struct *prev,
561 struct task_struct *task)
563 struct perf_cgroup *cgrp1;
564 struct perf_cgroup *cgrp2 = NULL;
567 * we come here when we know perf_cgroup_events > 0
569 cgrp1 = perf_cgroup_from_task(task);
571 /* prev can never be NULL */
572 cgrp2 = perf_cgroup_from_task(prev);
575 * only need to schedule in cgroup events if we are changing
576 * cgroup during ctxsw. Cgroup events were not scheduled
577 * out of ctxsw out if that was not the case.
580 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
583 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
584 struct perf_event_attr *attr,
585 struct perf_event *group_leader)
587 struct perf_cgroup *cgrp;
588 struct cgroup_subsys_state *css;
589 struct fd f = fdget(fd);
595 css = cgroup_css_from_dir(f.file, perf_subsys_id);
601 cgrp = container_of(css, struct perf_cgroup, css);
604 /* must be done before we fput() the file */
605 if (!perf_tryget_cgroup(event)) {
612 * all events in a group must monitor
613 * the same cgroup because a task belongs
614 * to only one perf cgroup at a time
616 if (group_leader && group_leader->cgrp != cgrp) {
617 perf_detach_cgroup(event);
626 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
628 struct perf_cgroup_info *t;
629 t = per_cpu_ptr(event->cgrp->info, event->cpu);
630 event->shadow_ctx_time = now - t->timestamp;
634 perf_cgroup_defer_enabled(struct perf_event *event)
637 * when the current task's perf cgroup does not match
638 * the event's, we need to remember to call the
639 * perf_mark_enable() function the first time a task with
640 * a matching perf cgroup is scheduled in.
642 if (is_cgroup_event(event) && !perf_cgroup_match(event))
643 event->cgrp_defer_enabled = 1;
647 perf_cgroup_mark_enabled(struct perf_event *event,
648 struct perf_event_context *ctx)
650 struct perf_event *sub;
651 u64 tstamp = perf_event_time(event);
653 if (!event->cgrp_defer_enabled)
656 event->cgrp_defer_enabled = 0;
658 event->tstamp_enabled = tstamp - event->total_time_enabled;
659 list_for_each_entry(sub, &event->sibling_list, group_entry) {
660 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
661 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
662 sub->cgrp_defer_enabled = 0;
666 #else /* !CONFIG_CGROUP_PERF */
669 perf_cgroup_match(struct perf_event *event)
674 static inline void perf_detach_cgroup(struct perf_event *event)
677 static inline int is_cgroup_event(struct perf_event *event)
682 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
687 static inline void update_cgrp_time_from_event(struct perf_event *event)
691 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
695 static inline void perf_cgroup_sched_out(struct task_struct *task,
696 struct task_struct *next)
700 static inline void perf_cgroup_sched_in(struct task_struct *prev,
701 struct task_struct *task)
705 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
706 struct perf_event_attr *attr,
707 struct perf_event *group_leader)
713 perf_cgroup_set_timestamp(struct task_struct *task,
714 struct perf_event_context *ctx)
719 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
724 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
728 static inline u64 perf_cgroup_event_time(struct perf_event *event)
734 perf_cgroup_defer_enabled(struct perf_event *event)
739 perf_cgroup_mark_enabled(struct perf_event *event,
740 struct perf_event_context *ctx)
746 * set default to be dependent on timer tick just
749 #define PERF_CPU_HRTIMER (1000 / HZ)
751 * function must be called with interrupts disbled
753 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
755 struct perf_cpu_context *cpuctx;
756 enum hrtimer_restart ret = HRTIMER_NORESTART;
759 WARN_ON(!irqs_disabled());
761 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
763 rotations = perf_rotate_context(cpuctx);
766 * arm timer if needed
769 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
770 ret = HRTIMER_RESTART;
776 /* CPU is going down */
777 void perf_cpu_hrtimer_cancel(int cpu)
779 struct perf_cpu_context *cpuctx;
783 if (WARN_ON(cpu != smp_processor_id()))
786 local_irq_save(flags);
790 list_for_each_entry_rcu(pmu, &pmus, entry) {
791 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
793 if (pmu->task_ctx_nr == perf_sw_context)
796 hrtimer_cancel(&cpuctx->hrtimer);
801 local_irq_restore(flags);
804 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
806 struct hrtimer *hr = &cpuctx->hrtimer;
807 struct pmu *pmu = cpuctx->ctx.pmu;
810 /* no multiplexing needed for SW PMU */
811 if (pmu->task_ctx_nr == perf_sw_context)
815 * check default is sane, if not set then force to
816 * default interval (1/tick)
818 timer = pmu->hrtimer_interval_ms;
820 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
822 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
824 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
825 hr->function = perf_cpu_hrtimer_handler;
828 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
830 struct hrtimer *hr = &cpuctx->hrtimer;
831 struct pmu *pmu = cpuctx->ctx.pmu;
834 if (pmu->task_ctx_nr == perf_sw_context)
837 if (hrtimer_active(hr))
840 if (!hrtimer_callback_running(hr))
841 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
842 0, HRTIMER_MODE_REL_PINNED, 0);
845 void perf_pmu_disable(struct pmu *pmu)
847 int *count = this_cpu_ptr(pmu->pmu_disable_count);
849 pmu->pmu_disable(pmu);
852 void perf_pmu_enable(struct pmu *pmu)
854 int *count = this_cpu_ptr(pmu->pmu_disable_count);
856 pmu->pmu_enable(pmu);
859 static DEFINE_PER_CPU(struct list_head, rotation_list);
862 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
863 * because they're strictly cpu affine and rotate_start is called with IRQs
864 * disabled, while rotate_context is called from IRQ context.
866 static void perf_pmu_rotate_start(struct pmu *pmu)
868 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
869 struct list_head *head = &__get_cpu_var(rotation_list);
871 WARN_ON(!irqs_disabled());
873 if (list_empty(&cpuctx->rotation_list))
874 list_add(&cpuctx->rotation_list, head);
877 static void get_ctx(struct perf_event_context *ctx)
879 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
882 static void put_ctx(struct perf_event_context *ctx)
884 if (atomic_dec_and_test(&ctx->refcount)) {
886 put_ctx(ctx->parent_ctx);
888 put_task_struct(ctx->task);
889 kfree_rcu(ctx, rcu_head);
893 static void unclone_ctx(struct perf_event_context *ctx)
895 if (ctx->parent_ctx) {
896 put_ctx(ctx->parent_ctx);
897 ctx->parent_ctx = NULL;
901 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
904 * only top level events have the pid namespace they were created in
907 event = event->parent;
909 return task_tgid_nr_ns(p, event->ns);
912 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
915 * only top level events have the pid namespace they were created in
918 event = event->parent;
920 return task_pid_nr_ns(p, event->ns);
924 * If we inherit events we want to return the parent event id
927 static u64 primary_event_id(struct perf_event *event)
932 id = event->parent->id;
938 * Get the perf_event_context for a task and lock it.
939 * This has to cope with with the fact that until it is locked,
940 * the context could get moved to another task.
942 static struct perf_event_context *
943 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
945 struct perf_event_context *ctx;
949 * One of the few rules of preemptible RCU is that one cannot do
950 * rcu_read_unlock() while holding a scheduler (or nested) lock when
951 * part of the read side critical section was preemptible -- see
952 * rcu_read_unlock_special().
954 * Since ctx->lock nests under rq->lock we must ensure the entire read
955 * side critical section is non-preemptible.
959 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
962 * If this context is a clone of another, it might
963 * get swapped for another underneath us by
964 * perf_event_task_sched_out, though the
965 * rcu_read_lock() protects us from any context
966 * getting freed. Lock the context and check if it
967 * got swapped before we could get the lock, and retry
968 * if so. If we locked the right context, then it
969 * can't get swapped on us any more.
971 raw_spin_lock_irqsave(&ctx->lock, *flags);
972 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
973 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
979 if (!atomic_inc_not_zero(&ctx->refcount)) {
980 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
990 * Get the context for a task and increment its pin_count so it
991 * can't get swapped to another task. This also increments its
992 * reference count so that the context can't get freed.
994 static struct perf_event_context *
995 perf_pin_task_context(struct task_struct *task, int ctxn)
997 struct perf_event_context *ctx;
1000 ctx = perf_lock_task_context(task, ctxn, &flags);
1003 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1008 static void perf_unpin_context(struct perf_event_context *ctx)
1010 unsigned long flags;
1012 raw_spin_lock_irqsave(&ctx->lock, flags);
1014 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1018 * Update the record of the current time in a context.
1020 static void update_context_time(struct perf_event_context *ctx)
1022 u64 now = perf_clock();
1024 ctx->time += now - ctx->timestamp;
1025 ctx->timestamp = now;
1028 static u64 perf_event_time(struct perf_event *event)
1030 struct perf_event_context *ctx = event->ctx;
1032 if (is_cgroup_event(event))
1033 return perf_cgroup_event_time(event);
1035 return ctx ? ctx->time : 0;
1039 * Update the total_time_enabled and total_time_running fields for a event.
1040 * The caller of this function needs to hold the ctx->lock.
1042 static void update_event_times(struct perf_event *event)
1044 struct perf_event_context *ctx = event->ctx;
1047 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1048 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1051 * in cgroup mode, time_enabled represents
1052 * the time the event was enabled AND active
1053 * tasks were in the monitored cgroup. This is
1054 * independent of the activity of the context as
1055 * there may be a mix of cgroup and non-cgroup events.
1057 * That is why we treat cgroup events differently
1060 if (is_cgroup_event(event))
1061 run_end = perf_cgroup_event_time(event);
1062 else if (ctx->is_active)
1063 run_end = ctx->time;
1065 run_end = event->tstamp_stopped;
1067 event->total_time_enabled = run_end - event->tstamp_enabled;
1069 if (event->state == PERF_EVENT_STATE_INACTIVE)
1070 run_end = event->tstamp_stopped;
1072 run_end = perf_event_time(event);
1074 event->total_time_running = run_end - event->tstamp_running;
1079 * Update total_time_enabled and total_time_running for all events in a group.
1081 static void update_group_times(struct perf_event *leader)
1083 struct perf_event *event;
1085 update_event_times(leader);
1086 list_for_each_entry(event, &leader->sibling_list, group_entry)
1087 update_event_times(event);
1090 static struct list_head *
1091 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1093 if (event->attr.pinned)
1094 return &ctx->pinned_groups;
1096 return &ctx->flexible_groups;
1100 * Add a event from the lists for its context.
1101 * Must be called with ctx->mutex and ctx->lock held.
1104 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1106 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1107 event->attach_state |= PERF_ATTACH_CONTEXT;
1110 * If we're a stand alone event or group leader, we go to the context
1111 * list, group events are kept attached to the group so that
1112 * perf_group_detach can, at all times, locate all siblings.
1114 if (event->group_leader == event) {
1115 struct list_head *list;
1117 if (is_software_event(event))
1118 event->group_flags |= PERF_GROUP_SOFTWARE;
1120 list = ctx_group_list(event, ctx);
1121 list_add_tail(&event->group_entry, list);
1124 if (is_cgroup_event(event))
1127 if (has_branch_stack(event))
1128 ctx->nr_branch_stack++;
1130 list_add_rcu(&event->event_entry, &ctx->event_list);
1131 if (!ctx->nr_events)
1132 perf_pmu_rotate_start(ctx->pmu);
1134 if (event->attr.inherit_stat)
1139 * Initialize event state based on the perf_event_attr::disabled.
1141 static inline void perf_event__state_init(struct perf_event *event)
1143 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1144 PERF_EVENT_STATE_INACTIVE;
1148 * Called at perf_event creation and when events are attached/detached from a
1151 static void perf_event__read_size(struct perf_event *event)
1153 int entry = sizeof(u64); /* value */
1157 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1158 size += sizeof(u64);
1160 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1161 size += sizeof(u64);
1163 if (event->attr.read_format & PERF_FORMAT_ID)
1164 entry += sizeof(u64);
1166 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1167 nr += event->group_leader->nr_siblings;
1168 size += sizeof(u64);
1172 event->read_size = size;
1175 static void perf_event__header_size(struct perf_event *event)
1177 struct perf_sample_data *data;
1178 u64 sample_type = event->attr.sample_type;
1181 perf_event__read_size(event);
1183 if (sample_type & PERF_SAMPLE_IP)
1184 size += sizeof(data->ip);
1186 if (sample_type & PERF_SAMPLE_ADDR)
1187 size += sizeof(data->addr);
1189 if (sample_type & PERF_SAMPLE_PERIOD)
1190 size += sizeof(data->period);
1192 if (sample_type & PERF_SAMPLE_WEIGHT)
1193 size += sizeof(data->weight);
1195 if (sample_type & PERF_SAMPLE_READ)
1196 size += event->read_size;
1198 if (sample_type & PERF_SAMPLE_DATA_SRC)
1199 size += sizeof(data->data_src.val);
1201 event->header_size = size;
1204 static void perf_event__id_header_size(struct perf_event *event)
1206 struct perf_sample_data *data;
1207 u64 sample_type = event->attr.sample_type;
1210 if (sample_type & PERF_SAMPLE_TID)
1211 size += sizeof(data->tid_entry);
1213 if (sample_type & PERF_SAMPLE_TIME)
1214 size += sizeof(data->time);
1216 if (sample_type & PERF_SAMPLE_ID)
1217 size += sizeof(data->id);
1219 if (sample_type & PERF_SAMPLE_STREAM_ID)
1220 size += sizeof(data->stream_id);
1222 if (sample_type & PERF_SAMPLE_CPU)
1223 size += sizeof(data->cpu_entry);
1225 event->id_header_size = size;
1228 static void perf_group_attach(struct perf_event *event)
1230 struct perf_event *group_leader = event->group_leader, *pos;
1233 * We can have double attach due to group movement in perf_event_open.
1235 if (event->attach_state & PERF_ATTACH_GROUP)
1238 event->attach_state |= PERF_ATTACH_GROUP;
1240 if (group_leader == event)
1243 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1244 !is_software_event(event))
1245 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1247 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1248 group_leader->nr_siblings++;
1250 perf_event__header_size(group_leader);
1252 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1253 perf_event__header_size(pos);
1257 * Remove a event from the lists for its context.
1258 * Must be called with ctx->mutex and ctx->lock held.
1261 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1263 struct perf_cpu_context *cpuctx;
1265 * We can have double detach due to exit/hot-unplug + close.
1267 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1270 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1272 if (is_cgroup_event(event)) {
1274 cpuctx = __get_cpu_context(ctx);
1276 * if there are no more cgroup events
1277 * then cler cgrp to avoid stale pointer
1278 * in update_cgrp_time_from_cpuctx()
1280 if (!ctx->nr_cgroups)
1281 cpuctx->cgrp = NULL;
1284 if (has_branch_stack(event))
1285 ctx->nr_branch_stack--;
1288 if (event->attr.inherit_stat)
1291 list_del_rcu(&event->event_entry);
1293 if (event->group_leader == event)
1294 list_del_init(&event->group_entry);
1296 update_group_times(event);
1299 * If event was in error state, then keep it
1300 * that way, otherwise bogus counts will be
1301 * returned on read(). The only way to get out
1302 * of error state is by explicit re-enabling
1305 if (event->state > PERF_EVENT_STATE_OFF)
1306 event->state = PERF_EVENT_STATE_OFF;
1309 static void perf_group_detach(struct perf_event *event)
1311 struct perf_event *sibling, *tmp;
1312 struct list_head *list = NULL;
1315 * We can have double detach due to exit/hot-unplug + close.
1317 if (!(event->attach_state & PERF_ATTACH_GROUP))
1320 event->attach_state &= ~PERF_ATTACH_GROUP;
1323 * If this is a sibling, remove it from its group.
1325 if (event->group_leader != event) {
1326 list_del_init(&event->group_entry);
1327 event->group_leader->nr_siblings--;
1331 if (!list_empty(&event->group_entry))
1332 list = &event->group_entry;
1335 * If this was a group event with sibling events then
1336 * upgrade the siblings to singleton events by adding them
1337 * to whatever list we are on.
1339 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1341 list_move_tail(&sibling->group_entry, list);
1342 sibling->group_leader = sibling;
1344 /* Inherit group flags from the previous leader */
1345 sibling->group_flags = event->group_flags;
1349 perf_event__header_size(event->group_leader);
1351 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1352 perf_event__header_size(tmp);
1356 event_filter_match(struct perf_event *event)
1358 return (event->cpu == -1 || event->cpu == smp_processor_id())
1359 && perf_cgroup_match(event);
1363 event_sched_out(struct perf_event *event,
1364 struct perf_cpu_context *cpuctx,
1365 struct perf_event_context *ctx)
1367 u64 tstamp = perf_event_time(event);
1370 * An event which could not be activated because of
1371 * filter mismatch still needs to have its timings
1372 * maintained, otherwise bogus information is return
1373 * via read() for time_enabled, time_running:
1375 if (event->state == PERF_EVENT_STATE_INACTIVE
1376 && !event_filter_match(event)) {
1377 delta = tstamp - event->tstamp_stopped;
1378 event->tstamp_running += delta;
1379 event->tstamp_stopped = tstamp;
1382 if (event->state != PERF_EVENT_STATE_ACTIVE)
1385 event->state = PERF_EVENT_STATE_INACTIVE;
1386 if (event->pending_disable) {
1387 event->pending_disable = 0;
1388 event->state = PERF_EVENT_STATE_OFF;
1390 event->tstamp_stopped = tstamp;
1391 event->pmu->del(event, 0);
1394 if (!is_software_event(event))
1395 cpuctx->active_oncpu--;
1397 if (event->attr.freq && event->attr.sample_freq)
1399 if (event->attr.exclusive || !cpuctx->active_oncpu)
1400 cpuctx->exclusive = 0;
1404 group_sched_out(struct perf_event *group_event,
1405 struct perf_cpu_context *cpuctx,
1406 struct perf_event_context *ctx)
1408 struct perf_event *event;
1409 int state = group_event->state;
1411 event_sched_out(group_event, cpuctx, ctx);
1414 * Schedule out siblings (if any):
1416 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1417 event_sched_out(event, cpuctx, ctx);
1419 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1420 cpuctx->exclusive = 0;
1424 * Cross CPU call to remove a performance event
1426 * We disable the event on the hardware level first. After that we
1427 * remove it from the context list.
1429 static int __perf_remove_from_context(void *info)
1431 struct perf_event *event = info;
1432 struct perf_event_context *ctx = event->ctx;
1433 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1435 raw_spin_lock(&ctx->lock);
1436 event_sched_out(event, cpuctx, ctx);
1437 list_del_event(event, ctx);
1438 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1440 cpuctx->task_ctx = NULL;
1442 raw_spin_unlock(&ctx->lock);
1449 * Remove the event from a task's (or a CPU's) list of events.
1451 * CPU events are removed with a smp call. For task events we only
1452 * call when the task is on a CPU.
1454 * If event->ctx is a cloned context, callers must make sure that
1455 * every task struct that event->ctx->task could possibly point to
1456 * remains valid. This is OK when called from perf_release since
1457 * that only calls us on the top-level context, which can't be a clone.
1458 * When called from perf_event_exit_task, it's OK because the
1459 * context has been detached from its task.
1461 static void perf_remove_from_context(struct perf_event *event)
1463 struct perf_event_context *ctx = event->ctx;
1464 struct task_struct *task = ctx->task;
1466 lockdep_assert_held(&ctx->mutex);
1470 * Per cpu events are removed via an smp call and
1471 * the removal is always successful.
1473 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1478 if (!task_function_call(task, __perf_remove_from_context, event))
1481 raw_spin_lock_irq(&ctx->lock);
1483 * If we failed to find a running task, but find the context active now
1484 * that we've acquired the ctx->lock, retry.
