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/cgroup.h>
38 #include <linux/perf_event.h>
39 #include <linux/ftrace_event.h>
40 #include <linux/hw_breakpoint.h>
41 #include <linux/mm_types.h>
42 #include <linux/module.h>
43 #include <linux/mman.h>
44 #include <linux/compat.h>
45 #include <linux/bpf.h>
46 #include <linux/filter.h>
50 #include <asm/irq_regs.h>
52 static struct workqueue_struct *perf_wq;
54 typedef int (*remote_function_f)(void *);
56 struct remote_function_call {
57 struct task_struct *p;
58 remote_function_f func;
63 static void remote_function(void *data)
65 struct remote_function_call *tfc = data;
66 struct task_struct *p = tfc->p;
70 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
74 tfc->ret = tfc->func(tfc->info);
78 * task_function_call - call a function on the cpu on which a task runs
79 * @p: the task to evaluate
80 * @func: the function to be called
81 * @info: the function call argument
83 * Calls the function @func when the task is currently running. This might
84 * be on the current CPU, which just calls the function directly
86 * returns: @func return value, or
87 * -ESRCH - when the process isn't running
88 * -EAGAIN - when the process moved away
91 task_function_call(struct task_struct *p, remote_function_f func, void *info)
93 struct remote_function_call data = {
97 .ret = -ESRCH, /* No such (running) process */
101 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
107 * cpu_function_call - call a function on the cpu
108 * @func: the function to be called
109 * @info: the function call argument
111 * Calls the function @func on the remote cpu.
113 * returns: @func return value or -ENXIO when the cpu is offline
115 static int cpu_function_call(int cpu, remote_function_f func, void *info)
117 struct remote_function_call data = {
121 .ret = -ENXIO, /* No such CPU */
124 smp_call_function_single(cpu, remote_function, &data, 1);
129 #define EVENT_OWNER_KERNEL ((void *) -1)
131 static bool is_kernel_event(struct perf_event *event)
133 return event->owner == EVENT_OWNER_KERNEL;
136 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
137 PERF_FLAG_FD_OUTPUT |\
138 PERF_FLAG_PID_CGROUP |\
139 PERF_FLAG_FD_CLOEXEC)
142 * branch priv levels that need permission checks
144 #define PERF_SAMPLE_BRANCH_PERM_PLM \
145 (PERF_SAMPLE_BRANCH_KERNEL |\
146 PERF_SAMPLE_BRANCH_HV)
149 EVENT_FLEXIBLE = 0x1,
151 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
155 * perf_sched_events : >0 events exist
156 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
158 struct static_key_deferred perf_sched_events __read_mostly;
159 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
160 static DEFINE_PER_CPU(int, perf_sched_cb_usages);
162 static atomic_t nr_mmap_events __read_mostly;
163 static atomic_t nr_comm_events __read_mostly;
164 static atomic_t nr_task_events __read_mostly;
165 static atomic_t nr_freq_events __read_mostly;
167 static LIST_HEAD(pmus);
168 static DEFINE_MUTEX(pmus_lock);
169 static struct srcu_struct pmus_srcu;
172 * perf event paranoia level:
173 * -1 - not paranoid at all
174 * 0 - disallow raw tracepoint access for unpriv
175 * 1 - disallow cpu events for unpriv
176 * 2 - disallow kernel profiling for unpriv
178 int sysctl_perf_event_paranoid __read_mostly = 1;
180 /* Minimum for 512 kiB + 1 user control page */
181 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
184 * max perf event sample rate
186 #define DEFAULT_MAX_SAMPLE_RATE 100000
187 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
188 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
190 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
192 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
193 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
195 static int perf_sample_allowed_ns __read_mostly =
196 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
198 void update_perf_cpu_limits(void)
200 u64 tmp = perf_sample_period_ns;
202 tmp *= sysctl_perf_cpu_time_max_percent;
204 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
207 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
209 int perf_proc_update_handler(struct ctl_table *table, int write,
210 void __user *buffer, size_t *lenp,
213 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
218 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
219 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
220 update_perf_cpu_limits();
225 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
227 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
228 void __user *buffer, size_t *lenp,
231 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
236 update_perf_cpu_limits();
242 * perf samples are done in some very critical code paths (NMIs).
243 * If they take too much CPU time, the system can lock up and not
244 * get any real work done. This will drop the sample rate when
245 * we detect that events are taking too long.
247 #define NR_ACCUMULATED_SAMPLES 128
248 static DEFINE_PER_CPU(u64, running_sample_length);
250 static void perf_duration_warn(struct irq_work *w)
252 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
253 u64 avg_local_sample_len;
254 u64 local_samples_len;
256 local_samples_len = __this_cpu_read(running_sample_length);
257 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
259 printk_ratelimited(KERN_WARNING
260 "perf interrupt took too long (%lld > %lld), lowering "
261 "kernel.perf_event_max_sample_rate to %d\n",
262 avg_local_sample_len, allowed_ns >> 1,
263 sysctl_perf_event_sample_rate);
266 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
268 void perf_sample_event_took(u64 sample_len_ns)
270 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
271 u64 avg_local_sample_len;
272 u64 local_samples_len;
277 /* decay the counter by 1 average sample */
278 local_samples_len = __this_cpu_read(running_sample_length);
279 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
280 local_samples_len += sample_len_ns;
281 __this_cpu_write(running_sample_length, local_samples_len);
284 * note: this will be biased artifically low until we have
285 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
286 * from having to maintain a count.
288 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
290 if (avg_local_sample_len <= allowed_ns)
293 if (max_samples_per_tick <= 1)
296 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
297 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
298 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
300 update_perf_cpu_limits();
302 if (!irq_work_queue(&perf_duration_work)) {
303 early_printk("perf interrupt took too long (%lld > %lld), lowering "
304 "kernel.perf_event_max_sample_rate to %d\n",
305 avg_local_sample_len, allowed_ns >> 1,
306 sysctl_perf_event_sample_rate);
310 static atomic64_t perf_event_id;
312 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
313 enum event_type_t event_type);
315 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
316 enum event_type_t event_type,
317 struct task_struct *task);
319 static void update_context_time(struct perf_event_context *ctx);
320 static u64 perf_event_time(struct perf_event *event);
322 void __weak perf_event_print_debug(void) { }
324 extern __weak const char *perf_pmu_name(void)
329 static inline u64 perf_clock(void)
331 return local_clock();
334 static inline u64 perf_event_clock(struct perf_event *event)
336 return event->clock();
339 static inline struct perf_cpu_context *
340 __get_cpu_context(struct perf_event_context *ctx)
342 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
345 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
346 struct perf_event_context *ctx)
348 raw_spin_lock(&cpuctx->ctx.lock);
350 raw_spin_lock(&ctx->lock);
353 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
354 struct perf_event_context *ctx)
357 raw_spin_unlock(&ctx->lock);
358 raw_spin_unlock(&cpuctx->ctx.lock);
361 #ifdef CONFIG_CGROUP_PERF
364 perf_cgroup_match(struct perf_event *event)
366 struct perf_event_context *ctx = event->ctx;
367 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
369 /* @event doesn't care about cgroup */
373 /* wants specific cgroup scope but @cpuctx isn't associated with any */
378 * Cgroup scoping is recursive. An event enabled for a cgroup is
379 * also enabled for all its descendant cgroups. If @cpuctx's
380 * cgroup is a descendant of @event's (the test covers identity
381 * case), it's a match.
383 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
384 event->cgrp->css.cgroup);
387 static inline void perf_detach_cgroup(struct perf_event *event)
389 css_put(&event->cgrp->css);
393 static inline int is_cgroup_event(struct perf_event *event)
395 return event->cgrp != NULL;
398 static inline u64 perf_cgroup_event_time(struct perf_event *event)
400 struct perf_cgroup_info *t;
402 t = per_cpu_ptr(event->cgrp->info, event->cpu);
406 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
408 struct perf_cgroup_info *info;
413 info = this_cpu_ptr(cgrp->info);
415 info->time += now - info->timestamp;
416 info->timestamp = now;
419 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
421 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
423 __update_cgrp_time(cgrp_out);
426 static inline void update_cgrp_time_from_event(struct perf_event *event)
428 struct perf_cgroup *cgrp;
431 * ensure we access cgroup data only when needed and
432 * when we know the cgroup is pinned (css_get)
434 if (!is_cgroup_event(event))
437 cgrp = perf_cgroup_from_task(current);
439 * Do not update time when cgroup is not active
441 if (cgrp == event->cgrp)
442 __update_cgrp_time(event->cgrp);
446 perf_cgroup_set_timestamp(struct task_struct *task,
447 struct perf_event_context *ctx)
449 struct perf_cgroup *cgrp;
450 struct perf_cgroup_info *info;
453 * ctx->lock held by caller
454 * ensure we do not access cgroup data
455 * unless we have the cgroup pinned (css_get)
457 if (!task || !ctx->nr_cgroups)
460 cgrp = perf_cgroup_from_task(task);
461 info = this_cpu_ptr(cgrp->info);
462 info->timestamp = ctx->timestamp;
465 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
466 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
469 * reschedule events based on the cgroup constraint of task.
471 * mode SWOUT : schedule out everything
472 * mode SWIN : schedule in based on cgroup for next
474 void perf_cgroup_switch(struct task_struct *task, int mode)
476 struct perf_cpu_context *cpuctx;
481 * disable interrupts to avoid geting nr_cgroup
482 * changes via __perf_event_disable(). Also
485 local_irq_save(flags);
488 * we reschedule only in the presence of cgroup
489 * constrained events.
493 list_for_each_entry_rcu(pmu, &pmus, entry) {
494 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
495 if (cpuctx->unique_pmu != pmu)
496 continue; /* ensure we process each cpuctx once */
499 * perf_cgroup_events says at least one
500 * context on this CPU has cgroup events.
502 * ctx->nr_cgroups reports the number of cgroup
503 * events for a context.
505 if (cpuctx->ctx.nr_cgroups > 0) {
506 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
507 perf_pmu_disable(cpuctx->ctx.pmu);
509 if (mode & PERF_CGROUP_SWOUT) {
510 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
512 * must not be done before ctxswout due
513 * to event_filter_match() in event_sched_out()
518 if (mode & PERF_CGROUP_SWIN) {
519 WARN_ON_ONCE(cpuctx->cgrp);
521 * set cgrp before ctxsw in to allow
522 * event_filter_match() to not have to pass
525 cpuctx->cgrp = perf_cgroup_from_task(task);
526 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
528 perf_pmu_enable(cpuctx->ctx.pmu);
529 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
535 local_irq_restore(flags);
538 static inline void perf_cgroup_sched_out(struct task_struct *task,
539 struct task_struct *next)
541 struct perf_cgroup *cgrp1;
542 struct perf_cgroup *cgrp2 = NULL;
545 * we come here when we know perf_cgroup_events > 0
547 cgrp1 = perf_cgroup_from_task(task);
550 * next is NULL when called from perf_event_enable_on_exec()
551 * that will systematically cause a cgroup_switch()
554 cgrp2 = perf_cgroup_from_task(next);
557 * only schedule out current cgroup events if we know
558 * that we are switching to a different cgroup. Otherwise,
559 * do no touch the cgroup events.
562 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
565 static inline void perf_cgroup_sched_in(struct task_struct *prev,
566 struct task_struct *task)
568 struct perf_cgroup *cgrp1;
569 struct perf_cgroup *cgrp2 = NULL;
572 * we come here when we know perf_cgroup_events > 0
574 cgrp1 = perf_cgroup_from_task(task);
576 /* prev can never be NULL */
577 cgrp2 = perf_cgroup_from_task(prev);
580 * only need to schedule in cgroup events if we are changing
581 * cgroup during ctxsw. Cgroup events were not scheduled
582 * out of ctxsw out if that was not the case.
585 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
588 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
589 struct perf_event_attr *attr,
590 struct perf_event *group_leader)
592 struct perf_cgroup *cgrp;
593 struct cgroup_subsys_state *css;
594 struct fd f = fdget(fd);
600 css = css_tryget_online_from_dir(f.file->f_path.dentry,
601 &perf_event_cgrp_subsys);
607 cgrp = container_of(css, struct perf_cgroup, css);
611 * all events in a group must monitor
612 * the same cgroup because a task belongs
613 * to only one perf cgroup at a time
615 if (group_leader && group_leader->cgrp != cgrp) {
616 perf_detach_cgroup(event);
625 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
627 struct perf_cgroup_info *t;
628 t = per_cpu_ptr(event->cgrp->info, event->cpu);
629 event->shadow_ctx_time = now - t->timestamp;
633 perf_cgroup_defer_enabled(struct perf_event *event)
636 * when the current task's perf cgroup does not match
637 * the event's, we need to remember to call the
638 * perf_mark_enable() function the first time a task with
639 * a matching perf cgroup is scheduled in.
641 if (is_cgroup_event(event) && !perf_cgroup_match(event))
642 event->cgrp_defer_enabled = 1;
646 perf_cgroup_mark_enabled(struct perf_event *event,
647 struct perf_event_context *ctx)
649 struct perf_event *sub;
650 u64 tstamp = perf_event_time(event);
652 if (!event->cgrp_defer_enabled)
655 event->cgrp_defer_enabled = 0;
657 event->tstamp_enabled = tstamp - event->total_time_enabled;
658 list_for_each_entry(sub, &event->sibling_list, group_entry) {
659 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
660 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
661 sub->cgrp_defer_enabled = 0;
665 #else /* !CONFIG_CGROUP_PERF */
668 perf_cgroup_match(struct perf_event *event)
673 static inline void perf_detach_cgroup(struct perf_event *event)
676 static inline int is_cgroup_event(struct perf_event *event)
681 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
686 static inline void update_cgrp_time_from_event(struct perf_event *event)
690 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
694 static inline void perf_cgroup_sched_out(struct task_struct *task,
695 struct task_struct *next)
699 static inline void perf_cgroup_sched_in(struct task_struct *prev,
700 struct task_struct *task)
704 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
705 struct perf_event_attr *attr,
706 struct perf_event *group_leader)
712 perf_cgroup_set_timestamp(struct task_struct *task,
713 struct perf_event_context *ctx)
718 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
723 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
727 static inline u64 perf_cgroup_event_time(struct perf_event *event)
733 perf_cgroup_defer_enabled(struct perf_event *event)
738 perf_cgroup_mark_enabled(struct perf_event *event,
739 struct perf_event_context *ctx)
745 * set default to be dependent on timer tick just
748 #define PERF_CPU_HRTIMER (1000 / HZ)
750 * function must be called with interrupts disbled
752 static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr)
754 struct perf_cpu_context *cpuctx;
755 enum hrtimer_restart ret = HRTIMER_NORESTART;
758 WARN_ON(!irqs_disabled());
760 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
762 rotations = perf_rotate_context(cpuctx);
765 * arm timer if needed
768 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
769 ret = HRTIMER_RESTART;
775 /* CPU is going down */
776 void perf_mux_hrtimer_cancel(int cpu)
778 struct perf_cpu_context *cpuctx;
782 if (WARN_ON(cpu != smp_processor_id()))
785 local_irq_save(flags);
789 list_for_each_entry_rcu(pmu, &pmus, entry) {
790 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
792 if (pmu->task_ctx_nr == perf_sw_context)
795 hrtimer_cancel(&cpuctx->hrtimer);
800 local_irq_restore(flags);
803 static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
805 struct hrtimer *timer = &cpuctx->hrtimer;
806 struct pmu *pmu = cpuctx->ctx.pmu;
809 /* no multiplexing needed for SW PMU */
810 if (pmu->task_ctx_nr == perf_sw_context)
814 * check default is sane, if not set then force to
815 * default interval (1/tick)
817 interval = pmu->hrtimer_interval_ms;
819 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
821 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval);
823 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
824 timer->function = perf_mux_hrtimer_handler;
827 static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
829 struct hrtimer *timer = &cpuctx->hrtimer;
830 struct pmu *pmu = cpuctx->ctx.pmu;
833 if (pmu->task_ctx_nr == perf_sw_context)
836 if (hrtimer_is_queued(timer))
839 hrtimer_start(timer, cpuctx->hrtimer_interval, HRTIMER_MODE_REL_PINNED);
843 void perf_pmu_disable(struct pmu *pmu)
845 int *count = this_cpu_ptr(pmu->pmu_disable_count);
847 pmu->pmu_disable(pmu);
850 void perf_pmu_enable(struct pmu *pmu)
852 int *count = this_cpu_ptr(pmu->pmu_disable_count);
854 pmu->pmu_enable(pmu);
857 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
860 * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
861 * perf_event_task_tick() are fully serialized because they're strictly cpu
862 * affine and perf_event_ctx{activate,deactivate} are called with IRQs
863 * disabled, while perf_event_task_tick is called from IRQ context.
865 static void perf_event_ctx_activate(struct perf_event_context *ctx)
867 struct list_head *head = this_cpu_ptr(&active_ctx_list);
869 WARN_ON(!irqs_disabled());
871 WARN_ON(!list_empty(&ctx->active_ctx_list));
873 list_add(&ctx->active_ctx_list, head);
876 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
878 WARN_ON(!irqs_disabled());
880 WARN_ON(list_empty(&ctx->active_ctx_list));
882 list_del_init(&ctx->active_ctx_list);
885 static void get_ctx(struct perf_event_context *ctx)
887 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
890 static void free_ctx(struct rcu_head *head)
892 struct perf_event_context *ctx;
894 ctx = container_of(head, struct perf_event_context, rcu_head);
895 kfree(ctx->task_ctx_data);
899 static void put_ctx(struct perf_event_context *ctx)
901 if (atomic_dec_and_test(&ctx->refcount)) {
903 put_ctx(ctx->parent_ctx);
905 put_task_struct(ctx->task);
906 call_rcu(&ctx->rcu_head, free_ctx);
911 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
912 * perf_pmu_migrate_context() we need some magic.
914 * Those places that change perf_event::ctx will hold both
915 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
917 * Lock ordering is by mutex address. There is one other site where
918 * perf_event_context::mutex nests and that is put_event(). But remember that
919 * that is a parent<->child context relation, and migration does not affect
920 * children, therefore these two orderings should not interact.
922 * The change in perf_event::ctx does not affect children (as claimed above)
923 * because the sys_perf_event_open() case will install a new event and break
924 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
925 * concerned with cpuctx and that doesn't have children.
927 * The places that change perf_event::ctx will issue:
929 * perf_remove_from_context();
931 * perf_install_in_context();
933 * to affect the change. The remove_from_context() + synchronize_rcu() should
934 * quiesce the event, after which we can install it in the new location. This
935 * means that only external vectors (perf_fops, prctl) can perturb the event
936 * while in transit. Therefore all such accessors should also acquire
937 * perf_event_context::mutex to serialize against this.
939 * However; because event->ctx can change while we're waiting to acquire
940 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
944 * task_struct::perf_event_mutex
945 * perf_event_context::mutex
946 * perf_event_context::lock
947 * perf_event::child_mutex;
948 * perf_event::mmap_mutex
951 static struct perf_event_context *
952 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
954 struct perf_event_context *ctx;
958 ctx = ACCESS_ONCE(event->ctx);
959 if (!atomic_inc_not_zero(&ctx->refcount)) {
965 mutex_lock_nested(&ctx->mutex, nesting);
966 if (event->ctx != ctx) {
967 mutex_unlock(&ctx->mutex);
975 static inline struct perf_event_context *
976 perf_event_ctx_lock(struct perf_event *event)
978 return perf_event_ctx_lock_nested(event, 0);
981 static void perf_event_ctx_unlock(struct perf_event *event,
982 struct perf_event_context *ctx)
984 mutex_unlock(&ctx->mutex);
989 * This must be done under the ctx->lock, such as to serialize against
990 * context_equiv(), therefore we cannot call put_ctx() since that might end up
991 * calling scheduler related locks and ctx->lock nests inside those.
993 static __must_check struct perf_event_context *
994 unclone_ctx(struct perf_event_context *ctx)
996 struct perf_event_context *parent_ctx = ctx->parent_ctx;
998 lockdep_assert_held(&ctx->lock);
1001 ctx->parent_ctx = NULL;
1007 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1010 * only top level events have the pid namespace they were created in
1013 event = event->parent;
1015 return task_tgid_nr_ns(p, event->ns);
1018 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1021 * only top level events have the pid namespace they were created in
1024 event = event->parent;
1026 return task_pid_nr_ns(p, event->ns);
1030 * If we inherit events we want to return the parent event id
1033 static u64 primary_event_id(struct perf_event *event)
1038 id = event->parent->id;
1044 * Get the perf_event_context for a task and lock it.
1045 * This has to cope with with the fact that until it is locked,
1046 * the context could get moved to another task.
1048 static struct perf_event_context *
1049 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1051 struct perf_event_context *ctx;
1055 * One of the few rules of preemptible RCU is that one cannot do
1056 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1057 * part of the read side critical section was preemptible -- see
1058 * rcu_read_unlock_special().
1060 * Since ctx->lock nests under rq->lock we must ensure the entire read
1061 * side critical section is non-preemptible.
1065 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1068 * If this context is a clone of another, it might
1069 * get swapped for another underneath us by
1070 * perf_event_task_sched_out, though the
1071 * rcu_read_lock() protects us from any context
1072 * getting freed. Lock the context and check if it
1073 * got swapped before we could get the lock, and retry
1074 * if so. If we locked the right context, then it
1075 * can't get swapped on us any more.
1077 raw_spin_lock_irqsave(&ctx->lock, *flags);
1078 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1079 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1085 if (!atomic_inc_not_zero(&ctx->refcount)) {
1086 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1096 * Get the context for a task and increment its pin_count so it
1097 * can't get swapped to another task. This also increments its
1098 * reference count so that the context can't get freed.
1100 static struct perf_event_context *
1101 perf_pin_task_context(struct task_struct *task, int ctxn)
1103 struct perf_event_context *ctx;
1104 unsigned long flags;
1106 ctx = perf_lock_task_context(task, ctxn, &flags);
1109 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1114 static void perf_unpin_context(struct perf_event_context *ctx)
1116 unsigned long flags;
1118 raw_spin_lock_irqsave(&ctx->lock, flags);
1120 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1124 * Update the record of the current time in a context.
1126 static void update_context_time(struct perf_event_context *ctx)
1128 u64 now = perf_clock();
1130 ctx->time += now - ctx->timestamp;
1131 ctx->timestamp = now;
1134 static u64 perf_event_time(struct perf_event *event)
1136 struct perf_event_context *ctx = event->ctx;
1138 if (is_cgroup_event(event))
1139 return perf_cgroup_event_time(event);
1141 return ctx ? ctx->time : 0;
1145 * Update the total_time_enabled and total_time_running fields for a event.
1146 * The caller of this function needs to hold the ctx->lock.
1148 static void update_event_times(struct perf_event *event)
1150 struct perf_event_context *ctx = event->ctx;
1153 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1154 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1157 * in cgroup mode, time_enabled represents
1158 * the time the event was enabled AND active
1159 * tasks were in the monitored cgroup. This is
1160 * independent of the activity of the context as
1161 * there may be a mix of cgroup and non-cgroup events.
1163 * That is why we treat cgroup events differently
1166 if (is_cgroup_event(event))
1167 run_end = perf_cgroup_event_time(event);
1168 else if (ctx->is_active)
1169 run_end = ctx->time;
1171 run_end = event->tstamp_stopped;
1173 event->total_time_enabled = run_end - event->tstamp_enabled;
1175 if (event->state == PERF_EVENT_STATE_INACTIVE)
1176 run_end = event->tstamp_stopped;
1178 run_end = perf_event_time(event);
1180 event->total_time_running = run_end - event->tstamp_running;
1185 * Update total_time_enabled and total_time_running for all events in a group.
1187 static void update_group_times(struct perf_event *leader)
1189 struct perf_event *event;
1191 update_event_times(leader);
1192 list_for_each_entry(event, &leader->sibling_list, group_entry)
1193 update_event_times(event);
1196 static struct list_head *
1197 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1199 if (event->attr.pinned)
1200 return &ctx->pinned_groups;
1202 return &ctx->flexible_groups;
1206 * Add a event from the lists for its context.
1207 * Must be called with ctx->mutex and ctx->lock held.
