2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP)
125 * branch priv levels that need permission checks
127 #define PERF_SAMPLE_BRANCH_PERM_PLM \
128 (PERF_SAMPLE_BRANCH_KERNEL |\
129 PERF_SAMPLE_BRANCH_HV)
132 EVENT_FLEXIBLE = 0x1,
134 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
138 * perf_sched_events : >0 events exist
139 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 struct static_key_deferred perf_sched_events __read_mostly;
142 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
143 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
145 static atomic_t nr_mmap_events __read_mostly;
146 static atomic_t nr_comm_events __read_mostly;
147 static atomic_t nr_task_events __read_mostly;
149 static LIST_HEAD(pmus);
150 static DEFINE_MUTEX(pmus_lock);
151 static struct srcu_struct pmus_srcu;
154 * perf event paranoia level:
155 * -1 - not paranoid at all
156 * 0 - disallow raw tracepoint access for unpriv
157 * 1 - disallow cpu events for unpriv
158 * 2 - disallow kernel profiling for unpriv
160 int sysctl_perf_event_paranoid __read_mostly = 1;
162 /* Minimum for 512 kiB + 1 user control page */
163 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
166 * max perf event sample rate
168 #define DEFAULT_MAX_SAMPLE_RATE 100000
169 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
170 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
172 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
174 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
175 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
177 static atomic_t perf_sample_allowed_ns __read_mostly =
178 ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100);
180 void update_perf_cpu_limits(void)
182 u64 tmp = perf_sample_period_ns;
184 tmp *= sysctl_perf_cpu_time_max_percent;
186 atomic_set(&perf_sample_allowed_ns, tmp);
189 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
191 int perf_proc_update_handler(struct ctl_table *table, int write,
192 void __user *buffer, size_t *lenp,
195 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
200 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
201 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
202 update_perf_cpu_limits();
207 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
209 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
210 void __user *buffer, size_t *lenp,
213 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
218 update_perf_cpu_limits();
224 * perf samples are done in some very critical code paths (NMIs).
225 * If they take too much CPU time, the system can lock up and not
226 * get any real work done. This will drop the sample rate when
227 * we detect that events are taking too long.
229 #define NR_ACCUMULATED_SAMPLES 128
230 DEFINE_PER_CPU(u64, running_sample_length);
232 void perf_sample_event_took(u64 sample_len_ns)
234 u64 avg_local_sample_len;
235 u64 local_samples_len;
237 if (atomic_read(&perf_sample_allowed_ns) == 0)
240 /* decay the counter by 1 average sample */
241 local_samples_len = __get_cpu_var(running_sample_length);
242 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
243 local_samples_len += sample_len_ns;
244 __get_cpu_var(running_sample_length) = local_samples_len;
247 * note: this will be biased artifically low until we have
248 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
249 * from having to maintain a count.
251 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
253 if (avg_local_sample_len <= atomic_read(&perf_sample_allowed_ns))
256 if (max_samples_per_tick <= 1)
259 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
260 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
261 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
263 printk_ratelimited(KERN_WARNING
264 "perf samples too long (%lld > %d), lowering "
265 "kernel.perf_event_max_sample_rate to %d\n",
266 avg_local_sample_len,
267 atomic_read(&perf_sample_allowed_ns),
268 sysctl_perf_event_sample_rate);
270 update_perf_cpu_limits();
273 static atomic64_t perf_event_id;
275 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
276 enum event_type_t event_type);
278 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
279 enum event_type_t event_type,
280 struct task_struct *task);
282 static void update_context_time(struct perf_event_context *ctx);
283 static u64 perf_event_time(struct perf_event *event);
285 void __weak perf_event_print_debug(void) { }
287 extern __weak const char *perf_pmu_name(void)
292 static inline u64 perf_clock(void)
294 return local_clock();
297 static inline struct perf_cpu_context *
298 __get_cpu_context(struct perf_event_context *ctx)
300 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
303 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
304 struct perf_event_context *ctx)
306 raw_spin_lock(&cpuctx->ctx.lock);
308 raw_spin_lock(&ctx->lock);
311 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
312 struct perf_event_context *ctx)
315 raw_spin_unlock(&ctx->lock);
316 raw_spin_unlock(&cpuctx->ctx.lock);
319 #ifdef CONFIG_CGROUP_PERF
322 * perf_cgroup_info keeps track of time_enabled for a cgroup.
323 * This is a per-cpu dynamically allocated data structure.
325 struct perf_cgroup_info {
331 struct cgroup_subsys_state css;
332 struct perf_cgroup_info __percpu *info;
336 * Must ensure cgroup is pinned (css_get) before calling
337 * this function. In other words, we cannot call this function
338 * if there is no cgroup event for the current CPU context.
340 static inline struct perf_cgroup *
341 perf_cgroup_from_task(struct task_struct *task)
343 return container_of(task_subsys_state(task, perf_subsys_id),
344 struct perf_cgroup, css);
348 perf_cgroup_match(struct perf_event *event)
350 struct perf_event_context *ctx = event->ctx;
351 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
353 /* @event doesn't care about cgroup */
357 /* wants specific cgroup scope but @cpuctx isn't associated with any */
362 * Cgroup scoping is recursive. An event enabled for a cgroup is
363 * also enabled for all its descendant cgroups. If @cpuctx's
364 * cgroup is a descendant of @event's (the test covers identity
365 * case), it's a match.
367 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
368 event->cgrp->css.cgroup);
371 static inline bool perf_tryget_cgroup(struct perf_event *event)
373 return css_tryget(&event->cgrp->css);
376 static inline void perf_put_cgroup(struct perf_event *event)
378 css_put(&event->cgrp->css);
381 static inline void perf_detach_cgroup(struct perf_event *event)
383 perf_put_cgroup(event);
387 static inline int is_cgroup_event(struct perf_event *event)
389 return event->cgrp != NULL;
392 static inline u64 perf_cgroup_event_time(struct perf_event *event)
394 struct perf_cgroup_info *t;
396 t = per_cpu_ptr(event->cgrp->info, event->cpu);
400 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
402 struct perf_cgroup_info *info;
407 info = this_cpu_ptr(cgrp->info);
409 info->time += now - info->timestamp;
410 info->timestamp = now;
413 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
415 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
417 __update_cgrp_time(cgrp_out);
420 static inline void update_cgrp_time_from_event(struct perf_event *event)
422 struct perf_cgroup *cgrp;
425 * ensure we access cgroup data only when needed and
426 * when we know the cgroup is pinned (css_get)
428 if (!is_cgroup_event(event))
431 cgrp = perf_cgroup_from_task(current);
433 * Do not update time when cgroup is not active
435 if (cgrp == event->cgrp)
436 __update_cgrp_time(event->cgrp);
440 perf_cgroup_set_timestamp(struct task_struct *task,
441 struct perf_event_context *ctx)
443 struct perf_cgroup *cgrp;
444 struct perf_cgroup_info *info;
447 * ctx->lock held by caller
448 * ensure we do not access cgroup data
449 * unless we have the cgroup pinned (css_get)
451 if (!task || !ctx->nr_cgroups)
454 cgrp = perf_cgroup_from_task(task);
455 info = this_cpu_ptr(cgrp->info);
456 info->timestamp = ctx->timestamp;
459 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
460 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
463 * reschedule events based on the cgroup constraint of task.
