2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
42 #include <linux/module.h>
43 #include <linux/mman.h>
44 #include <linux/compat.h>
48 #include <asm/irq_regs.h>
50 static struct workqueue_struct *perf_wq;
52 struct remote_function_call {
53 struct task_struct *p;
54 int (*func)(void *info);
59 static void remote_function(void *data)
61 struct remote_function_call *tfc = data;
62 struct task_struct *p = tfc->p;
66 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
70 tfc->ret = tfc->func(tfc->info);
74 * task_function_call - call a function on the cpu on which a task runs
75 * @p: the task to evaluate
76 * @func: the function to be called
77 * @info: the function call argument
79 * Calls the function @func when the task is currently running. This might
80 * be on the current CPU, which just calls the function directly
82 * returns: @func return value, or
83 * -ESRCH - when the process isn't running
84 * -EAGAIN - when the process moved away
87 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
89 struct remote_function_call data = {
93 .ret = -ESRCH, /* No such (running) process */
97 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
103 * cpu_function_call - call a function on the cpu
104 * @func: the function to be called
105 * @info: the function call argument
107 * Calls the function @func on the remote cpu.
109 * returns: @func return value or -ENXIO when the cpu is offline
111 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
113 struct remote_function_call data = {
117 .ret = -ENXIO, /* No such CPU */
120 smp_call_function_single(cpu, remote_function, &data, 1);
125 #define EVENT_OWNER_KERNEL ((void *) -1)
127 static bool is_kernel_event(struct perf_event *event)
129 return event->owner == EVENT_OWNER_KERNEL;
132 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
133 PERF_FLAG_FD_OUTPUT |\
134 PERF_FLAG_PID_CGROUP |\
135 PERF_FLAG_FD_CLOEXEC)
138 * branch priv levels that need permission checks
140 #define PERF_SAMPLE_BRANCH_PERM_PLM \
141 (PERF_SAMPLE_BRANCH_KERNEL |\
142 PERF_SAMPLE_BRANCH_HV)
145 EVENT_FLEXIBLE = 0x1,
147 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
151 * perf_sched_events : >0 events exist
152 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
154 struct static_key_deferred perf_sched_events __read_mostly;
155 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
156 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
158 static atomic_t nr_mmap_events __read_mostly;
159 static atomic_t nr_comm_events __read_mostly;
160 static atomic_t nr_task_events __read_mostly;
161 static atomic_t nr_freq_events __read_mostly;
163 static LIST_HEAD(pmus);
164 static DEFINE_MUTEX(pmus_lock);
165 static struct srcu_struct pmus_srcu;
168 * perf event paranoia level:
169 * -1 - not paranoid at all
170 * 0 - disallow raw tracepoint access for unpriv
171 * 1 - disallow cpu events for unpriv
172 * 2 - disallow kernel profiling for unpriv
174 int sysctl_perf_event_paranoid __read_mostly = 1;
176 /* Minimum for 512 kiB + 1 user control page */
177 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
180 * max perf event sample rate
182 #define DEFAULT_MAX_SAMPLE_RATE 100000
183 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
184 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
186 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
188 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
189 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
191 static int perf_sample_allowed_ns __read_mostly =
192 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
194 void update_perf_cpu_limits(void)
196 u64 tmp = perf_sample_period_ns;
198 tmp *= sysctl_perf_cpu_time_max_percent;
200 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
203 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
205 int perf_proc_update_handler(struct ctl_table *table, int write,
206 void __user *buffer, size_t *lenp,
209 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
214 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
215 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
216 update_perf_cpu_limits();
221 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
223 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
224 void __user *buffer, size_t *lenp,
227 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
232 update_perf_cpu_limits();
238 * perf samples are done in some very critical code paths (NMIs).
239 * If they take too much CPU time, the system can lock up and not
240 * get any real work done. This will drop the sample rate when
241 * we detect that events are taking too long.
243 #define NR_ACCUMULATED_SAMPLES 128
244 static DEFINE_PER_CPU(u64, running_sample_length);
246 static void perf_duration_warn(struct irq_work *w)
248 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
249 u64 avg_local_sample_len;
250 u64 local_samples_len;
252 local_samples_len = __this_cpu_read(running_sample_length);
253 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
255 printk_ratelimited(KERN_WARNING
256 "perf interrupt took too long (%lld > %lld), lowering "
257 "kernel.perf_event_max_sample_rate to %d\n",
258 avg_local_sample_len, allowed_ns >> 1,
259 sysctl_perf_event_sample_rate);
262 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
264 void perf_sample_event_took(u64 sample_len_ns)
266 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
267 u64 avg_local_sample_len;
268 u64 local_samples_len;
273 /* decay the counter by 1 average sample */
274 local_samples_len = __this_cpu_read(running_sample_length);
275 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
276 local_samples_len += sample_len_ns;
277 __this_cpu_write(running_sample_length, local_samples_len);
280 * note: this will be biased artifically low until we have
281 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
282 * from having to maintain a count.
284 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
286 if (avg_local_sample_len <= allowed_ns)
289 if (max_samples_per_tick <= 1)
292 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
293 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
294 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
296 update_perf_cpu_limits();
298 if (!irq_work_queue(&perf_duration_work)) {
299 early_printk("perf interrupt took too long (%lld > %lld), lowering "
300 "kernel.perf_event_max_sample_rate to %d\n",
301 avg_local_sample_len, allowed_ns >> 1,
302 sysctl_perf_event_sample_rate);
306 static atomic64_t perf_event_id;
308 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
309 enum event_type_t event_type);
311 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
312 enum event_type_t event_type,
313 struct task_struct *task);
315 static void update_context_time(struct perf_event_context *ctx);
316 static u64 perf_event_time(struct perf_event *event);
318 void __weak perf_event_print_debug(void) { }
320 extern __weak const char *perf_pmu_name(void)
325 static inline u64 perf_clock(void)
327 return local_clock();
330 static inline struct perf_cpu_context *
331 __get_cpu_context(struct perf_event_context *ctx)
333 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
336 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
337 struct perf_event_context *ctx)
339 raw_spin_lock(&cpuctx->ctx.lock);
341 raw_spin_lock(&ctx->lock);
344 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
345 struct perf_event_context *ctx)
348 raw_spin_unlock(&ctx->lock);
349 raw_spin_unlock(&cpuctx->ctx.lock);
352 #ifdef CONFIG_CGROUP_PERF
355 * perf_cgroup_info keeps track of time_enabled for a cgroup.
356 * This is a per-cpu dynamically allocated data structure.
358 struct perf_cgroup_info {
364 struct cgroup_subsys_state css;
365 struct perf_cgroup_info __percpu *info;
369 * Must ensure cgroup is pinned (css_get) before calling
370 * this function. In other words, we cannot call this function
371 * if there is no cgroup event for the current CPU context.
373 static inline struct perf_cgroup *
374 perf_cgroup_from_task(struct task_struct *task)
376 return container_of(task_css(task, perf_event_cgrp_id),
377 struct perf_cgroup, css);
381 perf_cgroup_match(struct perf_event *event)
383 struct perf_event_context *ctx = event->ctx;
384 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
386 /* @event doesn't care about cgroup */
390 /* wants specific cgroup scope but @cpuctx isn't associated with any */
395 * Cgroup scoping is recursive. An event enabled for a cgroup is
396 * also enabled for all its descendant cgroups. If @cpuctx's
397 * cgroup is a descendant of @event's (the test covers identity
398 * case), it's a match.
400 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
401 event->cgrp->css.cgroup);
404 static inline void perf_detach_cgroup(struct perf_event *event)
406 css_put(&event->cgrp->css);
410 static inline int is_cgroup_event(struct perf_event *event)
412 return event->cgrp != NULL;
415 static inline u64 perf_cgroup_event_time(struct perf_event *event)
417 struct perf_cgroup_info *t;
419 t = per_cpu_ptr(event->cgrp->info, event->cpu);
423 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
425 struct perf_cgroup_info *info;
430 info = this_cpu_ptr(cgrp->info);
432 info->time += now - info->timestamp;
433 info->timestamp = now;
436 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
438 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
440 __update_cgrp_time(cgrp_out);
443 static inline void update_cgrp_time_from_event(struct perf_event *event)
445 struct perf_cgroup *cgrp;
448 * ensure we access cgroup data only when needed and
449 * when we know the cgroup is pinned (css_get)
451 if (!is_cgroup_event(event))
454 cgrp = perf_cgroup_from_task(current);
456 * Do not update time when cgroup is not active
458 if (cgrp == event->cgrp)
459 __update_cgrp_time(event->cgrp);
463 perf_cgroup_set_timestamp(struct task_struct *task,
464 struct perf_event_context *ctx)
466 struct perf_cgroup *cgrp;
467 struct perf_cgroup_info *info;
470 * ctx->lock held by caller
471 * ensure we do not access cgroup data
472 * unless we have the cgroup pinned (css_get)
474 if (!task || !ctx->nr_cgroups)
477 cgrp = perf_cgroup_from_task(task);
478 info = this_cpu_ptr(cgrp->info);
479 info->timestamp = ctx->timestamp;
482 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
483 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
486 * reschedule events based on the cgroup constraint of task.
488 * mode SWOUT : schedule out everything
489 * mode SWIN : schedule in based on cgroup for next
491 void perf_cgroup_switch(struct task_struct *task, int mode)
493 struct perf_cpu_context *cpuctx;
498 * disable interrupts to avoid geting nr_cgroup
499 * changes via __perf_event_disable(). Also
502 local_irq_save(flags);
505 * we reschedule only in the presence of cgroup
506 * constrained events.
510 list_for_each_entry_rcu(pmu, &pmus, entry) {
511 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
512 if (cpuctx->unique_pmu != pmu)
513 continue; /* ensure we process each cpuctx once */
516 * perf_cgroup_events says at least one
517 * context on this CPU has cgroup events.
519 * ctx->nr_cgroups reports the number of cgroup
520 * events for a context.
522 if (cpuctx->ctx.nr_cgroups > 0) {
523 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
524 perf_pmu_disable(cpuctx->ctx.pmu);
526 if (mode & PERF_CGROUP_SWOUT) {
527 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
529 * must not be done before ctxswout due
530 * to event_filter_match() in event_sched_out()
535 if (mode & PERF_CGROUP_SWIN) {
536 WARN_ON_ONCE(cpuctx->cgrp);
538 * set cgrp before ctxsw in to allow
539 * event_filter_match() to not have to pass
542 cpuctx->cgrp = perf_cgroup_from_task(task);
543 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
545 perf_pmu_enable(cpuctx->ctx.pmu);
546 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
552 local_irq_restore(flags);
555 static inline void perf_cgroup_sched_out(struct task_struct *task,
556 struct task_struct *next)
558 struct perf_cgroup *cgrp1;
559 struct perf_cgroup *cgrp2 = NULL;
562 * we come here when we know perf_cgroup_events > 0
564 cgrp1 = perf_cgroup_from_task(task);
567 * next is NULL when called from perf_event_enable_on_exec()
568 * that will systematically cause a cgroup_switch()
571 cgrp2 = perf_cgroup_from_task(next);
574 * only schedule out current cgroup events if we know
575 * that we are switching to a different cgroup. Otherwise,
576 * do no touch the cgroup events.
579 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
582 static inline void perf_cgroup_sched_in(struct task_struct *prev,
583 struct task_struct *task)
585 struct perf_cgroup *cgrp1;
586 struct perf_cgroup *cgrp2 = NULL;
589 * we come here when we know perf_cgroup_events > 0
591 cgrp1 = perf_cgroup_from_task(task);
593 /* prev can never be NULL */
594 cgrp2 = perf_cgroup_from_task(prev);
597 * only need to schedule in cgroup events if we are changing
598 * cgroup during ctxsw. Cgroup events were not scheduled
599 * out of ctxsw out if that was not the case.
602 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
605 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
606 struct perf_event_attr *attr,
607 struct perf_event *group_leader)
609 struct perf_cgroup *cgrp;
610 struct cgroup_subsys_state *css;
611 struct fd f = fdget(fd);
617 css = css_tryget_online_from_dir(f.file->f_path.dentry,
618 &perf_event_cgrp_subsys);
624 cgrp = container_of(css, struct perf_cgroup, css);
628 * all events in a group must monitor
629 * the same cgroup because a task belongs
630 * to only one perf cgroup at a time
632 if (group_leader && group_leader->cgrp != cgrp) {
633 perf_detach_cgroup(event);
642 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
644 struct perf_cgroup_info *t;
645 t = per_cpu_ptr(event->cgrp->info, event->cpu);
646 event->shadow_ctx_time = now - t->timestamp;
650 perf_cgroup_defer_enabled(struct perf_event *event)
653 * when the current task's perf cgroup does not match
654 * the event's, we need to remember to call the
655 * perf_mark_enable() function the first time a task with
656 * a matching perf cgroup is scheduled in.
658 if (is_cgroup_event(event) && !perf_cgroup_match(event))
659 event->cgrp_defer_enabled = 1;
663 perf_cgroup_mark_enabled(struct perf_event *event,
664 struct perf_event_context *ctx)
666 struct perf_event *sub;
667 u64 tstamp = perf_event_time(event);
669 if (!event->cgrp_defer_enabled)
672 event->cgrp_defer_enabled = 0;
674 event->tstamp_enabled = tstamp - event->total_time_enabled;
675 list_for_each_entry(sub, &event->sibling_list, group_entry) {
676 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
677 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
678 sub->cgrp_defer_enabled = 0;
682 #else /* !CONFIG_CGROUP_PERF */
685 perf_cgroup_match(struct perf_event *event)
690 static inline void perf_detach_cgroup(struct perf_event *event)
693 static inline int is_cgroup_event(struct perf_event *event)
698 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
703 static inline void update_cgrp_time_from_event(struct perf_event *event)
707 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
711 static inline void perf_cgroup_sched_out(struct task_struct *task,
712 struct task_struct *next)
716 static inline void perf_cgroup_sched_in(struct task_struct *prev,
717 struct task_struct *task)
721 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
722 struct perf_event_attr *attr,
723 struct perf_event *group_leader)
729 perf_cgroup_set_timestamp(struct task_struct *task,
730 struct perf_event_context *ctx)
735 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
740 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
744 static inline u64 perf_cgroup_event_time(struct perf_event *event)
750 perf_cgroup_defer_enabled(struct perf_event *event)
755 perf_cgroup_mark_enabled(struct perf_event *event,
756 struct perf_event_context *ctx)
762 * set default to be dependent on timer tick just
765 #define PERF_CPU_HRTIMER (1000 / HZ)
767 * function must be called with interrupts disbled
769 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
771 struct perf_cpu_context *cpuctx;
772 enum hrtimer_restart ret = HRTIMER_NORESTART;
775 WARN_ON(!irqs_disabled());
777 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
779 rotations = perf_rotate_context(cpuctx);
782 * arm timer if needed
785 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
786 ret = HRTIMER_RESTART;
792 /* CPU is going down */
793 void perf_cpu_hrtimer_cancel(int cpu)
795 struct perf_cpu_context *cpuctx;
799 if (WARN_ON(cpu != smp_processor_id()))
802 local_irq_save(flags);
806 list_for_each_entry_rcu(pmu, &pmus, entry) {
807 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
809 if (pmu->task_ctx_nr == perf_sw_context)
812 hrtimer_cancel(&cpuctx->hrtimer);
817 local_irq_restore(flags);
820 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
822 struct hrtimer *hr = &cpuctx->hrtimer;
823 struct pmu *pmu = cpuctx->ctx.pmu;
826 /* no multiplexing needed for SW PMU */
827 if (pmu->task_ctx_nr == perf_sw_context)
831 * check default is sane, if not set then force to
832 * default interval (1/tick)
834 timer = pmu->hrtimer_interval_ms;
836 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
838 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
840 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
841 hr->function = perf_cpu_hrtimer_handler;
844 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
846 struct hrtimer *hr = &cpuctx->hrtimer;
847 struct pmu *pmu = cpuctx->ctx.pmu;
850 if (pmu->task_ctx_nr == perf_sw_context)
853 if (hrtimer_active(hr))
856 if (!hrtimer_callback_running(hr))
857 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
858 0, HRTIMER_MODE_REL_PINNED, 0);
861 void perf_pmu_disable(struct pmu *pmu)
863 int *count = this_cpu_ptr(pmu->pmu_disable_count);
865 pmu->pmu_disable(pmu);
868 void perf_pmu_enable(struct pmu *pmu)
870 int *count = this_cpu_ptr(pmu->pmu_disable_count);
872 pmu->pmu_enable(pmu);
875 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
878 * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
879 * perf_event_task_tick() are fully serialized because they're strictly cpu
880 * affine and perf_event_ctx{activate,deactivate} are called with IRQs
881 * disabled, while perf_event_task_tick is called from IRQ context.
883 static void perf_event_ctx_activate(struct perf_event_context *ctx)
885 struct list_head *head = this_cpu_ptr(&active_ctx_list);
887 WARN_ON(!irqs_disabled());
889 WARN_ON(!list_empty(&ctx->active_ctx_list));
891 list_add(&ctx->active_ctx_list, head);
894 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
896 WARN_ON(!irqs_disabled());
898 WARN_ON(list_empty(&ctx->active_ctx_list));
900 list_del_init(&ctx->active_ctx_list);
903 static void get_ctx(struct perf_event_context *ctx)
905 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
908 static void put_ctx(struct perf_event_context *ctx)
910 if (atomic_dec_and_test(&ctx->refcount)) {
912 put_ctx(ctx->parent_ctx);
914 put_task_struct(ctx->task);
915 kfree_rcu(ctx, rcu_head);
920 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
921 * perf_pmu_migrate_context() we need some magic.
923 * Those places that change perf_event::ctx will hold both
924 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
926 * Lock ordering is by mutex address. There is one other site where
927 * perf_event_context::mutex nests and that is put_event(). But remember that
928 * that is a parent<->child context relation, and migration does not affect
929 * children, therefore these two orderings should not interact.
931 * The change in perf_event::ctx does not affect children (as claimed above)
932 * because the sys_perf_event_open() case will install a new event and break
933 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
934 * concerned with cpuctx and that doesn't have children.
936 * The places that change perf_event::ctx will issue:
938 * perf_remove_from_context();
940 * perf_install_in_context();
942 * to affect the change. The remove_from_context() + synchronize_rcu() should
943 * quiesce the event, after which we can install it in the new location. This
944 * means that only external vectors (perf_fops, prctl) can perturb the event
945 * while in transit. Therefore all such accessors should also acquire
946 * perf_event_context::mutex to serialize against this.
948 * However; because event->ctx can change while we're waiting to acquire
949 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
953 * task_struct::perf_event_mutex
954 * perf_event_context::mutex
955 * perf_event_context::lock
956 * perf_event::child_mutex;
957 * perf_event::mmap_mutex
960 static struct perf_event_context *
961 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
963 struct perf_event_context *ctx;
967 ctx = ACCESS_ONCE(event->ctx);
968 if (!atomic_inc_not_zero(&ctx->refcount)) {
974 mutex_lock_nested(&ctx->mutex, nesting);
975 if (event->ctx != ctx) {
976 mutex_unlock(&ctx->mutex);
984 static inline struct perf_event_context *
985 perf_event_ctx_lock(struct perf_event *event)
987 return perf_event_ctx_lock_nested(event, 0);
990 static void perf_event_ctx_unlock(struct perf_event *event,
991 struct perf_event_context *ctx)
993 mutex_unlock(&ctx->mutex);
998 * This must be done under the ctx->lock, such as to serialize against
999 * context_equiv(), therefore we cannot call put_ctx() since that might end up
1000 * calling scheduler related locks and ctx->lock nests inside those.
1002 static __must_check struct perf_event_context *
1003 unclone_ctx(struct perf_event_context *ctx)
1005 struct perf_event_context *parent_ctx = ctx->parent_ctx;
1007 lockdep_assert_held(&ctx->lock);
1010 ctx->parent_ctx = NULL;
1016 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1019 * only top level events have the pid namespace they were created in
1022 event = event->parent;
1024 return task_tgid_nr_ns(p, event->ns);
1027 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1030 * only top level events have the pid namespace they were created in
1033 event = event->parent;
1035 return task_pid_nr_ns(p, event->ns);
1039 * If we inherit events we want to return the parent event id
1042 static u64 primary_event_id(struct perf_event *event)
1047 id = event->parent->id;
1053 * Get the perf_event_context for a task and lock it.
1054 * This has to cope with with the fact that until it is locked,
1055 * the context could get moved to another task.
1057 static struct perf_event_context *
1058 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1060 struct perf_event_context *ctx;
1064 * One of the few rules of preemptible RCU is that one cannot do
1065 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1066 * part of the read side critical section was preemptible -- see
1067 * rcu_read_unlock_special().
1069 * Since ctx->lock nests under rq->lock we must ensure the entire read
1070 * side critical section is non-preemptible.
1074 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1077 * If this context is a clone of another, it might
1078 * get swapped for another underneath us by
1079 * perf_event_task_sched_out, though the
1080 * rcu_read_lock() protects us from any context
1081 * getting freed. Lock the context and check if it
1082 * got swapped before we could get the lock, and retry
1083 * if so. If we locked the right context, then it
1084 * can't get swapped on us any more.
1086 raw_spin_lock_irqsave(&ctx->lock, *flags);
1087 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1088 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1094 if (!atomic_inc_not_zero(&ctx->refcount)) {
1095 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1105 * Get the context for a task and increment its pin_count so it
1106 * can't get swapped to another task. This also increments its
1107 * reference count so that the context can't get freed.
1109 static struct perf_event_context *
1110 perf_pin_task_context(struct task_struct *task, int ctxn)
1112 struct perf_event_context *ctx;
1113 unsigned long flags;
1115 ctx = perf_lock_task_context(task, ctxn, &flags);
1118 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1123 static void perf_unpin_context(struct perf_event_context *ctx)
1125 unsigned long flags;
1127 raw_spin_lock_irqsave(&ctx->lock, flags);
1129 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1133 * Update the record of the current time in a context.
1135 static void update_context_time(struct perf_event_context *ctx)
1137 u64 now = perf_clock();
1139 ctx->time += now - ctx->timestamp;
1140 ctx->timestamp = now;
1143 static u64 perf_event_time(struct perf_event *event)
1145 struct perf_event_context *ctx = event->ctx;
1147 if (is_cgroup_event(event))
1148 return perf_cgroup_event_time(event);
1150 return ctx ? ctx->time : 0;
1154 * Update the total_time_enabled and total_time_running fields for a event.
1155 * The caller of this function needs to hold the ctx->lock.
1157 static void update_event_times(struct perf_event *event)
1159 struct perf_event_context *ctx = event->ctx;
1162 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1163 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1166 * in cgroup mode, time_enabled represents
1167 * the time the event was enabled AND active
1168 * tasks were in the monitored cgroup. This is
1169 * independent of the activity of the context as
1170 * there may be a mix of cgroup and non-cgroup events.