1210 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1212 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1213 event->attach_state |= PERF_ATTACH_CONTEXT;
1216 * If we're a stand alone event or group leader, we go to the context
1217 * list, group events are kept attached to the group so that
1218 * perf_group_detach can, at all times, locate all siblings.
1220 if (event->group_leader == event) {
1221 struct list_head *list;
1223 if (is_software_event(event))
1224 event->group_flags |= PERF_GROUP_SOFTWARE;
1226 list = ctx_group_list(event, ctx);
1227 list_add_tail(&event->group_entry, list);
1230 if (is_cgroup_event(event))
1233 list_add_rcu(&event->event_entry, &ctx->event_list);
1235 if (event->attr.inherit_stat)
1242 * Initialize event state based on the perf_event_attr::disabled.
1244 static inline void perf_event__state_init(struct perf_event *event)
1246 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1247 PERF_EVENT_STATE_INACTIVE;
1251 * Called at perf_event creation and when events are attached/detached from a
1254 static void perf_event__read_size(struct perf_event *event)
1256 int entry = sizeof(u64); /* value */
1260 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1261 size += sizeof(u64);
1263 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1264 size += sizeof(u64);
1266 if (event->attr.read_format & PERF_FORMAT_ID)
1267 entry += sizeof(u64);
1269 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1270 nr += event->group_leader->nr_siblings;
1271 size += sizeof(u64);
1275 event->read_size = size;
1278 static void perf_event__header_size(struct perf_event *event)
1280 struct perf_sample_data *data;
1281 u64 sample_type = event->attr.sample_type;
1284 perf_event__read_size(event);
1286 if (sample_type & PERF_SAMPLE_IP)
1287 size += sizeof(data->ip);
1289 if (sample_type & PERF_SAMPLE_ADDR)
1290 size += sizeof(data->addr);
1292 if (sample_type & PERF_SAMPLE_PERIOD)
1293 size += sizeof(data->period);
1295 if (sample_type & PERF_SAMPLE_WEIGHT)
1296 size += sizeof(data->weight);
1298 if (sample_type & PERF_SAMPLE_READ)
1299 size += event->read_size;
1301 if (sample_type & PERF_SAMPLE_DATA_SRC)
1302 size += sizeof(data->data_src.val);
1304 if (sample_type & PERF_SAMPLE_TRANSACTION)
1305 size += sizeof(data->txn);
1307 event->header_size = size;
1310 static void perf_event__id_header_size(struct perf_event *event)
1312 struct perf_sample_data *data;
1313 u64 sample_type = event->attr.sample_type;
1316 if (sample_type & PERF_SAMPLE_TID)
1317 size += sizeof(data->tid_entry);
1319 if (sample_type & PERF_SAMPLE_TIME)
1320 size += sizeof(data->time);
1322 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1323 size += sizeof(data->id);
1325 if (sample_type & PERF_SAMPLE_ID)
1326 size += sizeof(data->id);
1328 if (sample_type & PERF_SAMPLE_STREAM_ID)
1329 size += sizeof(data->stream_id);
1331 if (sample_type & PERF_SAMPLE_CPU)
1332 size += sizeof(data->cpu_entry);
1334 event->id_header_size = size;
1337 static void perf_group_attach(struct perf_event *event)
1339 struct perf_event *group_leader = event->group_leader, *pos;
1342 * We can have double attach due to group movement in perf_event_open.
1344 if (event->attach_state & PERF_ATTACH_GROUP)
1347 event->attach_state |= PERF_ATTACH_GROUP;
1349 if (group_leader == event)
1352 WARN_ON_ONCE(group_leader->ctx != event->ctx);
1354 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1355 !is_software_event(event))
1356 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1358 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1359 group_leader->nr_siblings++;
1361 perf_event__header_size(group_leader);
1363 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1364 perf_event__header_size(pos);
1368 * Remove a event from the lists for its context.
1369 * Must be called with ctx->mutex and ctx->lock held.
1372 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1374 struct perf_cpu_context *cpuctx;
1376 WARN_ON_ONCE(event->ctx != ctx);
1377 lockdep_assert_held(&ctx->lock);
1380 * We can have double detach due to exit/hot-unplug + close.
1382 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1385 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1387 if (is_cgroup_event(event)) {
1389 cpuctx = __get_cpu_context(ctx);
1391 * if there are no more cgroup events
1392 * then cler cgrp to avoid stale pointer
1393 * in update_cgrp_time_from_cpuctx()
1395 if (!ctx->nr_cgroups)
1396 cpuctx->cgrp = NULL;
1400 if (event->attr.inherit_stat)
1403 list_del_rcu(&event->event_entry);
1405 if (event->group_leader == event)
1406 list_del_init(&event->group_entry);
1408 update_group_times(event);
1411 * If event was in error state, then keep it
1412 * that way, otherwise bogus counts will be
1413 * returned on read(). The only way to get out
1414 * of error state is by explicit re-enabling
1417 if (event->state > PERF_EVENT_STATE_OFF)
1418 event->state = PERF_EVENT_STATE_OFF;
1423 static void perf_group_detach(struct perf_event *event)
1425 struct perf_event *sibling, *tmp;
1426 struct list_head *list = NULL;
1429 * We can have double detach due to exit/hot-unplug + close.
1431 if (!(event->attach_state & PERF_ATTACH_GROUP))
1434 event->attach_state &= ~PERF_ATTACH_GROUP;
1437 * If this is a sibling, remove it from its group.
1439 if (event->group_leader != event) {
1440 list_del_init(&event->group_entry);
1441 event->group_leader->nr_siblings--;
1445 if (!list_empty(&event->group_entry))
1446 list = &event->group_entry;
1449 * If this was a group event with sibling events then
1450 * upgrade the siblings to singleton events by adding them
1451 * to whatever list we are on.
1453 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1455 list_move_tail(&sibling->group_entry, list);
1456 sibling->group_leader = sibling;
1458 /* Inherit group flags from the previous leader */
1459 sibling->group_flags = event->group_flags;
1461 WARN_ON_ONCE(sibling->ctx != event->ctx);
1465 perf_event__header_size(event->group_leader);
1467 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1468 perf_event__header_size(tmp);
1472 * User event without the task.
1474 static bool is_orphaned_event(struct perf_event *event)
1476 return event && !is_kernel_event(event) && !event->owner;
1480 * Event has a parent but parent's task finished and it's
1481 * alive only because of children holding refference.
1483 static bool is_orphaned_child(struct perf_event *event)
1485 return is_orphaned_event(event->parent);
1488 static void orphans_remove_work(struct work_struct *work);
1490 static void schedule_orphans_remove(struct perf_event_context *ctx)
1492 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1495 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1497 ctx->orphans_remove_sched = true;
1501 static int __init perf_workqueue_init(void)
1503 perf_wq = create_singlethread_workqueue("perf");
1504 WARN(!perf_wq, "failed to create perf workqueue\n");
1505 return perf_wq ? 0 : -1;
1508 core_initcall(perf_workqueue_init);
1511 event_filter_match(struct perf_event *event)
1513 return (event->cpu == -1 || event->cpu == smp_processor_id())
1514 && perf_cgroup_match(event);
1518 event_sched_out(struct perf_event *event,
1519 struct perf_cpu_context *cpuctx,
1520 struct perf_event_context *ctx)
1522 u64 tstamp = perf_event_time(event);
1525 WARN_ON_ONCE(event->ctx != ctx);
1526 lockdep_assert_held(&ctx->lock);
1529 * An event which could not be activated because of
1530 * filter mismatch still needs to have its timings
1531 * maintained, otherwise bogus information is return
1532 * via read() for time_enabled, time_running:
1534 if (event->state == PERF_EVENT_STATE_INACTIVE
1535 && !event_filter_match(event)) {
1536 delta = tstamp - event->tstamp_stopped;
1537 event->tstamp_running += delta;
1538 event->tstamp_stopped = tstamp;
1541 if (event->state != PERF_EVENT_STATE_ACTIVE)
1544 perf_pmu_disable(event->pmu);
1546 event->state = PERF_EVENT_STATE_INACTIVE;
1547 if (event->pending_disable) {
1548 event->pending_disable = 0;
1549 event->state = PERF_EVENT_STATE_OFF;
1551 event->tstamp_stopped = tstamp;
1552 event->pmu->del(event, 0);
1555 if (!is_software_event(event))
1556 cpuctx->active_oncpu--;
1557 if (!--ctx->nr_active)
1558 perf_event_ctx_deactivate(ctx);
1559 if (event->attr.freq && event->attr.sample_freq)
1561 if (event->attr.exclusive || !cpuctx->active_oncpu)
1562 cpuctx->exclusive = 0;
1564 if (is_orphaned_child(event))
1565 schedule_orphans_remove(ctx);
1567 perf_pmu_enable(event->pmu);
1571 group_sched_out(struct perf_event *group_event,
1572 struct perf_cpu_context *cpuctx,
1573 struct perf_event_context *ctx)
1575 struct perf_event *event;
1576 int state = group_event->state;
1578 event_sched_out(group_event, cpuctx, ctx);
1581 * Schedule out siblings (if any):
1583 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1584 event_sched_out(event, cpuctx, ctx);
1586 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1587 cpuctx->exclusive = 0;
1590 struct remove_event {
1591 struct perf_event *event;
1596 * Cross CPU call to remove a performance event
1598 * We disable the event on the hardware level first. After that we
1599 * remove it from the context list.
1601 static int __perf_remove_from_context(void *info)
1603 struct remove_event *re = info;
1604 struct perf_event *event = re->event;
1605 struct perf_event_context *ctx = event->ctx;
1606 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1608 raw_spin_lock(&ctx->lock);
1609 event_sched_out(event, cpuctx, ctx);
1610 if (re->detach_group)
1611 perf_group_detach(event);
1612 list_del_event(event, ctx);
1613 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1615 cpuctx->task_ctx = NULL;
1617 raw_spin_unlock(&ctx->lock);
1624 * Remove the event from a task's (or a CPU's) list of events.
1626 * CPU events are removed with a smp call. For task events we only
1627 * call when the task is on a CPU.
1629 * If event->ctx is a cloned context, callers must make sure that
1630 * every task struct that event->ctx->task could possibly point to
1631 * remains valid. This is OK when called from perf_release since
1632 * that only calls us on the top-level context, which can't be a clone.
1633 * When called from perf_event_exit_task, it's OK because the
1634 * context has been detached from its task.
1636 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1638 struct perf_event_context *ctx = event->ctx;
1639 struct task_struct *task = ctx->task;
1640 struct remove_event re = {
1642 .detach_group = detach_group,
1645 lockdep_assert_held(&ctx->mutex);
1649 * Per cpu events are removed via an smp call. The removal can
1650 * fail if the CPU is currently offline, but in that case we
1651 * already called __perf_remove_from_context from
1652 * perf_event_exit_cpu.
1654 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1659 if (!task_function_call(task, __perf_remove_from_context, &re))
1662 raw_spin_lock_irq(&ctx->lock);
1664 * If we failed to find a running task, but find the context active now
1665 * that we've acquired the ctx->lock, retry.
1667 if (ctx->is_active) {
1668 raw_spin_unlock_irq(&ctx->lock);
1670 * Reload the task pointer, it might have been changed by
1671 * a concurrent perf_event_context_sched_out().
1678 * Since the task isn't running, its safe to remove the event, us
1679 * holding the ctx->lock ensures the task won't get scheduled in.
1682 perf_group_detach(event);
1683 list_del_event(event, ctx);
1684 raw_spin_unlock_irq(&ctx->lock);
1688 * Cross CPU call to disable a performance event
1690 int __perf_event_disable(void *info)
1692 struct perf_event *event = info;
1693 struct perf_event_context *ctx = event->ctx;
1694 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1697 * If this is a per-task event, need to check whether this
1698 * event's task is the current task on this cpu.
1700 * Can trigger due to concurrent perf_event_context_sched_out()
1701 * flipping contexts around.
1703 if (ctx->task && cpuctx->task_ctx != ctx)
1706 raw_spin_lock(&ctx->lock);
1709 * If the event is on, turn it off.
1710 * If it is in error state, leave it in error state.
1712 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1713 update_context_time(ctx);
1714 update_cgrp_time_from_event(event);
1715 update_group_times(event);
1716 if (event == event->group_leader)
1717 group_sched_out(event, cpuctx, ctx);
1719 event_sched_out(event, cpuctx, ctx);
1720 event->state = PERF_EVENT_STATE_OFF;
1723 raw_spin_unlock(&ctx->lock);
1731 * If event->ctx is a cloned context, callers must make sure that
1732 * every task struct that event->ctx->task could possibly point to
1733 * remains valid. This condition is satisifed when called through
1734 * perf_event_for_each_child or perf_event_for_each because they
1735 * hold the top-level event's child_mutex, so any descendant that
1736 * goes to exit will block in sync_child_event.
1737 * When called from perf_pending_event it's OK because event->ctx
1738 * is the current context on this CPU and preemption is disabled,
1739 * hence we can't get into perf_event_task_sched_out for this context.
1741 static void _perf_event_disable(struct perf_event *event)
1743 struct perf_event_context *ctx = event->ctx;
1744 struct task_struct *task = ctx->task;
1748 * Disable the event on the cpu that it's on
1750 cpu_function_call(event->cpu, __perf_event_disable, event);
1755 if (!task_function_call(task, __perf_event_disable, event))
1758 raw_spin_lock_irq(&ctx->lock);
1760 * If the event is still active, we need to retry the cross-call.
1762 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1763 raw_spin_unlock_irq(&ctx->lock);
1765 * Reload the task pointer, it might have been changed by
1766 * a concurrent perf_event_context_sched_out().
1773 * Since we have the lock this context can't be scheduled
1774 * in, so we can change the state safely.
1776 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1777 update_group_times(event);
1778 event->state = PERF_EVENT_STATE_OFF;
1780 raw_spin_unlock_irq(&ctx->lock);
1784 * Strictly speaking kernel users cannot create groups and therefore this
1785 * interface does not need the perf_event_ctx_lock() magic.
1787 void perf_event_disable(struct perf_event *event)
1789 struct perf_event_context *ctx;
1791 ctx = perf_event_ctx_lock(event);
1792 _perf_event_disable(event);
1793 perf_event_ctx_unlock(event, ctx);
1795 EXPORT_SYMBOL_GPL(perf_event_disable);
1797 static void perf_set_shadow_time(struct perf_event *event,
1798 struct perf_event_context *ctx,
1802 * use the correct time source for the time snapshot
1804 * We could get by without this by leveraging the
1805 * fact that to get to this function, the caller
1806 * has most likely already called update_context_time()
1807 * and update_cgrp_time_xx() and thus both timestamp
1808 * are identical (or very close). Given that tstamp is,
1809 * already adjusted for cgroup, we could say that:
1810 * tstamp - ctx->timestamp
1812 * tstamp - cgrp->timestamp.
1814 * Then, in perf_output_read(), the calculation would
1815 * work with no changes because:
1816 * - event is guaranteed scheduled in
1817 * - no scheduled out in between
1818 * - thus the timestamp would be the same
1820 * But this is a bit hairy.
1822 * So instead, we have an explicit cgroup call to remain
1823 * within the time time source all along. We believe it
1824 * is cleaner and simpler to understand.
1826 if (is_cgroup_event(event))
1827 perf_cgroup_set_shadow_time(event, tstamp);
1829 event->shadow_ctx_time = tstamp - ctx->timestamp;
1832 #define MAX_INTERRUPTS (~0ULL)
1834 static void perf_log_throttle(struct perf_event *event, int enable);
1835 static void perf_log_itrace_start(struct perf_event *event);
1838 event_sched_in(struct perf_event *event,
1839 struct perf_cpu_context *cpuctx,
1840 struct perf_event_context *ctx)
1842 u64 tstamp = perf_event_time(event);
1845 lockdep_assert_held(&ctx->lock);
1847 if (event->state <= PERF_EVENT_STATE_OFF)
1850 event->state = PERF_EVENT_STATE_ACTIVE;
1851 event->oncpu = smp_processor_id();
1854 * Unthrottle events, since we scheduled we might have missed several
1855 * ticks already, also for a heavily scheduling task there is little
1856 * guarantee it'll get a tick in a timely manner.
1858 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1859 perf_log_throttle(event, 1);
1860 event->hw.interrupts = 0;
1864 * The new state must be visible before we turn it on in the hardware:
1868 perf_pmu_disable(event->pmu);
1870 event->tstamp_running += tstamp - event->tstamp_stopped;
1872 perf_set_shadow_time(event, ctx, tstamp);
1874 perf_log_itrace_start(event);
1876 if (event->pmu->add(event, PERF_EF_START)) {
1877 event->state = PERF_EVENT_STATE_INACTIVE;
1883 if (!is_software_event(event))
1884 cpuctx->active_oncpu++;
1885 if (!ctx->nr_active++)
1886 perf_event_ctx_activate(ctx);
1887 if (event->attr.freq && event->attr.sample_freq)
1890 if (event->attr.exclusive)
1891 cpuctx->exclusive = 1;
1893 if (is_orphaned_child(event))
1894 schedule_orphans_remove(ctx);
1897 perf_pmu_enable(event->pmu);
1903 group_sched_in(struct perf_event *group_event,
1904 struct perf_cpu_context *cpuctx,
1905 struct perf_event_context *ctx)
1907 struct perf_event *event, *partial_group = NULL;
1908 struct pmu *pmu = ctx->pmu;
1909 u64 now = ctx->time;
1910 bool simulate = false;
1912 if (group_event->state == PERF_EVENT_STATE_OFF)
1915 pmu->start_txn(pmu);
1917 if (event_sched_in(group_event, cpuctx, ctx)) {
1918 pmu->cancel_txn(pmu);
1919 perf_mux_hrtimer_restart(cpuctx);
1924 * Schedule in siblings as one group (if any):
1926 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1927 if (event_sched_in(event, cpuctx, ctx)) {
1928 partial_group = event;
1933 if (!pmu->commit_txn(pmu))
1938 * Groups can be scheduled in as one unit only, so undo any
1939 * partial group before returning:
1940 * The events up to the failed event are scheduled out normally,
1941 * tstamp_stopped will be updated.
1943 * The failed events and the remaining siblings need to have
1944 * their timings updated as if they had gone thru event_sched_in()
1945 * and event_sched_out(). This is required to get consistent timings
1946 * across the group. This also takes care of the case where the group
1947 * could never be scheduled by ensuring tstamp_stopped is set to mark
1948 * the time the event was actually stopped, such that time delta
1949 * calculation in update_event_times() is correct.
1951 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1952 if (event == partial_group)
1956 event->tstamp_running += now - event->tstamp_stopped;
1957 event->tstamp_stopped = now;
1959 event_sched_out(event, cpuctx, ctx);
1962 event_sched_out(group_event, cpuctx, ctx);
1964 pmu->cancel_txn(pmu);
1966 perf_mux_hrtimer_restart(cpuctx);
1972 * Work out whether we can put this event group on the CPU now.
1974 static int group_can_go_on(struct perf_event *event,
1975 struct perf_cpu_context *cpuctx,
1979 * Groups consisting entirely of software events can always go on.
1981 if (event->group_flags & PERF_GROUP_SOFTWARE)
1984 * If an exclusive group is already on, no other hardware
1987 if (cpuctx->exclusive)
1990 * If this group is exclusive and there are already
1991 * events on the CPU, it can't go on.
1993 if (event->attr.exclusive && cpuctx->active_oncpu)
1996 * Otherwise, try to add it if all previous groups were able
2002 static void add_event_to_ctx(struct perf_event *event,
2003 struct perf_event_context *ctx)
2005 u64 tstamp = perf_event_time(event);
2007 list_add_event(event, ctx);
2008 perf_group_attach(event);
2009 event->tstamp_enabled = tstamp;
2010 event->tstamp_running = tstamp;
2011 event->tstamp_stopped = tstamp;
2014 static void task_ctx_sched_out(struct perf_event_context *ctx);
2016 ctx_sched_in(struct perf_event_context *ctx,
2017 struct perf_cpu_context *cpuctx,
2018 enum event_type_t event_type,
2019 struct task_struct *task);
2021 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2022 struct perf_event_context *ctx,
2023 struct task_struct *task)
2025 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2027 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2028 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2030 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2034 * Cross CPU call to install and enable a performance event
2036 * Must be called with ctx->mutex held
2038 static int __perf_install_in_context(void *info)
2040 struct perf_event *event = info;
2041 struct perf_event_context *ctx = event->ctx;
2042 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2043 struct perf_event_context *task_ctx = cpuctx->task_ctx;
2044 struct task_struct *task = current;
2046 perf_ctx_lock(cpuctx, task_ctx);
2047 perf_pmu_disable(cpuctx->ctx.pmu);
2050 * If there was an active task_ctx schedule it out.
2053 task_ctx_sched_out(task_ctx);
2056 * If the context we're installing events in is not the
2057 * active task_ctx, flip them.
2059 if (ctx->task && task_ctx != ctx) {
2061 raw_spin_unlock(&task_ctx->lock);
2062 raw_spin_lock(&ctx->lock);
2067 cpuctx->task_ctx = task_ctx;
2068 task = task_ctx->task;
2071 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
2073 update_context_time(ctx);
2075 * update cgrp time only if current cgrp
2076 * matches event->cgrp. Must be done before
2077 * calling add_event_to_ctx()
2079 update_cgrp_time_from_event(event);
2081 add_event_to_ctx(event, ctx);
2084 * Schedule everything back in
2086 perf_event_sched_in(cpuctx, task_ctx, task);
2088 perf_pmu_enable(cpuctx->ctx.pmu);
2089 perf_ctx_unlock(cpuctx, task_ctx);
2095 * Attach a performance event to a context
2097 * First we add the event to the list with the hardware enable bit
2098 * in event->hw_config cleared.
2100 * If the event is attached to a task which is on a CPU we use a smp
2101 * call to enable it in the task context. The task might have been
2102 * scheduled away, but we check this in the smp call again.
2105 perf_install_in_context(struct perf_event_context *ctx,
2106 struct perf_event *event,
2109 struct task_struct *task = ctx->task;
2111 lockdep_assert_held(&ctx->mutex);
2114 if (event->cpu != -1)
2119 * Per cpu events are installed via an smp call and
2120 * the install is always successful.
2122 cpu_function_call(cpu, __perf_install_in_context, event);
2127 if (!task_function_call(task, __perf_install_in_context, event))
2130 raw_spin_lock_irq(&ctx->lock);
2132 * If we failed to find a running task, but find the context active now
2133 * that we've acquired the ctx->lock, retry.
2135 if (ctx->is_active) {
2136 raw_spin_unlock_irq(&ctx->lock);
2138 * Reload the task pointer, it might have been changed by
2139 * a concurrent perf_event_context_sched_out().
2146 * Since the task isn't running, its safe to add the event, us holding
2147 * the ctx->lock ensures the task won't get scheduled in.
2149 add_event_to_ctx(event, ctx);
2150 raw_spin_unlock_irq(&ctx->lock);
2154 * Put a event into inactive state and update time fields.
2155 * Enabling the leader of a group effectively enables all
2156 * the group members that aren't explicitly disabled, so we
2157 * have to update their ->tstamp_enabled also.
2158 * Note: this works for group members as well as group leaders
2159 * since the non-leader members' sibling_lists will be empty.
2161 static void __perf_event_mark_enabled(struct perf_event *event)
2163 struct perf_event *sub;
2164 u64 tstamp = perf_event_time(event);
2166 event->state = PERF_EVENT_STATE_INACTIVE;
2167 event->tstamp_enabled = tstamp - event->total_time_enabled;
2168 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2169 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2170 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2175 * Cross CPU call to enable a performance event
2177 static int __perf_event_enable(void *info)
2179 struct perf_event *event = info;
2180 struct perf_event_context *ctx = event->ctx;
2181 struct perf_event *leader = event->group_leader;
2182 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2186 * There's a time window between 'ctx->is_active' check
2187 * in perf_event_enable function and this place having:
2189 * - ctx->lock unlocked
2191 * where the task could be killed and 'ctx' deactivated
2192 * by perf_event_exit_task.
2194 if (!ctx->is_active)
2197 raw_spin_lock(&ctx->lock);
2198 update_context_time(ctx);
2200 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2204 * set current task's cgroup time reference point
2206 perf_cgroup_set_timestamp(current, ctx);
2208 __perf_event_mark_enabled(event);
2210 if (!event_filter_match(event)) {
2211 if (is_cgroup_event(event))
2212 perf_cgroup_defer_enabled(event);
2217 * If the event is in a group and isn't the group leader,
2218 * then don't put it on unless the group is on.
