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 = __get_cpu_var(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 = __get_cpu_var(running_sample_length);
275 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
276 local_samples_len += sample_len_ns;
277 __get_cpu_var(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_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, rotation_list);
878 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
879 * because they're strictly cpu affine and rotate_start is called with IRQs
880 * disabled, while rotate_context is called from IRQ context.
882 static void perf_pmu_rotate_start(struct pmu *pmu)
884 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
885 struct list_head *head = &__get_cpu_var(rotation_list);
887 WARN_ON(!irqs_disabled());
889 if (list_empty(&cpuctx->rotation_list))
890 list_add(&cpuctx->rotation_list, head);
893 static void get_ctx(struct perf_event_context *ctx)
895 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
898 static void put_ctx(struct perf_event_context *ctx)
900 if (atomic_dec_and_test(&ctx->refcount)) {
902 put_ctx(ctx->parent_ctx);
904 put_task_struct(ctx->task);
905 kfree_rcu(ctx, rcu_head);
909 static void unclone_ctx(struct perf_event_context *ctx)
911 if (ctx->parent_ctx) {
912 put_ctx(ctx->parent_ctx);
913 ctx->parent_ctx = NULL;
918 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
921 * only top level events have the pid namespace they were created in
924 event = event->parent;
926 return task_tgid_nr_ns(p, event->ns);
929 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
932 * only top level events have the pid namespace they were created in
935 event = event->parent;
937 return task_pid_nr_ns(p, event->ns);
941 * If we inherit events we want to return the parent event id
944 static u64 primary_event_id(struct perf_event *event)
949 id = event->parent->id;
955 * Get the perf_event_context for a task and lock it.
956 * This has to cope with with the fact that until it is locked,
957 * the context could get moved to another task.
959 static struct perf_event_context *
960 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
962 struct perf_event_context *ctx;
966 * One of the few rules of preemptible RCU is that one cannot do
967 * rcu_read_unlock() while holding a scheduler (or nested) lock when
968 * part of the read side critical section was preemptible -- see
969 * rcu_read_unlock_special().
971 * Since ctx->lock nests under rq->lock we must ensure the entire read
972 * side critical section is non-preemptible.
976 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
979 * If this context is a clone of another, it might
980 * get swapped for another underneath us by
981 * perf_event_task_sched_out, though the
982 * rcu_read_lock() protects us from any context
983 * getting freed. Lock the context and check if it
984 * got swapped before we could get the lock, and retry
985 * if so. If we locked the right context, then it
986 * can't get swapped on us any more.
988 raw_spin_lock_irqsave(&ctx->lock, *flags);
989 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
990 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
996 if (!atomic_inc_not_zero(&ctx->refcount)) {
997 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1007 * Get the context for a task and increment its pin_count so it
1008 * can't get swapped to another task. This also increments its
1009 * reference count so that the context can't get freed.
1011 static struct perf_event_context *
1012 perf_pin_task_context(struct task_struct *task, int ctxn)
1014 struct perf_event_context *ctx;
1015 unsigned long flags;
1017 ctx = perf_lock_task_context(task, ctxn, &flags);
1020 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1025 static void perf_unpin_context(struct perf_event_context *ctx)
1027 unsigned long flags;
1029 raw_spin_lock_irqsave(&ctx->lock, flags);
1031 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1035 * Update the record of the current time in a context.
1037 static void update_context_time(struct perf_event_context *ctx)
1039 u64 now = perf_clock();
1041 ctx->time += now - ctx->timestamp;
1042 ctx->timestamp = now;
1045 static u64 perf_event_time(struct perf_event *event)
1047 struct perf_event_context *ctx = event->ctx;
1049 if (is_cgroup_event(event))
1050 return perf_cgroup_event_time(event);
1052 return ctx ? ctx->time : 0;
1056 * Update the total_time_enabled and total_time_running fields for a event.
1057 * The caller of this function needs to hold the ctx->lock.
1059 static void update_event_times(struct perf_event *event)
1061 struct perf_event_context *ctx = event->ctx;
1064 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1065 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1068 * in cgroup mode, time_enabled represents
1069 * the time the event was enabled AND active
1070 * tasks were in the monitored cgroup. This is
1071 * independent of the activity of the context as
1072 * there may be a mix of cgroup and non-cgroup events.
1074 * That is why we treat cgroup events differently
1077 if (is_cgroup_event(event))
1078 run_end = perf_cgroup_event_time(event);
1079 else if (ctx->is_active)
1080 run_end = ctx->time;
1082 run_end = event->tstamp_stopped;
1084 event->total_time_enabled = run_end - event->tstamp_enabled;
1086 if (event->state == PERF_EVENT_STATE_INACTIVE)
1087 run_end = event->tstamp_stopped;
1089 run_end = perf_event_time(event);
1091 event->total_time_running = run_end - event->tstamp_running;
1096 * Update total_time_enabled and total_time_running for all events in a group.
1098 static void update_group_times(struct perf_event *leader)
1100 struct perf_event *event;
1102 update_event_times(leader);
1103 list_for_each_entry(event, &leader->sibling_list, group_entry)
1104 update_event_times(event);
1107 static struct list_head *
1108 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1110 if (event->attr.pinned)
1111 return &ctx->pinned_groups;
1113 return &ctx->flexible_groups;
1117 * Add a event from the lists for its context.
1118 * Must be called with ctx->mutex and ctx->lock held.
1121 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1123 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1124 event->attach_state |= PERF_ATTACH_CONTEXT;
1127 * If we're a stand alone event or group leader, we go to the context
1128 * list, group events are kept attached to the group so that
1129 * perf_group_detach can, at all times, locate all siblings.
1131 if (event->group_leader == event) {
1132 struct list_head *list;
1134 if (is_software_event(event))
1135 event->group_flags |= PERF_GROUP_SOFTWARE;
1137 list = ctx_group_list(event, ctx);
1138 list_add_tail(&event->group_entry, list);
1141 if (is_cgroup_event(event))
1144 if (has_branch_stack(event))
1145 ctx->nr_branch_stack++;
1147 list_add_rcu(&event->event_entry, &ctx->event_list);
1148 if (!ctx->nr_events)
1149 perf_pmu_rotate_start(ctx->pmu);
1151 if (event->attr.inherit_stat)
1158 * Initialize event state based on the perf_event_attr::disabled.
1160 static inline void perf_event__state_init(struct perf_event *event)
1162 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1163 PERF_EVENT_STATE_INACTIVE;
1167 * Called at perf_event creation and when events are attached/detached from a
1170 static void perf_event__read_size(struct perf_event *event)
1172 int entry = sizeof(u64); /* value */
1176 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1177 size += sizeof(u64);
1179 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1180 size += sizeof(u64);
1182 if (event->attr.read_format & PERF_FORMAT_ID)
1183 entry += sizeof(u64);
1185 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1186 nr += event->group_leader->nr_siblings;
1187 size += sizeof(u64);
1191 event->read_size = size;
1194 static void perf_event__header_size(struct perf_event *event)
1196 struct perf_sample_data *data;
1197 u64 sample_type = event->attr.sample_type;
1200 perf_event__read_size(event);
1202 if (sample_type & PERF_SAMPLE_IP)
1203 size += sizeof(data->ip);
1205 if (sample_type & PERF_SAMPLE_ADDR)
1206 size += sizeof(data->addr);
1208 if (sample_type & PERF_SAMPLE_PERIOD)
1209 size += sizeof(data->period);
1211 if (sample_type & PERF_SAMPLE_WEIGHT)
1212 size += sizeof(data->weight);
1214 if (sample_type & PERF_SAMPLE_READ)
1215 size += event->read_size;
1217 if (sample_type & PERF_SAMPLE_DATA_SRC)
1218 size += sizeof(data->data_src.val);
1220 if (sample_type & PERF_SAMPLE_TRANSACTION)
1221 size += sizeof(data->txn);
1223 event->header_size = size;
1226 static void perf_event__id_header_size(struct perf_event *event)
1228 struct perf_sample_data *data;
1229 u64 sample_type = event->attr.sample_type;
1232 if (sample_type & PERF_SAMPLE_TID)
1233 size += sizeof(data->tid_entry);
1235 if (sample_type & PERF_SAMPLE_TIME)
1236 size += sizeof(data->time);
1238 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1239 size += sizeof(data->id);
1241 if (sample_type & PERF_SAMPLE_ID)
1242 size += sizeof(data->id);
1244 if (sample_type & PERF_SAMPLE_STREAM_ID)
1245 size += sizeof(data->stream_id);
1247 if (sample_type & PERF_SAMPLE_CPU)
1248 size += sizeof(data->cpu_entry);
1250 event->id_header_size = size;
1253 static void perf_group_attach(struct perf_event *event)
1255 struct perf_event *group_leader = event->group_leader, *pos;
1258 * We can have double attach due to group movement in perf_event_open.
1260 if (event->attach_state & PERF_ATTACH_GROUP)
1263 event->attach_state |= PERF_ATTACH_GROUP;
1265 if (group_leader == event)
1268 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1269 !is_software_event(event))
1270 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1272 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1273 group_leader->nr_siblings++;
1275 perf_event__header_size(group_leader);
1277 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1278 perf_event__header_size(pos);
1282 * Remove a event from the lists for its context.
1283 * Must be called with ctx->mutex and ctx->lock held.
1286 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1288 struct perf_cpu_context *cpuctx;
1290 * We can have double detach due to exit/hot-unplug + close.
1292 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1295 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1297 if (is_cgroup_event(event)) {
1299 cpuctx = __get_cpu_context(ctx);
1301 * if there are no more cgroup events
1302 * then cler cgrp to avoid stale pointer
1303 * in update_cgrp_time_from_cpuctx()
1305 if (!ctx->nr_cgroups)
1306 cpuctx->cgrp = NULL;
1309 if (has_branch_stack(event))
1310 ctx->nr_branch_stack--;
1313 if (event->attr.inherit_stat)
1316 list_del_rcu(&event->event_entry);
1318 if (event->group_leader == event)
1319 list_del_init(&event->group_entry);
1321 update_group_times(event);
1324 * If event was in error state, then keep it
1325 * that way, otherwise bogus counts will be
1326 * returned on read(). The only way to get out
1327 * of error state is by explicit re-enabling
1330 if (event->state > PERF_EVENT_STATE_OFF)
1331 event->state = PERF_EVENT_STATE_OFF;
1336 static void perf_group_detach(struct perf_event *event)
1338 struct perf_event *sibling, *tmp;
1339 struct list_head *list = NULL;
1342 * We can have double detach due to exit/hot-unplug + close.
1344 if (!(event->attach_state & PERF_ATTACH_GROUP))
1347 event->attach_state &= ~PERF_ATTACH_GROUP;
1350 * If this is a sibling, remove it from its group.
1352 if (event->group_leader != event) {
1353 list_del_init(&event->group_entry);
1354 event->group_leader->nr_siblings--;
1358 if (!list_empty(&event->group_entry))
1359 list = &event->group_entry;
1362 * If this was a group event with sibling events then
1363 * upgrade the siblings to singleton events by adding them
1364 * to whatever list we are on.
1366 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1368 list_move_tail(&sibling->group_entry, list);
1369 sibling->group_leader = sibling;
1371 /* Inherit group flags from the previous leader */
1372 sibling->group_flags = event->group_flags;
1376 perf_event__header_size(event->group_leader);
1378 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1379 perf_event__header_size(tmp);
1383 * User event without the task.
1385 static bool is_orphaned_event(struct perf_event *event)
1387 return event && !is_kernel_event(event) && !event->owner;
1391 * Event has a parent but parent's task finished and it's
1392 * alive only because of children holding refference.
1394 static bool is_orphaned_child(struct perf_event *event)
1396 return is_orphaned_event(event->parent);
1399 static void orphans_remove_work(struct work_struct *work);
1401 static void schedule_orphans_remove(struct perf_event_context *ctx)
1403 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1406 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1408 ctx->orphans_remove_sched = true;
1412 static int __init perf_workqueue_init(void)
1414 perf_wq = create_singlethread_workqueue("perf");
1415 WARN(!perf_wq, "failed to create perf workqueue\n");
1416 return perf_wq ? 0 : -1;
1419 core_initcall(perf_workqueue_init);
1422 event_filter_match(struct perf_event *event)
1424 return (event->cpu == -1 || event->cpu == smp_processor_id())
1425 && perf_cgroup_match(event);
1429 event_sched_out(struct perf_event *event,
1430 struct perf_cpu_context *cpuctx,
1431 struct perf_event_context *ctx)
1433 u64 tstamp = perf_event_time(event);
1436 * An event which could not be activated because of
1437 * filter mismatch still needs to have its timings
1438 * maintained, otherwise bogus information is return
1439 * via read() for time_enabled, time_running:
1441 if (event->state == PERF_EVENT_STATE_INACTIVE
1442 && !event_filter_match(event)) {
1443 delta = tstamp - event->tstamp_stopped;
1444 event->tstamp_running += delta;
1445 event->tstamp_stopped = tstamp;
1448 if (event->state != PERF_EVENT_STATE_ACTIVE)
1451 perf_pmu_disable(event->pmu);
1453 event->state = PERF_EVENT_STATE_INACTIVE;
1454 if (event->pending_disable) {
1455 event->pending_disable = 0;
1456 event->state = PERF_EVENT_STATE_OFF;
1458 event->tstamp_stopped = tstamp;
1459 event->pmu->del(event, 0);
1462 if (!is_software_event(event))
1463 cpuctx->active_oncpu--;
1465 if (event->attr.freq && event->attr.sample_freq)
1467 if (event->attr.exclusive || !cpuctx->active_oncpu)
1468 cpuctx->exclusive = 0;
1470 if (is_orphaned_child(event))
1471 schedule_orphans_remove(ctx);
1473 perf_pmu_enable(event->pmu);
1477 group_sched_out(struct perf_event *group_event,
1478 struct perf_cpu_context *cpuctx,
1479 struct perf_event_context *ctx)
1481 struct perf_event *event;
1482 int state = group_event->state;
1484 event_sched_out(group_event, cpuctx, ctx);
1487 * Schedule out siblings (if any):
1489 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1490 event_sched_out(event, cpuctx, ctx);
1492 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1493 cpuctx->exclusive = 0;
1496 struct remove_event {
1497 struct perf_event *event;
1502 * Cross CPU call to remove a performance event
1504 * We disable the event on the hardware level first. After that we
1505 * remove it from the context list.
1507 static int __perf_remove_from_context(void *info)
1509 struct remove_event *re = info;
1510 struct perf_event *event = re->event;
1511 struct perf_event_context *ctx = event->ctx;
1512 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1514 raw_spin_lock(&ctx->lock);
1515 event_sched_out(event, cpuctx, ctx);
1516 if (re->detach_group)
1517 perf_group_detach(event);
1518 list_del_event(event, ctx);
1519 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1521 cpuctx->task_ctx = NULL;
1523 raw_spin_unlock(&ctx->lock);
1530 * Remove the event from a task's (or a CPU's) list of events.
1532 * CPU events are removed with a smp call. For task events we only
1533 * call when the task is on a CPU.
1535 * If event->ctx is a cloned context, callers must make sure that
1536 * every task struct that event->ctx->task could possibly point to
1537 * remains valid. This is OK when called from perf_release since
1538 * that only calls us on the top-level context, which can't be a clone.
1539 * When called from perf_event_exit_task, it's OK because the
1540 * context has been detached from its task.
1542 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1544 struct perf_event_context *ctx = event->ctx;
1545 struct task_struct *task = ctx->task;
1546 struct remove_event re = {
1548 .detach_group = detach_group,
1551 lockdep_assert_held(&ctx->mutex);
1555 * Per cpu events are removed via an smp call and
1556 * the removal is always successful.
1558 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1563 if (!task_function_call(task, __perf_remove_from_context, &re))
1566 raw_spin_lock_irq(&ctx->lock);
1568 * If we failed to find a running task, but find the context active now
1569 * that we've acquired the ctx->lock, retry.
1571 if (ctx->is_active) {
1572 raw_spin_unlock_irq(&ctx->lock);
1574 * Reload the task pointer, it might have been changed by
1575 * a concurrent perf_event_context_sched_out().
1582 * Since the task isn't running, its safe to remove the event, us
1583 * holding the ctx->lock ensures the task won't get scheduled in.
1586 perf_group_detach(event);
1587 list_del_event(event, ctx);
1588 raw_spin_unlock_irq(&ctx->lock);
1592 * Cross CPU call to disable a performance event
1594 int __perf_event_disable(void *info)
1596 struct perf_event *event = info;
1597 struct perf_event_context *ctx = event->ctx;
1598 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1601 * If this is a per-task event, need to check whether this
1602 * event's task is the current task on this cpu.
1604 * Can trigger due to concurrent perf_event_context_sched_out()
1605 * flipping contexts around.
1607 if (ctx->task && cpuctx->task_ctx != ctx)
1610 raw_spin_lock(&ctx->lock);
1613 * If the event is on, turn it off.
1614 * If it is in error state, leave it in error state.
1616 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1617 update_context_time(ctx);
1618 update_cgrp_time_from_event(event);
1619 update_group_times(event);
1620 if (event == event->group_leader)
1621 group_sched_out(event, cpuctx, ctx);
1623 event_sched_out(event, cpuctx, ctx);
1624 event->state = PERF_EVENT_STATE_OFF;
1627 raw_spin_unlock(&ctx->lock);
1635 * If event->ctx is a cloned context, callers must make sure that
1636 * every task struct that event->ctx->task could possibly point to
1637 * remains valid. This condition is satisifed when called through
1638 * perf_event_for_each_child or perf_event_for_each because they
1639 * hold the top-level event's child_mutex, so any descendant that
1640 * goes to exit will block in sync_child_event.
1641 * When called from perf_pending_event it's OK because event->ctx
1642 * is the current context on this CPU and preemption is disabled,
1643 * hence we can't get into perf_event_task_sched_out for this context.
1645 void perf_event_disable(struct perf_event *event)
1647 struct perf_event_context *ctx = event->ctx;
1648 struct task_struct *task = ctx->task;
1652 * Disable the event on the cpu that it's on
1654 cpu_function_call(event->cpu, __perf_event_disable, event);
1659 if (!task_function_call(task, __perf_event_disable, event))
1662 raw_spin_lock_irq(&ctx->lock);
1664 * If the event is still active, we need to retry the cross-call.
1666 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1667 raw_spin_unlock_irq(&ctx->lock);
1669 * Reload the task pointer, it might have been changed by
1670 * a concurrent perf_event_context_sched_out().
1677 * Since we have the lock this context can't be scheduled
1678 * in, so we can change the state safely.
1680 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1681 update_group_times(event);
1682 event->state = PERF_EVENT_STATE_OFF;
1684 raw_spin_unlock_irq(&ctx->lock);
1686 EXPORT_SYMBOL_GPL(perf_event_disable);
1688 static void perf_set_shadow_time(struct perf_event *event,
1689 struct perf_event_context *ctx,
1693 * use the correct time source for the time snapshot
1695 * We could get by without this by leveraging the
1696 * fact that to get to this function, the caller
1697 * has most likely already called update_context_time()
1698 * and update_cgrp_time_xx() and thus both timestamp
1699 * are identical (or very close). Given that tstamp is,
1700 * already adjusted for cgroup, we could say that:
1701 * tstamp - ctx->timestamp
1703 * tstamp - cgrp->timestamp.
1705 * Then, in perf_output_read(), the calculation would
1706 * work with no changes because:
1707 * - event is guaranteed scheduled in
1708 * - no scheduled out in between
1709 * - thus the timestamp would be the same
1711 * But this is a bit hairy.
1713 * So instead, we have an explicit cgroup call to remain
1714 * within the time time source all along. We believe it
1715 * is cleaner and simpler to understand.