1172 * That is why we treat cgroup events differently
1175 if (is_cgroup_event(event))
1176 run_end = perf_cgroup_event_time(event);
1177 else if (ctx->is_active)
1178 run_end = ctx->time;
1180 run_end = event->tstamp_stopped;
1182 event->total_time_enabled = run_end - event->tstamp_enabled;
1184 if (event->state == PERF_EVENT_STATE_INACTIVE)
1185 run_end = event->tstamp_stopped;
1187 run_end = perf_event_time(event);
1189 event->total_time_running = run_end - event->tstamp_running;
1194 * Update total_time_enabled and total_time_running for all events in a group.
1196 static void update_group_times(struct perf_event *leader)
1198 struct perf_event *event;
1200 update_event_times(leader);
1201 list_for_each_entry(event, &leader->sibling_list, group_entry)
1202 update_event_times(event);
1205 static struct list_head *
1206 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1208 if (event->attr.pinned)
1209 return &ctx->pinned_groups;
1211 return &ctx->flexible_groups;
1215 * Add a event from the lists for its context.
1216 * Must be called with ctx->mutex and ctx->lock held.
1219 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1221 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1222 event->attach_state |= PERF_ATTACH_CONTEXT;
1225 * If we're a stand alone event or group leader, we go to the context
1226 * list, group events are kept attached to the group so that
1227 * perf_group_detach can, at all times, locate all siblings.
1229 if (event->group_leader == event) {
1230 struct list_head *list;
1232 if (is_software_event(event))
1233 event->group_flags |= PERF_GROUP_SOFTWARE;
1235 list = ctx_group_list(event, ctx);
1236 list_add_tail(&event->group_entry, list);
1239 if (is_cgroup_event(event))
1242 if (has_branch_stack(event))
1243 ctx->nr_branch_stack++;
1245 list_add_rcu(&event->event_entry, &ctx->event_list);
1247 if (event->attr.inherit_stat)
1254 * Initialize event state based on the perf_event_attr::disabled.
1256 static inline void perf_event__state_init(struct perf_event *event)
1258 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1259 PERF_EVENT_STATE_INACTIVE;
1263 * Called at perf_event creation and when events are attached/detached from a
1266 static void perf_event__read_size(struct perf_event *event)
1268 int entry = sizeof(u64); /* value */
1272 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1273 size += sizeof(u64);
1275 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1276 size += sizeof(u64);
1278 if (event->attr.read_format & PERF_FORMAT_ID)
1279 entry += sizeof(u64);
1281 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1282 nr += event->group_leader->nr_siblings;
1283 size += sizeof(u64);
1287 event->read_size = size;
1290 static void perf_event__header_size(struct perf_event *event)
1292 struct perf_sample_data *data;
1293 u64 sample_type = event->attr.sample_type;
1296 perf_event__read_size(event);
1298 if (sample_type & PERF_SAMPLE_IP)
1299 size += sizeof(data->ip);
1301 if (sample_type & PERF_SAMPLE_ADDR)
1302 size += sizeof(data->addr);
1304 if (sample_type & PERF_SAMPLE_PERIOD)
1305 size += sizeof(data->period);
1307 if (sample_type & PERF_SAMPLE_WEIGHT)
1308 size += sizeof(data->weight);
1310 if (sample_type & PERF_SAMPLE_READ)
1311 size += event->read_size;
1313 if (sample_type & PERF_SAMPLE_DATA_SRC)
1314 size += sizeof(data->data_src.val);
1316 if (sample_type & PERF_SAMPLE_TRANSACTION)
1317 size += sizeof(data->txn);
1319 event->header_size = size;
1322 static void perf_event__id_header_size(struct perf_event *event)
1324 struct perf_sample_data *data;
1325 u64 sample_type = event->attr.sample_type;
1328 if (sample_type & PERF_SAMPLE_TID)
1329 size += sizeof(data->tid_entry);
1331 if (sample_type & PERF_SAMPLE_TIME)
1332 size += sizeof(data->time);
1334 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1335 size += sizeof(data->id);
1337 if (sample_type & PERF_SAMPLE_ID)
1338 size += sizeof(data->id);
1340 if (sample_type & PERF_SAMPLE_STREAM_ID)
1341 size += sizeof(data->stream_id);
1343 if (sample_type & PERF_SAMPLE_CPU)
1344 size += sizeof(data->cpu_entry);
1346 event->id_header_size = size;
1349 static void perf_group_attach(struct perf_event *event)
1351 struct perf_event *group_leader = event->group_leader, *pos;
1354 * We can have double attach due to group movement in perf_event_open.
1356 if (event->attach_state & PERF_ATTACH_GROUP)
1359 event->attach_state |= PERF_ATTACH_GROUP;
1361 if (group_leader == event)
1364 WARN_ON_ONCE(group_leader->ctx != event->ctx);
1366 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1367 !is_software_event(event))
1368 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1370 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1371 group_leader->nr_siblings++;
1373 perf_event__header_size(group_leader);
1375 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1376 perf_event__header_size(pos);
1380 * Remove a event from the lists for its context.
1381 * Must be called with ctx->mutex and ctx->lock held.
1384 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1386 struct perf_cpu_context *cpuctx;
1388 WARN_ON_ONCE(event->ctx != ctx);
1389 lockdep_assert_held(&ctx->lock);
1392 * We can have double detach due to exit/hot-unplug + close.
1394 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1397 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1399 if (is_cgroup_event(event)) {
1401 cpuctx = __get_cpu_context(ctx);
1403 * if there are no more cgroup events
1404 * then cler cgrp to avoid stale pointer
1405 * in update_cgrp_time_from_cpuctx()
1407 if (!ctx->nr_cgroups)
1408 cpuctx->cgrp = NULL;
1411 if (has_branch_stack(event))
1412 ctx->nr_branch_stack--;
1415 if (event->attr.inherit_stat)
1418 list_del_rcu(&event->event_entry);
1420 if (event->group_leader == event)
1421 list_del_init(&event->group_entry);
1423 update_group_times(event);
1426 * If event was in error state, then keep it
1427 * that way, otherwise bogus counts will be
1428 * returned on read(). The only way to get out
1429 * of error state is by explicit re-enabling
1432 if (event->state > PERF_EVENT_STATE_OFF)
1433 event->state = PERF_EVENT_STATE_OFF;
1438 static void perf_group_detach(struct perf_event *event)
1440 struct perf_event *sibling, *tmp;
1441 struct list_head *list = NULL;
1444 * We can have double detach due to exit/hot-unplug + close.
1446 if (!(event->attach_state & PERF_ATTACH_GROUP))
1449 event->attach_state &= ~PERF_ATTACH_GROUP;
1452 * If this is a sibling, remove it from its group.
1454 if (event->group_leader != event) {
1455 list_del_init(&event->group_entry);
1456 event->group_leader->nr_siblings--;
1460 if (!list_empty(&event->group_entry))
1461 list = &event->group_entry;
1464 * If this was a group event with sibling events then
1465 * upgrade the siblings to singleton events by adding them
1466 * to whatever list we are on.
1468 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1470 list_move_tail(&sibling->group_entry, list);
1471 sibling->group_leader = sibling;
1473 /* Inherit group flags from the previous leader */
1474 sibling->group_flags = event->group_flags;
1476 WARN_ON_ONCE(sibling->ctx != event->ctx);
1480 perf_event__header_size(event->group_leader);
1482 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1483 perf_event__header_size(tmp);
1487 * User event without the task.
1489 static bool is_orphaned_event(struct perf_event *event)
1491 return event && !is_kernel_event(event) && !event->owner;
1495 * Event has a parent but parent's task finished and it's
1496 * alive only because of children holding refference.
1498 static bool is_orphaned_child(struct perf_event *event)
1500 return is_orphaned_event(event->parent);
1503 static void orphans_remove_work(struct work_struct *work);
1505 static void schedule_orphans_remove(struct perf_event_context *ctx)
1507 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1510 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1512 ctx->orphans_remove_sched = true;
1516 static int __init perf_workqueue_init(void)
1518 perf_wq = create_singlethread_workqueue("perf");
1519 WARN(!perf_wq, "failed to create perf workqueue\n");
1520 return perf_wq ? 0 : -1;
1523 core_initcall(perf_workqueue_init);
1526 event_filter_match(struct perf_event *event)
1528 return (event->cpu == -1 || event->cpu == smp_processor_id())
1529 && perf_cgroup_match(event);
1533 event_sched_out(struct perf_event *event,
1534 struct perf_cpu_context *cpuctx,
1535 struct perf_event_context *ctx)
1537 u64 tstamp = perf_event_time(event);
1540 WARN_ON_ONCE(event->ctx != ctx);
1541 lockdep_assert_held(&ctx->lock);
1544 * An event which could not be activated because of
1545 * filter mismatch still needs to have its timings
1546 * maintained, otherwise bogus information is return
1547 * via read() for time_enabled, time_running:
1549 if (event->state == PERF_EVENT_STATE_INACTIVE
1550 && !event_filter_match(event)) {
1551 delta = tstamp - event->tstamp_stopped;
1552 event->tstamp_running += delta;
1553 event->tstamp_stopped = tstamp;
1556 if (event->state != PERF_EVENT_STATE_ACTIVE)
1559 perf_pmu_disable(event->pmu);
1561 event->state = PERF_EVENT_STATE_INACTIVE;
1562 if (event->pending_disable) {
1563 event->pending_disable = 0;
1564 event->state = PERF_EVENT_STATE_OFF;
1566 event->tstamp_stopped = tstamp;
1567 event->pmu->del(event, 0);
1570 if (!is_software_event(event))
1571 cpuctx->active_oncpu--;
1572 if (!--ctx->nr_active)
1573 perf_event_ctx_deactivate(ctx);
1574 if (event->attr.freq && event->attr.sample_freq)
1576 if (event->attr.exclusive || !cpuctx->active_oncpu)
1577 cpuctx->exclusive = 0;
1579 if (is_orphaned_child(event))
1580 schedule_orphans_remove(ctx);
1582 perf_pmu_enable(event->pmu);
1586 group_sched_out(struct perf_event *group_event,
1587 struct perf_cpu_context *cpuctx,
1588 struct perf_event_context *ctx)
1590 struct perf_event *event;
1591 int state = group_event->state;
1593 event_sched_out(group_event, cpuctx, ctx);
1596 * Schedule out siblings (if any):
1598 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1599 event_sched_out(event, cpuctx, ctx);
1601 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1602 cpuctx->exclusive = 0;
1605 struct remove_event {
1606 struct perf_event *event;
1611 * Cross CPU call to remove a performance event
1613 * We disable the event on the hardware level first. After that we
1614 * remove it from the context list.
1616 static int __perf_remove_from_context(void *info)
1618 struct remove_event *re = info;
1619 struct perf_event *event = re->event;
1620 struct perf_event_context *ctx = event->ctx;
1621 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1623 raw_spin_lock(&ctx->lock);
1624 event_sched_out(event, cpuctx, ctx);
1625 if (re->detach_group)
1626 perf_group_detach(event);
1627 list_del_event(event, ctx);
1628 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1630 cpuctx->task_ctx = NULL;
1632 raw_spin_unlock(&ctx->lock);
1639 * Remove the event from a task's (or a CPU's) list of events.
1641 * CPU events are removed with a smp call. For task events we only
1642 * call when the task is on a CPU.
1644 * If event->ctx is a cloned context, callers must make sure that
1645 * every task struct that event->ctx->task could possibly point to
1646 * remains valid. This is OK when called from perf_release since
1647 * that only calls us on the top-level context, which can't be a clone.
1648 * When called from perf_event_exit_task, it's OK because the
1649 * context has been detached from its task.
1651 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1653 struct perf_event_context *ctx = event->ctx;
1654 struct task_struct *task = ctx->task;
1655 struct remove_event re = {
1657 .detach_group = detach_group,
1660 lockdep_assert_held(&ctx->mutex);
1664 * Per cpu events are removed via an smp call. The removal can
1665 * fail if the CPU is currently offline, but in that case we
1666 * already called __perf_remove_from_context from
1667 * perf_event_exit_cpu.
1669 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1674 if (!task_function_call(task, __perf_remove_from_context, &re))
1677 raw_spin_lock_irq(&ctx->lock);
1679 * If we failed to find a running task, but find the context active now
1680 * that we've acquired the ctx->lock, retry.
1682 if (ctx->is_active) {
1683 raw_spin_unlock_irq(&ctx->lock);
1685 * Reload the task pointer, it might have been changed by
1686 * a concurrent perf_event_context_sched_out().
1693 * Since the task isn't running, its safe to remove the event, us
1694 * holding the ctx->lock ensures the task won't get scheduled in.
1697 perf_group_detach(event);
1698 list_del_event(event, ctx);
1699 raw_spin_unlock_irq(&ctx->lock);
1703 * Cross CPU call to disable a performance event
1705 int __perf_event_disable(void *info)
1707 struct perf_event *event = info;
1708 struct perf_event_context *ctx = event->ctx;
1709 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1712 * If this is a per-task event, need to check whether this
1713 * event's task is the current task on this cpu.
1715 * Can trigger due to concurrent perf_event_context_sched_out()
1716 * flipping contexts around.
1718 if (ctx->task && cpuctx->task_ctx != ctx)
1721 raw_spin_lock(&ctx->lock);
1724 * If the event is on, turn it off.
1725 * If it is in error state, leave it in error state.
1727 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1728 update_context_time(ctx);
1729 update_cgrp_time_from_event(event);
1730 update_group_times(event);
1731 if (event == event->group_leader)
1732 group_sched_out(event, cpuctx, ctx);
1734 event_sched_out(event, cpuctx, ctx);
1735 event->state = PERF_EVENT_STATE_OFF;
1738 raw_spin_unlock(&ctx->lock);
1746 * If event->ctx is a cloned context, callers must make sure that
1747 * every task struct that event->ctx->task could possibly point to
1748 * remains valid. This condition is satisifed when called through
1749 * perf_event_for_each_child or perf_event_for_each because they
1750 * hold the top-level event's child_mutex, so any descendant that
1751 * goes to exit will block in sync_child_event.
1752 * When called from perf_pending_event it's OK because event->ctx
1753 * is the current context on this CPU and preemption is disabled,
1754 * hence we can't get into perf_event_task_sched_out for this context.
1756 static void _perf_event_disable(struct perf_event *event)
1758 struct perf_event_context *ctx = event->ctx;
1759 struct task_struct *task = ctx->task;
1763 * Disable the event on the cpu that it's on
1765 cpu_function_call(event->cpu, __perf_event_disable, event);
1770 if (!task_function_call(task, __perf_event_disable, event))
1773 raw_spin_lock_irq(&ctx->lock);
1775 * If the event is still active, we need to retry the cross-call.
1777 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1778 raw_spin_unlock_irq(&ctx->lock);
1780 * Reload the task pointer, it might have been changed by
1781 * a concurrent perf_event_context_sched_out().
1788 * Since we have the lock this context can't be scheduled
1789 * in, so we can change the state safely.
1791 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1792 update_group_times(event);
1793 event->state = PERF_EVENT_STATE_OFF;
1795 raw_spin_unlock_irq(&ctx->lock);
1799 * Strictly speaking kernel users cannot create groups and therefore this
1800 * interface does not need the perf_event_ctx_lock() magic.
1802 void perf_event_disable(struct perf_event *event)
1804 struct perf_event_context *ctx;
1806 ctx = perf_event_ctx_lock(event);
1807 _perf_event_disable(event);
1808 perf_event_ctx_unlock(event, ctx);
1810 EXPORT_SYMBOL_GPL(perf_event_disable);
1812 static void perf_set_shadow_time(struct perf_event *event,
1813 struct perf_event_context *ctx,
1817 * use the correct time source for the time snapshot
1819 * We could get by without this by leveraging the
1820 * fact that to get to this function, the caller
1821 * has most likely already called update_context_time()
1822 * and update_cgrp_time_xx() and thus both timestamp
1823 * are identical (or very close). Given that tstamp is,
1824 * already adjusted for cgroup, we could say that:
1825 * tstamp - ctx->timestamp
1827 * tstamp - cgrp->timestamp.
1829 * Then, in perf_output_read(), the calculation would
1830 * work with no changes because:
1831 * - event is guaranteed scheduled in
1832 * - no scheduled out in between
1833 * - thus the timestamp would be the same
1835 * But this is a bit hairy.
1837 * So instead, we have an explicit cgroup call to remain
1838 * within the time time source all along. We believe it
1839 * is cleaner and simpler to understand.
1841 if (is_cgroup_event(event))
1842 perf_cgroup_set_shadow_time(event, tstamp);
1844 event->shadow_ctx_time = tstamp - ctx->timestamp;
1847 #define MAX_INTERRUPTS (~0ULL)
1849 static void perf_log_throttle(struct perf_event *event, int enable);
1852 event_sched_in(struct perf_event *event,
1853 struct perf_cpu_context *cpuctx,
1854 struct perf_event_context *ctx)
1856 u64 tstamp = perf_event_time(event);
1859 lockdep_assert_held(&ctx->lock);
1861 if (event->state <= PERF_EVENT_STATE_OFF)
1864 event->state = PERF_EVENT_STATE_ACTIVE;
1865 event->oncpu = smp_processor_id();
1868 * Unthrottle events, since we scheduled we might have missed several
1869 * ticks already, also for a heavily scheduling task there is little
1870 * guarantee it'll get a tick in a timely manner.
1872 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1873 perf_log_throttle(event, 1);
1874 event->hw.interrupts = 0;
1878 * The new state must be visible before we turn it on in the hardware:
1882 perf_pmu_disable(event->pmu);
1884 if (event->pmu->add(event, PERF_EF_START)) {
1885 event->state = PERF_EVENT_STATE_INACTIVE;
1891 event->tstamp_running += tstamp - event->tstamp_stopped;
1893 perf_set_shadow_time(event, ctx, tstamp);
1895 if (!is_software_event(event))
1896 cpuctx->active_oncpu++;
1897 if (!ctx->nr_active++)
1898 perf_event_ctx_activate(ctx);
1899 if (event->attr.freq && event->attr.sample_freq)
1902 if (event->attr.exclusive)
1903 cpuctx->exclusive = 1;
1905 if (is_orphaned_child(event))
1906 schedule_orphans_remove(ctx);
1909 perf_pmu_enable(event->pmu);
1915 group_sched_in(struct perf_event *group_event,
1916 struct perf_cpu_context *cpuctx,
1917 struct perf_event_context *ctx)
1919 struct perf_event *event, *partial_group = NULL;
1920 struct pmu *pmu = ctx->pmu;
1921 u64 now = ctx->time;
1922 bool simulate = false;
1924 if (group_event->state == PERF_EVENT_STATE_OFF)
1927 pmu->start_txn(pmu);
1929 if (event_sched_in(group_event, cpuctx, ctx)) {
1930 pmu->cancel_txn(pmu);
1931 perf_cpu_hrtimer_restart(cpuctx);
1936 * Schedule in siblings as one group (if any):
1938 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1939 if (event_sched_in(event, cpuctx, ctx)) {
1940 partial_group = event;
1945 if (!pmu->commit_txn(pmu))
1950 * Groups can be scheduled in as one unit only, so undo any
1951 * partial group before returning:
1952 * The events up to the failed event are scheduled out normally,
1953 * tstamp_stopped will be updated.
1955 * The failed events and the remaining siblings need to have
1956 * their timings updated as if they had gone thru event_sched_in()
1957 * and event_sched_out(). This is required to get consistent timings
1958 * across the group. This also takes care of the case where the group
1959 * could never be scheduled by ensuring tstamp_stopped is set to mark
1960 * the time the event was actually stopped, such that time delta
1961 * calculation in update_event_times() is correct.
1963 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1964 if (event == partial_group)
1968 event->tstamp_running += now - event->tstamp_stopped;
1969 event->tstamp_stopped = now;
1971 event_sched_out(event, cpuctx, ctx);
1974 event_sched_out(group_event, cpuctx, ctx);
1976 pmu->cancel_txn(pmu);
1978 perf_cpu_hrtimer_restart(cpuctx);
1984 * Work out whether we can put this event group on the CPU now.
1986 static int group_can_go_on(struct perf_event *event,
1987 struct perf_cpu_context *cpuctx,
1991 * Groups consisting entirely of software events can always go on.
1993 if (event->group_flags & PERF_GROUP_SOFTWARE)
1996 * If an exclusive group is already on, no other hardware
1999 if (cpuctx->exclusive)
2002 * If this group is exclusive and there are already
2003 * events on the CPU, it can't go on.
2005 if (event->attr.exclusive && cpuctx->active_oncpu)
2008 * Otherwise, try to add it if all previous groups were able
2014 static void add_event_to_ctx(struct perf_event *event,
2015 struct perf_event_context *ctx)
2017 u64 tstamp = perf_event_time(event);
2019 list_add_event(event, ctx);
2020 perf_group_attach(event);
2021 event->tstamp_enabled = tstamp;
2022 event->tstamp_running = tstamp;
2023 event->tstamp_stopped = tstamp;
2026 static void task_ctx_sched_out(struct perf_event_context *ctx);
2028 ctx_sched_in(struct perf_event_context *ctx,
2029 struct perf_cpu_context *cpuctx,
2030 enum event_type_t event_type,
2031 struct task_struct *task);
2033 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2034 struct perf_event_context *ctx,
2035 struct task_struct *task)
2037 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2039 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2040 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2042 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2046 * Cross CPU call to install and enable a performance event
2048 * Must be called with ctx->mutex held
2050 static int __perf_install_in_context(void *info)
2052 struct perf_event *event = info;
2053 struct perf_event_context *ctx = event->ctx;
2054 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2055 struct perf_event_context *task_ctx = cpuctx->task_ctx;
2056 struct task_struct *task = current;
2058 perf_ctx_lock(cpuctx, task_ctx);
2059 perf_pmu_disable(cpuctx->ctx.pmu);
2062 * If there was an active task_ctx schedule it out.
2065 task_ctx_sched_out(task_ctx);
2068 * If the context we're installing events in is not the
2069 * active task_ctx, flip them.
2071 if (ctx->task && task_ctx != ctx) {
2073 raw_spin_unlock(&task_ctx->lock);
2074 raw_spin_lock(&ctx->lock);
2079 cpuctx->task_ctx = task_ctx;
2080 task = task_ctx->task;
2083 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
2085 update_context_time(ctx);
2087 * update cgrp time only if current cgrp
2088 * matches event->cgrp. Must be done before
2089 * calling add_event_to_ctx()
2091 update_cgrp_time_from_event(event);
2093 add_event_to_ctx(event, ctx);
2096 * Schedule everything back in
2098 perf_event_sched_in(cpuctx, task_ctx, task);
2100 perf_pmu_enable(cpuctx->ctx.pmu);
2101 perf_ctx_unlock(cpuctx, task_ctx);
2107 * Attach a performance event to a context
2109 * First we add the event to the list with the hardware enable bit
2110 * in event->hw_config cleared.
2112 * If the event is attached to a task which is on a CPU we use a smp
2113 * call to enable it in the task context. The task might have been
2114 * scheduled away, but we check this in the smp call again.
2117 perf_install_in_context(struct perf_event_context *ctx,
2118 struct perf_event *event,
2121 struct task_struct *task = ctx->task;
2123 lockdep_assert_held(&ctx->mutex);
2126 if (event->cpu != -1)
2131 * Per cpu events are installed via an smp call and
2132 * the install is always successful.