2220 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2223 if (!group_can_go_on(event, cpuctx, 1)) {
2226 if (event == leader)
2227 err = group_sched_in(event, cpuctx, ctx);
2229 err = event_sched_in(event, cpuctx, ctx);
2234 * If this event can't go on and it's part of a
2235 * group, then the whole group has to come off.
2237 if (leader != event) {
2238 group_sched_out(leader, cpuctx, ctx);
2239 perf_mux_hrtimer_restart(cpuctx);
2241 if (leader->attr.pinned) {
2242 update_group_times(leader);
2243 leader->state = PERF_EVENT_STATE_ERROR;
2248 raw_spin_unlock(&ctx->lock);
2256 * If event->ctx is a cloned context, callers must make sure that
2257 * every task struct that event->ctx->task could possibly point to
2258 * remains valid. This condition is satisfied when called through
2259 * perf_event_for_each_child or perf_event_for_each as described
2260 * for perf_event_disable.
2262 static void _perf_event_enable(struct perf_event *event)
2264 struct perf_event_context *ctx = event->ctx;
2265 struct task_struct *task = ctx->task;
2269 * Enable the event on the cpu that it's on
2271 cpu_function_call(event->cpu, __perf_event_enable, event);
2275 raw_spin_lock_irq(&ctx->lock);
2276 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2280 * If the event is in error state, clear that first.
2281 * That way, if we see the event in error state below, we
2282 * know that it has gone back into error state, as distinct
2283 * from the task having been scheduled away before the
2284 * cross-call arrived.
2286 if (event->state == PERF_EVENT_STATE_ERROR)
2287 event->state = PERF_EVENT_STATE_OFF;
2290 if (!ctx->is_active) {
2291 __perf_event_mark_enabled(event);
2295 raw_spin_unlock_irq(&ctx->lock);
2297 if (!task_function_call(task, __perf_event_enable, event))
2300 raw_spin_lock_irq(&ctx->lock);
2303 * If the context is active and the event is still off,
2304 * we need to retry the cross-call.
2306 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2308 * task could have been flipped by a concurrent
2309 * perf_event_context_sched_out()
2316 raw_spin_unlock_irq(&ctx->lock);
2320 * See perf_event_disable();
2322 void perf_event_enable(struct perf_event *event)
2324 struct perf_event_context *ctx;
2326 ctx = perf_event_ctx_lock(event);
2327 _perf_event_enable(event);
2328 perf_event_ctx_unlock(event, ctx);
2330 EXPORT_SYMBOL_GPL(perf_event_enable);
2332 static int _perf_event_refresh(struct perf_event *event, int refresh)
2335 * not supported on inherited events
2337 if (event->attr.inherit || !is_sampling_event(event))
2340 atomic_add(refresh, &event->event_limit);
2341 _perf_event_enable(event);
2347 * See perf_event_disable()
2349 int perf_event_refresh(struct perf_event *event, int refresh)
2351 struct perf_event_context *ctx;
2354 ctx = perf_event_ctx_lock(event);
2355 ret = _perf_event_refresh(event, refresh);
2356 perf_event_ctx_unlock(event, ctx);
2360 EXPORT_SYMBOL_GPL(perf_event_refresh);
2362 static void ctx_sched_out(struct perf_event_context *ctx,
2363 struct perf_cpu_context *cpuctx,
2364 enum event_type_t event_type)
2366 struct perf_event *event;
2367 int is_active = ctx->is_active;
2369 ctx->is_active &= ~event_type;
2370 if (likely(!ctx->nr_events))
2373 update_context_time(ctx);
2374 update_cgrp_time_from_cpuctx(cpuctx);
2375 if (!ctx->nr_active)
2378 perf_pmu_disable(ctx->pmu);
2379 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2380 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2381 group_sched_out(event, cpuctx, ctx);
2384 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2385 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2386 group_sched_out(event, cpuctx, ctx);
2388 perf_pmu_enable(ctx->pmu);
2392 * Test whether two contexts are equivalent, i.e. whether they have both been
2393 * cloned from the same version of the same context.
2395 * Equivalence is measured using a generation number in the context that is
2396 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2397 * and list_del_event().
2399 static int context_equiv(struct perf_event_context *ctx1,
2400 struct perf_event_context *ctx2)
2402 lockdep_assert_held(&ctx1->lock);
2403 lockdep_assert_held(&ctx2->lock);
2405 /* Pinning disables the swap optimization */
2406 if (ctx1->pin_count || ctx2->pin_count)
2409 /* If ctx1 is the parent of ctx2 */
2410 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2413 /* If ctx2 is the parent of ctx1 */
2414 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2418 * If ctx1 and ctx2 have the same parent; we flatten the parent
2419 * hierarchy, see perf_event_init_context().
2421 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2422 ctx1->parent_gen == ctx2->parent_gen)
2429 static void __perf_event_sync_stat(struct perf_event *event,
2430 struct perf_event *next_event)
2434 if (!event->attr.inherit_stat)
2438 * Update the event value, we cannot use perf_event_read()
2439 * because we're in the middle of a context switch and have IRQs
2440 * disabled, which upsets smp_call_function_single(), however
2441 * we know the event must be on the current CPU, therefore we
2442 * don't need to use it.
2444 switch (event->state) {
2445 case PERF_EVENT_STATE_ACTIVE:
2446 event->pmu->read(event);
2449 case PERF_EVENT_STATE_INACTIVE:
2450 update_event_times(event);
2458 * In order to keep per-task stats reliable we need to flip the event
2459 * values when we flip the contexts.
2461 value = local64_read(&next_event->count);
2462 value = local64_xchg(&event->count, value);
2463 local64_set(&next_event->count, value);
2465 swap(event->total_time_enabled, next_event->total_time_enabled);
2466 swap(event->total_time_running, next_event->total_time_running);
2469 * Since we swizzled the values, update the user visible data too.
2471 perf_event_update_userpage(event);
2472 perf_event_update_userpage(next_event);
2475 static void perf_event_sync_stat(struct perf_event_context *ctx,
2476 struct perf_event_context *next_ctx)
2478 struct perf_event *event, *next_event;
2483 update_context_time(ctx);
2485 event = list_first_entry(&ctx->event_list,
2486 struct perf_event, event_entry);
2488 next_event = list_first_entry(&next_ctx->event_list,
2489 struct perf_event, event_entry);
2491 while (&event->event_entry != &ctx->event_list &&
2492 &next_event->event_entry != &next_ctx->event_list) {
2494 __perf_event_sync_stat(event, next_event);
2496 event = list_next_entry(event, event_entry);
2497 next_event = list_next_entry(next_event, event_entry);
2501 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2502 struct task_struct *next)
2504 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2505 struct perf_event_context *next_ctx;
2506 struct perf_event_context *parent, *next_parent;
2507 struct perf_cpu_context *cpuctx;
2513 cpuctx = __get_cpu_context(ctx);
2514 if (!cpuctx->task_ctx)
2518 next_ctx = next->perf_event_ctxp[ctxn];
2522 parent = rcu_dereference(ctx->parent_ctx);
2523 next_parent = rcu_dereference(next_ctx->parent_ctx);
2525 /* If neither context have a parent context; they cannot be clones. */
2526 if (!parent && !next_parent)
2529 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2531 * Looks like the two contexts are clones, so we might be
2532 * able to optimize the context switch. We lock both
2533 * contexts and check that they are clones under the
2534 * lock (including re-checking that neither has been
2535 * uncloned in the meantime). It doesn't matter which
2536 * order we take the locks because no other cpu could
2537 * be trying to lock both of these tasks.
2539 raw_spin_lock(&ctx->lock);
2540 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2541 if (context_equiv(ctx, next_ctx)) {
2543 * XXX do we need a memory barrier of sorts
2544 * wrt to rcu_dereference() of perf_event_ctxp
2546 task->perf_event_ctxp[ctxn] = next_ctx;
2547 next->perf_event_ctxp[ctxn] = ctx;
2549 next_ctx->task = task;
2551 swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
2555 perf_event_sync_stat(ctx, next_ctx);
2557 raw_spin_unlock(&next_ctx->lock);
2558 raw_spin_unlock(&ctx->lock);
2564 raw_spin_lock(&ctx->lock);
2565 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2566 cpuctx->task_ctx = NULL;
2567 raw_spin_unlock(&ctx->lock);
2571 void perf_sched_cb_dec(struct pmu *pmu)
2573 this_cpu_dec(perf_sched_cb_usages);
2576 void perf_sched_cb_inc(struct pmu *pmu)
2578 this_cpu_inc(perf_sched_cb_usages);
2582 * This function provides the context switch callback to the lower code
2583 * layer. It is invoked ONLY when the context switch callback is enabled.
2585 static void perf_pmu_sched_task(struct task_struct *prev,
2586 struct task_struct *next,
2589 struct perf_cpu_context *cpuctx;
2591 unsigned long flags;
2596 local_irq_save(flags);
2600 list_for_each_entry_rcu(pmu, &pmus, entry) {
2601 if (pmu->sched_task) {
2602 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2604 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2606 perf_pmu_disable(pmu);
2608 pmu->sched_task(cpuctx->task_ctx, sched_in);
2610 perf_pmu_enable(pmu);
2612 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2618 local_irq_restore(flags);
2621 #define for_each_task_context_nr(ctxn) \
2622 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2625 * Called from scheduler to remove the events of the current task,
2626 * with interrupts disabled.
2628 * We stop each event and update the event value in event->count.
2630 * This does not protect us against NMI, but disable()
2631 * sets the disabled bit in the control field of event _before_
2632 * accessing the event control register. If a NMI hits, then it will
2633 * not restart the event.
2635 void __perf_event_task_sched_out(struct task_struct *task,
2636 struct task_struct *next)
2640 if (__this_cpu_read(perf_sched_cb_usages))
2641 perf_pmu_sched_task(task, next, false);
2643 for_each_task_context_nr(ctxn)
2644 perf_event_context_sched_out(task, ctxn, next);
2647 * if cgroup events exist on this CPU, then we need
2648 * to check if we have to switch out PMU state.
2649 * cgroup event are system-wide mode only
2651 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2652 perf_cgroup_sched_out(task, next);
2655 static void task_ctx_sched_out(struct perf_event_context *ctx)
2657 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2659 if (!cpuctx->task_ctx)
2662 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2665 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2666 cpuctx->task_ctx = NULL;
2670 * Called with IRQs disabled
2672 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2673 enum event_type_t event_type)
2675 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2679 ctx_pinned_sched_in(struct perf_event_context *ctx,
2680 struct perf_cpu_context *cpuctx)
2682 struct perf_event *event;
2684 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2685 if (event->state <= PERF_EVENT_STATE_OFF)
2687 if (!event_filter_match(event))
2690 /* may need to reset tstamp_enabled */
2691 if (is_cgroup_event(event))
2692 perf_cgroup_mark_enabled(event, ctx);
2694 if (group_can_go_on(event, cpuctx, 1))
2695 group_sched_in(event, cpuctx, ctx);
2698 * If this pinned group hasn't been scheduled,
2699 * put it in error state.
2701 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2702 update_group_times(event);
2703 event->state = PERF_EVENT_STATE_ERROR;
2709 ctx_flexible_sched_in(struct perf_event_context *ctx,
2710 struct perf_cpu_context *cpuctx)
2712 struct perf_event *event;
2715 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2716 /* Ignore events in OFF or ERROR state */
2717 if (event->state <= PERF_EVENT_STATE_OFF)
2720 * Listen to the 'cpu' scheduling filter constraint
2723 if (!event_filter_match(event))
2726 /* may need to reset tstamp_enabled */
2727 if (is_cgroup_event(event))
2728 perf_cgroup_mark_enabled(event, ctx);
2730 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2731 if (group_sched_in(event, cpuctx, ctx))
2738 ctx_sched_in(struct perf_event_context *ctx,
2739 struct perf_cpu_context *cpuctx,
2740 enum event_type_t event_type,
2741 struct task_struct *task)
2744 int is_active = ctx->is_active;
2746 ctx->is_active |= event_type;
2747 if (likely(!ctx->nr_events))
2751 ctx->timestamp = now;
2752 perf_cgroup_set_timestamp(task, ctx);
2754 * First go through the list and put on any pinned groups
2755 * in order to give them the best chance of going on.
2757 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2758 ctx_pinned_sched_in(ctx, cpuctx);
2760 /* Then walk through the lower prio flexible groups */
2761 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2762 ctx_flexible_sched_in(ctx, cpuctx);
2765 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2766 enum event_type_t event_type,
2767 struct task_struct *task)
2769 struct perf_event_context *ctx = &cpuctx->ctx;
2771 ctx_sched_in(ctx, cpuctx, event_type, task);
2774 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2775 struct task_struct *task)
2777 struct perf_cpu_context *cpuctx;
2779 cpuctx = __get_cpu_context(ctx);
2780 if (cpuctx->task_ctx == ctx)
2783 perf_ctx_lock(cpuctx, ctx);
2784 perf_pmu_disable(ctx->pmu);
2786 * We want to keep the following priority order:
2787 * cpu pinned (that don't need to move), task pinned,
2788 * cpu flexible, task flexible.
2790 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2793 cpuctx->task_ctx = ctx;
2795 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2797 perf_pmu_enable(ctx->pmu);
2798 perf_ctx_unlock(cpuctx, ctx);
2802 * Called from scheduler to add the events of the current task
2803 * with interrupts disabled.
2805 * We restore the event value and then enable it.
2807 * This does not protect us against NMI, but enable()
2808 * sets the enabled bit in the control field of event _before_
2809 * accessing the event control register. If a NMI hits, then it will
2810 * keep the event running.
2812 void __perf_event_task_sched_in(struct task_struct *prev,
2813 struct task_struct *task)
2815 struct perf_event_context *ctx;
2818 for_each_task_context_nr(ctxn) {
2819 ctx = task->perf_event_ctxp[ctxn];
2823 perf_event_context_sched_in(ctx, task);
2826 * if cgroup events exist on this CPU, then we need
2827 * to check if we have to switch in PMU state.
2828 * cgroup event are system-wide mode only
2830 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2831 perf_cgroup_sched_in(prev, task);
2833 if (__this_cpu_read(perf_sched_cb_usages))
2834 perf_pmu_sched_task(prev, task, true);
2837 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2839 u64 frequency = event->attr.sample_freq;
2840 u64 sec = NSEC_PER_SEC;
2841 u64 divisor, dividend;
2843 int count_fls, nsec_fls, frequency_fls, sec_fls;
2845 count_fls = fls64(count);
2846 nsec_fls = fls64(nsec);
2847 frequency_fls = fls64(frequency);
2851 * We got @count in @nsec, with a target of sample_freq HZ
2852 * the target period becomes:
2855 * period = -------------------
2856 * @nsec * sample_freq
2861 * Reduce accuracy by one bit such that @a and @b converge
2862 * to a similar magnitude.
2864 #define REDUCE_FLS(a, b) \
2866 if (a##_fls > b##_fls) { \
2876 * Reduce accuracy until either term fits in a u64, then proceed with
2877 * the other, so that finally we can do a u64/u64 division.
2879 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2880 REDUCE_FLS(nsec, frequency);
2881 REDUCE_FLS(sec, count);
2884 if (count_fls + sec_fls > 64) {
2885 divisor = nsec * frequency;
2887 while (count_fls + sec_fls > 64) {
2888 REDUCE_FLS(count, sec);
2892 dividend = count * sec;
2894 dividend = count * sec;
2896 while (nsec_fls + frequency_fls > 64) {
2897 REDUCE_FLS(nsec, frequency);
2901 divisor = nsec * frequency;
2907 return div64_u64(dividend, divisor);
2910 static DEFINE_PER_CPU(int, perf_throttled_count);
2911 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2913 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2915 struct hw_perf_event *hwc = &event->hw;
2916 s64 period, sample_period;
2919 period = perf_calculate_period(event, nsec, count);
2921 delta = (s64)(period - hwc->sample_period);
2922 delta = (delta + 7) / 8; /* low pass filter */
2924 sample_period = hwc->sample_period + delta;
2929 hwc->sample_period = sample_period;
2931 if (local64_read(&hwc->period_left) > 8*sample_period) {
2933 event->pmu->stop(event, PERF_EF_UPDATE);
2935 local64_set(&hwc->period_left, 0);
2938 event->pmu->start(event, PERF_EF_RELOAD);
2943 * combine freq adjustment with unthrottling to avoid two passes over the
2944 * events. At the same time, make sure, having freq events does not change
2945 * the rate of unthrottling as that would introduce bias.
2947 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2950 struct perf_event *event;
2951 struct hw_perf_event *hwc;
2952 u64 now, period = TICK_NSEC;
2956 * only need to iterate over all events iff:
2957 * - context have events in frequency mode (needs freq adjust)
2958 * - there are events to unthrottle on this cpu
2960 if (!(ctx->nr_freq || needs_unthr))
2963 raw_spin_lock(&ctx->lock);
2964 perf_pmu_disable(ctx->pmu);
2966 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2967 if (event->state != PERF_EVENT_STATE_ACTIVE)
2970 if (!event_filter_match(event))
2973 perf_pmu_disable(event->pmu);
2977 if (hwc->interrupts == MAX_INTERRUPTS) {
2978 hwc->interrupts = 0;
2979 perf_log_throttle(event, 1);
2980 event->pmu->start(event, 0);
2983 if (!event->attr.freq || !event->attr.sample_freq)
2987 * stop the event and update event->count
2989 event->pmu->stop(event, PERF_EF_UPDATE);
2991 now = local64_read(&event->count);
2992 delta = now - hwc->freq_count_stamp;
2993 hwc->freq_count_stamp = now;
2997 * reload only if value has changed
2998 * we have stopped the event so tell that
2999 * to perf_adjust_period() to avoid stopping it
3003 perf_adjust_period(event, period, delta, false);
3005 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3007 perf_pmu_enable(event->pmu);
3010 perf_pmu_enable(ctx->pmu);
3011 raw_spin_unlock(&ctx->lock);
3015 * Round-robin a context's events:
3017 static void rotate_ctx(struct perf_event_context *ctx)
3020 * Rotate the first entry last of non-pinned groups. Rotation might be
3021 * disabled by the inheritance code.
3023 if (!ctx->rotate_disable)
3024 list_rotate_left(&ctx->flexible_groups);
3027 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
3029 struct perf_event_context *ctx = NULL;
3032 if (cpuctx->ctx.nr_events) {
3033 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3037 ctx = cpuctx->task_ctx;
3038 if (ctx && ctx->nr_events) {
3039 if (ctx->nr_events != ctx->nr_active)
3046 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3047 perf_pmu_disable(cpuctx->ctx.pmu);
3049 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3051 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3053 rotate_ctx(&cpuctx->ctx);
3057 perf_event_sched_in(cpuctx, ctx, current);
3059 perf_pmu_enable(cpuctx->ctx.pmu);
3060 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3066 #ifdef CONFIG_NO_HZ_FULL
3067 bool perf_event_can_stop_tick(void)
3069 if (atomic_read(&nr_freq_events) ||
3070 __this_cpu_read(perf_throttled_count))
3077 void perf_event_task_tick(void)
3079 struct list_head *head = this_cpu_ptr(&active_ctx_list);
3080 struct perf_event_context *ctx, *tmp;
3083 WARN_ON(!irqs_disabled());
3085 __this_cpu_inc(perf_throttled_seq);
3086 throttled = __this_cpu_xchg(perf_throttled_count, 0);
3088 list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3089 perf_adjust_freq_unthr_context(ctx, throttled);
3092 static int event_enable_on_exec(struct perf_event *event,
3093 struct perf_event_context *ctx)
3095 if (!event->attr.enable_on_exec)
3098 event->attr.enable_on_exec = 0;
3099 if (event->state >= PERF_EVENT_STATE_INACTIVE)
3102 __perf_event_mark_enabled(event);
3108 * Enable all of a task's events that have been marked enable-on-exec.
3109 * This expects task == current.
3111 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3113 struct perf_event_context *clone_ctx = NULL;
3114 struct perf_event *event;
3115 unsigned long flags;
3119 local_irq_save(flags);
3120 if (!ctx || !ctx->nr_events)
3124 * We must ctxsw out cgroup events to avoid conflict
3125 * when invoking perf_task_event_sched_in() later on
3126 * in this function. Otherwise we end up trying to
3127 * ctxswin cgroup events which are already scheduled
3130 perf_cgroup_sched_out(current, NULL);
3132 raw_spin_lock(&ctx->lock);
3133 task_ctx_sched_out(ctx);
3135 list_for_each_entry(event, &ctx->event_list, event_entry) {
3136 ret = event_enable_on_exec(event, ctx);
3142 * Unclone this context if we enabled any event.
3145 clone_ctx = unclone_ctx(ctx);
3147 raw_spin_unlock(&ctx->lock);
3150 * Also calls ctxswin for cgroup events, if any:
3152 perf_event_context_sched_in(ctx, ctx->task);
3154 local_irq_restore(flags);
3160 void perf_event_exec(void)
3162 struct perf_event_context *ctx;
3166 for_each_task_context_nr(ctxn) {
3167 ctx = current->perf_event_ctxp[ctxn];
3171 perf_event_enable_on_exec(ctx);
3177 * Cross CPU call to read the hardware event
3179 static void __perf_event_read(void *info)
3181 struct perf_event *event = info;
3182 struct perf_event_context *ctx = event->ctx;
3183 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3186 * If this is a task context, we need to check whether it is
3187 * the current task context of this cpu. If not it has been
3188 * scheduled out before the smp call arrived. In that case
3189 * event->count would have been updated to a recent sample
3190 * when the event was scheduled out.
3192 if (ctx->task && cpuctx->task_ctx != ctx)
3195 raw_spin_lock(&ctx->lock);
3196 if (ctx->is_active) {
3197 update_context_time(ctx);
3198 update_cgrp_time_from_event(event);
3200 update_event_times(event);
3201 if (event->state == PERF_EVENT_STATE_ACTIVE)
3202 event->pmu->read(event);
3203 raw_spin_unlock(&ctx->lock);
3206 static inline u64 perf_event_count(struct perf_event *event)
3208 if (event->pmu->count)
3209 return event->pmu->count(event);
3211 return __perf_event_count(event);
3214 static u64 perf_event_read(struct perf_event *event)
3217 * If event is enabled and currently active on a CPU, update the
3218 * value in the event structure:
3220 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3221 smp_call_function_single(event->oncpu,
3222 __perf_event_read, event, 1);
3223 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3224 struct perf_event_context *ctx = event->ctx;
3225 unsigned long flags;
3227 raw_spin_lock_irqsave(&ctx->lock, flags);
3229 * may read while context is not active
3230 * (e.g., thread is blocked), in that case
3231 * we cannot update context time
3233 if (ctx->is_active) {
3234 update_context_time(ctx);
3235 update_cgrp_time_from_event(event);
3237 update_event_times(event);
3238 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3241 return perf_event_count(event);
3245 * Initialize the perf_event context in a task_struct:
3247 static void __perf_event_init_context(struct perf_event_context *ctx)
3249 raw_spin_lock_init(&ctx->lock);
3250 mutex_init(&ctx->mutex);
3251 INIT_LIST_HEAD(&ctx->active_ctx_list);
3252 INIT_LIST_HEAD(&ctx->pinned_groups);
3253 INIT_LIST_HEAD(&ctx->flexible_groups);
3254 INIT_LIST_HEAD(&ctx->event_list);
3255 atomic_set(&ctx->refcount, 1);
3256 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3259 static struct perf_event_context *
3260 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3262 struct perf_event_context *ctx;
3264 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3268 __perf_event_init_context(ctx);
3271 get_task_struct(task);
3278 static struct task_struct *
3279 find_lively_task_by_vpid(pid_t vpid)
3281 struct task_struct *task;
3288 task = find_task_by_vpid(vpid);
3290 get_task_struct(task);
3294 return ERR_PTR(-ESRCH);
3296 /* Reuse ptrace permission checks for now. */
3298 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3303 put_task_struct(task);
3304 return ERR_PTR(err);
3309 * Returns a matching context with refcount and pincount.