1717 if (is_cgroup_event(event))
1718 perf_cgroup_set_shadow_time(event, tstamp);
1720 event->shadow_ctx_time = tstamp - ctx->timestamp;
1723 #define MAX_INTERRUPTS (~0ULL)
1725 static void perf_log_throttle(struct perf_event *event, int enable);
1728 event_sched_in(struct perf_event *event,
1729 struct perf_cpu_context *cpuctx,
1730 struct perf_event_context *ctx)
1732 u64 tstamp = perf_event_time(event);
1735 lockdep_assert_held(&ctx->lock);
1737 if (event->state <= PERF_EVENT_STATE_OFF)
1740 event->state = PERF_EVENT_STATE_ACTIVE;
1741 event->oncpu = smp_processor_id();
1744 * Unthrottle events, since we scheduled we might have missed several
1745 * ticks already, also for a heavily scheduling task there is little
1746 * guarantee it'll get a tick in a timely manner.
1748 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1749 perf_log_throttle(event, 1);
1750 event->hw.interrupts = 0;
1754 * The new state must be visible before we turn it on in the hardware:
1758 perf_pmu_disable(event->pmu);
1760 if (event->pmu->add(event, PERF_EF_START)) {
1761 event->state = PERF_EVENT_STATE_INACTIVE;
1767 event->tstamp_running += tstamp - event->tstamp_stopped;
1769 perf_set_shadow_time(event, ctx, tstamp);
1771 if (!is_software_event(event))
1772 cpuctx->active_oncpu++;
1774 if (event->attr.freq && event->attr.sample_freq)
1777 if (event->attr.exclusive)
1778 cpuctx->exclusive = 1;
1780 if (is_orphaned_child(event))
1781 schedule_orphans_remove(ctx);
1784 perf_pmu_enable(event->pmu);
1790 group_sched_in(struct perf_event *group_event,
1791 struct perf_cpu_context *cpuctx,
1792 struct perf_event_context *ctx)
1794 struct perf_event *event, *partial_group = NULL;
1795 struct pmu *pmu = ctx->pmu;
1796 u64 now = ctx->time;
1797 bool simulate = false;
1799 if (group_event->state == PERF_EVENT_STATE_OFF)
1802 pmu->start_txn(pmu);
1804 if (event_sched_in(group_event, cpuctx, ctx)) {
1805 pmu->cancel_txn(pmu);
1806 perf_cpu_hrtimer_restart(cpuctx);
1811 * Schedule in siblings as one group (if any):
1813 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1814 if (event_sched_in(event, cpuctx, ctx)) {
1815 partial_group = event;
1820 if (!pmu->commit_txn(pmu))
1825 * Groups can be scheduled in as one unit only, so undo any
1826 * partial group before returning:
1827 * The events up to the failed event are scheduled out normally,
1828 * tstamp_stopped will be updated.
1830 * The failed events and the remaining siblings need to have
1831 * their timings updated as if they had gone thru event_sched_in()
1832 * and event_sched_out(). This is required to get consistent timings
1833 * across the group. This also takes care of the case where the group
1834 * could never be scheduled by ensuring tstamp_stopped is set to mark
1835 * the time the event was actually stopped, such that time delta
1836 * calculation in update_event_times() is correct.
1838 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1839 if (event == partial_group)
1843 event->tstamp_running += now - event->tstamp_stopped;
1844 event->tstamp_stopped = now;
1846 event_sched_out(event, cpuctx, ctx);
1849 event_sched_out(group_event, cpuctx, ctx);
1851 pmu->cancel_txn(pmu);
1853 perf_cpu_hrtimer_restart(cpuctx);
1859 * Work out whether we can put this event group on the CPU now.
1861 static int group_can_go_on(struct perf_event *event,
1862 struct perf_cpu_context *cpuctx,
1866 * Groups consisting entirely of software events can always go on.
1868 if (event->group_flags & PERF_GROUP_SOFTWARE)
1871 * If an exclusive group is already on, no other hardware
1874 if (cpuctx->exclusive)
1877 * If this group is exclusive and there are already
1878 * events on the CPU, it can't go on.
1880 if (event->attr.exclusive && cpuctx->active_oncpu)
1883 * Otherwise, try to add it if all previous groups were able
1889 static void add_event_to_ctx(struct perf_event *event,
1890 struct perf_event_context *ctx)
1892 u64 tstamp = perf_event_time(event);
1894 list_add_event(event, ctx);
1895 perf_group_attach(event);
1896 event->tstamp_enabled = tstamp;
1897 event->tstamp_running = tstamp;
1898 event->tstamp_stopped = tstamp;
1901 static void task_ctx_sched_out(struct perf_event_context *ctx);
1903 ctx_sched_in(struct perf_event_context *ctx,
1904 struct perf_cpu_context *cpuctx,
1905 enum event_type_t event_type,
1906 struct task_struct *task);
1908 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1909 struct perf_event_context *ctx,
1910 struct task_struct *task)
1912 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1914 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1915 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1917 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1921 * Cross CPU call to install and enable a performance event
1923 * Must be called with ctx->mutex held
1925 static int __perf_install_in_context(void *info)
1927 struct perf_event *event = info;
1928 struct perf_event_context *ctx = event->ctx;
1929 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1930 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1931 struct task_struct *task = current;
1933 perf_ctx_lock(cpuctx, task_ctx);
1934 perf_pmu_disable(cpuctx->ctx.pmu);
1937 * If there was an active task_ctx schedule it out.
1940 task_ctx_sched_out(task_ctx);
1943 * If the context we're installing events in is not the
1944 * active task_ctx, flip them.
1946 if (ctx->task && task_ctx != ctx) {
1948 raw_spin_unlock(&task_ctx->lock);
1949 raw_spin_lock(&ctx->lock);
1954 cpuctx->task_ctx = task_ctx;
1955 task = task_ctx->task;
1958 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1960 update_context_time(ctx);
1962 * update cgrp time only if current cgrp
1963 * matches event->cgrp. Must be done before
1964 * calling add_event_to_ctx()
1966 update_cgrp_time_from_event(event);
1968 add_event_to_ctx(event, ctx);
1971 * Schedule everything back in
1973 perf_event_sched_in(cpuctx, task_ctx, task);
1975 perf_pmu_enable(cpuctx->ctx.pmu);
1976 perf_ctx_unlock(cpuctx, task_ctx);
1982 * Attach a performance event to a context
1984 * First we add the event to the list with the hardware enable bit
1985 * in event->hw_config cleared.
1987 * If the event is attached to a task which is on a CPU we use a smp
1988 * call to enable it in the task context. The task might have been
1989 * scheduled away, but we check this in the smp call again.
1992 perf_install_in_context(struct perf_event_context *ctx,
1993 struct perf_event *event,
1996 struct task_struct *task = ctx->task;
1998 lockdep_assert_held(&ctx->mutex);
2001 if (event->cpu != -1)
2006 * Per cpu events are installed via an smp call and
2007 * the install is always successful.
2009 cpu_function_call(cpu, __perf_install_in_context, event);
2014 if (!task_function_call(task, __perf_install_in_context, event))
2017 raw_spin_lock_irq(&ctx->lock);
2019 * If we failed to find a running task, but find the context active now
2020 * that we've acquired the ctx->lock, retry.
2022 if (ctx->is_active) {
2023 raw_spin_unlock_irq(&ctx->lock);
2025 * Reload the task pointer, it might have been changed by
2026 * a concurrent perf_event_context_sched_out().
2033 * Since the task isn't running, its safe to add the event, us holding
2034 * the ctx->lock ensures the task won't get scheduled in.
2036 add_event_to_ctx(event, ctx);
2037 raw_spin_unlock_irq(&ctx->lock);
2041 * Put a event into inactive state and update time fields.
2042 * Enabling the leader of a group effectively enables all
2043 * the group members that aren't explicitly disabled, so we
2044 * have to update their ->tstamp_enabled also.
2045 * Note: this works for group members as well as group leaders
2046 * since the non-leader members' sibling_lists will be empty.
2048 static void __perf_event_mark_enabled(struct perf_event *event)
2050 struct perf_event *sub;
2051 u64 tstamp = perf_event_time(event);
2053 event->state = PERF_EVENT_STATE_INACTIVE;
2054 event->tstamp_enabled = tstamp - event->total_time_enabled;
2055 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2056 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2057 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2062 * Cross CPU call to enable a performance event
2064 static int __perf_event_enable(void *info)
2066 struct perf_event *event = info;
2067 struct perf_event_context *ctx = event->ctx;
2068 struct perf_event *leader = event->group_leader;
2069 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2073 * There's a time window between 'ctx->is_active' check
2074 * in perf_event_enable function and this place having:
2076 * - ctx->lock unlocked
2078 * where the task could be killed and 'ctx' deactivated
2079 * by perf_event_exit_task.
2081 if (!ctx->is_active)
2084 raw_spin_lock(&ctx->lock);
2085 update_context_time(ctx);
2087 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2091 * set current task's cgroup time reference point
2093 perf_cgroup_set_timestamp(current, ctx);
2095 __perf_event_mark_enabled(event);
2097 if (!event_filter_match(event)) {
2098 if (is_cgroup_event(event))
2099 perf_cgroup_defer_enabled(event);
2104 * If the event is in a group and isn't the group leader,
2105 * then don't put it on unless the group is on.
2107 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2110 if (!group_can_go_on(event, cpuctx, 1)) {
2113 if (event == leader)
2114 err = group_sched_in(event, cpuctx, ctx);
2116 err = event_sched_in(event, cpuctx, ctx);
2121 * If this event can't go on and it's part of a
2122 * group, then the whole group has to come off.
2124 if (leader != event) {
2125 group_sched_out(leader, cpuctx, ctx);
2126 perf_cpu_hrtimer_restart(cpuctx);
2128 if (leader->attr.pinned) {
2129 update_group_times(leader);
2130 leader->state = PERF_EVENT_STATE_ERROR;
2135 raw_spin_unlock(&ctx->lock);
2143 * If event->ctx is a cloned context, callers must make sure that
2144 * every task struct that event->ctx->task could possibly point to
2145 * remains valid. This condition is satisfied when called through
2146 * perf_event_for_each_child or perf_event_for_each as described
2147 * for perf_event_disable.
2149 void perf_event_enable(struct perf_event *event)
2151 struct perf_event_context *ctx = event->ctx;
2152 struct task_struct *task = ctx->task;
2156 * Enable the event on the cpu that it's on
2158 cpu_function_call(event->cpu, __perf_event_enable, event);
2162 raw_spin_lock_irq(&ctx->lock);
2163 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2167 * If the event is in error state, clear that first.
2168 * That way, if we see the event in error state below, we
2169 * know that it has gone back into error state, as distinct
2170 * from the task having been scheduled away before the
2171 * cross-call arrived.
2173 if (event->state == PERF_EVENT_STATE_ERROR)
2174 event->state = PERF_EVENT_STATE_OFF;
2177 if (!ctx->is_active) {
2178 __perf_event_mark_enabled(event);
2182 raw_spin_unlock_irq(&ctx->lock);
2184 if (!task_function_call(task, __perf_event_enable, event))
2187 raw_spin_lock_irq(&ctx->lock);
2190 * If the context is active and the event is still off,
2191 * we need to retry the cross-call.
2193 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2195 * task could have been flipped by a concurrent
2196 * perf_event_context_sched_out()
2203 raw_spin_unlock_irq(&ctx->lock);
2205 EXPORT_SYMBOL_GPL(perf_event_enable);
2207 int perf_event_refresh(struct perf_event *event, int refresh)
2210 * not supported on inherited events
2212 if (event->attr.inherit || !is_sampling_event(event))
2215 atomic_add(refresh, &event->event_limit);
2216 perf_event_enable(event);
2220 EXPORT_SYMBOL_GPL(perf_event_refresh);
2222 static void ctx_sched_out(struct perf_event_context *ctx,
2223 struct perf_cpu_context *cpuctx,
2224 enum event_type_t event_type)
2226 struct perf_event *event;
2227 int is_active = ctx->is_active;
2229 ctx->is_active &= ~event_type;
2230 if (likely(!ctx->nr_events))
2233 update_context_time(ctx);
2234 update_cgrp_time_from_cpuctx(cpuctx);
2235 if (!ctx->nr_active)
2238 perf_pmu_disable(ctx->pmu);
2239 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2240 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2241 group_sched_out(event, cpuctx, ctx);
2244 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2245 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2246 group_sched_out(event, cpuctx, ctx);
2248 perf_pmu_enable(ctx->pmu);
2252 * Test whether two contexts are equivalent, i.e. whether they have both been
2253 * cloned from the same version of the same context.
2255 * Equivalence is measured using a generation number in the context that is
2256 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2257 * and list_del_event().
2259 static int context_equiv(struct perf_event_context *ctx1,
2260 struct perf_event_context *ctx2)
2262 /* Pinning disables the swap optimization */
2263 if (ctx1->pin_count || ctx2->pin_count)
2266 /* If ctx1 is the parent of ctx2 */
2267 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2270 /* If ctx2 is the parent of ctx1 */
2271 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2275 * If ctx1 and ctx2 have the same parent; we flatten the parent
2276 * hierarchy, see perf_event_init_context().
2278 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2279 ctx1->parent_gen == ctx2->parent_gen)
2286 static void __perf_event_sync_stat(struct perf_event *event,
2287 struct perf_event *next_event)
2291 if (!event->attr.inherit_stat)
2295 * Update the event value, we cannot use perf_event_read()
2296 * because we're in the middle of a context switch and have IRQs
2297 * disabled, which upsets smp_call_function_single(), however
2298 * we know the event must be on the current CPU, therefore we
2299 * don't need to use it.
2301 switch (event->state) {
2302 case PERF_EVENT_STATE_ACTIVE:
2303 event->pmu->read(event);
2306 case PERF_EVENT_STATE_INACTIVE:
2307 update_event_times(event);
2315 * In order to keep per-task stats reliable we need to flip the event
2316 * values when we flip the contexts.
2318 value = local64_read(&next_event->count);
2319 value = local64_xchg(&event->count, value);
2320 local64_set(&next_event->count, value);
2322 swap(event->total_time_enabled, next_event->total_time_enabled);
2323 swap(event->total_time_running, next_event->total_time_running);
2326 * Since we swizzled the values, update the user visible data too.
2328 perf_event_update_userpage(event);
2329 perf_event_update_userpage(next_event);
2332 static void perf_event_sync_stat(struct perf_event_context *ctx,
2333 struct perf_event_context *next_ctx)
2335 struct perf_event *event, *next_event;
2340 update_context_time(ctx);
2342 event = list_first_entry(&ctx->event_list,
2343 struct perf_event, event_entry);
2345 next_event = list_first_entry(&next_ctx->event_list,
2346 struct perf_event, event_entry);
2348 while (&event->event_entry != &ctx->event_list &&
2349 &next_event->event_entry != &next_ctx->event_list) {
2351 __perf_event_sync_stat(event, next_event);
2353 event = list_next_entry(event, event_entry);
2354 next_event = list_next_entry(next_event, event_entry);
2358 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2359 struct task_struct *next)
2361 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2362 struct perf_event_context *next_ctx;
2363 struct perf_event_context *parent, *next_parent;
2364 struct perf_cpu_context *cpuctx;
2370 cpuctx = __get_cpu_context(ctx);
2371 if (!cpuctx->task_ctx)
2375 next_ctx = next->perf_event_ctxp[ctxn];
2379 parent = rcu_dereference(ctx->parent_ctx);
2380 next_parent = rcu_dereference(next_ctx->parent_ctx);
2382 /* If neither context have a parent context; they cannot be clones. */
2383 if (!parent && !next_parent)
2386 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2388 * Looks like the two contexts are clones, so we might be
2389 * able to optimize the context switch. We lock both
2390 * contexts and check that they are clones under the
2391 * lock (including re-checking that neither has been
2392 * uncloned in the meantime). It doesn't matter which
2393 * order we take the locks because no other cpu could
2394 * be trying to lock both of these tasks.
2396 raw_spin_lock(&ctx->lock);
2397 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2398 if (context_equiv(ctx, next_ctx)) {
2400 * XXX do we need a memory barrier of sorts
2401 * wrt to rcu_dereference() of perf_event_ctxp
2403 task->perf_event_ctxp[ctxn] = next_ctx;
2404 next->perf_event_ctxp[ctxn] = ctx;
2406 next_ctx->task = task;
2409 perf_event_sync_stat(ctx, next_ctx);
2411 raw_spin_unlock(&next_ctx->lock);
2412 raw_spin_unlock(&ctx->lock);
2418 raw_spin_lock(&ctx->lock);
2419 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2420 cpuctx->task_ctx = NULL;
2421 raw_spin_unlock(&ctx->lock);
2425 #define for_each_task_context_nr(ctxn) \
2426 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2429 * Called from scheduler to remove the events of the current task,
2430 * with interrupts disabled.
2432 * We stop each event and update the event value in event->count.
2434 * This does not protect us against NMI, but disable()
2435 * sets the disabled bit in the control field of event _before_
2436 * accessing the event control register. If a NMI hits, then it will
2437 * not restart the event.
2439 void __perf_event_task_sched_out(struct task_struct *task,
2440 struct task_struct *next)
2444 for_each_task_context_nr(ctxn)
2445 perf_event_context_sched_out(task, ctxn, next);
2448 * if cgroup events exist on this CPU, then we need
2449 * to check if we have to switch out PMU state.
2450 * cgroup event are system-wide mode only
2452 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2453 perf_cgroup_sched_out(task, next);
2456 static void task_ctx_sched_out(struct perf_event_context *ctx)
2458 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2460 if (!cpuctx->task_ctx)
2463 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2466 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2467 cpuctx->task_ctx = NULL;
2471 * Called with IRQs disabled
2473 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2474 enum event_type_t event_type)
2476 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2480 ctx_pinned_sched_in(struct perf_event_context *ctx,
2481 struct perf_cpu_context *cpuctx)
2483 struct perf_event *event;
2485 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2486 if (event->state <= PERF_EVENT_STATE_OFF)
2488 if (!event_filter_match(event))
2491 /* may need to reset tstamp_enabled */
2492 if (is_cgroup_event(event))
2493 perf_cgroup_mark_enabled(event, ctx);
2495 if (group_can_go_on(event, cpuctx, 1))
2496 group_sched_in(event, cpuctx, ctx);
2499 * If this pinned group hasn't been scheduled,
2500 * put it in error state.
2502 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2503 update_group_times(event);
2504 event->state = PERF_EVENT_STATE_ERROR;
2510 ctx_flexible_sched_in(struct perf_event_context *ctx,
2511 struct perf_cpu_context *cpuctx)
2513 struct perf_event *event;
2516 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2517 /* Ignore events in OFF or ERROR state */
2518 if (event->state <= PERF_EVENT_STATE_OFF)
2521 * Listen to the 'cpu' scheduling filter constraint
2524 if (!event_filter_match(event))
2527 /* may need to reset tstamp_enabled */
2528 if (is_cgroup_event(event))
2529 perf_cgroup_mark_enabled(event, ctx);
2531 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2532 if (group_sched_in(event, cpuctx, ctx))
2539 ctx_sched_in(struct perf_event_context *ctx,
2540 struct perf_cpu_context *cpuctx,
2541 enum event_type_t event_type,
2542 struct task_struct *task)
2545 int is_active = ctx->is_active;
2547 ctx->is_active |= event_type;
2548 if (likely(!ctx->nr_events))
2552 ctx->timestamp = now;
2553 perf_cgroup_set_timestamp(task, ctx);
2555 * First go through the list and put on any pinned groups
2556 * in order to give them the best chance of going on.
2558 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2559 ctx_pinned_sched_in(ctx, cpuctx);
2561 /* Then walk through the lower prio flexible groups */
2562 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2563 ctx_flexible_sched_in(ctx, cpuctx);
2566 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2567 enum event_type_t event_type,
2568 struct task_struct *task)
2570 struct perf_event_context *ctx = &cpuctx->ctx;
2572 ctx_sched_in(ctx, cpuctx, event_type, task);
2575 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2576 struct task_struct *task)
2578 struct perf_cpu_context *cpuctx;
2580 cpuctx = __get_cpu_context(ctx);
2581 if (cpuctx->task_ctx == ctx)
2584 perf_ctx_lock(cpuctx, ctx);
2585 perf_pmu_disable(ctx->pmu);
2587 * We want to keep the following priority order:
2588 * cpu pinned (that don't need to move), task pinned,
2589 * cpu flexible, task flexible.