2134 cpu_function_call(cpu, __perf_install_in_context, event);
2139 if (!task_function_call(task, __perf_install_in_context, event))
2142 raw_spin_lock_irq(&ctx->lock);
2144 * If we failed to find a running task, but find the context active now
2145 * that we've acquired the ctx->lock, retry.
2147 if (ctx->is_active) {
2148 raw_spin_unlock_irq(&ctx->lock);
2150 * Reload the task pointer, it might have been changed by
2151 * a concurrent perf_event_context_sched_out().
2158 * Since the task isn't running, its safe to add the event, us holding
2159 * the ctx->lock ensures the task won't get scheduled in.
2161 add_event_to_ctx(event, ctx);
2162 raw_spin_unlock_irq(&ctx->lock);
2166 * Put a event into inactive state and update time fields.
2167 * Enabling the leader of a group effectively enables all
2168 * the group members that aren't explicitly disabled, so we
2169 * have to update their ->tstamp_enabled also.
2170 * Note: this works for group members as well as group leaders
2171 * since the non-leader members' sibling_lists will be empty.
2173 static void __perf_event_mark_enabled(struct perf_event *event)
2175 struct perf_event *sub;
2176 u64 tstamp = perf_event_time(event);
2178 event->state = PERF_EVENT_STATE_INACTIVE;
2179 event->tstamp_enabled = tstamp - event->total_time_enabled;
2180 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2181 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2182 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2187 * Cross CPU call to enable a performance event
2189 static int __perf_event_enable(void *info)
2191 struct perf_event *event = info;
2192 struct perf_event_context *ctx = event->ctx;
2193 struct perf_event *leader = event->group_leader;
2194 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2198 * There's a time window between 'ctx->is_active' check
2199 * in perf_event_enable function and this place having:
2201 * - ctx->lock unlocked
2203 * where the task could be killed and 'ctx' deactivated
2204 * by perf_event_exit_task.
2206 if (!ctx->is_active)
2209 raw_spin_lock(&ctx->lock);
2210 update_context_time(ctx);
2212 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2216 * set current task's cgroup time reference point
2218 perf_cgroup_set_timestamp(current, ctx);
2220 __perf_event_mark_enabled(event);
2222 if (!event_filter_match(event)) {
2223 if (is_cgroup_event(event))
2224 perf_cgroup_defer_enabled(event);
2229 * If the event is in a group and isn't the group leader,
2230 * then don't put it on unless the group is on.
2232 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2235 if (!group_can_go_on(event, cpuctx, 1)) {
2238 if (event == leader)
2239 err = group_sched_in(event, cpuctx, ctx);
2241 err = event_sched_in(event, cpuctx, ctx);
2246 * If this event can't go on and it's part of a
2247 * group, then the whole group has to come off.
2249 if (leader != event) {
2250 group_sched_out(leader, cpuctx, ctx);
2251 perf_cpu_hrtimer_restart(cpuctx);
2253 if (leader->attr.pinned) {
2254 update_group_times(leader);
2255 leader->state = PERF_EVENT_STATE_ERROR;
2260 raw_spin_unlock(&ctx->lock);
2268 * If event->ctx is a cloned context, callers must make sure that
2269 * every task struct that event->ctx->task could possibly point to
2270 * remains valid. This condition is satisfied when called through
2271 * perf_event_for_each_child or perf_event_for_each as described
2272 * for perf_event_disable.
2274 static void _perf_event_enable(struct perf_event *event)
2276 struct perf_event_context *ctx = event->ctx;
2277 struct task_struct *task = ctx->task;
2281 * Enable the event on the cpu that it's on
2283 cpu_function_call(event->cpu, __perf_event_enable, event);
2287 raw_spin_lock_irq(&ctx->lock);
2288 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2292 * If the event is in error state, clear that first.
2293 * That way, if we see the event in error state below, we
2294 * know that it has gone back into error state, as distinct
2295 * from the task having been scheduled away before the
2296 * cross-call arrived.
2298 if (event->state == PERF_EVENT_STATE_ERROR)
2299 event->state = PERF_EVENT_STATE_OFF;
2302 if (!ctx->is_active) {
2303 __perf_event_mark_enabled(event);
2307 raw_spin_unlock_irq(&ctx->lock);
2309 if (!task_function_call(task, __perf_event_enable, event))
2312 raw_spin_lock_irq(&ctx->lock);
2315 * If the context is active and the event is still off,
2316 * we need to retry the cross-call.
2318 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2320 * task could have been flipped by a concurrent
2321 * perf_event_context_sched_out()
2328 raw_spin_unlock_irq(&ctx->lock);
2332 * See perf_event_disable();
2334 void perf_event_enable(struct perf_event *event)
2336 struct perf_event_context *ctx;
2338 ctx = perf_event_ctx_lock(event);
2339 _perf_event_enable(event);
2340 perf_event_ctx_unlock(event, ctx);
2342 EXPORT_SYMBOL_GPL(perf_event_enable);
2344 static int _perf_event_refresh(struct perf_event *event, int refresh)
2347 * not supported on inherited events
2349 if (event->attr.inherit || !is_sampling_event(event))
2352 atomic_add(refresh, &event->event_limit);
2353 _perf_event_enable(event);
2359 * See perf_event_disable()
2361 int perf_event_refresh(struct perf_event *event, int refresh)
2363 struct perf_event_context *ctx;
2366 ctx = perf_event_ctx_lock(event);
2367 ret = _perf_event_refresh(event, refresh);
2368 perf_event_ctx_unlock(event, ctx);
2372 EXPORT_SYMBOL_GPL(perf_event_refresh);
2374 static void ctx_sched_out(struct perf_event_context *ctx,
2375 struct perf_cpu_context *cpuctx,
2376 enum event_type_t event_type)
2378 struct perf_event *event;
2379 int is_active = ctx->is_active;
2381 ctx->is_active &= ~event_type;
2382 if (likely(!ctx->nr_events))
2385 update_context_time(ctx);
2386 update_cgrp_time_from_cpuctx(cpuctx);
2387 if (!ctx->nr_active)
2390 perf_pmu_disable(ctx->pmu);
2391 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2392 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2393 group_sched_out(event, cpuctx, ctx);
2396 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2397 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2398 group_sched_out(event, cpuctx, ctx);
2400 perf_pmu_enable(ctx->pmu);
2404 * Test whether two contexts are equivalent, i.e. whether they have both been
2405 * cloned from the same version of the same context.
2407 * Equivalence is measured using a generation number in the context that is
2408 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2409 * and list_del_event().
2411 static int context_equiv(struct perf_event_context *ctx1,
2412 struct perf_event_context *ctx2)
2414 lockdep_assert_held(&ctx1->lock);
2415 lockdep_assert_held(&ctx2->lock);
2417 /* Pinning disables the swap optimization */
2418 if (ctx1->pin_count || ctx2->pin_count)
2421 /* If ctx1 is the parent of ctx2 */
2422 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2425 /* If ctx2 is the parent of ctx1 */
2426 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2430 * If ctx1 and ctx2 have the same parent; we flatten the parent
2431 * hierarchy, see perf_event_init_context().
2433 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2434 ctx1->parent_gen == ctx2->parent_gen)
2441 static void __perf_event_sync_stat(struct perf_event *event,
2442 struct perf_event *next_event)
2446 if (!event->attr.inherit_stat)
2450 * Update the event value, we cannot use perf_event_read()
2451 * because we're in the middle of a context switch and have IRQs
2452 * disabled, which upsets smp_call_function_single(), however
2453 * we know the event must be on the current CPU, therefore we
2454 * don't need to use it.
2456 switch (event->state) {
2457 case PERF_EVENT_STATE_ACTIVE:
2458 event->pmu->read(event);
2461 case PERF_EVENT_STATE_INACTIVE:
2462 update_event_times(event);
2470 * In order to keep per-task stats reliable we need to flip the event
2471 * values when we flip the contexts.
2473 value = local64_read(&next_event->count);
2474 value = local64_xchg(&event->count, value);
2475 local64_set(&next_event->count, value);
2477 swap(event->total_time_enabled, next_event->total_time_enabled);
2478 swap(event->total_time_running, next_event->total_time_running);
2481 * Since we swizzled the values, update the user visible data too.
2483 perf_event_update_userpage(event);
2484 perf_event_update_userpage(next_event);
2487 static void perf_event_sync_stat(struct perf_event_context *ctx,
2488 struct perf_event_context *next_ctx)
2490 struct perf_event *event, *next_event;
2495 update_context_time(ctx);
2497 event = list_first_entry(&ctx->event_list,
2498 struct perf_event, event_entry);
2500 next_event = list_first_entry(&next_ctx->event_list,
2501 struct perf_event, event_entry);
2503 while (&event->event_entry != &ctx->event_list &&
2504 &next_event->event_entry != &next_ctx->event_list) {
2506 __perf_event_sync_stat(event, next_event);
2508 event = list_next_entry(event, event_entry);
2509 next_event = list_next_entry(next_event, event_entry);
2513 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2514 struct task_struct *next)
2516 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2517 struct perf_event_context *next_ctx;
2518 struct perf_event_context *parent, *next_parent;
2519 struct perf_cpu_context *cpuctx;
2525 cpuctx = __get_cpu_context(ctx);
2526 if (!cpuctx->task_ctx)
2530 next_ctx = next->perf_event_ctxp[ctxn];
2534 parent = rcu_dereference(ctx->parent_ctx);
2535 next_parent = rcu_dereference(next_ctx->parent_ctx);
2537 /* If neither context have a parent context; they cannot be clones. */
2538 if (!parent && !next_parent)
2541 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2543 * Looks like the two contexts are clones, so we might be
2544 * able to optimize the context switch. We lock both
2545 * contexts and check that they are clones under the
2546 * lock (including re-checking that neither has been
2547 * uncloned in the meantime). It doesn't matter which
2548 * order we take the locks because no other cpu could
2549 * be trying to lock both of these tasks.
2551 raw_spin_lock(&ctx->lock);
2552 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2553 if (context_equiv(ctx, next_ctx)) {
2555 * XXX do we need a memory barrier of sorts
2556 * wrt to rcu_dereference() of perf_event_ctxp
2558 task->perf_event_ctxp[ctxn] = next_ctx;
2559 next->perf_event_ctxp[ctxn] = ctx;
2561 next_ctx->task = task;
2564 perf_event_sync_stat(ctx, next_ctx);
2566 raw_spin_unlock(&next_ctx->lock);
2567 raw_spin_unlock(&ctx->lock);
2573 raw_spin_lock(&ctx->lock);
2574 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2575 cpuctx->task_ctx = NULL;
2576 raw_spin_unlock(&ctx->lock);
2580 #define for_each_task_context_nr(ctxn) \
2581 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2584 * Called from scheduler to remove the events of the current task,
2585 * with interrupts disabled.
2587 * We stop each event and update the event value in event->count.
2589 * This does not protect us against NMI, but disable()
2590 * sets the disabled bit in the control field of event _before_
2591 * accessing the event control register. If a NMI hits, then it will
2592 * not restart the event.
2594 void __perf_event_task_sched_out(struct task_struct *task,
2595 struct task_struct *next)
2599 for_each_task_context_nr(ctxn)
2600 perf_event_context_sched_out(task, ctxn, next);
2603 * if cgroup events exist on this CPU, then we need
2604 * to check if we have to switch out PMU state.
2605 * cgroup event are system-wide mode only
2607 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2608 perf_cgroup_sched_out(task, next);
2611 static void task_ctx_sched_out(struct perf_event_context *ctx)
2613 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2615 if (!cpuctx->task_ctx)
2618 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2621 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2622 cpuctx->task_ctx = NULL;
2626 * Called with IRQs disabled
2628 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2629 enum event_type_t event_type)
2631 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2635 ctx_pinned_sched_in(struct perf_event_context *ctx,
2636 struct perf_cpu_context *cpuctx)
2638 struct perf_event *event;
2640 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2641 if (event->state <= PERF_EVENT_STATE_OFF)
2643 if (!event_filter_match(event))
2646 /* may need to reset tstamp_enabled */
2647 if (is_cgroup_event(event))
2648 perf_cgroup_mark_enabled(event, ctx);
2650 if (group_can_go_on(event, cpuctx, 1))
2651 group_sched_in(event, cpuctx, ctx);
2654 * If this pinned group hasn't been scheduled,
2655 * put it in error state.
2657 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2658 update_group_times(event);
2659 event->state = PERF_EVENT_STATE_ERROR;
2665 ctx_flexible_sched_in(struct perf_event_context *ctx,
2666 struct perf_cpu_context *cpuctx)
2668 struct perf_event *event;
2671 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2672 /* Ignore events in OFF or ERROR state */
2673 if (event->state <= PERF_EVENT_STATE_OFF)
2676 * Listen to the 'cpu' scheduling filter constraint
2679 if (!event_filter_match(event))
2682 /* may need to reset tstamp_enabled */
2683 if (is_cgroup_event(event))
2684 perf_cgroup_mark_enabled(event, ctx);
2686 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2687 if (group_sched_in(event, cpuctx, ctx))
2694 ctx_sched_in(struct perf_event_context *ctx,
2695 struct perf_cpu_context *cpuctx,
2696 enum event_type_t event_type,
2697 struct task_struct *task)
2700 int is_active = ctx->is_active;
2702 ctx->is_active |= event_type;
2703 if (likely(!ctx->nr_events))
2707 ctx->timestamp = now;
2708 perf_cgroup_set_timestamp(task, ctx);
2710 * First go through the list and put on any pinned groups
2711 * in order to give them the best chance of going on.
2713 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2714 ctx_pinned_sched_in(ctx, cpuctx);
2716 /* Then walk through the lower prio flexible groups */
2717 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2718 ctx_flexible_sched_in(ctx, cpuctx);
2721 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2722 enum event_type_t event_type,
2723 struct task_struct *task)
2725 struct perf_event_context *ctx = &cpuctx->ctx;
2727 ctx_sched_in(ctx, cpuctx, event_type, task);
2730 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2731 struct task_struct *task)
2733 struct perf_cpu_context *cpuctx;
2735 cpuctx = __get_cpu_context(ctx);
2736 if (cpuctx->task_ctx == ctx)
2739 perf_ctx_lock(cpuctx, ctx);
2740 perf_pmu_disable(ctx->pmu);
2742 * We want to keep the following priority order:
2743 * cpu pinned (that don't need to move), task pinned,
2744 * cpu flexible, task flexible.
2746 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2749 cpuctx->task_ctx = ctx;
2751 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2753 perf_pmu_enable(ctx->pmu);
2754 perf_ctx_unlock(cpuctx, ctx);
2758 * When sampling the branck stack in system-wide, it may be necessary
2759 * to flush the stack on context switch. This happens when the branch
2760 * stack does not tag its entries with the pid of the current task.
2761 * Otherwise it becomes impossible to associate a branch entry with a
2762 * task. This ambiguity is more likely to appear when the branch stack
2763 * supports priv level filtering and the user sets it to monitor only
2764 * at the user level (which could be a useful measurement in system-wide
2765 * mode). In that case, the risk is high of having a branch stack with
2766 * branch from multiple tasks. Flushing may mean dropping the existing
2767 * entries or stashing them somewhere in the PMU specific code layer.
2769 * This function provides the context switch callback to the lower code
2770 * layer. It is invoked ONLY when there is at least one system-wide context
2771 * with at least one active event using taken branch sampling.
2773 static void perf_branch_stack_sched_in(struct task_struct *prev,
2774 struct task_struct *task)
2776 struct perf_cpu_context *cpuctx;
2778 unsigned long flags;
2780 /* no need to flush branch stack if not changing task */
2784 local_irq_save(flags);
2788 list_for_each_entry_rcu(pmu, &pmus, entry) {
2789 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2792 * check if the context has at least one
2793 * event using PERF_SAMPLE_BRANCH_STACK
2795 if (cpuctx->ctx.nr_branch_stack > 0
2796 && pmu->flush_branch_stack) {
2798 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2800 perf_pmu_disable(pmu);
2802 pmu->flush_branch_stack();
2804 perf_pmu_enable(pmu);
2806 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2812 local_irq_restore(flags);
2816 * Called from scheduler to add the events of the current task
2817 * with interrupts disabled.
2819 * We restore the event value and then enable it.
2821 * This does not protect us against NMI, but enable()
2822 * sets the enabled bit in the control field of event _before_
2823 * accessing the event control register. If a NMI hits, then it will
2824 * keep the event running.
2826 void __perf_event_task_sched_in(struct task_struct *prev,
2827 struct task_struct *task)
2829 struct perf_event_context *ctx;
2832 for_each_task_context_nr(ctxn) {
2833 ctx = task->perf_event_ctxp[ctxn];
2837 perf_event_context_sched_in(ctx, task);
2840 * if cgroup events exist on this CPU, then we need
2841 * to check if we have to switch in PMU state.
2842 * cgroup event are system-wide mode only
2844 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2845 perf_cgroup_sched_in(prev, task);
2847 /* check for system-wide branch_stack events */
2848 if (atomic_read(this_cpu_ptr(&perf_branch_stack_events)))
2849 perf_branch_stack_sched_in(prev, task);
2852 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2854 u64 frequency = event->attr.sample_freq;
2855 u64 sec = NSEC_PER_SEC;
2856 u64 divisor, dividend;
2858 int count_fls, nsec_fls, frequency_fls, sec_fls;
2860 count_fls = fls64(count);
2861 nsec_fls = fls64(nsec);
2862 frequency_fls = fls64(frequency);
2866 * We got @count in @nsec, with a target of sample_freq HZ
2867 * the target period becomes:
2870 * period = -------------------
2871 * @nsec * sample_freq
2876 * Reduce accuracy by one bit such that @a and @b converge
2877 * to a similar magnitude.
2879 #define REDUCE_FLS(a, b) \
2881 if (a##_fls > b##_fls) { \
2891 * Reduce accuracy until either term fits in a u64, then proceed with
2892 * the other, so that finally we can do a u64/u64 division.
2894 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2895 REDUCE_FLS(nsec, frequency);
2896 REDUCE_FLS(sec, count);
2899 if (count_fls + sec_fls > 64) {
2900 divisor = nsec * frequency;
2902 while (count_fls + sec_fls > 64) {
2903 REDUCE_FLS(count, sec);
2907 dividend = count * sec;
2909 dividend = count * sec;
2911 while (nsec_fls + frequency_fls > 64) {
2912 REDUCE_FLS(nsec, frequency);
2916 divisor = nsec * frequency;
2922 return div64_u64(dividend, divisor);
2925 static DEFINE_PER_CPU(int, perf_throttled_count);
2926 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2928 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2930 struct hw_perf_event *hwc = &event->hw;
2931 s64 period, sample_period;
2934 period = perf_calculate_period(event, nsec, count);
2936 delta = (s64)(period - hwc->sample_period);
2937 delta = (delta + 7) / 8; /* low pass filter */
2939 sample_period = hwc->sample_period + delta;
2944 hwc->sample_period = sample_period;
2946 if (local64_read(&hwc->period_left) > 8*sample_period) {
2948 event->pmu->stop(event, PERF_EF_UPDATE);
2950 local64_set(&hwc->period_left, 0);
2953 event->pmu->start(event, PERF_EF_RELOAD);
2958 * combine freq adjustment with unthrottling to avoid two passes over the
2959 * events. At the same time, make sure, having freq events does not change
2960 * the rate of unthrottling as that would introduce bias.
2962 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2965 struct perf_event *event;
2966 struct hw_perf_event *hwc;
2967 u64 now, period = TICK_NSEC;
2971 * only need to iterate over all events iff:
2972 * - context have events in frequency mode (needs freq adjust)
2973 * - there are events to unthrottle on this cpu
2975 if (!(ctx->nr_freq || needs_unthr))
2978 raw_spin_lock(&ctx->lock);
2979 perf_pmu_disable(ctx->pmu);
2981 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2982 if (event->state != PERF_EVENT_STATE_ACTIVE)
2985 if (!event_filter_match(event))
2988 perf_pmu_disable(event->pmu);
2992 if (hwc->interrupts == MAX_INTERRUPTS) {
2993 hwc->interrupts = 0;
2994 perf_log_throttle(event, 1);
2995 event->pmu->start(event, 0);
2998 if (!event->attr.freq || !event->attr.sample_freq)
3002 * stop the event and update event->count
3004 event->pmu->stop(event, PERF_EF_UPDATE);
3006 now = local64_read(&event->count);
3007 delta = now - hwc->freq_count_stamp;
3008 hwc->freq_count_stamp = now;
3012 * reload only if value has changed
3013 * we have stopped the event so tell that
3014 * to perf_adjust_period() to avoid stopping it
3018 perf_adjust_period(event, period, delta, false);
3020 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3022 perf_pmu_enable(event->pmu);
3025 perf_pmu_enable(ctx->pmu);
3026 raw_spin_unlock(&ctx->lock);
3030 * Round-robin a context's events:
3032 static void rotate_ctx(struct perf_event_context *ctx)
3035 * Rotate the first entry last of non-pinned groups. Rotation might be
3036 * disabled by the inheritance code.
3038 if (!ctx->rotate_disable)
3039 list_rotate_left(&ctx->flexible_groups);
3042 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
3044 struct perf_event_context *ctx = NULL;
3047 if (cpuctx->ctx.nr_events) {
3048 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3052 ctx = cpuctx->task_ctx;
3053 if (ctx && ctx->nr_events) {
3054 if (ctx->nr_events != ctx->nr_active)
3061 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3062 perf_pmu_disable(cpuctx->ctx.pmu);
3064 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3066 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3068 rotate_ctx(&cpuctx->ctx);
3072 perf_event_sched_in(cpuctx, ctx, current);
3074 perf_pmu_enable(cpuctx->ctx.pmu);
3075 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3081 #ifdef CONFIG_NO_HZ_FULL
3082 bool perf_event_can_stop_tick(void)
3084 if (atomic_read(&nr_freq_events) ||
3085 __this_cpu_read(perf_throttled_count))
3092 void perf_event_task_tick(void)
3094 struct list_head *head = this_cpu_ptr(&active_ctx_list);
3095 struct perf_event_context *ctx, *tmp;
3098 WARN_ON(!irqs_disabled());
3100 __this_cpu_inc(perf_throttled_seq);
3101 throttled = __this_cpu_xchg(perf_throttled_count, 0);
3103 list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3104 perf_adjust_freq_unthr_context(ctx, throttled);
3107 static int event_enable_on_exec(struct perf_event *event,
3108 struct perf_event_context *ctx)
3110 if (!event->attr.enable_on_exec)
3113 event->attr.enable_on_exec = 0;
3114 if (event->state >= PERF_EVENT_STATE_INACTIVE)
3117 __perf_event_mark_enabled(event);
3123 * Enable all of a task's events that have been marked enable-on-exec.
3124 * This expects task == current.
3126 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3128 struct perf_event_context *clone_ctx = NULL;
3129 struct perf_event *event;
3130 unsigned long flags;
3134 local_irq_save(flags);
3135 if (!ctx || !ctx->nr_events)
3139 * We must ctxsw out cgroup events to avoid conflict
3140 * when invoking perf_task_event_sched_in() later on
3141 * in this function. Otherwise we end up trying to
3142 * ctxswin cgroup events which are already scheduled
3145 perf_cgroup_sched_out(current, NULL);
3147 raw_spin_lock(&ctx->lock);
3148 task_ctx_sched_out(ctx);
3150 list_for_each_entry(event, &ctx->event_list, event_entry) {
3151 ret = event_enable_on_exec(event, ctx);
3157 * Unclone this context if we enabled any event.