465 * mode SWOUT : schedule out everything
466 * mode SWIN : schedule in based on cgroup for next
468 void perf_cgroup_switch(struct task_struct *task, int mode)
470 struct perf_cpu_context *cpuctx;
475 * disable interrupts to avoid geting nr_cgroup
476 * changes via __perf_event_disable(). Also
479 local_irq_save(flags);
482 * we reschedule only in the presence of cgroup
483 * constrained events.
487 list_for_each_entry_rcu(pmu, &pmus, entry) {
488 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
489 if (cpuctx->unique_pmu != pmu)
490 continue; /* ensure we process each cpuctx once */
493 * perf_cgroup_events says at least one
494 * context on this CPU has cgroup events.
496 * ctx->nr_cgroups reports the number of cgroup
497 * events for a context.
499 if (cpuctx->ctx.nr_cgroups > 0) {
500 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
501 perf_pmu_disable(cpuctx->ctx.pmu);
503 if (mode & PERF_CGROUP_SWOUT) {
504 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
506 * must not be done before ctxswout due
507 * to event_filter_match() in event_sched_out()
512 if (mode & PERF_CGROUP_SWIN) {
513 WARN_ON_ONCE(cpuctx->cgrp);
515 * set cgrp before ctxsw in to allow
516 * event_filter_match() to not have to pass
519 cpuctx->cgrp = perf_cgroup_from_task(task);
520 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
522 perf_pmu_enable(cpuctx->ctx.pmu);
523 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
529 local_irq_restore(flags);
532 static inline void perf_cgroup_sched_out(struct task_struct *task,
533 struct task_struct *next)
535 struct perf_cgroup *cgrp1;
536 struct perf_cgroup *cgrp2 = NULL;
539 * we come here when we know perf_cgroup_events > 0
541 cgrp1 = perf_cgroup_from_task(task);
544 * next is NULL when called from perf_event_enable_on_exec()
545 * that will systematically cause a cgroup_switch()
548 cgrp2 = perf_cgroup_from_task(next);
551 * only schedule out current cgroup events if we know
552 * that we are switching to a different cgroup. Otherwise,
553 * do no touch the cgroup events.
556 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
559 static inline void perf_cgroup_sched_in(struct task_struct *prev,
560 struct task_struct *task)
562 struct perf_cgroup *cgrp1;
563 struct perf_cgroup *cgrp2 = NULL;
566 * we come here when we know perf_cgroup_events > 0
568 cgrp1 = perf_cgroup_from_task(task);
570 /* prev can never be NULL */
571 cgrp2 = perf_cgroup_from_task(prev);
574 * only need to schedule in cgroup events if we are changing
575 * cgroup during ctxsw. Cgroup events were not scheduled
576 * out of ctxsw out if that was not the case.
579 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
582 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
583 struct perf_event_attr *attr,
584 struct perf_event *group_leader)
586 struct perf_cgroup *cgrp;
587 struct cgroup_subsys_state *css;
588 struct fd f = fdget(fd);
594 css = cgroup_css_from_dir(f.file, perf_subsys_id);
600 cgrp = container_of(css, struct perf_cgroup, css);
603 /* must be done before we fput() the file */
604 if (!perf_tryget_cgroup(event)) {
611 * all events in a group must monitor
612 * the same cgroup because a task belongs
613 * to only one perf cgroup at a time
615 if (group_leader && group_leader->cgrp != cgrp) {
616 perf_detach_cgroup(event);
625 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
627 struct perf_cgroup_info *t;
628 t = per_cpu_ptr(event->cgrp->info, event->cpu);
629 event->shadow_ctx_time = now - t->timestamp;
633 perf_cgroup_defer_enabled(struct perf_event *event)
636 * when the current task's perf cgroup does not match
637 * the event's, we need to remember to call the
638 * perf_mark_enable() function the first time a task with
639 * a matching perf cgroup is scheduled in.