2591 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2594 cpuctx->task_ctx = ctx;
2596 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2598 perf_pmu_enable(ctx->pmu);
2599 perf_ctx_unlock(cpuctx, ctx);
2602 * Since these rotations are per-cpu, we need to ensure the
2603 * cpu-context we got scheduled on is actually rotating.
2605 perf_pmu_rotate_start(ctx->pmu);
2609 * When sampling the branck stack in system-wide, it may be necessary
2610 * to flush the stack on context switch. This happens when the branch
2611 * stack does not tag its entries with the pid of the current task.
2612 * Otherwise it becomes impossible to associate a branch entry with a
2613 * task. This ambiguity is more likely to appear when the branch stack
2614 * supports priv level filtering and the user sets it to monitor only
2615 * at the user level (which could be a useful measurement in system-wide
2616 * mode). In that case, the risk is high of having a branch stack with
2617 * branch from multiple tasks. Flushing may mean dropping the existing
2618 * entries or stashing them somewhere in the PMU specific code layer.
2620 * This function provides the context switch callback to the lower code
2621 * layer. It is invoked ONLY when there is at least one system-wide context
2622 * with at least one active event using taken branch sampling.
2624 static void perf_branch_stack_sched_in(struct task_struct *prev,
2625 struct task_struct *task)
2627 struct perf_cpu_context *cpuctx;
2629 unsigned long flags;
2631 /* no need to flush branch stack if not changing task */
2635 local_irq_save(flags);
2639 list_for_each_entry_rcu(pmu, &pmus, entry) {
2640 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2643 * check if the context has at least one
2644 * event using PERF_SAMPLE_BRANCH_STACK
2646 if (cpuctx->ctx.nr_branch_stack > 0
2647 && pmu->flush_branch_stack) {
2649 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2651 perf_pmu_disable(pmu);
2653 pmu->flush_branch_stack();
2655 perf_pmu_enable(pmu);
2657 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2663 local_irq_restore(flags);
2667 * Called from scheduler to add the events of the current task
2668 * with interrupts disabled.
2670 * We restore the event value and then enable it.
2672 * This does not protect us against NMI, but enable()
2673 * sets the enabled bit in the control field of event _before_
2674 * accessing the event control register. If a NMI hits, then it will
2675 * keep the event running.
2677 void __perf_event_task_sched_in(struct task_struct *prev,
2678 struct task_struct *task)
2680 struct perf_event_context *ctx;
2683 for_each_task_context_nr(ctxn) {
2684 ctx = task->perf_event_ctxp[ctxn];
2688 perf_event_context_sched_in(ctx, task);
2691 * if cgroup events exist on this CPU, then we need
2692 * to check if we have to switch in PMU state.
2693 * cgroup event are system-wide mode only
2695 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2696 perf_cgroup_sched_in(prev, task);
2698 /* check for system-wide branch_stack events */
2699 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2700 perf_branch_stack_sched_in(prev, task);
2703 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2705 u64 frequency = event->attr.sample_freq;
2706 u64 sec = NSEC_PER_SEC;
2707 u64 divisor, dividend;
2709 int count_fls, nsec_fls, frequency_fls, sec_fls;
2711 count_fls = fls64(count);
2712 nsec_fls = fls64(nsec);
2713 frequency_fls = fls64(frequency);
2717 * We got @count in @nsec, with a target of sample_freq HZ
2718 * the target period becomes:
2721 * period = -------------------
2722 * @nsec * sample_freq
2727 * Reduce accuracy by one bit such that @a and @b converge
2728 * to a similar magnitude.
2730 #define REDUCE_FLS(a, b) \
2732 if (a##_fls > b##_fls) { \
2742 * Reduce accuracy until either term fits in a u64, then proceed with
2743 * the other, so that finally we can do a u64/u64 division.
2745 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2746 REDUCE_FLS(nsec, frequency);
2747 REDUCE_FLS(sec, count);
2750 if (count_fls + sec_fls > 64) {
2751 divisor = nsec * frequency;
2753 while (count_fls + sec_fls > 64) {
2754 REDUCE_FLS(count, sec);
2758 dividend = count * sec;
2760 dividend = count * sec;
2762 while (nsec_fls + frequency_fls > 64) {
2763 REDUCE_FLS(nsec, frequency);
2767 divisor = nsec * frequency;
2773 return div64_u64(dividend, divisor);
2776 static DEFINE_PER_CPU(int, perf_throttled_count);
2777 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2779 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2781 struct hw_perf_event *hwc = &event->hw;
2782 s64 period, sample_period;
2785 period = perf_calculate_period(event, nsec, count);
2787 delta = (s64)(period - hwc->sample_period);
2788 delta = (delta + 7) / 8; /* low pass filter */
2790 sample_period = hwc->sample_period + delta;
2795 hwc->sample_period = sample_period;
2797 if (local64_read(&hwc->period_left) > 8*sample_period) {
2799 event->pmu->stop(event, PERF_EF_UPDATE);
2801 local64_set(&hwc->period_left, 0);
2804 event->pmu->start(event, PERF_EF_RELOAD);
2809 * combine freq adjustment with unthrottling to avoid two passes over the
2810 * events. At the same time, make sure, having freq events does not change
2811 * the rate of unthrottling as that would introduce bias.
2813 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2816 struct perf_event *event;
2817 struct hw_perf_event *hwc;
2818 u64 now, period = TICK_NSEC;
2822 * only need to iterate over all events iff:
2823 * - context have events in frequency mode (needs freq adjust)
2824 * - there are events to unthrottle on this cpu
2826 if (!(ctx->nr_freq || needs_unthr))
2829 raw_spin_lock(&ctx->lock);
2830 perf_pmu_disable(ctx->pmu);
2832 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2833 if (event->state != PERF_EVENT_STATE_ACTIVE)
2836 if (!event_filter_match(event))
2839 perf_pmu_disable(event->pmu);
2843 if (hwc->interrupts == MAX_INTERRUPTS) {
2844 hwc->interrupts = 0;
2845 perf_log_throttle(event, 1);
2846 event->pmu->start(event, 0);
2849 if (!event->attr.freq || !event->attr.sample_freq)
2853 * stop the event and update event->count
2855 event->pmu->stop(event, PERF_EF_UPDATE);
2857 now = local64_read(&event->count);
2858 delta = now - hwc->freq_count_stamp;
2859 hwc->freq_count_stamp = now;
2863 * reload only if value has changed
2864 * we have stopped the event so tell that
2865 * to perf_adjust_period() to avoid stopping it
2869 perf_adjust_period(event, period, delta, false);
2871 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2873 perf_pmu_enable(event->pmu);
2876 perf_pmu_enable(ctx->pmu);
2877 raw_spin_unlock(&ctx->lock);
2881 * Round-robin a context's events:
2883 static void rotate_ctx(struct perf_event_context *ctx)
2886 * Rotate the first entry last of non-pinned groups. Rotation might be
2887 * disabled by the inheritance code.
2889 if (!ctx->rotate_disable)
2890 list_rotate_left(&ctx->flexible_groups);
2894 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2895 * because they're strictly cpu affine and rotate_start is called with IRQs
2896 * disabled, while rotate_context is called from IRQ context.
2898 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2900 struct perf_event_context *ctx = NULL;
2901 int rotate = 0, remove = 1;
2903 if (cpuctx->ctx.nr_events) {
2905 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2909 ctx = cpuctx->task_ctx;
2910 if (ctx && ctx->nr_events) {
2912 if (ctx->nr_events != ctx->nr_active)
2919 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2920 perf_pmu_disable(cpuctx->ctx.pmu);
2922 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2924 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2926 rotate_ctx(&cpuctx->ctx);
2930 perf_event_sched_in(cpuctx, ctx, current);
2932 perf_pmu_enable(cpuctx->ctx.pmu);
2933 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2936 list_del_init(&cpuctx->rotation_list);
2941 #ifdef CONFIG_NO_HZ_FULL
2942 bool perf_event_can_stop_tick(void)
2944 if (atomic_read(&nr_freq_events) ||
2945 __this_cpu_read(perf_throttled_count))
2952 void perf_event_task_tick(void)
2954 struct list_head *head = &__get_cpu_var(rotation_list);
2955 struct perf_cpu_context *cpuctx, *tmp;
2956 struct perf_event_context *ctx;
2959 WARN_ON(!irqs_disabled());
2961 __this_cpu_inc(perf_throttled_seq);
2962 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2964 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2966 perf_adjust_freq_unthr_context(ctx, throttled);
2968 ctx = cpuctx->task_ctx;
2970 perf_adjust_freq_unthr_context(ctx, throttled);
2974 static int event_enable_on_exec(struct perf_event *event,
2975 struct perf_event_context *ctx)
2977 if (!event->attr.enable_on_exec)
2980 event->attr.enable_on_exec = 0;
2981 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2984 __perf_event_mark_enabled(event);
2990 * Enable all of a task's events that have been marked enable-on-exec.
2991 * This expects task == current.
2993 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2995 struct perf_event *event;
2996 unsigned long flags;
3000 local_irq_save(flags);
3001 if (!ctx || !ctx->nr_events)
3005 * We must ctxsw out cgroup events to avoid conflict
3006 * when invoking perf_task_event_sched_in() later on
3007 * in this function. Otherwise we end up trying to
3008 * ctxswin cgroup events which are already scheduled
3011 perf_cgroup_sched_out(current, NULL);
3013 raw_spin_lock(&ctx->lock);
3014 task_ctx_sched_out(ctx);
3016 list_for_each_entry(event, &ctx->event_list, event_entry) {
3017 ret = event_enable_on_exec(event, ctx);
3023 * Unclone this context if we enabled any event.
3028 raw_spin_unlock(&ctx->lock);
3031 * Also calls ctxswin for cgroup events, if any:
3033 perf_event_context_sched_in(ctx, ctx->task);
3035 local_irq_restore(flags);
3038 void perf_event_exec(void)
3040 struct perf_event_context *ctx;
3044 for_each_task_context_nr(ctxn) {
3045 ctx = current->perf_event_ctxp[ctxn];
3049 perf_event_enable_on_exec(ctx);
3055 * Cross CPU call to read the hardware event
3057 static void __perf_event_read(void *info)
3059 struct perf_event *event = info;
3060 struct perf_event_context *ctx = event->ctx;
3061 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3064 * If this is a task context, we need to check whether it is
3065 * the current task context of this cpu. If not it has been
3066 * scheduled out before the smp call arrived. In that case
3067 * event->count would have been updated to a recent sample
3068 * when the event was scheduled out.
3070 if (ctx->task && cpuctx->task_ctx != ctx)
3073 raw_spin_lock(&ctx->lock);
3074 if (ctx->is_active) {
3075 update_context_time(ctx);
3076 update_cgrp_time_from_event(event);
3078 update_event_times(event);
3079 if (event->state == PERF_EVENT_STATE_ACTIVE)
3080 event->pmu->read(event);
3081 raw_spin_unlock(&ctx->lock);
3084 static inline u64 perf_event_count(struct perf_event *event)
3086 return local64_read(&event->count) + atomic64_read(&event->child_count);
3089 static u64 perf_event_read(struct perf_event *event)
3092 * If event is enabled and currently active on a CPU, update the
3093 * value in the event structure:
3095 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3096 smp_call_function_single(event->oncpu,
3097 __perf_event_read, event, 1);
3098 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3099 struct perf_event_context *ctx = event->ctx;
3100 unsigned long flags;
3102 raw_spin_lock_irqsave(&ctx->lock, flags);
3104 * may read while context is not active
3105 * (e.g., thread is blocked), in that case
3106 * we cannot update context time
3108 if (ctx->is_active) {
3109 update_context_time(ctx);
3110 update_cgrp_time_from_event(event);
3112 update_event_times(event);
3113 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3116 return perf_event_count(event);
3120 * Initialize the perf_event context in a task_struct:
3122 static void __perf_event_init_context(struct perf_event_context *ctx)
3124 raw_spin_lock_init(&ctx->lock);
3125 mutex_init(&ctx->mutex);
3126 INIT_LIST_HEAD(&ctx->pinned_groups);
3127 INIT_LIST_HEAD(&ctx->flexible_groups);
3128 INIT_LIST_HEAD(&ctx->event_list);
3129 atomic_set(&ctx->refcount, 1);
3130 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3133 static struct perf_event_context *
3134 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3136 struct perf_event_context *ctx;
3138 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3142 __perf_event_init_context(ctx);
3145 get_task_struct(task);
3152 static struct task_struct *
3153 find_lively_task_by_vpid(pid_t vpid)
3155 struct task_struct *task;
3162 task = find_task_by_vpid(vpid);
3164 get_task_struct(task);
3168 return ERR_PTR(-ESRCH);
3170 /* Reuse ptrace permission checks for now. */
3172 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3177 put_task_struct(task);
3178 return ERR_PTR(err);
3183 * Returns a matching context with refcount and pincount.
3185 static struct perf_event_context *
3186 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3188 struct perf_event_context *ctx;
3189 struct perf_cpu_context *cpuctx;
3190 unsigned long flags;
3194 /* Must be root to operate on a CPU event: */
3195 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3196 return ERR_PTR(-EACCES);
3199 * We could be clever and allow to attach a event to an
3200 * offline CPU and activate it when the CPU comes up, but
3203 if (!cpu_online(cpu))
3204 return ERR_PTR(-ENODEV);
3206 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3215 ctxn = pmu->task_ctx_nr;
3220 ctx = perf_lock_task_context(task, ctxn, &flags);
3224 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3226 ctx = alloc_perf_context(pmu, task);
3232 mutex_lock(&task->perf_event_mutex);
3234 * If it has already passed perf_event_exit_task().
3235 * we must see PF_EXITING, it takes this mutex too.
3237 if (task->flags & PF_EXITING)
3239 else if (task->perf_event_ctxp[ctxn])
3244 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3246 mutex_unlock(&task->perf_event_mutex);
3248 if (unlikely(err)) {
3260 return ERR_PTR(err);
3263 static void perf_event_free_filter(struct perf_event *event);
3265 static void free_event_rcu(struct rcu_head *head)
3267 struct perf_event *event;
3269 event = container_of(head, struct perf_event, rcu_head);
3271 put_pid_ns(event->ns);
3272 perf_event_free_filter(event);
3276 static void ring_buffer_put(struct ring_buffer *rb);
3277 static void ring_buffer_attach(struct perf_event *event,
3278 struct ring_buffer *rb);
3280 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3285 if (has_branch_stack(event)) {
3286 if (!(event->attach_state & PERF_ATTACH_TASK))
3287 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3289 if (is_cgroup_event(event))
3290 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3293 static void unaccount_event(struct perf_event *event)
3298 if (event->attach_state & PERF_ATTACH_TASK)
3299 static_key_slow_dec_deferred(&perf_sched_events);
3300 if (event->attr.mmap || event->attr.mmap_data)
3301 atomic_dec(&nr_mmap_events);
3302 if (event->attr.comm)
3303 atomic_dec(&nr_comm_events);
3304 if (event->attr.task)
3305 atomic_dec(&nr_task_events);
3306 if (event->attr.freq)
3307 atomic_dec(&nr_freq_events);
3308 if (is_cgroup_event(event))
3309 static_key_slow_dec_deferred(&perf_sched_events);
3310 if (has_branch_stack(event))
3311 static_key_slow_dec_deferred(&perf_sched_events);
3313 unaccount_event_cpu(event, event->cpu);
3316 static void __free_event(struct perf_event *event)
3318 if (!event->parent) {
3319 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3320 put_callchain_buffers();
3324 event->destroy(event);
3327 put_ctx(event->ctx);
3330 module_put(event->pmu->module);
3332 call_rcu(&event->rcu_head, free_event_rcu);
3335 static void _free_event(struct perf_event *event)
3337 irq_work_sync(&event->pending);
3339 unaccount_event(event);
3343 * Can happen when we close an event with re-directed output.
3345 * Since we have a 0 refcount, perf_mmap_close() will skip
3346 * over us; possibly making our ring_buffer_put() the last.
3348 mutex_lock(&event->mmap_mutex);
3349 ring_buffer_attach(event, NULL);
3350 mutex_unlock(&event->mmap_mutex);
3353 if (is_cgroup_event(event))
3354 perf_detach_cgroup(event);
3356 __free_event(event);
3360 * Used to free events which have a known refcount of 1, such as in error paths
3361 * where the event isn't exposed yet and inherited events.
3363 static void free_event(struct perf_event *event)
3365 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3366 "unexpected event refcount: %ld; ptr=%p\n",
3367 atomic_long_read(&event->refcount), event)) {
3368 /* leak to avoid use-after-free */
3376 * Remove user event from the owner task.
3378 static void perf_remove_from_owner(struct perf_event *event)
3380 struct task_struct *owner;
3383 owner = ACCESS_ONCE(event->owner);
3385 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3386 * !owner it means the list deletion is complete and we can indeed
3387 * free this event, otherwise we need to serialize on
3388 * owner->perf_event_mutex.
3390 smp_read_barrier_depends();
3393 * Since delayed_put_task_struct() also drops the last
3394 * task reference we can safely take a new reference
3395 * while holding the rcu_read_lock().
3397 get_task_struct(owner);
3402 mutex_lock(&owner->perf_event_mutex);
3404 * We have to re-check the event->owner field, if it is cleared
3405 * we raced with perf_event_exit_task(), acquiring the mutex
3406 * ensured they're done, and we can proceed with freeing the
3410 list_del_init(&event->owner_entry);
3411 mutex_unlock(&owner->perf_event_mutex);
3412 put_task_struct(owner);
3417 * Called when the last reference to the file is gone.
3419 static void put_event(struct perf_event *event)
3421 struct perf_event_context *ctx = event->ctx;
3423 if (!atomic_long_dec_and_test(&event->refcount))
3426 if (!is_kernel_event(event))
3427 perf_remove_from_owner(event);
3429 WARN_ON_ONCE(ctx->parent_ctx);
3431 * There are two ways this annotation is useful:
3433 * 1) there is a lock recursion from perf_event_exit_task
3434 * see the comment there.
3436 * 2) there is a lock-inversion with mmap_sem through
3437 * perf_event_read_group(), which takes faults while
3438 * holding ctx->mutex, however this is called after
3439 * the last filedesc died, so there is no possibility
3440 * to trigger the AB-BA case.
3442 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3443 perf_remove_from_context(event, true);
3444 mutex_unlock(&ctx->mutex);
3449 int perf_event_release_kernel(struct perf_event *event)
3454 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3456 static int perf_release(struct inode *inode, struct file *file)
3458 put_event(file->private_data);
3463 * Remove all orphanes events from the context.