1486 if (ctx->is_active) {
1487 raw_spin_unlock_irq(&ctx->lock);
1492 * Since the task isn't running, its safe to remove the event, us
1493 * holding the ctx->lock ensures the task won't get scheduled in.
1495 list_del_event(event, ctx);
1496 raw_spin_unlock_irq(&ctx->lock);
1500 * Cross CPU call to disable a performance event
1502 int __perf_event_disable(void *info)
1504 struct perf_event *event = info;
1505 struct perf_event_context *ctx = event->ctx;
1506 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1509 * If this is a per-task event, need to check whether this
1510 * event's task is the current task on this cpu.
1512 * Can trigger due to concurrent perf_event_context_sched_out()
1513 * flipping contexts around.
1515 if (ctx->task && cpuctx->task_ctx != ctx)
1518 raw_spin_lock(&ctx->lock);
1521 * If the event is on, turn it off.
1522 * If it is in error state, leave it in error state.
1524 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1525 update_context_time(ctx);
1526 update_cgrp_time_from_event(event);
1527 update_group_times(event);
1528 if (event == event->group_leader)
1529 group_sched_out(event, cpuctx, ctx);
1531 event_sched_out(event, cpuctx, ctx);
1532 event->state = PERF_EVENT_STATE_OFF;
1535 raw_spin_unlock(&ctx->lock);
1543 * If event->ctx is a cloned context, callers must make sure that
1544 * every task struct that event->ctx->task could possibly point to
1545 * remains valid. This condition is satisifed when called through
1546 * perf_event_for_each_child or perf_event_for_each because they
1547 * hold the top-level event's child_mutex, so any descendant that
1548 * goes to exit will block in sync_child_event.
1549 * When called from perf_pending_event it's OK because event->ctx
1550 * is the current context on this CPU and preemption is disabled,
1551 * hence we can't get into perf_event_task_sched_out for this context.
1553 void perf_event_disable(struct perf_event *event)
1555 struct perf_event_context *ctx = event->ctx;
1556 struct task_struct *task = ctx->task;
1560 * Disable the event on the cpu that it's on
1562 cpu_function_call(event->cpu, __perf_event_disable, event);
1567 if (!task_function_call(task, __perf_event_disable, event))
1570 raw_spin_lock_irq(&ctx->lock);
1572 * If the event is still active, we need to retry the cross-call.
1574 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1575 raw_spin_unlock_irq(&ctx->lock);
1577 * Reload the task pointer, it might have been changed by
1578 * a concurrent perf_event_context_sched_out().
1585 * Since we have the lock this context can't be scheduled
1586 * in, so we can change the state safely.
1588 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1589 update_group_times(event);
1590 event->state = PERF_EVENT_STATE_OFF;
1592 raw_spin_unlock_irq(&ctx->lock);
1594 EXPORT_SYMBOL_GPL(perf_event_disable);
1596 static void perf_set_shadow_time(struct perf_event *event,
1597 struct perf_event_context *ctx,
1601 * use the correct time source for the time snapshot
1603 * We could get by without this by leveraging the
1604 * fact that to get to this function, the caller
1605 * has most likely already called update_context_time()
1606 * and update_cgrp_time_xx() and thus both timestamp
1607 * are identical (or very close). Given that tstamp is,
1608 * already adjusted for cgroup, we could say that:
1609 * tstamp - ctx->timestamp
1611 * tstamp - cgrp->timestamp.
1613 * Then, in perf_output_read(), the calculation would
1614 * work with no changes because:
1615 * - event is guaranteed scheduled in
1616 * - no scheduled out in between
1617 * - thus the timestamp would be the same
1619 * But this is a bit hairy.
1621 * So instead, we have an explicit cgroup call to remain
1622 * within the time time source all along. We believe it
1623 * is cleaner and simpler to understand.
1625 if (is_cgroup_event(event))
1626 perf_cgroup_set_shadow_time(event, tstamp);
1628 event->shadow_ctx_time = tstamp - ctx->timestamp;
1631 #define MAX_INTERRUPTS (~0ULL)
1633 static void perf_log_throttle(struct perf_event *event, int enable);
1636 event_sched_in(struct perf_event *event,
1637 struct perf_cpu_context *cpuctx,
1638 struct perf_event_context *ctx)
1640 u64 tstamp = perf_event_time(event);
1642 if (event->state <= PERF_EVENT_STATE_OFF)
1645 event->state = PERF_EVENT_STATE_ACTIVE;
1646 event->oncpu = smp_processor_id();
1649 * Unthrottle events, since we scheduled we might have missed several
1650 * ticks already, also for a heavily scheduling task there is little
1651 * guarantee it'll get a tick in a timely manner.
1653 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1654 perf_log_throttle(event, 1);
1655 event->hw.interrupts = 0;
1659 * The new state must be visible before we turn it on in the hardware:
1663 if (event->pmu->add(event, PERF_EF_START)) {
1664 event->state = PERF_EVENT_STATE_INACTIVE;
1669 event->tstamp_running += tstamp - event->tstamp_stopped;
1671 perf_set_shadow_time(event, ctx, tstamp);
1673 if (!is_software_event(event))
1674 cpuctx->active_oncpu++;
1676 if (event->attr.freq && event->attr.sample_freq)
1679 if (event->attr.exclusive)
1680 cpuctx->exclusive = 1;
1686 group_sched_in(struct perf_event *group_event,
1687 struct perf_cpu_context *cpuctx,
1688 struct perf_event_context *ctx)
1690 struct perf_event *event, *partial_group = NULL;
1691 struct pmu *pmu = group_event->pmu;
1692 u64 now = ctx->time;
1693 bool simulate = false;
1695 if (group_event->state == PERF_EVENT_STATE_OFF)
1698 pmu->start_txn(pmu);
1700 if (event_sched_in(group_event, cpuctx, ctx)) {
1701 pmu->cancel_txn(pmu);
1702 perf_cpu_hrtimer_restart(cpuctx);
1707 * Schedule in siblings as one group (if any):
1709 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1710 if (event_sched_in(event, cpuctx, ctx)) {
1711 partial_group = event;
1716 if (!pmu->commit_txn(pmu))
1721 * Groups can be scheduled in as one unit only, so undo any
1722 * partial group before returning:
1723 * The events up to the failed event are scheduled out normally,
1724 * tstamp_stopped will be updated.
1726 * The failed events and the remaining siblings need to have
1727 * their timings updated as if they had gone thru event_sched_in()
1728 * and event_sched_out(). This is required to get consistent timings
1729 * across the group. This also takes care of the case where the group
1730 * could never be scheduled by ensuring tstamp_stopped is set to mark
1731 * the time the event was actually stopped, such that time delta
1732 * calculation in update_event_times() is correct.
1734 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1735 if (event == partial_group)
1739 event->tstamp_running += now - event->tstamp_stopped;
1740 event->tstamp_stopped = now;
1742 event_sched_out(event, cpuctx, ctx);
1745 event_sched_out(group_event, cpuctx, ctx);
1747 pmu->cancel_txn(pmu);
1749 perf_cpu_hrtimer_restart(cpuctx);
1755 * Work out whether we can put this event group on the CPU now.
1757 static int group_can_go_on(struct perf_event *event,
1758 struct perf_cpu_context *cpuctx,
1762 * Groups consisting entirely of software events can always go on.
1764 if (event->group_flags & PERF_GROUP_SOFTWARE)
1767 * If an exclusive group is already on, no other hardware
1770 if (cpuctx->exclusive)
1773 * If this group is exclusive and there are already
1774 * events on the CPU, it can't go on.
1776 if (event->attr.exclusive && cpuctx->active_oncpu)
1779 * Otherwise, try to add it if all previous groups were able
1785 static void add_event_to_ctx(struct perf_event *event,
1786 struct perf_event_context *ctx)
1788 u64 tstamp = perf_event_time(event);
1790 list_add_event(event, ctx);
1791 perf_group_attach(event);
1792 event->tstamp_enabled = tstamp;
1793 event->tstamp_running = tstamp;
1794 event->tstamp_stopped = tstamp;
1797 static void task_ctx_sched_out(struct perf_event_context *ctx);
1799 ctx_sched_in(struct perf_event_context *ctx,
1800 struct perf_cpu_context *cpuctx,
1801 enum event_type_t event_type,
1802 struct task_struct *task);
1804 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1805 struct perf_event_context *ctx,
1806 struct task_struct *task)
1808 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1810 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1811 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1813 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1817 * Cross CPU call to install and enable a performance event
1819 * Must be called with ctx->mutex held
1821 static int __perf_install_in_context(void *info)
1823 struct perf_event *event = info;
1824 struct perf_event_context *ctx = event->ctx;
1825 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1826 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1827 struct task_struct *task = current;
1829 perf_ctx_lock(cpuctx, task_ctx);
1830 perf_pmu_disable(cpuctx->ctx.pmu);
1833 * If there was an active task_ctx schedule it out.
1836 task_ctx_sched_out(task_ctx);
1839 * If the context we're installing events in is not the
1840 * active task_ctx, flip them.
1842 if (ctx->task && task_ctx != ctx) {
1844 raw_spin_unlock(&task_ctx->lock);
1845 raw_spin_lock(&ctx->lock);
1850 cpuctx->task_ctx = task_ctx;
1851 task = task_ctx->task;
1854 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1856 update_context_time(ctx);
1858 * update cgrp time only if current cgrp
1859 * matches event->cgrp. Must be done before
1860 * calling add_event_to_ctx()
1862 update_cgrp_time_from_event(event);
1864 add_event_to_ctx(event, ctx);
1867 * Schedule everything back in
1869 perf_event_sched_in(cpuctx, task_ctx, task);
1871 perf_pmu_enable(cpuctx->ctx.pmu);
1872 perf_ctx_unlock(cpuctx, task_ctx);
1878 * Attach a performance event to a context
1880 * First we add the event to the list with the hardware enable bit
1881 * in event->hw_config cleared.
1883 * If the event is attached to a task which is on a CPU we use a smp
1884 * call to enable it in the task context. The task might have been
1885 * scheduled away, but we check this in the smp call again.
1888 perf_install_in_context(struct perf_event_context *ctx,
1889 struct perf_event *event,
1892 struct task_struct *task = ctx->task;
1894 lockdep_assert_held(&ctx->mutex);
1897 if (event->cpu != -1)
1902 * Per cpu events are installed via an smp call and
1903 * the install is always successful.
1905 cpu_function_call(cpu, __perf_install_in_context, event);
1910 if (!task_function_call(task, __perf_install_in_context, event))
1913 raw_spin_lock_irq(&ctx->lock);
1915 * If we failed to find a running task, but find the context active now
1916 * that we've acquired the ctx->lock, retry.
1918 if (ctx->is_active) {
1919 raw_spin_unlock_irq(&ctx->lock);
1924 * Since the task isn't running, its safe to add the event, us holding
1925 * the ctx->lock ensures the task won't get scheduled in.
1927 add_event_to_ctx(event, ctx);
1928 raw_spin_unlock_irq(&ctx->lock);
1932 * Put a event into inactive state and update time fields.
1933 * Enabling the leader of a group effectively enables all
1934 * the group members that aren't explicitly disabled, so we
1935 * have to update their ->tstamp_enabled also.
1936 * Note: this works for group members as well as group leaders
1937 * since the non-leader members' sibling_lists will be empty.
1939 static void __perf_event_mark_enabled(struct perf_event *event)
1941 struct perf_event *sub;
1942 u64 tstamp = perf_event_time(event);
1944 event->state = PERF_EVENT_STATE_INACTIVE;
1945 event->tstamp_enabled = tstamp - event->total_time_enabled;
1946 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1947 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1948 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1953 * Cross CPU call to enable a performance event
1955 static int __perf_event_enable(void *info)
1957 struct perf_event *event = info;
1958 struct perf_event_context *ctx = event->ctx;
1959 struct perf_event *leader = event->group_leader;
1960 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1964 * There's a time window between 'ctx->is_active' check
1965 * in perf_event_enable function and this place having:
1967 * - ctx->lock unlocked
1969 * where the task could be killed and 'ctx' deactivated
1970 * by perf_event_exit_task.
1972 if (!ctx->is_active)
1975 raw_spin_lock(&ctx->lock);
1976 update_context_time(ctx);
1978 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1982 * set current task's cgroup time reference point
1984 perf_cgroup_set_timestamp(current, ctx);
1986 __perf_event_mark_enabled(event);
1988 if (!event_filter_match(event)) {
1989 if (is_cgroup_event(event))
1990 perf_cgroup_defer_enabled(event);
1995 * If the event is in a group and isn't the group leader,
1996 * then don't put it on unless the group is on.
1998 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2001 if (!group_can_go_on(event, cpuctx, 1)) {
2004 if (event == leader)
2005 err = group_sched_in(event, cpuctx, ctx);
2007 err = event_sched_in(event, cpuctx, ctx);
2012 * If this event can't go on and it's part of a
2013 * group, then the whole group has to come off.
2015 if (leader != event) {
2016 group_sched_out(leader, cpuctx, ctx);
2017 perf_cpu_hrtimer_restart(cpuctx);
2019 if (leader->attr.pinned) {
2020 update_group_times(leader);
2021 leader->state = PERF_EVENT_STATE_ERROR;
2026 raw_spin_unlock(&ctx->lock);
2034 * If event->ctx is a cloned context, callers must make sure that
2035 * every task struct that event->ctx->task could possibly point to
2036 * remains valid. This condition is satisfied when called through
2037 * perf_event_for_each_child or perf_event_for_each as described
2038 * for perf_event_disable.
2040 void perf_event_enable(struct perf_event *event)
2042 struct perf_event_context *ctx = event->ctx;
2043 struct task_struct *task = ctx->task;
2047 * Enable the event on the cpu that it's on
2049 cpu_function_call(event->cpu, __perf_event_enable, event);
2053 raw_spin_lock_irq(&ctx->lock);
2054 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2058 * If the event is in error state, clear that first.
2059 * That way, if we see the event in error state below, we
2060 * know that it has gone back into error state, as distinct
2061 * from the task having been scheduled away before the
2062 * cross-call arrived.
2064 if (event->state == PERF_EVENT_STATE_ERROR)
2065 event->state = PERF_EVENT_STATE_OFF;
2068 if (!ctx->is_active) {
2069 __perf_event_mark_enabled(event);
2073 raw_spin_unlock_irq(&ctx->lock);
2075 if (!task_function_call(task, __perf_event_enable, event))
2078 raw_spin_lock_irq(&ctx->lock);
2081 * If the context is active and the event is still off,
2082 * we need to retry the cross-call.
2084 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2086 * task could have been flipped by a concurrent
2087 * perf_event_context_sched_out()
2094 raw_spin_unlock_irq(&ctx->lock);
2096 EXPORT_SYMBOL_GPL(perf_event_enable);
2098 int perf_event_refresh(struct perf_event *event, int refresh)
2101 * not supported on inherited events
2103 if (event->attr.inherit || !is_sampling_event(event))
2106 atomic_add(refresh, &event->event_limit);
2107 perf_event_enable(event);
2111 EXPORT_SYMBOL_GPL(perf_event_refresh);
2113 static void ctx_sched_out(struct perf_event_context *ctx,
2114 struct perf_cpu_context *cpuctx,
2115 enum event_type_t event_type)
2117 struct perf_event *event;
2118 int is_active = ctx->is_active;
2120 ctx->is_active &= ~event_type;
2121 if (likely(!ctx->nr_events))
2124 update_context_time(ctx);
2125 update_cgrp_time_from_cpuctx(cpuctx);
2126 if (!ctx->nr_active)
2129 perf_pmu_disable(ctx->pmu);
2130 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2131 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2132 group_sched_out(event, cpuctx, ctx);
2135 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2136 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2137 group_sched_out(event, cpuctx, ctx);
2139 perf_pmu_enable(ctx->pmu);
2143 * Test whether two contexts are equivalent, i.e. whether they
2144 * have both been cloned from the same version of the same context
2145 * and they both have the same number of enabled events.
2146 * If the number of enabled events is the same, then the set
2147 * of enabled events should be the same, because these are both
2148 * inherited contexts, therefore we can't access individual events
2149 * in them directly with an fd; we can only enable/disable all
2150 * events via prctl, or enable/disable all events in a family
2151 * via ioctl, which will have the same effect on both contexts.
2153 static int context_equiv(struct perf_event_context *ctx1,
2154 struct perf_event_context *ctx2)
2156 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
2157 && ctx1->parent_gen == ctx2->parent_gen
2158 && !ctx1->pin_count && !ctx2->pin_count;
2161 static void __perf_event_sync_stat(struct perf_event *event,
2162 struct perf_event *next_event)
2166 if (!event->attr.inherit_stat)
2170 * Update the event value, we cannot use perf_event_read()
2171 * because we're in the middle of a context switch and have IRQs
2172 * disabled, which upsets smp_call_function_single(), however
2173 * we know the event must be on the current CPU, therefore we
2174 * don't need to use it.
2176 switch (event->state) {
2177 case PERF_EVENT_STATE_ACTIVE:
2178 event->pmu->read(event);
2181 case PERF_EVENT_STATE_INACTIVE:
2182 update_event_times(event);
2190 * In order to keep per-task stats reliable we need to flip the event
2191 * values when we flip the contexts.
2193 value = local64_read(&next_event->count);
2194 value = local64_xchg(&event->count, value);
2195 local64_set(&next_event->count, value);
2197 swap(event->total_time_enabled, next_event->total_time_enabled);
2198 swap(event->total_time_running, next_event->total_time_running);
2201 * Since we swizzled the values, update the user visible data too.
2203 perf_event_update_userpage(event);
2204 perf_event_update_userpage(next_event);
2207 #define list_next_entry(pos, member) \
2208 list_entry(pos->member.next, typeof(*pos), member)
2210 static void perf_event_sync_stat(struct perf_event_context *ctx,
2211 struct perf_event_context *next_ctx)
2213 struct perf_event *event, *next_event;
2218 update_context_time(ctx);
2220 event = list_first_entry(&ctx->event_list,
2221 struct perf_event, event_entry);
2223 next_event = list_first_entry(&next_ctx->event_list,
2224 struct perf_event, event_entry);
2226 while (&event->event_entry != &ctx->event_list &&
2227 &next_event->event_entry != &next_ctx->event_list) {
2229 __perf_event_sync_stat(event, next_event);
2231 event = list_next_entry(event, event_entry);
2232 next_event = list_next_entry(next_event, event_entry);
2236 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2237 struct task_struct *next)
2239 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2240 struct perf_event_context *next_ctx;
2241 struct perf_event_context *parent;
2242 struct perf_cpu_context *cpuctx;
2248 cpuctx = __get_cpu_context(ctx);
2249 if (!cpuctx->task_ctx)
2253 parent = rcu_dereference(ctx->parent_ctx);
2254 next_ctx = next->perf_event_ctxp[ctxn];
2255 if (parent && next_ctx &&
2256 rcu_dereference(next_ctx->parent_ctx) == parent) {
2258 * Looks like the two contexts are clones, so we might be
2259 * able to optimize the context switch. We lock both
2260 * contexts and check that they are clones under the
2261 * lock (including re-checking that neither has been
2262 * uncloned in the meantime). It doesn't matter which
2263 * order we take the locks because no other cpu could
2264 * be trying to lock both of these tasks.
2266 raw_spin_lock(&ctx->lock);
2267 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2268 if (context_equiv(ctx, next_ctx)) {
2270 * XXX do we need a memory barrier of sorts
2271 * wrt to rcu_dereference() of perf_event_ctxp
2273 task->perf_event_ctxp[ctxn] = next_ctx;
2274 next->perf_event_ctxp[ctxn] = ctx;
2276 next_ctx->task = task;
2279 perf_event_sync_stat(ctx, next_ctx);
2281 raw_spin_unlock(&next_ctx->lock);
2282 raw_spin_unlock(&ctx->lock);
2287 raw_spin_lock(&ctx->lock);
2288 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2289 cpuctx->task_ctx = NULL;
2290 raw_spin_unlock(&ctx->lock);
2294 #define for_each_task_context_nr(ctxn) \
2295 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2298 * Called from scheduler to remove the events of the current task,
2299 * with interrupts disabled.