641 if (is_cgroup_event(event) && !perf_cgroup_match(event))
642 event->cgrp_defer_enabled = 1;
646 perf_cgroup_mark_enabled(struct perf_event *event,
647 struct perf_event_context *ctx)
649 struct perf_event *sub;
650 u64 tstamp = perf_event_time(event);
652 if (!event->cgrp_defer_enabled)
655 event->cgrp_defer_enabled = 0;
657 event->tstamp_enabled = tstamp - event->total_time_enabled;
658 list_for_each_entry(sub, &event->sibling_list, group_entry) {
659 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
660 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
661 sub->cgrp_defer_enabled = 0;
665 #else /* !CONFIG_CGROUP_PERF */
668 perf_cgroup_match(struct perf_event *event)
673 static inline void perf_detach_cgroup(struct perf_event *event)
676 static inline int is_cgroup_event(struct perf_event *event)
681 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
686 static inline void update_cgrp_time_from_event(struct perf_event *event)
690 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
694 static inline void perf_cgroup_sched_out(struct task_struct *task,
695 struct task_struct *next)
699 static inline void perf_cgroup_sched_in(struct task_struct *prev,
700 struct task_struct *task)
704 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
705 struct perf_event_attr *attr,
706 struct perf_event *group_leader)
712 perf_cgroup_set_timestamp(struct task_struct *task,
713 struct perf_event_context *ctx)
718 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
723 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
727 static inline u64 perf_cgroup_event_time(struct perf_event *event)
733 perf_cgroup_defer_enabled(struct perf_event *event)
738 perf_cgroup_mark_enabled(struct perf_event *event,
739 struct perf_event_context *ctx)
745 * set default to be dependent on timer tick just
748 #define PERF_CPU_HRTIMER (1000 / HZ)
750 * function must be called with interrupts disbled
752 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
754 struct perf_cpu_context *cpuctx;
755 enum hrtimer_restart ret = HRTIMER_NORESTART;
758 WARN_ON(!irqs_disabled());
760 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
762 rotations = perf_rotate_context(cpuctx);
765 * arm timer if needed
768 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
769 ret = HRTIMER_RESTART;
775 /* CPU is going down */
776 void perf_cpu_hrtimer_cancel(int cpu)
778 struct perf_cpu_context *cpuctx;
782 if (WARN_ON(cpu != smp_processor_id()))
785 local_irq_save(flags);
789 list_for_each_entry_rcu(pmu, &pmus, entry) {
790 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
792 if (pmu->task_ctx_nr == perf_sw_context)
795 hrtimer_cancel(&cpuctx->hrtimer);
800 local_irq_restore(flags);
803 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
805 struct hrtimer *hr = &cpuctx->hrtimer;
806 struct pmu *pmu = cpuctx->ctx.pmu;
809 /* no multiplexing needed for SW PMU */
810 if (pmu->task_ctx_nr == perf_sw_context)
814 * check default is sane, if not set then force to
815 * default interval (1/tick)
817 timer = pmu->hrtimer_interval_ms;
819 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
821 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
823 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
824 hr->function = perf_cpu_hrtimer_handler;
827 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
829 struct hrtimer *hr = &cpuctx->hrtimer;
830 struct pmu *pmu = cpuctx->ctx.pmu;
833 if (pmu->task_ctx_nr == perf_sw_context)
836 if (hrtimer_active(hr))
839 if (!hrtimer_callback_running(hr))
840 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
841 0, HRTIMER_MODE_REL_PINNED, 0);
844 void perf_pmu_disable(struct pmu *pmu)
846 int *count = this_cpu_ptr(pmu->pmu_disable_count);
848 pmu->pmu_disable(pmu);
851 void perf_pmu_enable(struct pmu *pmu)
853 int *count = this_cpu_ptr(pmu->pmu_disable_count);
855 pmu->pmu_enable(pmu);
858 static DEFINE_PER_CPU(struct list_head, rotation_list);
861 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
862 * because they're strictly cpu affine and rotate_start is called with IRQs
863 * disabled, while rotate_context is called from IRQ context.
865 static void perf_pmu_rotate_start(struct pmu *pmu)
867 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
868 struct list_head *head = &__get_cpu_var(rotation_list);
870 WARN_ON(!irqs_disabled());
872 if (list_empty(&cpuctx->rotation_list)) {
873 int was_empty = list_empty(head);
874 list_add(&cpuctx->rotation_list, head);
876 tick_nohz_full_kick();
880 static void get_ctx(struct perf_event_context *ctx)
882 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
885 static void put_ctx(struct perf_event_context *ctx)
887 if (atomic_dec_and_test(&ctx->refcount)) {
889 put_ctx(ctx->parent_ctx);
891 put_task_struct(ctx->task);
892 kfree_rcu(ctx, rcu_head);
896 static void unclone_ctx(struct perf_event_context *ctx)
898 if (ctx->parent_ctx) {
899 put_ctx(ctx->parent_ctx);
900 ctx->parent_ctx = NULL;
904 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
907 * only top level events have the pid namespace they were created in
910 event = event->parent;
912 return task_tgid_nr_ns(p, event->ns);
915 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
918 * only top level events have the pid namespace they were created in
921 event = event->parent;
923 return task_pid_nr_ns(p, event->ns);
927 * If we inherit events we want to return the parent event id
930 static u64 primary_event_id(struct perf_event *event)
935 id = event->parent->id;
941 * Get the perf_event_context for a task and lock it.
942 * This has to cope with with the fact that until it is locked,
943 * the context could get moved to another task.
945 static struct perf_event_context *
946 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
948 struct perf_event_context *ctx;
952 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
955 * If this context is a clone of another, it might
956 * get swapped for another underneath us by
957 * perf_event_task_sched_out, though the
958 * rcu_read_lock() protects us from any context
959 * getting freed. Lock the context and check if it
960 * got swapped before we could get the lock, and retry
961 * if so. If we locked the right context, then it
962 * can't get swapped on us any more.
964 raw_spin_lock_irqsave(&ctx->lock, *flags);
965 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
966 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
970 if (!atomic_inc_not_zero(&ctx->refcount)) {
971 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
980 * Get the context for a task and increment its pin_count so it
981 * can't get swapped to another task. This also increments its
982 * reference count so that the context can't get freed.
984 static struct perf_event_context *
985 perf_pin_task_context(struct task_struct *task, int ctxn)
987 struct perf_event_context *ctx;
990 ctx = perf_lock_task_context(task, ctxn, &flags);
993 raw_spin_unlock_irqrestore(&ctx->lock, flags);
998 static void perf_unpin_context(struct perf_event_context *ctx)
1000 unsigned long flags;
1002 raw_spin_lock_irqsave(&ctx->lock, flags);
1004 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1008 * Update the record of the current time in a context.
1010 static void update_context_time(struct perf_event_context *ctx)
1012 u64 now = perf_clock();
1014 ctx->time += now - ctx->timestamp;
1015 ctx->timestamp = now;
1018 static u64 perf_event_time(struct perf_event *event)
1020 struct perf_event_context *ctx = event->ctx;
1022 if (is_cgroup_event(event))
1023 return perf_cgroup_event_time(event);
1025 return ctx ? ctx->time : 0;
1029 * Update the total_time_enabled and total_time_running fields for a event.
1030 * The caller of this function needs to hold the ctx->lock.
1032 static void update_event_times(struct perf_event *event)
1034 struct perf_event_context *ctx = event->ctx;
1037 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1038 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1041 * in cgroup mode, time_enabled represents
1042 * the time the event was enabled AND active
1043 * tasks were in the monitored cgroup. This is
1044 * independent of the activity of the context as
1045 * there may be a mix of cgroup and non-cgroup events.
1047 * That is why we treat cgroup events differently
1050 if (is_cgroup_event(event))
1051 run_end = perf_cgroup_event_time(event);
1052 else if (ctx->is_active)
1053 run_end = ctx->time;
1055 run_end = event->tstamp_stopped;
1057 event->total_time_enabled = run_end - event->tstamp_enabled;
1059 if (event->state == PERF_EVENT_STATE_INACTIVE)
1060 run_end = event->tstamp_stopped;
1062 run_end = perf_event_time(event);
1064 event->total_time_running = run_end - event->tstamp_running;
1069 * Update total_time_enabled and total_time_running for all events in a group.