3465 static void orphans_remove_work(struct work_struct *work)
3467 struct perf_event_context *ctx;
3468 struct perf_event *event, *tmp;
3470 ctx = container_of(work, struct perf_event_context,
3471 orphans_remove.work);
3473 mutex_lock(&ctx->mutex);
3474 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3475 struct perf_event *parent_event = event->parent;
3477 if (!is_orphaned_child(event))
3480 perf_remove_from_context(event, true);
3482 mutex_lock(&parent_event->child_mutex);
3483 list_del_init(&event->child_list);
3484 mutex_unlock(&parent_event->child_mutex);
3487 put_event(parent_event);
3490 raw_spin_lock_irq(&ctx->lock);
3491 ctx->orphans_remove_sched = false;
3492 raw_spin_unlock_irq(&ctx->lock);
3493 mutex_unlock(&ctx->mutex);
3498 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3500 struct perf_event *child;
3506 mutex_lock(&event->child_mutex);
3507 total += perf_event_read(event);
3508 *enabled += event->total_time_enabled +
3509 atomic64_read(&event->child_total_time_enabled);
3510 *running += event->total_time_running +
3511 atomic64_read(&event->child_total_time_running);
3513 list_for_each_entry(child, &event->child_list, child_list) {
3514 total += perf_event_read(child);
3515 *enabled += child->total_time_enabled;
3516 *running += child->total_time_running;
3518 mutex_unlock(&event->child_mutex);
3522 EXPORT_SYMBOL_GPL(perf_event_read_value);
3524 static int perf_event_read_group(struct perf_event *event,
3525 u64 read_format, char __user *buf)
3527 struct perf_event *leader = event->group_leader, *sub;
3528 int n = 0, size = 0, ret = -EFAULT;
3529 struct perf_event_context *ctx = leader->ctx;
3531 u64 count, enabled, running;
3533 mutex_lock(&ctx->mutex);
3534 count = perf_event_read_value(leader, &enabled, &running);
3536 values[n++] = 1 + leader->nr_siblings;
3537 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3538 values[n++] = enabled;
3539 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3540 values[n++] = running;
3541 values[n++] = count;
3542 if (read_format & PERF_FORMAT_ID)
3543 values[n++] = primary_event_id(leader);
3545 size = n * sizeof(u64);
3547 if (copy_to_user(buf, values, size))
3552 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3555 values[n++] = perf_event_read_value(sub, &enabled, &running);
3556 if (read_format & PERF_FORMAT_ID)
3557 values[n++] = primary_event_id(sub);
3559 size = n * sizeof(u64);
3561 if (copy_to_user(buf + ret, values, size)) {
3569 mutex_unlock(&ctx->mutex);
3574 static int perf_event_read_one(struct perf_event *event,
3575 u64 read_format, char __user *buf)
3577 u64 enabled, running;
3581 values[n++] = perf_event_read_value(event, &enabled, &running);
3582 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3583 values[n++] = enabled;
3584 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3585 values[n++] = running;
3586 if (read_format & PERF_FORMAT_ID)
3587 values[n++] = primary_event_id(event);
3589 if (copy_to_user(buf, values, n * sizeof(u64)))
3592 return n * sizeof(u64);
3595 static bool is_event_hup(struct perf_event *event)
3599 if (event->state != PERF_EVENT_STATE_EXIT)
3602 mutex_lock(&event->child_mutex);
3603 no_children = list_empty(&event->child_list);
3604 mutex_unlock(&event->child_mutex);
3609 * Read the performance event - simple non blocking version for now
3612 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3614 u64 read_format = event->attr.read_format;
3618 * Return end-of-file for a read on a event that is in
3619 * error state (i.e. because it was pinned but it couldn't be
3620 * scheduled on to the CPU at some point).
3622 if (event->state == PERF_EVENT_STATE_ERROR)
3625 if (count < event->read_size)
3628 WARN_ON_ONCE(event->ctx->parent_ctx);
3629 if (read_format & PERF_FORMAT_GROUP)
3630 ret = perf_event_read_group(event, read_format, buf);
3632 ret = perf_event_read_one(event, read_format, buf);
3638 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3640 struct perf_event *event = file->private_data;
3642 return perf_read_hw(event, buf, count);
3645 static unsigned int perf_poll(struct file *file, poll_table *wait)
3647 struct perf_event *event = file->private_data;
3648 struct ring_buffer *rb;
3649 unsigned int events = POLLHUP;
3651 poll_wait(file, &event->waitq, wait);
3653 if (is_event_hup(event))
3657 * Pin the event->rb by taking event->mmap_mutex; otherwise
3658 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3660 mutex_lock(&event->mmap_mutex);
3663 events = atomic_xchg(&rb->poll, 0);
3664 mutex_unlock(&event->mmap_mutex);
3668 static void perf_event_reset(struct perf_event *event)
3670 (void)perf_event_read(event);
3671 local64_set(&event->count, 0);
3672 perf_event_update_userpage(event);
3676 * Holding the top-level event's child_mutex means that any
3677 * descendant process that has inherited this event will block
3678 * in sync_child_event if it goes to exit, thus satisfying the
3679 * task existence requirements of perf_event_enable/disable.
3681 static void perf_event_for_each_child(struct perf_event *event,
3682 void (*func)(struct perf_event *))
3684 struct perf_event *child;
3686 WARN_ON_ONCE(event->ctx->parent_ctx);
3687 mutex_lock(&event->child_mutex);
3689 list_for_each_entry(child, &event->child_list, child_list)
3691 mutex_unlock(&event->child_mutex);
3694 static void perf_event_for_each(struct perf_event *event,
3695 void (*func)(struct perf_event *))
3697 struct perf_event_context *ctx = event->ctx;
3698 struct perf_event *sibling;
3700 WARN_ON_ONCE(ctx->parent_ctx);
3701 mutex_lock(&ctx->mutex);
3702 event = event->group_leader;
3704 perf_event_for_each_child(event, func);
3705 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3706 perf_event_for_each_child(sibling, func);
3707 mutex_unlock(&ctx->mutex);
3710 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3712 struct perf_event_context *ctx = event->ctx;
3713 int ret = 0, active;
3716 if (!is_sampling_event(event))
3719 if (copy_from_user(&value, arg, sizeof(value)))
3725 raw_spin_lock_irq(&ctx->lock);
3726 if (event->attr.freq) {
3727 if (value > sysctl_perf_event_sample_rate) {
3732 event->attr.sample_freq = value;
3734 event->attr.sample_period = value;
3735 event->hw.sample_period = value;
3738 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3740 perf_pmu_disable(ctx->pmu);
3741 event->pmu->stop(event, PERF_EF_UPDATE);
3744 local64_set(&event->hw.period_left, 0);
3747 event->pmu->start(event, PERF_EF_RELOAD);
3748 perf_pmu_enable(ctx->pmu);
3752 raw_spin_unlock_irq(&ctx->lock);
3757 static const struct file_operations perf_fops;
3759 static inline int perf_fget_light(int fd, struct fd *p)
3761 struct fd f = fdget(fd);
3765 if (f.file->f_op != &perf_fops) {
3773 static int perf_event_set_output(struct perf_event *event,
3774 struct perf_event *output_event);
3775 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3777 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3779 struct perf_event *event = file->private_data;
3780 void (*func)(struct perf_event *);
3784 case PERF_EVENT_IOC_ENABLE:
3785 func = perf_event_enable;
3787 case PERF_EVENT_IOC_DISABLE:
3788 func = perf_event_disable;
3790 case PERF_EVENT_IOC_RESET:
3791 func = perf_event_reset;
3794 case PERF_EVENT_IOC_REFRESH:
3795 return perf_event_refresh(event, arg);
3797 case PERF_EVENT_IOC_PERIOD:
3798 return perf_event_period(event, (u64 __user *)arg);
3800 case PERF_EVENT_IOC_ID:
3802 u64 id = primary_event_id(event);
3804 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3809 case PERF_EVENT_IOC_SET_OUTPUT:
3813 struct perf_event *output_event;
3815 ret = perf_fget_light(arg, &output);
3818 output_event = output.file->private_data;
3819 ret = perf_event_set_output(event, output_event);
3822 ret = perf_event_set_output(event, NULL);
3827 case PERF_EVENT_IOC_SET_FILTER:
3828 return perf_event_set_filter(event, (void __user *)arg);
3834 if (flags & PERF_IOC_FLAG_GROUP)
3835 perf_event_for_each(event, func);
3837 perf_event_for_each_child(event, func);
3842 #ifdef CONFIG_COMPAT
3843 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
3846 switch (_IOC_NR(cmd)) {
3847 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
3848 case _IOC_NR(PERF_EVENT_IOC_ID):
3849 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
3850 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
3851 cmd &= ~IOCSIZE_MASK;
3852 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
3856 return perf_ioctl(file, cmd, arg);
3859 # define perf_compat_ioctl NULL
3862 int perf_event_task_enable(void)
3864 struct perf_event *event;
3866 mutex_lock(¤t->perf_event_mutex);
3867 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3868 perf_event_for_each_child(event, perf_event_enable);
3869 mutex_unlock(¤t->perf_event_mutex);
3874 int perf_event_task_disable(void)
3876 struct perf_event *event;
3878 mutex_lock(¤t->perf_event_mutex);
3879 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3880 perf_event_for_each_child(event, perf_event_disable);
3881 mutex_unlock(¤t->perf_event_mutex);
3886 static int perf_event_index(struct perf_event *event)
3888 if (event->hw.state & PERF_HES_STOPPED)
3891 if (event->state != PERF_EVENT_STATE_ACTIVE)
3894 return event->pmu->event_idx(event);
3897 static void calc_timer_values(struct perf_event *event,
3904 *now = perf_clock();
3905 ctx_time = event->shadow_ctx_time + *now;
3906 *enabled = ctx_time - event->tstamp_enabled;
3907 *running = ctx_time - event->tstamp_running;
3910 static void perf_event_init_userpage(struct perf_event *event)
3912 struct perf_event_mmap_page *userpg;
3913 struct ring_buffer *rb;
3916 rb = rcu_dereference(event->rb);
3920 userpg = rb->user_page;
3922 /* Allow new userspace to detect that bit 0 is deprecated */
3923 userpg->cap_bit0_is_deprecated = 1;
3924 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3930 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3935 * Callers need to ensure there can be no nesting of this function, otherwise
3936 * the seqlock logic goes bad. We can not serialize this because the arch
3937 * code calls this from NMI context.
3939 void perf_event_update_userpage(struct perf_event *event)
3941 struct perf_event_mmap_page *userpg;
3942 struct ring_buffer *rb;
3943 u64 enabled, running, now;
3946 rb = rcu_dereference(event->rb);
3951 * compute total_time_enabled, total_time_running
3952 * based on snapshot values taken when the event
3953 * was last scheduled in.
3955 * we cannot simply called update_context_time()
3956 * because of locking issue as we can be called in
3959 calc_timer_values(event, &now, &enabled, &running);
3961 userpg = rb->user_page;
3963 * Disable preemption so as to not let the corresponding user-space
3964 * spin too long if we get preempted.
3969 userpg->index = perf_event_index(event);
3970 userpg->offset = perf_event_count(event);
3972 userpg->offset -= local64_read(&event->hw.prev_count);
3974 userpg->time_enabled = enabled +
3975 atomic64_read(&event->child_total_time_enabled);
3977 userpg->time_running = running +
3978 atomic64_read(&event->child_total_time_running);
3980 arch_perf_update_userpage(userpg, now);
3989 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3991 struct perf_event *event = vma->vm_file->private_data;
3992 struct ring_buffer *rb;
3993 int ret = VM_FAULT_SIGBUS;
3995 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3996 if (vmf->pgoff == 0)
4002 rb = rcu_dereference(event->rb);
4006 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4009 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4013 get_page(vmf->page);
4014 vmf->page->mapping = vma->vm_file->f_mapping;
4015 vmf->page->index = vmf->pgoff;
4024 static void ring_buffer_attach(struct perf_event *event,
4025 struct ring_buffer *rb)
4027 struct ring_buffer *old_rb = NULL;
4028 unsigned long flags;
4032 * Should be impossible, we set this when removing
4033 * event->rb_entry and wait/clear when adding event->rb_entry.
4035 WARN_ON_ONCE(event->rcu_pending);
4038 event->rcu_batches = get_state_synchronize_rcu();
4039 event->rcu_pending = 1;
4041 spin_lock_irqsave(&old_rb->event_lock, flags);
4042 list_del_rcu(&event->rb_entry);
4043 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4046 if (event->rcu_pending && rb) {
4047 cond_synchronize_rcu(event->rcu_batches);
4048 event->rcu_pending = 0;
4052 spin_lock_irqsave(&rb->event_lock, flags);
4053 list_add_rcu(&event->rb_entry, &rb->event_list);
4054 spin_unlock_irqrestore(&rb->event_lock, flags);
4057 rcu_assign_pointer(event->rb, rb);
4060 ring_buffer_put(old_rb);
4062 * Since we detached before setting the new rb, so that we
4063 * could attach the new rb, we could have missed a wakeup.
4066 wake_up_all(&event->waitq);
4070 static void ring_buffer_wakeup(struct perf_event *event)
4072 struct ring_buffer *rb;
4075 rb = rcu_dereference(event->rb);
4077 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4078 wake_up_all(&event->waitq);
4083 static void rb_free_rcu(struct rcu_head *rcu_head)
4085 struct ring_buffer *rb;
4087 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4091 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
4093 struct ring_buffer *rb;
4096 rb = rcu_dereference(event->rb);
4098 if (!atomic_inc_not_zero(&rb->refcount))
4106 static void ring_buffer_put(struct ring_buffer *rb)
4108 if (!atomic_dec_and_test(&rb->refcount))
4111 WARN_ON_ONCE(!list_empty(&rb->event_list));
4113 call_rcu(&rb->rcu_head, rb_free_rcu);
4116 static void perf_mmap_open(struct vm_area_struct *vma)
4118 struct perf_event *event = vma->vm_file->private_data;
4120 atomic_inc(&event->mmap_count);
4121 atomic_inc(&event->rb->mmap_count);
4125 * A buffer can be mmap()ed multiple times; either directly through the same
4126 * event, or through other events by use of perf_event_set_output().
4128 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4129 * the buffer here, where we still have a VM context. This means we need
4130 * to detach all events redirecting to us.
4132 static void perf_mmap_close(struct vm_area_struct *vma)
4134 struct perf_event *event = vma->vm_file->private_data;
4136 struct ring_buffer *rb = ring_buffer_get(event);
4137 struct user_struct *mmap_user = rb->mmap_user;
4138 int mmap_locked = rb->mmap_locked;
4139 unsigned long size = perf_data_size(rb);
4141 atomic_dec(&rb->mmap_count);
4143 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4146 ring_buffer_attach(event, NULL);
4147 mutex_unlock(&event->mmap_mutex);
4149 /* If there's still other mmap()s of this buffer, we're done. */
4150 if (atomic_read(&rb->mmap_count))
4154 * No other mmap()s, detach from all other events that might redirect
4155 * into the now unreachable buffer. Somewhat complicated by the
4156 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4160 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4161 if (!atomic_long_inc_not_zero(&event->refcount)) {
4163 * This event is en-route to free_event() which will
4164 * detach it and remove it from the list.
4170 mutex_lock(&event->mmap_mutex);
4172 * Check we didn't race with perf_event_set_output() which can
4173 * swizzle the rb from under us while we were waiting to
4174 * acquire mmap_mutex.
4176 * If we find a different rb; ignore this event, a next
4177 * iteration will no longer find it on the list. We have to
4178 * still restart the iteration to make sure we're not now
4179 * iterating the wrong list.
4181 if (event->rb == rb)
4182 ring_buffer_attach(event, NULL);
4184 mutex_unlock(&event->mmap_mutex);
4188 * Restart the iteration; either we're on the wrong list or
4189 * destroyed its integrity by doing a deletion.
4196 * It could be there's still a few 0-ref events on the list; they'll
4197 * get cleaned up by free_event() -- they'll also still have their
4198 * ref on the rb and will free it whenever they are done with it.
4200 * Aside from that, this buffer is 'fully' detached and unmapped,
4201 * undo the VM accounting.
4204 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4205 vma->vm_mm->pinned_vm -= mmap_locked;
4206 free_uid(mmap_user);
4209 ring_buffer_put(rb); /* could be last */
4212 static const struct vm_operations_struct perf_mmap_vmops = {
4213 .open = perf_mmap_open,
4214 .close = perf_mmap_close,
4215 .fault = perf_mmap_fault,
4216 .page_mkwrite = perf_mmap_fault,
4219 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4221 struct perf_event *event = file->private_data;
4222 unsigned long user_locked, user_lock_limit;
4223 struct user_struct *user = current_user();
4224 unsigned long locked, lock_limit;
4225 struct ring_buffer *rb;
4226 unsigned long vma_size;
4227 unsigned long nr_pages;
4228 long user_extra, extra;
4229 int ret = 0, flags = 0;
4232 * Don't allow mmap() of inherited per-task counters. This would
4233 * create a performance issue due to all children writing to the
4236 if (event->cpu == -1 && event->attr.inherit)
4239 if (!(vma->vm_flags & VM_SHARED))
4242 vma_size = vma->vm_end - vma->vm_start;
4243 nr_pages = (vma_size / PAGE_SIZE) - 1;
4246 * If we have rb pages ensure they're a power-of-two number, so we
4247 * can do bitmasks instead of modulo.
4249 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4252 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4255 if (vma->vm_pgoff != 0)
4258 WARN_ON_ONCE(event->ctx->parent_ctx);
4260 mutex_lock(&event->mmap_mutex);
4262 if (event->rb->nr_pages != nr_pages) {
4267 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4269 * Raced against perf_mmap_close() through
4270 * perf_event_set_output(). Try again, hope for better
4273 mutex_unlock(&event->mmap_mutex);
4280 user_extra = nr_pages + 1;
4281 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4284 * Increase the limit linearly with more CPUs:
4286 user_lock_limit *= num_online_cpus();
4288 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4291 if (user_locked > user_lock_limit)
4292 extra = user_locked - user_lock_limit;
4294 lock_limit = rlimit(RLIMIT_MEMLOCK);
4295 lock_limit >>= PAGE_SHIFT;
4296 locked = vma->vm_mm->pinned_vm + extra;
4298 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4299 !capable(CAP_IPC_LOCK)) {
4306 if (vma->vm_flags & VM_WRITE)
4307 flags |= RING_BUFFER_WRITABLE;
4309 rb = rb_alloc(nr_pages,
4310 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4318 atomic_set(&rb->mmap_count, 1);
4319 rb->mmap_locked = extra;
4320 rb->mmap_user = get_current_user();
4322 atomic_long_add(user_extra, &user->locked_vm);
4323 vma->vm_mm->pinned_vm += extra;
4325 ring_buffer_attach(event, rb);
4327 perf_event_init_userpage(event);
4328 perf_event_update_userpage(event);
4332 atomic_inc(&event->mmap_count);
4333 mutex_unlock(&event->mmap_mutex);
4336 * Since pinned accounting is per vm we cannot allow fork() to copy our
4339 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4340 vma->vm_ops = &perf_mmap_vmops;
4345 static int perf_fasync(int fd, struct file *filp, int on)
4347 struct inode *inode = file_inode(filp);
4348 struct perf_event *event = filp->private_data;
4351 mutex_lock(&inode->i_mutex);
4352 retval = fasync_helper(fd, filp, on, &event->fasync);
4353 mutex_unlock(&inode->i_mutex);
4361 static const struct file_operations perf_fops = {
4362 .llseek = no_llseek,
4363 .release = perf_release,
4366 .unlocked_ioctl = perf_ioctl,
4367 .compat_ioctl = perf_compat_ioctl,
4369 .fasync = perf_fasync,
4375 * If there's data, ensure we set the poll() state and publish everything
4376 * to user-space before waking everybody up.
4379 void perf_event_wakeup(struct perf_event *event)
4381 ring_buffer_wakeup(event);
4383 if (event->pending_kill) {
4384 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4385 event->pending_kill = 0;
4389 static void perf_pending_event(struct irq_work *entry)
4391 struct perf_event *event = container_of(entry,
4392 struct perf_event, pending);
4394 if (event->pending_disable) {
4395 event->pending_disable = 0;
4396 __perf_event_disable(event);
4399 if (event->pending_wakeup) {
4400 event->pending_wakeup = 0;
4401 perf_event_wakeup(event);
4406 * We assume there is only KVM supporting the callbacks.
4407 * Later on, we might change it to a list if there is
4408 * another virtualization implementation supporting the callbacks.
4410 struct perf_guest_info_callbacks *perf_guest_cbs;
4412 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4414 perf_guest_cbs = cbs;
4417 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4419 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4421 perf_guest_cbs = NULL;
4424 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4427 perf_output_sample_regs(struct perf_output_handle *handle,
4428 struct pt_regs *regs, u64 mask)
4432 for_each_set_bit(bit, (const unsigned long *) &mask,
4433 sizeof(mask) * BITS_PER_BYTE) {
4436 val = perf_reg_value(regs, bit);
4437 perf_output_put(handle, val);
4441 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4442 struct pt_regs *regs)
4444 if (!user_mode(regs)) {
4446 regs = task_pt_regs(current);
4452 regs_user->regs = regs;
4453 regs_user->abi = perf_reg_abi(current);
4458 * Get remaining task size from user stack pointer.
4460 * It'd be better to take stack vma map and limit this more
4461 * precisly, but there's no way to get it safely under interrupt,
4462 * so using TASK_SIZE as limit.
4464 static u64 perf_ustack_task_size(struct pt_regs *regs)
4466 unsigned long addr = perf_user_stack_pointer(regs);
4468 if (!addr || addr >= TASK_SIZE)
4471 return TASK_SIZE - addr;
4475 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4476 struct pt_regs *regs)
4480 /* No regs, no stack pointer, no dump. */
4485 * Check if we fit in with the requested stack size into the:
4487 * If we don't, we limit the size to the TASK_SIZE.