3160 clone_ctx = unclone_ctx(ctx);
3162 raw_spin_unlock(&ctx->lock);
3165 * Also calls ctxswin for cgroup events, if any:
3167 perf_event_context_sched_in(ctx, ctx->task);
3169 local_irq_restore(flags);
3175 void perf_event_exec(void)
3177 struct perf_event_context *ctx;
3181 for_each_task_context_nr(ctxn) {
3182 ctx = current->perf_event_ctxp[ctxn];
3186 perf_event_enable_on_exec(ctx);
3192 * Cross CPU call to read the hardware event
3194 static void __perf_event_read(void *info)
3196 struct perf_event *event = info;
3197 struct perf_event_context *ctx = event->ctx;
3198 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3201 * If this is a task context, we need to check whether it is
3202 * the current task context of this cpu. If not it has been
3203 * scheduled out before the smp call arrived. In that case
3204 * event->count would have been updated to a recent sample
3205 * when the event was scheduled out.
3207 if (ctx->task && cpuctx->task_ctx != ctx)
3210 raw_spin_lock(&ctx->lock);
3211 if (ctx->is_active) {
3212 update_context_time(ctx);
3213 update_cgrp_time_from_event(event);
3215 update_event_times(event);
3216 if (event->state == PERF_EVENT_STATE_ACTIVE)
3217 event->pmu->read(event);
3218 raw_spin_unlock(&ctx->lock);
3221 static inline u64 perf_event_count(struct perf_event *event)
3223 return local64_read(&event->count) + atomic64_read(&event->child_count);
3226 static u64 perf_event_read(struct perf_event *event)
3229 * If event is enabled and currently active on a CPU, update the
3230 * value in the event structure:
3232 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3233 smp_call_function_single(event->oncpu,
3234 __perf_event_read, event, 1);
3235 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3236 struct perf_event_context *ctx = event->ctx;
3237 unsigned long flags;
3239 raw_spin_lock_irqsave(&ctx->lock, flags);
3241 * may read while context is not active
3242 * (e.g., thread is blocked), in that case
3243 * we cannot update context time
3245 if (ctx->is_active) {
3246 update_context_time(ctx);
3247 update_cgrp_time_from_event(event);
3249 update_event_times(event);
3250 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3253 return perf_event_count(event);
3257 * Initialize the perf_event context in a task_struct:
3259 static void __perf_event_init_context(struct perf_event_context *ctx)
3261 raw_spin_lock_init(&ctx->lock);
3262 mutex_init(&ctx->mutex);
3263 INIT_LIST_HEAD(&ctx->active_ctx_list);
3264 INIT_LIST_HEAD(&ctx->pinned_groups);
3265 INIT_LIST_HEAD(&ctx->flexible_groups);
3266 INIT_LIST_HEAD(&ctx->event_list);
3267 atomic_set(&ctx->refcount, 1);
3268 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3271 static struct perf_event_context *
3272 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3274 struct perf_event_context *ctx;
3276 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3280 __perf_event_init_context(ctx);
3283 get_task_struct(task);
3290 static struct task_struct *
3291 find_lively_task_by_vpid(pid_t vpid)
3293 struct task_struct *task;
3300 task = find_task_by_vpid(vpid);
3302 get_task_struct(task);
3306 return ERR_PTR(-ESRCH);
3308 /* Reuse ptrace permission checks for now. */
3310 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3315 put_task_struct(task);
3316 return ERR_PTR(err);
3321 * Returns a matching context with refcount and pincount.
3323 static struct perf_event_context *
3324 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3326 struct perf_event_context *ctx, *clone_ctx = NULL;
3327 struct perf_cpu_context *cpuctx;
3328 unsigned long flags;
3332 /* Must be root to operate on a CPU event: */
3333 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3334 return ERR_PTR(-EACCES);
3337 * We could be clever and allow to attach a event to an
3338 * offline CPU and activate it when the CPU comes up, but
3341 if (!cpu_online(cpu))
3342 return ERR_PTR(-ENODEV);
3344 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3353 ctxn = pmu->task_ctx_nr;
3358 ctx = perf_lock_task_context(task, ctxn, &flags);
3360 clone_ctx = unclone_ctx(ctx);
3362 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3367 ctx = alloc_perf_context(pmu, task);
3373 mutex_lock(&task->perf_event_mutex);
3375 * If it has already passed perf_event_exit_task().
3376 * we must see PF_EXITING, it takes this mutex too.
3378 if (task->flags & PF_EXITING)
3380 else if (task->perf_event_ctxp[ctxn])
3385 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3387 mutex_unlock(&task->perf_event_mutex);
3389 if (unlikely(err)) {
3401 return ERR_PTR(err);
3404 static void perf_event_free_filter(struct perf_event *event);
3406 static void free_event_rcu(struct rcu_head *head)
3408 struct perf_event *event;
3410 event = container_of(head, struct perf_event, rcu_head);
3412 put_pid_ns(event->ns);
3413 perf_event_free_filter(event);
3417 static void ring_buffer_put(struct ring_buffer *rb);
3418 static void ring_buffer_attach(struct perf_event *event,
3419 struct ring_buffer *rb);
3421 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3426 if (has_branch_stack(event)) {
3427 if (!(event->attach_state & PERF_ATTACH_TASK))
3428 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3430 if (is_cgroup_event(event))
3431 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3434 static void unaccount_event(struct perf_event *event)
3439 if (event->attach_state & PERF_ATTACH_TASK)
3440 static_key_slow_dec_deferred(&perf_sched_events);
3441 if (event->attr.mmap || event->attr.mmap_data)
3442 atomic_dec(&nr_mmap_events);
3443 if (event->attr.comm)
3444 atomic_dec(&nr_comm_events);
3445 if (event->attr.task)
3446 atomic_dec(&nr_task_events);
3447 if (event->attr.freq)
3448 atomic_dec(&nr_freq_events);
3449 if (is_cgroup_event(event))
3450 static_key_slow_dec_deferred(&perf_sched_events);
3451 if (has_branch_stack(event))
3452 static_key_slow_dec_deferred(&perf_sched_events);
3454 unaccount_event_cpu(event, event->cpu);
3457 static void __free_event(struct perf_event *event)
3459 if (!event->parent) {
3460 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3461 put_callchain_buffers();
3465 event->destroy(event);
3468 put_ctx(event->ctx);
3471 module_put(event->pmu->module);
3473 call_rcu(&event->rcu_head, free_event_rcu);
3476 static void _free_event(struct perf_event *event)
3478 irq_work_sync(&event->pending);
3480 unaccount_event(event);
3484 * Can happen when we close an event with re-directed output.
3486 * Since we have a 0 refcount, perf_mmap_close() will skip
3487 * over us; possibly making our ring_buffer_put() the last.
3489 mutex_lock(&event->mmap_mutex);
3490 ring_buffer_attach(event, NULL);
3491 mutex_unlock(&event->mmap_mutex);
3494 if (is_cgroup_event(event))
3495 perf_detach_cgroup(event);
3497 __free_event(event);
3501 * Used to free events which have a known refcount of 1, such as in error paths
3502 * where the event isn't exposed yet and inherited events.
3504 static void free_event(struct perf_event *event)
3506 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3507 "unexpected event refcount: %ld; ptr=%p\n",
3508 atomic_long_read(&event->refcount), event)) {
3509 /* leak to avoid use-after-free */
3517 * Remove user event from the owner task.
3519 static void perf_remove_from_owner(struct perf_event *event)
3521 struct task_struct *owner;
3524 owner = ACCESS_ONCE(event->owner);
3526 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3527 * !owner it means the list deletion is complete and we can indeed
3528 * free this event, otherwise we need to serialize on
3529 * owner->perf_event_mutex.
3531 smp_read_barrier_depends();
3534 * Since delayed_put_task_struct() also drops the last
3535 * task reference we can safely take a new reference
3536 * while holding the rcu_read_lock().
3538 get_task_struct(owner);
3544 * If we're here through perf_event_exit_task() we're already
3545 * holding ctx->mutex which would be an inversion wrt. the
3546 * normal lock order.
3548 * However we can safely take this lock because its the child
3551 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
3554 * We have to re-check the event->owner field, if it is cleared
3555 * we raced with perf_event_exit_task(), acquiring the mutex
3556 * ensured they're done, and we can proceed with freeing the
3560 list_del_init(&event->owner_entry);
3561 mutex_unlock(&owner->perf_event_mutex);
3562 put_task_struct(owner);
3567 * Called when the last reference to the file is gone.
3569 static void put_event(struct perf_event *event)
3571 struct perf_event_context *ctx;
3573 if (!atomic_long_dec_and_test(&event->refcount))
3576 if (!is_kernel_event(event))
3577 perf_remove_from_owner(event);
3580 * There are two ways this annotation is useful:
3582 * 1) there is a lock recursion from perf_event_exit_task
3583 * see the comment there.
3585 * 2) there is a lock-inversion with mmap_sem through
3586 * perf_event_read_group(), which takes faults while
3587 * holding ctx->mutex, however this is called after
3588 * the last filedesc died, so there is no possibility
3589 * to trigger the AB-BA case.
3591 ctx = perf_event_ctx_lock_nested(event, SINGLE_DEPTH_NESTING);
3592 WARN_ON_ONCE(ctx->parent_ctx);
3593 perf_remove_from_context(event, true);
3594 perf_event_ctx_unlock(event, ctx);
3599 int perf_event_release_kernel(struct perf_event *event)
3604 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3606 static int perf_release(struct inode *inode, struct file *file)
3608 put_event(file->private_data);
3613 * Remove all orphanes events from the context.
3615 static void orphans_remove_work(struct work_struct *work)
3617 struct perf_event_context *ctx;
3618 struct perf_event *event, *tmp;
3620 ctx = container_of(work, struct perf_event_context,
3621 orphans_remove.work);
3623 mutex_lock(&ctx->mutex);
3624 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3625 struct perf_event *parent_event = event->parent;
3627 if (!is_orphaned_child(event))
3630 perf_remove_from_context(event, true);
3632 mutex_lock(&parent_event->child_mutex);
3633 list_del_init(&event->child_list);
3634 mutex_unlock(&parent_event->child_mutex);
3637 put_event(parent_event);
3640 raw_spin_lock_irq(&ctx->lock);
3641 ctx->orphans_remove_sched = false;
3642 raw_spin_unlock_irq(&ctx->lock);
3643 mutex_unlock(&ctx->mutex);
3648 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3650 struct perf_event *child;
3656 mutex_lock(&event->child_mutex);
3657 total += perf_event_read(event);
3658 *enabled += event->total_time_enabled +
3659 atomic64_read(&event->child_total_time_enabled);
3660 *running += event->total_time_running +
3661 atomic64_read(&event->child_total_time_running);
3663 list_for_each_entry(child, &event->child_list, child_list) {
3664 total += perf_event_read(child);
3665 *enabled += child->total_time_enabled;
3666 *running += child->total_time_running;
3668 mutex_unlock(&event->child_mutex);
3672 EXPORT_SYMBOL_GPL(perf_event_read_value);
3674 static int perf_event_read_group(struct perf_event *event,
3675 u64 read_format, char __user *buf)
3677 struct perf_event *leader = event->group_leader, *sub;
3678 struct perf_event_context *ctx = leader->ctx;
3679 int n = 0, size = 0, ret;
3680 u64 count, enabled, running;
3683 lockdep_assert_held(&ctx->mutex);
3685 count = perf_event_read_value(leader, &enabled, &running);
3687 values[n++] = 1 + leader->nr_siblings;
3688 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3689 values[n++] = enabled;
3690 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3691 values[n++] = running;
3692 values[n++] = count;
3693 if (read_format & PERF_FORMAT_ID)
3694 values[n++] = primary_event_id(leader);
3696 size = n * sizeof(u64);
3698 if (copy_to_user(buf, values, size))
3703 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3706 values[n++] = perf_event_read_value(sub, &enabled, &running);
3707 if (read_format & PERF_FORMAT_ID)
3708 values[n++] = primary_event_id(sub);
3710 size = n * sizeof(u64);
3712 if (copy_to_user(buf + ret, values, size)) {
3722 static int perf_event_read_one(struct perf_event *event,
3723 u64 read_format, char __user *buf)
3725 u64 enabled, running;
3729 values[n++] = perf_event_read_value(event, &enabled, &running);
3730 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3731 values[n++] = enabled;
3732 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3733 values[n++] = running;
3734 if (read_format & PERF_FORMAT_ID)
3735 values[n++] = primary_event_id(event);
3737 if (copy_to_user(buf, values, n * sizeof(u64)))
3740 return n * sizeof(u64);
3743 static bool is_event_hup(struct perf_event *event)
3747 if (event->state != PERF_EVENT_STATE_EXIT)
3750 mutex_lock(&event->child_mutex);
3751 no_children = list_empty(&event->child_list);
3752 mutex_unlock(&event->child_mutex);
3757 * Read the performance event - simple non blocking version for now
3760 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3762 u64 read_format = event->attr.read_format;
3766 * Return end-of-file for a read on a event that is in
3767 * error state (i.e. because it was pinned but it couldn't be
3768 * scheduled on to the CPU at some point).
3770 if (event->state == PERF_EVENT_STATE_ERROR)
3773 if (count < event->read_size)
3776 WARN_ON_ONCE(event->ctx->parent_ctx);
3777 if (read_format & PERF_FORMAT_GROUP)
3778 ret = perf_event_read_group(event, read_format, buf);
3780 ret = perf_event_read_one(event, read_format, buf);
3786 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3788 struct perf_event *event = file->private_data;
3789 struct perf_event_context *ctx;
3792 ctx = perf_event_ctx_lock(event);
3793 ret = perf_read_hw(event, buf, count);
3794 perf_event_ctx_unlock(event, ctx);
3799 static unsigned int perf_poll(struct file *file, poll_table *wait)
3801 struct perf_event *event = file->private_data;
3802 struct ring_buffer *rb;
3803 unsigned int events = POLLHUP;
3805 poll_wait(file, &event->waitq, wait);
3807 if (is_event_hup(event))
3811 * Pin the event->rb by taking event->mmap_mutex; otherwise
3812 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3814 mutex_lock(&event->mmap_mutex);
3817 events = atomic_xchg(&rb->poll, 0);
3818 mutex_unlock(&event->mmap_mutex);
3822 static void _perf_event_reset(struct perf_event *event)
3824 (void)perf_event_read(event);
3825 local64_set(&event->count, 0);
3826 perf_event_update_userpage(event);
3830 * Holding the top-level event's child_mutex means that any
3831 * descendant process that has inherited this event will block
3832 * in sync_child_event if it goes to exit, thus satisfying the
3833 * task existence requirements of perf_event_enable/disable.
3835 static void perf_event_for_each_child(struct perf_event *event,
3836 void (*func)(struct perf_event *))
3838 struct perf_event *child;
3840 WARN_ON_ONCE(event->ctx->parent_ctx);
3842 mutex_lock(&event->child_mutex);
3844 list_for_each_entry(child, &event->child_list, child_list)
3846 mutex_unlock(&event->child_mutex);
3849 static void perf_event_for_each(struct perf_event *event,
3850 void (*func)(struct perf_event *))
3852 struct perf_event_context *ctx = event->ctx;
3853 struct perf_event *sibling;
3855 lockdep_assert_held(&ctx->mutex);
3857 event = event->group_leader;
3859 perf_event_for_each_child(event, func);
3860 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3861 perf_event_for_each_child(sibling, func);
3864 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3866 struct perf_event_context *ctx = event->ctx;
3867 int ret = 0, active;
3870 if (!is_sampling_event(event))
3873 if (copy_from_user(&value, arg, sizeof(value)))
3879 raw_spin_lock_irq(&ctx->lock);
3880 if (event->attr.freq) {
3881 if (value > sysctl_perf_event_sample_rate) {
3886 event->attr.sample_freq = value;
3888 event->attr.sample_period = value;
3889 event->hw.sample_period = value;
3892 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3894 perf_pmu_disable(ctx->pmu);
3895 event->pmu->stop(event, PERF_EF_UPDATE);
3898 local64_set(&event->hw.period_left, 0);
3901 event->pmu->start(event, PERF_EF_RELOAD);
3902 perf_pmu_enable(ctx->pmu);
3906 raw_spin_unlock_irq(&ctx->lock);
3911 static const struct file_operations perf_fops;
3913 static inline int perf_fget_light(int fd, struct fd *p)
3915 struct fd f = fdget(fd);
3919 if (f.file->f_op != &perf_fops) {
3927 static int perf_event_set_output(struct perf_event *event,
3928 struct perf_event *output_event);
3929 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3931 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
3933 void (*func)(struct perf_event *);
3937 case PERF_EVENT_IOC_ENABLE:
3938 func = _perf_event_enable;
3940 case PERF_EVENT_IOC_DISABLE:
3941 func = _perf_event_disable;
3943 case PERF_EVENT_IOC_RESET:
3944 func = _perf_event_reset;
3947 case PERF_EVENT_IOC_REFRESH:
3948 return _perf_event_refresh(event, arg);
3950 case PERF_EVENT_IOC_PERIOD:
3951 return perf_event_period(event, (u64 __user *)arg);
3953 case PERF_EVENT_IOC_ID:
3955 u64 id = primary_event_id(event);
3957 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3962 case PERF_EVENT_IOC_SET_OUTPUT:
3966 struct perf_event *output_event;
3968 ret = perf_fget_light(arg, &output);
3971 output_event = output.file->private_data;
3972 ret = perf_event_set_output(event, output_event);
3975 ret = perf_event_set_output(event, NULL);
3980 case PERF_EVENT_IOC_SET_FILTER:
3981 return perf_event_set_filter(event, (void __user *)arg);
3987 if (flags & PERF_IOC_FLAG_GROUP)
3988 perf_event_for_each(event, func);
3990 perf_event_for_each_child(event, func);
3995 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3997 struct perf_event *event = file->private_data;
3998 struct perf_event_context *ctx;
4001 ctx = perf_event_ctx_lock(event);
4002 ret = _perf_ioctl(event, cmd, arg);
4003 perf_event_ctx_unlock(event, ctx);
4008 #ifdef CONFIG_COMPAT
4009 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4012 switch (_IOC_NR(cmd)) {
4013 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4014 case _IOC_NR(PERF_EVENT_IOC_ID):
4015 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4016 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4017 cmd &= ~IOCSIZE_MASK;
4018 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4022 return perf_ioctl(file, cmd, arg);
4025 # define perf_compat_ioctl NULL
4028 int perf_event_task_enable(void)
4030 struct perf_event_context *ctx;
4031 struct perf_event *event;
4033 mutex_lock(¤t->perf_event_mutex);
4034 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4035 ctx = perf_event_ctx_lock(event);
4036 perf_event_for_each_child(event, _perf_event_enable);
4037 perf_event_ctx_unlock(event, ctx);
4039 mutex_unlock(¤t->perf_event_mutex);
4044 int perf_event_task_disable(void)
4046 struct perf_event_context *ctx;
4047 struct perf_event *event;
4049 mutex_lock(¤t->perf_event_mutex);
4050 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4051 ctx = perf_event_ctx_lock(event);
4052 perf_event_for_each_child(event, _perf_event_disable);
4053 perf_event_ctx_unlock(event, ctx);
4055 mutex_unlock(¤t->perf_event_mutex);
4060 static int perf_event_index(struct perf_event *event)
4062 if (event->hw.state & PERF_HES_STOPPED)
4065 if (event->state != PERF_EVENT_STATE_ACTIVE)
4068 return event->pmu->event_idx(event);
4071 static void calc_timer_values(struct perf_event *event,
4078 *now = perf_clock();
4079 ctx_time = event->shadow_ctx_time + *now;
4080 *enabled = ctx_time - event->tstamp_enabled;
4081 *running = ctx_time - event->tstamp_running;
4084 static void perf_event_init_userpage(struct perf_event *event)
4086 struct perf_event_mmap_page *userpg;
4087 struct ring_buffer *rb;
4090 rb = rcu_dereference(event->rb);
4094 userpg = rb->user_page;
4096 /* Allow new userspace to detect that bit 0 is deprecated */
4097 userpg->cap_bit0_is_deprecated = 1;
4098 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4104 void __weak arch_perf_update_userpage(
4105 struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
4110 * Callers need to ensure there can be no nesting of this function, otherwise
4111 * the seqlock logic goes bad. We can not serialize this because the arch
4112 * code calls this from NMI context.
4114 void perf_event_update_userpage(struct perf_event *event)
4116 struct perf_event_mmap_page *userpg;
4117 struct ring_buffer *rb;
4118 u64 enabled, running, now;
4121 rb = rcu_dereference(event->rb);
4126 * compute total_time_enabled, total_time_running
4127 * based on snapshot values taken when the event
4128 * was last scheduled in.
4130 * we cannot simply called update_context_time()
4131 * because of locking issue as we can be called in
4134 calc_timer_values(event, &now, &enabled, &running);
4136 userpg = rb->user_page;
4138 * Disable preemption so as to not let the corresponding user-space
4139 * spin too long if we get preempted.
4144 userpg->index = perf_event_index(event);
4145 userpg->offset = perf_event_count(event);
4147 userpg->offset -= local64_read(&event->hw.prev_count);
4149 userpg->time_enabled = enabled +
4150 atomic64_read(&event->child_total_time_enabled);
4152 userpg->time_running = running +
4153 atomic64_read(&event->child_total_time_running);
4155 arch_perf_update_userpage(event, userpg, now);
4164 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4166 struct perf_event *event = vma->vm_file->private_data;
4167 struct ring_buffer *rb;
4168 int ret = VM_FAULT_SIGBUS;
4170 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4171 if (vmf->pgoff == 0)
4177 rb = rcu_dereference(event->rb);
4181 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4184 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4188 get_page(vmf->page);
4189 vmf->page->mapping = vma->vm_file->f_mapping;
4190 vmf->page->index = vmf->pgoff;
4199 static void ring_buffer_attach(struct perf_event *event,
4200 struct ring_buffer *rb)
4202 struct ring_buffer *old_rb = NULL;
4203 unsigned long flags;
4207 * Should be impossible, we set this when removing
4208 * event->rb_entry and wait/clear when adding event->rb_entry.
4210 WARN_ON_ONCE(event->rcu_pending);
4213 event->rcu_batches = get_state_synchronize_rcu();
4214 event->rcu_pending = 1;
4216 spin_lock_irqsave(&old_rb->event_lock, flags);
4217 list_del_rcu(&event->rb_entry);
4218 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4221 if (event->rcu_pending && rb) {
4222 cond_synchronize_rcu(event->rcu_batches);
4223 event->rcu_pending = 0;
4227 spin_lock_irqsave(&rb->event_lock, flags);
4228 list_add_rcu(&event->rb_entry, &rb->event_list);
4229 spin_unlock_irqrestore(&rb->event_lock, flags);
4232 rcu_assign_pointer(event->rb, rb);
4235 ring_buffer_put(old_rb);
4237 * Since we detached before setting the new rb, so that we
4238 * could attach the new rb, we could have missed a wakeup.