1071 static void update_group_times(struct perf_event *leader)
1073 struct perf_event *event;
1075 update_event_times(leader);
1076 list_for_each_entry(event, &leader->sibling_list, group_entry)
1077 update_event_times(event);
1080 static struct list_head *
1081 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1083 if (event->attr.pinned)
1084 return &ctx->pinned_groups;
1086 return &ctx->flexible_groups;
1090 * Add a event from the lists for its context.
1091 * Must be called with ctx->mutex and ctx->lock held.
1094 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1096 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1097 event->attach_state |= PERF_ATTACH_CONTEXT;
1100 * If we're a stand alone event or group leader, we go to the context
1101 * list, group events are kept attached to the group so that
1102 * perf_group_detach can, at all times, locate all siblings.
1104 if (event->group_leader == event) {
1105 struct list_head *list;
1107 if (is_software_event(event))
1108 event->group_flags |= PERF_GROUP_SOFTWARE;
1110 list = ctx_group_list(event, ctx);
1111 list_add_tail(&event->group_entry, list);
1114 if (is_cgroup_event(event))
1117 if (has_branch_stack(event))
1118 ctx->nr_branch_stack++;
1120 list_add_rcu(&event->event_entry, &ctx->event_list);
1121 if (!ctx->nr_events)
1122 perf_pmu_rotate_start(ctx->pmu);
1124 if (event->attr.inherit_stat)
1129 * Initialize event state based on the perf_event_attr::disabled.
1131 static inline void perf_event__state_init(struct perf_event *event)
1133 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1134 PERF_EVENT_STATE_INACTIVE;
1138 * Called at perf_event creation and when events are attached/detached from a
1141 static void perf_event__read_size(struct perf_event *event)
1143 int entry = sizeof(u64); /* value */
1147 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1148 size += sizeof(u64);
1150 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1151 size += sizeof(u64);
1153 if (event->attr.read_format & PERF_FORMAT_ID)
1154 entry += sizeof(u64);
1156 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1157 nr += event->group_leader->nr_siblings;
1158 size += sizeof(u64);
1162 event->read_size = size;
1165 static void perf_event__header_size(struct perf_event *event)
1167 struct perf_sample_data *data;
1168 u64 sample_type = event->attr.sample_type;
1171 perf_event__read_size(event);
1173 if (sample_type & PERF_SAMPLE_IP)
1174 size += sizeof(data->ip);
1176 if (sample_type & PERF_SAMPLE_ADDR)
1177 size += sizeof(data->addr);
1179 if (sample_type & PERF_SAMPLE_PERIOD)
1180 size += sizeof(data->period);
1182 if (sample_type & PERF_SAMPLE_WEIGHT)
1183 size += sizeof(data->weight);
1185 if (sample_type & PERF_SAMPLE_READ)
1186 size += event->read_size;
1188 if (sample_type & PERF_SAMPLE_DATA_SRC)
1189 size += sizeof(data->data_src.val);
1191 event->header_size = size;
1194 static void perf_event__id_header_size(struct perf_event *event)
1196 struct perf_sample_data *data;
1197 u64 sample_type = event->attr.sample_type;
1200 if (sample_type & PERF_SAMPLE_TID)
1201 size += sizeof(data->tid_entry);
1203 if (sample_type & PERF_SAMPLE_TIME)
1204 size += sizeof(data->time);
1206 if (sample_type & PERF_SAMPLE_ID)
1207 size += sizeof(data->id);
1209 if (sample_type & PERF_SAMPLE_STREAM_ID)
1210 size += sizeof(data->stream_id);
1212 if (sample_type & PERF_SAMPLE_CPU)
1213 size += sizeof(data->cpu_entry);
1215 event->id_header_size = size;
1218 static void perf_group_attach(struct perf_event *event)
1220 struct perf_event *group_leader = event->group_leader, *pos;
1223 * We can have double attach due to group movement in perf_event_open.
1225 if (event->attach_state & PERF_ATTACH_GROUP)
1228 event->attach_state |= PERF_ATTACH_GROUP;
1230 if (group_leader == event)
1233 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1234 !is_software_event(event))
1235 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1237 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1238 group_leader->nr_siblings++;
1240 perf_event__header_size(group_leader);
1242 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1243 perf_event__header_size(pos);
1247 * Remove a event from the lists for its context.
1248 * Must be called with ctx->mutex and ctx->lock held.
1251 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1253 struct perf_cpu_context *cpuctx;
1255 * We can have double detach due to exit/hot-unplug + close.
1257 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1260 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1262 if (is_cgroup_event(event)) {
1264 cpuctx = __get_cpu_context(ctx);
1266 * if there are no more cgroup events
1267 * then cler cgrp to avoid stale pointer
1268 * in update_cgrp_time_from_cpuctx()
1270 if (!ctx->nr_cgroups)
1271 cpuctx->cgrp = NULL;
1274 if (has_branch_stack(event))
1275 ctx->nr_branch_stack--;
1278 if (event->attr.inherit_stat)
1281 list_del_rcu(&event->event_entry);
1283 if (event->group_leader == event)
1284 list_del_init(&event->group_entry);
1286 update_group_times(event);
1289 * If event was in error state, then keep it
1290 * that way, otherwise bogus counts will be
1291 * returned on read(). The only way to get out
1292 * of error state is by explicit re-enabling
1295 if (event->state > PERF_EVENT_STATE_OFF)
1296 event->state = PERF_EVENT_STATE_OFF;
1299 static void perf_group_detach(struct perf_event *event)
1301 struct perf_event *sibling, *tmp;
1302 struct list_head *list = NULL;
1305 * We can have double detach due to exit/hot-unplug + close.
1307 if (!(event->attach_state & PERF_ATTACH_GROUP))
1310 event->attach_state &= ~PERF_ATTACH_GROUP;
1313 * If this is a sibling, remove it from its group.
1315 if (event->group_leader != event) {
1316 list_del_init(&event->group_entry);
1317 event->group_leader->nr_siblings--;
1321 if (!list_empty(&event->group_entry))
1322 list = &event->group_entry;
1325 * If this was a group event with sibling events then
1326 * upgrade the siblings to singleton events by adding them
1327 * to whatever list we are on.