4489 * - remaining sample size
4490 * If we don't, we customize the stack size to
4491 * fit in to the remaining sample size.
4494 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4495 stack_size = min(stack_size, (u16) task_size);
4497 /* Current header size plus static size and dynamic size. */
4498 header_size += 2 * sizeof(u64);
4500 /* Do we fit in with the current stack dump size? */
4501 if ((u16) (header_size + stack_size) < header_size) {
4503 * If we overflow the maximum size for the sample,
4504 * we customize the stack dump size to fit in.
4506 stack_size = USHRT_MAX - header_size - sizeof(u64);
4507 stack_size = round_up(stack_size, sizeof(u64));
4514 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4515 struct pt_regs *regs)
4517 /* Case of a kernel thread, nothing to dump */
4520 perf_output_put(handle, size);
4529 * - the size requested by user or the best one we can fit
4530 * in to the sample max size
4532 * - user stack dump data
4534 * - the actual dumped size
4538 perf_output_put(handle, dump_size);
4541 sp = perf_user_stack_pointer(regs);
4542 rem = __output_copy_user(handle, (void *) sp, dump_size);
4543 dyn_size = dump_size - rem;
4545 perf_output_skip(handle, rem);
4548 perf_output_put(handle, dyn_size);
4552 static void __perf_event_header__init_id(struct perf_event_header *header,
4553 struct perf_sample_data *data,
4554 struct perf_event *event)
4556 u64 sample_type = event->attr.sample_type;
4558 data->type = sample_type;
4559 header->size += event->id_header_size;
4561 if (sample_type & PERF_SAMPLE_TID) {
4562 /* namespace issues */
4563 data->tid_entry.pid = perf_event_pid(event, current);
4564 data->tid_entry.tid = perf_event_tid(event, current);
4567 if (sample_type & PERF_SAMPLE_TIME)
4568 data->time = perf_clock();
4570 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4571 data->id = primary_event_id(event);
4573 if (sample_type & PERF_SAMPLE_STREAM_ID)
4574 data->stream_id = event->id;
4576 if (sample_type & PERF_SAMPLE_CPU) {
4577 data->cpu_entry.cpu = raw_smp_processor_id();
4578 data->cpu_entry.reserved = 0;
4582 void perf_event_header__init_id(struct perf_event_header *header,
4583 struct perf_sample_data *data,
4584 struct perf_event *event)
4586 if (event->attr.sample_id_all)
4587 __perf_event_header__init_id(header, data, event);
4590 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4591 struct perf_sample_data *data)
4593 u64 sample_type = data->type;
4595 if (sample_type & PERF_SAMPLE_TID)
4596 perf_output_put(handle, data->tid_entry);
4598 if (sample_type & PERF_SAMPLE_TIME)
4599 perf_output_put(handle, data->time);
4601 if (sample_type & PERF_SAMPLE_ID)
4602 perf_output_put(handle, data->id);
4604 if (sample_type & PERF_SAMPLE_STREAM_ID)
4605 perf_output_put(handle, data->stream_id);
4607 if (sample_type & PERF_SAMPLE_CPU)
4608 perf_output_put(handle, data->cpu_entry);
4610 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4611 perf_output_put(handle, data->id);
4614 void perf_event__output_id_sample(struct perf_event *event,
4615 struct perf_output_handle *handle,
4616 struct perf_sample_data *sample)
4618 if (event->attr.sample_id_all)
4619 __perf_event__output_id_sample(handle, sample);
4622 static void perf_output_read_one(struct perf_output_handle *handle,
4623 struct perf_event *event,
4624 u64 enabled, u64 running)
4626 u64 read_format = event->attr.read_format;
4630 values[n++] = perf_event_count(event);
4631 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4632 values[n++] = enabled +
4633 atomic64_read(&event->child_total_time_enabled);
4635 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4636 values[n++] = running +
4637 atomic64_read(&event->child_total_time_running);
4639 if (read_format & PERF_FORMAT_ID)
4640 values[n++] = primary_event_id(event);
4642 __output_copy(handle, values, n * sizeof(u64));
4646 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4648 static void perf_output_read_group(struct perf_output_handle *handle,
4649 struct perf_event *event,
4650 u64 enabled, u64 running)
4652 struct perf_event *leader = event->group_leader, *sub;
4653 u64 read_format = event->attr.read_format;
4657 values[n++] = 1 + leader->nr_siblings;
4659 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4660 values[n++] = enabled;
4662 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4663 values[n++] = running;
4665 if (leader != event)
4666 leader->pmu->read(leader);
4668 values[n++] = perf_event_count(leader);
4669 if (read_format & PERF_FORMAT_ID)
4670 values[n++] = primary_event_id(leader);
4672 __output_copy(handle, values, n * sizeof(u64));
4674 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4677 if ((sub != event) &&
4678 (sub->state == PERF_EVENT_STATE_ACTIVE))
4679 sub->pmu->read(sub);
4681 values[n++] = perf_event_count(sub);
4682 if (read_format & PERF_FORMAT_ID)
4683 values[n++] = primary_event_id(sub);
4685 __output_copy(handle, values, n * sizeof(u64));
4689 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4690 PERF_FORMAT_TOTAL_TIME_RUNNING)
4692 static void perf_output_read(struct perf_output_handle *handle,
4693 struct perf_event *event)
4695 u64 enabled = 0, running = 0, now;
4696 u64 read_format = event->attr.read_format;
4699 * compute total_time_enabled, total_time_running
4700 * based on snapshot values taken when the event
4701 * was last scheduled in.
4703 * we cannot simply called update_context_time()
4704 * because of locking issue as we are called in
4707 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4708 calc_timer_values(event, &now, &enabled, &running);
4710 if (event->attr.read_format & PERF_FORMAT_GROUP)
4711 perf_output_read_group(handle, event, enabled, running);
4713 perf_output_read_one(handle, event, enabled, running);
4716 void perf_output_sample(struct perf_output_handle *handle,
4717 struct perf_event_header *header,
4718 struct perf_sample_data *data,
4719 struct perf_event *event)
4721 u64 sample_type = data->type;
4723 perf_output_put(handle, *header);
4725 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4726 perf_output_put(handle, data->id);
4728 if (sample_type & PERF_SAMPLE_IP)
4729 perf_output_put(handle, data->ip);
4731 if (sample_type & PERF_SAMPLE_TID)
4732 perf_output_put(handle, data->tid_entry);
4734 if (sample_type & PERF_SAMPLE_TIME)
4735 perf_output_put(handle, data->time);
4737 if (sample_type & PERF_SAMPLE_ADDR)
4738 perf_output_put(handle, data->addr);
4740 if (sample_type & PERF_SAMPLE_ID)
4741 perf_output_put(handle, data->id);
4743 if (sample_type & PERF_SAMPLE_STREAM_ID)
4744 perf_output_put(handle, data->stream_id);
4746 if (sample_type & PERF_SAMPLE_CPU)
4747 perf_output_put(handle, data->cpu_entry);
4749 if (sample_type & PERF_SAMPLE_PERIOD)
4750 perf_output_put(handle, data->period);
4752 if (sample_type & PERF_SAMPLE_READ)
4753 perf_output_read(handle, event);
4755 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4756 if (data->callchain) {
4759 if (data->callchain)
4760 size += data->callchain->nr;
4762 size *= sizeof(u64);
4764 __output_copy(handle, data->callchain, size);
4767 perf_output_put(handle, nr);
4771 if (sample_type & PERF_SAMPLE_RAW) {
4773 perf_output_put(handle, data->raw->size);
4774 __output_copy(handle, data->raw->data,
4781 .size = sizeof(u32),
4784 perf_output_put(handle, raw);
4788 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4789 if (data->br_stack) {
4792 size = data->br_stack->nr
4793 * sizeof(struct perf_branch_entry);
4795 perf_output_put(handle, data->br_stack->nr);
4796 perf_output_copy(handle, data->br_stack->entries, size);
4799 * we always store at least the value of nr
4802 perf_output_put(handle, nr);
4806 if (sample_type & PERF_SAMPLE_REGS_USER) {
4807 u64 abi = data->regs_user.abi;
4810 * If there are no regs to dump, notice it through
4811 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4813 perf_output_put(handle, abi);
4816 u64 mask = event->attr.sample_regs_user;
4817 perf_output_sample_regs(handle,
4818 data->regs_user.regs,
4823 if (sample_type & PERF_SAMPLE_STACK_USER) {
4824 perf_output_sample_ustack(handle,
4825 data->stack_user_size,
4826 data->regs_user.regs);
4829 if (sample_type & PERF_SAMPLE_WEIGHT)
4830 perf_output_put(handle, data->weight);
4832 if (sample_type & PERF_SAMPLE_DATA_SRC)
4833 perf_output_put(handle, data->data_src.val);
4835 if (sample_type & PERF_SAMPLE_TRANSACTION)
4836 perf_output_put(handle, data->txn);
4838 if (!event->attr.watermark) {
4839 int wakeup_events = event->attr.wakeup_events;
4841 if (wakeup_events) {
4842 struct ring_buffer *rb = handle->rb;
4843 int events = local_inc_return(&rb->events);
4845 if (events >= wakeup_events) {
4846 local_sub(wakeup_events, &rb->events);
4847 local_inc(&rb->wakeup);
4853 void perf_prepare_sample(struct perf_event_header *header,
4854 struct perf_sample_data *data,
4855 struct perf_event *event,
4856 struct pt_regs *regs)
4858 u64 sample_type = event->attr.sample_type;
4860 header->type = PERF_RECORD_SAMPLE;
4861 header->size = sizeof(*header) + event->header_size;
4864 header->misc |= perf_misc_flags(regs);
4866 __perf_event_header__init_id(header, data, event);
4868 if (sample_type & PERF_SAMPLE_IP)
4869 data->ip = perf_instruction_pointer(regs);
4871 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4874 data->callchain = perf_callchain(event, regs);
4876 if (data->callchain)
4877 size += data->callchain->nr;
4879 header->size += size * sizeof(u64);
4882 if (sample_type & PERF_SAMPLE_RAW) {
4883 int size = sizeof(u32);
4886 size += data->raw->size;
4888 size += sizeof(u32);
4890 WARN_ON_ONCE(size & (sizeof(u64)-1));
4891 header->size += size;
4894 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4895 int size = sizeof(u64); /* nr */
4896 if (data->br_stack) {
4897 size += data->br_stack->nr
4898 * sizeof(struct perf_branch_entry);
4900 header->size += size;
4903 if (sample_type & PERF_SAMPLE_REGS_USER) {
4904 /* regs dump ABI info */
4905 int size = sizeof(u64);
4907 perf_sample_regs_user(&data->regs_user, regs);
4909 if (data->regs_user.regs) {
4910 u64 mask = event->attr.sample_regs_user;
4911 size += hweight64(mask) * sizeof(u64);
4914 header->size += size;
4917 if (sample_type & PERF_SAMPLE_STACK_USER) {
4919 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4920 * processed as the last one or have additional check added
4921 * in case new sample type is added, because we could eat
4922 * up the rest of the sample size.
4924 struct perf_regs_user *uregs = &data->regs_user;
4925 u16 stack_size = event->attr.sample_stack_user;
4926 u16 size = sizeof(u64);
4929 perf_sample_regs_user(uregs, regs);
4931 stack_size = perf_sample_ustack_size(stack_size, header->size,
4935 * If there is something to dump, add space for the dump
4936 * itself and for the field that tells the dynamic size,
4937 * which is how many have been actually dumped.
4940 size += sizeof(u64) + stack_size;
4942 data->stack_user_size = stack_size;
4943 header->size += size;
4947 static void perf_event_output(struct perf_event *event,
4948 struct perf_sample_data *data,
4949 struct pt_regs *regs)
4951 struct perf_output_handle handle;
4952 struct perf_event_header header;
4954 /* protect the callchain buffers */
4957 perf_prepare_sample(&header, data, event, regs);
4959 if (perf_output_begin(&handle, event, header.size))
4962 perf_output_sample(&handle, &header, data, event);
4964 perf_output_end(&handle);
4974 struct perf_read_event {
4975 struct perf_event_header header;
4982 perf_event_read_event(struct perf_event *event,
4983 struct task_struct *task)
4985 struct perf_output_handle handle;
4986 struct perf_sample_data sample;
4987 struct perf_read_event read_event = {
4989 .type = PERF_RECORD_READ,
4991 .size = sizeof(read_event) + event->read_size,
4993 .pid = perf_event_pid(event, task),
4994 .tid = perf_event_tid(event, task),
4998 perf_event_header__init_id(&read_event.header, &sample, event);
4999 ret = perf_output_begin(&handle, event, read_event.header.size);
5003 perf_output_put(&handle, read_event);
5004 perf_output_read(&handle, event);
5005 perf_event__output_id_sample(event, &handle, &sample);
5007 perf_output_end(&handle);
5010 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5013 perf_event_aux_ctx(struct perf_event_context *ctx,
5014 perf_event_aux_output_cb output,
5017 struct perf_event *event;
5019 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5020 if (event->state < PERF_EVENT_STATE_INACTIVE)
5022 if (!event_filter_match(event))
5024 output(event, data);
5029 perf_event_aux(perf_event_aux_output_cb output, void *data,
5030 struct perf_event_context *task_ctx)
5032 struct perf_cpu_context *cpuctx;
5033 struct perf_event_context *ctx;
5038 list_for_each_entry_rcu(pmu, &pmus, entry) {
5039 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5040 if (cpuctx->unique_pmu != pmu)
5042 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5045 ctxn = pmu->task_ctx_nr;
5048 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5050 perf_event_aux_ctx(ctx, output, data);
5052 put_cpu_ptr(pmu->pmu_cpu_context);
5057 perf_event_aux_ctx(task_ctx, output, data);
5064 * task tracking -- fork/exit
5066 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5069 struct perf_task_event {
5070 struct task_struct *task;
5071 struct perf_event_context *task_ctx;
5074 struct perf_event_header header;
5084 static int perf_event_task_match(struct perf_event *event)
5086 return event->attr.comm || event->attr.mmap ||
5087 event->attr.mmap2 || event->attr.mmap_data ||
5091 static void perf_event_task_output(struct perf_event *event,
5094 struct perf_task_event *task_event = data;
5095 struct perf_output_handle handle;
5096 struct perf_sample_data sample;
5097 struct task_struct *task = task_event->task;
5098 int ret, size = task_event->event_id.header.size;
5100 if (!perf_event_task_match(event))
5103 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5105 ret = perf_output_begin(&handle, event,
5106 task_event->event_id.header.size);
5110 task_event->event_id.pid = perf_event_pid(event, task);
5111 task_event->event_id.ppid = perf_event_pid(event, current);
5113 task_event->event_id.tid = perf_event_tid(event, task);
5114 task_event->event_id.ptid = perf_event_tid(event, current);
5116 perf_output_put(&handle, task_event->event_id);
5118 perf_event__output_id_sample(event, &handle, &sample);
5120 perf_output_end(&handle);
5122 task_event->event_id.header.size = size;
5125 static void perf_event_task(struct task_struct *task,
5126 struct perf_event_context *task_ctx,
5129 struct perf_task_event task_event;
5131 if (!atomic_read(&nr_comm_events) &&
5132 !atomic_read(&nr_mmap_events) &&
5133 !atomic_read(&nr_task_events))
5136 task_event = (struct perf_task_event){
5138 .task_ctx = task_ctx,
5141 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5143 .size = sizeof(task_event.event_id),
5149 .time = perf_clock(),
5153 perf_event_aux(perf_event_task_output,
5158 void perf_event_fork(struct task_struct *task)
5160 perf_event_task(task, NULL, 1);
5167 struct perf_comm_event {
5168 struct task_struct *task;
5173 struct perf_event_header header;
5180 static int perf_event_comm_match(struct perf_event *event)
5182 return event->attr.comm;
5185 static void perf_event_comm_output(struct perf_event *event,
5188 struct perf_comm_event *comm_event = data;
5189 struct perf_output_handle handle;
5190 struct perf_sample_data sample;
5191 int size = comm_event->event_id.header.size;
5194 if (!perf_event_comm_match(event))
5197 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5198 ret = perf_output_begin(&handle, event,
5199 comm_event->event_id.header.size);
5204 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5205 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5207 perf_output_put(&handle, comm_event->event_id);
5208 __output_copy(&handle, comm_event->comm,
5209 comm_event->comm_size);
5211 perf_event__output_id_sample(event, &handle, &sample);
5213 perf_output_end(&handle);
5215 comm_event->event_id.header.size = size;
5218 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5220 char comm[TASK_COMM_LEN];
5223 memset(comm, 0, sizeof(comm));
5224 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5225 size = ALIGN(strlen(comm)+1, sizeof(u64));
5227 comm_event->comm = comm;
5228 comm_event->comm_size = size;
5230 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5232 perf_event_aux(perf_event_comm_output,
5237 void perf_event_comm(struct task_struct *task, bool exec)
5239 struct perf_comm_event comm_event;
5241 if (!atomic_read(&nr_comm_events))
5244 comm_event = (struct perf_comm_event){
5250 .type = PERF_RECORD_COMM,
5251 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5259 perf_event_comm_event(&comm_event);
5266 struct perf_mmap_event {
5267 struct vm_area_struct *vma;
5269 const char *file_name;
5277 struct perf_event_header header;
5287 static int perf_event_mmap_match(struct perf_event *event,
5290 struct perf_mmap_event *mmap_event = data;
5291 struct vm_area_struct *vma = mmap_event->vma;
5292 int executable = vma->vm_flags & VM_EXEC;
5294 return (!executable && event->attr.mmap_data) ||
5295 (executable && (event->attr.mmap || event->attr.mmap2));
5298 static void perf_event_mmap_output(struct perf_event *event,
5301 struct perf_mmap_event *mmap_event = data;
5302 struct perf_output_handle handle;
5303 struct perf_sample_data sample;
5304 int size = mmap_event->event_id.header.size;
5307 if (!perf_event_mmap_match(event, data))
5310 if (event->attr.mmap2) {
5311 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5312 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5313 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5314 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5315 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5316 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5317 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5320 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5321 ret = perf_output_begin(&handle, event,
5322 mmap_event->event_id.header.size);
5326 mmap_event->event_id.pid = perf_event_pid(event, current);
5327 mmap_event->event_id.tid = perf_event_tid(event, current);
5329 perf_output_put(&handle, mmap_event->event_id);
5331 if (event->attr.mmap2) {
5332 perf_output_put(&handle, mmap_event->maj);
5333 perf_output_put(&handle, mmap_event->min);
5334 perf_output_put(&handle, mmap_event->ino);
5335 perf_output_put(&handle, mmap_event->ino_generation);
5336 perf_output_put(&handle, mmap_event->prot);
5337 perf_output_put(&handle, mmap_event->flags);
5340 __output_copy(&handle, mmap_event->file_name,
5341 mmap_event->file_size);
5343 perf_event__output_id_sample(event, &handle, &sample);
5345 perf_output_end(&handle);
5347 mmap_event->event_id.header.size = size;
5350 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5352 struct vm_area_struct *vma = mmap_event->vma;
5353 struct file *file = vma->vm_file;
5354 int maj = 0, min = 0;
5355 u64 ino = 0, gen = 0;
5356 u32 prot = 0, flags = 0;
5363 struct inode *inode;
5366 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5372 * d_path() works from the end of the rb backwards, so we
5373 * need to add enough zero bytes after the string to handle
5374 * the 64bit alignment we do later.