3311 static struct perf_event_context *
3312 find_get_context(struct pmu *pmu, struct task_struct *task,
3313 struct perf_event *event)
3315 struct perf_event_context *ctx, *clone_ctx = NULL;
3316 struct perf_cpu_context *cpuctx;
3317 void *task_ctx_data = NULL;
3318 unsigned long flags;
3320 int cpu = event->cpu;
3323 /* Must be root to operate on a CPU event: */
3324 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3325 return ERR_PTR(-EACCES);
3328 * We could be clever and allow to attach a event to an
3329 * offline CPU and activate it when the CPU comes up, but
3332 if (!cpu_online(cpu))
3333 return ERR_PTR(-ENODEV);
3335 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3344 ctxn = pmu->task_ctx_nr;
3348 if (event->attach_state & PERF_ATTACH_TASK_DATA) {
3349 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
3350 if (!task_ctx_data) {
3357 ctx = perf_lock_task_context(task, ctxn, &flags);
3359 clone_ctx = unclone_ctx(ctx);
3362 if (task_ctx_data && !ctx->task_ctx_data) {
3363 ctx->task_ctx_data = task_ctx_data;
3364 task_ctx_data = NULL;
3366 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3371 ctx = alloc_perf_context(pmu, task);
3376 if (task_ctx_data) {
3377 ctx->task_ctx_data = task_ctx_data;
3378 task_ctx_data = NULL;
3382 mutex_lock(&task->perf_event_mutex);
3384 * If it has already passed perf_event_exit_task().
3385 * we must see PF_EXITING, it takes this mutex too.
3387 if (task->flags & PF_EXITING)
3389 else if (task->perf_event_ctxp[ctxn])
3394 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3396 mutex_unlock(&task->perf_event_mutex);
3398 if (unlikely(err)) {
3407 kfree(task_ctx_data);
3411 kfree(task_ctx_data);
3412 return ERR_PTR(err);
3415 static void perf_event_free_filter(struct perf_event *event);
3416 static void perf_event_free_bpf_prog(struct perf_event *event);
3418 static void free_event_rcu(struct rcu_head *head)
3420 struct perf_event *event;
3422 event = container_of(head, struct perf_event, rcu_head);
3424 put_pid_ns(event->ns);
3425 perf_event_free_filter(event);
3426 perf_event_free_bpf_prog(event);
3430 static void ring_buffer_attach(struct perf_event *event,
3431 struct ring_buffer *rb);
3433 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3438 if (is_cgroup_event(event))
3439 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3442 static void unaccount_event(struct perf_event *event)
3447 if (event->attach_state & PERF_ATTACH_TASK)
3448 static_key_slow_dec_deferred(&perf_sched_events);
3449 if (event->attr.mmap || event->attr.mmap_data)
3450 atomic_dec(&nr_mmap_events);
3451 if (event->attr.comm)
3452 atomic_dec(&nr_comm_events);
3453 if (event->attr.task)
3454 atomic_dec(&nr_task_events);
3455 if (event->attr.freq)
3456 atomic_dec(&nr_freq_events);
3457 if (is_cgroup_event(event))
3458 static_key_slow_dec_deferred(&perf_sched_events);
3459 if (has_branch_stack(event))
3460 static_key_slow_dec_deferred(&perf_sched_events);
3462 unaccount_event_cpu(event, event->cpu);
3466 * The following implement mutual exclusion of events on "exclusive" pmus
3467 * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
3468 * at a time, so we disallow creating events that might conflict, namely:
3470 * 1) cpu-wide events in the presence of per-task events,
3471 * 2) per-task events in the presence of cpu-wide events,
3472 * 3) two matching events on the same context.
3474 * The former two cases are handled in the allocation path (perf_event_alloc(),
3475 * __free_event()), the latter -- before the first perf_install_in_context().
3477 static int exclusive_event_init(struct perf_event *event)
3479 struct pmu *pmu = event->pmu;
3481 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3485 * Prevent co-existence of per-task and cpu-wide events on the
3486 * same exclusive pmu.
3488 * Negative pmu::exclusive_cnt means there are cpu-wide
3489 * events on this "exclusive" pmu, positive means there are
3492 * Since this is called in perf_event_alloc() path, event::ctx
3493 * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
3494 * to mean "per-task event", because unlike other attach states it
3495 * never gets cleared.
3497 if (event->attach_state & PERF_ATTACH_TASK) {
3498 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
3501 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
3508 static void exclusive_event_destroy(struct perf_event *event)
3510 struct pmu *pmu = event->pmu;
3512 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3515 /* see comment in exclusive_event_init() */
3516 if (event->attach_state & PERF_ATTACH_TASK)
3517 atomic_dec(&pmu->exclusive_cnt);
3519 atomic_inc(&pmu->exclusive_cnt);
3522 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
3524 if ((e1->pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) &&
3525 (e1->cpu == e2->cpu ||
3532 /* Called under the same ctx::mutex as perf_install_in_context() */
3533 static bool exclusive_event_installable(struct perf_event *event,
3534 struct perf_event_context *ctx)
3536 struct perf_event *iter_event;
3537 struct pmu *pmu = event->pmu;
3539 if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
3542 list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
3543 if (exclusive_event_match(iter_event, event))
3550 static void __free_event(struct perf_event *event)
3552 if (!event->parent) {
3553 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3554 put_callchain_buffers();
3558 event->destroy(event);
3561 put_ctx(event->ctx);
3564 exclusive_event_destroy(event);
3565 module_put(event->pmu->module);
3568 call_rcu(&event->rcu_head, free_event_rcu);
3571 static void _free_event(struct perf_event *event)
3573 irq_work_sync(&event->pending);
3575 unaccount_event(event);
3579 * Can happen when we close an event with re-directed output.
3581 * Since we have a 0 refcount, perf_mmap_close() will skip
3582 * over us; possibly making our ring_buffer_put() the last.
3584 mutex_lock(&event->mmap_mutex);
3585 ring_buffer_attach(event, NULL);
3586 mutex_unlock(&event->mmap_mutex);
3589 if (is_cgroup_event(event))
3590 perf_detach_cgroup(event);
3592 __free_event(event);
3596 * Used to free events which have a known refcount of 1, such as in error paths
3597 * where the event isn't exposed yet and inherited events.
3599 static void free_event(struct perf_event *event)
3601 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3602 "unexpected event refcount: %ld; ptr=%p\n",
3603 atomic_long_read(&event->refcount), event)) {
3604 /* leak to avoid use-after-free */
3612 * Remove user event from the owner task.
3614 static void perf_remove_from_owner(struct perf_event *event)
3616 struct task_struct *owner;
3619 owner = ACCESS_ONCE(event->owner);
3621 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3622 * !owner it means the list deletion is complete and we can indeed
3623 * free this event, otherwise we need to serialize on
3624 * owner->perf_event_mutex.
3626 smp_read_barrier_depends();
3629 * Since delayed_put_task_struct() also drops the last
3630 * task reference we can safely take a new reference
3631 * while holding the rcu_read_lock().
3633 get_task_struct(owner);
3639 * If we're here through perf_event_exit_task() we're already
3640 * holding ctx->mutex which would be an inversion wrt. the
3641 * normal lock order.
3643 * However we can safely take this lock because its the child
3646 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
3649 * We have to re-check the event->owner field, if it is cleared
3650 * we raced with perf_event_exit_task(), acquiring the mutex
3651 * ensured they're done, and we can proceed with freeing the
3655 list_del_init(&event->owner_entry);
3656 mutex_unlock(&owner->perf_event_mutex);
3657 put_task_struct(owner);
3662 * Called when the last reference to the file is gone.
3664 static void put_event(struct perf_event *event)
3666 struct perf_event_context *ctx;
3668 if (!atomic_long_dec_and_test(&event->refcount))
3671 if (!is_kernel_event(event))
3672 perf_remove_from_owner(event);
3675 * There are two ways this annotation is useful:
3677 * 1) there is a lock recursion from perf_event_exit_task
3678 * see the comment there.
3680 * 2) there is a lock-inversion with mmap_sem through
3681 * perf_event_read_group(), which takes faults while
3682 * holding ctx->mutex, however this is called after
3683 * the last filedesc died, so there is no possibility
3684 * to trigger the AB-BA case.
3686 ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING);
3687 WARN_ON_ONCE(ctx->parent_ctx);
3688 perf_remove_from_context(event, true);
3689 perf_event_ctx_unlock(event, ctx);
3694 int perf_event_release_kernel(struct perf_event *event)
3699 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3701 static int perf_release(struct inode *inode, struct file *file)
3703 put_event(file->private_data);
3708 * Remove all orphanes events from the context.
3710 static void orphans_remove_work(struct work_struct *work)
3712 struct perf_event_context *ctx;
3713 struct perf_event *event, *tmp;
3715 ctx = container_of(work, struct perf_event_context,
3716 orphans_remove.work);
3718 mutex_lock(&ctx->mutex);
3719 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3720 struct perf_event *parent_event = event->parent;
3722 if (!is_orphaned_child(event))
3725 perf_remove_from_context(event, true);
3727 mutex_lock(&parent_event->child_mutex);
3728 list_del_init(&event->child_list);
3729 mutex_unlock(&parent_event->child_mutex);
3732 put_event(parent_event);
3735 raw_spin_lock_irq(&ctx->lock);
3736 ctx->orphans_remove_sched = false;
3737 raw_spin_unlock_irq(&ctx->lock);
3738 mutex_unlock(&ctx->mutex);
3743 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3745 struct perf_event *child;
3751 mutex_lock(&event->child_mutex);
3752 total += perf_event_read(event);
3753 *enabled += event->total_time_enabled +
3754 atomic64_read(&event->child_total_time_enabled);
3755 *running += event->total_time_running +
3756 atomic64_read(&event->child_total_time_running);
3758 list_for_each_entry(child, &event->child_list, child_list) {
3759 total += perf_event_read(child);
3760 *enabled += child->total_time_enabled;
3761 *running += child->total_time_running;
3763 mutex_unlock(&event->child_mutex);
3767 EXPORT_SYMBOL_GPL(perf_event_read_value);
3769 static int perf_event_read_group(struct perf_event *event,
3770 u64 read_format, char __user *buf)
3772 struct perf_event *leader = event->group_leader, *sub;
3773 struct perf_event_context *ctx = leader->ctx;
3774 int n = 0, size = 0, ret;
3775 u64 count, enabled, running;
3778 lockdep_assert_held(&ctx->mutex);
3780 count = perf_event_read_value(leader, &enabled, &running);
3782 values[n++] = 1 + leader->nr_siblings;
3783 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3784 values[n++] = enabled;
3785 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3786 values[n++] = running;
3787 values[n++] = count;
3788 if (read_format & PERF_FORMAT_ID)
3789 values[n++] = primary_event_id(leader);
3791 size = n * sizeof(u64);
3793 if (copy_to_user(buf, values, size))
3798 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3801 values[n++] = perf_event_read_value(sub, &enabled, &running);
3802 if (read_format & PERF_FORMAT_ID)
3803 values[n++] = primary_event_id(sub);
3805 size = n * sizeof(u64);
3807 if (copy_to_user(buf + ret, values, size)) {
3817 static int perf_event_read_one(struct perf_event *event,
3818 u64 read_format, char __user *buf)
3820 u64 enabled, running;
3824 values[n++] = perf_event_read_value(event, &enabled, &running);
3825 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3826 values[n++] = enabled;
3827 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3828 values[n++] = running;
3829 if (read_format & PERF_FORMAT_ID)
3830 values[n++] = primary_event_id(event);
3832 if (copy_to_user(buf, values, n * sizeof(u64)))
3835 return n * sizeof(u64);
3838 static bool is_event_hup(struct perf_event *event)
3842 if (event->state != PERF_EVENT_STATE_EXIT)
3845 mutex_lock(&event->child_mutex);
3846 no_children = list_empty(&event->child_list);
3847 mutex_unlock(&event->child_mutex);
3852 * Read the performance event - simple non blocking version for now
3855 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3857 u64 read_format = event->attr.read_format;
3861 * Return end-of-file for a read on a event that is in
3862 * error state (i.e. because it was pinned but it couldn't be
3863 * scheduled on to the CPU at some point).
3865 if (event->state == PERF_EVENT_STATE_ERROR)
3868 if (count < event->read_size)
3871 WARN_ON_ONCE(event->ctx->parent_ctx);
3872 if (read_format & PERF_FORMAT_GROUP)
3873 ret = perf_event_read_group(event, read_format, buf);
3875 ret = perf_event_read_one(event, read_format, buf);
3881 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3883 struct perf_event *event = file->private_data;
3884 struct perf_event_context *ctx;
3887 ctx = perf_event_ctx_lock(event);
3888 ret = perf_read_hw(event, buf, count);
3889 perf_event_ctx_unlock(event, ctx);
3894 static unsigned int perf_poll(struct file *file, poll_table *wait)
3896 struct perf_event *event = file->private_data;
3897 struct ring_buffer *rb;
3898 unsigned int events = POLLHUP;
3900 poll_wait(file, &event->waitq, wait);
3902 if (is_event_hup(event))
3906 * Pin the event->rb by taking event->mmap_mutex; otherwise
3907 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3909 mutex_lock(&event->mmap_mutex);
3912 events = atomic_xchg(&rb->poll, 0);
3913 mutex_unlock(&event->mmap_mutex);
3917 static void _perf_event_reset(struct perf_event *event)
3919 (void)perf_event_read(event);
3920 local64_set(&event->count, 0);
3921 perf_event_update_userpage(event);
3925 * Holding the top-level event's child_mutex means that any
3926 * descendant process that has inherited this event will block
3927 * in sync_child_event if it goes to exit, thus satisfying the
3928 * task existence requirements of perf_event_enable/disable.
3930 static void perf_event_for_each_child(struct perf_event *event,
3931 void (*func)(struct perf_event *))
3933 struct perf_event *child;
3935 WARN_ON_ONCE(event->ctx->parent_ctx);
3937 mutex_lock(&event->child_mutex);
3939 list_for_each_entry(child, &event->child_list, child_list)
3941 mutex_unlock(&event->child_mutex);
3944 static void perf_event_for_each(struct perf_event *event,
3945 void (*func)(struct perf_event *))
3947 struct perf_event_context *ctx = event->ctx;
3948 struct perf_event *sibling;
3950 lockdep_assert_held(&ctx->mutex);
3952 event = event->group_leader;
3954 perf_event_for_each_child(event, func);
3955 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3956 perf_event_for_each_child(sibling, func);
3959 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3961 struct perf_event_context *ctx = event->ctx;
3962 int ret = 0, active;
3965 if (!is_sampling_event(event))
3968 if (copy_from_user(&value, arg, sizeof(value)))
3974 raw_spin_lock_irq(&ctx->lock);
3975 if (event->attr.freq) {
3976 if (value > sysctl_perf_event_sample_rate) {
3981 event->attr.sample_freq = value;
3983 event->attr.sample_period = value;
3984 event->hw.sample_period = value;
3987 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3989 perf_pmu_disable(ctx->pmu);
3990 event->pmu->stop(event, PERF_EF_UPDATE);
3993 local64_set(&event->hw.period_left, 0);
3996 event->pmu->start(event, PERF_EF_RELOAD);
3997 perf_pmu_enable(ctx->pmu);
4001 raw_spin_unlock_irq(&ctx->lock);
4006 static const struct file_operations perf_fops;
4008 static inline int perf_fget_light(int fd, struct fd *p)
4010 struct fd f = fdget(fd);
4014 if (f.file->f_op != &perf_fops) {
4022 static int perf_event_set_output(struct perf_event *event,
4023 struct perf_event *output_event);
4024 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
4025 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
4027 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
4029 void (*func)(struct perf_event *);
4033 case PERF_EVENT_IOC_ENABLE:
4034 func = _perf_event_enable;
4036 case PERF_EVENT_IOC_DISABLE:
4037 func = _perf_event_disable;
4039 case PERF_EVENT_IOC_RESET:
4040 func = _perf_event_reset;
4043 case PERF_EVENT_IOC_REFRESH:
4044 return _perf_event_refresh(event, arg);
4046 case PERF_EVENT_IOC_PERIOD:
4047 return perf_event_period(event, (u64 __user *)arg);
4049 case PERF_EVENT_IOC_ID:
4051 u64 id = primary_event_id(event);
4053 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
4058 case PERF_EVENT_IOC_SET_OUTPUT:
4062 struct perf_event *output_event;
4064 ret = perf_fget_light(arg, &output);
4067 output_event = output.file->private_data;
4068 ret = perf_event_set_output(event, output_event);
4071 ret = perf_event_set_output(event, NULL);
4076 case PERF_EVENT_IOC_SET_FILTER:
4077 return perf_event_set_filter(event, (void __user *)arg);
4079 case PERF_EVENT_IOC_SET_BPF:
4080 return perf_event_set_bpf_prog(event, arg);
4086 if (flags & PERF_IOC_FLAG_GROUP)
4087 perf_event_for_each(event, func);
4089 perf_event_for_each_child(event, func);
4094 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
4096 struct perf_event *event = file->private_data;
4097 struct perf_event_context *ctx;
4100 ctx = perf_event_ctx_lock(event);
4101 ret = _perf_ioctl(event, cmd, arg);
4102 perf_event_ctx_unlock(event, ctx);
4107 #ifdef CONFIG_COMPAT
4108 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4111 switch (_IOC_NR(cmd)) {
4112 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4113 case _IOC_NR(PERF_EVENT_IOC_ID):
4114 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4115 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4116 cmd &= ~IOCSIZE_MASK;
4117 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4121 return perf_ioctl(file, cmd, arg);
4124 # define perf_compat_ioctl NULL
4127 int perf_event_task_enable(void)
4129 struct perf_event_context *ctx;
4130 struct perf_event *event;
4132 mutex_lock(¤t->perf_event_mutex);
4133 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4134 ctx = perf_event_ctx_lock(event);
4135 perf_event_for_each_child(event, _perf_event_enable);
4136 perf_event_ctx_unlock(event, ctx);
4138 mutex_unlock(¤t->perf_event_mutex);
4143 int perf_event_task_disable(void)
4145 struct perf_event_context *ctx;
4146 struct perf_event *event;
4148 mutex_lock(¤t->perf_event_mutex);
4149 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4150 ctx = perf_event_ctx_lock(event);
4151 perf_event_for_each_child(event, _perf_event_disable);
4152 perf_event_ctx_unlock(event, ctx);
4154 mutex_unlock(¤t->perf_event_mutex);
4159 static int perf_event_index(struct perf_event *event)
4161 if (event->hw.state & PERF_HES_STOPPED)
4164 if (event->state != PERF_EVENT_STATE_ACTIVE)
4167 return event->pmu->event_idx(event);
4170 static void calc_timer_values(struct perf_event *event,
4177 *now = perf_clock();
4178 ctx_time = event->shadow_ctx_time + *now;
4179 *enabled = ctx_time - event->tstamp_enabled;
4180 *running = ctx_time - event->tstamp_running;
4183 static void perf_event_init_userpage(struct perf_event *event)
4185 struct perf_event_mmap_page *userpg;
4186 struct ring_buffer *rb;
4189 rb = rcu_dereference(event->rb);
4193 userpg = rb->user_page;
4195 /* Allow new userspace to detect that bit 0 is deprecated */
4196 userpg->cap_bit0_is_deprecated = 1;
4197 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4198 userpg->data_offset = PAGE_SIZE;
4199 userpg->data_size = perf_data_size(rb);
4205 void __weak arch_perf_update_userpage(
4206 struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
4211 * Callers need to ensure there can be no nesting of this function, otherwise
4212 * the seqlock logic goes bad. We can not serialize this because the arch
4213 * code calls this from NMI context.
4215 void perf_event_update_userpage(struct perf_event *event)
4217 struct perf_event_mmap_page *userpg;
4218 struct ring_buffer *rb;
4219 u64 enabled, running, now;
4222 rb = rcu_dereference(event->rb);
4227 * compute total_time_enabled, total_time_running
4228 * based on snapshot values taken when the event
4229 * was last scheduled in.
4231 * we cannot simply called update_context_time()
4232 * because of locking issue as we can be called in
4235 calc_timer_values(event, &now, &enabled, &running);
4237 userpg = rb->user_page;
4239 * Disable preemption so as to not let the corresponding user-space
4240 * spin too long if we get preempted.
4245 userpg->index = perf_event_index(event);
4246 userpg->offset = perf_event_count(event);
4248 userpg->offset -= local64_read(&event->hw.prev_count);
4250 userpg->time_enabled = enabled +
4251 atomic64_read(&event->child_total_time_enabled);
4253 userpg->time_running = running +
4254 atomic64_read(&event->child_total_time_running);
4256 arch_perf_update_userpage(event, userpg, now);
4265 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4267 struct perf_event *event = vma->vm_file->private_data;
4268 struct ring_buffer *rb;
4269 int ret = VM_FAULT_SIGBUS;
4271 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4272 if (vmf->pgoff == 0)
4278 rb = rcu_dereference(event->rb);
4282 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4285 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4289 get_page(vmf->page);
4290 vmf->page->mapping = vma->vm_file->f_mapping;
4291 vmf->page->index = vmf->pgoff;
4300 static void ring_buffer_attach(struct perf_event *event,
4301 struct ring_buffer *rb)
4303 struct ring_buffer *old_rb = NULL;
4304 unsigned long flags;
4308 * Should be impossible, we set this when removing
4309 * event->rb_entry and wait/clear when adding event->rb_entry.
4311 WARN_ON_ONCE(event->rcu_pending);
4314 event->rcu_batches = get_state_synchronize_rcu();
4315 event->rcu_pending = 1;
4317 spin_lock_irqsave(&old_rb->event_lock, flags);
4318 list_del_rcu(&event->rb_entry);
4319 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4322 if (event->rcu_pending && rb) {
4323 cond_synchronize_rcu(event->rcu_batches);
4324 event->rcu_pending = 0;
4328 spin_lock_irqsave(&rb->event_lock, flags);
4329 list_add_rcu(&event->rb_entry, &rb->event_list);
4330 spin_unlock_irqrestore(&rb->event_lock, flags);
4333 rcu_assign_pointer(event->rb, rb);
4336 ring_buffer_put(old_rb);
4338 * Since we detached before setting the new rb, so that we
4339 * could attach the new rb, we could have missed a wakeup.
4342 wake_up_all(&event->waitq);
4346 static void ring_buffer_wakeup(struct perf_event *event)
4348 struct ring_buffer *rb;
4351 rb = rcu_dereference(event->rb);
4353 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4354 wake_up_all(&event->waitq);
4359 static void rb_free_rcu(struct rcu_head *rcu_head)
4361 struct ring_buffer *rb;
4363 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4367 struct ring_buffer *ring_buffer_get(struct perf_event *event)
4369 struct ring_buffer *rb;
4372 rb = rcu_dereference(event->rb);
4374 if (!atomic_inc_not_zero(&rb->refcount))
4382 void ring_buffer_put(struct ring_buffer *rb)
4384 if (!atomic_dec_and_test(&rb->refcount))
4387 WARN_ON_ONCE(!list_empty(&rb->event_list));
4389 call_rcu(&rb->rcu_head, rb_free_rcu);
4392 static void perf_mmap_open(struct vm_area_struct *vma)
4394 struct perf_event *event = vma->vm_file->private_data;
4396 atomic_inc(&event->mmap_count);
4397 atomic_inc(&event->rb->mmap_count);
4400 atomic_inc(&event->rb->aux_mmap_count);
4402 if (event->pmu->event_mapped)
4403 event->pmu->event_mapped(event);
4407 * A buffer can be mmap()ed multiple times; either directly through the same
4408 * event, or through other events by use of perf_event_set_output().
4410 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4411 * the buffer here, where we still have a VM context. This means we need
4412 * to detach all events redirecting to us.
4414 static void perf_mmap_close(struct vm_area_struct *vma)
4416 struct perf_event *event = vma->vm_file->private_data;
4418 struct ring_buffer *rb = ring_buffer_get(event);
4419 struct user_struct *mmap_user = rb->mmap_user;
4420 int mmap_locked = rb->mmap_locked;
4421 unsigned long size = perf_data_size(rb);
4423 if (event->pmu->event_unmapped)
4424 event->pmu->event_unmapped(event);
4427 * rb->aux_mmap_count will always drop before rb->mmap_count and
4428 * event->mmap_count, so it is ok to use event->mmap_mutex to
4429 * serialize with perf_mmap here.