1329 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1331 list_move_tail(&sibling->group_entry, list);
1332 sibling->group_leader = sibling;
1334 /* Inherit group flags from the previous leader */
1335 sibling->group_flags = event->group_flags;
1339 perf_event__header_size(event->group_leader);
1341 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1342 perf_event__header_size(tmp);
1346 event_filter_match(struct perf_event *event)
1348 return (event->cpu == -1 || event->cpu == smp_processor_id())
1349 && perf_cgroup_match(event);
1353 event_sched_out(struct perf_event *event,
1354 struct perf_cpu_context *cpuctx,
1355 struct perf_event_context *ctx)
1357 u64 tstamp = perf_event_time(event);
1360 * An event which could not be activated because of
1361 * filter mismatch still needs to have its timings
1362 * maintained, otherwise bogus information is return
1363 * via read() for time_enabled, time_running:
1365 if (event->state == PERF_EVENT_STATE_INACTIVE
1366 && !event_filter_match(event)) {
1367 delta = tstamp - event->tstamp_stopped;
1368 event->tstamp_running += delta;
1369 event->tstamp_stopped = tstamp;
1372 if (event->state != PERF_EVENT_STATE_ACTIVE)
1375 event->state = PERF_EVENT_STATE_INACTIVE;
1376 if (event->pending_disable) {
1377 event->pending_disable = 0;
1378 event->state = PERF_EVENT_STATE_OFF;
1380 event->tstamp_stopped = tstamp;
1381 event->pmu->del(event, 0);
1384 if (!is_software_event(event))
1385 cpuctx->active_oncpu--;
1387 if (event->attr.freq && event->attr.sample_freq)
1389 if (event->attr.exclusive || !cpuctx->active_oncpu)
1390 cpuctx->exclusive = 0;
1394 group_sched_out(struct perf_event *group_event,
1395 struct perf_cpu_context *cpuctx,
1396 struct perf_event_context *ctx)
1398 struct perf_event *event;
1399 int state = group_event->state;
1401 event_sched_out(group_event, cpuctx, ctx);
1404 * Schedule out siblings (if any):
1406 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1407 event_sched_out(event, cpuctx, ctx);
1409 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1410 cpuctx->exclusive = 0;
1414 * Cross CPU call to remove a performance event
1416 * We disable the event on the hardware level first. After that we
1417 * remove it from the context list.
1419 static int __perf_remove_from_context(void *info)
1421 struct perf_event *event = info;
1422 struct perf_event_context *ctx = event->ctx;
1423 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1425 raw_spin_lock(&ctx->lock);
1426 event_sched_out(event, cpuctx, ctx);
1427 list_del_event(event, ctx);
1428 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1430 cpuctx->task_ctx = NULL;
1432 raw_spin_unlock(&ctx->lock);
1439 * Remove the event from a task's (or a CPU's) list of events.
1441 * CPU events are removed with a smp call. For task events we only
1442 * call when the task is on a CPU.
1444 * If event->ctx is a cloned context, callers must make sure that
1445 * every task struct that event->ctx->task could possibly point to
1446 * remains valid. This is OK when called from perf_release since
1447 * that only calls us on the top-level context, which can't be a clone.
1448 * When called from perf_event_exit_task, it's OK because the
1449 * context has been detached from its task.
1451 static void perf_remove_from_context(struct perf_event *event)
1453 struct perf_event_context *ctx = event->ctx;
1454 struct task_struct *task = ctx->task;
1456 lockdep_assert_held(&ctx->mutex);
1460 * Per cpu events are removed via an smp call and
1461 * the removal is always successful.
1463 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1468 if (!task_function_call(task, __perf_remove_from_context, event))
1471 raw_spin_lock_irq(&ctx->lock);
1473 * If we failed to find a running task, but find the context active now
1474 * that we've acquired the ctx->lock, retry.
1476 if (ctx->is_active) {
1477 raw_spin_unlock_irq(&ctx->lock);
1482 * Since the task isn't running, its safe to remove the event, us
1483 * holding the ctx->lock ensures the task won't get scheduled in.
1485 list_del_event(event, ctx);
1486 raw_spin_unlock_irq(&ctx->lock);
1490 * Cross CPU call to disable a performance event
1492 int __perf_event_disable(void *info)
1494 struct perf_event *event = info;
1495 struct perf_event_context *ctx = event->ctx;
1496 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1499 * If this is a per-task event, need to check whether this
1500 * event's task is the current task on this cpu.
1502 * Can trigger due to concurrent perf_event_context_sched_out()
1503 * flipping contexts around.
1505 if (ctx->task && cpuctx->task_ctx != ctx)
1508 raw_spin_lock(&ctx->lock);
1511 * If the event is on, turn it off.
1512 * If it is in error state, leave it in error state.
1514 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1515 update_context_time(ctx);
1516 update_cgrp_time_from_event(event);
1517 update_group_times(event);
1518 if (event == event->group_leader)
1519 group_sched_out(event, cpuctx, ctx);
1521 event_sched_out(event, cpuctx, ctx);
1522 event->state = PERF_EVENT_STATE_OFF;
1525 raw_spin_unlock(&ctx->lock);
1533 * If event->ctx is a cloned context, callers must make sure that
1534 * every task struct that event->ctx->task could possibly point to
1535 * remains valid. This condition is satisifed when called through
1536 * perf_event_for_each_child or perf_event_for_each because they
1537 * hold the top-level event's child_mutex, so any descendant that
1538 * goes to exit will block in sync_child_event.
1539 * When called from perf_pending_event it's OK because event->ctx
1540 * is the current context on this CPU and preemption is disabled,
1541 * hence we can't get into perf_event_task_sched_out for this context.
1543 void perf_event_disable(struct perf_event *event)
1545 struct perf_event_context *ctx = event->ctx;
1546 struct task_struct *task = ctx->task;
1550 * Disable the event on the cpu that it's on
1552 cpu_function_call(event->cpu, __perf_event_disable, event);
1557 if (!task_function_call(task, __perf_event_disable, event))
1560 raw_spin_lock_irq(&ctx->lock);
1562 * If the event is still active, we need to retry the cross-call.
1564 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1565 raw_spin_unlock_irq(&ctx->lock);
1567 * Reload the task pointer, it might have been changed by
1568 * a concurrent perf_event_context_sched_out().
1575 * Since we have the lock this context can't be scheduled
1576 * in, so we can change the state safely.
1578 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1579 update_group_times(event);
1580 event->state = PERF_EVENT_STATE_OFF;
1582 raw_spin_unlock_irq(&ctx->lock);
1584 EXPORT_SYMBOL_GPL(perf_event_disable);
1586 static void perf_set_shadow_time(struct perf_event *event,
1587 struct perf_event_context *ctx,
1591 * use the correct time source for the time snapshot
1593 * We could get by without this by leveraging the
1594 * fact that to get to this function, the caller
1595 * has most likely already called update_context_time()
1596 * and update_cgrp_time_xx() and thus both timestamp
1597 * are identical (or very close). Given that tstamp is,
1598 * already adjusted for cgroup, we could say that:
1599 * tstamp - ctx->timestamp
1601 * tstamp - cgrp->timestamp.