2301 * We stop each event and update the event value in event->count.
2303 * This does not protect us against NMI, but disable()
2304 * sets the disabled bit in the control field of event _before_
2305 * accessing the event control register. If a NMI hits, then it will
2306 * not restart the event.
2308 void __perf_event_task_sched_out(struct task_struct *task,
2309 struct task_struct *next)
2313 for_each_task_context_nr(ctxn)
2314 perf_event_context_sched_out(task, ctxn, next);
2317 * if cgroup events exist on this CPU, then we need
2318 * to check if we have to switch out PMU state.
2319 * cgroup event are system-wide mode only
2321 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2322 perf_cgroup_sched_out(task, next);
2325 static void task_ctx_sched_out(struct perf_event_context *ctx)
2327 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2329 if (!cpuctx->task_ctx)
2332 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2335 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2336 cpuctx->task_ctx = NULL;
2340 * Called with IRQs disabled
2342 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2343 enum event_type_t event_type)
2345 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2349 ctx_pinned_sched_in(struct perf_event_context *ctx,
2350 struct perf_cpu_context *cpuctx)
2352 struct perf_event *event;
2354 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2355 if (event->state <= PERF_EVENT_STATE_OFF)
2357 if (!event_filter_match(event))
2360 /* may need to reset tstamp_enabled */
2361 if (is_cgroup_event(event))
2362 perf_cgroup_mark_enabled(event, ctx);
2364 if (group_can_go_on(event, cpuctx, 1))
2365 group_sched_in(event, cpuctx, ctx);
2368 * If this pinned group hasn't been scheduled,
2369 * put it in error state.
2371 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2372 update_group_times(event);
2373 event->state = PERF_EVENT_STATE_ERROR;
2379 ctx_flexible_sched_in(struct perf_event_context *ctx,
2380 struct perf_cpu_context *cpuctx)
2382 struct perf_event *event;
2385 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2386 /* Ignore events in OFF or ERROR state */
2387 if (event->state <= PERF_EVENT_STATE_OFF)
2390 * Listen to the 'cpu' scheduling filter constraint
2393 if (!event_filter_match(event))
2396 /* may need to reset tstamp_enabled */
2397 if (is_cgroup_event(event))
2398 perf_cgroup_mark_enabled(event, ctx);
2400 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2401 if (group_sched_in(event, cpuctx, ctx))
2408 ctx_sched_in(struct perf_event_context *ctx,
2409 struct perf_cpu_context *cpuctx,
2410 enum event_type_t event_type,
2411 struct task_struct *task)
2414 int is_active = ctx->is_active;
2416 ctx->is_active |= event_type;
2417 if (likely(!ctx->nr_events))
2421 ctx->timestamp = now;
2422 perf_cgroup_set_timestamp(task, ctx);
2424 * First go through the list and put on any pinned groups
2425 * in order to give them the best chance of going on.
2427 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2428 ctx_pinned_sched_in(ctx, cpuctx);
2430 /* Then walk through the lower prio flexible groups */
2431 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2432 ctx_flexible_sched_in(ctx, cpuctx);
2435 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2436 enum event_type_t event_type,
2437 struct task_struct *task)
2439 struct perf_event_context *ctx = &cpuctx->ctx;
2441 ctx_sched_in(ctx, cpuctx, event_type, task);
2444 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2445 struct task_struct *task)
2447 struct perf_cpu_context *cpuctx;
2449 cpuctx = __get_cpu_context(ctx);
2450 if (cpuctx->task_ctx == ctx)
2453 perf_ctx_lock(cpuctx, ctx);
2454 perf_pmu_disable(ctx->pmu);
2456 * We want to keep the following priority order:
2457 * cpu pinned (that don't need to move), task pinned,
2458 * cpu flexible, task flexible.
2460 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2463 cpuctx->task_ctx = ctx;
2465 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2467 perf_pmu_enable(ctx->pmu);
2468 perf_ctx_unlock(cpuctx, ctx);
2471 * Since these rotations are per-cpu, we need to ensure the
2472 * cpu-context we got scheduled on is actually rotating.
2474 perf_pmu_rotate_start(ctx->pmu);
2478 * When sampling the branck stack in system-wide, it may be necessary
2479 * to flush the stack on context switch. This happens when the branch
2480 * stack does not tag its entries with the pid of the current task.
2481 * Otherwise it becomes impossible to associate a branch entry with a
2482 * task. This ambiguity is more likely to appear when the branch stack
2483 * supports priv level filtering and the user sets it to monitor only
2484 * at the user level (which could be a useful measurement in system-wide
2485 * mode). In that case, the risk is high of having a branch stack with
2486 * branch from multiple tasks. Flushing may mean dropping the existing
2487 * entries or stashing them somewhere in the PMU specific code layer.
2489 * This function provides the context switch callback to the lower code
2490 * layer. It is invoked ONLY when there is at least one system-wide context
2491 * with at least one active event using taken branch sampling.
2493 static void perf_branch_stack_sched_in(struct task_struct *prev,
2494 struct task_struct *task)
2496 struct perf_cpu_context *cpuctx;
2498 unsigned long flags;
2500 /* no need to flush branch stack if not changing task */
2504 local_irq_save(flags);
2508 list_for_each_entry_rcu(pmu, &pmus, entry) {
2509 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2512 * check if the context has at least one
2513 * event using PERF_SAMPLE_BRANCH_STACK
2515 if (cpuctx->ctx.nr_branch_stack > 0
2516 && pmu->flush_branch_stack) {
2518 pmu = cpuctx->ctx.pmu;
2520 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2522 perf_pmu_disable(pmu);
2524 pmu->flush_branch_stack();
2526 perf_pmu_enable(pmu);
2528 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2534 local_irq_restore(flags);
2538 * Called from scheduler to add the events of the current task
2539 * with interrupts disabled.
2541 * We restore the event value and then enable it.
2543 * This does not protect us against NMI, but enable()
2544 * sets the enabled bit in the control field of event _before_
2545 * accessing the event control register. If a NMI hits, then it will
2546 * keep the event running.
2548 void __perf_event_task_sched_in(struct task_struct *prev,
2549 struct task_struct *task)
2551 struct perf_event_context *ctx;
2554 for_each_task_context_nr(ctxn) {
2555 ctx = task->perf_event_ctxp[ctxn];
2559 perf_event_context_sched_in(ctx, task);
2562 * if cgroup events exist on this CPU, then we need
2563 * to check if we have to switch in PMU state.
2564 * cgroup event are system-wide mode only
2566 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2567 perf_cgroup_sched_in(prev, task);
2569 /* check for system-wide branch_stack events */
2570 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2571 perf_branch_stack_sched_in(prev, task);
2574 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2576 u64 frequency = event->attr.sample_freq;
2577 u64 sec = NSEC_PER_SEC;
2578 u64 divisor, dividend;
2580 int count_fls, nsec_fls, frequency_fls, sec_fls;
2582 count_fls = fls64(count);
2583 nsec_fls = fls64(nsec);
2584 frequency_fls = fls64(frequency);
2588 * We got @count in @nsec, with a target of sample_freq HZ
2589 * the target period becomes:
2592 * period = -------------------
2593 * @nsec * sample_freq
2598 * Reduce accuracy by one bit such that @a and @b converge
2599 * to a similar magnitude.
2601 #define REDUCE_FLS(a, b) \
2603 if (a##_fls > b##_fls) { \
2613 * Reduce accuracy until either term fits in a u64, then proceed with
2614 * the other, so that finally we can do a u64/u64 division.
2616 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2617 REDUCE_FLS(nsec, frequency);
2618 REDUCE_FLS(sec, count);
2621 if (count_fls + sec_fls > 64) {
2622 divisor = nsec * frequency;
2624 while (count_fls + sec_fls > 64) {
2625 REDUCE_FLS(count, sec);
2629 dividend = count * sec;
2631 dividend = count * sec;
2633 while (nsec_fls + frequency_fls > 64) {
2634 REDUCE_FLS(nsec, frequency);
2638 divisor = nsec * frequency;
2644 return div64_u64(dividend, divisor);
2647 static DEFINE_PER_CPU(int, perf_throttled_count);
2648 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2650 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2652 struct hw_perf_event *hwc = &event->hw;
2653 s64 period, sample_period;
2656 period = perf_calculate_period(event, nsec, count);
2658 delta = (s64)(period - hwc->sample_period);
2659 delta = (delta + 7) / 8; /* low pass filter */
2661 sample_period = hwc->sample_period + delta;
2666 hwc->sample_period = sample_period;
2668 if (local64_read(&hwc->period_left) > 8*sample_period) {
2670 event->pmu->stop(event, PERF_EF_UPDATE);
2672 local64_set(&hwc->period_left, 0);
2675 event->pmu->start(event, PERF_EF_RELOAD);
2680 * combine freq adjustment with unthrottling to avoid two passes over the
2681 * events. At the same time, make sure, having freq events does not change
2682 * the rate of unthrottling as that would introduce bias.
2684 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2687 struct perf_event *event;
2688 struct hw_perf_event *hwc;
2689 u64 now, period = TICK_NSEC;
2693 * only need to iterate over all events iff:
2694 * - context have events in frequency mode (needs freq adjust)
2695 * - there are events to unthrottle on this cpu
2697 if (!(ctx->nr_freq || needs_unthr))
2700 raw_spin_lock(&ctx->lock);
2701 perf_pmu_disable(ctx->pmu);
2703 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2704 if (event->state != PERF_EVENT_STATE_ACTIVE)
2707 if (!event_filter_match(event))
2712 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2713 hwc->interrupts = 0;
2714 perf_log_throttle(event, 1);
2715 event->pmu->start(event, 0);
2718 if (!event->attr.freq || !event->attr.sample_freq)
2722 * stop the event and update event->count
2724 event->pmu->stop(event, PERF_EF_UPDATE);
2726 now = local64_read(&event->count);
2727 delta = now - hwc->freq_count_stamp;
2728 hwc->freq_count_stamp = now;
2732 * reload only if value has changed
2733 * we have stopped the event so tell that
2734 * to perf_adjust_period() to avoid stopping it
2738 perf_adjust_period(event, period, delta, false);
2740 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2743 perf_pmu_enable(ctx->pmu);
2744 raw_spin_unlock(&ctx->lock);
2748 * Round-robin a context's events:
2750 static void rotate_ctx(struct perf_event_context *ctx)
2753 * Rotate the first entry last of non-pinned groups. Rotation might be
2754 * disabled by the inheritance code.
2756 if (!ctx->rotate_disable)
2757 list_rotate_left(&ctx->flexible_groups);
2761 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2762 * because they're strictly cpu affine and rotate_start is called with IRQs
2763 * disabled, while rotate_context is called from IRQ context.
2765 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2767 struct perf_event_context *ctx = NULL;
2768 int rotate = 0, remove = 1;
2770 if (cpuctx->ctx.nr_events) {
2772 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2776 ctx = cpuctx->task_ctx;
2777 if (ctx && ctx->nr_events) {
2779 if (ctx->nr_events != ctx->nr_active)
2786 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2787 perf_pmu_disable(cpuctx->ctx.pmu);
2789 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2791 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2793 rotate_ctx(&cpuctx->ctx);
2797 perf_event_sched_in(cpuctx, ctx, current);
2799 perf_pmu_enable(cpuctx->ctx.pmu);
2800 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2803 list_del_init(&cpuctx->rotation_list);
2808 #ifdef CONFIG_NO_HZ_FULL
2809 bool perf_event_can_stop_tick(void)
2811 if (atomic_read(&nr_freq_events) ||
2812 __this_cpu_read(perf_throttled_count))
2819 void perf_event_task_tick(void)
2821 struct list_head *head = &__get_cpu_var(rotation_list);
2822 struct perf_cpu_context *cpuctx, *tmp;
2823 struct perf_event_context *ctx;
2826 WARN_ON(!irqs_disabled());
2828 __this_cpu_inc(perf_throttled_seq);
2829 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2831 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2833 perf_adjust_freq_unthr_context(ctx, throttled);
2835 ctx = cpuctx->task_ctx;
2837 perf_adjust_freq_unthr_context(ctx, throttled);
2841 static int event_enable_on_exec(struct perf_event *event,
2842 struct perf_event_context *ctx)
2844 if (!event->attr.enable_on_exec)
2847 event->attr.enable_on_exec = 0;
2848 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2851 __perf_event_mark_enabled(event);
2857 * Enable all of a task's events that have been marked enable-on-exec.
2858 * This expects task == current.
2860 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2862 struct perf_event *event;
2863 unsigned long flags;
2867 local_irq_save(flags);
2868 if (!ctx || !ctx->nr_events)
2872 * We must ctxsw out cgroup events to avoid conflict
2873 * when invoking perf_task_event_sched_in() later on
2874 * in this function. Otherwise we end up trying to
2875 * ctxswin cgroup events which are already scheduled
2878 perf_cgroup_sched_out(current, NULL);
2880 raw_spin_lock(&ctx->lock);
2881 task_ctx_sched_out(ctx);
2883 list_for_each_entry(event, &ctx->event_list, event_entry) {
2884 ret = event_enable_on_exec(event, ctx);
2890 * Unclone this context if we enabled any event.
2895 raw_spin_unlock(&ctx->lock);
2898 * Also calls ctxswin for cgroup events, if any:
2900 perf_event_context_sched_in(ctx, ctx->task);
2902 local_irq_restore(flags);
2906 * Cross CPU call to read the hardware event
2908 static void __perf_event_read(void *info)
2910 struct perf_event *event = info;
2911 struct perf_event_context *ctx = event->ctx;
2912 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2915 * If this is a task context, we need to check whether it is
2916 * the current task context of this cpu. If not it has been
2917 * scheduled out before the smp call arrived. In that case
2918 * event->count would have been updated to a recent sample
2919 * when the event was scheduled out.
2921 if (ctx->task && cpuctx->task_ctx != ctx)
2924 raw_spin_lock(&ctx->lock);
2925 if (ctx->is_active) {
2926 update_context_time(ctx);
2927 update_cgrp_time_from_event(event);
2929 update_event_times(event);
2930 if (event->state == PERF_EVENT_STATE_ACTIVE)
2931 event->pmu->read(event);
2932 raw_spin_unlock(&ctx->lock);
2935 static inline u64 perf_event_count(struct perf_event *event)
2937 return local64_read(&event->count) + atomic64_read(&event->child_count);
2940 static u64 perf_event_read(struct perf_event *event)
2943 * If event is enabled and currently active on a CPU, update the
2944 * value in the event structure:
2946 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2947 smp_call_function_single(event->oncpu,
2948 __perf_event_read, event, 1);
2949 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2950 struct perf_event_context *ctx = event->ctx;
2951 unsigned long flags;
2953 raw_spin_lock_irqsave(&ctx->lock, flags);
2955 * may read while context is not active
2956 * (e.g., thread is blocked), in that case
2957 * we cannot update context time
2959 if (ctx->is_active) {
2960 update_context_time(ctx);
2961 update_cgrp_time_from_event(event);
2963 update_event_times(event);
2964 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2967 return perf_event_count(event);
2971 * Initialize the perf_event context in a task_struct:
2973 static void __perf_event_init_context(struct perf_event_context *ctx)
2975 raw_spin_lock_init(&ctx->lock);
2976 mutex_init(&ctx->mutex);
2977 INIT_LIST_HEAD(&ctx->pinned_groups);
2978 INIT_LIST_HEAD(&ctx->flexible_groups);
2979 INIT_LIST_HEAD(&ctx->event_list);
2980 atomic_set(&ctx->refcount, 1);
2983 static struct perf_event_context *
2984 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2986 struct perf_event_context *ctx;
2988 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2992 __perf_event_init_context(ctx);
2995 get_task_struct(task);
3002 static struct task_struct *
3003 find_lively_task_by_vpid(pid_t vpid)
3005 struct task_struct *task;
3012 task = find_task_by_vpid(vpid);
3014 get_task_struct(task);
3018 return ERR_PTR(-ESRCH);
3020 /* Reuse ptrace permission checks for now. */
3022 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3027 put_task_struct(task);
3028 return ERR_PTR(err);
3033 * Returns a matching context with refcount and pincount.
3035 static struct perf_event_context *
3036 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3038 struct perf_event_context *ctx;
3039 struct perf_cpu_context *cpuctx;
3040 unsigned long flags;
3044 /* Must be root to operate on a CPU event: */
3045 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3046 return ERR_PTR(-EACCES);
3049 * We could be clever and allow to attach a event to an
3050 * offline CPU and activate it when the CPU comes up, but
3053 if (!cpu_online(cpu))
3054 return ERR_PTR(-ENODEV);
3056 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3065 ctxn = pmu->task_ctx_nr;
3070 ctx = perf_lock_task_context(task, ctxn, &flags);
3074 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3076 ctx = alloc_perf_context(pmu, task);
3082 mutex_lock(&task->perf_event_mutex);
3084 * If it has already passed perf_event_exit_task().
3085 * we must see PF_EXITING, it takes this mutex too.
3087 if (task->flags & PF_EXITING)
3089 else if (task->perf_event_ctxp[ctxn])
3094 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3096 mutex_unlock(&task->perf_event_mutex);
3098 if (unlikely(err)) {
3110 return ERR_PTR(err);
3113 static void perf_event_free_filter(struct perf_event *event);
3115 static void free_event_rcu(struct rcu_head *head)
3117 struct perf_event *event;
3119 event = container_of(head, struct perf_event, rcu_head);
3121 put_pid_ns(event->ns);
3122 perf_event_free_filter(event);
3126 static void ring_buffer_put(struct ring_buffer *rb);
3127 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3129 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3134 if (has_branch_stack(event)) {
3135 if (!(event->attach_state & PERF_ATTACH_TASK))
3136 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3138 if (is_cgroup_event(event))
3139 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3142 static void unaccount_event(struct perf_event *event)
3147 if (event->attach_state & PERF_ATTACH_TASK)
3148 static_key_slow_dec_deferred(&perf_sched_events);
3149 if (event->attr.mmap || event->attr.mmap_data)
3150 atomic_dec(&nr_mmap_events);
3151 if (event->attr.comm)
3152 atomic_dec(&nr_comm_events);
3153 if (event->attr.task)
3154 atomic_dec(&nr_task_events);
3155 if (event->attr.freq)
3156 atomic_dec(&nr_freq_events);
3157 if (is_cgroup_event(event))
3158 static_key_slow_dec_deferred(&perf_sched_events);
3159 if (has_branch_stack(event))
3160 static_key_slow_dec_deferred(&perf_sched_events);
3162 unaccount_event_cpu(event, event->cpu);
3165 static void __free_event(struct perf_event *event)
3167 if (!event->parent) {
3168 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3169 put_callchain_buffers();
3173 event->destroy(event);
3176 put_ctx(event->ctx);
3178 call_rcu(&event->rcu_head, free_event_rcu);
3180 static void free_event(struct perf_event *event)
3182 irq_work_sync(&event->pending);
3184 unaccount_event(event);
3187 struct ring_buffer *rb;
3190 * Can happen when we close an event with re-directed output.
3192 * Since we have a 0 refcount, perf_mmap_close() will skip
3193 * over us; possibly making our ring_buffer_put() the last.