4431 if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
4432 atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
4433 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
4434 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
4437 mutex_unlock(&event->mmap_mutex);
4440 atomic_dec(&rb->mmap_count);
4442 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4445 ring_buffer_attach(event, NULL);
4446 mutex_unlock(&event->mmap_mutex);
4448 /* If there's still other mmap()s of this buffer, we're done. */
4449 if (atomic_read(&rb->mmap_count))
4453 * No other mmap()s, detach from all other events that might redirect
4454 * into the now unreachable buffer. Somewhat complicated by the
4455 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4459 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4460 if (!atomic_long_inc_not_zero(&event->refcount)) {
4462 * This event is en-route to free_event() which will
4463 * detach it and remove it from the list.
4469 mutex_lock(&event->mmap_mutex);
4471 * Check we didn't race with perf_event_set_output() which can
4472 * swizzle the rb from under us while we were waiting to
4473 * acquire mmap_mutex.
4475 * If we find a different rb; ignore this event, a next
4476 * iteration will no longer find it on the list. We have to
4477 * still restart the iteration to make sure we're not now
4478 * iterating the wrong list.
4480 if (event->rb == rb)
4481 ring_buffer_attach(event, NULL);
4483 mutex_unlock(&event->mmap_mutex);
4487 * Restart the iteration; either we're on the wrong list or
4488 * destroyed its integrity by doing a deletion.
4495 * It could be there's still a few 0-ref events on the list; they'll
4496 * get cleaned up by free_event() -- they'll also still have their
4497 * ref on the rb and will free it whenever they are done with it.
4499 * Aside from that, this buffer is 'fully' detached and unmapped,
4500 * undo the VM accounting.
4503 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4504 vma->vm_mm->pinned_vm -= mmap_locked;
4505 free_uid(mmap_user);
4508 ring_buffer_put(rb); /* could be last */
4511 static const struct vm_operations_struct perf_mmap_vmops = {
4512 .open = perf_mmap_open,
4513 .close = perf_mmap_close, /* non mergable */
4514 .fault = perf_mmap_fault,
4515 .page_mkwrite = perf_mmap_fault,
4518 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4520 struct perf_event *event = file->private_data;
4521 unsigned long user_locked, user_lock_limit;
4522 struct user_struct *user = current_user();
4523 unsigned long locked, lock_limit;
4524 struct ring_buffer *rb = NULL;
4525 unsigned long vma_size;
4526 unsigned long nr_pages;
4527 long user_extra = 0, extra = 0;
4528 int ret = 0, flags = 0;
4531 * Don't allow mmap() of inherited per-task counters. This would
4532 * create a performance issue due to all children writing to the
4535 if (event->cpu == -1 && event->attr.inherit)
4538 if (!(vma->vm_flags & VM_SHARED))
4541 vma_size = vma->vm_end - vma->vm_start;
4543 if (vma->vm_pgoff == 0) {
4544 nr_pages = (vma_size / PAGE_SIZE) - 1;
4547 * AUX area mapping: if rb->aux_nr_pages != 0, it's already
4548 * mapped, all subsequent mappings should have the same size
4549 * and offset. Must be above the normal perf buffer.
4551 u64 aux_offset, aux_size;
4556 nr_pages = vma_size / PAGE_SIZE;
4558 mutex_lock(&event->mmap_mutex);
4565 aux_offset = ACCESS_ONCE(rb->user_page->aux_offset);
4566 aux_size = ACCESS_ONCE(rb->user_page->aux_size);
4568 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
4571 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
4574 /* already mapped with a different offset */
4575 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
4578 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
4581 /* already mapped with a different size */
4582 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
4585 if (!is_power_of_2(nr_pages))
4588 if (!atomic_inc_not_zero(&rb->mmap_count))
4591 if (rb_has_aux(rb)) {
4592 atomic_inc(&rb->aux_mmap_count);
4597 atomic_set(&rb->aux_mmap_count, 1);
4598 user_extra = nr_pages;
4604 * If we have rb pages ensure they're a power-of-two number, so we
4605 * can do bitmasks instead of modulo.
4607 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4610 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4613 WARN_ON_ONCE(event->ctx->parent_ctx);
4615 mutex_lock(&event->mmap_mutex);
4617 if (event->rb->nr_pages != nr_pages) {
4622 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4624 * Raced against perf_mmap_close() through
4625 * perf_event_set_output(). Try again, hope for better
4628 mutex_unlock(&event->mmap_mutex);
4635 user_extra = nr_pages + 1;
4638 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4641 * Increase the limit linearly with more CPUs:
4643 user_lock_limit *= num_online_cpus();
4645 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4647 if (user_locked > user_lock_limit)
4648 extra = user_locked - user_lock_limit;
4650 lock_limit = rlimit(RLIMIT_MEMLOCK);
4651 lock_limit >>= PAGE_SHIFT;
4652 locked = vma->vm_mm->pinned_vm + extra;
4654 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4655 !capable(CAP_IPC_LOCK)) {
4660 WARN_ON(!rb && event->rb);
4662 if (vma->vm_flags & VM_WRITE)
4663 flags |= RING_BUFFER_WRITABLE;
4666 rb = rb_alloc(nr_pages,
4667 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4675 atomic_set(&rb->mmap_count, 1);
4676 rb->mmap_user = get_current_user();
4677 rb->mmap_locked = extra;
4679 ring_buffer_attach(event, rb);
4681 perf_event_init_userpage(event);
4682 perf_event_update_userpage(event);
4684 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
4685 event->attr.aux_watermark, flags);
4687 rb->aux_mmap_locked = extra;
4692 atomic_long_add(user_extra, &user->locked_vm);
4693 vma->vm_mm->pinned_vm += extra;
4695 atomic_inc(&event->mmap_count);
4697 atomic_dec(&rb->mmap_count);
4700 mutex_unlock(&event->mmap_mutex);
4703 * Since pinned accounting is per vm we cannot allow fork() to copy our
4706 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4707 vma->vm_ops = &perf_mmap_vmops;
4709 if (event->pmu->event_mapped)
4710 event->pmu->event_mapped(event);
4715 static int perf_fasync(int fd, struct file *filp, int on)
4717 struct inode *inode = file_inode(filp);
4718 struct perf_event *event = filp->private_data;
4721 mutex_lock(&inode->i_mutex);
4722 retval = fasync_helper(fd, filp, on, &event->fasync);
4723 mutex_unlock(&inode->i_mutex);
4731 static const struct file_operations perf_fops = {
4732 .llseek = no_llseek,
4733 .release = perf_release,
4736 .unlocked_ioctl = perf_ioctl,
4737 .compat_ioctl = perf_compat_ioctl,
4739 .fasync = perf_fasync,
4745 * If there's data, ensure we set the poll() state and publish everything
4746 * to user-space before waking everybody up.
4749 void perf_event_wakeup(struct perf_event *event)
4751 ring_buffer_wakeup(event);
4753 if (event->pending_kill) {
4754 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4755 event->pending_kill = 0;
4759 static void perf_pending_event(struct irq_work *entry)
4761 struct perf_event *event = container_of(entry,
4762 struct perf_event, pending);
4765 rctx = perf_swevent_get_recursion_context();
4767 * If we 'fail' here, that's OK, it means recursion is already disabled
4768 * and we won't recurse 'further'.
4771 if (event->pending_disable) {
4772 event->pending_disable = 0;
4773 __perf_event_disable(event);
4776 if (event->pending_wakeup) {
4777 event->pending_wakeup = 0;
4778 perf_event_wakeup(event);
4782 perf_swevent_put_recursion_context(rctx);
4786 * We assume there is only KVM supporting the callbacks.
4787 * Later on, we might change it to a list if there is
4788 * another virtualization implementation supporting the callbacks.
4790 struct perf_guest_info_callbacks *perf_guest_cbs;
4792 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4794 perf_guest_cbs = cbs;
4797 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4799 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4801 perf_guest_cbs = NULL;
4804 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4807 perf_output_sample_regs(struct perf_output_handle *handle,
4808 struct pt_regs *regs, u64 mask)
4812 for_each_set_bit(bit, (const unsigned long *) &mask,
4813 sizeof(mask) * BITS_PER_BYTE) {
4816 val = perf_reg_value(regs, bit);
4817 perf_output_put(handle, val);
4821 static void perf_sample_regs_user(struct perf_regs *regs_user,
4822 struct pt_regs *regs,
4823 struct pt_regs *regs_user_copy)
4825 if (user_mode(regs)) {
4826 regs_user->abi = perf_reg_abi(current);
4827 regs_user->regs = regs;
4828 } else if (current->mm) {
4829 perf_get_regs_user(regs_user, regs, regs_user_copy);
4831 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4832 regs_user->regs = NULL;
4836 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4837 struct pt_regs *regs)
4839 regs_intr->regs = regs;
4840 regs_intr->abi = perf_reg_abi(current);
4845 * Get remaining task size from user stack pointer.
4847 * It'd be better to take stack vma map and limit this more
4848 * precisly, but there's no way to get it safely under interrupt,
4849 * so using TASK_SIZE as limit.
4851 static u64 perf_ustack_task_size(struct pt_regs *regs)
4853 unsigned long addr = perf_user_stack_pointer(regs);
4855 if (!addr || addr >= TASK_SIZE)
4858 return TASK_SIZE - addr;
4862 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4863 struct pt_regs *regs)
4867 /* No regs, no stack pointer, no dump. */
4872 * Check if we fit in with the requested stack size into the:
4874 * If we don't, we limit the size to the TASK_SIZE.
4876 * - remaining sample size
4877 * If we don't, we customize the stack size to
4878 * fit in to the remaining sample size.
4881 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4882 stack_size = min(stack_size, (u16) task_size);
4884 /* Current header size plus static size and dynamic size. */
4885 header_size += 2 * sizeof(u64);
4887 /* Do we fit in with the current stack dump size? */
4888 if ((u16) (header_size + stack_size) < header_size) {
4890 * If we overflow the maximum size for the sample,
4891 * we customize the stack dump size to fit in.
4893 stack_size = USHRT_MAX - header_size - sizeof(u64);
4894 stack_size = round_up(stack_size, sizeof(u64));
4901 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4902 struct pt_regs *regs)
4904 /* Case of a kernel thread, nothing to dump */
4907 perf_output_put(handle, size);
4916 * - the size requested by user or the best one we can fit
4917 * in to the sample max size
4919 * - user stack dump data
4921 * - the actual dumped size
4925 perf_output_put(handle, dump_size);
4928 sp = perf_user_stack_pointer(regs);
4929 rem = __output_copy_user(handle, (void *) sp, dump_size);
4930 dyn_size = dump_size - rem;
4932 perf_output_skip(handle, rem);
4935 perf_output_put(handle, dyn_size);
4939 static void __perf_event_header__init_id(struct perf_event_header *header,
4940 struct perf_sample_data *data,
4941 struct perf_event *event)
4943 u64 sample_type = event->attr.sample_type;
4945 data->type = sample_type;
4946 header->size += event->id_header_size;
4948 if (sample_type & PERF_SAMPLE_TID) {
4949 /* namespace issues */
4950 data->tid_entry.pid = perf_event_pid(event, current);
4951 data->tid_entry.tid = perf_event_tid(event, current);
4954 if (sample_type & PERF_SAMPLE_TIME)
4955 data->time = perf_event_clock(event);
4957 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4958 data->id = primary_event_id(event);
4960 if (sample_type & PERF_SAMPLE_STREAM_ID)
4961 data->stream_id = event->id;
4963 if (sample_type & PERF_SAMPLE_CPU) {
4964 data->cpu_entry.cpu = raw_smp_processor_id();
4965 data->cpu_entry.reserved = 0;
4969 void perf_event_header__init_id(struct perf_event_header *header,
4970 struct perf_sample_data *data,
4971 struct perf_event *event)
4973 if (event->attr.sample_id_all)
4974 __perf_event_header__init_id(header, data, event);
4977 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4978 struct perf_sample_data *data)
4980 u64 sample_type = data->type;
4982 if (sample_type & PERF_SAMPLE_TID)
4983 perf_output_put(handle, data->tid_entry);
4985 if (sample_type & PERF_SAMPLE_TIME)
4986 perf_output_put(handle, data->time);
4988 if (sample_type & PERF_SAMPLE_ID)
4989 perf_output_put(handle, data->id);
4991 if (sample_type & PERF_SAMPLE_STREAM_ID)
4992 perf_output_put(handle, data->stream_id);
4994 if (sample_type & PERF_SAMPLE_CPU)
4995 perf_output_put(handle, data->cpu_entry);
4997 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4998 perf_output_put(handle, data->id);
5001 void perf_event__output_id_sample(struct perf_event *event,
5002 struct perf_output_handle *handle,
5003 struct perf_sample_data *sample)
5005 if (event->attr.sample_id_all)
5006 __perf_event__output_id_sample(handle, sample);
5009 static void perf_output_read_one(struct perf_output_handle *handle,
5010 struct perf_event *event,
5011 u64 enabled, u64 running)
5013 u64 read_format = event->attr.read_format;
5017 values[n++] = perf_event_count(event);
5018 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
5019 values[n++] = enabled +
5020 atomic64_read(&event->child_total_time_enabled);
5022 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
5023 values[n++] = running +
5024 atomic64_read(&event->child_total_time_running);
5026 if (read_format & PERF_FORMAT_ID)
5027 values[n++] = primary_event_id(event);
5029 __output_copy(handle, values, n * sizeof(u64));
5033 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
5035 static void perf_output_read_group(struct perf_output_handle *handle,
5036 struct perf_event *event,
5037 u64 enabled, u64 running)
5039 struct perf_event *leader = event->group_leader, *sub;
5040 u64 read_format = event->attr.read_format;
5044 values[n++] = 1 + leader->nr_siblings;
5046 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
5047 values[n++] = enabled;
5049 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
5050 values[n++] = running;
5052 if (leader != event)
5053 leader->pmu->read(leader);
5055 values[n++] = perf_event_count(leader);
5056 if (read_format & PERF_FORMAT_ID)
5057 values[n++] = primary_event_id(leader);
5059 __output_copy(handle, values, n * sizeof(u64));
5061 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
5064 if ((sub != event) &&
5065 (sub->state == PERF_EVENT_STATE_ACTIVE))
5066 sub->pmu->read(sub);
5068 values[n++] = perf_event_count(sub);
5069 if (read_format & PERF_FORMAT_ID)
5070 values[n++] = primary_event_id(sub);
5072 __output_copy(handle, values, n * sizeof(u64));
5076 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
5077 PERF_FORMAT_TOTAL_TIME_RUNNING)
5079 static void perf_output_read(struct perf_output_handle *handle,
5080 struct perf_event *event)
5082 u64 enabled = 0, running = 0, now;
5083 u64 read_format = event->attr.read_format;
5086 * compute total_time_enabled, total_time_running
5087 * based on snapshot values taken when the event
5088 * was last scheduled in.
5090 * we cannot simply called update_context_time()
5091 * because of locking issue as we are called in
5094 if (read_format & PERF_FORMAT_TOTAL_TIMES)
5095 calc_timer_values(event, &now, &enabled, &running);
5097 if (event->attr.read_format & PERF_FORMAT_GROUP)
5098 perf_output_read_group(handle, event, enabled, running);
5100 perf_output_read_one(handle, event, enabled, running);
5103 void perf_output_sample(struct perf_output_handle *handle,
5104 struct perf_event_header *header,
5105 struct perf_sample_data *data,
5106 struct perf_event *event)
5108 u64 sample_type = data->type;
5110 perf_output_put(handle, *header);
5112 if (sample_type & PERF_SAMPLE_IDENTIFIER)
5113 perf_output_put(handle, data->id);
5115 if (sample_type & PERF_SAMPLE_IP)
5116 perf_output_put(handle, data->ip);
5118 if (sample_type & PERF_SAMPLE_TID)
5119 perf_output_put(handle, data->tid_entry);
5121 if (sample_type & PERF_SAMPLE_TIME)
5122 perf_output_put(handle, data->time);
5124 if (sample_type & PERF_SAMPLE_ADDR)
5125 perf_output_put(handle, data->addr);
5127 if (sample_type & PERF_SAMPLE_ID)
5128 perf_output_put(handle, data->id);
5130 if (sample_type & PERF_SAMPLE_STREAM_ID)
5131 perf_output_put(handle, data->stream_id);
5133 if (sample_type & PERF_SAMPLE_CPU)
5134 perf_output_put(handle, data->cpu_entry);
5136 if (sample_type & PERF_SAMPLE_PERIOD)
5137 perf_output_put(handle, data->period);
5139 if (sample_type & PERF_SAMPLE_READ)
5140 perf_output_read(handle, event);
5142 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5143 if (data->callchain) {
5146 if (data->callchain)
5147 size += data->callchain->nr;
5149 size *= sizeof(u64);
5151 __output_copy(handle, data->callchain, size);
5154 perf_output_put(handle, nr);
5158 if (sample_type & PERF_SAMPLE_RAW) {
5160 perf_output_put(handle, data->raw->size);
5161 __output_copy(handle, data->raw->data,
5168 .size = sizeof(u32),
5171 perf_output_put(handle, raw);
5175 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5176 if (data->br_stack) {
5179 size = data->br_stack->nr
5180 * sizeof(struct perf_branch_entry);
5182 perf_output_put(handle, data->br_stack->nr);
5183 perf_output_copy(handle, data->br_stack->entries, size);
5186 * we always store at least the value of nr
5189 perf_output_put(handle, nr);
5193 if (sample_type & PERF_SAMPLE_REGS_USER) {
5194 u64 abi = data->regs_user.abi;
5197 * If there are no regs to dump, notice it through
5198 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5200 perf_output_put(handle, abi);
5203 u64 mask = event->attr.sample_regs_user;
5204 perf_output_sample_regs(handle,
5205 data->regs_user.regs,
5210 if (sample_type & PERF_SAMPLE_STACK_USER) {
5211 perf_output_sample_ustack(handle,
5212 data->stack_user_size,
5213 data->regs_user.regs);
5216 if (sample_type & PERF_SAMPLE_WEIGHT)
5217 perf_output_put(handle, data->weight);
5219 if (sample_type & PERF_SAMPLE_DATA_SRC)
5220 perf_output_put(handle, data->data_src.val);
5222 if (sample_type & PERF_SAMPLE_TRANSACTION)
5223 perf_output_put(handle, data->txn);
5225 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5226 u64 abi = data->regs_intr.abi;
5228 * If there are no regs to dump, notice it through
5229 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5231 perf_output_put(handle, abi);
5234 u64 mask = event->attr.sample_regs_intr;
5236 perf_output_sample_regs(handle,
5237 data->regs_intr.regs,
5242 if (!event->attr.watermark) {
5243 int wakeup_events = event->attr.wakeup_events;
5245 if (wakeup_events) {
5246 struct ring_buffer *rb = handle->rb;
5247 int events = local_inc_return(&rb->events);
5249 if (events >= wakeup_events) {
5250 local_sub(wakeup_events, &rb->events);
5251 local_inc(&rb->wakeup);
5257 void perf_prepare_sample(struct perf_event_header *header,
5258 struct perf_sample_data *data,
5259 struct perf_event *event,
5260 struct pt_regs *regs)
5262 u64 sample_type = event->attr.sample_type;
5264 header->type = PERF_RECORD_SAMPLE;
5265 header->size = sizeof(*header) + event->header_size;
5268 header->misc |= perf_misc_flags(regs);
5270 __perf_event_header__init_id(header, data, event);
5272 if (sample_type & PERF_SAMPLE_IP)
5273 data->ip = perf_instruction_pointer(regs);
5275 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5278 data->callchain = perf_callchain(event, regs);
5280 if (data->callchain)
5281 size += data->callchain->nr;
5283 header->size += size * sizeof(u64);
5286 if (sample_type & PERF_SAMPLE_RAW) {
5287 int size = sizeof(u32);
5290 size += data->raw->size;
5292 size += sizeof(u32);
5294 WARN_ON_ONCE(size & (sizeof(u64)-1));
5295 header->size += size;
5298 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5299 int size = sizeof(u64); /* nr */
5300 if (data->br_stack) {
5301 size += data->br_stack->nr
5302 * sizeof(struct perf_branch_entry);
5304 header->size += size;
5307 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
5308 perf_sample_regs_user(&data->regs_user, regs,
5309 &data->regs_user_copy);
5311 if (sample_type & PERF_SAMPLE_REGS_USER) {
5312 /* regs dump ABI info */
5313 int size = sizeof(u64);
5315 if (data->regs_user.regs) {
5316 u64 mask = event->attr.sample_regs_user;
5317 size += hweight64(mask) * sizeof(u64);
5320 header->size += size;
5323 if (sample_type & PERF_SAMPLE_STACK_USER) {
5325 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5326 * processed as the last one or have additional check added
5327 * in case new sample type is added, because we could eat
5328 * up the rest of the sample size.
5330 u16 stack_size = event->attr.sample_stack_user;
5331 u16 size = sizeof(u64);
5333 stack_size = perf_sample_ustack_size(stack_size, header->size,
5334 data->regs_user.regs);
5337 * If there is something to dump, add space for the dump
5338 * itself and for the field that tells the dynamic size,
5339 * which is how many have been actually dumped.