1603 * Then, in perf_output_read(), the calculation would
1604 * work with no changes because:
1605 * - event is guaranteed scheduled in
1606 * - no scheduled out in between
1607 * - thus the timestamp would be the same
1609 * But this is a bit hairy.
1611 * So instead, we have an explicit cgroup call to remain
1612 * within the time time source all along. We believe it
1613 * is cleaner and simpler to understand.
1615 if (is_cgroup_event(event))
1616 perf_cgroup_set_shadow_time(event, tstamp);
1618 event->shadow_ctx_time = tstamp - ctx->timestamp;
1621 #define MAX_INTERRUPTS (~0ULL)
1623 static void perf_log_throttle(struct perf_event *event, int enable);
1626 event_sched_in(struct perf_event *event,
1627 struct perf_cpu_context *cpuctx,
1628 struct perf_event_context *ctx)
1630 u64 tstamp = perf_event_time(event);
1632 if (event->state <= PERF_EVENT_STATE_OFF)
1635 event->state = PERF_EVENT_STATE_ACTIVE;
1636 event->oncpu = smp_processor_id();
1639 * Unthrottle events, since we scheduled we might have missed several
1640 * ticks already, also for a heavily scheduling task there is little
1641 * guarantee it'll get a tick in a timely manner.
1643 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1644 perf_log_throttle(event, 1);
1645 event->hw.interrupts = 0;
1649 * The new state must be visible before we turn it on in the hardware:
1653 if (event->pmu->add(event, PERF_EF_START)) {
1654 event->state = PERF_EVENT_STATE_INACTIVE;
1659 event->tstamp_running += tstamp - event->tstamp_stopped;
1661 perf_set_shadow_time(event, ctx, tstamp);
1663 if (!is_software_event(event))
1664 cpuctx->active_oncpu++;
1666 if (event->attr.freq && event->attr.sample_freq)
1669 if (event->attr.exclusive)
1670 cpuctx->exclusive = 1;
1676 group_sched_in(struct perf_event *group_event,
1677 struct perf_cpu_context *cpuctx,
1678 struct perf_event_context *ctx)
1680 struct perf_event *event, *partial_group = NULL;
1681 struct pmu *pmu = group_event->pmu;
1682 u64 now = ctx->time;
1683 bool simulate = false;
1685 if (group_event->state == PERF_EVENT_STATE_OFF)
1688 pmu->start_txn(pmu);
1690 if (event_sched_in(group_event, cpuctx, ctx)) {
1691 pmu->cancel_txn(pmu);
1692 perf_cpu_hrtimer_restart(cpuctx);
1697 * Schedule in siblings as one group (if any):
1699 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1700 if (event_sched_in(event, cpuctx, ctx)) {
1701 partial_group = event;
1706 if (!pmu->commit_txn(pmu))
1711 * Groups can be scheduled in as one unit only, so undo any
1712 * partial group before returning:
1713 * The events up to the failed event are scheduled out normally,
1714 * tstamp_stopped will be updated.
1716 * The failed events and the remaining siblings need to have
1717 * their timings updated as if they had gone thru event_sched_in()
1718 * and event_sched_out(). This is required to get consistent timings
1719 * across the group. This also takes care of the case where the group
1720 * could never be scheduled by ensuring tstamp_stopped is set to mark
1721 * the time the event was actually stopped, such that time delta
1722 * calculation in update_event_times() is correct.
1724 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1725 if (event == partial_group)
1729 event->tstamp_running += now - event->tstamp_stopped;
1730 event->tstamp_stopped = now;
1732 event_sched_out(event, cpuctx, ctx);
1735 event_sched_out(group_event, cpuctx, ctx);
1737 pmu->cancel_txn(pmu);
1739 perf_cpu_hrtimer_restart(cpuctx);
1745 * Work out whether we can put this event group on the CPU now.
1747 static int group_can_go_on(struct perf_event *event,
1748 struct perf_cpu_context *cpuctx,
1752 * Groups consisting entirely of software events can always go on.
1754 if (event->group_flags & PERF_GROUP_SOFTWARE)
1757 * If an exclusive group is already on, no other hardware
1760 if (cpuctx->exclusive)
1763 * If this group is exclusive and there are already
1764 * events on the CPU, it can't go on.
1766 if (event->attr.exclusive && cpuctx->active_oncpu)
1769 * Otherwise, try to add it if all previous groups were able
1775 static void add_event_to_ctx(struct perf_event *event,
1776 struct perf_event_context *ctx)
1778 u64 tstamp = perf_event_time(event);
1780 list_add_event(event, ctx);
1781 perf_group_attach(event);
1782 event->tstamp_enabled = tstamp;
1783 event->tstamp_running = tstamp;
1784 event->tstamp_stopped = tstamp;
1787 static void task_ctx_sched_out(struct perf_event_context *ctx);
1789 ctx_sched_in(struct perf_event_context *ctx,
1790 struct perf_cpu_context *cpuctx,
1791 enum event_type_t event_type,
1792 struct task_struct *task);
1794 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1795 struct perf_event_context *ctx,
1796 struct task_struct *task)
1798 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1800 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1801 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1803 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1807 * Cross CPU call to install and enable a performance event
1809 * Must be called with ctx->mutex held
1811 static int __perf_install_in_context(void *info)
1813 struct perf_event *event = info;
1814 struct perf_event_context *ctx = event->ctx;
1815 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1816 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1817 struct task_struct *task = current;
1819 perf_ctx_lock(cpuctx, task_ctx);
1820 perf_pmu_disable(cpuctx->ctx.pmu);
1823 * If there was an active task_ctx schedule it out.
1826 task_ctx_sched_out(task_ctx);
1829 * If the context we're installing events in is not the
1830 * active task_ctx, flip them.
1832 if (ctx->task && task_ctx != ctx) {
1834 raw_spin_unlock(&task_ctx->lock);
1835 raw_spin_lock(&ctx->lock);
1840 cpuctx->task_ctx = task_ctx;
1841 task = task_ctx->task;
1844 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1846 update_context_time(ctx);
1848 * update cgrp time only if current cgrp
1849 * matches event->cgrp. Must be done before
1850 * calling add_event_to_ctx()
1852 update_cgrp_time_from_event(event);
1854 add_event_to_ctx(event, ctx);
1857 * Schedule everything back in
1859 perf_event_sched_in(cpuctx, task_ctx, task);
1861 perf_pmu_enable(cpuctx->ctx.pmu);
1862 perf_ctx_unlock(cpuctx, task_ctx);
1868 * Attach a performance event to a context
1870 * First we add the event to the list with the hardware enable bit
1871 * in event->hw_config cleared.