3195 mutex_lock(&event->mmap_mutex);
3198 rcu_assign_pointer(event->rb, NULL);
3199 ring_buffer_detach(event, rb);
3200 ring_buffer_put(rb); /* could be last */
3202 mutex_unlock(&event->mmap_mutex);
3205 if (is_cgroup_event(event))
3206 perf_detach_cgroup(event);
3209 __free_event(event);
3212 int perf_event_release_kernel(struct perf_event *event)
3214 struct perf_event_context *ctx = event->ctx;
3216 WARN_ON_ONCE(ctx->parent_ctx);
3218 * There are two ways this annotation is useful:
3220 * 1) there is a lock recursion from perf_event_exit_task
3221 * see the comment there.
3223 * 2) there is a lock-inversion with mmap_sem through
3224 * perf_event_read_group(), which takes faults while
3225 * holding ctx->mutex, however this is called after
3226 * the last filedesc died, so there is no possibility
3227 * to trigger the AB-BA case.
3229 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3230 raw_spin_lock_irq(&ctx->lock);
3231 perf_group_detach(event);
3232 raw_spin_unlock_irq(&ctx->lock);
3233 perf_remove_from_context(event);
3234 mutex_unlock(&ctx->mutex);
3240 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3243 * Called when the last reference to the file is gone.
3245 static void put_event(struct perf_event *event)
3247 struct task_struct *owner;
3249 if (!atomic_long_dec_and_test(&event->refcount))
3253 owner = ACCESS_ONCE(event->owner);
3255 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3256 * !owner it means the list deletion is complete and we can indeed
3257 * free this event, otherwise we need to serialize on
3258 * owner->perf_event_mutex.
3260 smp_read_barrier_depends();
3263 * Since delayed_put_task_struct() also drops the last
3264 * task reference we can safely take a new reference
3265 * while holding the rcu_read_lock().
3267 get_task_struct(owner);
3272 mutex_lock(&owner->perf_event_mutex);
3274 * We have to re-check the event->owner field, if it is cleared
3275 * we raced with perf_event_exit_task(), acquiring the mutex
3276 * ensured they're done, and we can proceed with freeing the
3280 list_del_init(&event->owner_entry);
3281 mutex_unlock(&owner->perf_event_mutex);
3282 put_task_struct(owner);
3285 perf_event_release_kernel(event);
3288 static int perf_release(struct inode *inode, struct file *file)
3290 put_event(file->private_data);
3294 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3296 struct perf_event *child;
3302 mutex_lock(&event->child_mutex);
3303 total += perf_event_read(event);
3304 *enabled += event->total_time_enabled +
3305 atomic64_read(&event->child_total_time_enabled);
3306 *running += event->total_time_running +
3307 atomic64_read(&event->child_total_time_running);
3309 list_for_each_entry(child, &event->child_list, child_list) {
3310 total += perf_event_read(child);
3311 *enabled += child->total_time_enabled;
3312 *running += child->total_time_running;
3314 mutex_unlock(&event->child_mutex);
3318 EXPORT_SYMBOL_GPL(perf_event_read_value);
3320 static int perf_event_read_group(struct perf_event *event,
3321 u64 read_format, char __user *buf)
3323 struct perf_event *leader = event->group_leader, *sub;
3324 int n = 0, size = 0, ret = -EFAULT;
3325 struct perf_event_context *ctx = leader->ctx;
3327 u64 count, enabled, running;
3329 mutex_lock(&ctx->mutex);
3330 count = perf_event_read_value(leader, &enabled, &running);
3332 values[n++] = 1 + leader->nr_siblings;
3333 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3334 values[n++] = enabled;
3335 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3336 values[n++] = running;
3337 values[n++] = count;
3338 if (read_format & PERF_FORMAT_ID)
3339 values[n++] = primary_event_id(leader);
3341 size = n * sizeof(u64);
3343 if (copy_to_user(buf, values, size))
3348 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3351 values[n++] = perf_event_read_value(sub, &enabled, &running);
3352 if (read_format & PERF_FORMAT_ID)
3353 values[n++] = primary_event_id(sub);
3355 size = n * sizeof(u64);
3357 if (copy_to_user(buf + ret, values, size)) {
3365 mutex_unlock(&ctx->mutex);
3370 static int perf_event_read_one(struct perf_event *event,
3371 u64 read_format, char __user *buf)
3373 u64 enabled, running;
3377 values[n++] = perf_event_read_value(event, &enabled, &running);
3378 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3379 values[n++] = enabled;
3380 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3381 values[n++] = running;
3382 if (read_format & PERF_FORMAT_ID)
3383 values[n++] = primary_event_id(event);
3385 if (copy_to_user(buf, values, n * sizeof(u64)))
3388 return n * sizeof(u64);
3392 * Read the performance event - simple non blocking version for now
3395 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3397 u64 read_format = event->attr.read_format;
3401 * Return end-of-file for a read on a event that is in
3402 * error state (i.e. because it was pinned but it couldn't be
3403 * scheduled on to the CPU at some point).
3405 if (event->state == PERF_EVENT_STATE_ERROR)
3408 if (count < event->read_size)
3411 WARN_ON_ONCE(event->ctx->parent_ctx);
3412 if (read_format & PERF_FORMAT_GROUP)
3413 ret = perf_event_read_group(event, read_format, buf);
3415 ret = perf_event_read_one(event, read_format, buf);
3421 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3423 struct perf_event *event = file->private_data;
3425 return perf_read_hw(event, buf, count);
3428 static unsigned int perf_poll(struct file *file, poll_table *wait)
3430 struct perf_event *event = file->private_data;
3431 struct ring_buffer *rb;
3432 unsigned int events = POLL_HUP;
3435 * Pin the event->rb by taking event->mmap_mutex; otherwise
3436 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3438 mutex_lock(&event->mmap_mutex);
3441 events = atomic_xchg(&rb->poll, 0);
3442 mutex_unlock(&event->mmap_mutex);
3444 poll_wait(file, &event->waitq, wait);
3449 static void perf_event_reset(struct perf_event *event)
3451 (void)perf_event_read(event);
3452 local64_set(&event->count, 0);
3453 perf_event_update_userpage(event);
3457 * Holding the top-level event's child_mutex means that any
3458 * descendant process that has inherited this event will block
3459 * in sync_child_event if it goes to exit, thus satisfying the
3460 * task existence requirements of perf_event_enable/disable.
3462 static void perf_event_for_each_child(struct perf_event *event,
3463 void (*func)(struct perf_event *))
3465 struct perf_event *child;
3467 WARN_ON_ONCE(event->ctx->parent_ctx);
3468 mutex_lock(&event->child_mutex);
3470 list_for_each_entry(child, &event->child_list, child_list)
3472 mutex_unlock(&event->child_mutex);
3475 static void perf_event_for_each(struct perf_event *event,
3476 void (*func)(struct perf_event *))
3478 struct perf_event_context *ctx = event->ctx;
3479 struct perf_event *sibling;
3481 WARN_ON_ONCE(ctx->parent_ctx);
3482 mutex_lock(&ctx->mutex);
3483 event = event->group_leader;
3485 perf_event_for_each_child(event, func);
3486 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3487 perf_event_for_each_child(sibling, func);
3488 mutex_unlock(&ctx->mutex);
3491 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3493 struct perf_event_context *ctx = event->ctx;
3497 if (!is_sampling_event(event))
3500 if (copy_from_user(&value, arg, sizeof(value)))
3506 raw_spin_lock_irq(&ctx->lock);
3507 if (event->attr.freq) {
3508 if (value > sysctl_perf_event_sample_rate) {
3513 event->attr.sample_freq = value;
3515 event->attr.sample_period = value;
3516 event->hw.sample_period = value;
3519 raw_spin_unlock_irq(&ctx->lock);
3524 static const struct file_operations perf_fops;
3526 static inline int perf_fget_light(int fd, struct fd *p)
3528 struct fd f = fdget(fd);
3532 if (f.file->f_op != &perf_fops) {
3540 static int perf_event_set_output(struct perf_event *event,
3541 struct perf_event *output_event);
3542 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3544 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3546 struct perf_event *event = file->private_data;
3547 void (*func)(struct perf_event *);
3551 case PERF_EVENT_IOC_ENABLE:
3552 func = perf_event_enable;
3554 case PERF_EVENT_IOC_DISABLE:
3555 func = perf_event_disable;
3557 case PERF_EVENT_IOC_RESET:
3558 func = perf_event_reset;
3561 case PERF_EVENT_IOC_REFRESH:
3562 return perf_event_refresh(event, arg);
3564 case PERF_EVENT_IOC_PERIOD:
3565 return perf_event_period(event, (u64 __user *)arg);
3567 case PERF_EVENT_IOC_ID:
3569 u64 id = primary_event_id(event);
3571 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3576 case PERF_EVENT_IOC_SET_OUTPUT:
3580 struct perf_event *output_event;
3582 ret = perf_fget_light(arg, &output);
3585 output_event = output.file->private_data;
3586 ret = perf_event_set_output(event, output_event);
3589 ret = perf_event_set_output(event, NULL);
3594 case PERF_EVENT_IOC_SET_FILTER:
3595 return perf_event_set_filter(event, (void __user *)arg);
3601 if (flags & PERF_IOC_FLAG_GROUP)
3602 perf_event_for_each(event, func);
3604 perf_event_for_each_child(event, func);
3609 int perf_event_task_enable(void)
3611 struct perf_event *event;
3613 mutex_lock(¤t->perf_event_mutex);
3614 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3615 perf_event_for_each_child(event, perf_event_enable);
3616 mutex_unlock(¤t->perf_event_mutex);
3621 int perf_event_task_disable(void)
3623 struct perf_event *event;
3625 mutex_lock(¤t->perf_event_mutex);
3626 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3627 perf_event_for_each_child(event, perf_event_disable);
3628 mutex_unlock(¤t->perf_event_mutex);
3633 static int perf_event_index(struct perf_event *event)
3635 if (event->hw.state & PERF_HES_STOPPED)
3638 if (event->state != PERF_EVENT_STATE_ACTIVE)
3641 return event->pmu->event_idx(event);
3644 static void calc_timer_values(struct perf_event *event,
3651 *now = perf_clock();
3652 ctx_time = event->shadow_ctx_time + *now;
3653 *enabled = ctx_time - event->tstamp_enabled;
3654 *running = ctx_time - event->tstamp_running;
3657 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3662 * Callers need to ensure there can be no nesting of this function, otherwise
3663 * the seqlock logic goes bad. We can not serialize this because the arch
3664 * code calls this from NMI context.
3666 void perf_event_update_userpage(struct perf_event *event)
3668 struct perf_event_mmap_page *userpg;
3669 struct ring_buffer *rb;
3670 u64 enabled, running, now;
3674 * compute total_time_enabled, total_time_running
3675 * based on snapshot values taken when the event
3676 * was last scheduled in.
3678 * we cannot simply called update_context_time()
3679 * because of locking issue as we can be called in
3682 calc_timer_values(event, &now, &enabled, &running);
3683 rb = rcu_dereference(event->rb);
3687 userpg = rb->user_page;
3690 * Disable preemption so as to not let the corresponding user-space
3691 * spin too long if we get preempted.
3696 userpg->index = perf_event_index(event);
3697 userpg->offset = perf_event_count(event);
3699 userpg->offset -= local64_read(&event->hw.prev_count);
3701 userpg->time_enabled = enabled +
3702 atomic64_read(&event->child_total_time_enabled);
3704 userpg->time_running = running +
3705 atomic64_read(&event->child_total_time_running);
3707 arch_perf_update_userpage(userpg, now);
3716 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3718 struct perf_event *event = vma->vm_file->private_data;
3719 struct ring_buffer *rb;
3720 int ret = VM_FAULT_SIGBUS;
3722 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3723 if (vmf->pgoff == 0)
3729 rb = rcu_dereference(event->rb);
3733 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3736 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3740 get_page(vmf->page);
3741 vmf->page->mapping = vma->vm_file->f_mapping;
3742 vmf->page->index = vmf->pgoff;
3751 static void ring_buffer_attach(struct perf_event *event,
3752 struct ring_buffer *rb)
3754 unsigned long flags;
3756 if (!list_empty(&event->rb_entry))
3759 spin_lock_irqsave(&rb->event_lock, flags);
3760 if (list_empty(&event->rb_entry))
3761 list_add(&event->rb_entry, &rb->event_list);
3762 spin_unlock_irqrestore(&rb->event_lock, flags);
3765 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3767 unsigned long flags;
3769 if (list_empty(&event->rb_entry))
3772 spin_lock_irqsave(&rb->event_lock, flags);
3773 list_del_init(&event->rb_entry);
3774 wake_up_all(&event->waitq);
3775 spin_unlock_irqrestore(&rb->event_lock, flags);
3778 static void ring_buffer_wakeup(struct perf_event *event)
3780 struct ring_buffer *rb;
3783 rb = rcu_dereference(event->rb);
3785 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3786 wake_up_all(&event->waitq);
3791 static void rb_free_rcu(struct rcu_head *rcu_head)
3793 struct ring_buffer *rb;
3795 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3799 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3801 struct ring_buffer *rb;
3804 rb = rcu_dereference(event->rb);
3806 if (!atomic_inc_not_zero(&rb->refcount))
3814 static void ring_buffer_put(struct ring_buffer *rb)
3816 if (!atomic_dec_and_test(&rb->refcount))
3819 WARN_ON_ONCE(!list_empty(&rb->event_list));
3821 call_rcu(&rb->rcu_head, rb_free_rcu);
3824 static void perf_mmap_open(struct vm_area_struct *vma)
3826 struct perf_event *event = vma->vm_file->private_data;
3828 atomic_inc(&event->mmap_count);
3829 atomic_inc(&event->rb->mmap_count);
3833 * A buffer can be mmap()ed multiple times; either directly through the same
3834 * event, or through other events by use of perf_event_set_output().
3836 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3837 * the buffer here, where we still have a VM context. This means we need
3838 * to detach all events redirecting to us.
3840 static void perf_mmap_close(struct vm_area_struct *vma)
3842 struct perf_event *event = vma->vm_file->private_data;
3844 struct ring_buffer *rb = event->rb;
3845 struct user_struct *mmap_user = rb->mmap_user;
3846 int mmap_locked = rb->mmap_locked;
3847 unsigned long size = perf_data_size(rb);
3849 atomic_dec(&rb->mmap_count);
3851 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3854 /* Detach current event from the buffer. */
3855 rcu_assign_pointer(event->rb, NULL);
3856 ring_buffer_detach(event, rb);
3857 mutex_unlock(&event->mmap_mutex);
3859 /* If there's still other mmap()s of this buffer, we're done. */
3860 if (atomic_read(&rb->mmap_count)) {
3861 ring_buffer_put(rb); /* can't be last */
3866 * No other mmap()s, detach from all other events that might redirect
3867 * into the now unreachable buffer. Somewhat complicated by the
3868 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3872 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3873 if (!atomic_long_inc_not_zero(&event->refcount)) {
3875 * This event is en-route to free_event() which will
3876 * detach it and remove it from the list.
3882 mutex_lock(&event->mmap_mutex);
3884 * Check we didn't race with perf_event_set_output() which can
3885 * swizzle the rb from under us while we were waiting to
3886 * acquire mmap_mutex.
3888 * If we find a different rb; ignore this event, a next
3889 * iteration will no longer find it on the list. We have to
3890 * still restart the iteration to make sure we're not now
3891 * iterating the wrong list.
3893 if (event->rb == rb) {
3894 rcu_assign_pointer(event->rb, NULL);
3895 ring_buffer_detach(event, rb);
3896 ring_buffer_put(rb); /* can't be last, we still have one */
3898 mutex_unlock(&event->mmap_mutex);
3902 * Restart the iteration; either we're on the wrong list or
3903 * destroyed its integrity by doing a deletion.
3910 * It could be there's still a few 0-ref events on the list; they'll
3911 * get cleaned up by free_event() -- they'll also still have their
3912 * ref on the rb and will free it whenever they are done with it.
3914 * Aside from that, this buffer is 'fully' detached and unmapped,
3915 * undo the VM accounting.
3918 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3919 vma->vm_mm->pinned_vm -= mmap_locked;
3920 free_uid(mmap_user);
3922 ring_buffer_put(rb); /* could be last */
3925 static const struct vm_operations_struct perf_mmap_vmops = {
3926 .open = perf_mmap_open,
3927 .close = perf_mmap_close,
3928 .fault = perf_mmap_fault,
3929 .page_mkwrite = perf_mmap_fault,
3932 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3934 struct perf_event *event = file->private_data;
3935 unsigned long user_locked, user_lock_limit;
3936 struct user_struct *user = current_user();
3937 unsigned long locked, lock_limit;
3938 struct ring_buffer *rb;
3939 unsigned long vma_size;
3940 unsigned long nr_pages;
3941 long user_extra, extra;
3942 int ret = 0, flags = 0;
3945 * Don't allow mmap() of inherited per-task counters. This would
3946 * create a performance issue due to all children writing to the
3949 if (event->cpu == -1 && event->attr.inherit)
3952 if (!(vma->vm_flags & VM_SHARED))
3955 vma_size = vma->vm_end - vma->vm_start;
3956 nr_pages = (vma_size / PAGE_SIZE) - 1;
3959 * If we have rb pages ensure they're a power-of-two number, so we
3960 * can do bitmasks instead of modulo.
3962 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3965 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3968 if (vma->vm_pgoff != 0)
3971 WARN_ON_ONCE(event->ctx->parent_ctx);
3973 mutex_lock(&event->mmap_mutex);
3975 if (event->rb->nr_pages != nr_pages) {
3980 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
3982 * Raced against perf_mmap_close() through
3983 * perf_event_set_output(). Try again, hope for better
3986 mutex_unlock(&event->mmap_mutex);
3993 user_extra = nr_pages + 1;
3994 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3997 * Increase the limit linearly with more CPUs:
3999 user_lock_limit *= num_online_cpus();
4001 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4004 if (user_locked > user_lock_limit)
4005 extra = user_locked - user_lock_limit;
4007 lock_limit = rlimit(RLIMIT_MEMLOCK);
4008 lock_limit >>= PAGE_SHIFT;
4009 locked = vma->vm_mm->pinned_vm + extra;
4011 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4012 !capable(CAP_IPC_LOCK)) {
4019 if (vma->vm_flags & VM_WRITE)
4020 flags |= RING_BUFFER_WRITABLE;
4022 rb = rb_alloc(nr_pages,
4023 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4031 atomic_set(&rb->mmap_count, 1);
4032 rb->mmap_locked = extra;
4033 rb->mmap_user = get_current_user();
4035 atomic_long_add(user_extra, &user->locked_vm);
4036 vma->vm_mm->pinned_vm += extra;
4038 ring_buffer_attach(event, rb);
4039 rcu_assign_pointer(event->rb, rb);
4041 perf_event_update_userpage(event);
4045 atomic_inc(&event->mmap_count);
4046 mutex_unlock(&event->mmap_mutex);
4049 * Since pinned accounting is per vm we cannot allow fork() to copy our
4052 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4053 vma->vm_ops = &perf_mmap_vmops;
4058 static int perf_fasync(int fd, struct file *filp, int on)
4060 struct inode *inode = file_inode(filp);
4061 struct perf_event *event = filp->private_data;
4064 mutex_lock(&inode->i_mutex);
4065 retval = fasync_helper(fd, filp, on, &event->fasync);
4066 mutex_unlock(&inode->i_mutex);
4074 static const struct file_operations perf_fops = {
4075 .llseek = no_llseek,
4076 .release = perf_release,
4079 .unlocked_ioctl = perf_ioctl,
4080 .compat_ioctl = perf_ioctl,
4082 .fasync = perf_fasync,
4088 * If there's data, ensure we set the poll() state and publish everything
4089 * to user-space before waking everybody up.
4092 void perf_event_wakeup(struct perf_event *event)
4094 ring_buffer_wakeup(event);
4096 if (event->pending_kill) {
4097 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4098 event->pending_kill = 0;
4102 static void perf_pending_event(struct irq_work *entry)
4104 struct perf_event *event = container_of(entry,
4105 struct perf_event, pending);
4107 if (event->pending_disable) {
4108 event->pending_disable = 0;
4109 __perf_event_disable(event);
4112 if (event->pending_wakeup) {
4113 event->pending_wakeup = 0;
4114 perf_event_wakeup(event);
4119 * We assume there is only KVM supporting the callbacks.