4241 wake_up_all(&event->waitq);
4245 static void ring_buffer_wakeup(struct perf_event *event)
4247 struct ring_buffer *rb;
4250 rb = rcu_dereference(event->rb);
4252 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4253 wake_up_all(&event->waitq);
4258 static void rb_free_rcu(struct rcu_head *rcu_head)
4260 struct ring_buffer *rb;
4262 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4266 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
4268 struct ring_buffer *rb;
4271 rb = rcu_dereference(event->rb);
4273 if (!atomic_inc_not_zero(&rb->refcount))
4281 static void ring_buffer_put(struct ring_buffer *rb)
4283 if (!atomic_dec_and_test(&rb->refcount))
4286 WARN_ON_ONCE(!list_empty(&rb->event_list));
4288 call_rcu(&rb->rcu_head, rb_free_rcu);
4291 static void perf_mmap_open(struct vm_area_struct *vma)
4293 struct perf_event *event = vma->vm_file->private_data;
4295 atomic_inc(&event->mmap_count);
4296 atomic_inc(&event->rb->mmap_count);
4298 if (event->pmu->event_mapped)
4299 event->pmu->event_mapped(event);
4303 * A buffer can be mmap()ed multiple times; either directly through the same
4304 * event, or through other events by use of perf_event_set_output().
4306 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4307 * the buffer here, where we still have a VM context. This means we need
4308 * to detach all events redirecting to us.
4310 static void perf_mmap_close(struct vm_area_struct *vma)
4312 struct perf_event *event = vma->vm_file->private_data;
4314 struct ring_buffer *rb = ring_buffer_get(event);
4315 struct user_struct *mmap_user = rb->mmap_user;
4316 int mmap_locked = rb->mmap_locked;
4317 unsigned long size = perf_data_size(rb);
4319 if (event->pmu->event_unmapped)
4320 event->pmu->event_unmapped(event);
4322 atomic_dec(&rb->mmap_count);
4324 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4327 ring_buffer_attach(event, NULL);
4328 mutex_unlock(&event->mmap_mutex);
4330 /* If there's still other mmap()s of this buffer, we're done. */
4331 if (atomic_read(&rb->mmap_count))
4335 * No other mmap()s, detach from all other events that might redirect
4336 * into the now unreachable buffer. Somewhat complicated by the
4337 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4341 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4342 if (!atomic_long_inc_not_zero(&event->refcount)) {
4344 * This event is en-route to free_event() which will
4345 * detach it and remove it from the list.
4351 mutex_lock(&event->mmap_mutex);
4353 * Check we didn't race with perf_event_set_output() which can
4354 * swizzle the rb from under us while we were waiting to
4355 * acquire mmap_mutex.
4357 * If we find a different rb; ignore this event, a next
4358 * iteration will no longer find it on the list. We have to
4359 * still restart the iteration to make sure we're not now
4360 * iterating the wrong list.
4362 if (event->rb == rb)
4363 ring_buffer_attach(event, NULL);
4365 mutex_unlock(&event->mmap_mutex);
4369 * Restart the iteration; either we're on the wrong list or
4370 * destroyed its integrity by doing a deletion.
4377 * It could be there's still a few 0-ref events on the list; they'll
4378 * get cleaned up by free_event() -- they'll also still have their
4379 * ref on the rb and will free it whenever they are done with it.
4381 * Aside from that, this buffer is 'fully' detached and unmapped,
4382 * undo the VM accounting.
4385 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4386 vma->vm_mm->pinned_vm -= mmap_locked;
4387 free_uid(mmap_user);
4390 ring_buffer_put(rb); /* could be last */
4393 static const struct vm_operations_struct perf_mmap_vmops = {
4394 .open = perf_mmap_open,
4395 .close = perf_mmap_close,
4396 .fault = perf_mmap_fault,
4397 .page_mkwrite = perf_mmap_fault,
4400 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4402 struct perf_event *event = file->private_data;
4403 unsigned long user_locked, user_lock_limit;
4404 struct user_struct *user = current_user();
4405 unsigned long locked, lock_limit;
4406 struct ring_buffer *rb;
4407 unsigned long vma_size;
4408 unsigned long nr_pages;
4409 long user_extra, extra;
4410 int ret = 0, flags = 0;
4413 * Don't allow mmap() of inherited per-task counters. This would
4414 * create a performance issue due to all children writing to the
4417 if (event->cpu == -1 && event->attr.inherit)
4420 if (!(vma->vm_flags & VM_SHARED))
4423 vma_size = vma->vm_end - vma->vm_start;
4424 nr_pages = (vma_size / PAGE_SIZE) - 1;
4427 * If we have rb pages ensure they're a power-of-two number, so we
4428 * can do bitmasks instead of modulo.
4430 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4433 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4436 if (vma->vm_pgoff != 0)
4439 WARN_ON_ONCE(event->ctx->parent_ctx);
4441 mutex_lock(&event->mmap_mutex);
4443 if (event->rb->nr_pages != nr_pages) {
4448 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4450 * Raced against perf_mmap_close() through
4451 * perf_event_set_output(). Try again, hope for better
4454 mutex_unlock(&event->mmap_mutex);
4461 user_extra = nr_pages + 1;
4462 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4465 * Increase the limit linearly with more CPUs:
4467 user_lock_limit *= num_online_cpus();
4469 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4472 if (user_locked > user_lock_limit)
4473 extra = user_locked - user_lock_limit;
4475 lock_limit = rlimit(RLIMIT_MEMLOCK);
4476 lock_limit >>= PAGE_SHIFT;
4477 locked = vma->vm_mm->pinned_vm + extra;
4479 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4480 !capable(CAP_IPC_LOCK)) {
4487 if (vma->vm_flags & VM_WRITE)
4488 flags |= RING_BUFFER_WRITABLE;
4490 rb = rb_alloc(nr_pages,
4491 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4499 atomic_set(&rb->mmap_count, 1);
4500 rb->mmap_locked = extra;
4501 rb->mmap_user = get_current_user();
4503 atomic_long_add(user_extra, &user->locked_vm);
4504 vma->vm_mm->pinned_vm += extra;
4506 ring_buffer_attach(event, rb);
4508 perf_event_init_userpage(event);
4509 perf_event_update_userpage(event);
4513 atomic_inc(&event->mmap_count);
4514 mutex_unlock(&event->mmap_mutex);
4517 * Since pinned accounting is per vm we cannot allow fork() to copy our
4520 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4521 vma->vm_ops = &perf_mmap_vmops;
4523 if (event->pmu->event_mapped)
4524 event->pmu->event_mapped(event);
4529 static int perf_fasync(int fd, struct file *filp, int on)
4531 struct inode *inode = file_inode(filp);
4532 struct perf_event *event = filp->private_data;
4535 mutex_lock(&inode->i_mutex);
4536 retval = fasync_helper(fd, filp, on, &event->fasync);
4537 mutex_unlock(&inode->i_mutex);
4545 static const struct file_operations perf_fops = {
4546 .llseek = no_llseek,
4547 .release = perf_release,
4550 .unlocked_ioctl = perf_ioctl,
4551 .compat_ioctl = perf_compat_ioctl,
4553 .fasync = perf_fasync,
4559 * If there's data, ensure we set the poll() state and publish everything
4560 * to user-space before waking everybody up.
4563 void perf_event_wakeup(struct perf_event *event)
4565 ring_buffer_wakeup(event);
4567 if (event->pending_kill) {
4568 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4569 event->pending_kill = 0;
4573 static void perf_pending_event(struct irq_work *entry)
4575 struct perf_event *event = container_of(entry,
4576 struct perf_event, pending);
4579 rctx = perf_swevent_get_recursion_context();
4581 * If we 'fail' here, that's OK, it means recursion is already disabled
4582 * and we won't recurse 'further'.
4585 if (event->pending_disable) {
4586 event->pending_disable = 0;
4587 __perf_event_disable(event);
4590 if (event->pending_wakeup) {
4591 event->pending_wakeup = 0;
4592 perf_event_wakeup(event);
4596 perf_swevent_put_recursion_context(rctx);
4600 * We assume there is only KVM supporting the callbacks.
4601 * Later on, we might change it to a list if there is
4602 * another virtualization implementation supporting the callbacks.
4604 struct perf_guest_info_callbacks *perf_guest_cbs;
4606 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4608 perf_guest_cbs = cbs;
4611 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4613 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4615 perf_guest_cbs = NULL;
4618 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4621 perf_output_sample_regs(struct perf_output_handle *handle,
4622 struct pt_regs *regs, u64 mask)
4626 for_each_set_bit(bit, (const unsigned long *) &mask,
4627 sizeof(mask) * BITS_PER_BYTE) {
4630 val = perf_reg_value(regs, bit);
4631 perf_output_put(handle, val);
4635 static void perf_sample_regs_user(struct perf_regs *regs_user,
4636 struct pt_regs *regs,
4637 struct pt_regs *regs_user_copy)
4639 if (user_mode(regs)) {
4640 regs_user->abi = perf_reg_abi(current);
4641 regs_user->regs = regs;
4642 } else if (current->mm) {
4643 perf_get_regs_user(regs_user, regs, regs_user_copy);
4645 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4646 regs_user->regs = NULL;
4650 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4651 struct pt_regs *regs)
4653 regs_intr->regs = regs;
4654 regs_intr->abi = perf_reg_abi(current);
4659 * Get remaining task size from user stack pointer.
4661 * It'd be better to take stack vma map and limit this more
4662 * precisly, but there's no way to get it safely under interrupt,
4663 * so using TASK_SIZE as limit.
4665 static u64 perf_ustack_task_size(struct pt_regs *regs)
4667 unsigned long addr = perf_user_stack_pointer(regs);
4669 if (!addr || addr >= TASK_SIZE)
4672 return TASK_SIZE - addr;
4676 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4677 struct pt_regs *regs)
4681 /* No regs, no stack pointer, no dump. */
4686 * Check if we fit in with the requested stack size into the:
4688 * If we don't, we limit the size to the TASK_SIZE.
4690 * - remaining sample size
4691 * If we don't, we customize the stack size to
4692 * fit in to the remaining sample size.
4695 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4696 stack_size = min(stack_size, (u16) task_size);
4698 /* Current header size plus static size and dynamic size. */
4699 header_size += 2 * sizeof(u64);
4701 /* Do we fit in with the current stack dump size? */
4702 if ((u16) (header_size + stack_size) < header_size) {
4704 * If we overflow the maximum size for the sample,
4705 * we customize the stack dump size to fit in.
4707 stack_size = USHRT_MAX - header_size - sizeof(u64);
4708 stack_size = round_up(stack_size, sizeof(u64));
4715 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4716 struct pt_regs *regs)
4718 /* Case of a kernel thread, nothing to dump */
4721 perf_output_put(handle, size);
4730 * - the size requested by user or the best one we can fit
4731 * in to the sample max size
4733 * - user stack dump data
4735 * - the actual dumped size
4739 perf_output_put(handle, dump_size);
4742 sp = perf_user_stack_pointer(regs);
4743 rem = __output_copy_user(handle, (void *) sp, dump_size);
4744 dyn_size = dump_size - rem;
4746 perf_output_skip(handle, rem);
4749 perf_output_put(handle, dyn_size);
4753 static void __perf_event_header__init_id(struct perf_event_header *header,
4754 struct perf_sample_data *data,
4755 struct perf_event *event)
4757 u64 sample_type = event->attr.sample_type;
4759 data->type = sample_type;
4760 header->size += event->id_header_size;
4762 if (sample_type & PERF_SAMPLE_TID) {
4763 /* namespace issues */
4764 data->tid_entry.pid = perf_event_pid(event, current);
4765 data->tid_entry.tid = perf_event_tid(event, current);
4768 if (sample_type & PERF_SAMPLE_TIME)
4769 data->time = perf_clock();
4771 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4772 data->id = primary_event_id(event);
4774 if (sample_type & PERF_SAMPLE_STREAM_ID)
4775 data->stream_id = event->id;
4777 if (sample_type & PERF_SAMPLE_CPU) {
4778 data->cpu_entry.cpu = raw_smp_processor_id();
4779 data->cpu_entry.reserved = 0;
4783 void perf_event_header__init_id(struct perf_event_header *header,
4784 struct perf_sample_data *data,
4785 struct perf_event *event)
4787 if (event->attr.sample_id_all)
4788 __perf_event_header__init_id(header, data, event);
4791 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4792 struct perf_sample_data *data)
4794 u64 sample_type = data->type;
4796 if (sample_type & PERF_SAMPLE_TID)
4797 perf_output_put(handle, data->tid_entry);
4799 if (sample_type & PERF_SAMPLE_TIME)
4800 perf_output_put(handle, data->time);
4802 if (sample_type & PERF_SAMPLE_ID)
4803 perf_output_put(handle, data->id);
4805 if (sample_type & PERF_SAMPLE_STREAM_ID)
4806 perf_output_put(handle, data->stream_id);
4808 if (sample_type & PERF_SAMPLE_CPU)
4809 perf_output_put(handle, data->cpu_entry);
4811 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4812 perf_output_put(handle, data->id);
4815 void perf_event__output_id_sample(struct perf_event *event,
4816 struct perf_output_handle *handle,
4817 struct perf_sample_data *sample)
4819 if (event->attr.sample_id_all)
4820 __perf_event__output_id_sample(handle, sample);
4823 static void perf_output_read_one(struct perf_output_handle *handle,
4824 struct perf_event *event,
4825 u64 enabled, u64 running)
4827 u64 read_format = event->attr.read_format;
4831 values[n++] = perf_event_count(event);
4832 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4833 values[n++] = enabled +
4834 atomic64_read(&event->child_total_time_enabled);
4836 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4837 values[n++] = running +
4838 atomic64_read(&event->child_total_time_running);
4840 if (read_format & PERF_FORMAT_ID)
4841 values[n++] = primary_event_id(event);
4843 __output_copy(handle, values, n * sizeof(u64));
4847 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4849 static void perf_output_read_group(struct perf_output_handle *handle,
4850 struct perf_event *event,
4851 u64 enabled, u64 running)
4853 struct perf_event *leader = event->group_leader, *sub;
4854 u64 read_format = event->attr.read_format;
4858 values[n++] = 1 + leader->nr_siblings;
4860 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4861 values[n++] = enabled;
4863 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4864 values[n++] = running;
4866 if (leader != event)
4867 leader->pmu->read(leader);
4869 values[n++] = perf_event_count(leader);
4870 if (read_format & PERF_FORMAT_ID)
4871 values[n++] = primary_event_id(leader);
4873 __output_copy(handle, values, n * sizeof(u64));
4875 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4878 if ((sub != event) &&
4879 (sub->state == PERF_EVENT_STATE_ACTIVE))
4880 sub->pmu->read(sub);
4882 values[n++] = perf_event_count(sub);
4883 if (read_format & PERF_FORMAT_ID)
4884 values[n++] = primary_event_id(sub);
4886 __output_copy(handle, values, n * sizeof(u64));
4890 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4891 PERF_FORMAT_TOTAL_TIME_RUNNING)
4893 static void perf_output_read(struct perf_output_handle *handle,
4894 struct perf_event *event)
4896 u64 enabled = 0, running = 0, now;
4897 u64 read_format = event->attr.read_format;
4900 * compute total_time_enabled, total_time_running
4901 * based on snapshot values taken when the event
4902 * was last scheduled in.
4904 * we cannot simply called update_context_time()
4905 * because of locking issue as we are called in
4908 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4909 calc_timer_values(event, &now, &enabled, &running);
4911 if (event->attr.read_format & PERF_FORMAT_GROUP)
4912 perf_output_read_group(handle, event, enabled, running);
4914 perf_output_read_one(handle, event, enabled, running);
4917 void perf_output_sample(struct perf_output_handle *handle,
4918 struct perf_event_header *header,
4919 struct perf_sample_data *data,
4920 struct perf_event *event)
4922 u64 sample_type = data->type;
4924 perf_output_put(handle, *header);
4926 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4927 perf_output_put(handle, data->id);
4929 if (sample_type & PERF_SAMPLE_IP)
4930 perf_output_put(handle, data->ip);
4932 if (sample_type & PERF_SAMPLE_TID)
4933 perf_output_put(handle, data->tid_entry);
4935 if (sample_type & PERF_SAMPLE_TIME)
4936 perf_output_put(handle, data->time);
4938 if (sample_type & PERF_SAMPLE_ADDR)
4939 perf_output_put(handle, data->addr);
4941 if (sample_type & PERF_SAMPLE_ID)
4942 perf_output_put(handle, data->id);
4944 if (sample_type & PERF_SAMPLE_STREAM_ID)
4945 perf_output_put(handle, data->stream_id);
4947 if (sample_type & PERF_SAMPLE_CPU)
4948 perf_output_put(handle, data->cpu_entry);
4950 if (sample_type & PERF_SAMPLE_PERIOD)
4951 perf_output_put(handle, data->period);
4953 if (sample_type & PERF_SAMPLE_READ)
4954 perf_output_read(handle, event);
4956 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4957 if (data->callchain) {
4960 if (data->callchain)
4961 size += data->callchain->nr;
4963 size *= sizeof(u64);
4965 __output_copy(handle, data->callchain, size);
4968 perf_output_put(handle, nr);
4972 if (sample_type & PERF_SAMPLE_RAW) {
4974 perf_output_put(handle, data->raw->size);
4975 __output_copy(handle, data->raw->data,
4982 .size = sizeof(u32),
4985 perf_output_put(handle, raw);
4989 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4990 if (data->br_stack) {
4993 size = data->br_stack->nr
4994 * sizeof(struct perf_branch_entry);
4996 perf_output_put(handle, data->br_stack->nr);
4997 perf_output_copy(handle, data->br_stack->entries, size);
5000 * we always store at least the value of nr
5003 perf_output_put(handle, nr);
5007 if (sample_type & PERF_SAMPLE_REGS_USER) {
5008 u64 abi = data->regs_user.abi;
5011 * If there are no regs to dump, notice it through
5012 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5014 perf_output_put(handle, abi);
5017 u64 mask = event->attr.sample_regs_user;
5018 perf_output_sample_regs(handle,
5019 data->regs_user.regs,
5024 if (sample_type & PERF_SAMPLE_STACK_USER) {
5025 perf_output_sample_ustack(handle,
5026 data->stack_user_size,
5027 data->regs_user.regs);
5030 if (sample_type & PERF_SAMPLE_WEIGHT)
5031 perf_output_put(handle, data->weight);
5033 if (sample_type & PERF_SAMPLE_DATA_SRC)
5034 perf_output_put(handle, data->data_src.val);
5036 if (sample_type & PERF_SAMPLE_TRANSACTION)
5037 perf_output_put(handle, data->txn);
5039 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5040 u64 abi = data->regs_intr.abi;
5042 * If there are no regs to dump, notice it through
5043 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5045 perf_output_put(handle, abi);
5048 u64 mask = event->attr.sample_regs_intr;
5050 perf_output_sample_regs(handle,
5051 data->regs_intr.regs,
5056 if (!event->attr.watermark) {
5057 int wakeup_events = event->attr.wakeup_events;
5059 if (wakeup_events) {
5060 struct ring_buffer *rb = handle->rb;
5061 int events = local_inc_return(&rb->events);
5063 if (events >= wakeup_events) {
5064 local_sub(wakeup_events, &rb->events);
5065 local_inc(&rb->wakeup);
5071 void perf_prepare_sample(struct perf_event_header *header,
5072 struct perf_sample_data *data,
5073 struct perf_event *event,
5074 struct pt_regs *regs)
5076 u64 sample_type = event->attr.sample_type;
5078 header->type = PERF_RECORD_SAMPLE;
5079 header->size = sizeof(*header) + event->header_size;
5082 header->misc |= perf_misc_flags(regs);
5084 __perf_event_header__init_id(header, data, event);
5086 if (sample_type & PERF_SAMPLE_IP)
5087 data->ip = perf_instruction_pointer(regs);
5089 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5092 data->callchain = perf_callchain(event, regs);
5094 if (data->callchain)
5095 size += data->callchain->nr;
5097 header->size += size * sizeof(u64);
5100 if (sample_type & PERF_SAMPLE_RAW) {
5101 int size = sizeof(u32);
5104 size += data->raw->size;
5106 size += sizeof(u32);
5108 WARN_ON_ONCE(size & (sizeof(u64)-1));
5109 header->size += size;
5112 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5113 int size = sizeof(u64); /* nr */
5114 if (data->br_stack) {
5115 size += data->br_stack->nr
5116 * sizeof(struct perf_branch_entry);
5118 header->size += size;
5121 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
5122 perf_sample_regs_user(&data->regs_user, regs,
5123 &data->regs_user_copy);
5125 if (sample_type & PERF_SAMPLE_REGS_USER) {
5126 /* regs dump ABI info */
5127 int size = sizeof(u64);
5129 if (data->regs_user.regs) {
5130 u64 mask = event->attr.sample_regs_user;
5131 size += hweight64(mask) * sizeof(u64);
5134 header->size += size;
5137 if (sample_type & PERF_SAMPLE_STACK_USER) {
5139 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5140 * processed as the last one or have additional check added
5141 * in case new sample type is added, because we could eat
5142 * up the rest of the sample size.
5144 u16 stack_size = event->attr.sample_stack_user;
5145 u16 size = sizeof(u64);
5147 stack_size = perf_sample_ustack_size(stack_size, header->size,
5148 data->regs_user.regs);
5151 * If there is something to dump, add space for the dump
5152 * itself and for the field that tells the dynamic size,
5153 * which is how many have been actually dumped.