1873 * If the event is attached to a task which is on a CPU we use a smp
1874 * call to enable it in the task context. The task might have been
1875 * scheduled away, but we check this in the smp call again.
1878 perf_install_in_context(struct perf_event_context *ctx,
1879 struct perf_event *event,
1882 struct task_struct *task = ctx->task;
1884 lockdep_assert_held(&ctx->mutex);
1887 if (event->cpu != -1)
1892 * Per cpu events are installed via an smp call and
1893 * the install is always successful.
1895 cpu_function_call(cpu, __perf_install_in_context, event);
1900 if (!task_function_call(task, __perf_install_in_context, event))
1903 raw_spin_lock_irq(&ctx->lock);
1905 * If we failed to find a running task, but find the context active now
1906 * that we've acquired the ctx->lock, retry.
1908 if (ctx->is_active) {
1909 raw_spin_unlock_irq(&ctx->lock);
1914 * Since the task isn't running, its safe to add the event, us holding
1915 * the ctx->lock ensures the task won't get scheduled in.
1917 add_event_to_ctx(event, ctx);
1918 raw_spin_unlock_irq(&ctx->lock);
1922 * Put a event into inactive state and update time fields.
1923 * Enabling the leader of a group effectively enables all
1924 * the group members that aren't explicitly disabled, so we
1925 * have to update their ->tstamp_enabled also.
1926 * Note: this works for group members as well as group leaders
1927 * since the non-leader members' sibling_lists will be empty.
1929 static void __perf_event_mark_enabled(struct perf_event *event)
1931 struct perf_event *sub;
1932 u64 tstamp = perf_event_time(event);
1934 event->state = PERF_EVENT_STATE_INACTIVE;
1935 event->tstamp_enabled = tstamp - event->total_time_enabled;
1936 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1937 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1938 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1943 * Cross CPU call to enable a performance event
1945 static int __perf_event_enable(void *info)
1947 struct perf_event *event = info;
1948 struct perf_event_context *ctx = event->ctx;
1949 struct perf_event *leader = event->group_leader;
1950 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1954 * There's a time window between 'ctx->is_active' check
1955 * in perf_event_enable function and this place having:
1957 * - ctx->lock unlocked
1959 * where the task could be killed and 'ctx' deactivated
1960 * by perf_event_exit_task.
1962 if (!ctx->is_active)
1965 raw_spin_lock(&ctx->lock);
1966 update_context_time(ctx);
1968 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1972 * set current task's cgroup time reference point
1974 perf_cgroup_set_timestamp(current, ctx);
1976 __perf_event_mark_enabled(event);
1978 if (!event_filter_match(event)) {
1979 if (is_cgroup_event(event))
1980 perf_cgroup_defer_enabled(event);
1985 * If the event is in a group and isn't the group leader,
1986 * then don't put it on unless the group is on.
1988 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1991 if (!group_can_go_on(event, cpuctx, 1)) {
1994 if (event == leader)
1995 err = group_sched_in(event, cpuctx, ctx);
1997 err = event_sched_in(event, cpuctx, ctx);
2002 * If this event can't go on and it's part of a
2003 * group, then the whole group has to come off.
2005 if (leader != event) {
2006 group_sched_out(leader, cpuctx, ctx);
2007 perf_cpu_hrtimer_restart(cpuctx);
2009 if (leader->attr.pinned) {
2010 update_group_times(leader);
2011 leader->state = PERF_EVENT_STATE_ERROR;
2016 raw_spin_unlock(&ctx->lock);
2024 * If event->ctx is a cloned context, callers must make sure that
2025 * every task struct that event->ctx->task could possibly point to
2026 * remains valid. This condition is satisfied when called through
2027 * perf_event_for_each_child or perf_event_for_each as described
2028 * for perf_event_disable.
2030 void perf_event_enable(struct perf_event *event)
2032 struct perf_event_context *ctx = event->ctx;
2033 struct task_struct *task = ctx->task;
2037 * Enable the event on the cpu that it's on
2039 cpu_function_call(event->cpu, __perf_event_enable, event);
2043 raw_spin_lock_irq(&ctx->lock);
2044 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2048 * If the event is in error state, clear that first.
2049 * That way, if we see the event in error state below, we
2050 * know that it has gone back into error state, as distinct
2051 * from the task having been scheduled away before the
2052 * cross-call arrived.
2054 if (event->state == PERF_EVENT_STATE_ERROR)
2055 event->state = PERF_EVENT_STATE_OFF;
2058 if (!ctx->is_active) {
2059 __perf_event_mark_enabled(event);
2063 raw_spin_unlock_irq(&ctx->lock);
2065 if (!task_function_call(task, __perf_event_enable, event))
2068 raw_spin_lock_irq(&ctx->lock);
2071 * If the context is active and the event is still off,
2072 * we need to retry the cross-call.
2074 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2076 * task could have been flipped by a concurrent
2077 * perf_event_context_sched_out()
2084 raw_spin_unlock_irq(&ctx->lock);
2086 EXPORT_SYMBOL_GPL(perf_event_enable);
2088 int perf_event_refresh(struct perf_event *event, int refresh)
2091 * not supported on inherited events
2093 if (event->attr.inherit || !is_sampling_event(event))
2096 atomic_add(refresh, &event->event_limit);
2097 perf_event_enable(event);
2101 EXPORT_SYMBOL_GPL(perf_event_refresh);
2103 static void ctx_sched_out(struct perf_event_context *ctx,
2104 struct perf_cpu_context *cpuctx,
2105 enum event_type_t event_type)
2107 struct perf_event *event;
2108 int is_active = ctx->is_active;
2110 ctx->is_active &= ~event_type;
2111 if (likely(!ctx->nr_events))
2114 update_context_time(ctx);
2115 update_cgrp_time_from_cpuctx(cpuctx);
2116 if (!ctx->nr_active)
2119 perf_pmu_disable(ctx->pmu);
2120 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2121 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2122 group_sched_out(event, cpuctx, ctx);
2125 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2126 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2127 group_sched_out(event, cpuctx, ctx);
2129 perf_pmu_enable(ctx->pmu);
2133 * Test whether two contexts are equivalent, i.e. whether they
2134 * have both been cloned from the same version of the same context
2135 * and they both have the same number of enabled events.
2136 * If the number of enabled events is the same, then the set
2137 * of enabled events should be the same, because these are both
2138 * inherited contexts, therefore we can't access individual events
2139 * in them directly with an fd; we can only enable/disable all
2140 * events via prctl, or enable/disable all events in a family
2141 * via ioctl, which will have the same effect on both contexts.