5376 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5381 inode = file_inode(vma->vm_file);
5382 dev = inode->i_sb->s_dev;
5384 gen = inode->i_generation;
5388 if (vma->vm_flags & VM_READ)
5390 if (vma->vm_flags & VM_WRITE)
5392 if (vma->vm_flags & VM_EXEC)
5395 if (vma->vm_flags & VM_MAYSHARE)
5398 flags = MAP_PRIVATE;
5400 if (vma->vm_flags & VM_DENYWRITE)
5401 flags |= MAP_DENYWRITE;
5402 if (vma->vm_flags & VM_MAYEXEC)
5403 flags |= MAP_EXECUTABLE;
5404 if (vma->vm_flags & VM_LOCKED)
5405 flags |= MAP_LOCKED;
5406 if (vma->vm_flags & VM_HUGETLB)
5407 flags |= MAP_HUGETLB;
5411 if (vma->vm_ops && vma->vm_ops->name) {
5412 name = (char *) vma->vm_ops->name(vma);
5417 name = (char *)arch_vma_name(vma);
5421 if (vma->vm_start <= vma->vm_mm->start_brk &&
5422 vma->vm_end >= vma->vm_mm->brk) {
5426 if (vma->vm_start <= vma->vm_mm->start_stack &&
5427 vma->vm_end >= vma->vm_mm->start_stack) {
5437 strlcpy(tmp, name, sizeof(tmp));
5441 * Since our buffer works in 8 byte units we need to align our string
5442 * size to a multiple of 8. However, we must guarantee the tail end is
5443 * zero'd out to avoid leaking random bits to userspace.
5445 size = strlen(name)+1;
5446 while (!IS_ALIGNED(size, sizeof(u64)))
5447 name[size++] = '\0';
5449 mmap_event->file_name = name;
5450 mmap_event->file_size = size;
5451 mmap_event->maj = maj;
5452 mmap_event->min = min;
5453 mmap_event->ino = ino;
5454 mmap_event->ino_generation = gen;
5455 mmap_event->prot = prot;
5456 mmap_event->flags = flags;
5458 if (!(vma->vm_flags & VM_EXEC))
5459 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5461 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5463 perf_event_aux(perf_event_mmap_output,
5470 void perf_event_mmap(struct vm_area_struct *vma)
5472 struct perf_mmap_event mmap_event;
5474 if (!atomic_read(&nr_mmap_events))
5477 mmap_event = (struct perf_mmap_event){
5483 .type = PERF_RECORD_MMAP,
5484 .misc = PERF_RECORD_MISC_USER,
5489 .start = vma->vm_start,
5490 .len = vma->vm_end - vma->vm_start,
5491 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5493 /* .maj (attr_mmap2 only) */
5494 /* .min (attr_mmap2 only) */
5495 /* .ino (attr_mmap2 only) */
5496 /* .ino_generation (attr_mmap2 only) */
5497 /* .prot (attr_mmap2 only) */
5498 /* .flags (attr_mmap2 only) */
5501 perf_event_mmap_event(&mmap_event);
5505 * IRQ throttle logging
5508 static void perf_log_throttle(struct perf_event *event, int enable)
5510 struct perf_output_handle handle;
5511 struct perf_sample_data sample;
5515 struct perf_event_header header;
5519 } throttle_event = {
5521 .type = PERF_RECORD_THROTTLE,
5523 .size = sizeof(throttle_event),
5525 .time = perf_clock(),
5526 .id = primary_event_id(event),
5527 .stream_id = event->id,
5531 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5533 perf_event_header__init_id(&throttle_event.header, &sample, event);
5535 ret = perf_output_begin(&handle, event,
5536 throttle_event.header.size);
5540 perf_output_put(&handle, throttle_event);
5541 perf_event__output_id_sample(event, &handle, &sample);
5542 perf_output_end(&handle);
5546 * Generic event overflow handling, sampling.
5549 static int __perf_event_overflow(struct perf_event *event,
5550 int throttle, struct perf_sample_data *data,
5551 struct pt_regs *regs)
5553 int events = atomic_read(&event->event_limit);
5554 struct hw_perf_event *hwc = &event->hw;
5559 * Non-sampling counters might still use the PMI to fold short
5560 * hardware counters, ignore those.
5562 if (unlikely(!is_sampling_event(event)))
5565 seq = __this_cpu_read(perf_throttled_seq);
5566 if (seq != hwc->interrupts_seq) {
5567 hwc->interrupts_seq = seq;
5568 hwc->interrupts = 1;
5571 if (unlikely(throttle
5572 && hwc->interrupts >= max_samples_per_tick)) {
5573 __this_cpu_inc(perf_throttled_count);
5574 hwc->interrupts = MAX_INTERRUPTS;
5575 perf_log_throttle(event, 0);
5576 tick_nohz_full_kick();
5581 if (event->attr.freq) {
5582 u64 now = perf_clock();
5583 s64 delta = now - hwc->freq_time_stamp;
5585 hwc->freq_time_stamp = now;
5587 if (delta > 0 && delta < 2*TICK_NSEC)
5588 perf_adjust_period(event, delta, hwc->last_period, true);
5592 * XXX event_limit might not quite work as expected on inherited
5596 event->pending_kill = POLL_IN;
5597 if (events && atomic_dec_and_test(&event->event_limit)) {
5599 event->pending_kill = POLL_HUP;
5600 event->pending_disable = 1;
5601 irq_work_queue(&event->pending);
5604 if (event->overflow_handler)
5605 event->overflow_handler(event, data, regs);
5607 perf_event_output(event, data, regs);
5609 if (event->fasync && event->pending_kill) {
5610 event->pending_wakeup = 1;
5611 irq_work_queue(&event->pending);
5617 int perf_event_overflow(struct perf_event *event,
5618 struct perf_sample_data *data,
5619 struct pt_regs *regs)
5621 return __perf_event_overflow(event, 1, data, regs);
5625 * Generic software event infrastructure
5628 struct swevent_htable {
5629 struct swevent_hlist *swevent_hlist;
5630 struct mutex hlist_mutex;
5633 /* Recursion avoidance in each contexts */
5634 int recursion[PERF_NR_CONTEXTS];
5636 /* Keeps track of cpu being initialized/exited */
5640 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5643 * We directly increment event->count and keep a second value in
5644 * event->hw.period_left to count intervals. This period event
5645 * is kept in the range [-sample_period, 0] so that we can use the
5649 u64 perf_swevent_set_period(struct perf_event *event)
5651 struct hw_perf_event *hwc = &event->hw;
5652 u64 period = hwc->last_period;
5656 hwc->last_period = hwc->sample_period;
5659 old = val = local64_read(&hwc->period_left);
5663 nr = div64_u64(period + val, period);
5664 offset = nr * period;
5666 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5672 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5673 struct perf_sample_data *data,
5674 struct pt_regs *regs)
5676 struct hw_perf_event *hwc = &event->hw;
5680 overflow = perf_swevent_set_period(event);
5682 if (hwc->interrupts == MAX_INTERRUPTS)
5685 for (; overflow; overflow--) {
5686 if (__perf_event_overflow(event, throttle,
5689 * We inhibit the overflow from happening when
5690 * hwc->interrupts == MAX_INTERRUPTS.
5698 static void perf_swevent_event(struct perf_event *event, u64 nr,
5699 struct perf_sample_data *data,
5700 struct pt_regs *regs)
5702 struct hw_perf_event *hwc = &event->hw;
5704 local64_add(nr, &event->count);
5709 if (!is_sampling_event(event))
5712 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5714 return perf_swevent_overflow(event, 1, data, regs);
5716 data->period = event->hw.last_period;
5718 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5719 return perf_swevent_overflow(event, 1, data, regs);
5721 if (local64_add_negative(nr, &hwc->period_left))
5724 perf_swevent_overflow(event, 0, data, regs);
5727 static int perf_exclude_event(struct perf_event *event,
5728 struct pt_regs *regs)
5730 if (event->hw.state & PERF_HES_STOPPED)
5734 if (event->attr.exclude_user && user_mode(regs))
5737 if (event->attr.exclude_kernel && !user_mode(regs))
5744 static int perf_swevent_match(struct perf_event *event,
5745 enum perf_type_id type,
5747 struct perf_sample_data *data,
5748 struct pt_regs *regs)
5750 if (event->attr.type != type)
5753 if (event->attr.config != event_id)
5756 if (perf_exclude_event(event, regs))
5762 static inline u64 swevent_hash(u64 type, u32 event_id)
5764 u64 val = event_id | (type << 32);
5766 return hash_64(val, SWEVENT_HLIST_BITS);
5769 static inline struct hlist_head *
5770 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5772 u64 hash = swevent_hash(type, event_id);
5774 return &hlist->heads[hash];
5777 /* For the read side: events when they trigger */
5778 static inline struct hlist_head *
5779 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5781 struct swevent_hlist *hlist;
5783 hlist = rcu_dereference(swhash->swevent_hlist);
5787 return __find_swevent_head(hlist, type, event_id);
5790 /* For the event head insertion and removal in the hlist */
5791 static inline struct hlist_head *
5792 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5794 struct swevent_hlist *hlist;
5795 u32 event_id = event->attr.config;
5796 u64 type = event->attr.type;
5799 * Event scheduling is always serialized against hlist allocation
5800 * and release. Which makes the protected version suitable here.
5801 * The context lock guarantees that.
5803 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5804 lockdep_is_held(&event->ctx->lock));
5808 return __find_swevent_head(hlist, type, event_id);
5811 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5813 struct perf_sample_data *data,
5814 struct pt_regs *regs)
5816 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5817 struct perf_event *event;
5818 struct hlist_head *head;
5821 head = find_swevent_head_rcu(swhash, type, event_id);
5825 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5826 if (perf_swevent_match(event, type, event_id, data, regs))
5827 perf_swevent_event(event, nr, data, regs);
5833 int perf_swevent_get_recursion_context(void)
5835 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5837 return get_recursion_context(swhash->recursion);
5839 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5841 inline void perf_swevent_put_recursion_context(int rctx)
5843 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5845 put_recursion_context(swhash->recursion, rctx);
5848 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5850 struct perf_sample_data data;
5853 preempt_disable_notrace();
5854 rctx = perf_swevent_get_recursion_context();
5858 perf_sample_data_init(&data, addr, 0);
5860 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5862 perf_swevent_put_recursion_context(rctx);
5863 preempt_enable_notrace();
5866 static void perf_swevent_read(struct perf_event *event)
5870 static int perf_swevent_add(struct perf_event *event, int flags)
5872 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5873 struct hw_perf_event *hwc = &event->hw;
5874 struct hlist_head *head;
5876 if (is_sampling_event(event)) {
5877 hwc->last_period = hwc->sample_period;
5878 perf_swevent_set_period(event);
5881 hwc->state = !(flags & PERF_EF_START);
5883 head = find_swevent_head(swhash, event);
5886 * We can race with cpu hotplug code. Do not
5887 * WARN if the cpu just got unplugged.
5889 WARN_ON_ONCE(swhash->online);
5893 hlist_add_head_rcu(&event->hlist_entry, head);
5898 static void perf_swevent_del(struct perf_event *event, int flags)
5900 hlist_del_rcu(&event->hlist_entry);
5903 static void perf_swevent_start(struct perf_event *event, int flags)
5905 event->hw.state = 0;
5908 static void perf_swevent_stop(struct perf_event *event, int flags)
5910 event->hw.state = PERF_HES_STOPPED;
5913 /* Deref the hlist from the update side */
5914 static inline struct swevent_hlist *
5915 swevent_hlist_deref(struct swevent_htable *swhash)
5917 return rcu_dereference_protected(swhash->swevent_hlist,
5918 lockdep_is_held(&swhash->hlist_mutex));
5921 static void swevent_hlist_release(struct swevent_htable *swhash)
5923 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5928 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
5929 kfree_rcu(hlist, rcu_head);
5932 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5934 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5936 mutex_lock(&swhash->hlist_mutex);
5938 if (!--swhash->hlist_refcount)
5939 swevent_hlist_release(swhash);
5941 mutex_unlock(&swhash->hlist_mutex);
5944 static void swevent_hlist_put(struct perf_event *event)
5948 for_each_possible_cpu(cpu)
5949 swevent_hlist_put_cpu(event, cpu);
5952 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5954 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5957 mutex_lock(&swhash->hlist_mutex);
5959 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5960 struct swevent_hlist *hlist;
5962 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5967 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5969 swhash->hlist_refcount++;
5971 mutex_unlock(&swhash->hlist_mutex);
5976 static int swevent_hlist_get(struct perf_event *event)
5979 int cpu, failed_cpu;
5982 for_each_possible_cpu(cpu) {
5983 err = swevent_hlist_get_cpu(event, cpu);
5993 for_each_possible_cpu(cpu) {
5994 if (cpu == failed_cpu)
5996 swevent_hlist_put_cpu(event, cpu);
6003 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6005 static void sw_perf_event_destroy(struct perf_event *event)
6007 u64 event_id = event->attr.config;
6009 WARN_ON(event->parent);
6011 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6012 swevent_hlist_put(event);
6015 static int perf_swevent_init(struct perf_event *event)
6017 u64 event_id = event->attr.config;
6019 if (event->attr.type != PERF_TYPE_SOFTWARE)
6023 * no branch sampling for software events
6025 if (has_branch_stack(event))
6029 case PERF_COUNT_SW_CPU_CLOCK:
6030 case PERF_COUNT_SW_TASK_CLOCK:
6037 if (event_id >= PERF_COUNT_SW_MAX)
6040 if (!event->parent) {
6043 err = swevent_hlist_get(event);
6047 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6048 event->destroy = sw_perf_event_destroy;
6054 static int perf_swevent_event_idx(struct perf_event *event)
6059 static struct pmu perf_swevent = {
6060 .task_ctx_nr = perf_sw_context,
6062 .event_init = perf_swevent_init,
6063 .add = perf_swevent_add,
6064 .del = perf_swevent_del,
6065 .start = perf_swevent_start,
6066 .stop = perf_swevent_stop,
6067 .read = perf_swevent_read,
6069 .event_idx = perf_swevent_event_idx,
6072 #ifdef CONFIG_EVENT_TRACING
6074 static int perf_tp_filter_match(struct perf_event *event,
6075 struct perf_sample_data *data)
6077 void *record = data->raw->data;
6079 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6084 static int perf_tp_event_match(struct perf_event *event,
6085 struct perf_sample_data *data,
6086 struct pt_regs *regs)
6088 if (event->hw.state & PERF_HES_STOPPED)
6091 * All tracepoints are from kernel-space.
6093 if (event->attr.exclude_kernel)
6096 if (!perf_tp_filter_match(event, data))
6102 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6103 struct pt_regs *regs, struct hlist_head *head, int rctx,
6104 struct task_struct *task)
6106 struct perf_sample_data data;
6107 struct perf_event *event;
6109 struct perf_raw_record raw = {
6114 perf_sample_data_init(&data, addr, 0);
6117 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6118 if (perf_tp_event_match(event, &data, regs))
6119 perf_swevent_event(event, count, &data, regs);
6123 * If we got specified a target task, also iterate its context and
6124 * deliver this event there too.
6126 if (task && task != current) {
6127 struct perf_event_context *ctx;
6128 struct trace_entry *entry = record;
6131 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6135 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6136 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6138 if (event->attr.config != entry->type)
6140 if (perf_tp_event_match(event, &data, regs))
6141 perf_swevent_event(event, count, &data, regs);
6147 perf_swevent_put_recursion_context(rctx);
6149 EXPORT_SYMBOL_GPL(perf_tp_event);
6151 static void tp_perf_event_destroy(struct perf_event *event)
6153 perf_trace_destroy(event);
6156 static int perf_tp_event_init(struct perf_event *event)
6160 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6164 * no branch sampling for tracepoint events
6166 if (has_branch_stack(event))
6169 err = perf_trace_init(event);
6173 event->destroy = tp_perf_event_destroy;
6178 static struct pmu perf_tracepoint = {
6179 .task_ctx_nr = perf_sw_context,
6181 .event_init = perf_tp_event_init,
6182 .add = perf_trace_add,
6183 .del = perf_trace_del,
6184 .start = perf_swevent_start,
6185 .stop = perf_swevent_stop,
6186 .read = perf_swevent_read,
6188 .event_idx = perf_swevent_event_idx,
6191 static inline void perf_tp_register(void)
6193 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6196 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6201 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6204 filter_str = strndup_user(arg, PAGE_SIZE);
6205 if (IS_ERR(filter_str))
6206 return PTR_ERR(filter_str);
6208 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6214 static void perf_event_free_filter(struct perf_event *event)
6216 ftrace_profile_free_filter(event);
6221 static inline void perf_tp_register(void)
6225 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6230 static void perf_event_free_filter(struct perf_event *event)
6234 #endif /* CONFIG_EVENT_TRACING */
6236 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6237 void perf_bp_event(struct perf_event *bp, void *data)
6239 struct perf_sample_data sample;
6240 struct pt_regs *regs = data;
6242 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6244 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6245 perf_swevent_event(bp, 1, &sample, regs);
6250 * hrtimer based swevent callback
6253 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6255 enum hrtimer_restart ret = HRTIMER_RESTART;
6256 struct perf_sample_data data;
6257 struct pt_regs *regs;
6258 struct perf_event *event;
6261 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6263 if (event->state != PERF_EVENT_STATE_ACTIVE)
6264 return HRTIMER_NORESTART;
6266 event->pmu->read(event);
6268 perf_sample_data_init(&data, 0, event->hw.last_period);
6269 regs = get_irq_regs();
6271 if (regs && !perf_exclude_event(event, regs)) {
6272 if (!(event->attr.exclude_idle && is_idle_task(current)))
6273 if (__perf_event_overflow(event, 1, &data, regs))
6274 ret = HRTIMER_NORESTART;
6277 period = max_t(u64, 10000, event->hw.sample_period);
6278 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6283 static void perf_swevent_start_hrtimer(struct perf_event *event)
6285 struct hw_perf_event *hwc = &event->hw;
6288 if (!is_sampling_event(event))
6291 period = local64_read(&hwc->period_left);
6296 local64_set(&hwc->period_left, 0);
6298 period = max_t(u64, 10000, hwc->sample_period);
6300 __hrtimer_start_range_ns(&hwc->hrtimer,
6301 ns_to_ktime(period), 0,
6302 HRTIMER_MODE_REL_PINNED, 0);
6305 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6307 struct hw_perf_event *hwc = &event->hw;
6309 if (is_sampling_event(event)) {
6310 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6311 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6313 hrtimer_cancel(&hwc->hrtimer);
6317 static void perf_swevent_init_hrtimer(struct perf_event *event)
6319 struct hw_perf_event *hwc = &event->hw;
6321 if (!is_sampling_event(event))
6324 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6325 hwc->hrtimer.function = perf_swevent_hrtimer;
6328 * Since hrtimers have a fixed rate, we can do a static freq->period
6329 * mapping and avoid the whole period adjust feedback stuff.