4120 * Later on, we might change it to a list if there is
4121 * another virtualization implementation supporting the callbacks.
4123 struct perf_guest_info_callbacks *perf_guest_cbs;
4125 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4127 perf_guest_cbs = cbs;
4130 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4132 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4134 perf_guest_cbs = NULL;
4137 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4140 perf_output_sample_regs(struct perf_output_handle *handle,
4141 struct pt_regs *regs, u64 mask)
4145 for_each_set_bit(bit, (const unsigned long *) &mask,
4146 sizeof(mask) * BITS_PER_BYTE) {
4149 val = perf_reg_value(regs, bit);
4150 perf_output_put(handle, val);
4154 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4155 struct pt_regs *regs)
4157 if (!user_mode(regs)) {
4159 regs = task_pt_regs(current);
4165 regs_user->regs = regs;
4166 regs_user->abi = perf_reg_abi(current);
4171 * Get remaining task size from user stack pointer.
4173 * It'd be better to take stack vma map and limit this more
4174 * precisly, but there's no way to get it safely under interrupt,
4175 * so using TASK_SIZE as limit.
4177 static u64 perf_ustack_task_size(struct pt_regs *regs)
4179 unsigned long addr = perf_user_stack_pointer(regs);
4181 if (!addr || addr >= TASK_SIZE)
4184 return TASK_SIZE - addr;
4188 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4189 struct pt_regs *regs)
4193 /* No regs, no stack pointer, no dump. */
4198 * Check if we fit in with the requested stack size into the:
4200 * If we don't, we limit the size to the TASK_SIZE.
4202 * - remaining sample size
4203 * If we don't, we customize the stack size to
4204 * fit in to the remaining sample size.
4207 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4208 stack_size = min(stack_size, (u16) task_size);
4210 /* Current header size plus static size and dynamic size. */
4211 header_size += 2 * sizeof(u64);
4213 /* Do we fit in with the current stack dump size? */
4214 if ((u16) (header_size + stack_size) < header_size) {
4216 * If we overflow the maximum size for the sample,
4217 * we customize the stack dump size to fit in.
4219 stack_size = USHRT_MAX - header_size - sizeof(u64);
4220 stack_size = round_up(stack_size, sizeof(u64));
4227 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4228 struct pt_regs *regs)
4230 /* Case of a kernel thread, nothing to dump */
4233 perf_output_put(handle, size);
4242 * - the size requested by user or the best one we can fit
4243 * in to the sample max size
4245 * - user stack dump data
4247 * - the actual dumped size
4251 perf_output_put(handle, dump_size);
4254 sp = perf_user_stack_pointer(regs);
4255 rem = __output_copy_user(handle, (void *) sp, dump_size);
4256 dyn_size = dump_size - rem;
4258 perf_output_skip(handle, rem);
4261 perf_output_put(handle, dyn_size);
4265 static void __perf_event_header__init_id(struct perf_event_header *header,
4266 struct perf_sample_data *data,
4267 struct perf_event *event)
4269 u64 sample_type = event->attr.sample_type;
4271 data->type = sample_type;
4272 header->size += event->id_header_size;
4274 if (sample_type & PERF_SAMPLE_TID) {
4275 /* namespace issues */
4276 data->tid_entry.pid = perf_event_pid(event, current);
4277 data->tid_entry.tid = perf_event_tid(event, current);
4280 if (sample_type & PERF_SAMPLE_TIME)
4281 data->time = perf_clock();
4283 if (sample_type & PERF_SAMPLE_ID)
4284 data->id = primary_event_id(event);
4286 if (sample_type & PERF_SAMPLE_STREAM_ID)
4287 data->stream_id = event->id;
4289 if (sample_type & PERF_SAMPLE_CPU) {
4290 data->cpu_entry.cpu = raw_smp_processor_id();
4291 data->cpu_entry.reserved = 0;
4295 void perf_event_header__init_id(struct perf_event_header *header,
4296 struct perf_sample_data *data,
4297 struct perf_event *event)
4299 if (event->attr.sample_id_all)
4300 __perf_event_header__init_id(header, data, event);
4303 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4304 struct perf_sample_data *data)
4306 u64 sample_type = data->type;
4308 if (sample_type & PERF_SAMPLE_TID)
4309 perf_output_put(handle, data->tid_entry);
4311 if (sample_type & PERF_SAMPLE_TIME)
4312 perf_output_put(handle, data->time);
4314 if (sample_type & PERF_SAMPLE_ID)
4315 perf_output_put(handle, data->id);
4317 if (sample_type & PERF_SAMPLE_STREAM_ID)
4318 perf_output_put(handle, data->stream_id);
4320 if (sample_type & PERF_SAMPLE_CPU)
4321 perf_output_put(handle, data->cpu_entry);
4324 void perf_event__output_id_sample(struct perf_event *event,
4325 struct perf_output_handle *handle,
4326 struct perf_sample_data *sample)
4328 if (event->attr.sample_id_all)
4329 __perf_event__output_id_sample(handle, sample);
4332 static void perf_output_read_one(struct perf_output_handle *handle,
4333 struct perf_event *event,
4334 u64 enabled, u64 running)
4336 u64 read_format = event->attr.read_format;
4340 values[n++] = perf_event_count(event);
4341 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4342 values[n++] = enabled +
4343 atomic64_read(&event->child_total_time_enabled);
4345 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4346 values[n++] = running +
4347 atomic64_read(&event->child_total_time_running);
4349 if (read_format & PERF_FORMAT_ID)
4350 values[n++] = primary_event_id(event);
4352 __output_copy(handle, values, n * sizeof(u64));
4356 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4358 static void perf_output_read_group(struct perf_output_handle *handle,
4359 struct perf_event *event,
4360 u64 enabled, u64 running)
4362 struct perf_event *leader = event->group_leader, *sub;
4363 u64 read_format = event->attr.read_format;
4367 values[n++] = 1 + leader->nr_siblings;
4369 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4370 values[n++] = enabled;
4372 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4373 values[n++] = running;
4375 if (leader != event)
4376 leader->pmu->read(leader);
4378 values[n++] = perf_event_count(leader);
4379 if (read_format & PERF_FORMAT_ID)
4380 values[n++] = primary_event_id(leader);
4382 __output_copy(handle, values, n * sizeof(u64));
4384 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4387 if ((sub != event) &&
4388 (sub->state == PERF_EVENT_STATE_ACTIVE))
4389 sub->pmu->read(sub);
4391 values[n++] = perf_event_count(sub);
4392 if (read_format & PERF_FORMAT_ID)
4393 values[n++] = primary_event_id(sub);
4395 __output_copy(handle, values, n * sizeof(u64));
4399 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4400 PERF_FORMAT_TOTAL_TIME_RUNNING)
4402 static void perf_output_read(struct perf_output_handle *handle,
4403 struct perf_event *event)
4405 u64 enabled = 0, running = 0, now;
4406 u64 read_format = event->attr.read_format;
4409 * compute total_time_enabled, total_time_running
4410 * based on snapshot values taken when the event
4411 * was last scheduled in.
4413 * we cannot simply called update_context_time()
4414 * because of locking issue as we are called in
4417 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4418 calc_timer_values(event, &now, &enabled, &running);
4420 if (event->attr.read_format & PERF_FORMAT_GROUP)
4421 perf_output_read_group(handle, event, enabled, running);
4423 perf_output_read_one(handle, event, enabled, running);
4426 void perf_output_sample(struct perf_output_handle *handle,
4427 struct perf_event_header *header,
4428 struct perf_sample_data *data,
4429 struct perf_event *event)
4431 u64 sample_type = data->type;
4433 perf_output_put(handle, *header);
4435 if (sample_type & PERF_SAMPLE_IP)
4436 perf_output_put(handle, data->ip);
4438 if (sample_type & PERF_SAMPLE_TID)
4439 perf_output_put(handle, data->tid_entry);
4441 if (sample_type & PERF_SAMPLE_TIME)
4442 perf_output_put(handle, data->time);
4444 if (sample_type & PERF_SAMPLE_ADDR)
4445 perf_output_put(handle, data->addr);
4447 if (sample_type & PERF_SAMPLE_ID)
4448 perf_output_put(handle, data->id);
4450 if (sample_type & PERF_SAMPLE_STREAM_ID)
4451 perf_output_put(handle, data->stream_id);
4453 if (sample_type & PERF_SAMPLE_CPU)
4454 perf_output_put(handle, data->cpu_entry);
4456 if (sample_type & PERF_SAMPLE_PERIOD)
4457 perf_output_put(handle, data->period);
4459 if (sample_type & PERF_SAMPLE_READ)
4460 perf_output_read(handle, event);
4462 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4463 if (data->callchain) {
4466 if (data->callchain)
4467 size += data->callchain->nr;
4469 size *= sizeof(u64);
4471 __output_copy(handle, data->callchain, size);
4474 perf_output_put(handle, nr);
4478 if (sample_type & PERF_SAMPLE_RAW) {
4480 perf_output_put(handle, data->raw->size);
4481 __output_copy(handle, data->raw->data,
4488 .size = sizeof(u32),
4491 perf_output_put(handle, raw);
4495 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4496 if (data->br_stack) {
4499 size = data->br_stack->nr
4500 * sizeof(struct perf_branch_entry);
4502 perf_output_put(handle, data->br_stack->nr);
4503 perf_output_copy(handle, data->br_stack->entries, size);
4506 * we always store at least the value of nr
4509 perf_output_put(handle, nr);
4513 if (sample_type & PERF_SAMPLE_REGS_USER) {
4514 u64 abi = data->regs_user.abi;
4517 * If there are no regs to dump, notice it through
4518 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4520 perf_output_put(handle, abi);
4523 u64 mask = event->attr.sample_regs_user;
4524 perf_output_sample_regs(handle,
4525 data->regs_user.regs,
4530 if (sample_type & PERF_SAMPLE_STACK_USER) {
4531 perf_output_sample_ustack(handle,
4532 data->stack_user_size,
4533 data->regs_user.regs);
4536 if (sample_type & PERF_SAMPLE_WEIGHT)
4537 perf_output_put(handle, data->weight);
4539 if (sample_type & PERF_SAMPLE_DATA_SRC)
4540 perf_output_put(handle, data->data_src.val);
4542 if (!event->attr.watermark) {
4543 int wakeup_events = event->attr.wakeup_events;
4545 if (wakeup_events) {
4546 struct ring_buffer *rb = handle->rb;
4547 int events = local_inc_return(&rb->events);
4549 if (events >= wakeup_events) {
4550 local_sub(wakeup_events, &rb->events);
4551 local_inc(&rb->wakeup);
4557 void perf_prepare_sample(struct perf_event_header *header,
4558 struct perf_sample_data *data,
4559 struct perf_event *event,
4560 struct pt_regs *regs)
4562 u64 sample_type = event->attr.sample_type;
4564 header->type = PERF_RECORD_SAMPLE;
4565 header->size = sizeof(*header) + event->header_size;
4568 header->misc |= perf_misc_flags(regs);
4570 __perf_event_header__init_id(header, data, event);
4572 if (sample_type & PERF_SAMPLE_IP)
4573 data->ip = perf_instruction_pointer(regs);
4575 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4578 data->callchain = perf_callchain(event, regs);
4580 if (data->callchain)
4581 size += data->callchain->nr;
4583 header->size += size * sizeof(u64);
4586 if (sample_type & PERF_SAMPLE_RAW) {
4587 int size = sizeof(u32);
4590 size += data->raw->size;
4592 size += sizeof(u32);
4594 WARN_ON_ONCE(size & (sizeof(u64)-1));
4595 header->size += size;
4598 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4599 int size = sizeof(u64); /* nr */
4600 if (data->br_stack) {
4601 size += data->br_stack->nr
4602 * sizeof(struct perf_branch_entry);
4604 header->size += size;
4607 if (sample_type & PERF_SAMPLE_REGS_USER) {
4608 /* regs dump ABI info */
4609 int size = sizeof(u64);
4611 perf_sample_regs_user(&data->regs_user, regs);
4613 if (data->regs_user.regs) {
4614 u64 mask = event->attr.sample_regs_user;
4615 size += hweight64(mask) * sizeof(u64);
4618 header->size += size;
4621 if (sample_type & PERF_SAMPLE_STACK_USER) {
4623 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4624 * processed as the last one or have additional check added
4625 * in case new sample type is added, because we could eat
4626 * up the rest of the sample size.
4628 struct perf_regs_user *uregs = &data->regs_user;
4629 u16 stack_size = event->attr.sample_stack_user;
4630 u16 size = sizeof(u64);
4633 perf_sample_regs_user(uregs, regs);
4635 stack_size = perf_sample_ustack_size(stack_size, header->size,
4639 * If there is something to dump, add space for the dump
4640 * itself and for the field that tells the dynamic size,
4641 * which is how many have been actually dumped.
4644 size += sizeof(u64) + stack_size;
4646 data->stack_user_size = stack_size;
4647 header->size += size;
4651 static void perf_event_output(struct perf_event *event,
4652 struct perf_sample_data *data,
4653 struct pt_regs *regs)
4655 struct perf_output_handle handle;
4656 struct perf_event_header header;
4658 /* protect the callchain buffers */
4661 perf_prepare_sample(&header, data, event, regs);
4663 if (perf_output_begin(&handle, event, header.size))
4666 perf_output_sample(&handle, &header, data, event);
4668 perf_output_end(&handle);
4678 struct perf_read_event {
4679 struct perf_event_header header;
4686 perf_event_read_event(struct perf_event *event,
4687 struct task_struct *task)
4689 struct perf_output_handle handle;
4690 struct perf_sample_data sample;
4691 struct perf_read_event read_event = {
4693 .type = PERF_RECORD_READ,
4695 .size = sizeof(read_event) + event->read_size,
4697 .pid = perf_event_pid(event, task),
4698 .tid = perf_event_tid(event, task),
4702 perf_event_header__init_id(&read_event.header, &sample, event);
4703 ret = perf_output_begin(&handle, event, read_event.header.size);
4707 perf_output_put(&handle, read_event);
4708 perf_output_read(&handle, event);
4709 perf_event__output_id_sample(event, &handle, &sample);
4711 perf_output_end(&handle);
4714 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4717 perf_event_aux_ctx(struct perf_event_context *ctx,
4718 perf_event_aux_output_cb output,
4721 struct perf_event *event;
4723 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4724 if (event->state < PERF_EVENT_STATE_INACTIVE)
4726 if (!event_filter_match(event))
4728 output(event, data);
4733 perf_event_aux(perf_event_aux_output_cb output, void *data,
4734 struct perf_event_context *task_ctx)
4736 struct perf_cpu_context *cpuctx;
4737 struct perf_event_context *ctx;
4742 list_for_each_entry_rcu(pmu, &pmus, entry) {
4743 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4744 if (cpuctx->unique_pmu != pmu)
4746 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4749 ctxn = pmu->task_ctx_nr;
4752 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4754 perf_event_aux_ctx(ctx, output, data);
4756 put_cpu_ptr(pmu->pmu_cpu_context);
4761 perf_event_aux_ctx(task_ctx, output, data);
4768 * task tracking -- fork/exit
4770 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4773 struct perf_task_event {
4774 struct task_struct *task;
4775 struct perf_event_context *task_ctx;
4778 struct perf_event_header header;
4788 static int perf_event_task_match(struct perf_event *event)
4790 return event->attr.comm || event->attr.mmap ||
4791 event->attr.mmap_data || event->attr.task;
4794 static void perf_event_task_output(struct perf_event *event,
4797 struct perf_task_event *task_event = data;
4798 struct perf_output_handle handle;
4799 struct perf_sample_data sample;
4800 struct task_struct *task = task_event->task;
4801 int ret, size = task_event->event_id.header.size;
4803 if (!perf_event_task_match(event))
4806 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4808 ret = perf_output_begin(&handle, event,
4809 task_event->event_id.header.size);
4813 task_event->event_id.pid = perf_event_pid(event, task);
4814 task_event->event_id.ppid = perf_event_pid(event, current);
4816 task_event->event_id.tid = perf_event_tid(event, task);
4817 task_event->event_id.ptid = perf_event_tid(event, current);
4819 perf_output_put(&handle, task_event->event_id);
4821 perf_event__output_id_sample(event, &handle, &sample);
4823 perf_output_end(&handle);
4825 task_event->event_id.header.size = size;
4828 static void perf_event_task(struct task_struct *task,
4829 struct perf_event_context *task_ctx,
4832 struct perf_task_event task_event;
4834 if (!atomic_read(&nr_comm_events) &&
4835 !atomic_read(&nr_mmap_events) &&
4836 !atomic_read(&nr_task_events))
4839 task_event = (struct perf_task_event){
4841 .task_ctx = task_ctx,
4844 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4846 .size = sizeof(task_event.event_id),
4852 .time = perf_clock(),
4856 perf_event_aux(perf_event_task_output,
4861 void perf_event_fork(struct task_struct *task)
4863 perf_event_task(task, NULL, 1);
4870 struct perf_comm_event {
4871 struct task_struct *task;
4876 struct perf_event_header header;
4883 static int perf_event_comm_match(struct perf_event *event)
4885 return event->attr.comm;
4888 static void perf_event_comm_output(struct perf_event *event,
4891 struct perf_comm_event *comm_event = data;
4892 struct perf_output_handle handle;
4893 struct perf_sample_data sample;
4894 int size = comm_event->event_id.header.size;
4897 if (!perf_event_comm_match(event))
4900 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4901 ret = perf_output_begin(&handle, event,
4902 comm_event->event_id.header.size);
4907 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4908 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4910 perf_output_put(&handle, comm_event->event_id);
4911 __output_copy(&handle, comm_event->comm,
4912 comm_event->comm_size);
4914 perf_event__output_id_sample(event, &handle, &sample);
4916 perf_output_end(&handle);
4918 comm_event->event_id.header.size = size;
4921 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4923 char comm[TASK_COMM_LEN];
4926 memset(comm, 0, sizeof(comm));
4927 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4928 size = ALIGN(strlen(comm)+1, sizeof(u64));
4930 comm_event->comm = comm;
4931 comm_event->comm_size = size;
4933 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4935 perf_event_aux(perf_event_comm_output,
4940 void perf_event_comm(struct task_struct *task)
4942 struct perf_comm_event comm_event;
4943 struct perf_event_context *ctx;
4947 for_each_task_context_nr(ctxn) {
4948 ctx = task->perf_event_ctxp[ctxn];
4952 perf_event_enable_on_exec(ctx);
4956 if (!atomic_read(&nr_comm_events))
4959 comm_event = (struct perf_comm_event){
4965 .type = PERF_RECORD_COMM,
4974 perf_event_comm_event(&comm_event);
4981 struct perf_mmap_event {
4982 struct vm_area_struct *vma;
4984 const char *file_name;
4988 struct perf_event_header header;
4998 static int perf_event_mmap_match(struct perf_event *event,
5001 struct perf_mmap_event *mmap_event = data;
5002 struct vm_area_struct *vma = mmap_event->vma;
5003 int executable = vma->vm_flags & VM_EXEC;
5005 return (!executable && event->attr.mmap_data) ||
5006 (executable && event->attr.mmap);
5009 static void perf_event_mmap_output(struct perf_event *event,
5012 struct perf_mmap_event *mmap_event = data;
5013 struct perf_output_handle handle;
5014 struct perf_sample_data sample;
5015 int size = mmap_event->event_id.header.size;
5018 if (!perf_event_mmap_match(event, data))
5021 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5022 ret = perf_output_begin(&handle, event,
5023 mmap_event->event_id.header.size);
5027 mmap_event->event_id.pid = perf_event_pid(event, current);
5028 mmap_event->event_id.tid = perf_event_tid(event, current);
5030 perf_output_put(&handle, mmap_event->event_id);
5031 __output_copy(&handle, mmap_event->file_name,
5032 mmap_event->file_size);
5034 perf_event__output_id_sample(event, &handle, &sample);
5036 perf_output_end(&handle);
5038 mmap_event->event_id.header.size = size;
5041 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5043 struct vm_area_struct *vma = mmap_event->vma;
5044 struct file *file = vma->vm_file;
5050 memset(tmp, 0, sizeof(tmp));
5054 * d_path works from the end of the rb backwards, so we
5055 * need to add enough zero bytes after the string to handle
5056 * the 64bit alignment we do later.