5342 size += sizeof(u64) + stack_size;
5344 data->stack_user_size = stack_size;
5345 header->size += size;
5348 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5349 /* regs dump ABI info */
5350 int size = sizeof(u64);
5352 perf_sample_regs_intr(&data->regs_intr, regs);
5354 if (data->regs_intr.regs) {
5355 u64 mask = event->attr.sample_regs_intr;
5357 size += hweight64(mask) * sizeof(u64);
5360 header->size += size;
5364 static void perf_event_output(struct perf_event *event,
5365 struct perf_sample_data *data,
5366 struct pt_regs *regs)
5368 struct perf_output_handle handle;
5369 struct perf_event_header header;
5371 /* protect the callchain buffers */
5374 perf_prepare_sample(&header, data, event, regs);
5376 if (perf_output_begin(&handle, event, header.size))
5379 perf_output_sample(&handle, &header, data, event);
5381 perf_output_end(&handle);
5391 struct perf_read_event {
5392 struct perf_event_header header;
5399 perf_event_read_event(struct perf_event *event,
5400 struct task_struct *task)
5402 struct perf_output_handle handle;
5403 struct perf_sample_data sample;
5404 struct perf_read_event read_event = {
5406 .type = PERF_RECORD_READ,
5408 .size = sizeof(read_event) + event->read_size,
5410 .pid = perf_event_pid(event, task),
5411 .tid = perf_event_tid(event, task),
5415 perf_event_header__init_id(&read_event.header, &sample, event);
5416 ret = perf_output_begin(&handle, event, read_event.header.size);
5420 perf_output_put(&handle, read_event);
5421 perf_output_read(&handle, event);
5422 perf_event__output_id_sample(event, &handle, &sample);
5424 perf_output_end(&handle);
5427 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5430 perf_event_aux_ctx(struct perf_event_context *ctx,
5431 perf_event_aux_output_cb output,
5434 struct perf_event *event;
5436 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5437 if (event->state < PERF_EVENT_STATE_INACTIVE)
5439 if (!event_filter_match(event))
5441 output(event, data);
5446 perf_event_aux(perf_event_aux_output_cb output, void *data,
5447 struct perf_event_context *task_ctx)
5449 struct perf_cpu_context *cpuctx;
5450 struct perf_event_context *ctx;
5455 list_for_each_entry_rcu(pmu, &pmus, entry) {
5456 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5457 if (cpuctx->unique_pmu != pmu)
5459 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5462 ctxn = pmu->task_ctx_nr;
5465 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5467 perf_event_aux_ctx(ctx, output, data);
5469 put_cpu_ptr(pmu->pmu_cpu_context);
5474 perf_event_aux_ctx(task_ctx, output, data);
5481 * task tracking -- fork/exit
5483 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5486 struct perf_task_event {
5487 struct task_struct *task;
5488 struct perf_event_context *task_ctx;
5491 struct perf_event_header header;
5501 static int perf_event_task_match(struct perf_event *event)
5503 return event->attr.comm || event->attr.mmap ||
5504 event->attr.mmap2 || event->attr.mmap_data ||
5508 static void perf_event_task_output(struct perf_event *event,
5511 struct perf_task_event *task_event = data;
5512 struct perf_output_handle handle;
5513 struct perf_sample_data sample;
5514 struct task_struct *task = task_event->task;
5515 int ret, size = task_event->event_id.header.size;
5517 if (!perf_event_task_match(event))
5520 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5522 ret = perf_output_begin(&handle, event,
5523 task_event->event_id.header.size);
5527 task_event->event_id.pid = perf_event_pid(event, task);
5528 task_event->event_id.ppid = perf_event_pid(event, current);
5530 task_event->event_id.tid = perf_event_tid(event, task);
5531 task_event->event_id.ptid = perf_event_tid(event, current);
5533 task_event->event_id.time = perf_event_clock(event);
5535 perf_output_put(&handle, task_event->event_id);
5537 perf_event__output_id_sample(event, &handle, &sample);
5539 perf_output_end(&handle);
5541 task_event->event_id.header.size = size;
5544 static void perf_event_task(struct task_struct *task,
5545 struct perf_event_context *task_ctx,
5548 struct perf_task_event task_event;
5550 if (!atomic_read(&nr_comm_events) &&
5551 !atomic_read(&nr_mmap_events) &&
5552 !atomic_read(&nr_task_events))
5555 task_event = (struct perf_task_event){
5557 .task_ctx = task_ctx,
5560 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5562 .size = sizeof(task_event.event_id),
5572 perf_event_aux(perf_event_task_output,
5577 void perf_event_fork(struct task_struct *task)
5579 perf_event_task(task, NULL, 1);
5586 struct perf_comm_event {
5587 struct task_struct *task;
5592 struct perf_event_header header;
5599 static int perf_event_comm_match(struct perf_event *event)
5601 return event->attr.comm;
5604 static void perf_event_comm_output(struct perf_event *event,
5607 struct perf_comm_event *comm_event = data;
5608 struct perf_output_handle handle;
5609 struct perf_sample_data sample;
5610 int size = comm_event->event_id.header.size;
5613 if (!perf_event_comm_match(event))
5616 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5617 ret = perf_output_begin(&handle, event,
5618 comm_event->event_id.header.size);
5623 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5624 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5626 perf_output_put(&handle, comm_event->event_id);
5627 __output_copy(&handle, comm_event->comm,
5628 comm_event->comm_size);
5630 perf_event__output_id_sample(event, &handle, &sample);
5632 perf_output_end(&handle);
5634 comm_event->event_id.header.size = size;
5637 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5639 char comm[TASK_COMM_LEN];
5642 memset(comm, 0, sizeof(comm));
5643 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5644 size = ALIGN(strlen(comm)+1, sizeof(u64));
5646 comm_event->comm = comm;
5647 comm_event->comm_size = size;
5649 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5651 perf_event_aux(perf_event_comm_output,
5656 void perf_event_comm(struct task_struct *task, bool exec)
5658 struct perf_comm_event comm_event;
5660 if (!atomic_read(&nr_comm_events))
5663 comm_event = (struct perf_comm_event){
5669 .type = PERF_RECORD_COMM,
5670 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5678 perf_event_comm_event(&comm_event);
5685 struct perf_mmap_event {
5686 struct vm_area_struct *vma;
5688 const char *file_name;
5696 struct perf_event_header header;
5706 static int perf_event_mmap_match(struct perf_event *event,
5709 struct perf_mmap_event *mmap_event = data;
5710 struct vm_area_struct *vma = mmap_event->vma;
5711 int executable = vma->vm_flags & VM_EXEC;
5713 return (!executable && event->attr.mmap_data) ||
5714 (executable && (event->attr.mmap || event->attr.mmap2));
5717 static void perf_event_mmap_output(struct perf_event *event,
5720 struct perf_mmap_event *mmap_event = data;
5721 struct perf_output_handle handle;
5722 struct perf_sample_data sample;
5723 int size = mmap_event->event_id.header.size;
5726 if (!perf_event_mmap_match(event, data))
5729 if (event->attr.mmap2) {
5730 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5731 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5732 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5733 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5734 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5735 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5736 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5739 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5740 ret = perf_output_begin(&handle, event,
5741 mmap_event->event_id.header.size);
5745 mmap_event->event_id.pid = perf_event_pid(event, current);
5746 mmap_event->event_id.tid = perf_event_tid(event, current);
5748 perf_output_put(&handle, mmap_event->event_id);
5750 if (event->attr.mmap2) {
5751 perf_output_put(&handle, mmap_event->maj);
5752 perf_output_put(&handle, mmap_event->min);
5753 perf_output_put(&handle, mmap_event->ino);
5754 perf_output_put(&handle, mmap_event->ino_generation);
5755 perf_output_put(&handle, mmap_event->prot);
5756 perf_output_put(&handle, mmap_event->flags);
5759 __output_copy(&handle, mmap_event->file_name,
5760 mmap_event->file_size);
5762 perf_event__output_id_sample(event, &handle, &sample);
5764 perf_output_end(&handle);
5766 mmap_event->event_id.header.size = size;
5769 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5771 struct vm_area_struct *vma = mmap_event->vma;
5772 struct file *file = vma->vm_file;
5773 int maj = 0, min = 0;
5774 u64 ino = 0, gen = 0;
5775 u32 prot = 0, flags = 0;
5782 struct inode *inode;
5785 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5791 * d_path() works from the end of the rb backwards, so we
5792 * need to add enough zero bytes after the string to handle
5793 * the 64bit alignment we do later.
5795 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5800 inode = file_inode(vma->vm_file);
5801 dev = inode->i_sb->s_dev;
5803 gen = inode->i_generation;
5807 if (vma->vm_flags & VM_READ)
5809 if (vma->vm_flags & VM_WRITE)
5811 if (vma->vm_flags & VM_EXEC)
5814 if (vma->vm_flags & VM_MAYSHARE)
5817 flags = MAP_PRIVATE;
5819 if (vma->vm_flags & VM_DENYWRITE)
5820 flags |= MAP_DENYWRITE;
5821 if (vma->vm_flags & VM_MAYEXEC)
5822 flags |= MAP_EXECUTABLE;
5823 if (vma->vm_flags & VM_LOCKED)
5824 flags |= MAP_LOCKED;
5825 if (vma->vm_flags & VM_HUGETLB)
5826 flags |= MAP_HUGETLB;
5830 if (vma->vm_ops && vma->vm_ops->name) {
5831 name = (char *) vma->vm_ops->name(vma);
5836 name = (char *)arch_vma_name(vma);
5840 if (vma->vm_start <= vma->vm_mm->start_brk &&
5841 vma->vm_end >= vma->vm_mm->brk) {
5845 if (vma->vm_start <= vma->vm_mm->start_stack &&
5846 vma->vm_end >= vma->vm_mm->start_stack) {
5856 strlcpy(tmp, name, sizeof(tmp));
5860 * Since our buffer works in 8 byte units we need to align our string
5861 * size to a multiple of 8. However, we must guarantee the tail end is
5862 * zero'd out to avoid leaking random bits to userspace.
5864 size = strlen(name)+1;
5865 while (!IS_ALIGNED(size, sizeof(u64)))
5866 name[size++] = '\0';
5868 mmap_event->file_name = name;
5869 mmap_event->file_size = size;
5870 mmap_event->maj = maj;
5871 mmap_event->min = min;
5872 mmap_event->ino = ino;
5873 mmap_event->ino_generation = gen;
5874 mmap_event->prot = prot;
5875 mmap_event->flags = flags;
5877 if (!(vma->vm_flags & VM_EXEC))
5878 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5880 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5882 perf_event_aux(perf_event_mmap_output,
5889 void perf_event_mmap(struct vm_area_struct *vma)
5891 struct perf_mmap_event mmap_event;
5893 if (!atomic_read(&nr_mmap_events))
5896 mmap_event = (struct perf_mmap_event){
5902 .type = PERF_RECORD_MMAP,
5903 .misc = PERF_RECORD_MISC_USER,
5908 .start = vma->vm_start,
5909 .len = vma->vm_end - vma->vm_start,
5910 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5912 /* .maj (attr_mmap2 only) */
5913 /* .min (attr_mmap2 only) */
5914 /* .ino (attr_mmap2 only) */
5915 /* .ino_generation (attr_mmap2 only) */
5916 /* .prot (attr_mmap2 only) */
5917 /* .flags (attr_mmap2 only) */
5920 perf_event_mmap_event(&mmap_event);
5923 void perf_event_aux_event(struct perf_event *event, unsigned long head,
5924 unsigned long size, u64 flags)
5926 struct perf_output_handle handle;
5927 struct perf_sample_data sample;
5928 struct perf_aux_event {
5929 struct perf_event_header header;
5935 .type = PERF_RECORD_AUX,
5937 .size = sizeof(rec),
5945 perf_event_header__init_id(&rec.header, &sample, event);
5946 ret = perf_output_begin(&handle, event, rec.header.size);
5951 perf_output_put(&handle, rec);
5952 perf_event__output_id_sample(event, &handle, &sample);
5954 perf_output_end(&handle);
5958 * IRQ throttle logging
5961 static void perf_log_throttle(struct perf_event *event, int enable)
5963 struct perf_output_handle handle;
5964 struct perf_sample_data sample;
5968 struct perf_event_header header;
5972 } throttle_event = {
5974 .type = PERF_RECORD_THROTTLE,
5976 .size = sizeof(throttle_event),
5978 .time = perf_event_clock(event),
5979 .id = primary_event_id(event),
5980 .stream_id = event->id,
5984 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5986 perf_event_header__init_id(&throttle_event.header, &sample, event);
5988 ret = perf_output_begin(&handle, event,
5989 throttle_event.header.size);
5993 perf_output_put(&handle, throttle_event);
5994 perf_event__output_id_sample(event, &handle, &sample);
5995 perf_output_end(&handle);
5998 static void perf_log_itrace_start(struct perf_event *event)
6000 struct perf_output_handle handle;
6001 struct perf_sample_data sample;
6002 struct perf_aux_event {
6003 struct perf_event_header header;
6010 event = event->parent;
6012 if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
6013 event->hw.itrace_started)
6016 event->hw.itrace_started = 1;
6018 rec.header.type = PERF_RECORD_ITRACE_START;
6019 rec.header.misc = 0;
6020 rec.header.size = sizeof(rec);
6021 rec.pid = perf_event_pid(event, current);
6022 rec.tid = perf_event_tid(event, current);
6024 perf_event_header__init_id(&rec.header, &sample, event);
6025 ret = perf_output_begin(&handle, event, rec.header.size);
6030 perf_output_put(&handle, rec);
6031 perf_event__output_id_sample(event, &handle, &sample);
6033 perf_output_end(&handle);
6037 * Generic event overflow handling, sampling.
6040 static int __perf_event_overflow(struct perf_event *event,
6041 int throttle, struct perf_sample_data *data,
6042 struct pt_regs *regs)
6044 int events = atomic_read(&event->event_limit);
6045 struct hw_perf_event *hwc = &event->hw;
6050 * Non-sampling counters might still use the PMI to fold short
6051 * hardware counters, ignore those.
6053 if (unlikely(!is_sampling_event(event)))
6056 seq = __this_cpu_read(perf_throttled_seq);
6057 if (seq != hwc->interrupts_seq) {
6058 hwc->interrupts_seq = seq;
6059 hwc->interrupts = 1;
6062 if (unlikely(throttle
6063 && hwc->interrupts >= max_samples_per_tick)) {
6064 __this_cpu_inc(perf_throttled_count);
6065 hwc->interrupts = MAX_INTERRUPTS;
6066 perf_log_throttle(event, 0);
6067 tick_nohz_full_kick();
6072 if (event->attr.freq) {
6073 u64 now = perf_clock();
6074 s64 delta = now - hwc->freq_time_stamp;
6076 hwc->freq_time_stamp = now;
6078 if (delta > 0 && delta < 2*TICK_NSEC)
6079 perf_adjust_period(event, delta, hwc->last_period, true);
6083 * XXX event_limit might not quite work as expected on inherited
6087 event->pending_kill = POLL_IN;
6088 if (events && atomic_dec_and_test(&event->event_limit)) {
6090 event->pending_kill = POLL_HUP;
6091 event->pending_disable = 1;
6092 irq_work_queue(&event->pending);
6095 if (event->overflow_handler)
6096 event->overflow_handler(event, data, regs);
6098 perf_event_output(event, data, regs);
6100 if (event->fasync && event->pending_kill) {
6101 event->pending_wakeup = 1;
6102 irq_work_queue(&event->pending);
6108 int perf_event_overflow(struct perf_event *event,
6109 struct perf_sample_data *data,
6110 struct pt_regs *regs)
6112 return __perf_event_overflow(event, 1, data, regs);
6116 * Generic software event infrastructure
6119 struct swevent_htable {
6120 struct swevent_hlist *swevent_hlist;
6121 struct mutex hlist_mutex;
6124 /* Recursion avoidance in each contexts */
6125 int recursion[PERF_NR_CONTEXTS];
6127 /* Keeps track of cpu being initialized/exited */
6131 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
6134 * We directly increment event->count and keep a second value in
6135 * event->hw.period_left to count intervals. This period event
6136 * is kept in the range [-sample_period, 0] so that we can use the
6140 u64 perf_swevent_set_period(struct perf_event *event)
6142 struct hw_perf_event *hwc = &event->hw;
6143 u64 period = hwc->last_period;
6147 hwc->last_period = hwc->sample_period;
6150 old = val = local64_read(&hwc->period_left);
6154 nr = div64_u64(period + val, period);
6155 offset = nr * period;
6157 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
6163 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
6164 struct perf_sample_data *data,
6165 struct pt_regs *regs)
6167 struct hw_perf_event *hwc = &event->hw;
6171 overflow = perf_swevent_set_period(event);
6173 if (hwc->interrupts == MAX_INTERRUPTS)
6176 for (; overflow; overflow--) {
6177 if (__perf_event_overflow(event, throttle,
6180 * We inhibit the overflow from happening when
6181 * hwc->interrupts == MAX_INTERRUPTS.
6189 static void perf_swevent_event(struct perf_event *event, u64 nr,
6190 struct perf_sample_data *data,
6191 struct pt_regs *regs)
6193 struct hw_perf_event *hwc = &event->hw;
6195 local64_add(nr, &event->count);
6200 if (!is_sampling_event(event))
6203 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
6205 return perf_swevent_overflow(event, 1, data, regs);
6207 data->period = event->hw.last_period;
6209 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
6210 return perf_swevent_overflow(event, 1, data, regs);
6212 if (local64_add_negative(nr, &hwc->period_left))
6215 perf_swevent_overflow(event, 0, data, regs);
6218 static int perf_exclude_event(struct perf_event *event,
6219 struct pt_regs *regs)
6221 if (event->hw.state & PERF_HES_STOPPED)
6225 if (event->attr.exclude_user && user_mode(regs))
6228 if (event->attr.exclude_kernel && !user_mode(regs))
6235 static int perf_swevent_match(struct perf_event *event,
6236 enum perf_type_id type,
6238 struct perf_sample_data *data,
6239 struct pt_regs *regs)
6241 if (event->attr.type != type)
6244 if (event->attr.config != event_id)
6247 if (perf_exclude_event(event, regs))
6253 static inline u64 swevent_hash(u64 type, u32 event_id)
6255 u64 val = event_id | (type << 32);
6257 return hash_64(val, SWEVENT_HLIST_BITS);
6260 static inline struct hlist_head *
6261 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
6263 u64 hash = swevent_hash(type, event_id);
6265 return &hlist->heads[hash];
6268 /* For the read side: events when they trigger */
6269 static inline struct hlist_head *
6270 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
6272 struct swevent_hlist *hlist;
6274 hlist = rcu_dereference(swhash->swevent_hlist);
6278 return __find_swevent_head(hlist, type, event_id);
6281 /* For the event head insertion and removal in the hlist */
6282 static inline struct hlist_head *
6283 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
6285 struct swevent_hlist *hlist;
6286 u32 event_id = event->attr.config;
6287 u64 type = event->attr.type;
6290 * Event scheduling is always serialized against hlist allocation
6291 * and release. Which makes the protected version suitable here.
6292 * The context lock guarantees that.
6294 hlist = rcu_dereference_protected(swhash->swevent_hlist,
6295 lockdep_is_held(&event->ctx->lock));
6299 return __find_swevent_head(hlist, type, event_id);
6302 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
6304 struct perf_sample_data *data,
6305 struct pt_regs *regs)
6307 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6308 struct perf_event *event;
6309 struct hlist_head *head;
6312 head = find_swevent_head_rcu(swhash, type, event_id);
6316 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6317 if (perf_swevent_match(event, type, event_id, data, regs))
6318 perf_swevent_event(event, nr, data, regs);
6324 DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);
6326 int perf_swevent_get_recursion_context(void)
6328 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6330 return get_recursion_context(swhash->recursion);
6332 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
6334 inline void perf_swevent_put_recursion_context(int rctx)
6336 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6338 put_recursion_context(swhash->recursion, rctx);
6341 void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6343 struct perf_sample_data data;
6345 if (WARN_ON_ONCE(!regs))
6348 perf_sample_data_init(&data, addr, 0);
6349 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
6352 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6356 preempt_disable_notrace();
6357 rctx = perf_swevent_get_recursion_context();
6358 if (unlikely(rctx < 0))
6361 ___perf_sw_event(event_id, nr, regs, addr);
6363 perf_swevent_put_recursion_context(rctx);
6365 preempt_enable_notrace();
6368 static void perf_swevent_read(struct perf_event *event)
6372 static int perf_swevent_add(struct perf_event *event, int flags)
6374 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6375 struct hw_perf_event *hwc = &event->hw;
6376 struct hlist_head *head;
6378 if (is_sampling_event(event)) {
6379 hwc->last_period = hwc->sample_period;
6380 perf_swevent_set_period(event);
6383 hwc->state = !(flags & PERF_EF_START);
6385 head = find_swevent_head(swhash, event);
6388 * We can race with cpu hotplug code. Do not
6389 * WARN if the cpu just got unplugged.
6391 WARN_ON_ONCE(swhash->online);
6395 hlist_add_head_rcu(&event->hlist_entry, head);
6396 perf_event_update_userpage(event);
6401 static void perf_swevent_del(struct perf_event *event, int flags)
6403 hlist_del_rcu(&event->hlist_entry);
6406 static void perf_swevent_start(struct perf_event *event, int flags)
6408 event->hw.state = 0;
6411 static void perf_swevent_stop(struct perf_event *event, int flags)
6413 event->hw.state = PERF_HES_STOPPED;
6416 /* Deref the hlist from the update side */
6417 static inline struct swevent_hlist *
6418 swevent_hlist_deref(struct swevent_htable *swhash)
6420 return rcu_dereference_protected(swhash->swevent_hlist,
6421 lockdep_is_held(&swhash->hlist_mutex));
6424 static void swevent_hlist_release(struct swevent_htable *swhash)
6426 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
6431 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
6432 kfree_rcu(hlist, rcu_head);
6435 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
6437 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6439 mutex_lock(&swhash->hlist_mutex);
6441 if (!--swhash->hlist_refcount)
6442 swevent_hlist_release(swhash);
6444 mutex_unlock(&swhash->hlist_mutex);
6447 static void swevent_hlist_put(struct perf_event *event)
6451 for_each_possible_cpu(cpu)
6452 swevent_hlist_put_cpu(event, cpu);
6455 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6457 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6460 mutex_lock(&swhash->hlist_mutex);
6462 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6463 struct swevent_hlist *hlist;
6465 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6470 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6472 swhash->hlist_refcount++;
6474 mutex_unlock(&swhash->hlist_mutex);
6479 static int swevent_hlist_get(struct perf_event *event)
6482 int cpu, failed_cpu;
6485 for_each_possible_cpu(cpu) {
6486 err = swevent_hlist_get_cpu(event, cpu);
6496 for_each_possible_cpu(cpu) {
6497 if (cpu == failed_cpu)
6499 swevent_hlist_put_cpu(event, cpu);
6506 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6508 static void sw_perf_event_destroy(struct perf_event *event)
6510 u64 event_id = event->attr.config;
6512 WARN_ON(event->parent);
6514 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6515 swevent_hlist_put(event);
6518 static int perf_swevent_init(struct perf_event *event)
6520 u64 event_id = event->attr.config;
6522 if (event->attr.type != PERF_TYPE_SOFTWARE)
6526 * no branch sampling for software events
6528 if (has_branch_stack(event))
6532 case PERF_COUNT_SW_CPU_CLOCK:
6533 case PERF_COUNT_SW_TASK_CLOCK:
6540 if (event_id >= PERF_COUNT_SW_MAX)
6543 if (!event->parent) {
6546 err = swevent_hlist_get(event);
6550 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6551 event->destroy = sw_perf_event_destroy;
6557 static struct pmu perf_swevent = {
6558 .task_ctx_nr = perf_sw_context,
6560 .capabilities = PERF_PMU_CAP_NO_NMI,
6562 .event_init = perf_swevent_init,
6563 .add = perf_swevent_add,
6564 .del = perf_swevent_del,
6565 .start = perf_swevent_start,
6566 .stop = perf_swevent_stop,
6567 .read = perf_swevent_read,
6570 #ifdef CONFIG_EVENT_TRACING
6572 static int perf_tp_filter_match(struct perf_event *event,
6573 struct perf_sample_data *data)
6575 void *record = data->raw->data;
6577 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6582 static int perf_tp_event_match(struct perf_event *event,
6583 struct perf_sample_data *data,
6584 struct pt_regs *regs)
6586 if (event->hw.state & PERF_HES_STOPPED)
6589 * All tracepoints are from kernel-space.
6591 if (event->attr.exclude_kernel)
6594 if (!perf_tp_filter_match(event, data))
6600 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6601 struct pt_regs *regs, struct hlist_head *head, int rctx,
6602 struct task_struct *task)
6604 struct perf_sample_data data;
6605 struct perf_event *event;
6607 struct perf_raw_record raw = {
6612 perf_sample_data_init(&data, addr, 0);
6615 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6616 if (perf_tp_event_match(event, &data, regs))
6617 perf_swevent_event(event, count, &data, regs);
6621 * If we got specified a target task, also iterate its context and
6622 * deliver this event there too.