5156 size += sizeof(u64) + stack_size;
5158 data->stack_user_size = stack_size;
5159 header->size += size;
5162 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5163 /* regs dump ABI info */
5164 int size = sizeof(u64);
5166 perf_sample_regs_intr(&data->regs_intr, regs);
5168 if (data->regs_intr.regs) {
5169 u64 mask = event->attr.sample_regs_intr;
5171 size += hweight64(mask) * sizeof(u64);
5174 header->size += size;
5178 static void perf_event_output(struct perf_event *event,
5179 struct perf_sample_data *data,
5180 struct pt_regs *regs)
5182 struct perf_output_handle handle;
5183 struct perf_event_header header;
5185 /* protect the callchain buffers */
5188 perf_prepare_sample(&header, data, event, regs);
5190 if (perf_output_begin(&handle, event, header.size))
5193 perf_output_sample(&handle, &header, data, event);
5195 perf_output_end(&handle);
5205 struct perf_read_event {
5206 struct perf_event_header header;
5213 perf_event_read_event(struct perf_event *event,
5214 struct task_struct *task)
5216 struct perf_output_handle handle;
5217 struct perf_sample_data sample;
5218 struct perf_read_event read_event = {
5220 .type = PERF_RECORD_READ,
5222 .size = sizeof(read_event) + event->read_size,
5224 .pid = perf_event_pid(event, task),
5225 .tid = perf_event_tid(event, task),
5229 perf_event_header__init_id(&read_event.header, &sample, event);
5230 ret = perf_output_begin(&handle, event, read_event.header.size);
5234 perf_output_put(&handle, read_event);
5235 perf_output_read(&handle, event);
5236 perf_event__output_id_sample(event, &handle, &sample);
5238 perf_output_end(&handle);
5241 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5244 perf_event_aux_ctx(struct perf_event_context *ctx,
5245 perf_event_aux_output_cb output,
5248 struct perf_event *event;
5250 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5251 if (event->state < PERF_EVENT_STATE_INACTIVE)
5253 if (!event_filter_match(event))
5255 output(event, data);
5260 perf_event_aux(perf_event_aux_output_cb output, void *data,
5261 struct perf_event_context *task_ctx)
5263 struct perf_cpu_context *cpuctx;
5264 struct perf_event_context *ctx;
5269 list_for_each_entry_rcu(pmu, &pmus, entry) {
5270 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5271 if (cpuctx->unique_pmu != pmu)
5273 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5276 ctxn = pmu->task_ctx_nr;
5279 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5281 perf_event_aux_ctx(ctx, output, data);
5283 put_cpu_ptr(pmu->pmu_cpu_context);
5288 perf_event_aux_ctx(task_ctx, output, data);
5295 * task tracking -- fork/exit
5297 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5300 struct perf_task_event {
5301 struct task_struct *task;
5302 struct perf_event_context *task_ctx;
5305 struct perf_event_header header;
5315 static int perf_event_task_match(struct perf_event *event)
5317 return event->attr.comm || event->attr.mmap ||
5318 event->attr.mmap2 || event->attr.mmap_data ||
5322 static void perf_event_task_output(struct perf_event *event,
5325 struct perf_task_event *task_event = data;
5326 struct perf_output_handle handle;
5327 struct perf_sample_data sample;
5328 struct task_struct *task = task_event->task;
5329 int ret, size = task_event->event_id.header.size;
5331 if (!perf_event_task_match(event))
5334 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5336 ret = perf_output_begin(&handle, event,
5337 task_event->event_id.header.size);
5341 task_event->event_id.pid = perf_event_pid(event, task);
5342 task_event->event_id.ppid = perf_event_pid(event, current);
5344 task_event->event_id.tid = perf_event_tid(event, task);
5345 task_event->event_id.ptid = perf_event_tid(event, current);
5347 perf_output_put(&handle, task_event->event_id);
5349 perf_event__output_id_sample(event, &handle, &sample);
5351 perf_output_end(&handle);
5353 task_event->event_id.header.size = size;
5356 static void perf_event_task(struct task_struct *task,
5357 struct perf_event_context *task_ctx,
5360 struct perf_task_event task_event;
5362 if (!atomic_read(&nr_comm_events) &&
5363 !atomic_read(&nr_mmap_events) &&
5364 !atomic_read(&nr_task_events))
5367 task_event = (struct perf_task_event){
5369 .task_ctx = task_ctx,
5372 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5374 .size = sizeof(task_event.event_id),
5380 .time = perf_clock(),
5384 perf_event_aux(perf_event_task_output,
5389 void perf_event_fork(struct task_struct *task)
5391 perf_event_task(task, NULL, 1);
5398 struct perf_comm_event {
5399 struct task_struct *task;
5404 struct perf_event_header header;
5411 static int perf_event_comm_match(struct perf_event *event)
5413 return event->attr.comm;
5416 static void perf_event_comm_output(struct perf_event *event,
5419 struct perf_comm_event *comm_event = data;
5420 struct perf_output_handle handle;
5421 struct perf_sample_data sample;
5422 int size = comm_event->event_id.header.size;
5425 if (!perf_event_comm_match(event))
5428 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5429 ret = perf_output_begin(&handle, event,
5430 comm_event->event_id.header.size);
5435 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5436 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5438 perf_output_put(&handle, comm_event->event_id);
5439 __output_copy(&handle, comm_event->comm,
5440 comm_event->comm_size);
5442 perf_event__output_id_sample(event, &handle, &sample);
5444 perf_output_end(&handle);
5446 comm_event->event_id.header.size = size;
5449 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5451 char comm[TASK_COMM_LEN];
5454 memset(comm, 0, sizeof(comm));
5455 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5456 size = ALIGN(strlen(comm)+1, sizeof(u64));
5458 comm_event->comm = comm;
5459 comm_event->comm_size = size;
5461 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5463 perf_event_aux(perf_event_comm_output,
5468 void perf_event_comm(struct task_struct *task, bool exec)
5470 struct perf_comm_event comm_event;
5472 if (!atomic_read(&nr_comm_events))
5475 comm_event = (struct perf_comm_event){
5481 .type = PERF_RECORD_COMM,
5482 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5490 perf_event_comm_event(&comm_event);
5497 struct perf_mmap_event {
5498 struct vm_area_struct *vma;
5500 const char *file_name;
5508 struct perf_event_header header;
5518 static int perf_event_mmap_match(struct perf_event *event,
5521 struct perf_mmap_event *mmap_event = data;
5522 struct vm_area_struct *vma = mmap_event->vma;
5523 int executable = vma->vm_flags & VM_EXEC;
5525 return (!executable && event->attr.mmap_data) ||
5526 (executable && (event->attr.mmap || event->attr.mmap2));
5529 static void perf_event_mmap_output(struct perf_event *event,
5532 struct perf_mmap_event *mmap_event = data;
5533 struct perf_output_handle handle;
5534 struct perf_sample_data sample;
5535 int size = mmap_event->event_id.header.size;
5538 if (!perf_event_mmap_match(event, data))
5541 if (event->attr.mmap2) {
5542 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5543 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5544 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5545 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5546 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5547 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5548 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5551 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5552 ret = perf_output_begin(&handle, event,
5553 mmap_event->event_id.header.size);
5557 mmap_event->event_id.pid = perf_event_pid(event, current);
5558 mmap_event->event_id.tid = perf_event_tid(event, current);
5560 perf_output_put(&handle, mmap_event->event_id);
5562 if (event->attr.mmap2) {
5563 perf_output_put(&handle, mmap_event->maj);
5564 perf_output_put(&handle, mmap_event->min);
5565 perf_output_put(&handle, mmap_event->ino);
5566 perf_output_put(&handle, mmap_event->ino_generation);
5567 perf_output_put(&handle, mmap_event->prot);
5568 perf_output_put(&handle, mmap_event->flags);
5571 __output_copy(&handle, mmap_event->file_name,
5572 mmap_event->file_size);
5574 perf_event__output_id_sample(event, &handle, &sample);
5576 perf_output_end(&handle);
5578 mmap_event->event_id.header.size = size;
5581 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5583 struct vm_area_struct *vma = mmap_event->vma;
5584 struct file *file = vma->vm_file;
5585 int maj = 0, min = 0;
5586 u64 ino = 0, gen = 0;
5587 u32 prot = 0, flags = 0;
5594 struct inode *inode;
5597 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5603 * d_path() works from the end of the rb backwards, so we
5604 * need to add enough zero bytes after the string to handle
5605 * the 64bit alignment we do later.
5607 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5612 inode = file_inode(vma->vm_file);
5613 dev = inode->i_sb->s_dev;
5615 gen = inode->i_generation;
5619 if (vma->vm_flags & VM_READ)
5621 if (vma->vm_flags & VM_WRITE)
5623 if (vma->vm_flags & VM_EXEC)
5626 if (vma->vm_flags & VM_MAYSHARE)
5629 flags = MAP_PRIVATE;
5631 if (vma->vm_flags & VM_DENYWRITE)
5632 flags |= MAP_DENYWRITE;
5633 if (vma->vm_flags & VM_MAYEXEC)
5634 flags |= MAP_EXECUTABLE;
5635 if (vma->vm_flags & VM_LOCKED)
5636 flags |= MAP_LOCKED;
5637 if (vma->vm_flags & VM_HUGETLB)
5638 flags |= MAP_HUGETLB;
5642 if (vma->vm_ops && vma->vm_ops->name) {
5643 name = (char *) vma->vm_ops->name(vma);
5648 name = (char *)arch_vma_name(vma);
5652 if (vma->vm_start <= vma->vm_mm->start_brk &&
5653 vma->vm_end >= vma->vm_mm->brk) {
5657 if (vma->vm_start <= vma->vm_mm->start_stack &&
5658 vma->vm_end >= vma->vm_mm->start_stack) {
5668 strlcpy(tmp, name, sizeof(tmp));
5672 * Since our buffer works in 8 byte units we need to align our string
5673 * size to a multiple of 8. However, we must guarantee the tail end is
5674 * zero'd out to avoid leaking random bits to userspace.
5676 size = strlen(name)+1;
5677 while (!IS_ALIGNED(size, sizeof(u64)))
5678 name[size++] = '\0';
5680 mmap_event->file_name = name;
5681 mmap_event->file_size = size;
5682 mmap_event->maj = maj;
5683 mmap_event->min = min;
5684 mmap_event->ino = ino;
5685 mmap_event->ino_generation = gen;
5686 mmap_event->prot = prot;
5687 mmap_event->flags = flags;
5689 if (!(vma->vm_flags & VM_EXEC))
5690 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5692 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5694 perf_event_aux(perf_event_mmap_output,
5701 void perf_event_mmap(struct vm_area_struct *vma)
5703 struct perf_mmap_event mmap_event;
5705 if (!atomic_read(&nr_mmap_events))
5708 mmap_event = (struct perf_mmap_event){
5714 .type = PERF_RECORD_MMAP,
5715 .misc = PERF_RECORD_MISC_USER,
5720 .start = vma->vm_start,
5721 .len = vma->vm_end - vma->vm_start,
5722 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5724 /* .maj (attr_mmap2 only) */
5725 /* .min (attr_mmap2 only) */
5726 /* .ino (attr_mmap2 only) */
5727 /* .ino_generation (attr_mmap2 only) */
5728 /* .prot (attr_mmap2 only) */
5729 /* .flags (attr_mmap2 only) */
5732 perf_event_mmap_event(&mmap_event);
5736 * IRQ throttle logging
5739 static void perf_log_throttle(struct perf_event *event, int enable)
5741 struct perf_output_handle handle;
5742 struct perf_sample_data sample;
5746 struct perf_event_header header;
5750 } throttle_event = {
5752 .type = PERF_RECORD_THROTTLE,
5754 .size = sizeof(throttle_event),
5756 .time = perf_clock(),
5757 .id = primary_event_id(event),
5758 .stream_id = event->id,
5762 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5764 perf_event_header__init_id(&throttle_event.header, &sample, event);
5766 ret = perf_output_begin(&handle, event,
5767 throttle_event.header.size);
5771 perf_output_put(&handle, throttle_event);
5772 perf_event__output_id_sample(event, &handle, &sample);
5773 perf_output_end(&handle);
5777 * Generic event overflow handling, sampling.
5780 static int __perf_event_overflow(struct perf_event *event,
5781 int throttle, struct perf_sample_data *data,
5782 struct pt_regs *regs)
5784 int events = atomic_read(&event->event_limit);
5785 struct hw_perf_event *hwc = &event->hw;
5790 * Non-sampling counters might still use the PMI to fold short
5791 * hardware counters, ignore those.
5793 if (unlikely(!is_sampling_event(event)))
5796 seq = __this_cpu_read(perf_throttled_seq);
5797 if (seq != hwc->interrupts_seq) {
5798 hwc->interrupts_seq = seq;
5799 hwc->interrupts = 1;
5802 if (unlikely(throttle
5803 && hwc->interrupts >= max_samples_per_tick)) {
5804 __this_cpu_inc(perf_throttled_count);
5805 hwc->interrupts = MAX_INTERRUPTS;
5806 perf_log_throttle(event, 0);
5807 tick_nohz_full_kick();
5812 if (event->attr.freq) {
5813 u64 now = perf_clock();
5814 s64 delta = now - hwc->freq_time_stamp;
5816 hwc->freq_time_stamp = now;
5818 if (delta > 0 && delta < 2*TICK_NSEC)
5819 perf_adjust_period(event, delta, hwc->last_period, true);
5823 * XXX event_limit might not quite work as expected on inherited
5827 event->pending_kill = POLL_IN;
5828 if (events && atomic_dec_and_test(&event->event_limit)) {
5830 event->pending_kill = POLL_HUP;
5831 event->pending_disable = 1;
5832 irq_work_queue(&event->pending);
5835 if (event->overflow_handler)
5836 event->overflow_handler(event, data, regs);
5838 perf_event_output(event, data, regs);
5840 if (event->fasync && event->pending_kill) {
5841 event->pending_wakeup = 1;
5842 irq_work_queue(&event->pending);
5848 int perf_event_overflow(struct perf_event *event,
5849 struct perf_sample_data *data,
5850 struct pt_regs *regs)
5852 return __perf_event_overflow(event, 1, data, regs);
5856 * Generic software event infrastructure
5859 struct swevent_htable {
5860 struct swevent_hlist *swevent_hlist;
5861 struct mutex hlist_mutex;
5864 /* Recursion avoidance in each contexts */
5865 int recursion[PERF_NR_CONTEXTS];
5867 /* Keeps track of cpu being initialized/exited */
5871 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5874 * We directly increment event->count and keep a second value in
5875 * event->hw.period_left to count intervals. This period event
5876 * is kept in the range [-sample_period, 0] so that we can use the
5880 u64 perf_swevent_set_period(struct perf_event *event)
5882 struct hw_perf_event *hwc = &event->hw;
5883 u64 period = hwc->last_period;
5887 hwc->last_period = hwc->sample_period;
5890 old = val = local64_read(&hwc->period_left);
5894 nr = div64_u64(period + val, period);
5895 offset = nr * period;
5897 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5903 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5904 struct perf_sample_data *data,
5905 struct pt_regs *regs)
5907 struct hw_perf_event *hwc = &event->hw;
5911 overflow = perf_swevent_set_period(event);
5913 if (hwc->interrupts == MAX_INTERRUPTS)
5916 for (; overflow; overflow--) {
5917 if (__perf_event_overflow(event, throttle,
5920 * We inhibit the overflow from happening when
5921 * hwc->interrupts == MAX_INTERRUPTS.
5929 static void perf_swevent_event(struct perf_event *event, u64 nr,
5930 struct perf_sample_data *data,
5931 struct pt_regs *regs)
5933 struct hw_perf_event *hwc = &event->hw;
5935 local64_add(nr, &event->count);
5940 if (!is_sampling_event(event))
5943 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5945 return perf_swevent_overflow(event, 1, data, regs);
5947 data->period = event->hw.last_period;
5949 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5950 return perf_swevent_overflow(event, 1, data, regs);
5952 if (local64_add_negative(nr, &hwc->period_left))
5955 perf_swevent_overflow(event, 0, data, regs);
5958 static int perf_exclude_event(struct perf_event *event,
5959 struct pt_regs *regs)
5961 if (event->hw.state & PERF_HES_STOPPED)
5965 if (event->attr.exclude_user && user_mode(regs))
5968 if (event->attr.exclude_kernel && !user_mode(regs))
5975 static int perf_swevent_match(struct perf_event *event,
5976 enum perf_type_id type,
5978 struct perf_sample_data *data,
5979 struct pt_regs *regs)
5981 if (event->attr.type != type)
5984 if (event->attr.config != event_id)
5987 if (perf_exclude_event(event, regs))
5993 static inline u64 swevent_hash(u64 type, u32 event_id)
5995 u64 val = event_id | (type << 32);
5997 return hash_64(val, SWEVENT_HLIST_BITS);
6000 static inline struct hlist_head *
6001 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
6003 u64 hash = swevent_hash(type, event_id);
6005 return &hlist->heads[hash];
6008 /* For the read side: events when they trigger */
6009 static inline struct hlist_head *
6010 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
6012 struct swevent_hlist *hlist;
6014 hlist = rcu_dereference(swhash->swevent_hlist);
6018 return __find_swevent_head(hlist, type, event_id);
6021 /* For the event head insertion and removal in the hlist */
6022 static inline struct hlist_head *
6023 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
6025 struct swevent_hlist *hlist;
6026 u32 event_id = event->attr.config;
6027 u64 type = event->attr.type;
6030 * Event scheduling is always serialized against hlist allocation
6031 * and release. Which makes the protected version suitable here.
6032 * The context lock guarantees that.
6034 hlist = rcu_dereference_protected(swhash->swevent_hlist,
6035 lockdep_is_held(&event->ctx->lock));
6039 return __find_swevent_head(hlist, type, event_id);
6042 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
6044 struct perf_sample_data *data,
6045 struct pt_regs *regs)
6047 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6048 struct perf_event *event;
6049 struct hlist_head *head;
6052 head = find_swevent_head_rcu(swhash, type, event_id);
6056 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6057 if (perf_swevent_match(event, type, event_id, data, regs))
6058 perf_swevent_event(event, nr, data, regs);
6064 DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);
6066 int perf_swevent_get_recursion_context(void)
6068 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6070 return get_recursion_context(swhash->recursion);
6072 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
6074 inline void perf_swevent_put_recursion_context(int rctx)
6076 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6078 put_recursion_context(swhash->recursion, rctx);
6081 void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6083 struct perf_sample_data data;
6085 if (WARN_ON_ONCE(!regs))
6088 perf_sample_data_init(&data, addr, 0);
6089 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
6092 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6096 preempt_disable_notrace();
6097 rctx = perf_swevent_get_recursion_context();
6098 if (unlikely(rctx < 0))
6101 ___perf_sw_event(event_id, nr, regs, addr);
6103 perf_swevent_put_recursion_context(rctx);
6105 preempt_enable_notrace();
6108 static void perf_swevent_read(struct perf_event *event)
6112 static int perf_swevent_add(struct perf_event *event, int flags)
6114 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6115 struct hw_perf_event *hwc = &event->hw;
6116 struct hlist_head *head;
6118 if (is_sampling_event(event)) {
6119 hwc->last_period = hwc->sample_period;
6120 perf_swevent_set_period(event);
6123 hwc->state = !(flags & PERF_EF_START);
6125 head = find_swevent_head(swhash, event);
6128 * We can race with cpu hotplug code. Do not
6129 * WARN if the cpu just got unplugged.
6131 WARN_ON_ONCE(swhash->online);
6135 hlist_add_head_rcu(&event->hlist_entry, head);
6140 static void perf_swevent_del(struct perf_event *event, int flags)
6142 hlist_del_rcu(&event->hlist_entry);
6145 static void perf_swevent_start(struct perf_event *event, int flags)
6147 event->hw.state = 0;
6150 static void perf_swevent_stop(struct perf_event *event, int flags)
6152 event->hw.state = PERF_HES_STOPPED;
6155 /* Deref the hlist from the update side */
6156 static inline struct swevent_hlist *
6157 swevent_hlist_deref(struct swevent_htable *swhash)
6159 return rcu_dereference_protected(swhash->swevent_hlist,
6160 lockdep_is_held(&swhash->hlist_mutex));
6163 static void swevent_hlist_release(struct swevent_htable *swhash)
6165 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
6170 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
6171 kfree_rcu(hlist, rcu_head);
6174 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
6176 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6178 mutex_lock(&swhash->hlist_mutex);
6180 if (!--swhash->hlist_refcount)
6181 swevent_hlist_release(swhash);
6183 mutex_unlock(&swhash->hlist_mutex);
6186 static void swevent_hlist_put(struct perf_event *event)
6190 for_each_possible_cpu(cpu)
6191 swevent_hlist_put_cpu(event, cpu);
6194 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6196 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6199 mutex_lock(&swhash->hlist_mutex);
6201 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6202 struct swevent_hlist *hlist;
6204 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6209 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6211 swhash->hlist_refcount++;
6213 mutex_unlock(&swhash->hlist_mutex);
6218 static int swevent_hlist_get(struct perf_event *event)
6221 int cpu, failed_cpu;
6224 for_each_possible_cpu(cpu) {
6225 err = swevent_hlist_get_cpu(event, cpu);
6235 for_each_possible_cpu(cpu) {
6236 if (cpu == failed_cpu)
6238 swevent_hlist_put_cpu(event, cpu);
6245 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6247 static void sw_perf_event_destroy(struct perf_event *event)
6249 u64 event_id = event->attr.config;
6251 WARN_ON(event->parent);
6253 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6254 swevent_hlist_put(event);
6257 static int perf_swevent_init(struct perf_event *event)
6259 u64 event_id = event->attr.config;
6261 if (event->attr.type != PERF_TYPE_SOFTWARE)
6265 * no branch sampling for software events
6267 if (has_branch_stack(event))
6271 case PERF_COUNT_SW_CPU_CLOCK:
6272 case PERF_COUNT_SW_TASK_CLOCK:
6279 if (event_id >= PERF_COUNT_SW_MAX)
6282 if (!event->parent) {
6285 err = swevent_hlist_get(event);
6289 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6290 event->destroy = sw_perf_event_destroy;
6296 static struct pmu perf_swevent = {
6297 .task_ctx_nr = perf_sw_context,
6299 .event_init = perf_swevent_init,
6300 .add = perf_swevent_add,
6301 .del = perf_swevent_del,
6302 .start = perf_swevent_start,
6303 .stop = perf_swevent_stop,
6304 .read = perf_swevent_read,
6307 #ifdef CONFIG_EVENT_TRACING
6309 static int perf_tp_filter_match(struct perf_event *event,
6310 struct perf_sample_data *data)
6312 void *record = data->raw->data;
6314 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6319 static int perf_tp_event_match(struct perf_event *event,
6320 struct perf_sample_data *data,
6321 struct pt_regs *regs)
6323 if (event->hw.state & PERF_HES_STOPPED)
6326 * All tracepoints are from kernel-space.
6328 if (event->attr.exclude_kernel)
6331 if (!perf_tp_filter_match(event, data))
6337 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6338 struct pt_regs *regs, struct hlist_head *head, int rctx,
6339 struct task_struct *task)
6341 struct perf_sample_data data;
6342 struct perf_event *event;
6344 struct perf_raw_record raw = {
6349 perf_sample_data_init(&data, addr, 0);
6352 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6353 if (perf_tp_event_match(event, &data, regs))
6354 perf_swevent_event(event, count, &data, regs);
6358 * If we got specified a target task, also iterate its context and
6359 * deliver this event there too.