5058 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
5060 name = strncpy(tmp, "//enomem", sizeof(tmp));
5063 name = d_path(&file->f_path, buf, PATH_MAX);
5065 name = strncpy(tmp, "//toolong", sizeof(tmp));
5069 if (arch_vma_name(mmap_event->vma)) {
5070 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
5072 tmp[sizeof(tmp) - 1] = '\0';
5077 name = strncpy(tmp, "[vdso]", sizeof(tmp));
5079 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
5080 vma->vm_end >= vma->vm_mm->brk) {
5081 name = strncpy(tmp, "[heap]", sizeof(tmp));
5083 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
5084 vma->vm_end >= vma->vm_mm->start_stack) {
5085 name = strncpy(tmp, "[stack]", sizeof(tmp));
5089 name = strncpy(tmp, "//anon", sizeof(tmp));
5094 size = ALIGN(strlen(name)+1, sizeof(u64));
5096 mmap_event->file_name = name;
5097 mmap_event->file_size = size;
5099 if (!(vma->vm_flags & VM_EXEC))
5100 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5102 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5104 perf_event_aux(perf_event_mmap_output,
5111 void perf_event_mmap(struct vm_area_struct *vma)
5113 struct perf_mmap_event mmap_event;
5115 if (!atomic_read(&nr_mmap_events))
5118 mmap_event = (struct perf_mmap_event){
5124 .type = PERF_RECORD_MMAP,
5125 .misc = PERF_RECORD_MISC_USER,
5130 .start = vma->vm_start,
5131 .len = vma->vm_end - vma->vm_start,
5132 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5136 perf_event_mmap_event(&mmap_event);
5140 * IRQ throttle logging
5143 static void perf_log_throttle(struct perf_event *event, int enable)
5145 struct perf_output_handle handle;
5146 struct perf_sample_data sample;
5150 struct perf_event_header header;
5154 } throttle_event = {
5156 .type = PERF_RECORD_THROTTLE,
5158 .size = sizeof(throttle_event),
5160 .time = perf_clock(),
5161 .id = primary_event_id(event),
5162 .stream_id = event->id,
5166 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5168 perf_event_header__init_id(&throttle_event.header, &sample, event);
5170 ret = perf_output_begin(&handle, event,
5171 throttle_event.header.size);
5175 perf_output_put(&handle, throttle_event);
5176 perf_event__output_id_sample(event, &handle, &sample);
5177 perf_output_end(&handle);
5181 * Generic event overflow handling, sampling.
5184 static int __perf_event_overflow(struct perf_event *event,
5185 int throttle, struct perf_sample_data *data,
5186 struct pt_regs *regs)
5188 int events = atomic_read(&event->event_limit);
5189 struct hw_perf_event *hwc = &event->hw;
5194 * Non-sampling counters might still use the PMI to fold short
5195 * hardware counters, ignore those.
5197 if (unlikely(!is_sampling_event(event)))
5200 seq = __this_cpu_read(perf_throttled_seq);
5201 if (seq != hwc->interrupts_seq) {
5202 hwc->interrupts_seq = seq;
5203 hwc->interrupts = 1;
5206 if (unlikely(throttle
5207 && hwc->interrupts >= max_samples_per_tick)) {
5208 __this_cpu_inc(perf_throttled_count);
5209 hwc->interrupts = MAX_INTERRUPTS;
5210 perf_log_throttle(event, 0);
5211 tick_nohz_full_kick();
5216 if (event->attr.freq) {
5217 u64 now = perf_clock();
5218 s64 delta = now - hwc->freq_time_stamp;
5220 hwc->freq_time_stamp = now;
5222 if (delta > 0 && delta < 2*TICK_NSEC)
5223 perf_adjust_period(event, delta, hwc->last_period, true);
5227 * XXX event_limit might not quite work as expected on inherited
5231 event->pending_kill = POLL_IN;
5232 if (events && atomic_dec_and_test(&event->event_limit)) {
5234 event->pending_kill = POLL_HUP;
5235 event->pending_disable = 1;
5236 irq_work_queue(&event->pending);
5239 if (event->overflow_handler)
5240 event->overflow_handler(event, data, regs);
5242 perf_event_output(event, data, regs);
5244 if (event->fasync && event->pending_kill) {
5245 event->pending_wakeup = 1;
5246 irq_work_queue(&event->pending);
5252 int perf_event_overflow(struct perf_event *event,
5253 struct perf_sample_data *data,
5254 struct pt_regs *regs)
5256 return __perf_event_overflow(event, 1, data, regs);
5260 * Generic software event infrastructure
5263 struct swevent_htable {
5264 struct swevent_hlist *swevent_hlist;
5265 struct mutex hlist_mutex;
5268 /* Recursion avoidance in each contexts */
5269 int recursion[PERF_NR_CONTEXTS];
5272 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5275 * We directly increment event->count and keep a second value in
5276 * event->hw.period_left to count intervals. This period event
5277 * is kept in the range [-sample_period, 0] so that we can use the
5281 u64 perf_swevent_set_period(struct perf_event *event)
5283 struct hw_perf_event *hwc = &event->hw;
5284 u64 period = hwc->last_period;
5288 hwc->last_period = hwc->sample_period;
5291 old = val = local64_read(&hwc->period_left);
5295 nr = div64_u64(period + val, period);
5296 offset = nr * period;
5298 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5304 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5305 struct perf_sample_data *data,
5306 struct pt_regs *regs)
5308 struct hw_perf_event *hwc = &event->hw;
5312 overflow = perf_swevent_set_period(event);
5314 if (hwc->interrupts == MAX_INTERRUPTS)
5317 for (; overflow; overflow--) {
5318 if (__perf_event_overflow(event, throttle,
5321 * We inhibit the overflow from happening when
5322 * hwc->interrupts == MAX_INTERRUPTS.
5330 static void perf_swevent_event(struct perf_event *event, u64 nr,
5331 struct perf_sample_data *data,
5332 struct pt_regs *regs)
5334 struct hw_perf_event *hwc = &event->hw;
5336 local64_add(nr, &event->count);
5341 if (!is_sampling_event(event))
5344 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5346 return perf_swevent_overflow(event, 1, data, regs);
5348 data->period = event->hw.last_period;
5350 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5351 return perf_swevent_overflow(event, 1, data, regs);
5353 if (local64_add_negative(nr, &hwc->period_left))
5356 perf_swevent_overflow(event, 0, data, regs);
5359 static int perf_exclude_event(struct perf_event *event,
5360 struct pt_regs *regs)
5362 if (event->hw.state & PERF_HES_STOPPED)
5366 if (event->attr.exclude_user && user_mode(regs))
5369 if (event->attr.exclude_kernel && !user_mode(regs))
5376 static int perf_swevent_match(struct perf_event *event,
5377 enum perf_type_id type,
5379 struct perf_sample_data *data,
5380 struct pt_regs *regs)
5382 if (event->attr.type != type)
5385 if (event->attr.config != event_id)
5388 if (perf_exclude_event(event, regs))
5394 static inline u64 swevent_hash(u64 type, u32 event_id)
5396 u64 val = event_id | (type << 32);
5398 return hash_64(val, SWEVENT_HLIST_BITS);
5401 static inline struct hlist_head *
5402 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5404 u64 hash = swevent_hash(type, event_id);
5406 return &hlist->heads[hash];
5409 /* For the read side: events when they trigger */
5410 static inline struct hlist_head *
5411 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5413 struct swevent_hlist *hlist;
5415 hlist = rcu_dereference(swhash->swevent_hlist);
5419 return __find_swevent_head(hlist, type, event_id);
5422 /* For the event head insertion and removal in the hlist */
5423 static inline struct hlist_head *
5424 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5426 struct swevent_hlist *hlist;
5427 u32 event_id = event->attr.config;
5428 u64 type = event->attr.type;
5431 * Event scheduling is always serialized against hlist allocation
5432 * and release. Which makes the protected version suitable here.
5433 * The context lock guarantees that.
5435 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5436 lockdep_is_held(&event->ctx->lock));
5440 return __find_swevent_head(hlist, type, event_id);
5443 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5445 struct perf_sample_data *data,
5446 struct pt_regs *regs)
5448 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5449 struct perf_event *event;
5450 struct hlist_head *head;
5453 head = find_swevent_head_rcu(swhash, type, event_id);
5457 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5458 if (perf_swevent_match(event, type, event_id, data, regs))
5459 perf_swevent_event(event, nr, data, regs);
5465 int perf_swevent_get_recursion_context(void)
5467 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5469 return get_recursion_context(swhash->recursion);
5471 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5473 inline void perf_swevent_put_recursion_context(int rctx)
5475 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5477 put_recursion_context(swhash->recursion, rctx);
5480 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5482 struct perf_sample_data data;
5485 preempt_disable_notrace();
5486 rctx = perf_swevent_get_recursion_context();
5490 perf_sample_data_init(&data, addr, 0);
5492 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5494 perf_swevent_put_recursion_context(rctx);
5495 preempt_enable_notrace();
5498 static void perf_swevent_read(struct perf_event *event)
5502 static int perf_swevent_add(struct perf_event *event, int flags)
5504 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5505 struct hw_perf_event *hwc = &event->hw;
5506 struct hlist_head *head;
5508 if (is_sampling_event(event)) {
5509 hwc->last_period = hwc->sample_period;
5510 perf_swevent_set_period(event);
5513 hwc->state = !(flags & PERF_EF_START);
5515 head = find_swevent_head(swhash, event);
5516 if (WARN_ON_ONCE(!head))
5519 hlist_add_head_rcu(&event->hlist_entry, head);
5524 static void perf_swevent_del(struct perf_event *event, int flags)
5526 hlist_del_rcu(&event->hlist_entry);
5529 static void perf_swevent_start(struct perf_event *event, int flags)
5531 event->hw.state = 0;
5534 static void perf_swevent_stop(struct perf_event *event, int flags)
5536 event->hw.state = PERF_HES_STOPPED;
5539 /* Deref the hlist from the update side */
5540 static inline struct swevent_hlist *
5541 swevent_hlist_deref(struct swevent_htable *swhash)
5543 return rcu_dereference_protected(swhash->swevent_hlist,
5544 lockdep_is_held(&swhash->hlist_mutex));
5547 static void swevent_hlist_release(struct swevent_htable *swhash)
5549 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5554 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5555 kfree_rcu(hlist, rcu_head);
5558 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5560 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5562 mutex_lock(&swhash->hlist_mutex);
5564 if (!--swhash->hlist_refcount)
5565 swevent_hlist_release(swhash);
5567 mutex_unlock(&swhash->hlist_mutex);
5570 static void swevent_hlist_put(struct perf_event *event)
5574 if (event->cpu != -1) {
5575 swevent_hlist_put_cpu(event, event->cpu);
5579 for_each_possible_cpu(cpu)
5580 swevent_hlist_put_cpu(event, cpu);
5583 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5585 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5588 mutex_lock(&swhash->hlist_mutex);
5590 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5591 struct swevent_hlist *hlist;
5593 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5598 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5600 swhash->hlist_refcount++;
5602 mutex_unlock(&swhash->hlist_mutex);
5607 static int swevent_hlist_get(struct perf_event *event)
5610 int cpu, failed_cpu;
5612 if (event->cpu != -1)
5613 return swevent_hlist_get_cpu(event, event->cpu);
5616 for_each_possible_cpu(cpu) {
5617 err = swevent_hlist_get_cpu(event, cpu);
5627 for_each_possible_cpu(cpu) {
5628 if (cpu == failed_cpu)
5630 swevent_hlist_put_cpu(event, cpu);
5637 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5639 static void sw_perf_event_destroy(struct perf_event *event)
5641 u64 event_id = event->attr.config;
5643 WARN_ON(event->parent);
5645 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5646 swevent_hlist_put(event);
5649 static int perf_swevent_init(struct perf_event *event)
5651 u64 event_id = event->attr.config;
5653 if (event->attr.type != PERF_TYPE_SOFTWARE)
5657 * no branch sampling for software events
5659 if (has_branch_stack(event))
5663 case PERF_COUNT_SW_CPU_CLOCK:
5664 case PERF_COUNT_SW_TASK_CLOCK:
5671 if (event_id >= PERF_COUNT_SW_MAX)
5674 if (!event->parent) {
5677 err = swevent_hlist_get(event);
5681 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5682 event->destroy = sw_perf_event_destroy;
5688 static int perf_swevent_event_idx(struct perf_event *event)
5693 static struct pmu perf_swevent = {
5694 .task_ctx_nr = perf_sw_context,
5696 .event_init = perf_swevent_init,
5697 .add = perf_swevent_add,
5698 .del = perf_swevent_del,
5699 .start = perf_swevent_start,
5700 .stop = perf_swevent_stop,
5701 .read = perf_swevent_read,
5703 .event_idx = perf_swevent_event_idx,
5706 #ifdef CONFIG_EVENT_TRACING
5708 static int perf_tp_filter_match(struct perf_event *event,
5709 struct perf_sample_data *data)
5711 void *record = data->raw->data;
5713 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5718 static int perf_tp_event_match(struct perf_event *event,
5719 struct perf_sample_data *data,
5720 struct pt_regs *regs)
5722 if (event->hw.state & PERF_HES_STOPPED)
5725 * All tracepoints are from kernel-space.
5727 if (event->attr.exclude_kernel)
5730 if (!perf_tp_filter_match(event, data))
5736 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5737 struct pt_regs *regs, struct hlist_head *head, int rctx,
5738 struct task_struct *task)
5740 struct perf_sample_data data;
5741 struct perf_event *event;
5743 struct perf_raw_record raw = {
5748 perf_sample_data_init(&data, addr, 0);
5751 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5752 if (perf_tp_event_match(event, &data, regs))
5753 perf_swevent_event(event, count, &data, regs);
5757 * If we got specified a target task, also iterate its context and
5758 * deliver this event there too.
5760 if (task && task != current) {
5761 struct perf_event_context *ctx;
5762 struct trace_entry *entry = record;
5765 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5769 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5770 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5772 if (event->attr.config != entry->type)
5774 if (perf_tp_event_match(event, &data, regs))
5775 perf_swevent_event(event, count, &data, regs);
5781 perf_swevent_put_recursion_context(rctx);
5783 EXPORT_SYMBOL_GPL(perf_tp_event);
5785 static void tp_perf_event_destroy(struct perf_event *event)
5787 perf_trace_destroy(event);
5790 static int perf_tp_event_init(struct perf_event *event)
5794 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5798 * no branch sampling for tracepoint events
5800 if (has_branch_stack(event))
5803 err = perf_trace_init(event);
5807 event->destroy = tp_perf_event_destroy;
5812 static struct pmu perf_tracepoint = {
5813 .task_ctx_nr = perf_sw_context,
5815 .event_init = perf_tp_event_init,
5816 .add = perf_trace_add,
5817 .del = perf_trace_del,
5818 .start = perf_swevent_start,
5819 .stop = perf_swevent_stop,
5820 .read = perf_swevent_read,
5822 .event_idx = perf_swevent_event_idx,
5825 static inline void perf_tp_register(void)
5827 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5830 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5835 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5838 filter_str = strndup_user(arg, PAGE_SIZE);
5839 if (IS_ERR(filter_str))
5840 return PTR_ERR(filter_str);
5842 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5848 static void perf_event_free_filter(struct perf_event *event)
5850 ftrace_profile_free_filter(event);
5855 static inline void perf_tp_register(void)
5859 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5864 static void perf_event_free_filter(struct perf_event *event)
5868 #endif /* CONFIG_EVENT_TRACING */
5870 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5871 void perf_bp_event(struct perf_event *bp, void *data)
5873 struct perf_sample_data sample;
5874 struct pt_regs *regs = data;
5876 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5878 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5879 perf_swevent_event(bp, 1, &sample, regs);
5884 * hrtimer based swevent callback
5887 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5889 enum hrtimer_restart ret = HRTIMER_RESTART;
5890 struct perf_sample_data data;
5891 struct pt_regs *regs;
5892 struct perf_event *event;
5895 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5897 if (event->state != PERF_EVENT_STATE_ACTIVE)
5898 return HRTIMER_NORESTART;
5900 event->pmu->read(event);
5902 perf_sample_data_init(&data, 0, event->hw.last_period);
5903 regs = get_irq_regs();
5905 if (regs && !perf_exclude_event(event, regs)) {
5906 if (!(event->attr.exclude_idle && is_idle_task(current)))
5907 if (__perf_event_overflow(event, 1, &data, regs))
5908 ret = HRTIMER_NORESTART;
5911 period = max_t(u64, 10000, event->hw.sample_period);
5912 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5917 static void perf_swevent_start_hrtimer(struct perf_event *event)
5919 struct hw_perf_event *hwc = &event->hw;
5922 if (!is_sampling_event(event))
5925 period = local64_read(&hwc->period_left);
5930 local64_set(&hwc->period_left, 0);
5932 period = max_t(u64, 10000, hwc->sample_period);
5934 __hrtimer_start_range_ns(&hwc->hrtimer,
5935 ns_to_ktime(period), 0,
5936 HRTIMER_MODE_REL_PINNED, 0);
5939 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5941 struct hw_perf_event *hwc = &event->hw;
5943 if (is_sampling_event(event)) {
5944 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5945 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5947 hrtimer_cancel(&hwc->hrtimer);
5951 static void perf_swevent_init_hrtimer(struct perf_event *event)
5953 struct hw_perf_event *hwc = &event->hw;
5955 if (!is_sampling_event(event))
5958 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5959 hwc->hrtimer.function = perf_swevent_hrtimer;
5962 * Since hrtimers have a fixed rate, we can do a static freq->period
5963 * mapping and avoid the whole period adjust feedback stuff.