6624 if (task && task != current) {
6625 struct perf_event_context *ctx;
6626 struct trace_entry *entry = record;
6629 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6633 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6634 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6636 if (event->attr.config != entry->type)
6638 if (perf_tp_event_match(event, &data, regs))
6639 perf_swevent_event(event, count, &data, regs);
6645 perf_swevent_put_recursion_context(rctx);
6647 EXPORT_SYMBOL_GPL(perf_tp_event);
6649 static void tp_perf_event_destroy(struct perf_event *event)
6651 perf_trace_destroy(event);
6654 static int perf_tp_event_init(struct perf_event *event)
6658 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6662 * no branch sampling for tracepoint events
6664 if (has_branch_stack(event))
6667 err = perf_trace_init(event);
6671 event->destroy = tp_perf_event_destroy;
6676 static struct pmu perf_tracepoint = {
6677 .task_ctx_nr = perf_sw_context,
6679 .event_init = perf_tp_event_init,
6680 .add = perf_trace_add,
6681 .del = perf_trace_del,
6682 .start = perf_swevent_start,
6683 .stop = perf_swevent_stop,
6684 .read = perf_swevent_read,
6687 static inline void perf_tp_register(void)
6689 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6692 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6697 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6700 filter_str = strndup_user(arg, PAGE_SIZE);
6701 if (IS_ERR(filter_str))
6702 return PTR_ERR(filter_str);
6704 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6710 static void perf_event_free_filter(struct perf_event *event)
6712 ftrace_profile_free_filter(event);
6715 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6717 struct bpf_prog *prog;
6719 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6722 if (event->tp_event->prog)
6725 if (!(event->tp_event->flags & TRACE_EVENT_FL_KPROBE))
6726 /* bpf programs can only be attached to kprobes */
6729 prog = bpf_prog_get(prog_fd);
6731 return PTR_ERR(prog);
6733 if (prog->type != BPF_PROG_TYPE_KPROBE) {
6734 /* valid fd, but invalid bpf program type */
6739 event->tp_event->prog = prog;
6744 static void perf_event_free_bpf_prog(struct perf_event *event)
6746 struct bpf_prog *prog;
6748 if (!event->tp_event)
6751 prog = event->tp_event->prog;
6753 event->tp_event->prog = NULL;
6760 static inline void perf_tp_register(void)
6764 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6769 static void perf_event_free_filter(struct perf_event *event)
6773 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd)
6778 static void perf_event_free_bpf_prog(struct perf_event *event)
6781 #endif /* CONFIG_EVENT_TRACING */
6783 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6784 void perf_bp_event(struct perf_event *bp, void *data)
6786 struct perf_sample_data sample;
6787 struct pt_regs *regs = data;
6789 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6791 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6792 perf_swevent_event(bp, 1, &sample, regs);
6797 * hrtimer based swevent callback
6800 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6802 enum hrtimer_restart ret = HRTIMER_RESTART;
6803 struct perf_sample_data data;
6804 struct pt_regs *regs;
6805 struct perf_event *event;
6808 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6810 if (event->state != PERF_EVENT_STATE_ACTIVE)
6811 return HRTIMER_NORESTART;
6813 event->pmu->read(event);
6815 perf_sample_data_init(&data, 0, event->hw.last_period);
6816 regs = get_irq_regs();
6818 if (regs && !perf_exclude_event(event, regs)) {
6819 if (!(event->attr.exclude_idle && is_idle_task(current)))
6820 if (__perf_event_overflow(event, 1, &data, regs))
6821 ret = HRTIMER_NORESTART;
6824 period = max_t(u64, 10000, event->hw.sample_period);
6825 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6830 static void perf_swevent_start_hrtimer(struct perf_event *event)
6832 struct hw_perf_event *hwc = &event->hw;
6835 if (!is_sampling_event(event))
6838 period = local64_read(&hwc->period_left);
6843 local64_set(&hwc->period_left, 0);
6845 period = max_t(u64, 10000, hwc->sample_period);
6847 hrtimer_start(&hwc->hrtimer, ns_to_ktime(period),
6848 HRTIMER_MODE_REL_PINNED);
6851 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6853 struct hw_perf_event *hwc = &event->hw;
6855 if (is_sampling_event(event)) {
6856 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6857 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6859 hrtimer_cancel(&hwc->hrtimer);
6863 static void perf_swevent_init_hrtimer(struct perf_event *event)
6865 struct hw_perf_event *hwc = &event->hw;
6867 if (!is_sampling_event(event))
6870 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6871 hwc->hrtimer.function = perf_swevent_hrtimer;
6874 * Since hrtimers have a fixed rate, we can do a static freq->period
6875 * mapping and avoid the whole period adjust feedback stuff.
6877 if (event->attr.freq) {
6878 long freq = event->attr.sample_freq;
6880 event->attr.sample_period = NSEC_PER_SEC / freq;
6881 hwc->sample_period = event->attr.sample_period;
6882 local64_set(&hwc->period_left, hwc->sample_period);
6883 hwc->last_period = hwc->sample_period;
6884 event->attr.freq = 0;
6889 * Software event: cpu wall time clock
6892 static void cpu_clock_event_update(struct perf_event *event)
6897 now = local_clock();
6898 prev = local64_xchg(&event->hw.prev_count, now);
6899 local64_add(now - prev, &event->count);
6902 static void cpu_clock_event_start(struct perf_event *event, int flags)
6904 local64_set(&event->hw.prev_count, local_clock());
6905 perf_swevent_start_hrtimer(event);
6908 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6910 perf_swevent_cancel_hrtimer(event);
6911 cpu_clock_event_update(event);
6914 static int cpu_clock_event_add(struct perf_event *event, int flags)
6916 if (flags & PERF_EF_START)
6917 cpu_clock_event_start(event, flags);
6918 perf_event_update_userpage(event);
6923 static void cpu_clock_event_del(struct perf_event *event, int flags)
6925 cpu_clock_event_stop(event, flags);
6928 static void cpu_clock_event_read(struct perf_event *event)
6930 cpu_clock_event_update(event);
6933 static int cpu_clock_event_init(struct perf_event *event)
6935 if (event->attr.type != PERF_TYPE_SOFTWARE)
6938 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6942 * no branch sampling for software events
6944 if (has_branch_stack(event))
6947 perf_swevent_init_hrtimer(event);
6952 static struct pmu perf_cpu_clock = {
6953 .task_ctx_nr = perf_sw_context,
6955 .capabilities = PERF_PMU_CAP_NO_NMI,
6957 .event_init = cpu_clock_event_init,
6958 .add = cpu_clock_event_add,
6959 .del = cpu_clock_event_del,
6960 .start = cpu_clock_event_start,
6961 .stop = cpu_clock_event_stop,
6962 .read = cpu_clock_event_read,
6966 * Software event: task time clock
6969 static void task_clock_event_update(struct perf_event *event, u64 now)
6974 prev = local64_xchg(&event->hw.prev_count, now);
6976 local64_add(delta, &event->count);
6979 static void task_clock_event_start(struct perf_event *event, int flags)
6981 local64_set(&event->hw.prev_count, event->ctx->time);
6982 perf_swevent_start_hrtimer(event);
6985 static void task_clock_event_stop(struct perf_event *event, int flags)
6987 perf_swevent_cancel_hrtimer(event);
6988 task_clock_event_update(event, event->ctx->time);
6991 static int task_clock_event_add(struct perf_event *event, int flags)
6993 if (flags & PERF_EF_START)
6994 task_clock_event_start(event, flags);
6995 perf_event_update_userpage(event);
7000 static void task_clock_event_del(struct perf_event *event, int flags)
7002 task_clock_event_stop(event, PERF_EF_UPDATE);
7005 static void task_clock_event_read(struct perf_event *event)
7007 u64 now = perf_clock();
7008 u64 delta = now - event->ctx->timestamp;
7009 u64 time = event->ctx->time + delta;
7011 task_clock_event_update(event, time);
7014 static int task_clock_event_init(struct perf_event *event)
7016 if (event->attr.type != PERF_TYPE_SOFTWARE)
7019 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
7023 * no branch sampling for software events
7025 if (has_branch_stack(event))
7028 perf_swevent_init_hrtimer(event);
7033 static struct pmu perf_task_clock = {
7034 .task_ctx_nr = perf_sw_context,
7036 .capabilities = PERF_PMU_CAP_NO_NMI,
7038 .event_init = task_clock_event_init,
7039 .add = task_clock_event_add,
7040 .del = task_clock_event_del,
7041 .start = task_clock_event_start,
7042 .stop = task_clock_event_stop,
7043 .read = task_clock_event_read,
7046 static void perf_pmu_nop_void(struct pmu *pmu)
7050 static int perf_pmu_nop_int(struct pmu *pmu)
7055 static void perf_pmu_start_txn(struct pmu *pmu)
7057 perf_pmu_disable(pmu);
7060 static int perf_pmu_commit_txn(struct pmu *pmu)
7062 perf_pmu_enable(pmu);
7066 static void perf_pmu_cancel_txn(struct pmu *pmu)
7068 perf_pmu_enable(pmu);
7071 static int perf_event_idx_default(struct perf_event *event)
7077 * Ensures all contexts with the same task_ctx_nr have the same
7078 * pmu_cpu_context too.
7080 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
7087 list_for_each_entry(pmu, &pmus, entry) {
7088 if (pmu->task_ctx_nr == ctxn)
7089 return pmu->pmu_cpu_context;
7095 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
7099 for_each_possible_cpu(cpu) {
7100 struct perf_cpu_context *cpuctx;
7102 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7104 if (cpuctx->unique_pmu == old_pmu)
7105 cpuctx->unique_pmu = pmu;
7109 static void free_pmu_context(struct pmu *pmu)
7113 mutex_lock(&pmus_lock);
7115 * Like a real lame refcount.
7117 list_for_each_entry(i, &pmus, entry) {
7118 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
7119 update_pmu_context(i, pmu);
7124 free_percpu(pmu->pmu_cpu_context);
7126 mutex_unlock(&pmus_lock);
7128 static struct idr pmu_idr;
7131 type_show(struct device *dev, struct device_attribute *attr, char *page)
7133 struct pmu *pmu = dev_get_drvdata(dev);
7135 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
7137 static DEVICE_ATTR_RO(type);
7140 perf_event_mux_interval_ms_show(struct device *dev,
7141 struct device_attribute *attr,
7144 struct pmu *pmu = dev_get_drvdata(dev);
7146 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
7149 static DEFINE_MUTEX(mux_interval_mutex);
7152 perf_event_mux_interval_ms_store(struct device *dev,
7153 struct device_attribute *attr,
7154 const char *buf, size_t count)
7156 struct pmu *pmu = dev_get_drvdata(dev);
7157 int timer, cpu, ret;
7159 ret = kstrtoint(buf, 0, &timer);
7166 /* same value, noting to do */
7167 if (timer == pmu->hrtimer_interval_ms)
7170 mutex_lock(&mux_interval_mutex);
7171 pmu->hrtimer_interval_ms = timer;
7173 /* update all cpuctx for this PMU */
7175 for_each_online_cpu(cpu) {
7176 struct perf_cpu_context *cpuctx;
7177 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7178 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
7180 cpu_function_call(cpu,
7181 (remote_function_f)perf_mux_hrtimer_restart, cpuctx);
7184 mutex_unlock(&mux_interval_mutex);
7188 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
7190 static struct attribute *pmu_dev_attrs[] = {
7191 &dev_attr_type.attr,
7192 &dev_attr_perf_event_mux_interval_ms.attr,
7195 ATTRIBUTE_GROUPS(pmu_dev);
7197 static int pmu_bus_running;
7198 static struct bus_type pmu_bus = {
7199 .name = "event_source",
7200 .dev_groups = pmu_dev_groups,
7203 static void pmu_dev_release(struct device *dev)
7208 static int pmu_dev_alloc(struct pmu *pmu)
7212 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
7216 pmu->dev->groups = pmu->attr_groups;
7217 device_initialize(pmu->dev);
7218 ret = dev_set_name(pmu->dev, "%s", pmu->name);
7222 dev_set_drvdata(pmu->dev, pmu);
7223 pmu->dev->bus = &pmu_bus;
7224 pmu->dev->release = pmu_dev_release;
7225 ret = device_add(pmu->dev);
7233 put_device(pmu->dev);
7237 static struct lock_class_key cpuctx_mutex;
7238 static struct lock_class_key cpuctx_lock;
7240 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
7244 mutex_lock(&pmus_lock);
7246 pmu->pmu_disable_count = alloc_percpu(int);
7247 if (!pmu->pmu_disable_count)
7256 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
7264 if (pmu_bus_running) {
7265 ret = pmu_dev_alloc(pmu);
7271 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
7272 if (pmu->pmu_cpu_context)
7273 goto got_cpu_context;
7276 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
7277 if (!pmu->pmu_cpu_context)
7280 for_each_possible_cpu(cpu) {
7281 struct perf_cpu_context *cpuctx;
7283 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
7284 __perf_event_init_context(&cpuctx->ctx);
7285 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
7286 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
7287 cpuctx->ctx.pmu = pmu;
7289 __perf_mux_hrtimer_init(cpuctx, cpu);
7291 cpuctx->unique_pmu = pmu;
7295 if (!pmu->start_txn) {
7296 if (pmu->pmu_enable) {
7298 * If we have pmu_enable/pmu_disable calls, install
7299 * transaction stubs that use that to try and batch
7300 * hardware accesses.
7302 pmu->start_txn = perf_pmu_start_txn;
7303 pmu->commit_txn = perf_pmu_commit_txn;
7304 pmu->cancel_txn = perf_pmu_cancel_txn;
7306 pmu->start_txn = perf_pmu_nop_void;
7307 pmu->commit_txn = perf_pmu_nop_int;
7308 pmu->cancel_txn = perf_pmu_nop_void;
7312 if (!pmu->pmu_enable) {
7313 pmu->pmu_enable = perf_pmu_nop_void;
7314 pmu->pmu_disable = perf_pmu_nop_void;
7317 if (!pmu->event_idx)
7318 pmu->event_idx = perf_event_idx_default;
7320 list_add_rcu(&pmu->entry, &pmus);
7321 atomic_set(&pmu->exclusive_cnt, 0);
7324 mutex_unlock(&pmus_lock);
7329 device_del(pmu->dev);
7330 put_device(pmu->dev);
7333 if (pmu->type >= PERF_TYPE_MAX)
7334 idr_remove(&pmu_idr, pmu->type);
7337 free_percpu(pmu->pmu_disable_count);
7340 EXPORT_SYMBOL_GPL(perf_pmu_register);
7342 void perf_pmu_unregister(struct pmu *pmu)
7344 mutex_lock(&pmus_lock);
7345 list_del_rcu(&pmu->entry);
7346 mutex_unlock(&pmus_lock);
7349 * We dereference the pmu list under both SRCU and regular RCU, so
7350 * synchronize against both of those.
7352 synchronize_srcu(&pmus_srcu);
7355 free_percpu(pmu->pmu_disable_count);
7356 if (pmu->type >= PERF_TYPE_MAX)
7357 idr_remove(&pmu_idr, pmu->type);
7358 device_del(pmu->dev);
7359 put_device(pmu->dev);
7360 free_pmu_context(pmu);
7362 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
7364 static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
7366 struct perf_event_context *ctx = NULL;
7369 if (!try_module_get(pmu->module))
7372 if (event->group_leader != event) {
7373 ctx = perf_event_ctx_lock(event->group_leader);
7378 ret = pmu->event_init(event);
7381 perf_event_ctx_unlock(event->group_leader, ctx);
7384 module_put(pmu->module);
7389 struct pmu *perf_init_event(struct perf_event *event)
7391 struct pmu *pmu = NULL;
7395 idx = srcu_read_lock(&pmus_srcu);
7398 pmu = idr_find(&pmu_idr, event->attr.type);
7401 ret = perf_try_init_event(pmu, event);
7407 list_for_each_entry_rcu(pmu, &pmus, entry) {
7408 ret = perf_try_init_event(pmu, event);
7412 if (ret != -ENOENT) {
7417 pmu = ERR_PTR(-ENOENT);
7419 srcu_read_unlock(&pmus_srcu, idx);
7424 static void account_event_cpu(struct perf_event *event, int cpu)
7429 if (is_cgroup_event(event))
7430 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
7433 static void account_event(struct perf_event *event)
7438 if (event->attach_state & PERF_ATTACH_TASK)
7439 static_key_slow_inc(&perf_sched_events.key);
7440 if (event->attr.mmap || event->attr.mmap_data)
7441 atomic_inc(&nr_mmap_events);
7442 if (event->attr.comm)
7443 atomic_inc(&nr_comm_events);
7444 if (event->attr.task)
7445 atomic_inc(&nr_task_events);
7446 if (event->attr.freq) {
7447 if (atomic_inc_return(&nr_freq_events) == 1)
7448 tick_nohz_full_kick_all();
7450 if (has_branch_stack(event))
7451 static_key_slow_inc(&perf_sched_events.key);
7452 if (is_cgroup_event(event))
7453 static_key_slow_inc(&perf_sched_events.key);
7455 account_event_cpu(event, event->cpu);
7459 * Allocate and initialize a event structure
7461 static struct perf_event *
7462 perf_event_alloc(struct perf_event_attr *attr, int cpu,
7463 struct task_struct *task,
7464 struct perf_event *group_leader,
7465 struct perf_event *parent_event,
7466 perf_overflow_handler_t overflow_handler,
7467 void *context, int cgroup_fd)
7470 struct perf_event *event;
7471 struct hw_perf_event *hwc;
7474 if ((unsigned)cpu >= nr_cpu_ids) {
7475 if (!task || cpu != -1)
7476 return ERR_PTR(-EINVAL);
7479 event = kzalloc(sizeof(*event), GFP_KERNEL);
7481 return ERR_PTR(-ENOMEM);
7484 * Single events are their own group leaders, with an
7485 * empty sibling list:
7488 group_leader = event;
7490 mutex_init(&event->child_mutex);
7491 INIT_LIST_HEAD(&event->child_list);
7493 INIT_LIST_HEAD(&event->group_entry);
7494 INIT_LIST_HEAD(&event->event_entry);
7495 INIT_LIST_HEAD(&event->sibling_list);
7496 INIT_LIST_HEAD(&event->rb_entry);
7497 INIT_LIST_HEAD(&event->active_entry);
7498 INIT_HLIST_NODE(&event->hlist_entry);
7501 init_waitqueue_head(&event->waitq);
7502 init_irq_work(&event->pending, perf_pending_event);
7504 mutex_init(&event->mmap_mutex);
7506 atomic_long_set(&event->refcount, 1);
7508 event->attr = *attr;
7509 event->group_leader = group_leader;
7513 event->parent = parent_event;
7515 event->ns = get_pid_ns(task_active_pid_ns(current));
7516 event->id = atomic64_inc_return(&perf_event_id);
7518 event->state = PERF_EVENT_STATE_INACTIVE;
7521 event->attach_state = PERF_ATTACH_TASK;
7523 * XXX pmu::event_init needs to know what task to account to
7524 * and we cannot use the ctx information because we need the
7525 * pmu before we get a ctx.
7527 event->hw.target = task;
7530 event->clock = &local_clock;
7532 event->clock = parent_event->clock;
7534 if (!overflow_handler && parent_event) {
7535 overflow_handler = parent_event->overflow_handler;
7536 context = parent_event->overflow_handler_context;
7539 event->overflow_handler = overflow_handler;
7540 event->overflow_handler_context = context;
7542 perf_event__state_init(event);
7547 hwc->sample_period = attr->sample_period;
7548 if (attr->freq && attr->sample_freq)
7549 hwc->sample_period = 1;
7550 hwc->last_period = hwc->sample_period;
7552 local64_set(&hwc->period_left, hwc->sample_period);
7555 * we currently do not support PERF_FORMAT_GROUP on inherited events
7557 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7560 if (!has_branch_stack(event))
7561 event->attr.branch_sample_type = 0;
7563 if (cgroup_fd != -1) {
7564 err = perf_cgroup_connect(cgroup_fd, event, attr, group_leader);
7569 pmu = perf_init_event(event);
7572 else if (IS_ERR(pmu)) {
7577 err = exclusive_event_init(event);
7581 if (!event->parent) {
7582 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7583 err = get_callchain_buffers();
7592 exclusive_event_destroy(event);
7596 event->destroy(event);
7597 module_put(pmu->module);
7599 if (is_cgroup_event(event))
7600 perf_detach_cgroup(event);
7602 put_pid_ns(event->ns);
7605 return ERR_PTR(err);
7608 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7609 struct perf_event_attr *attr)
7614 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7618 * zero the full structure, so that a short copy will be nice.
7620 memset(attr, 0, sizeof(*attr));
7622 ret = get_user(size, &uattr->size);
7626 if (size > PAGE_SIZE) /* silly large */
7629 if (!size) /* abi compat */
7630 size = PERF_ATTR_SIZE_VER0;
7632 if (size < PERF_ATTR_SIZE_VER0)
7636 * If we're handed a bigger struct than we know of,
7637 * ensure all the unknown bits are 0 - i.e. new
7638 * user-space does not rely on any kernel feature
7639 * extensions we dont know about yet.
7641 if (size > sizeof(*attr)) {
7642 unsigned char __user *addr;
7643 unsigned char __user *end;
7646 addr = (void __user *)uattr + sizeof(*attr);
7647 end = (void __user *)uattr + size;
7649 for (; addr < end; addr++) {
7650 ret = get_user(val, addr);
7656 size = sizeof(*attr);
7659 ret = copy_from_user(attr, uattr, size);
7663 if (attr->__reserved_1)
7666 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7669 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7672 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7673 u64 mask = attr->branch_sample_type;
7675 /* only using defined bits */
7676 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7679 /* at least one branch bit must be set */
7680 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7683 /* propagate priv level, when not set for branch */
7684 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7686 /* exclude_kernel checked on syscall entry */
7687 if (!attr->exclude_kernel)
7688 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7690 if (!attr->exclude_user)
7691 mask |= PERF_SAMPLE_BRANCH_USER;
7693 if (!attr->exclude_hv)
7694 mask |= PERF_SAMPLE_BRANCH_HV;
7696 * adjust user setting (for HW filter setup)
7698 attr->branch_sample_type = mask;
7700 /* privileged levels capture (kernel, hv): check permissions */
7701 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7702 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7706 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7707 ret = perf_reg_validate(attr->sample_regs_user);
7712 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7713 if (!arch_perf_have_user_stack_dump())
7717 * We have __u32 type for the size, but so far
7718 * we can only use __u16 as maximum due to the
7719 * __u16 sample size limit.
7721 if (attr->sample_stack_user >= USHRT_MAX)
7723 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7727 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7728 ret = perf_reg_validate(attr->sample_regs_intr);
7733 put_user(sizeof(*attr), &uattr->size);
7739 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7741 struct ring_buffer *rb = NULL;
7747 /* don't allow circular references */
7748 if (event == output_event)
7752 * Don't allow cross-cpu buffers
7754 if (output_event->cpu != event->cpu)
7758 * If its not a per-cpu rb, it must be the same task.
7760 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7764 * Mixing clocks in the same buffer is trouble you don't need.
7766 if (output_event->clock != event->clock)
7770 * If both events generate aux data, they must be on the same PMU
7772 if (has_aux(event) && has_aux(output_event) &&
7773 event->pmu != output_event->pmu)
7777 mutex_lock(&event->mmap_mutex);
7778 /* Can't redirect output if we've got an active mmap() */
7779 if (atomic_read(&event->mmap_count))
7783 /* get the rb we want to redirect to */
7784 rb = ring_buffer_get(output_event);
7789 ring_buffer_attach(event, rb);
7793 mutex_unlock(&event->mmap_mutex);
7799 static void mutex_lock_double(struct mutex *a, struct mutex *b)
7805 mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
7808 static int perf_event_set_clock(struct perf_event *event, clockid_t clk_id)
7810 bool nmi_safe = false;
7813 case CLOCK_MONOTONIC:
7814 event->clock = &ktime_get_mono_fast_ns;
7818 case CLOCK_MONOTONIC_RAW:
7819 event->clock = &ktime_get_raw_fast_ns;
7823 case CLOCK_REALTIME:
7824 event->clock = &ktime_get_real_ns;
7827 case CLOCK_BOOTTIME:
7828 event->clock = &ktime_get_boot_ns;
7832 event->clock = &ktime_get_tai_ns;
7839 if (!nmi_safe && !(event->pmu->capabilities & PERF_PMU_CAP_NO_NMI))
7846 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7848 * @attr_uptr: event_id type attributes for monitoring/sampling
7851 * @group_fd: group leader event fd
7853 SYSCALL_DEFINE5(perf_event_open,
7854 struct perf_event_attr __user *, attr_uptr,
7855 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7857 struct perf_event *group_leader = NULL, *output_event = NULL;
7858 struct perf_event *event, *sibling;
7859 struct perf_event_attr attr;
7860 struct perf_event_context *ctx, *uninitialized_var(gctx);
7861 struct file *event_file = NULL;
7862 struct fd group = {NULL, 0};
7863 struct task_struct *task = NULL;
7868 int f_flags = O_RDWR;
7871 /* for future expandability... */
7872 if (flags & ~PERF_FLAG_ALL)
7875 err = perf_copy_attr(attr_uptr, &attr);
7879 if (!attr.exclude_kernel) {
7880 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7885 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7888 if (attr.sample_period & (1ULL << 63))
7893 * In cgroup mode, the pid argument is used to pass the fd
7894 * opened to the cgroup directory in cgroupfs. The cpu argument
7895 * designates the cpu on which to monitor threads from that
7898 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7901 if (flags & PERF_FLAG_FD_CLOEXEC)
7902 f_flags |= O_CLOEXEC;
7904 event_fd = get_unused_fd_flags(f_flags);
7908 if (group_fd != -1) {
7909 err = perf_fget_light(group_fd, &group);
7912 group_leader = group.file->private_data;
7913 if (flags & PERF_FLAG_FD_OUTPUT)
7914 output_event = group_leader;
7915 if (flags & PERF_FLAG_FD_NO_GROUP)
7916 group_leader = NULL;
7919 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7920 task = find_lively_task_by_vpid(pid);
7922 err = PTR_ERR(task);
7927 if (task && group_leader &&
7928 group_leader->attr.inherit != attr.inherit) {
7935 if (flags & PERF_FLAG_PID_CGROUP)
7938 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7939 NULL, NULL, cgroup_fd);
7940 if (IS_ERR(event)) {
7941 err = PTR_ERR(event);
7945 if (is_sampling_event(event)) {
7946 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7952 account_event(event);
7955 * Special case software events and allow them to be part of
7956 * any hardware group.