6331 if (event->attr.freq) {
6332 long freq = event->attr.sample_freq;
6334 event->attr.sample_period = NSEC_PER_SEC / freq;
6335 hwc->sample_period = event->attr.sample_period;
6336 local64_set(&hwc->period_left, hwc->sample_period);
6337 hwc->last_period = hwc->sample_period;
6338 event->attr.freq = 0;
6343 * Software event: cpu wall time clock
6346 static void cpu_clock_event_update(struct perf_event *event)
6351 now = local_clock();
6352 prev = local64_xchg(&event->hw.prev_count, now);
6353 local64_add(now - prev, &event->count);
6356 static void cpu_clock_event_start(struct perf_event *event, int flags)
6358 local64_set(&event->hw.prev_count, local_clock());
6359 perf_swevent_start_hrtimer(event);
6362 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6364 perf_swevent_cancel_hrtimer(event);
6365 cpu_clock_event_update(event);
6368 static int cpu_clock_event_add(struct perf_event *event, int flags)
6370 if (flags & PERF_EF_START)
6371 cpu_clock_event_start(event, flags);
6376 static void cpu_clock_event_del(struct perf_event *event, int flags)
6378 cpu_clock_event_stop(event, flags);
6381 static void cpu_clock_event_read(struct perf_event *event)
6383 cpu_clock_event_update(event);
6386 static int cpu_clock_event_init(struct perf_event *event)
6388 if (event->attr.type != PERF_TYPE_SOFTWARE)
6391 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6395 * no branch sampling for software events
6397 if (has_branch_stack(event))
6400 perf_swevent_init_hrtimer(event);
6405 static struct pmu perf_cpu_clock = {
6406 .task_ctx_nr = perf_sw_context,
6408 .event_init = cpu_clock_event_init,
6409 .add = cpu_clock_event_add,
6410 .del = cpu_clock_event_del,
6411 .start = cpu_clock_event_start,
6412 .stop = cpu_clock_event_stop,
6413 .read = cpu_clock_event_read,
6415 .event_idx = perf_swevent_event_idx,
6419 * Software event: task time clock
6422 static void task_clock_event_update(struct perf_event *event, u64 now)
6427 prev = local64_xchg(&event->hw.prev_count, now);
6429 local64_add(delta, &event->count);
6432 static void task_clock_event_start(struct perf_event *event, int flags)
6434 local64_set(&event->hw.prev_count, event->ctx->time);
6435 perf_swevent_start_hrtimer(event);
6438 static void task_clock_event_stop(struct perf_event *event, int flags)
6440 perf_swevent_cancel_hrtimer(event);
6441 task_clock_event_update(event, event->ctx->time);
6444 static int task_clock_event_add(struct perf_event *event, int flags)
6446 if (flags & PERF_EF_START)
6447 task_clock_event_start(event, flags);
6452 static void task_clock_event_del(struct perf_event *event, int flags)
6454 task_clock_event_stop(event, PERF_EF_UPDATE);
6457 static void task_clock_event_read(struct perf_event *event)
6459 u64 now = perf_clock();
6460 u64 delta = now - event->ctx->timestamp;
6461 u64 time = event->ctx->time + delta;
6463 task_clock_event_update(event, time);
6466 static int task_clock_event_init(struct perf_event *event)
6468 if (event->attr.type != PERF_TYPE_SOFTWARE)
6471 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6475 * no branch sampling for software events
6477 if (has_branch_stack(event))
6480 perf_swevent_init_hrtimer(event);
6485 static struct pmu perf_task_clock = {
6486 .task_ctx_nr = perf_sw_context,
6488 .event_init = task_clock_event_init,
6489 .add = task_clock_event_add,
6490 .del = task_clock_event_del,
6491 .start = task_clock_event_start,
6492 .stop = task_clock_event_stop,
6493 .read = task_clock_event_read,
6495 .event_idx = perf_swevent_event_idx,
6498 static void perf_pmu_nop_void(struct pmu *pmu)
6502 static int perf_pmu_nop_int(struct pmu *pmu)
6507 static void perf_pmu_start_txn(struct pmu *pmu)
6509 perf_pmu_disable(pmu);
6512 static int perf_pmu_commit_txn(struct pmu *pmu)
6514 perf_pmu_enable(pmu);
6518 static void perf_pmu_cancel_txn(struct pmu *pmu)
6520 perf_pmu_enable(pmu);
6523 static int perf_event_idx_default(struct perf_event *event)
6525 return event->hw.idx + 1;
6529 * Ensures all contexts with the same task_ctx_nr have the same
6530 * pmu_cpu_context too.
6532 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6539 list_for_each_entry(pmu, &pmus, entry) {
6540 if (pmu->task_ctx_nr == ctxn)
6541 return pmu->pmu_cpu_context;
6547 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6551 for_each_possible_cpu(cpu) {
6552 struct perf_cpu_context *cpuctx;
6554 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6556 if (cpuctx->unique_pmu == old_pmu)
6557 cpuctx->unique_pmu = pmu;
6561 static void free_pmu_context(struct pmu *pmu)
6565 mutex_lock(&pmus_lock);
6567 * Like a real lame refcount.
6569 list_for_each_entry(i, &pmus, entry) {
6570 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6571 update_pmu_context(i, pmu);
6576 free_percpu(pmu->pmu_cpu_context);
6578 mutex_unlock(&pmus_lock);
6580 static struct idr pmu_idr;
6583 type_show(struct device *dev, struct device_attribute *attr, char *page)
6585 struct pmu *pmu = dev_get_drvdata(dev);
6587 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6589 static DEVICE_ATTR_RO(type);
6592 perf_event_mux_interval_ms_show(struct device *dev,
6593 struct device_attribute *attr,
6596 struct pmu *pmu = dev_get_drvdata(dev);
6598 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6602 perf_event_mux_interval_ms_store(struct device *dev,
6603 struct device_attribute *attr,
6604 const char *buf, size_t count)
6606 struct pmu *pmu = dev_get_drvdata(dev);
6607 int timer, cpu, ret;
6609 ret = kstrtoint(buf, 0, &timer);
6616 /* same value, noting to do */
6617 if (timer == pmu->hrtimer_interval_ms)
6620 pmu->hrtimer_interval_ms = timer;
6622 /* update all cpuctx for this PMU */
6623 for_each_possible_cpu(cpu) {
6624 struct perf_cpu_context *cpuctx;
6625 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6626 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6628 if (hrtimer_active(&cpuctx->hrtimer))
6629 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6634 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6636 static struct attribute *pmu_dev_attrs[] = {
6637 &dev_attr_type.attr,
6638 &dev_attr_perf_event_mux_interval_ms.attr,
6641 ATTRIBUTE_GROUPS(pmu_dev);
6643 static int pmu_bus_running;
6644 static struct bus_type pmu_bus = {
6645 .name = "event_source",
6646 .dev_groups = pmu_dev_groups,
6649 static void pmu_dev_release(struct device *dev)
6654 static int pmu_dev_alloc(struct pmu *pmu)
6658 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6662 pmu->dev->groups = pmu->attr_groups;
6663 device_initialize(pmu->dev);
6664 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6668 dev_set_drvdata(pmu->dev, pmu);
6669 pmu->dev->bus = &pmu_bus;
6670 pmu->dev->release = pmu_dev_release;
6671 ret = device_add(pmu->dev);
6679 put_device(pmu->dev);
6683 static struct lock_class_key cpuctx_mutex;
6684 static struct lock_class_key cpuctx_lock;
6686 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6690 mutex_lock(&pmus_lock);
6692 pmu->pmu_disable_count = alloc_percpu(int);
6693 if (!pmu->pmu_disable_count)
6702 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6710 if (pmu_bus_running) {
6711 ret = pmu_dev_alloc(pmu);
6717 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6718 if (pmu->pmu_cpu_context)
6719 goto got_cpu_context;
6722 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6723 if (!pmu->pmu_cpu_context)
6726 for_each_possible_cpu(cpu) {
6727 struct perf_cpu_context *cpuctx;
6729 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6730 __perf_event_init_context(&cpuctx->ctx);
6731 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6732 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6733 cpuctx->ctx.type = cpu_context;
6734 cpuctx->ctx.pmu = pmu;
6736 __perf_cpu_hrtimer_init(cpuctx, cpu);
6738 INIT_LIST_HEAD(&cpuctx->rotation_list);
6739 cpuctx->unique_pmu = pmu;
6743 if (!pmu->start_txn) {
6744 if (pmu->pmu_enable) {
6746 * If we have pmu_enable/pmu_disable calls, install
6747 * transaction stubs that use that to try and batch
6748 * hardware accesses.
6750 pmu->start_txn = perf_pmu_start_txn;
6751 pmu->commit_txn = perf_pmu_commit_txn;
6752 pmu->cancel_txn = perf_pmu_cancel_txn;
6754 pmu->start_txn = perf_pmu_nop_void;
6755 pmu->commit_txn = perf_pmu_nop_int;
6756 pmu->cancel_txn = perf_pmu_nop_void;
6760 if (!pmu->pmu_enable) {
6761 pmu->pmu_enable = perf_pmu_nop_void;
6762 pmu->pmu_disable = perf_pmu_nop_void;
6765 if (!pmu->event_idx)
6766 pmu->event_idx = perf_event_idx_default;
6768 list_add_rcu(&pmu->entry, &pmus);
6771 mutex_unlock(&pmus_lock);
6776 device_del(pmu->dev);
6777 put_device(pmu->dev);
6780 if (pmu->type >= PERF_TYPE_MAX)
6781 idr_remove(&pmu_idr, pmu->type);
6784 free_percpu(pmu->pmu_disable_count);
6787 EXPORT_SYMBOL_GPL(perf_pmu_register);
6789 void perf_pmu_unregister(struct pmu *pmu)
6791 mutex_lock(&pmus_lock);
6792 list_del_rcu(&pmu->entry);
6793 mutex_unlock(&pmus_lock);
6796 * We dereference the pmu list under both SRCU and regular RCU, so
6797 * synchronize against both of those.
6799 synchronize_srcu(&pmus_srcu);
6802 free_percpu(pmu->pmu_disable_count);
6803 if (pmu->type >= PERF_TYPE_MAX)
6804 idr_remove(&pmu_idr, pmu->type);
6805 device_del(pmu->dev);
6806 put_device(pmu->dev);
6807 free_pmu_context(pmu);
6809 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6811 struct pmu *perf_init_event(struct perf_event *event)
6813 struct pmu *pmu = NULL;
6817 idx = srcu_read_lock(&pmus_srcu);
6820 pmu = idr_find(&pmu_idr, event->attr.type);
6823 if (!try_module_get(pmu->module)) {
6824 pmu = ERR_PTR(-ENODEV);
6828 ret = pmu->event_init(event);
6834 list_for_each_entry_rcu(pmu, &pmus, entry) {
6835 if (!try_module_get(pmu->module)) {
6836 pmu = ERR_PTR(-ENODEV);
6840 ret = pmu->event_init(event);
6844 if (ret != -ENOENT) {
6849 pmu = ERR_PTR(-ENOENT);
6851 srcu_read_unlock(&pmus_srcu, idx);
6856 static void account_event_cpu(struct perf_event *event, int cpu)
6861 if (has_branch_stack(event)) {
6862 if (!(event->attach_state & PERF_ATTACH_TASK))
6863 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6865 if (is_cgroup_event(event))
6866 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6869 static void account_event(struct perf_event *event)
6874 if (event->attach_state & PERF_ATTACH_TASK)
6875 static_key_slow_inc(&perf_sched_events.key);
6876 if (event->attr.mmap || event->attr.mmap_data)
6877 atomic_inc(&nr_mmap_events);
6878 if (event->attr.comm)
6879 atomic_inc(&nr_comm_events);
6880 if (event->attr.task)
6881 atomic_inc(&nr_task_events);
6882 if (event->attr.freq) {
6883 if (atomic_inc_return(&nr_freq_events) == 1)
6884 tick_nohz_full_kick_all();
6886 if (has_branch_stack(event))
6887 static_key_slow_inc(&perf_sched_events.key);
6888 if (is_cgroup_event(event))
6889 static_key_slow_inc(&perf_sched_events.key);
6891 account_event_cpu(event, event->cpu);
6895 * Allocate and initialize a event structure
6897 static struct perf_event *
6898 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6899 struct task_struct *task,
6900 struct perf_event *group_leader,
6901 struct perf_event *parent_event,
6902 perf_overflow_handler_t overflow_handler,
6906 struct perf_event *event;
6907 struct hw_perf_event *hwc;
6910 if ((unsigned)cpu >= nr_cpu_ids) {
6911 if (!task || cpu != -1)
6912 return ERR_PTR(-EINVAL);
6915 event = kzalloc(sizeof(*event), GFP_KERNEL);
6917 return ERR_PTR(-ENOMEM);
6920 * Single events are their own group leaders, with an
6921 * empty sibling list:
6924 group_leader = event;
6926 mutex_init(&event->child_mutex);
6927 INIT_LIST_HEAD(&event->child_list);
6929 INIT_LIST_HEAD(&event->group_entry);
6930 INIT_LIST_HEAD(&event->event_entry);
6931 INIT_LIST_HEAD(&event->sibling_list);
6932 INIT_LIST_HEAD(&event->rb_entry);
6933 INIT_LIST_HEAD(&event->active_entry);
6934 INIT_HLIST_NODE(&event->hlist_entry);
6937 init_waitqueue_head(&event->waitq);
6938 init_irq_work(&event->pending, perf_pending_event);
6940 mutex_init(&event->mmap_mutex);
6942 atomic_long_set(&event->refcount, 1);
6944 event->attr = *attr;
6945 event->group_leader = group_leader;
6949 event->parent = parent_event;
6951 event->ns = get_pid_ns(task_active_pid_ns(current));
6952 event->id = atomic64_inc_return(&perf_event_id);
6954 event->state = PERF_EVENT_STATE_INACTIVE;
6957 event->attach_state = PERF_ATTACH_TASK;
6959 if (attr->type == PERF_TYPE_TRACEPOINT)
6960 event->hw.tp_target = task;
6961 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6963 * hw_breakpoint is a bit difficult here..
6965 else if (attr->type == PERF_TYPE_BREAKPOINT)
6966 event->hw.bp_target = task;
6970 if (!overflow_handler && parent_event) {
6971 overflow_handler = parent_event->overflow_handler;
6972 context = parent_event->overflow_handler_context;
6975 event->overflow_handler = overflow_handler;
6976 event->overflow_handler_context = context;
6978 perf_event__state_init(event);
6983 hwc->sample_period = attr->sample_period;
6984 if (attr->freq && attr->sample_freq)
6985 hwc->sample_period = 1;
6986 hwc->last_period = hwc->sample_period;
6988 local64_set(&hwc->period_left, hwc->sample_period);
6991 * we currently do not support PERF_FORMAT_GROUP on inherited events
6993 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6996 pmu = perf_init_event(event);
6999 else if (IS_ERR(pmu)) {
7004 if (!event->parent) {
7005 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7006 err = get_callchain_buffers();
7016 event->destroy(event);
7017 module_put(pmu->module);
7020 put_pid_ns(event->ns);
7023 return ERR_PTR(err);
7026 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7027 struct perf_event_attr *attr)
7032 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7036 * zero the full structure, so that a short copy will be nice.
7038 memset(attr, 0, sizeof(*attr));
7040 ret = get_user(size, &uattr->size);
7044 if (size > PAGE_SIZE) /* silly large */
7047 if (!size) /* abi compat */
7048 size = PERF_ATTR_SIZE_VER0;
7050 if (size < PERF_ATTR_SIZE_VER0)
7054 * If we're handed a bigger struct than we know of,
7055 * ensure all the unknown bits are 0 - i.e. new
7056 * user-space does not rely on any kernel feature
7057 * extensions we dont know about yet.
7059 if (size > sizeof(*attr)) {
7060 unsigned char __user *addr;
7061 unsigned char __user *end;
7064 addr = (void __user *)uattr + sizeof(*attr);
7065 end = (void __user *)uattr + size;
7067 for (; addr < end; addr++) {
7068 ret = get_user(val, addr);
7074 size = sizeof(*attr);
7077 ret = copy_from_user(attr, uattr, size);
7081 if (attr->__reserved_1)
7084 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7087 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7090 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7091 u64 mask = attr->branch_sample_type;
7093 /* only using defined bits */
7094 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7097 /* at least one branch bit must be set */
7098 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7101 /* propagate priv level, when not set for branch */
7102 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7104 /* exclude_kernel checked on syscall entry */
7105 if (!attr->exclude_kernel)
7106 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7108 if (!attr->exclude_user)
7109 mask |= PERF_SAMPLE_BRANCH_USER;
7111 if (!attr->exclude_hv)
7112 mask |= PERF_SAMPLE_BRANCH_HV;
7114 * adjust user setting (for HW filter setup)
7116 attr->branch_sample_type = mask;
7118 /* privileged levels capture (kernel, hv): check permissions */
7119 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7120 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7124 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7125 ret = perf_reg_validate(attr->sample_regs_user);
7130 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7131 if (!arch_perf_have_user_stack_dump())
7135 * We have __u32 type for the size, but so far
7136 * we can only use __u16 as maximum due to the
7137 * __u16 sample size limit.
7139 if (attr->sample_stack_user >= USHRT_MAX)
7141 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7149 put_user(sizeof(*attr), &uattr->size);
7155 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7157 struct ring_buffer *rb = NULL;
7163 /* don't allow circular references */
7164 if (event == output_event)
7168 * Don't allow cross-cpu buffers
7170 if (output_event->cpu != event->cpu)
7174 * If its not a per-cpu rb, it must be the same task.
7176 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7180 mutex_lock(&event->mmap_mutex);
7181 /* Can't redirect output if we've got an active mmap() */
7182 if (atomic_read(&event->mmap_count))
7186 /* get the rb we want to redirect to */
7187 rb = ring_buffer_get(output_event);
7192 ring_buffer_attach(event, rb);
7196 mutex_unlock(&event->mmap_mutex);
7203 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7205 * @attr_uptr: event_id type attributes for monitoring/sampling
7208 * @group_fd: group leader event fd
7210 SYSCALL_DEFINE5(perf_event_open,
7211 struct perf_event_attr __user *, attr_uptr,
7212 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7214 struct perf_event *group_leader = NULL, *output_event = NULL;
7215 struct perf_event *event, *sibling;
7216 struct perf_event_attr attr;
7217 struct perf_event_context *ctx;
7218 struct file *event_file = NULL;
7219 struct fd group = {NULL, 0};
7220 struct task_struct *task = NULL;
7225 int f_flags = O_RDWR;
7227 /* for future expandability... */
7228 if (flags & ~PERF_FLAG_ALL)
7231 err = perf_copy_attr(attr_uptr, &attr);
7235 if (!attr.exclude_kernel) {
7236 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7241 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7244 if (attr.sample_period & (1ULL << 63))
7249 * In cgroup mode, the pid argument is used to pass the fd
7250 * opened to the cgroup directory in cgroupfs. The cpu argument
7251 * designates the cpu on which to monitor threads from that
7254 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7257 if (flags & PERF_FLAG_FD_CLOEXEC)
7258 f_flags |= O_CLOEXEC;
7260 event_fd = get_unused_fd_flags(f_flags);
7264 if (group_fd != -1) {
7265 err = perf_fget_light(group_fd, &group);
7268 group_leader = group.file->private_data;
7269 if (flags & PERF_FLAG_FD_OUTPUT)
7270 output_event = group_leader;
7271 if (flags & PERF_FLAG_FD_NO_GROUP)
7272 group_leader = NULL;
7275 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7276 task = find_lively_task_by_vpid(pid);
7278 err = PTR_ERR(task);
7283 if (task && group_leader &&
7284 group_leader->attr.inherit != attr.inherit) {
7291 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7293 if (IS_ERR(event)) {
7294 err = PTR_ERR(event);
7298 if (flags & PERF_FLAG_PID_CGROUP) {
7299 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7301 __free_event(event);
7306 if (is_sampling_event(event)) {
7307 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7313 account_event(event);
7316 * Special case software events and allow them to be part of
7317 * any hardware group.
7322 (is_software_event(event) != is_software_event(group_leader))) {
7323 if (is_software_event(event)) {
7325 * If event and group_leader are not both a software
7326 * event, and event is, then group leader is not.
7328 * Allow the addition of software events to !software
7329 * groups, this is safe because software events never
7332 pmu = group_leader->pmu;
7333 } else if (is_software_event(group_leader) &&
7334 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7336 * In case the group is a pure software group, and we
7337 * try to add a hardware event, move the whole group to
7338 * the hardware context.
7345 * Get the target context (task or percpu):
7347 ctx = find_get_context(pmu, task, event->cpu);
7354 put_task_struct(task);
7359 * Look up the group leader (we will attach this event to it):
7365 * Do not allow a recursive hierarchy (this new sibling
7366 * becoming part of another group-sibling):
7368 if (group_leader->group_leader != group_leader)
7371 * Do not allow to attach to a group in a different
7372 * task or CPU context:
7375 if (group_leader->ctx->type != ctx->type)
7378 if (group_leader->ctx != ctx)
7383 * Only a group leader can be exclusive or pinned
7385 if (attr.exclusive || attr.pinned)
7390 err = perf_event_set_output(event, output_event);
7395 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7397 if (IS_ERR(event_file)) {
7398 err = PTR_ERR(event_file);
7403 struct perf_event_context *gctx = group_leader->ctx;
7405 mutex_lock(&gctx->mutex);
7406 perf_remove_from_context(group_leader, false);
7409 * Removing from the context ends up with disabled
7410 * event. What we want here is event in the initial
7411 * startup state, ready to be add into new context.