5965 if (event->attr.freq) {
5966 long freq = event->attr.sample_freq;
5968 event->attr.sample_period = NSEC_PER_SEC / freq;
5969 hwc->sample_period = event->attr.sample_period;
5970 local64_set(&hwc->period_left, hwc->sample_period);
5971 hwc->last_period = hwc->sample_period;
5972 event->attr.freq = 0;
5977 * Software event: cpu wall time clock
5980 static void cpu_clock_event_update(struct perf_event *event)
5985 now = local_clock();
5986 prev = local64_xchg(&event->hw.prev_count, now);
5987 local64_add(now - prev, &event->count);
5990 static void cpu_clock_event_start(struct perf_event *event, int flags)
5992 local64_set(&event->hw.prev_count, local_clock());
5993 perf_swevent_start_hrtimer(event);
5996 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5998 perf_swevent_cancel_hrtimer(event);
5999 cpu_clock_event_update(event);
6002 static int cpu_clock_event_add(struct perf_event *event, int flags)
6004 if (flags & PERF_EF_START)
6005 cpu_clock_event_start(event, flags);
6010 static void cpu_clock_event_del(struct perf_event *event, int flags)
6012 cpu_clock_event_stop(event, flags);
6015 static void cpu_clock_event_read(struct perf_event *event)
6017 cpu_clock_event_update(event);
6020 static int cpu_clock_event_init(struct perf_event *event)
6022 if (event->attr.type != PERF_TYPE_SOFTWARE)
6025 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6029 * no branch sampling for software events
6031 if (has_branch_stack(event))
6034 perf_swevent_init_hrtimer(event);
6039 static struct pmu perf_cpu_clock = {
6040 .task_ctx_nr = perf_sw_context,
6042 .event_init = cpu_clock_event_init,
6043 .add = cpu_clock_event_add,
6044 .del = cpu_clock_event_del,
6045 .start = cpu_clock_event_start,
6046 .stop = cpu_clock_event_stop,
6047 .read = cpu_clock_event_read,
6049 .event_idx = perf_swevent_event_idx,
6053 * Software event: task time clock
6056 static void task_clock_event_update(struct perf_event *event, u64 now)
6061 prev = local64_xchg(&event->hw.prev_count, now);
6063 local64_add(delta, &event->count);
6066 static void task_clock_event_start(struct perf_event *event, int flags)
6068 local64_set(&event->hw.prev_count, event->ctx->time);
6069 perf_swevent_start_hrtimer(event);
6072 static void task_clock_event_stop(struct perf_event *event, int flags)
6074 perf_swevent_cancel_hrtimer(event);
6075 task_clock_event_update(event, event->ctx->time);
6078 static int task_clock_event_add(struct perf_event *event, int flags)
6080 if (flags & PERF_EF_START)
6081 task_clock_event_start(event, flags);
6086 static void task_clock_event_del(struct perf_event *event, int flags)
6088 task_clock_event_stop(event, PERF_EF_UPDATE);
6091 static void task_clock_event_read(struct perf_event *event)
6093 u64 now = perf_clock();
6094 u64 delta = now - event->ctx->timestamp;
6095 u64 time = event->ctx->time + delta;
6097 task_clock_event_update(event, time);
6100 static int task_clock_event_init(struct perf_event *event)
6102 if (event->attr.type != PERF_TYPE_SOFTWARE)
6105 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6109 * no branch sampling for software events
6111 if (has_branch_stack(event))
6114 perf_swevent_init_hrtimer(event);
6119 static struct pmu perf_task_clock = {
6120 .task_ctx_nr = perf_sw_context,
6122 .event_init = task_clock_event_init,
6123 .add = task_clock_event_add,
6124 .del = task_clock_event_del,
6125 .start = task_clock_event_start,
6126 .stop = task_clock_event_stop,
6127 .read = task_clock_event_read,
6129 .event_idx = perf_swevent_event_idx,
6132 static void perf_pmu_nop_void(struct pmu *pmu)
6136 static int perf_pmu_nop_int(struct pmu *pmu)
6141 static void perf_pmu_start_txn(struct pmu *pmu)
6143 perf_pmu_disable(pmu);
6146 static int perf_pmu_commit_txn(struct pmu *pmu)
6148 perf_pmu_enable(pmu);
6152 static void perf_pmu_cancel_txn(struct pmu *pmu)
6154 perf_pmu_enable(pmu);
6157 static int perf_event_idx_default(struct perf_event *event)
6159 return event->hw.idx + 1;
6163 * Ensures all contexts with the same task_ctx_nr have the same
6164 * pmu_cpu_context too.
6166 static void *find_pmu_context(int ctxn)
6173 list_for_each_entry(pmu, &pmus, entry) {
6174 if (pmu->task_ctx_nr == ctxn)
6175 return pmu->pmu_cpu_context;
6181 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6185 for_each_possible_cpu(cpu) {
6186 struct perf_cpu_context *cpuctx;
6188 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6190 if (cpuctx->unique_pmu == old_pmu)
6191 cpuctx->unique_pmu = pmu;
6195 static void free_pmu_context(struct pmu *pmu)
6199 mutex_lock(&pmus_lock);
6201 * Like a real lame refcount.
6203 list_for_each_entry(i, &pmus, entry) {
6204 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6205 update_pmu_context(i, pmu);
6210 free_percpu(pmu->pmu_cpu_context);
6212 mutex_unlock(&pmus_lock);
6214 static struct idr pmu_idr;
6217 type_show(struct device *dev, struct device_attribute *attr, char *page)
6219 struct pmu *pmu = dev_get_drvdata(dev);
6221 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6225 perf_event_mux_interval_ms_show(struct device *dev,
6226 struct device_attribute *attr,
6229 struct pmu *pmu = dev_get_drvdata(dev);
6231 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6235 perf_event_mux_interval_ms_store(struct device *dev,
6236 struct device_attribute *attr,
6237 const char *buf, size_t count)
6239 struct pmu *pmu = dev_get_drvdata(dev);
6240 int timer, cpu, ret;
6242 ret = kstrtoint(buf, 0, &timer);
6249 /* same value, noting to do */
6250 if (timer == pmu->hrtimer_interval_ms)
6253 pmu->hrtimer_interval_ms = timer;
6255 /* update all cpuctx for this PMU */
6256 for_each_possible_cpu(cpu) {
6257 struct perf_cpu_context *cpuctx;
6258 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6259 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6261 if (hrtimer_active(&cpuctx->hrtimer))
6262 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6268 static struct device_attribute pmu_dev_attrs[] = {
6270 __ATTR_RW(perf_event_mux_interval_ms),
6274 static int pmu_bus_running;
6275 static struct bus_type pmu_bus = {
6276 .name = "event_source",
6277 .dev_attrs = pmu_dev_attrs,
6280 static void pmu_dev_release(struct device *dev)
6285 static int pmu_dev_alloc(struct pmu *pmu)
6289 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6293 pmu->dev->groups = pmu->attr_groups;
6294 device_initialize(pmu->dev);
6295 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6299 dev_set_drvdata(pmu->dev, pmu);
6300 pmu->dev->bus = &pmu_bus;
6301 pmu->dev->release = pmu_dev_release;
6302 ret = device_add(pmu->dev);
6310 put_device(pmu->dev);
6314 static struct lock_class_key cpuctx_mutex;
6315 static struct lock_class_key cpuctx_lock;
6317 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6321 mutex_lock(&pmus_lock);
6323 pmu->pmu_disable_count = alloc_percpu(int);
6324 if (!pmu->pmu_disable_count)
6333 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6341 if (pmu_bus_running) {
6342 ret = pmu_dev_alloc(pmu);
6348 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6349 if (pmu->pmu_cpu_context)
6350 goto got_cpu_context;
6353 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6354 if (!pmu->pmu_cpu_context)
6357 for_each_possible_cpu(cpu) {
6358 struct perf_cpu_context *cpuctx;
6360 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6361 __perf_event_init_context(&cpuctx->ctx);
6362 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6363 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6364 cpuctx->ctx.type = cpu_context;
6365 cpuctx->ctx.pmu = pmu;
6367 __perf_cpu_hrtimer_init(cpuctx, cpu);
6369 INIT_LIST_HEAD(&cpuctx->rotation_list);
6370 cpuctx->unique_pmu = pmu;
6374 if (!pmu->start_txn) {
6375 if (pmu->pmu_enable) {
6377 * If we have pmu_enable/pmu_disable calls, install
6378 * transaction stubs that use that to try and batch
6379 * hardware accesses.
6381 pmu->start_txn = perf_pmu_start_txn;
6382 pmu->commit_txn = perf_pmu_commit_txn;
6383 pmu->cancel_txn = perf_pmu_cancel_txn;
6385 pmu->start_txn = perf_pmu_nop_void;
6386 pmu->commit_txn = perf_pmu_nop_int;
6387 pmu->cancel_txn = perf_pmu_nop_void;
6391 if (!pmu->pmu_enable) {
6392 pmu->pmu_enable = perf_pmu_nop_void;
6393 pmu->pmu_disable = perf_pmu_nop_void;
6396 if (!pmu->event_idx)
6397 pmu->event_idx = perf_event_idx_default;
6399 list_add_rcu(&pmu->entry, &pmus);
6402 mutex_unlock(&pmus_lock);
6407 device_del(pmu->dev);
6408 put_device(pmu->dev);
6411 if (pmu->type >= PERF_TYPE_MAX)
6412 idr_remove(&pmu_idr, pmu->type);
6415 free_percpu(pmu->pmu_disable_count);
6419 void perf_pmu_unregister(struct pmu *pmu)
6421 mutex_lock(&pmus_lock);
6422 list_del_rcu(&pmu->entry);
6423 mutex_unlock(&pmus_lock);
6426 * We dereference the pmu list under both SRCU and regular RCU, so
6427 * synchronize against both of those.
6429 synchronize_srcu(&pmus_srcu);
6432 free_percpu(pmu->pmu_disable_count);
6433 if (pmu->type >= PERF_TYPE_MAX)
6434 idr_remove(&pmu_idr, pmu->type);
6435 device_del(pmu->dev);
6436 put_device(pmu->dev);
6437 free_pmu_context(pmu);
6440 struct pmu *perf_init_event(struct perf_event *event)
6442 struct pmu *pmu = NULL;
6446 idx = srcu_read_lock(&pmus_srcu);
6449 pmu = idr_find(&pmu_idr, event->attr.type);
6453 ret = pmu->event_init(event);
6459 list_for_each_entry_rcu(pmu, &pmus, entry) {
6461 ret = pmu->event_init(event);
6465 if (ret != -ENOENT) {
6470 pmu = ERR_PTR(-ENOENT);
6472 srcu_read_unlock(&pmus_srcu, idx);
6477 static void account_event_cpu(struct perf_event *event, int cpu)
6482 if (has_branch_stack(event)) {
6483 if (!(event->attach_state & PERF_ATTACH_TASK))
6484 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6486 if (is_cgroup_event(event))
6487 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6490 static void account_event(struct perf_event *event)
6495 if (event->attach_state & PERF_ATTACH_TASK)
6496 static_key_slow_inc(&perf_sched_events.key);
6497 if (event->attr.mmap || event->attr.mmap_data)
6498 atomic_inc(&nr_mmap_events);
6499 if (event->attr.comm)
6500 atomic_inc(&nr_comm_events);
6501 if (event->attr.task)
6502 atomic_inc(&nr_task_events);
6503 if (event->attr.freq) {
6504 if (atomic_inc_return(&nr_freq_events) == 1)
6505 tick_nohz_full_kick_all();
6507 if (has_branch_stack(event))
6508 static_key_slow_inc(&perf_sched_events.key);
6509 if (is_cgroup_event(event))
6510 static_key_slow_inc(&perf_sched_events.key);
6512 account_event_cpu(event, event->cpu);
6516 * Allocate and initialize a event structure
6518 static struct perf_event *
6519 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6520 struct task_struct *task,
6521 struct perf_event *group_leader,
6522 struct perf_event *parent_event,
6523 perf_overflow_handler_t overflow_handler,
6527 struct perf_event *event;
6528 struct hw_perf_event *hwc;
6531 if ((unsigned)cpu >= nr_cpu_ids) {
6532 if (!task || cpu != -1)
6533 return ERR_PTR(-EINVAL);
6536 event = kzalloc(sizeof(*event), GFP_KERNEL);
6538 return ERR_PTR(-ENOMEM);
6541 * Single events are their own group leaders, with an
6542 * empty sibling list:
6545 group_leader = event;
6547 mutex_init(&event->child_mutex);
6548 INIT_LIST_HEAD(&event->child_list);
6550 INIT_LIST_HEAD(&event->group_entry);
6551 INIT_LIST_HEAD(&event->event_entry);
6552 INIT_LIST_HEAD(&event->sibling_list);
6553 INIT_LIST_HEAD(&event->rb_entry);
6555 init_waitqueue_head(&event->waitq);
6556 init_irq_work(&event->pending, perf_pending_event);
6558 mutex_init(&event->mmap_mutex);
6560 atomic_long_set(&event->refcount, 1);
6562 event->attr = *attr;
6563 event->group_leader = group_leader;
6567 event->parent = parent_event;
6569 event->ns = get_pid_ns(task_active_pid_ns(current));
6570 event->id = atomic64_inc_return(&perf_event_id);
6572 event->state = PERF_EVENT_STATE_INACTIVE;
6575 event->attach_state = PERF_ATTACH_TASK;
6577 if (attr->type == PERF_TYPE_TRACEPOINT)
6578 event->hw.tp_target = task;
6579 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6581 * hw_breakpoint is a bit difficult here..
6583 else if (attr->type == PERF_TYPE_BREAKPOINT)
6584 event->hw.bp_target = task;
6588 if (!overflow_handler && parent_event) {
6589 overflow_handler = parent_event->overflow_handler;
6590 context = parent_event->overflow_handler_context;
6593 event->overflow_handler = overflow_handler;
6594 event->overflow_handler_context = context;
6596 perf_event__state_init(event);
6601 hwc->sample_period = attr->sample_period;
6602 if (attr->freq && attr->sample_freq)
6603 hwc->sample_period = 1;
6604 hwc->last_period = hwc->sample_period;
6606 local64_set(&hwc->period_left, hwc->sample_period);
6609 * we currently do not support PERF_FORMAT_GROUP on inherited events
6611 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6614 pmu = perf_init_event(event);
6617 else if (IS_ERR(pmu)) {
6622 if (!event->parent) {
6623 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6624 err = get_callchain_buffers();
6634 event->destroy(event);
6637 put_pid_ns(event->ns);
6640 return ERR_PTR(err);
6643 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6644 struct perf_event_attr *attr)
6649 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6653 * zero the full structure, so that a short copy will be nice.
6655 memset(attr, 0, sizeof(*attr));
6657 ret = get_user(size, &uattr->size);
6661 if (size > PAGE_SIZE) /* silly large */
6664 if (!size) /* abi compat */
6665 size = PERF_ATTR_SIZE_VER0;
6667 if (size < PERF_ATTR_SIZE_VER0)
6671 * If we're handed a bigger struct than we know of,
6672 * ensure all the unknown bits are 0 - i.e. new
6673 * user-space does not rely on any kernel feature
6674 * extensions we dont know about yet.
6676 if (size > sizeof(*attr)) {
6677 unsigned char __user *addr;
6678 unsigned char __user *end;
6681 addr = (void __user *)uattr + sizeof(*attr);
6682 end = (void __user *)uattr + size;
6684 for (; addr < end; addr++) {
6685 ret = get_user(val, addr);
6691 size = sizeof(*attr);
6694 ret = copy_from_user(attr, uattr, size);
6698 if (attr->__reserved_1)
6701 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6704 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6707 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6708 u64 mask = attr->branch_sample_type;
6710 /* only using defined bits */
6711 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6714 /* at least one branch bit must be set */
6715 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6718 /* propagate priv level, when not set for branch */
6719 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6721 /* exclude_kernel checked on syscall entry */
6722 if (!attr->exclude_kernel)
6723 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6725 if (!attr->exclude_user)
6726 mask |= PERF_SAMPLE_BRANCH_USER;
6728 if (!attr->exclude_hv)
6729 mask |= PERF_SAMPLE_BRANCH_HV;
6731 * adjust user setting (for HW filter setup)
6733 attr->branch_sample_type = mask;
6735 /* privileged levels capture (kernel, hv): check permissions */
6736 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6737 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6741 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6742 ret = perf_reg_validate(attr->sample_regs_user);
6747 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6748 if (!arch_perf_have_user_stack_dump())
6752 * We have __u32 type for the size, but so far
6753 * we can only use __u16 as maximum due to the
6754 * __u16 sample size limit.
6756 if (attr->sample_stack_user >= USHRT_MAX)
6758 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6766 put_user(sizeof(*attr), &uattr->size);
6772 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6774 struct ring_buffer *rb = NULL, *old_rb = NULL;
6780 /* don't allow circular references */
6781 if (event == output_event)
6785 * Don't allow cross-cpu buffers
6787 if (output_event->cpu != event->cpu)
6791 * If its not a per-cpu rb, it must be the same task.
6793 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6797 mutex_lock(&event->mmap_mutex);
6798 /* Can't redirect output if we've got an active mmap() */
6799 if (atomic_read(&event->mmap_count))
6805 /* get the rb we want to redirect to */
6806 rb = ring_buffer_get(output_event);
6812 ring_buffer_detach(event, old_rb);
6815 ring_buffer_attach(event, rb);
6817 rcu_assign_pointer(event->rb, rb);
6820 ring_buffer_put(old_rb);
6822 * Since we detached before setting the new rb, so that we
6823 * could attach the new rb, we could have missed a wakeup.
6826 wake_up_all(&event->waitq);
6831 mutex_unlock(&event->mmap_mutex);
6838 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6840 * @attr_uptr: event_id type attributes for monitoring/sampling
6843 * @group_fd: group leader event fd
6845 SYSCALL_DEFINE5(perf_event_open,
6846 struct perf_event_attr __user *, attr_uptr,
6847 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6849 struct perf_event *group_leader = NULL, *output_event = NULL;
6850 struct perf_event *event, *sibling;
6851 struct perf_event_attr attr;
6852 struct perf_event_context *ctx;
6853 struct file *event_file = NULL;
6854 struct fd group = {NULL, 0};
6855 struct task_struct *task = NULL;
6861 /* for future expandability... */
6862 if (flags & ~PERF_FLAG_ALL)
6865 err = perf_copy_attr(attr_uptr, &attr);
6869 if (!attr.exclude_kernel) {
6870 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6875 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6880 * In cgroup mode, the pid argument is used to pass the fd
6881 * opened to the cgroup directory in cgroupfs. The cpu argument
6882 * designates the cpu on which to monitor threads from that
6885 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6888 event_fd = get_unused_fd();
6892 if (group_fd != -1) {
6893 err = perf_fget_light(group_fd, &group);
6896 group_leader = group.file->private_data;
6897 if (flags & PERF_FLAG_FD_OUTPUT)
6898 output_event = group_leader;
6899 if (flags & PERF_FLAG_FD_NO_GROUP)
6900 group_leader = NULL;
6903 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6904 task = find_lively_task_by_vpid(pid);
6906 err = PTR_ERR(task);
6913 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6915 if (IS_ERR(event)) {
6916 err = PTR_ERR(event);
6920 if (flags & PERF_FLAG_PID_CGROUP) {
6921 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6923 __free_event(event);
6928 account_event(event);
6931 * Special case software events and allow them to be part of
6932 * any hardware group.
6937 (is_software_event(event) != is_software_event(group_leader))) {
6938 if (is_software_event(event)) {
6940 * If event and group_leader are not both a software
6941 * event, and event is, then group leader is not.
6943 * Allow the addition of software events to !software
6944 * groups, this is safe because software events never
6947 pmu = group_leader->pmu;
6948 } else if (is_software_event(group_leader) &&
6949 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6951 * In case the group is a pure software group, and we
6952 * try to add a hardware event, move the whole group to
6953 * the hardware context.
6960 * Get the target context (task or percpu):
6962 ctx = find_get_context(pmu, task, event->cpu);
6969 put_task_struct(task);
6974 * Look up the group leader (we will attach this event to it):
6980 * Do not allow a recursive hierarchy (this new sibling
6981 * becoming part of another group-sibling):
6983 if (group_leader->group_leader != group_leader)
6986 * Do not allow to attach to a group in a different
6987 * task or CPU context:
6990 if (group_leader->ctx->type != ctx->type)
6993 if (group_leader->ctx != ctx)
6998 * Only a group leader can be exclusive or pinned
7000 if (attr.exclusive || attr.pinned)
7005 err = perf_event_set_output(event, output_event);
7010 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
7011 if (IS_ERR(event_file)) {
7012 err = PTR_ERR(event_file);
7017 struct perf_event_context *gctx = group_leader->ctx;
7019 mutex_lock(&gctx->mutex);
7020 perf_remove_from_context(group_leader);
7023 * Removing from the context ends up with disabled
7024 * event. What we want here is event in the initial
7025 * startup state, ready to be add into new context.