6361 if (task && task != current) {
6362 struct perf_event_context *ctx;
6363 struct trace_entry *entry = record;
6366 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6370 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6371 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6373 if (event->attr.config != entry->type)
6375 if (perf_tp_event_match(event, &data, regs))
6376 perf_swevent_event(event, count, &data, regs);
6382 perf_swevent_put_recursion_context(rctx);
6384 EXPORT_SYMBOL_GPL(perf_tp_event);
6386 static void tp_perf_event_destroy(struct perf_event *event)
6388 perf_trace_destroy(event);
6391 static int perf_tp_event_init(struct perf_event *event)
6395 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6399 * no branch sampling for tracepoint events
6401 if (has_branch_stack(event))
6404 err = perf_trace_init(event);
6408 event->destroy = tp_perf_event_destroy;
6413 static struct pmu perf_tracepoint = {
6414 .task_ctx_nr = perf_sw_context,
6416 .event_init = perf_tp_event_init,
6417 .add = perf_trace_add,
6418 .del = perf_trace_del,
6419 .start = perf_swevent_start,
6420 .stop = perf_swevent_stop,
6421 .read = perf_swevent_read,
6424 static inline void perf_tp_register(void)
6426 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6429 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6434 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6437 filter_str = strndup_user(arg, PAGE_SIZE);
6438 if (IS_ERR(filter_str))
6439 return PTR_ERR(filter_str);
6441 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6447 static void perf_event_free_filter(struct perf_event *event)
6449 ftrace_profile_free_filter(event);
6454 static inline void perf_tp_register(void)
6458 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6463 static void perf_event_free_filter(struct perf_event *event)
6467 #endif /* CONFIG_EVENT_TRACING */
6469 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6470 void perf_bp_event(struct perf_event *bp, void *data)
6472 struct perf_sample_data sample;
6473 struct pt_regs *regs = data;
6475 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6477 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6478 perf_swevent_event(bp, 1, &sample, regs);
6483 * hrtimer based swevent callback
6486 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6488 enum hrtimer_restart ret = HRTIMER_RESTART;
6489 struct perf_sample_data data;
6490 struct pt_regs *regs;
6491 struct perf_event *event;
6494 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6496 if (event->state != PERF_EVENT_STATE_ACTIVE)
6497 return HRTIMER_NORESTART;
6499 event->pmu->read(event);
6501 perf_sample_data_init(&data, 0, event->hw.last_period);
6502 regs = get_irq_regs();
6504 if (regs && !perf_exclude_event(event, regs)) {
6505 if (!(event->attr.exclude_idle && is_idle_task(current)))
6506 if (__perf_event_overflow(event, 1, &data, regs))
6507 ret = HRTIMER_NORESTART;
6510 period = max_t(u64, 10000, event->hw.sample_period);
6511 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6516 static void perf_swevent_start_hrtimer(struct perf_event *event)
6518 struct hw_perf_event *hwc = &event->hw;
6521 if (!is_sampling_event(event))
6524 period = local64_read(&hwc->period_left);
6529 local64_set(&hwc->period_left, 0);
6531 period = max_t(u64, 10000, hwc->sample_period);
6533 __hrtimer_start_range_ns(&hwc->hrtimer,
6534 ns_to_ktime(period), 0,
6535 HRTIMER_MODE_REL_PINNED, 0);
6538 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6540 struct hw_perf_event *hwc = &event->hw;
6542 if (is_sampling_event(event)) {
6543 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6544 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6546 hrtimer_cancel(&hwc->hrtimer);
6550 static void perf_swevent_init_hrtimer(struct perf_event *event)
6552 struct hw_perf_event *hwc = &event->hw;
6554 if (!is_sampling_event(event))
6557 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6558 hwc->hrtimer.function = perf_swevent_hrtimer;
6561 * Since hrtimers have a fixed rate, we can do a static freq->period
6562 * mapping and avoid the whole period adjust feedback stuff.
6564 if (event->attr.freq) {
6565 long freq = event->attr.sample_freq;
6567 event->attr.sample_period = NSEC_PER_SEC / freq;
6568 hwc->sample_period = event->attr.sample_period;
6569 local64_set(&hwc->period_left, hwc->sample_period);
6570 hwc->last_period = hwc->sample_period;
6571 event->attr.freq = 0;
6576 * Software event: cpu wall time clock
6579 static void cpu_clock_event_update(struct perf_event *event)
6584 now = local_clock();
6585 prev = local64_xchg(&event->hw.prev_count, now);
6586 local64_add(now - prev, &event->count);
6589 static void cpu_clock_event_start(struct perf_event *event, int flags)
6591 local64_set(&event->hw.prev_count, local_clock());
6592 perf_swevent_start_hrtimer(event);
6595 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6597 perf_swevent_cancel_hrtimer(event);
6598 cpu_clock_event_update(event);
6601 static int cpu_clock_event_add(struct perf_event *event, int flags)
6603 if (flags & PERF_EF_START)
6604 cpu_clock_event_start(event, flags);
6609 static void cpu_clock_event_del(struct perf_event *event, int flags)
6611 cpu_clock_event_stop(event, flags);
6614 static void cpu_clock_event_read(struct perf_event *event)
6616 cpu_clock_event_update(event);
6619 static int cpu_clock_event_init(struct perf_event *event)
6621 if (event->attr.type != PERF_TYPE_SOFTWARE)
6624 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6628 * no branch sampling for software events
6630 if (has_branch_stack(event))
6633 perf_swevent_init_hrtimer(event);
6638 static struct pmu perf_cpu_clock = {
6639 .task_ctx_nr = perf_sw_context,
6641 .event_init = cpu_clock_event_init,
6642 .add = cpu_clock_event_add,
6643 .del = cpu_clock_event_del,
6644 .start = cpu_clock_event_start,
6645 .stop = cpu_clock_event_stop,
6646 .read = cpu_clock_event_read,
6650 * Software event: task time clock
6653 static void task_clock_event_update(struct perf_event *event, u64 now)
6658 prev = local64_xchg(&event->hw.prev_count, now);
6660 local64_add(delta, &event->count);
6663 static void task_clock_event_start(struct perf_event *event, int flags)
6665 local64_set(&event->hw.prev_count, event->ctx->time);
6666 perf_swevent_start_hrtimer(event);
6669 static void task_clock_event_stop(struct perf_event *event, int flags)
6671 perf_swevent_cancel_hrtimer(event);
6672 task_clock_event_update(event, event->ctx->time);
6675 static int task_clock_event_add(struct perf_event *event, int flags)
6677 if (flags & PERF_EF_START)
6678 task_clock_event_start(event, flags);
6683 static void task_clock_event_del(struct perf_event *event, int flags)
6685 task_clock_event_stop(event, PERF_EF_UPDATE);
6688 static void task_clock_event_read(struct perf_event *event)
6690 u64 now = perf_clock();
6691 u64 delta = now - event->ctx->timestamp;
6692 u64 time = event->ctx->time + delta;
6694 task_clock_event_update(event, time);
6697 static int task_clock_event_init(struct perf_event *event)
6699 if (event->attr.type != PERF_TYPE_SOFTWARE)
6702 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6706 * no branch sampling for software events
6708 if (has_branch_stack(event))
6711 perf_swevent_init_hrtimer(event);
6716 static struct pmu perf_task_clock = {
6717 .task_ctx_nr = perf_sw_context,
6719 .event_init = task_clock_event_init,
6720 .add = task_clock_event_add,
6721 .del = task_clock_event_del,
6722 .start = task_clock_event_start,
6723 .stop = task_clock_event_stop,
6724 .read = task_clock_event_read,
6727 static void perf_pmu_nop_void(struct pmu *pmu)
6731 static int perf_pmu_nop_int(struct pmu *pmu)
6736 static void perf_pmu_start_txn(struct pmu *pmu)
6738 perf_pmu_disable(pmu);
6741 static int perf_pmu_commit_txn(struct pmu *pmu)
6743 perf_pmu_enable(pmu);
6747 static void perf_pmu_cancel_txn(struct pmu *pmu)
6749 perf_pmu_enable(pmu);
6752 static int perf_event_idx_default(struct perf_event *event)
6758 * Ensures all contexts with the same task_ctx_nr have the same
6759 * pmu_cpu_context too.
6761 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6768 list_for_each_entry(pmu, &pmus, entry) {
6769 if (pmu->task_ctx_nr == ctxn)
6770 return pmu->pmu_cpu_context;
6776 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6780 for_each_possible_cpu(cpu) {
6781 struct perf_cpu_context *cpuctx;
6783 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6785 if (cpuctx->unique_pmu == old_pmu)
6786 cpuctx->unique_pmu = pmu;
6790 static void free_pmu_context(struct pmu *pmu)
6794 mutex_lock(&pmus_lock);
6796 * Like a real lame refcount.
6798 list_for_each_entry(i, &pmus, entry) {
6799 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6800 update_pmu_context(i, pmu);
6805 free_percpu(pmu->pmu_cpu_context);
6807 mutex_unlock(&pmus_lock);
6809 static struct idr pmu_idr;
6812 type_show(struct device *dev, struct device_attribute *attr, char *page)
6814 struct pmu *pmu = dev_get_drvdata(dev);
6816 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6818 static DEVICE_ATTR_RO(type);
6821 perf_event_mux_interval_ms_show(struct device *dev,
6822 struct device_attribute *attr,
6825 struct pmu *pmu = dev_get_drvdata(dev);
6827 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6831 perf_event_mux_interval_ms_store(struct device *dev,
6832 struct device_attribute *attr,
6833 const char *buf, size_t count)
6835 struct pmu *pmu = dev_get_drvdata(dev);
6836 int timer, cpu, ret;
6838 ret = kstrtoint(buf, 0, &timer);
6845 /* same value, noting to do */
6846 if (timer == pmu->hrtimer_interval_ms)
6849 pmu->hrtimer_interval_ms = timer;
6851 /* update all cpuctx for this PMU */
6852 for_each_possible_cpu(cpu) {
6853 struct perf_cpu_context *cpuctx;
6854 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6855 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6857 if (hrtimer_active(&cpuctx->hrtimer))
6858 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6863 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6865 static struct attribute *pmu_dev_attrs[] = {
6866 &dev_attr_type.attr,
6867 &dev_attr_perf_event_mux_interval_ms.attr,
6870 ATTRIBUTE_GROUPS(pmu_dev);
6872 static int pmu_bus_running;
6873 static struct bus_type pmu_bus = {
6874 .name = "event_source",
6875 .dev_groups = pmu_dev_groups,
6878 static void pmu_dev_release(struct device *dev)
6883 static int pmu_dev_alloc(struct pmu *pmu)
6887 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6891 pmu->dev->groups = pmu->attr_groups;
6892 device_initialize(pmu->dev);
6893 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6897 dev_set_drvdata(pmu->dev, pmu);
6898 pmu->dev->bus = &pmu_bus;
6899 pmu->dev->release = pmu_dev_release;
6900 ret = device_add(pmu->dev);
6908 put_device(pmu->dev);
6912 static struct lock_class_key cpuctx_mutex;
6913 static struct lock_class_key cpuctx_lock;
6915 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6919 mutex_lock(&pmus_lock);
6921 pmu->pmu_disable_count = alloc_percpu(int);
6922 if (!pmu->pmu_disable_count)
6931 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6939 if (pmu_bus_running) {
6940 ret = pmu_dev_alloc(pmu);
6946 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6947 if (pmu->pmu_cpu_context)
6948 goto got_cpu_context;
6951 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6952 if (!pmu->pmu_cpu_context)
6955 for_each_possible_cpu(cpu) {
6956 struct perf_cpu_context *cpuctx;
6958 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6959 __perf_event_init_context(&cpuctx->ctx);
6960 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6961 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6962 cpuctx->ctx.pmu = pmu;
6964 __perf_cpu_hrtimer_init(cpuctx, cpu);
6966 cpuctx->unique_pmu = pmu;
6970 if (!pmu->start_txn) {
6971 if (pmu->pmu_enable) {
6973 * If we have pmu_enable/pmu_disable calls, install
6974 * transaction stubs that use that to try and batch
6975 * hardware accesses.
6977 pmu->start_txn = perf_pmu_start_txn;
6978 pmu->commit_txn = perf_pmu_commit_txn;
6979 pmu->cancel_txn = perf_pmu_cancel_txn;
6981 pmu->start_txn = perf_pmu_nop_void;
6982 pmu->commit_txn = perf_pmu_nop_int;
6983 pmu->cancel_txn = perf_pmu_nop_void;
6987 if (!pmu->pmu_enable) {
6988 pmu->pmu_enable = perf_pmu_nop_void;
6989 pmu->pmu_disable = perf_pmu_nop_void;
6992 if (!pmu->event_idx)
6993 pmu->event_idx = perf_event_idx_default;
6995 list_add_rcu(&pmu->entry, &pmus);
6998 mutex_unlock(&pmus_lock);
7003 device_del(pmu->dev);
7004 put_device(pmu->dev);
7007 if (pmu->type >= PERF_TYPE_MAX)
7008 idr_remove(&pmu_idr, pmu->type);
7011 free_percpu(pmu->pmu_disable_count);
7014 EXPORT_SYMBOL_GPL(perf_pmu_register);
7016 void perf_pmu_unregister(struct pmu *pmu)
7018 mutex_lock(&pmus_lock);
7019 list_del_rcu(&pmu->entry);
7020 mutex_unlock(&pmus_lock);
7023 * We dereference the pmu list under both SRCU and regular RCU, so
7024 * synchronize against both of those.
7026 synchronize_srcu(&pmus_srcu);
7029 free_percpu(pmu->pmu_disable_count);
7030 if (pmu->type >= PERF_TYPE_MAX)
7031 idr_remove(&pmu_idr, pmu->type);
7032 device_del(pmu->dev);
7033 put_device(pmu->dev);
7034 free_pmu_context(pmu);
7036 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
7038 static int perf_try_init_event(struct pmu *pmu, struct perf_event *event)
7042 if (!try_module_get(pmu->module))
7045 ret = pmu->event_init(event);
7047 module_put(pmu->module);
7052 struct pmu *perf_init_event(struct perf_event *event)
7054 struct pmu *pmu = NULL;
7058 idx = srcu_read_lock(&pmus_srcu);
7061 pmu = idr_find(&pmu_idr, event->attr.type);
7064 ret = perf_try_init_event(pmu, event);
7070 list_for_each_entry_rcu(pmu, &pmus, entry) {
7071 ret = perf_try_init_event(pmu, event);
7075 if (ret != -ENOENT) {
7080 pmu = ERR_PTR(-ENOENT);
7082 srcu_read_unlock(&pmus_srcu, idx);
7087 static void account_event_cpu(struct perf_event *event, int cpu)
7092 if (has_branch_stack(event)) {
7093 if (!(event->attach_state & PERF_ATTACH_TASK))
7094 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
7096 if (is_cgroup_event(event))
7097 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
7100 static void account_event(struct perf_event *event)
7105 if (event->attach_state & PERF_ATTACH_TASK)
7106 static_key_slow_inc(&perf_sched_events.key);
7107 if (event->attr.mmap || event->attr.mmap_data)
7108 atomic_inc(&nr_mmap_events);
7109 if (event->attr.comm)
7110 atomic_inc(&nr_comm_events);
7111 if (event->attr.task)
7112 atomic_inc(&nr_task_events);
7113 if (event->attr.freq) {
7114 if (atomic_inc_return(&nr_freq_events) == 1)
7115 tick_nohz_full_kick_all();
7117 if (has_branch_stack(event))
7118 static_key_slow_inc(&perf_sched_events.key);
7119 if (is_cgroup_event(event))
7120 static_key_slow_inc(&perf_sched_events.key);
7122 account_event_cpu(event, event->cpu);
7126 * Allocate and initialize a event structure
7128 static struct perf_event *
7129 perf_event_alloc(struct perf_event_attr *attr, int cpu,
7130 struct task_struct *task,
7131 struct perf_event *group_leader,
7132 struct perf_event *parent_event,
7133 perf_overflow_handler_t overflow_handler,
7137 struct perf_event *event;
7138 struct hw_perf_event *hwc;
7141 if ((unsigned)cpu >= nr_cpu_ids) {
7142 if (!task || cpu != -1)
7143 return ERR_PTR(-EINVAL);
7146 event = kzalloc(sizeof(*event), GFP_KERNEL);
7148 return ERR_PTR(-ENOMEM);
7151 * Single events are their own group leaders, with an
7152 * empty sibling list:
7155 group_leader = event;
7157 mutex_init(&event->child_mutex);
7158 INIT_LIST_HEAD(&event->child_list);
7160 INIT_LIST_HEAD(&event->group_entry);
7161 INIT_LIST_HEAD(&event->event_entry);
7162 INIT_LIST_HEAD(&event->sibling_list);
7163 INIT_LIST_HEAD(&event->rb_entry);
7164 INIT_LIST_HEAD(&event->active_entry);
7165 INIT_HLIST_NODE(&event->hlist_entry);
7168 init_waitqueue_head(&event->waitq);
7169 init_irq_work(&event->pending, perf_pending_event);
7171 mutex_init(&event->mmap_mutex);
7173 atomic_long_set(&event->refcount, 1);
7175 event->attr = *attr;
7176 event->group_leader = group_leader;
7180 event->parent = parent_event;
7182 event->ns = get_pid_ns(task_active_pid_ns(current));
7183 event->id = atomic64_inc_return(&perf_event_id);
7185 event->state = PERF_EVENT_STATE_INACTIVE;
7188 event->attach_state = PERF_ATTACH_TASK;
7190 if (attr->type == PERF_TYPE_TRACEPOINT)
7191 event->hw.tp_target = task;
7192 #ifdef CONFIG_HAVE_HW_BREAKPOINT
7194 * hw_breakpoint is a bit difficult here..
7196 else if (attr->type == PERF_TYPE_BREAKPOINT)
7197 event->hw.bp_target = task;
7201 if (!overflow_handler && parent_event) {
7202 overflow_handler = parent_event->overflow_handler;
7203 context = parent_event->overflow_handler_context;
7206 event->overflow_handler = overflow_handler;
7207 event->overflow_handler_context = context;
7209 perf_event__state_init(event);
7214 hwc->sample_period = attr->sample_period;
7215 if (attr->freq && attr->sample_freq)
7216 hwc->sample_period = 1;
7217 hwc->last_period = hwc->sample_period;
7219 local64_set(&hwc->period_left, hwc->sample_period);
7222 * we currently do not support PERF_FORMAT_GROUP on inherited events
7224 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7227 pmu = perf_init_event(event);
7230 else if (IS_ERR(pmu)) {
7235 if (!event->parent) {
7236 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7237 err = get_callchain_buffers();
7247 event->destroy(event);
7248 module_put(pmu->module);
7251 put_pid_ns(event->ns);
7254 return ERR_PTR(err);
7257 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7258 struct perf_event_attr *attr)
7263 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7267 * zero the full structure, so that a short copy will be nice.
7269 memset(attr, 0, sizeof(*attr));
7271 ret = get_user(size, &uattr->size);
7275 if (size > PAGE_SIZE) /* silly large */
7278 if (!size) /* abi compat */
7279 size = PERF_ATTR_SIZE_VER0;
7281 if (size < PERF_ATTR_SIZE_VER0)
7285 * If we're handed a bigger struct than we know of,
7286 * ensure all the unknown bits are 0 - i.e. new
7287 * user-space does not rely on any kernel feature
7288 * extensions we dont know about yet.
7290 if (size > sizeof(*attr)) {
7291 unsigned char __user *addr;
7292 unsigned char __user *end;
7295 addr = (void __user *)uattr + sizeof(*attr);
7296 end = (void __user *)uattr + size;
7298 for (; addr < end; addr++) {
7299 ret = get_user(val, addr);
7305 size = sizeof(*attr);
7308 ret = copy_from_user(attr, uattr, size);
7312 if (attr->__reserved_1)
7315 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7318 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7321 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7322 u64 mask = attr->branch_sample_type;
7324 /* only using defined bits */
7325 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7328 /* at least one branch bit must be set */
7329 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7332 /* propagate priv level, when not set for branch */
7333 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7335 /* exclude_kernel checked on syscall entry */
7336 if (!attr->exclude_kernel)
7337 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7339 if (!attr->exclude_user)
7340 mask |= PERF_SAMPLE_BRANCH_USER;
7342 if (!attr->exclude_hv)
7343 mask |= PERF_SAMPLE_BRANCH_HV;
7345 * adjust user setting (for HW filter setup)
7347 attr->branch_sample_type = mask;
7349 /* privileged levels capture (kernel, hv): check permissions */
7350 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7351 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7355 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7356 ret = perf_reg_validate(attr->sample_regs_user);
7361 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7362 if (!arch_perf_have_user_stack_dump())
7366 * We have __u32 type for the size, but so far
7367 * we can only use __u16 as maximum due to the
7368 * __u16 sample size limit.
7370 if (attr->sample_stack_user >= USHRT_MAX)
7372 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7376 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7377 ret = perf_reg_validate(attr->sample_regs_intr);
7382 put_user(sizeof(*attr), &uattr->size);
7388 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7390 struct ring_buffer *rb = NULL;
7396 /* don't allow circular references */
7397 if (event == output_event)
7401 * Don't allow cross-cpu buffers
7403 if (output_event->cpu != event->cpu)
7407 * If its not a per-cpu rb, it must be the same task.
7409 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7413 mutex_lock(&event->mmap_mutex);
7414 /* Can't redirect output if we've got an active mmap() */
7415 if (atomic_read(&event->mmap_count))
7419 /* get the rb we want to redirect to */
7420 rb = ring_buffer_get(output_event);
7425 ring_buffer_attach(event, rb);
7429 mutex_unlock(&event->mmap_mutex);
7435 static void mutex_lock_double(struct mutex *a, struct mutex *b)
7441 mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
7445 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7447 * @attr_uptr: event_id type attributes for monitoring/sampling
7450 * @group_fd: group leader event fd
7452 SYSCALL_DEFINE5(perf_event_open,
7453 struct perf_event_attr __user *, attr_uptr,
7454 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7456 struct perf_event *group_leader = NULL, *output_event = NULL;
7457 struct perf_event *event, *sibling;
7458 struct perf_event_attr attr;
7459 struct perf_event_context *ctx, *uninitialized_var(gctx);
7460 struct file *event_file = NULL;
7461 struct fd group = {NULL, 0};
7462 struct task_struct *task = NULL;
7467 int f_flags = O_RDWR;
7469 /* for future expandability... */
7470 if (flags & ~PERF_FLAG_ALL)
7473 err = perf_copy_attr(attr_uptr, &attr);
7477 if (!attr.exclude_kernel) {
7478 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7483 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7486 if (attr.sample_period & (1ULL << 63))
7491 * In cgroup mode, the pid argument is used to pass the fd
7492 * opened to the cgroup directory in cgroupfs. The cpu argument
7493 * designates the cpu on which to monitor threads from that
7496 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7499 if (flags & PERF_FLAG_FD_CLOEXEC)
7500 f_flags |= O_CLOEXEC;
7502 event_fd = get_unused_fd_flags(f_flags);
7506 if (group_fd != -1) {
7507 err = perf_fget_light(group_fd, &group);
7510 group_leader = group.file->private_data;
7511 if (flags & PERF_FLAG_FD_OUTPUT)
7512 output_event = group_leader;
7513 if (flags & PERF_FLAG_FD_NO_GROUP)
7514 group_leader = NULL;
7517 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7518 task = find_lively_task_by_vpid(pid);
7520 err = PTR_ERR(task);
7525 if (task && group_leader &&
7526 group_leader->attr.inherit != attr.inherit) {
7533 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7535 if (IS_ERR(event)) {
7536 err = PTR_ERR(event);
7540 if (flags & PERF_FLAG_PID_CGROUP) {
7541 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7543 __free_event(event);
7548 if (is_sampling_event(event)) {
7549 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7555 account_event(event);
7558 * Special case software events and allow them to be part of
7559 * any hardware group.
7564 (is_software_event(event) != is_software_event(group_leader))) {
7565 if (is_software_event(event)) {
7567 * If event and group_leader are not both a software
7568 * event, and event is, then group leader is not.
7570 * Allow the addition of software events to !software
7571 * groups, this is safe because software events never
7574 pmu = group_leader->pmu;
7575 } else if (is_software_event(group_leader) &&
7576 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7578 * In case the group is a pure software group, and we
7579 * try to add a hardware event, move the whole group to
7580 * the hardware context.