7413 perf_event__state_init(group_leader);
7414 list_for_each_entry(sibling, &group_leader->sibling_list,
7416 perf_remove_from_context(sibling, false);
7417 perf_event__state_init(sibling);
7420 mutex_unlock(&gctx->mutex);
7424 WARN_ON_ONCE(ctx->parent_ctx);
7425 mutex_lock(&ctx->mutex);
7429 perf_install_in_context(ctx, group_leader, event->cpu);
7431 list_for_each_entry(sibling, &group_leader->sibling_list,
7433 perf_install_in_context(ctx, sibling, event->cpu);
7438 perf_install_in_context(ctx, event, event->cpu);
7439 perf_unpin_context(ctx);
7440 mutex_unlock(&ctx->mutex);
7444 event->owner = current;
7446 mutex_lock(¤t->perf_event_mutex);
7447 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7448 mutex_unlock(¤t->perf_event_mutex);
7451 * Precalculate sample_data sizes
7453 perf_event__header_size(event);
7454 perf_event__id_header_size(event);
7457 * Drop the reference on the group_event after placing the
7458 * new event on the sibling_list. This ensures destruction
7459 * of the group leader will find the pointer to itself in
7460 * perf_group_detach().
7463 fd_install(event_fd, event_file);
7467 perf_unpin_context(ctx);
7475 put_task_struct(task);
7479 put_unused_fd(event_fd);
7484 * perf_event_create_kernel_counter
7486 * @attr: attributes of the counter to create
7487 * @cpu: cpu in which the counter is bound
7488 * @task: task to profile (NULL for percpu)
7491 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7492 struct task_struct *task,
7493 perf_overflow_handler_t overflow_handler,
7496 struct perf_event_context *ctx;
7497 struct perf_event *event;
7501 * Get the target context (task or percpu):
7504 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7505 overflow_handler, context);
7506 if (IS_ERR(event)) {
7507 err = PTR_ERR(event);
7511 /* Mark owner so we could distinguish it from user events. */
7512 event->owner = EVENT_OWNER_KERNEL;
7514 account_event(event);
7516 ctx = find_get_context(event->pmu, task, cpu);
7522 WARN_ON_ONCE(ctx->parent_ctx);
7523 mutex_lock(&ctx->mutex);
7524 perf_install_in_context(ctx, event, cpu);
7525 perf_unpin_context(ctx);
7526 mutex_unlock(&ctx->mutex);
7533 return ERR_PTR(err);
7535 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7537 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7539 struct perf_event_context *src_ctx;
7540 struct perf_event_context *dst_ctx;
7541 struct perf_event *event, *tmp;
7544 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7545 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7547 mutex_lock(&src_ctx->mutex);
7548 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7550 perf_remove_from_context(event, false);
7551 unaccount_event_cpu(event, src_cpu);
7553 list_add(&event->migrate_entry, &events);
7555 mutex_unlock(&src_ctx->mutex);
7559 mutex_lock(&dst_ctx->mutex);
7560 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7561 list_del(&event->migrate_entry);
7562 if (event->state >= PERF_EVENT_STATE_OFF)
7563 event->state = PERF_EVENT_STATE_INACTIVE;
7564 account_event_cpu(event, dst_cpu);
7565 perf_install_in_context(dst_ctx, event, dst_cpu);
7568 mutex_unlock(&dst_ctx->mutex);
7570 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7572 static void sync_child_event(struct perf_event *child_event,
7573 struct task_struct *child)
7575 struct perf_event *parent_event = child_event->parent;
7578 if (child_event->attr.inherit_stat)
7579 perf_event_read_event(child_event, child);
7581 child_val = perf_event_count(child_event);
7584 * Add back the child's count to the parent's count:
7586 atomic64_add(child_val, &parent_event->child_count);
7587 atomic64_add(child_event->total_time_enabled,
7588 &parent_event->child_total_time_enabled);
7589 atomic64_add(child_event->total_time_running,
7590 &parent_event->child_total_time_running);
7593 * Remove this event from the parent's list
7595 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7596 mutex_lock(&parent_event->child_mutex);
7597 list_del_init(&child_event->child_list);
7598 mutex_unlock(&parent_event->child_mutex);
7601 * Make sure user/parent get notified, that we just
7604 perf_event_wakeup(parent_event);
7607 * Release the parent event, if this was the last
7610 put_event(parent_event);
7614 __perf_event_exit_task(struct perf_event *child_event,
7615 struct perf_event_context *child_ctx,
7616 struct task_struct *child)
7619 * Do not destroy the 'original' grouping; because of the context
7620 * switch optimization the original events could've ended up in a
7621 * random child task.
7623 * If we were to destroy the original group, all group related
7624 * operations would cease to function properly after this random
7627 * Do destroy all inherited groups, we don't care about those
7628 * and being thorough is better.
7630 perf_remove_from_context(child_event, !!child_event->parent);
7633 * It can happen that the parent exits first, and has events
7634 * that are still around due to the child reference. These
7635 * events need to be zapped.
7637 if (child_event->parent) {
7638 sync_child_event(child_event, child);
7639 free_event(child_event);
7641 child_event->state = PERF_EVENT_STATE_EXIT;
7642 perf_event_wakeup(child_event);
7646 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7648 struct perf_event *child_event, *next;
7649 struct perf_event_context *child_ctx, *parent_ctx;
7650 unsigned long flags;
7652 if (likely(!child->perf_event_ctxp[ctxn])) {
7653 perf_event_task(child, NULL, 0);
7657 local_irq_save(flags);
7659 * We can't reschedule here because interrupts are disabled,
7660 * and either child is current or it is a task that can't be
7661 * scheduled, so we are now safe from rescheduling changing
7664 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7667 * Take the context lock here so that if find_get_context is
7668 * reading child->perf_event_ctxp, we wait until it has
7669 * incremented the context's refcount before we do put_ctx below.
7671 raw_spin_lock(&child_ctx->lock);
7672 task_ctx_sched_out(child_ctx);
7673 child->perf_event_ctxp[ctxn] = NULL;
7676 * In order to avoid freeing: child_ctx->parent_ctx->task
7677 * under perf_event_context::lock, grab another reference.
7679 parent_ctx = child_ctx->parent_ctx;
7681 get_ctx(parent_ctx);
7684 * If this context is a clone; unclone it so it can't get
7685 * swapped to another process while we're removing all
7686 * the events from it.
7688 unclone_ctx(child_ctx);
7689 update_context_time(child_ctx);
7690 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7693 * Now that we no longer hold perf_event_context::lock, drop
7694 * our extra child_ctx->parent_ctx reference.
7697 put_ctx(parent_ctx);
7700 * Report the task dead after unscheduling the events so that we
7701 * won't get any samples after PERF_RECORD_EXIT. We can however still
7702 * get a few PERF_RECORD_READ events.
7704 perf_event_task(child, child_ctx, 0);
7707 * We can recurse on the same lock type through:
7709 * __perf_event_exit_task()
7710 * sync_child_event()
7712 * mutex_lock(&ctx->mutex)
7714 * But since its the parent context it won't be the same instance.
7716 mutex_lock(&child_ctx->mutex);
7718 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7719 __perf_event_exit_task(child_event, child_ctx, child);
7721 mutex_unlock(&child_ctx->mutex);
7727 * When a child task exits, feed back event values to parent events.
7729 void perf_event_exit_task(struct task_struct *child)
7731 struct perf_event *event, *tmp;
7734 mutex_lock(&child->perf_event_mutex);
7735 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7737 list_del_init(&event->owner_entry);
7740 * Ensure the list deletion is visible before we clear
7741 * the owner, closes a race against perf_release() where
7742 * we need to serialize on the owner->perf_event_mutex.
7745 event->owner = NULL;
7747 mutex_unlock(&child->perf_event_mutex);
7749 for_each_task_context_nr(ctxn)
7750 perf_event_exit_task_context(child, ctxn);
7753 static void perf_free_event(struct perf_event *event,
7754 struct perf_event_context *ctx)
7756 struct perf_event *parent = event->parent;
7758 if (WARN_ON_ONCE(!parent))
7761 mutex_lock(&parent->child_mutex);
7762 list_del_init(&event->child_list);
7763 mutex_unlock(&parent->child_mutex);
7767 perf_group_detach(event);
7768 list_del_event(event, ctx);
7773 * free an unexposed, unused context as created by inheritance by
7774 * perf_event_init_task below, used by fork() in case of fail.
7776 void perf_event_free_task(struct task_struct *task)
7778 struct perf_event_context *ctx;
7779 struct perf_event *event, *tmp;
7782 for_each_task_context_nr(ctxn) {
7783 ctx = task->perf_event_ctxp[ctxn];
7787 mutex_lock(&ctx->mutex);
7789 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7791 perf_free_event(event, ctx);
7793 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7795 perf_free_event(event, ctx);
7797 if (!list_empty(&ctx->pinned_groups) ||
7798 !list_empty(&ctx->flexible_groups))
7801 mutex_unlock(&ctx->mutex);
7807 void perf_event_delayed_put(struct task_struct *task)
7811 for_each_task_context_nr(ctxn)
7812 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7816 * inherit a event from parent task to child task:
7818 static struct perf_event *
7819 inherit_event(struct perf_event *parent_event,
7820 struct task_struct *parent,
7821 struct perf_event_context *parent_ctx,
7822 struct task_struct *child,
7823 struct perf_event *group_leader,
7824 struct perf_event_context *child_ctx)
7826 enum perf_event_active_state parent_state = parent_event->state;
7827 struct perf_event *child_event;
7828 unsigned long flags;
7831 * Instead of creating recursive hierarchies of events,
7832 * we link inherited events back to the original parent,
7833 * which has a filp for sure, which we use as the reference
7836 if (parent_event->parent)
7837 parent_event = parent_event->parent;
7839 child_event = perf_event_alloc(&parent_event->attr,
7842 group_leader, parent_event,
7844 if (IS_ERR(child_event))
7847 if (is_orphaned_event(parent_event) ||
7848 !atomic_long_inc_not_zero(&parent_event->refcount)) {
7849 free_event(child_event);
7856 * Make the child state follow the state of the parent event,
7857 * not its attr.disabled bit. We hold the parent's mutex,
7858 * so we won't race with perf_event_{en, dis}able_family.
7860 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
7861 child_event->state = PERF_EVENT_STATE_INACTIVE;
7863 child_event->state = PERF_EVENT_STATE_OFF;
7865 if (parent_event->attr.freq) {
7866 u64 sample_period = parent_event->hw.sample_period;
7867 struct hw_perf_event *hwc = &child_event->hw;
7869 hwc->sample_period = sample_period;
7870 hwc->last_period = sample_period;
7872 local64_set(&hwc->period_left, sample_period);
7875 child_event->ctx = child_ctx;
7876 child_event->overflow_handler = parent_event->overflow_handler;
7877 child_event->overflow_handler_context
7878 = parent_event->overflow_handler_context;
7881 * Precalculate sample_data sizes
7883 perf_event__header_size(child_event);
7884 perf_event__id_header_size(child_event);
7887 * Link it up in the child's context:
7889 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7890 add_event_to_ctx(child_event, child_ctx);
7891 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7894 * Link this into the parent event's child list
7896 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7897 mutex_lock(&parent_event->child_mutex);
7898 list_add_tail(&child_event->child_list, &parent_event->child_list);
7899 mutex_unlock(&parent_event->child_mutex);
7904 static int inherit_group(struct perf_event *parent_event,
7905 struct task_struct *parent,
7906 struct perf_event_context *parent_ctx,
7907 struct task_struct *child,
7908 struct perf_event_context *child_ctx)
7910 struct perf_event *leader;
7911 struct perf_event *sub;
7912 struct perf_event *child_ctr;
7914 leader = inherit_event(parent_event, parent, parent_ctx,
7915 child, NULL, child_ctx);
7917 return PTR_ERR(leader);
7918 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7919 child_ctr = inherit_event(sub, parent, parent_ctx,
7920 child, leader, child_ctx);
7921 if (IS_ERR(child_ctr))
7922 return PTR_ERR(child_ctr);
7928 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7929 struct perf_event_context *parent_ctx,
7930 struct task_struct *child, int ctxn,
7934 struct perf_event_context *child_ctx;
7936 if (!event->attr.inherit) {
7941 child_ctx = child->perf_event_ctxp[ctxn];
7944 * This is executed from the parent task context, so
7945 * inherit events that have been marked for cloning.
7946 * First allocate and initialize a context for the
7950 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7954 child->perf_event_ctxp[ctxn] = child_ctx;
7957 ret = inherit_group(event, parent, parent_ctx,
7967 * Initialize the perf_event context in task_struct
7969 static int perf_event_init_context(struct task_struct *child, int ctxn)
7971 struct perf_event_context *child_ctx, *parent_ctx;
7972 struct perf_event_context *cloned_ctx;
7973 struct perf_event *event;
7974 struct task_struct *parent = current;
7975 int inherited_all = 1;
7976 unsigned long flags;
7979 if (likely(!parent->perf_event_ctxp[ctxn]))
7983 * If the parent's context is a clone, pin it so it won't get
7986 parent_ctx = perf_pin_task_context(parent, ctxn);
7991 * No need to check if parent_ctx != NULL here; since we saw
7992 * it non-NULL earlier, the only reason for it to become NULL
7993 * is if we exit, and since we're currently in the middle of
7994 * a fork we can't be exiting at the same time.
7998 * Lock the parent list. No need to lock the child - not PID
7999 * hashed yet and not running, so nobody can access it.
8001 mutex_lock(&parent_ctx->mutex);
8004 * We dont have to disable NMIs - we are only looking at
8005 * the list, not manipulating it:
8007 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8008 ret = inherit_task_group(event, parent, parent_ctx,
8009 child, ctxn, &inherited_all);
8015 * We can't hold ctx->lock when iterating the ->flexible_group list due
8016 * to allocations, but we need to prevent rotation because
8017 * rotate_ctx() will change the list from interrupt context.
8019 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8020 parent_ctx->rotate_disable = 1;
8021 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8023 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8024 ret = inherit_task_group(event, parent, parent_ctx,
8025 child, ctxn, &inherited_all);
8030 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8031 parent_ctx->rotate_disable = 0;
8033 child_ctx = child->perf_event_ctxp[ctxn];
8035 if (child_ctx && inherited_all) {
8037 * Mark the child context as a clone of the parent
8038 * context, or of whatever the parent is a clone of.
8040 * Note that if the parent is a clone, the holding of
8041 * parent_ctx->lock avoids it from being uncloned.
8043 cloned_ctx = parent_ctx->parent_ctx;
8045 child_ctx->parent_ctx = cloned_ctx;
8046 child_ctx->parent_gen = parent_ctx->parent_gen;
8048 child_ctx->parent_ctx = parent_ctx;
8049 child_ctx->parent_gen = parent_ctx->generation;
8051 get_ctx(child_ctx->parent_ctx);
8054 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8055 mutex_unlock(&parent_ctx->mutex);
8057 perf_unpin_context(parent_ctx);
8058 put_ctx(parent_ctx);
8064 * Initialize the perf_event context in task_struct
8066 int perf_event_init_task(struct task_struct *child)
8070 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8071 mutex_init(&child->perf_event_mutex);
8072 INIT_LIST_HEAD(&child->perf_event_list);
8074 for_each_task_context_nr(ctxn) {
8075 ret = perf_event_init_context(child, ctxn);
8077 perf_event_free_task(child);
8085 static void __init perf_event_init_all_cpus(void)
8087 struct swevent_htable *swhash;
8090 for_each_possible_cpu(cpu) {
8091 swhash = &per_cpu(swevent_htable, cpu);
8092 mutex_init(&swhash->hlist_mutex);
8093 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
8097 static void perf_event_init_cpu(int cpu)
8099 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8101 mutex_lock(&swhash->hlist_mutex);
8102 swhash->online = true;
8103 if (swhash->hlist_refcount > 0) {
8104 struct swevent_hlist *hlist;
8106 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8108 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8110 mutex_unlock(&swhash->hlist_mutex);
8113 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8114 static void perf_pmu_rotate_stop(struct pmu *pmu)
8116 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
8118 WARN_ON(!irqs_disabled());
8120 list_del_init(&cpuctx->rotation_list);
8123 static void __perf_event_exit_context(void *__info)
8125 struct remove_event re = { .detach_group = false };
8126 struct perf_event_context *ctx = __info;
8128 perf_pmu_rotate_stop(ctx->pmu);
8131 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8132 __perf_remove_from_context(&re);
8136 static void perf_event_exit_cpu_context(int cpu)
8138 struct perf_event_context *ctx;
8142 idx = srcu_read_lock(&pmus_srcu);
8143 list_for_each_entry_rcu(pmu, &pmus, entry) {
8144 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8146 mutex_lock(&ctx->mutex);
8147 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8148 mutex_unlock(&ctx->mutex);
8150 srcu_read_unlock(&pmus_srcu, idx);
8153 static void perf_event_exit_cpu(int cpu)
8155 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8157 perf_event_exit_cpu_context(cpu);
8159 mutex_lock(&swhash->hlist_mutex);
8160 swhash->online = false;
8161 swevent_hlist_release(swhash);
8162 mutex_unlock(&swhash->hlist_mutex);
8165 static inline void perf_event_exit_cpu(int cpu) { }
8169 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8173 for_each_online_cpu(cpu)
8174 perf_event_exit_cpu(cpu);
8180 * Run the perf reboot notifier at the very last possible moment so that
8181 * the generic watchdog code runs as long as possible.
8183 static struct notifier_block perf_reboot_notifier = {
8184 .notifier_call = perf_reboot,
8185 .priority = INT_MIN,
8189 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8191 unsigned int cpu = (long)hcpu;
8193 switch (action & ~CPU_TASKS_FROZEN) {
8195 case CPU_UP_PREPARE:
8196 case CPU_DOWN_FAILED:
8197 perf_event_init_cpu(cpu);
8200 case CPU_UP_CANCELED:
8201 case CPU_DOWN_PREPARE:
8202 perf_event_exit_cpu(cpu);
8211 void __init perf_event_init(void)
8217 perf_event_init_all_cpus();
8218 init_srcu_struct(&pmus_srcu);
8219 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8220 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8221 perf_pmu_register(&perf_task_clock, NULL, -1);
8223 perf_cpu_notifier(perf_cpu_notify);
8224 register_reboot_notifier(&perf_reboot_notifier);
8226 ret = init_hw_breakpoint();
8227 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8229 /* do not patch jump label more than once per second */
8230 jump_label_rate_limit(&perf_sched_events, HZ);
8233 * Build time assertion that we keep the data_head at the intended
8234 * location. IOW, validation we got the __reserved[] size right.
8236 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8240 static int __init perf_event_sysfs_init(void)
8245 mutex_lock(&pmus_lock);
8247 ret = bus_register(&pmu_bus);
8251 list_for_each_entry(pmu, &pmus, entry) {
8252 if (!pmu->name || pmu->type < 0)
8255 ret = pmu_dev_alloc(pmu);
8256 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8258 pmu_bus_running = 1;
8262 mutex_unlock(&pmus_lock);
8266 device_initcall(perf_event_sysfs_init);
8268 #ifdef CONFIG_CGROUP_PERF
8269 static struct cgroup_subsys_state *
8270 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8272 struct perf_cgroup *jc;
8274 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8276 return ERR_PTR(-ENOMEM);
8278 jc->info = alloc_percpu(struct perf_cgroup_info);
8281 return ERR_PTR(-ENOMEM);
8287 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8289 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8291 free_percpu(jc->info);
8295 static int __perf_cgroup_move(void *info)
8297 struct task_struct *task = info;
8298 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8302 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8303 struct cgroup_taskset *tset)
8305 struct task_struct *task;
8307 cgroup_taskset_for_each(task, tset)
8308 task_function_call(task, __perf_cgroup_move, task);
8311 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8312 struct cgroup_subsys_state *old_css,
8313 struct task_struct *task)
8316 * cgroup_exit() is called in the copy_process() failure path.
8317 * Ignore this case since the task hasn't ran yet, this avoids
8318 * trying to poke a half freed task state from generic code.
8320 if (!(task->flags & PF_EXITING))
8323 task_function_call(task, __perf_cgroup_move, task);
8326 struct cgroup_subsys perf_event_cgrp_subsys = {
8327 .css_alloc = perf_cgroup_css_alloc,
8328 .css_free = perf_cgroup_css_free,
8329 .exit = perf_cgroup_exit,
8330 .attach = perf_cgroup_attach,
8332 #endif /* CONFIG_CGROUP_PERF */