7027 perf_event__state_init(group_leader);
7028 list_for_each_entry(sibling, &group_leader->sibling_list,
7030 perf_remove_from_context(sibling);
7031 perf_event__state_init(sibling);
7034 mutex_unlock(&gctx->mutex);
7038 WARN_ON_ONCE(ctx->parent_ctx);
7039 mutex_lock(&ctx->mutex);
7043 perf_install_in_context(ctx, group_leader, event->cpu);
7045 list_for_each_entry(sibling, &group_leader->sibling_list,
7047 perf_install_in_context(ctx, sibling, event->cpu);
7052 perf_install_in_context(ctx, event, event->cpu);
7054 perf_unpin_context(ctx);
7055 mutex_unlock(&ctx->mutex);
7059 event->owner = current;
7061 mutex_lock(¤t->perf_event_mutex);
7062 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7063 mutex_unlock(¤t->perf_event_mutex);
7066 * Precalculate sample_data sizes
7068 perf_event__header_size(event);
7069 perf_event__id_header_size(event);
7072 * Drop the reference on the group_event after placing the
7073 * new event on the sibling_list. This ensures destruction
7074 * of the group leader will find the pointer to itself in
7075 * perf_group_detach().
7078 fd_install(event_fd, event_file);
7082 perf_unpin_context(ctx);
7089 put_task_struct(task);
7093 put_unused_fd(event_fd);
7098 * perf_event_create_kernel_counter
7100 * @attr: attributes of the counter to create
7101 * @cpu: cpu in which the counter is bound
7102 * @task: task to profile (NULL for percpu)
7105 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7106 struct task_struct *task,
7107 perf_overflow_handler_t overflow_handler,
7110 struct perf_event_context *ctx;
7111 struct perf_event *event;
7115 * Get the target context (task or percpu):
7118 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7119 overflow_handler, context);
7120 if (IS_ERR(event)) {
7121 err = PTR_ERR(event);
7125 account_event(event);
7127 ctx = find_get_context(event->pmu, task, cpu);
7133 WARN_ON_ONCE(ctx->parent_ctx);
7134 mutex_lock(&ctx->mutex);
7135 perf_install_in_context(ctx, event, cpu);
7137 perf_unpin_context(ctx);
7138 mutex_unlock(&ctx->mutex);
7145 return ERR_PTR(err);
7147 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7149 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7151 struct perf_event_context *src_ctx;
7152 struct perf_event_context *dst_ctx;
7153 struct perf_event *event, *tmp;
7156 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7157 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7159 mutex_lock(&src_ctx->mutex);
7160 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7162 perf_remove_from_context(event);
7163 unaccount_event_cpu(event, src_cpu);
7165 list_add(&event->event_entry, &events);
7167 mutex_unlock(&src_ctx->mutex);
7171 mutex_lock(&dst_ctx->mutex);
7172 list_for_each_entry_safe(event, tmp, &events, event_entry) {
7173 list_del(&event->event_entry);
7174 if (event->state >= PERF_EVENT_STATE_OFF)
7175 event->state = PERF_EVENT_STATE_INACTIVE;
7176 account_event_cpu(event, dst_cpu);
7177 perf_install_in_context(dst_ctx, event, dst_cpu);
7180 mutex_unlock(&dst_ctx->mutex);
7182 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7184 static void sync_child_event(struct perf_event *child_event,
7185 struct task_struct *child)
7187 struct perf_event *parent_event = child_event->parent;
7190 if (child_event->attr.inherit_stat)
7191 perf_event_read_event(child_event, child);
7193 child_val = perf_event_count(child_event);
7196 * Add back the child's count to the parent's count:
7198 atomic64_add(child_val, &parent_event->child_count);
7199 atomic64_add(child_event->total_time_enabled,
7200 &parent_event->child_total_time_enabled);
7201 atomic64_add(child_event->total_time_running,
7202 &parent_event->child_total_time_running);
7205 * Remove this event from the parent's list
7207 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7208 mutex_lock(&parent_event->child_mutex);
7209 list_del_init(&child_event->child_list);
7210 mutex_unlock(&parent_event->child_mutex);
7213 * Release the parent event, if this was the last
7216 put_event(parent_event);
7220 __perf_event_exit_task(struct perf_event *child_event,
7221 struct perf_event_context *child_ctx,
7222 struct task_struct *child)
7224 if (child_event->parent) {
7225 raw_spin_lock_irq(&child_ctx->lock);
7226 perf_group_detach(child_event);
7227 raw_spin_unlock_irq(&child_ctx->lock);
7230 perf_remove_from_context(child_event);
7233 * It can happen that the parent exits first, and has events
7234 * that are still around due to the child reference. These
7235 * events need to be zapped.
7237 if (child_event->parent) {
7238 sync_child_event(child_event, child);
7239 free_event(child_event);
7243 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7245 struct perf_event *child_event, *tmp;
7246 struct perf_event_context *child_ctx;
7247 unsigned long flags;
7249 if (likely(!child->perf_event_ctxp[ctxn])) {
7250 perf_event_task(child, NULL, 0);
7254 local_irq_save(flags);
7256 * We can't reschedule here because interrupts are disabled,
7257 * and either child is current or it is a task that can't be
7258 * scheduled, so we are now safe from rescheduling changing
7261 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7264 * Take the context lock here so that if find_get_context is
7265 * reading child->perf_event_ctxp, we wait until it has
7266 * incremented the context's refcount before we do put_ctx below.
7268 raw_spin_lock(&child_ctx->lock);
7269 task_ctx_sched_out(child_ctx);
7270 child->perf_event_ctxp[ctxn] = NULL;
7272 * If this context is a clone; unclone it so it can't get
7273 * swapped to another process while we're removing all
7274 * the events from it.
7276 unclone_ctx(child_ctx);
7277 update_context_time(child_ctx);
7278 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7281 * Report the task dead after unscheduling the events so that we
7282 * won't get any samples after PERF_RECORD_EXIT. We can however still
7283 * get a few PERF_RECORD_READ events.
7285 perf_event_task(child, child_ctx, 0);
7288 * We can recurse on the same lock type through:
7290 * __perf_event_exit_task()
7291 * sync_child_event()
7293 * mutex_lock(&ctx->mutex)
7295 * But since its the parent context it won't be the same instance.
7297 mutex_lock(&child_ctx->mutex);
7300 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7302 __perf_event_exit_task(child_event, child_ctx, child);
7304 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7306 __perf_event_exit_task(child_event, child_ctx, child);
7309 * If the last event was a group event, it will have appended all
7310 * its siblings to the list, but we obtained 'tmp' before that which
7311 * will still point to the list head terminating the iteration.
7313 if (!list_empty(&child_ctx->pinned_groups) ||
7314 !list_empty(&child_ctx->flexible_groups))
7317 mutex_unlock(&child_ctx->mutex);
7323 * When a child task exits, feed back event values to parent events.
7325 void perf_event_exit_task(struct task_struct *child)
7327 struct perf_event *event, *tmp;
7330 mutex_lock(&child->perf_event_mutex);
7331 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7333 list_del_init(&event->owner_entry);
7336 * Ensure the list deletion is visible before we clear
7337 * the owner, closes a race against perf_release() where
7338 * we need to serialize on the owner->perf_event_mutex.
7341 event->owner = NULL;
7343 mutex_unlock(&child->perf_event_mutex);
7345 for_each_task_context_nr(ctxn)
7346 perf_event_exit_task_context(child, ctxn);
7349 static void perf_free_event(struct perf_event *event,
7350 struct perf_event_context *ctx)
7352 struct perf_event *parent = event->parent;
7354 if (WARN_ON_ONCE(!parent))
7357 mutex_lock(&parent->child_mutex);
7358 list_del_init(&event->child_list);
7359 mutex_unlock(&parent->child_mutex);
7363 perf_group_detach(event);
7364 list_del_event(event, ctx);
7369 * free an unexposed, unused context as created by inheritance by
7370 * perf_event_init_task below, used by fork() in case of fail.
7372 void perf_event_free_task(struct task_struct *task)
7374 struct perf_event_context *ctx;
7375 struct perf_event *event, *tmp;
7378 for_each_task_context_nr(ctxn) {
7379 ctx = task->perf_event_ctxp[ctxn];
7383 mutex_lock(&ctx->mutex);
7385 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7387 perf_free_event(event, ctx);
7389 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7391 perf_free_event(event, ctx);
7393 if (!list_empty(&ctx->pinned_groups) ||
7394 !list_empty(&ctx->flexible_groups))
7397 mutex_unlock(&ctx->mutex);
7403 void perf_event_delayed_put(struct task_struct *task)
7407 for_each_task_context_nr(ctxn)
7408 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7412 * inherit a event from parent task to child task:
7414 static struct perf_event *
7415 inherit_event(struct perf_event *parent_event,
7416 struct task_struct *parent,
7417 struct perf_event_context *parent_ctx,
7418 struct task_struct *child,
7419 struct perf_event *group_leader,
7420 struct perf_event_context *child_ctx)
7422 struct perf_event *child_event;
7423 unsigned long flags;
7426 * Instead of creating recursive hierarchies of events,
7427 * we link inherited events back to the original parent,
7428 * which has a filp for sure, which we use as the reference
7431 if (parent_event->parent)
7432 parent_event = parent_event->parent;
7434 child_event = perf_event_alloc(&parent_event->attr,
7437 group_leader, parent_event,
7439 if (IS_ERR(child_event))
7442 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7443 free_event(child_event);
7450 * Make the child state follow the state of the parent event,
7451 * not its attr.disabled bit. We hold the parent's mutex,
7452 * so we won't race with perf_event_{en, dis}able_family.
7454 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7455 child_event->state = PERF_EVENT_STATE_INACTIVE;
7457 child_event->state = PERF_EVENT_STATE_OFF;
7459 if (parent_event->attr.freq) {
7460 u64 sample_period = parent_event->hw.sample_period;
7461 struct hw_perf_event *hwc = &child_event->hw;
7463 hwc->sample_period = sample_period;
7464 hwc->last_period = sample_period;
7466 local64_set(&hwc->period_left, sample_period);
7469 child_event->ctx = child_ctx;
7470 child_event->overflow_handler = parent_event->overflow_handler;
7471 child_event->overflow_handler_context
7472 = parent_event->overflow_handler_context;
7475 * Precalculate sample_data sizes
7477 perf_event__header_size(child_event);
7478 perf_event__id_header_size(child_event);
7481 * Link it up in the child's context:
7483 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7484 add_event_to_ctx(child_event, child_ctx);
7485 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7488 * Link this into the parent event's child list
7490 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7491 mutex_lock(&parent_event->child_mutex);
7492 list_add_tail(&child_event->child_list, &parent_event->child_list);
7493 mutex_unlock(&parent_event->child_mutex);
7498 static int inherit_group(struct perf_event *parent_event,
7499 struct task_struct *parent,
7500 struct perf_event_context *parent_ctx,
7501 struct task_struct *child,
7502 struct perf_event_context *child_ctx)
7504 struct perf_event *leader;
7505 struct perf_event *sub;
7506 struct perf_event *child_ctr;
7508 leader = inherit_event(parent_event, parent, parent_ctx,
7509 child, NULL, child_ctx);
7511 return PTR_ERR(leader);
7512 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7513 child_ctr = inherit_event(sub, parent, parent_ctx,
7514 child, leader, child_ctx);
7515 if (IS_ERR(child_ctr))
7516 return PTR_ERR(child_ctr);
7522 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7523 struct perf_event_context *parent_ctx,
7524 struct task_struct *child, int ctxn,
7528 struct perf_event_context *child_ctx;
7530 if (!event->attr.inherit) {
7535 child_ctx = child->perf_event_ctxp[ctxn];
7538 * This is executed from the parent task context, so
7539 * inherit events that have been marked for cloning.
7540 * First allocate and initialize a context for the
7544 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7548 child->perf_event_ctxp[ctxn] = child_ctx;
7551 ret = inherit_group(event, parent, parent_ctx,
7561 * Initialize the perf_event context in task_struct
7563 int perf_event_init_context(struct task_struct *child, int ctxn)
7565 struct perf_event_context *child_ctx, *parent_ctx;
7566 struct perf_event_context *cloned_ctx;
7567 struct perf_event *event;
7568 struct task_struct *parent = current;
7569 int inherited_all = 1;
7570 unsigned long flags;
7573 if (likely(!parent->perf_event_ctxp[ctxn]))
7577 * If the parent's context is a clone, pin it so it won't get
7580 parent_ctx = perf_pin_task_context(parent, ctxn);
7583 * No need to check if parent_ctx != NULL here; since we saw
7584 * it non-NULL earlier, the only reason for it to become NULL
7585 * is if we exit, and since we're currently in the middle of
7586 * a fork we can't be exiting at the same time.
7590 * Lock the parent list. No need to lock the child - not PID
7591 * hashed yet and not running, so nobody can access it.
7593 mutex_lock(&parent_ctx->mutex);
7596 * We dont have to disable NMIs - we are only looking at
7597 * the list, not manipulating it:
7599 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7600 ret = inherit_task_group(event, parent, parent_ctx,
7601 child, ctxn, &inherited_all);
7607 * We can't hold ctx->lock when iterating the ->flexible_group list due
7608 * to allocations, but we need to prevent rotation because
7609 * rotate_ctx() will change the list from interrupt context.
7611 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7612 parent_ctx->rotate_disable = 1;
7613 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7615 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7616 ret = inherit_task_group(event, parent, parent_ctx,
7617 child, ctxn, &inherited_all);
7622 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7623 parent_ctx->rotate_disable = 0;
7625 child_ctx = child->perf_event_ctxp[ctxn];
7627 if (child_ctx && inherited_all) {
7629 * Mark the child context as a clone of the parent
7630 * context, or of whatever the parent is a clone of.
7632 * Note that if the parent is a clone, the holding of
7633 * parent_ctx->lock avoids it from being uncloned.
7635 cloned_ctx = parent_ctx->parent_ctx;
7637 child_ctx->parent_ctx = cloned_ctx;
7638 child_ctx->parent_gen = parent_ctx->parent_gen;
7640 child_ctx->parent_ctx = parent_ctx;
7641 child_ctx->parent_gen = parent_ctx->generation;
7643 get_ctx(child_ctx->parent_ctx);
7646 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7647 mutex_unlock(&parent_ctx->mutex);
7649 perf_unpin_context(parent_ctx);
7650 put_ctx(parent_ctx);
7656 * Initialize the perf_event context in task_struct
7658 int perf_event_init_task(struct task_struct *child)
7662 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7663 mutex_init(&child->perf_event_mutex);
7664 INIT_LIST_HEAD(&child->perf_event_list);
7666 for_each_task_context_nr(ctxn) {
7667 ret = perf_event_init_context(child, ctxn);
7675 static void __init perf_event_init_all_cpus(void)
7677 struct swevent_htable *swhash;
7680 for_each_possible_cpu(cpu) {
7681 swhash = &per_cpu(swevent_htable, cpu);
7682 mutex_init(&swhash->hlist_mutex);
7683 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7687 static void perf_event_init_cpu(int cpu)
7689 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7691 mutex_lock(&swhash->hlist_mutex);
7692 if (swhash->hlist_refcount > 0) {
7693 struct swevent_hlist *hlist;
7695 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7697 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7699 mutex_unlock(&swhash->hlist_mutex);
7702 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7703 static void perf_pmu_rotate_stop(struct pmu *pmu)
7705 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7707 WARN_ON(!irqs_disabled());
7709 list_del_init(&cpuctx->rotation_list);
7712 static void __perf_event_exit_context(void *__info)
7714 struct perf_event_context *ctx = __info;
7715 struct perf_event *event, *tmp;
7717 perf_pmu_rotate_stop(ctx->pmu);
7719 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7720 __perf_remove_from_context(event);
7721 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7722 __perf_remove_from_context(event);
7725 static void perf_event_exit_cpu_context(int cpu)
7727 struct perf_event_context *ctx;
7731 idx = srcu_read_lock(&pmus_srcu);
7732 list_for_each_entry_rcu(pmu, &pmus, entry) {
7733 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7735 mutex_lock(&ctx->mutex);
7736 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7737 mutex_unlock(&ctx->mutex);
7739 srcu_read_unlock(&pmus_srcu, idx);
7742 static void perf_event_exit_cpu(int cpu)
7744 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7746 mutex_lock(&swhash->hlist_mutex);
7747 swevent_hlist_release(swhash);
7748 mutex_unlock(&swhash->hlist_mutex);
7750 perf_event_exit_cpu_context(cpu);
7753 static inline void perf_event_exit_cpu(int cpu) { }
7757 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7761 for_each_online_cpu(cpu)
7762 perf_event_exit_cpu(cpu);
7768 * Run the perf reboot notifier at the very last possible moment so that
7769 * the generic watchdog code runs as long as possible.
7771 static struct notifier_block perf_reboot_notifier = {
7772 .notifier_call = perf_reboot,
7773 .priority = INT_MIN,
7777 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7779 unsigned int cpu = (long)hcpu;
7781 switch (action & ~CPU_TASKS_FROZEN) {
7783 case CPU_UP_PREPARE:
7784 case CPU_DOWN_FAILED:
7785 perf_event_init_cpu(cpu);
7788 case CPU_UP_CANCELED:
7789 case CPU_DOWN_PREPARE:
7790 perf_event_exit_cpu(cpu);
7799 void __init perf_event_init(void)
7805 perf_event_init_all_cpus();
7806 init_srcu_struct(&pmus_srcu);
7807 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7808 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7809 perf_pmu_register(&perf_task_clock, NULL, -1);
7811 perf_cpu_notifier(perf_cpu_notify);
7812 register_reboot_notifier(&perf_reboot_notifier);
7814 ret = init_hw_breakpoint();
7815 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7817 /* do not patch jump label more than once per second */
7818 jump_label_rate_limit(&perf_sched_events, HZ);
7821 * Build time assertion that we keep the data_head at the intended
7822 * location. IOW, validation we got the __reserved[] size right.
7824 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7828 static int __init perf_event_sysfs_init(void)
7833 mutex_lock(&pmus_lock);
7835 ret = bus_register(&pmu_bus);
7839 list_for_each_entry(pmu, &pmus, entry) {
7840 if (!pmu->name || pmu->type < 0)
7843 ret = pmu_dev_alloc(pmu);
7844 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7846 pmu_bus_running = 1;
7850 mutex_unlock(&pmus_lock);
7854 device_initcall(perf_event_sysfs_init);
7856 #ifdef CONFIG_CGROUP_PERF
7857 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7859 struct perf_cgroup *jc;
7861 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7863 return ERR_PTR(-ENOMEM);
7865 jc->info = alloc_percpu(struct perf_cgroup_info);
7868 return ERR_PTR(-ENOMEM);
7874 static void perf_cgroup_css_free(struct cgroup *cont)
7876 struct perf_cgroup *jc;
7877 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7878 struct perf_cgroup, css);
7879 free_percpu(jc->info);
7883 static int __perf_cgroup_move(void *info)
7885 struct task_struct *task = info;
7886 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7890 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7892 struct task_struct *task;
7894 cgroup_taskset_for_each(task, cgrp, tset)
7895 task_function_call(task, __perf_cgroup_move, task);
7898 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7899 struct task_struct *task)
7902 * cgroup_exit() is called in the copy_process() failure path.
7903 * Ignore this case since the task hasn't ran yet, this avoids
7904 * trying to poke a half freed task state from generic code.
7906 if (!(task->flags & PF_EXITING))
7909 task_function_call(task, __perf_cgroup_move, task);
7912 struct cgroup_subsys perf_subsys = {
7913 .name = "perf_event",
7914 .subsys_id = perf_subsys_id,
7915 .css_alloc = perf_cgroup_css_alloc,
7916 .css_free = perf_cgroup_css_free,
7917 .exit = perf_cgroup_exit,
7918 .attach = perf_cgroup_attach,
7920 #endif /* CONFIG_CGROUP_PERF */