7587 * Get the target context (task or percpu):
7589 ctx = find_get_context(pmu, task, event->cpu);
7596 put_task_struct(task);
7601 * Look up the group leader (we will attach this event to it):
7607 * Do not allow a recursive hierarchy (this new sibling
7608 * becoming part of another group-sibling):
7610 if (group_leader->group_leader != group_leader)
7613 * Do not allow to attach to a group in a different
7614 * task or CPU context:
7618 * Make sure we're both on the same task, or both
7621 if (group_leader->ctx->task != ctx->task)
7625 * Make sure we're both events for the same CPU;
7626 * grouping events for different CPUs is broken; since
7627 * you can never concurrently schedule them anyhow.
7629 if (group_leader->cpu != event->cpu)
7632 if (group_leader->ctx != ctx)
7637 * Only a group leader can be exclusive or pinned
7639 if (attr.exclusive || attr.pinned)
7644 err = perf_event_set_output(event, output_event);
7649 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7651 if (IS_ERR(event_file)) {
7652 err = PTR_ERR(event_file);
7657 gctx = group_leader->ctx;
7660 * See perf_event_ctx_lock() for comments on the details
7661 * of swizzling perf_event::ctx.
7663 mutex_lock_double(&gctx->mutex, &ctx->mutex);
7665 perf_remove_from_context(group_leader, false);
7667 list_for_each_entry(sibling, &group_leader->sibling_list,
7669 perf_remove_from_context(sibling, false);
7673 mutex_lock(&ctx->mutex);
7676 WARN_ON_ONCE(ctx->parent_ctx);
7680 * Wait for everybody to stop referencing the events through
7681 * the old lists, before installing it on new lists.
7686 * Install the group siblings before the group leader.
7688 * Because a group leader will try and install the entire group
7689 * (through the sibling list, which is still in-tact), we can
7690 * end up with siblings installed in the wrong context.
7692 * By installing siblings first we NO-OP because they're not
7693 * reachable through the group lists.
7695 list_for_each_entry(sibling, &group_leader->sibling_list,
7697 perf_event__state_init(sibling);
7698 perf_install_in_context(ctx, sibling, sibling->cpu);
7703 * Removing from the context ends up with disabled
7704 * event. What we want here is event in the initial
7705 * startup state, ready to be add into new context.
7707 perf_event__state_init(group_leader);
7708 perf_install_in_context(ctx, group_leader, group_leader->cpu);
7712 perf_install_in_context(ctx, event, event->cpu);
7713 perf_unpin_context(ctx);
7716 mutex_unlock(&gctx->mutex);
7719 mutex_unlock(&ctx->mutex);
7723 event->owner = current;
7725 mutex_lock(¤t->perf_event_mutex);
7726 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7727 mutex_unlock(¤t->perf_event_mutex);
7730 * Precalculate sample_data sizes
7732 perf_event__header_size(event);
7733 perf_event__id_header_size(event);
7736 * Drop the reference on the group_event after placing the
7737 * new event on the sibling_list. This ensures destruction
7738 * of the group leader will find the pointer to itself in
7739 * perf_group_detach().
7742 fd_install(event_fd, event_file);
7746 perf_unpin_context(ctx);
7754 put_task_struct(task);
7758 put_unused_fd(event_fd);
7763 * perf_event_create_kernel_counter
7765 * @attr: attributes of the counter to create
7766 * @cpu: cpu in which the counter is bound
7767 * @task: task to profile (NULL for percpu)
7770 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7771 struct task_struct *task,
7772 perf_overflow_handler_t overflow_handler,
7775 struct perf_event_context *ctx;
7776 struct perf_event *event;
7780 * Get the target context (task or percpu):
7783 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7784 overflow_handler, context);
7785 if (IS_ERR(event)) {
7786 err = PTR_ERR(event);
7790 /* Mark owner so we could distinguish it from user events. */
7791 event->owner = EVENT_OWNER_KERNEL;
7793 account_event(event);
7795 ctx = find_get_context(event->pmu, task, cpu);
7801 WARN_ON_ONCE(ctx->parent_ctx);
7802 mutex_lock(&ctx->mutex);
7803 perf_install_in_context(ctx, event, cpu);
7804 perf_unpin_context(ctx);
7805 mutex_unlock(&ctx->mutex);
7812 return ERR_PTR(err);
7814 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7816 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7818 struct perf_event_context *src_ctx;
7819 struct perf_event_context *dst_ctx;
7820 struct perf_event *event, *tmp;
7823 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7824 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7827 * See perf_event_ctx_lock() for comments on the details
7828 * of swizzling perf_event::ctx.
7830 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex);
7831 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7833 perf_remove_from_context(event, false);
7834 unaccount_event_cpu(event, src_cpu);
7836 list_add(&event->migrate_entry, &events);
7840 * Wait for the events to quiesce before re-instating them.
7845 * Re-instate events in 2 passes.
7847 * Skip over group leaders and only install siblings on this first
7848 * pass, siblings will not get enabled without a leader, however a
7849 * leader will enable its siblings, even if those are still on the old
7852 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7853 if (event->group_leader == event)
7856 list_del(&event->migrate_entry);
7857 if (event->state >= PERF_EVENT_STATE_OFF)
7858 event->state = PERF_EVENT_STATE_INACTIVE;
7859 account_event_cpu(event, dst_cpu);
7860 perf_install_in_context(dst_ctx, event, dst_cpu);
7865 * Once all the siblings are setup properly, install the group leaders
7868 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7869 list_del(&event->migrate_entry);
7870 if (event->state >= PERF_EVENT_STATE_OFF)
7871 event->state = PERF_EVENT_STATE_INACTIVE;
7872 account_event_cpu(event, dst_cpu);
7873 perf_install_in_context(dst_ctx, event, dst_cpu);
7876 mutex_unlock(&dst_ctx->mutex);
7877 mutex_unlock(&src_ctx->mutex);
7879 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7881 static void sync_child_event(struct perf_event *child_event,
7882 struct task_struct *child)
7884 struct perf_event *parent_event = child_event->parent;
7887 if (child_event->attr.inherit_stat)
7888 perf_event_read_event(child_event, child);
7890 child_val = perf_event_count(child_event);
7893 * Add back the child's count to the parent's count:
7895 atomic64_add(child_val, &parent_event->child_count);
7896 atomic64_add(child_event->total_time_enabled,
7897 &parent_event->child_total_time_enabled);
7898 atomic64_add(child_event->total_time_running,
7899 &parent_event->child_total_time_running);
7902 * Remove this event from the parent's list
7904 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7905 mutex_lock(&parent_event->child_mutex);
7906 list_del_init(&child_event->child_list);
7907 mutex_unlock(&parent_event->child_mutex);
7910 * Make sure user/parent get notified, that we just
7913 perf_event_wakeup(parent_event);
7916 * Release the parent event, if this was the last
7919 put_event(parent_event);
7923 __perf_event_exit_task(struct perf_event *child_event,
7924 struct perf_event_context *child_ctx,
7925 struct task_struct *child)
7928 * Do not destroy the 'original' grouping; because of the context
7929 * switch optimization the original events could've ended up in a
7930 * random child task.
7932 * If we were to destroy the original group, all group related
7933 * operations would cease to function properly after this random
7936 * Do destroy all inherited groups, we don't care about those
7937 * and being thorough is better.
7939 perf_remove_from_context(child_event, !!child_event->parent);
7942 * It can happen that the parent exits first, and has events
7943 * that are still around due to the child reference. These
7944 * events need to be zapped.
7946 if (child_event->parent) {
7947 sync_child_event(child_event, child);
7948 free_event(child_event);
7950 child_event->state = PERF_EVENT_STATE_EXIT;
7951 perf_event_wakeup(child_event);
7955 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7957 struct perf_event *child_event, *next;
7958 struct perf_event_context *child_ctx, *clone_ctx = NULL;
7959 unsigned long flags;
7961 if (likely(!child->perf_event_ctxp[ctxn])) {
7962 perf_event_task(child, NULL, 0);
7966 local_irq_save(flags);
7968 * We can't reschedule here because interrupts are disabled,
7969 * and either child is current or it is a task that can't be
7970 * scheduled, so we are now safe from rescheduling changing
7973 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7976 * Take the context lock here so that if find_get_context is
7977 * reading child->perf_event_ctxp, we wait until it has
7978 * incremented the context's refcount before we do put_ctx below.
7980 raw_spin_lock(&child_ctx->lock);
7981 task_ctx_sched_out(child_ctx);
7982 child->perf_event_ctxp[ctxn] = NULL;
7985 * If this context is a clone; unclone it so it can't get
7986 * swapped to another process while we're removing all
7987 * the events from it.
7989 clone_ctx = unclone_ctx(child_ctx);
7990 update_context_time(child_ctx);
7991 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7997 * Report the task dead after unscheduling the events so that we
7998 * won't get any samples after PERF_RECORD_EXIT. We can however still
7999 * get a few PERF_RECORD_READ events.
8001 perf_event_task(child, child_ctx, 0);
8004 * We can recurse on the same lock type through:
8006 * __perf_event_exit_task()
8007 * sync_child_event()
8009 * mutex_lock(&ctx->mutex)
8011 * But since its the parent context it won't be the same instance.
8013 mutex_lock(&child_ctx->mutex);
8015 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
8016 __perf_event_exit_task(child_event, child_ctx, child);
8018 mutex_unlock(&child_ctx->mutex);
8024 * When a child task exits, feed back event values to parent events.
8026 void perf_event_exit_task(struct task_struct *child)
8028 struct perf_event *event, *tmp;
8031 mutex_lock(&child->perf_event_mutex);
8032 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
8034 list_del_init(&event->owner_entry);
8037 * Ensure the list deletion is visible before we clear
8038 * the owner, closes a race against perf_release() where
8039 * we need to serialize on the owner->perf_event_mutex.
8042 event->owner = NULL;
8044 mutex_unlock(&child->perf_event_mutex);
8046 for_each_task_context_nr(ctxn)
8047 perf_event_exit_task_context(child, ctxn);
8050 static void perf_free_event(struct perf_event *event,
8051 struct perf_event_context *ctx)
8053 struct perf_event *parent = event->parent;
8055 if (WARN_ON_ONCE(!parent))
8058 mutex_lock(&parent->child_mutex);
8059 list_del_init(&event->child_list);
8060 mutex_unlock(&parent->child_mutex);
8064 raw_spin_lock_irq(&ctx->lock);
8065 perf_group_detach(event);
8066 list_del_event(event, ctx);
8067 raw_spin_unlock_irq(&ctx->lock);
8072 * Free an unexposed, unused context as created by inheritance by
8073 * perf_event_init_task below, used by fork() in case of fail.
8075 * Not all locks are strictly required, but take them anyway to be nice and
8076 * help out with the lockdep assertions.
8078 void perf_event_free_task(struct task_struct *task)
8080 struct perf_event_context *ctx;
8081 struct perf_event *event, *tmp;
8084 for_each_task_context_nr(ctxn) {
8085 ctx = task->perf_event_ctxp[ctxn];
8089 mutex_lock(&ctx->mutex);
8091 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
8093 perf_free_event(event, ctx);
8095 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
8097 perf_free_event(event, ctx);
8099 if (!list_empty(&ctx->pinned_groups) ||
8100 !list_empty(&ctx->flexible_groups))
8103 mutex_unlock(&ctx->mutex);
8109 void perf_event_delayed_put(struct task_struct *task)
8113 for_each_task_context_nr(ctxn)
8114 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
8118 * inherit a event from parent task to child task:
8120 static struct perf_event *
8121 inherit_event(struct perf_event *parent_event,
8122 struct task_struct *parent,
8123 struct perf_event_context *parent_ctx,
8124 struct task_struct *child,
8125 struct perf_event *group_leader,
8126 struct perf_event_context *child_ctx)
8128 enum perf_event_active_state parent_state = parent_event->state;
8129 struct perf_event *child_event;
8130 unsigned long flags;
8133 * Instead of creating recursive hierarchies of events,
8134 * we link inherited events back to the original parent,
8135 * which has a filp for sure, which we use as the reference
8138 if (parent_event->parent)
8139 parent_event = parent_event->parent;
8141 child_event = perf_event_alloc(&parent_event->attr,
8144 group_leader, parent_event,
8146 if (IS_ERR(child_event))
8149 if (is_orphaned_event(parent_event) ||
8150 !atomic_long_inc_not_zero(&parent_event->refcount)) {
8151 free_event(child_event);
8158 * Make the child state follow the state of the parent event,
8159 * not its attr.disabled bit. We hold the parent's mutex,
8160 * so we won't race with perf_event_{en, dis}able_family.
8162 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
8163 child_event->state = PERF_EVENT_STATE_INACTIVE;
8165 child_event->state = PERF_EVENT_STATE_OFF;
8167 if (parent_event->attr.freq) {
8168 u64 sample_period = parent_event->hw.sample_period;
8169 struct hw_perf_event *hwc = &child_event->hw;
8171 hwc->sample_period = sample_period;
8172 hwc->last_period = sample_period;
8174 local64_set(&hwc->period_left, sample_period);
8177 child_event->ctx = child_ctx;
8178 child_event->overflow_handler = parent_event->overflow_handler;
8179 child_event->overflow_handler_context
8180 = parent_event->overflow_handler_context;
8183 * Precalculate sample_data sizes
8185 perf_event__header_size(child_event);
8186 perf_event__id_header_size(child_event);
8189 * Link it up in the child's context:
8191 raw_spin_lock_irqsave(&child_ctx->lock, flags);
8192 add_event_to_ctx(child_event, child_ctx);
8193 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8196 * Link this into the parent event's child list
8198 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8199 mutex_lock(&parent_event->child_mutex);
8200 list_add_tail(&child_event->child_list, &parent_event->child_list);
8201 mutex_unlock(&parent_event->child_mutex);
8206 static int inherit_group(struct perf_event *parent_event,
8207 struct task_struct *parent,
8208 struct perf_event_context *parent_ctx,
8209 struct task_struct *child,
8210 struct perf_event_context *child_ctx)
8212 struct perf_event *leader;
8213 struct perf_event *sub;
8214 struct perf_event *child_ctr;
8216 leader = inherit_event(parent_event, parent, parent_ctx,
8217 child, NULL, child_ctx);
8219 return PTR_ERR(leader);
8220 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
8221 child_ctr = inherit_event(sub, parent, parent_ctx,
8222 child, leader, child_ctx);
8223 if (IS_ERR(child_ctr))
8224 return PTR_ERR(child_ctr);
8230 inherit_task_group(struct perf_event *event, struct task_struct *parent,
8231 struct perf_event_context *parent_ctx,
8232 struct task_struct *child, int ctxn,
8236 struct perf_event_context *child_ctx;
8238 if (!event->attr.inherit) {
8243 child_ctx = child->perf_event_ctxp[ctxn];
8246 * This is executed from the parent task context, so
8247 * inherit events that have been marked for cloning.
8248 * First allocate and initialize a context for the
8252 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8256 child->perf_event_ctxp[ctxn] = child_ctx;
8259 ret = inherit_group(event, parent, parent_ctx,
8269 * Initialize the perf_event context in task_struct
8271 static int perf_event_init_context(struct task_struct *child, int ctxn)
8273 struct perf_event_context *child_ctx, *parent_ctx;
8274 struct perf_event_context *cloned_ctx;
8275 struct perf_event *event;
8276 struct task_struct *parent = current;
8277 int inherited_all = 1;
8278 unsigned long flags;
8281 if (likely(!parent->perf_event_ctxp[ctxn]))
8285 * If the parent's context is a clone, pin it so it won't get
8288 parent_ctx = perf_pin_task_context(parent, ctxn);
8293 * No need to check if parent_ctx != NULL here; since we saw
8294 * it non-NULL earlier, the only reason for it to become NULL
8295 * is if we exit, and since we're currently in the middle of
8296 * a fork we can't be exiting at the same time.
8300 * Lock the parent list. No need to lock the child - not PID
8301 * hashed yet and not running, so nobody can access it.
8303 mutex_lock(&parent_ctx->mutex);
8306 * We dont have to disable NMIs - we are only looking at
8307 * the list, not manipulating it:
8309 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8310 ret = inherit_task_group(event, parent, parent_ctx,
8311 child, ctxn, &inherited_all);
8317 * We can't hold ctx->lock when iterating the ->flexible_group list due
8318 * to allocations, but we need to prevent rotation because
8319 * rotate_ctx() will change the list from interrupt context.
8321 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8322 parent_ctx->rotate_disable = 1;
8323 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8325 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8326 ret = inherit_task_group(event, parent, parent_ctx,
8327 child, ctxn, &inherited_all);
8332 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8333 parent_ctx->rotate_disable = 0;
8335 child_ctx = child->perf_event_ctxp[ctxn];
8337 if (child_ctx && inherited_all) {
8339 * Mark the child context as a clone of the parent
8340 * context, or of whatever the parent is a clone of.
8342 * Note that if the parent is a clone, the holding of
8343 * parent_ctx->lock avoids it from being uncloned.
8345 cloned_ctx = parent_ctx->parent_ctx;
8347 child_ctx->parent_ctx = cloned_ctx;
8348 child_ctx->parent_gen = parent_ctx->parent_gen;
8350 child_ctx->parent_ctx = parent_ctx;
8351 child_ctx->parent_gen = parent_ctx->generation;
8353 get_ctx(child_ctx->parent_ctx);
8356 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8357 mutex_unlock(&parent_ctx->mutex);
8359 perf_unpin_context(parent_ctx);
8360 put_ctx(parent_ctx);
8366 * Initialize the perf_event context in task_struct
8368 int perf_event_init_task(struct task_struct *child)
8372 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8373 mutex_init(&child->perf_event_mutex);
8374 INIT_LIST_HEAD(&child->perf_event_list);
8376 for_each_task_context_nr(ctxn) {
8377 ret = perf_event_init_context(child, ctxn);
8379 perf_event_free_task(child);
8387 static void __init perf_event_init_all_cpus(void)
8389 struct swevent_htable *swhash;
8392 for_each_possible_cpu(cpu) {
8393 swhash = &per_cpu(swevent_htable, cpu);
8394 mutex_init(&swhash->hlist_mutex);
8395 INIT_LIST_HEAD(&per_cpu(active_ctx_list, cpu));
8399 static void perf_event_init_cpu(int cpu)
8401 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8403 mutex_lock(&swhash->hlist_mutex);
8404 swhash->online = true;
8405 if (swhash->hlist_refcount > 0) {
8406 struct swevent_hlist *hlist;
8408 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8410 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8412 mutex_unlock(&swhash->hlist_mutex);
8415 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8416 static void __perf_event_exit_context(void *__info)
8418 struct remove_event re = { .detach_group = true };
8419 struct perf_event_context *ctx = __info;
8422 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8423 __perf_remove_from_context(&re);
8427 static void perf_event_exit_cpu_context(int cpu)
8429 struct perf_event_context *ctx;
8433 idx = srcu_read_lock(&pmus_srcu);
8434 list_for_each_entry_rcu(pmu, &pmus, entry) {
8435 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8437 mutex_lock(&ctx->mutex);
8438 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8439 mutex_unlock(&ctx->mutex);
8441 srcu_read_unlock(&pmus_srcu, idx);
8444 static void perf_event_exit_cpu(int cpu)
8446 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8448 perf_event_exit_cpu_context(cpu);
8450 mutex_lock(&swhash->hlist_mutex);
8451 swhash->online = false;
8452 swevent_hlist_release(swhash);
8453 mutex_unlock(&swhash->hlist_mutex);
8456 static inline void perf_event_exit_cpu(int cpu) { }
8460 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8464 for_each_online_cpu(cpu)
8465 perf_event_exit_cpu(cpu);
8471 * Run the perf reboot notifier at the very last possible moment so that
8472 * the generic watchdog code runs as long as possible.
8474 static struct notifier_block perf_reboot_notifier = {
8475 .notifier_call = perf_reboot,
8476 .priority = INT_MIN,
8480 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8482 unsigned int cpu = (long)hcpu;
8484 switch (action & ~CPU_TASKS_FROZEN) {
8486 case CPU_UP_PREPARE:
8487 case CPU_DOWN_FAILED:
8488 perf_event_init_cpu(cpu);
8491 case CPU_UP_CANCELED:
8492 case CPU_DOWN_PREPARE:
8493 perf_event_exit_cpu(cpu);
8502 void __init perf_event_init(void)
8508 perf_event_init_all_cpus();
8509 init_srcu_struct(&pmus_srcu);
8510 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8511 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8512 perf_pmu_register(&perf_task_clock, NULL, -1);
8514 perf_cpu_notifier(perf_cpu_notify);
8515 register_reboot_notifier(&perf_reboot_notifier);
8517 ret = init_hw_breakpoint();
8518 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8520 /* do not patch jump label more than once per second */
8521 jump_label_rate_limit(&perf_sched_events, HZ);
8524 * Build time assertion that we keep the data_head at the intended
8525 * location. IOW, validation we got the __reserved[] size right.
8527 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8531 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
8534 struct perf_pmu_events_attr *pmu_attr =
8535 container_of(attr, struct perf_pmu_events_attr, attr);
8537 if (pmu_attr->event_str)
8538 return sprintf(page, "%s\n", pmu_attr->event_str);
8543 static int __init perf_event_sysfs_init(void)
8548 mutex_lock(&pmus_lock);
8550 ret = bus_register(&pmu_bus);
8554 list_for_each_entry(pmu, &pmus, entry) {
8555 if (!pmu->name || pmu->type < 0)
8558 ret = pmu_dev_alloc(pmu);
8559 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8561 pmu_bus_running = 1;
8565 mutex_unlock(&pmus_lock);
8569 device_initcall(perf_event_sysfs_init);
8571 #ifdef CONFIG_CGROUP_PERF
8572 static struct cgroup_subsys_state *
8573 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8575 struct perf_cgroup *jc;
8577 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8579 return ERR_PTR(-ENOMEM);
8581 jc->info = alloc_percpu(struct perf_cgroup_info);
8584 return ERR_PTR(-ENOMEM);
8590 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8592 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8594 free_percpu(jc->info);
8598 static int __perf_cgroup_move(void *info)
8600 struct task_struct *task = info;
8601 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8605 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8606 struct cgroup_taskset *tset)
8608 struct task_struct *task;
8610 cgroup_taskset_for_each(task, tset)
8611 task_function_call(task, __perf_cgroup_move, task);
8614 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8615 struct cgroup_subsys_state *old_css,
8616 struct task_struct *task)
8619 * cgroup_exit() is called in the copy_process() failure path.
8620 * Ignore this case since the task hasn't ran yet, this avoids
8621 * trying to poke a half freed task state from generic code.
8623 if (!(task->flags & PF_EXITING))
8626 task_function_call(task, __perf_cgroup_move, task);
8629 struct cgroup_subsys perf_event_cgrp_subsys = {
8630 .css_alloc = perf_cgroup_css_alloc,
8631 .css_free = perf_cgroup_css_free,
8632 .exit = perf_cgroup_exit,
8633 .attach = perf_cgroup_attach,
8635 #endif /* CONFIG_CGROUP_PERF */