7960 if (attr.use_clockid) {
7961 err = perf_event_set_clock(event, attr.clockid);
7967 (is_software_event(event) != is_software_event(group_leader))) {
7968 if (is_software_event(event)) {
7970 * If event and group_leader are not both a software
7971 * event, and event is, then group leader is not.
7973 * Allow the addition of software events to !software
7974 * groups, this is safe because software events never
7977 pmu = group_leader->pmu;
7978 } else if (is_software_event(group_leader) &&
7979 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7981 * In case the group is a pure software group, and we
7982 * try to add a hardware event, move the whole group to
7983 * the hardware context.
7990 * Get the target context (task or percpu):
7992 ctx = find_get_context(pmu, task, event);
7998 if ((pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE) && group_leader) {
8004 put_task_struct(task);
8009 * Look up the group leader (we will attach this event to it):
8015 * Do not allow a recursive hierarchy (this new sibling
8016 * becoming part of another group-sibling):
8018 if (group_leader->group_leader != group_leader)
8021 /* All events in a group should have the same clock */
8022 if (group_leader->clock != event->clock)
8026 * Do not allow to attach to a group in a different
8027 * task or CPU context:
8031 * Make sure we're both on the same task, or both
8034 if (group_leader->ctx->task != ctx->task)
8038 * Make sure we're both events for the same CPU;
8039 * grouping events for different CPUs is broken; since
8040 * you can never concurrently schedule them anyhow.
8042 if (group_leader->cpu != event->cpu)
8045 if (group_leader->ctx != ctx)
8050 * Only a group leader can be exclusive or pinned
8052 if (attr.exclusive || attr.pinned)
8057 err = perf_event_set_output(event, output_event);
8062 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
8064 if (IS_ERR(event_file)) {
8065 err = PTR_ERR(event_file);
8070 gctx = group_leader->ctx;
8073 * See perf_event_ctx_lock() for comments on the details
8074 * of swizzling perf_event::ctx.
8076 mutex_lock_double(&gctx->mutex, &ctx->mutex);
8078 perf_remove_from_context(group_leader, false);
8080 list_for_each_entry(sibling, &group_leader->sibling_list,
8082 perf_remove_from_context(sibling, false);
8086 mutex_lock(&ctx->mutex);
8089 WARN_ON_ONCE(ctx->parent_ctx);
8093 * Wait for everybody to stop referencing the events through
8094 * the old lists, before installing it on new lists.
8099 * Install the group siblings before the group leader.
8101 * Because a group leader will try and install the entire group
8102 * (through the sibling list, which is still in-tact), we can
8103 * end up with siblings installed in the wrong context.
8105 * By installing siblings first we NO-OP because they're not
8106 * reachable through the group lists.
8108 list_for_each_entry(sibling, &group_leader->sibling_list,
8110 perf_event__state_init(sibling);
8111 perf_install_in_context(ctx, sibling, sibling->cpu);
8116 * Removing from the context ends up with disabled
8117 * event. What we want here is event in the initial
8118 * startup state, ready to be add into new context.
8120 perf_event__state_init(group_leader);
8121 perf_install_in_context(ctx, group_leader, group_leader->cpu);
8125 if (!exclusive_event_installable(event, ctx)) {
8127 mutex_unlock(&ctx->mutex);
8132 perf_install_in_context(ctx, event, event->cpu);
8133 perf_unpin_context(ctx);
8136 mutex_unlock(&gctx->mutex);
8139 mutex_unlock(&ctx->mutex);
8143 event->owner = current;
8145 mutex_lock(¤t->perf_event_mutex);
8146 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
8147 mutex_unlock(¤t->perf_event_mutex);
8150 * Precalculate sample_data sizes
8152 perf_event__header_size(event);
8153 perf_event__id_header_size(event);
8156 * Drop the reference on the group_event after placing the
8157 * new event on the sibling_list. This ensures destruction
8158 * of the group leader will find the pointer to itself in
8159 * perf_group_detach().
8162 fd_install(event_fd, event_file);
8166 perf_unpin_context(ctx);
8174 put_task_struct(task);
8178 put_unused_fd(event_fd);
8183 * perf_event_create_kernel_counter
8185 * @attr: attributes of the counter to create
8186 * @cpu: cpu in which the counter is bound
8187 * @task: task to profile (NULL for percpu)
8190 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
8191 struct task_struct *task,
8192 perf_overflow_handler_t overflow_handler,
8195 struct perf_event_context *ctx;
8196 struct perf_event *event;
8200 * Get the target context (task or percpu):
8203 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
8204 overflow_handler, context, -1);
8205 if (IS_ERR(event)) {
8206 err = PTR_ERR(event);
8210 /* Mark owner so we could distinguish it from user events. */
8211 event->owner = EVENT_OWNER_KERNEL;
8213 account_event(event);
8215 ctx = find_get_context(event->pmu, task, event);
8221 WARN_ON_ONCE(ctx->parent_ctx);
8222 mutex_lock(&ctx->mutex);
8223 if (!exclusive_event_installable(event, ctx)) {
8224 mutex_unlock(&ctx->mutex);
8225 perf_unpin_context(ctx);
8231 perf_install_in_context(ctx, event, cpu);
8232 perf_unpin_context(ctx);
8233 mutex_unlock(&ctx->mutex);
8240 return ERR_PTR(err);
8242 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
8244 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
8246 struct perf_event_context *src_ctx;
8247 struct perf_event_context *dst_ctx;
8248 struct perf_event *event, *tmp;
8251 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
8252 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
8255 * See perf_event_ctx_lock() for comments on the details
8256 * of swizzling perf_event::ctx.
8258 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex);
8259 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
8261 perf_remove_from_context(event, false);
8262 unaccount_event_cpu(event, src_cpu);
8264 list_add(&event->migrate_entry, &events);
8268 * Wait for the events to quiesce before re-instating them.
8273 * Re-instate events in 2 passes.
8275 * Skip over group leaders and only install siblings on this first
8276 * pass, siblings will not get enabled without a leader, however a
8277 * leader will enable its siblings, even if those are still on the old
8280 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8281 if (event->group_leader == event)
8284 list_del(&event->migrate_entry);
8285 if (event->state >= PERF_EVENT_STATE_OFF)
8286 event->state = PERF_EVENT_STATE_INACTIVE;
8287 account_event_cpu(event, dst_cpu);
8288 perf_install_in_context(dst_ctx, event, dst_cpu);
8293 * Once all the siblings are setup properly, install the group leaders
8296 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
8297 list_del(&event->migrate_entry);
8298 if (event->state >= PERF_EVENT_STATE_OFF)
8299 event->state = PERF_EVENT_STATE_INACTIVE;
8300 account_event_cpu(event, dst_cpu);
8301 perf_install_in_context(dst_ctx, event, dst_cpu);
8304 mutex_unlock(&dst_ctx->mutex);
8305 mutex_unlock(&src_ctx->mutex);
8307 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
8309 static void sync_child_event(struct perf_event *child_event,
8310 struct task_struct *child)
8312 struct perf_event *parent_event = child_event->parent;
8315 if (child_event->attr.inherit_stat)
8316 perf_event_read_event(child_event, child);
8318 child_val = perf_event_count(child_event);
8321 * Add back the child's count to the parent's count:
8323 atomic64_add(child_val, &parent_event->child_count);
8324 atomic64_add(child_event->total_time_enabled,
8325 &parent_event->child_total_time_enabled);
8326 atomic64_add(child_event->total_time_running,
8327 &parent_event->child_total_time_running);
8330 * Remove this event from the parent's list
8332 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8333 mutex_lock(&parent_event->child_mutex);
8334 list_del_init(&child_event->child_list);
8335 mutex_unlock(&parent_event->child_mutex);
8338 * Make sure user/parent get notified, that we just
8341 perf_event_wakeup(parent_event);
8344 * Release the parent event, if this was the last
8347 put_event(parent_event);
8351 __perf_event_exit_task(struct perf_event *child_event,
8352 struct perf_event_context *child_ctx,
8353 struct task_struct *child)
8356 * Do not destroy the 'original' grouping; because of the context
8357 * switch optimization the original events could've ended up in a
8358 * random child task.
8360 * If we were to destroy the original group, all group related
8361 * operations would cease to function properly after this random
8364 * Do destroy all inherited groups, we don't care about those
8365 * and being thorough is better.
8367 perf_remove_from_context(child_event, !!child_event->parent);
8370 * It can happen that the parent exits first, and has events
8371 * that are still around due to the child reference. These
8372 * events need to be zapped.
8374 if (child_event->parent) {
8375 sync_child_event(child_event, child);
8376 free_event(child_event);
8378 child_event->state = PERF_EVENT_STATE_EXIT;
8379 perf_event_wakeup(child_event);
8383 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
8385 struct perf_event *child_event, *next;
8386 struct perf_event_context *child_ctx, *clone_ctx = NULL;
8387 unsigned long flags;
8389 if (likely(!child->perf_event_ctxp[ctxn])) {
8390 perf_event_task(child, NULL, 0);
8394 local_irq_save(flags);
8396 * We can't reschedule here because interrupts are disabled,
8397 * and either child is current or it is a task that can't be
8398 * scheduled, so we are now safe from rescheduling changing
8401 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
8404 * Take the context lock here so that if find_get_context is
8405 * reading child->perf_event_ctxp, we wait until it has
8406 * incremented the context's refcount before we do put_ctx below.
8408 raw_spin_lock(&child_ctx->lock);
8409 task_ctx_sched_out(child_ctx);
8410 child->perf_event_ctxp[ctxn] = NULL;
8413 * If this context is a clone; unclone it so it can't get
8414 * swapped to another process while we're removing all
8415 * the events from it.
8417 clone_ctx = unclone_ctx(child_ctx);
8418 update_context_time(child_ctx);
8419 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8425 * Report the task dead after unscheduling the events so that we
8426 * won't get any samples after PERF_RECORD_EXIT. We can however still
8427 * get a few PERF_RECORD_READ events.
8429 perf_event_task(child, child_ctx, 0);
8432 * We can recurse on the same lock type through:
8434 * __perf_event_exit_task()
8435 * sync_child_event()
8437 * mutex_lock(&ctx->mutex)
8439 * But since its the parent context it won't be the same instance.
8441 mutex_lock(&child_ctx->mutex);
8443 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
8444 __perf_event_exit_task(child_event, child_ctx, child);
8446 mutex_unlock(&child_ctx->mutex);
8452 * When a child task exits, feed back event values to parent events.
8454 void perf_event_exit_task(struct task_struct *child)
8456 struct perf_event *event, *tmp;
8459 mutex_lock(&child->perf_event_mutex);
8460 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
8462 list_del_init(&event->owner_entry);
8465 * Ensure the list deletion is visible before we clear
8466 * the owner, closes a race against perf_release() where
8467 * we need to serialize on the owner->perf_event_mutex.
8470 event->owner = NULL;
8472 mutex_unlock(&child->perf_event_mutex);
8474 for_each_task_context_nr(ctxn)
8475 perf_event_exit_task_context(child, ctxn);
8478 static void perf_free_event(struct perf_event *event,
8479 struct perf_event_context *ctx)
8481 struct perf_event *parent = event->parent;
8483 if (WARN_ON_ONCE(!parent))
8486 mutex_lock(&parent->child_mutex);
8487 list_del_init(&event->child_list);
8488 mutex_unlock(&parent->child_mutex);
8492 raw_spin_lock_irq(&ctx->lock);
8493 perf_group_detach(event);
8494 list_del_event(event, ctx);
8495 raw_spin_unlock_irq(&ctx->lock);
8500 * Free an unexposed, unused context as created by inheritance by
8501 * perf_event_init_task below, used by fork() in case of fail.
8503 * Not all locks are strictly required, but take them anyway to be nice and
8504 * help out with the lockdep assertions.
8506 void perf_event_free_task(struct task_struct *task)
8508 struct perf_event_context *ctx;
8509 struct perf_event *event, *tmp;
8512 for_each_task_context_nr(ctxn) {
8513 ctx = task->perf_event_ctxp[ctxn];
8517 mutex_lock(&ctx->mutex);
8519 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
8521 perf_free_event(event, ctx);
8523 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
8525 perf_free_event(event, ctx);
8527 if (!list_empty(&ctx->pinned_groups) ||
8528 !list_empty(&ctx->flexible_groups))
8531 mutex_unlock(&ctx->mutex);
8537 void perf_event_delayed_put(struct task_struct *task)
8541 for_each_task_context_nr(ctxn)
8542 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
8546 * inherit a event from parent task to child task:
8548 static struct perf_event *
8549 inherit_event(struct perf_event *parent_event,
8550 struct task_struct *parent,
8551 struct perf_event_context *parent_ctx,
8552 struct task_struct *child,
8553 struct perf_event *group_leader,
8554 struct perf_event_context *child_ctx)
8556 enum perf_event_active_state parent_state = parent_event->state;
8557 struct perf_event *child_event;
8558 unsigned long flags;
8561 * Instead of creating recursive hierarchies of events,
8562 * we link inherited events back to the original parent,
8563 * which has a filp for sure, which we use as the reference
8566 if (parent_event->parent)
8567 parent_event = parent_event->parent;
8569 child_event = perf_event_alloc(&parent_event->attr,
8572 group_leader, parent_event,
8574 if (IS_ERR(child_event))
8577 if (is_orphaned_event(parent_event) ||
8578 !atomic_long_inc_not_zero(&parent_event->refcount)) {
8579 free_event(child_event);
8586 * Make the child state follow the state of the parent event,
8587 * not its attr.disabled bit. We hold the parent's mutex,
8588 * so we won't race with perf_event_{en, dis}able_family.
8590 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
8591 child_event->state = PERF_EVENT_STATE_INACTIVE;
8593 child_event->state = PERF_EVENT_STATE_OFF;
8595 if (parent_event->attr.freq) {
8596 u64 sample_period = parent_event->hw.sample_period;
8597 struct hw_perf_event *hwc = &child_event->hw;
8599 hwc->sample_period = sample_period;
8600 hwc->last_period = sample_period;
8602 local64_set(&hwc->period_left, sample_period);
8605 child_event->ctx = child_ctx;
8606 child_event->overflow_handler = parent_event->overflow_handler;
8607 child_event->overflow_handler_context
8608 = parent_event->overflow_handler_context;
8611 * Precalculate sample_data sizes
8613 perf_event__header_size(child_event);
8614 perf_event__id_header_size(child_event);
8617 * Link it up in the child's context:
8619 raw_spin_lock_irqsave(&child_ctx->lock, flags);
8620 add_event_to_ctx(child_event, child_ctx);
8621 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8624 * Link this into the parent event's child list
8626 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8627 mutex_lock(&parent_event->child_mutex);
8628 list_add_tail(&child_event->child_list, &parent_event->child_list);
8629 mutex_unlock(&parent_event->child_mutex);
8634 static int inherit_group(struct perf_event *parent_event,
8635 struct task_struct *parent,
8636 struct perf_event_context *parent_ctx,
8637 struct task_struct *child,
8638 struct perf_event_context *child_ctx)
8640 struct perf_event *leader;
8641 struct perf_event *sub;
8642 struct perf_event *child_ctr;
8644 leader = inherit_event(parent_event, parent, parent_ctx,
8645 child, NULL, child_ctx);
8647 return PTR_ERR(leader);
8648 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
8649 child_ctr = inherit_event(sub, parent, parent_ctx,
8650 child, leader, child_ctx);
8651 if (IS_ERR(child_ctr))
8652 return PTR_ERR(child_ctr);
8658 inherit_task_group(struct perf_event *event, struct task_struct *parent,
8659 struct perf_event_context *parent_ctx,
8660 struct task_struct *child, int ctxn,
8664 struct perf_event_context *child_ctx;
8666 if (!event->attr.inherit) {
8671 child_ctx = child->perf_event_ctxp[ctxn];
8674 * This is executed from the parent task context, so
8675 * inherit events that have been marked for cloning.
8676 * First allocate and initialize a context for the
8680 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8684 child->perf_event_ctxp[ctxn] = child_ctx;
8687 ret = inherit_group(event, parent, parent_ctx,
8697 * Initialize the perf_event context in task_struct
8699 static int perf_event_init_context(struct task_struct *child, int ctxn)
8701 struct perf_event_context *child_ctx, *parent_ctx;
8702 struct perf_event_context *cloned_ctx;
8703 struct perf_event *event;
8704 struct task_struct *parent = current;
8705 int inherited_all = 1;
8706 unsigned long flags;
8709 if (likely(!parent->perf_event_ctxp[ctxn]))
8713 * If the parent's context is a clone, pin it so it won't get
8716 parent_ctx = perf_pin_task_context(parent, ctxn);
8721 * No need to check if parent_ctx != NULL here; since we saw
8722 * it non-NULL earlier, the only reason for it to become NULL
8723 * is if we exit, and since we're currently in the middle of
8724 * a fork we can't be exiting at the same time.
8728 * Lock the parent list. No need to lock the child - not PID
8729 * hashed yet and not running, so nobody can access it.
8731 mutex_lock(&parent_ctx->mutex);
8734 * We dont have to disable NMIs - we are only looking at
8735 * the list, not manipulating it:
8737 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8738 ret = inherit_task_group(event, parent, parent_ctx,
8739 child, ctxn, &inherited_all);
8745 * We can't hold ctx->lock when iterating the ->flexible_group list due
8746 * to allocations, but we need to prevent rotation because
8747 * rotate_ctx() will change the list from interrupt context.
8749 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8750 parent_ctx->rotate_disable = 1;
8751 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8753 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8754 ret = inherit_task_group(event, parent, parent_ctx,
8755 child, ctxn, &inherited_all);
8760 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8761 parent_ctx->rotate_disable = 0;
8763 child_ctx = child->perf_event_ctxp[ctxn];
8765 if (child_ctx && inherited_all) {
8767 * Mark the child context as a clone of the parent
8768 * context, or of whatever the parent is a clone of.
8770 * Note that if the parent is a clone, the holding of
8771 * parent_ctx->lock avoids it from being uncloned.
8773 cloned_ctx = parent_ctx->parent_ctx;
8775 child_ctx->parent_ctx = cloned_ctx;
8776 child_ctx->parent_gen = parent_ctx->parent_gen;
8778 child_ctx->parent_ctx = parent_ctx;
8779 child_ctx->parent_gen = parent_ctx->generation;
8781 get_ctx(child_ctx->parent_ctx);
8784 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8785 mutex_unlock(&parent_ctx->mutex);
8787 perf_unpin_context(parent_ctx);
8788 put_ctx(parent_ctx);
8794 * Initialize the perf_event context in task_struct
8796 int perf_event_init_task(struct task_struct *child)
8800 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8801 mutex_init(&child->perf_event_mutex);
8802 INIT_LIST_HEAD(&child->perf_event_list);
8804 for_each_task_context_nr(ctxn) {
8805 ret = perf_event_init_context(child, ctxn);
8807 perf_event_free_task(child);
8815 static void __init perf_event_init_all_cpus(void)
8817 struct swevent_htable *swhash;
8820 for_each_possible_cpu(cpu) {
8821 swhash = &per_cpu(swevent_htable, cpu);
8822 mutex_init(&swhash->hlist_mutex);
8823 INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
8827 static void perf_event_init_cpu(int cpu)
8829 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8831 mutex_lock(&swhash->hlist_mutex);
8832 swhash->online = true;
8833 if (swhash->hlist_refcount > 0) {
8834 struct swevent_hlist *hlist;
8836 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8838 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8840 mutex_unlock(&swhash->hlist_mutex);
8843 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8844 static void __perf_event_exit_context(void *__info)
8846 struct remove_event re = { .detach_group = true };
8847 struct perf_event_context *ctx = __info;
8850 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8851 __perf_remove_from_context(&re);
8855 static void perf_event_exit_cpu_context(int cpu)
8857 struct perf_event_context *ctx;
8861 idx = srcu_read_lock(&pmus_srcu);
8862 list_for_each_entry_rcu(pmu, &pmus, entry) {
8863 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8865 mutex_lock(&ctx->mutex);
8866 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8867 mutex_unlock(&ctx->mutex);
8869 srcu_read_unlock(&pmus_srcu, idx);
8872 static void perf_event_exit_cpu(int cpu)
8874 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8876 perf_event_exit_cpu_context(cpu);
8878 mutex_lock(&swhash->hlist_mutex);
8879 swhash->online = false;
8880 swevent_hlist_release(swhash);
8881 mutex_unlock(&swhash->hlist_mutex);
8884 static inline void perf_event_exit_cpu(int cpu) { }
8888 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8892 for_each_online_cpu(cpu)
8893 perf_event_exit_cpu(cpu);
8899 * Run the perf reboot notifier at the very last possible moment so that
8900 * the generic watchdog code runs as long as possible.
8902 static struct notifier_block perf_reboot_notifier = {
8903 .notifier_call = perf_reboot,
8904 .priority = INT_MIN,
8908 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8910 unsigned int cpu = (long)hcpu;
8912 switch (action & ~CPU_TASKS_FROZEN) {
8914 case CPU_UP_PREPARE:
8915 case CPU_DOWN_FAILED:
8916 perf_event_init_cpu(cpu);
8919 case CPU_UP_CANCELED:
8920 case CPU_DOWN_PREPARE:
8921 perf_event_exit_cpu(cpu);
8930 void __init perf_event_init(void)
8936 perf_event_init_all_cpus();
8937 init_srcu_struct(&pmus_srcu);
8938 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8939 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8940 perf_pmu_register(&perf_task_clock, NULL, -1);
8942 perf_cpu_notifier(perf_cpu_notify);
8943 register_reboot_notifier(&perf_reboot_notifier);
8945 ret = init_hw_breakpoint();
8946 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8948 /* do not patch jump label more than once per second */
8949 jump_label_rate_limit(&perf_sched_events, HZ);
8952 * Build time assertion that we keep the data_head at the intended
8953 * location. IOW, validation we got the __reserved[] size right.
8955 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8959 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
8962 struct perf_pmu_events_attr *pmu_attr =
8963 container_of(attr, struct perf_pmu_events_attr, attr);
8965 if (pmu_attr->event_str)
8966 return sprintf(page, "%s\n", pmu_attr->event_str);
8971 static int __init perf_event_sysfs_init(void)
8976 mutex_lock(&pmus_lock);
8978 ret = bus_register(&pmu_bus);
8982 list_for_each_entry(pmu, &pmus, entry) {
8983 if (!pmu->name || pmu->type < 0)
8986 ret = pmu_dev_alloc(pmu);
8987 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8989 pmu_bus_running = 1;
8993 mutex_unlock(&pmus_lock);
8997 device_initcall(perf_event_sysfs_init);
8999 #ifdef CONFIG_CGROUP_PERF
9000 static struct cgroup_subsys_state *
9001 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
9003 struct perf_cgroup *jc;
9005 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
9007 return ERR_PTR(-ENOMEM);
9009 jc->info = alloc_percpu(struct perf_cgroup_info);
9012 return ERR_PTR(-ENOMEM);
9018 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
9020 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
9022 free_percpu(jc->info);
9026 static int __perf_cgroup_move(void *info)
9028 struct task_struct *task = info;
9029 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
9033 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
9034 struct cgroup_taskset *tset)
9036 struct task_struct *task;
9038 cgroup_taskset_for_each(task, tset)
9039 task_function_call(task, __perf_cgroup_move, task);
9042 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
9043 struct cgroup_subsys_state *old_css,
9044 struct task_struct *task)
9047 * cgroup_exit() is called in the copy_process() failure path.
9048 * Ignore this case since the task hasn't ran yet, this avoids
9049 * trying to poke a half freed task state from generic code.
9051 if (!(task->flags & PF_EXITING))
9054 task_function_call(task, __perf_cgroup_move, task);
9057 struct cgroup_subsys perf_event_cgrp_subsys = {
9058 .css_alloc = perf_cgroup_css_alloc,
9059 .css_free = perf_cgroup_css_free,
9060 .exit = perf_cgroup_exit,
9061 .attach = perf_cgroup_attach,
9063 #endif /* CONFIG_CGROUP_PERF */