2143 static int context_equiv(struct perf_event_context *ctx1,
2144 struct perf_event_context *ctx2)
2146 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
2147 && ctx1->parent_gen == ctx2->parent_gen
2148 && !ctx1->pin_count && !ctx2->pin_count;
2151 static void __perf_event_sync_stat(struct perf_event *event,
2152 struct perf_event *next_event)
2156 if (!event->attr.inherit_stat)
2160 * Update the event value, we cannot use perf_event_read()
2161 * because we're in the middle of a context switch and have IRQs
2162 * disabled, which upsets smp_call_function_single(), however
2163 * we know the event must be on the current CPU, therefore we
2164 * don't need to use it.
2166 switch (event->state) {
2167 case PERF_EVENT_STATE_ACTIVE:
2168 event->pmu->read(event);
2171 case PERF_EVENT_STATE_INACTIVE:
2172 update_event_times(event);
2180 * In order to keep per-task stats reliable we need to flip the event
2181 * values when we flip the contexts.
2183 value = local64_read(&next_event->count);
2184 value = local64_xchg(&event->count, value);
2185 local64_set(&next_event->count, value);
2187 swap(event->total_time_enabled, next_event->total_time_enabled);
2188 swap(event->total_time_running, next_event->total_time_running);
2191 * Since we swizzled the values, update the user visible data too.
2193 perf_event_update_userpage(event);
2194 perf_event_update_userpage(next_event);
2197 #define list_next_entry(pos, member) \
2198 list_entry(pos->member.next, typeof(*pos), member)
2200 static void perf_event_sync_stat(struct perf_event_context *ctx,
2201 struct perf_event_context *next_ctx)
2203 struct perf_event *event, *next_event;
2208 update_context_time(ctx);
2210 event = list_first_entry(&ctx->event_list,
2211 struct perf_event, event_entry);
2213 next_event = list_first_entry(&next_ctx->event_list,
2214 struct perf_event, event_entry);
2216 while (&event->event_entry != &ctx->event_list &&
2217 &next_event->event_entry != &next_ctx->event_list) {
2219 __perf_event_sync_stat(event, next_event);
2221 event = list_next_entry(event, event_entry);
2222 next_event = list_next_entry(next_event, event_entry);
2226 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2227 struct task_struct *next)
2229 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2230 struct perf_event_context *next_ctx;
2231 struct perf_event_context *parent;
2232 struct perf_cpu_context *cpuctx;
2238 cpuctx = __get_cpu_context(ctx);
2239 if (!cpuctx->task_ctx)
2243 parent = rcu_dereference(ctx->parent_ctx);
2244 next_ctx = next->perf_event_ctxp[ctxn];
2245 if (parent && next_ctx &&
2246 rcu_dereference(next_ctx->parent_ctx) == parent) {
2248 * Looks like the two contexts are clones, so we might be
2249 * able to optimize the context switch. We lock both
2250 * contexts and check that they are clones under the
2251 * lock (including re-checking that neither has been
2252 * uncloned in the meantime). It doesn't matter which
2253 * order we take the locks because no other cpu could
2254 * be trying to lock both of these tasks.
2256 raw_spin_lock(&ctx->lock);
2257 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2258 if (context_equiv(ctx, next_ctx)) {
2260 * XXX do we need a memory barrier of sorts
2261 * wrt to rcu_dereference() of perf_event_ctxp
2263 task->perf_event_ctxp[ctxn] = next_ctx;
2264 next->perf_event_ctxp[ctxn] = ctx;
2266 next_ctx->task = task;
2269 perf_event_sync_stat(ctx, next_ctx);
2271 raw_spin_unlock(&next_ctx->lock);
2272 raw_spin_unlock(&ctx->lock);
2277 raw_spin_lock(&ctx->lock);
2278 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2279 cpuctx->task_ctx = NULL;
2280 raw_spin_unlock(&ctx->lock);
2284 #define for_each_task_context_nr(ctxn) \
2285 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2288 * Called from scheduler to remove the events of the current task,
2289 * with interrupts disabled.
2291 * We stop each event and update the event value in event->count.
2293 * This does not protect us against NMI, but disable()
2294 * sets the disabled bit in the control field of event _before_
2295 * accessing the event control register. If a NMI hits, then it will
2296 * not restart the event.
2298 void __perf_event_task_sched_out(struct task_struct *task,
2299 struct task_struct *next)
2303 for_each_task_context_nr(ctxn)
2304 perf_event_context_sched_out(task, ctxn, next);
2307 * if cgroup events exist on this CPU, then we need
2308 * to check if we have to switch out PMU state.
2309 * cgroup event are system-wide mode only
2311 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2312 perf_cgroup_sched_out(task, next);
2315 static void task_ctx_sched_out(struct perf_event_context *ctx)
2317 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2319 if (!cpuctx->task_ctx)
2322 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2325 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2326 cpuctx->task_ctx = NULL;
2330 * Called with IRQs disabled
2332 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2333 enum event_type_t event_type)
2335 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2339 ctx_pinned_sched_in(struct perf_event_context *ctx,
2340 struct perf_cpu_context *cpuctx)
2342 struct perf_event *event;
2344 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2345 if (event->state <= PERF_EVENT_STATE_OFF)
2347 if (!event_filter_match(event))
2350 /* may need to reset tstamp_enabled */
2351 if (is_cgroup_event(event))
2352 perf_cgroup_mark_enabled(event, ctx);
2354 if (group_can_go_on(event, cpuctx, 1))
2355 group_sched_in(event, cpuctx, ctx);
2358 * If this pinned group hasn't been scheduled,
2359 * put it in error state.
2361 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2362 update_group_times(event);
2363 event->state = PERF_EVENT_STATE_ERROR;
2369 ctx_flexible_sched_in(struct perf_event_context *ctx,
2370 struct perf_cpu_context *cpuctx)
2372 struct perf_event *event;
2375 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2376 /* Ignore events in OFF or ERROR state */
2377 if (event->state <= PERF_EVENT_STATE_OFF)
2380 * Listen to the 'cpu' scheduling filter constraint
2383 if (!event_filter_match(event))
2386 /* may need to reset tstamp_enabled */
2387 if (is_cgroup_event(event))
2388 perf_cgroup_mark_enabled(event, ctx);
2390 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2391 if (group_sched_in(event, cpuctx, ctx))
2398 ctx_sched_in(struct perf_event_context *ctx,
2399 struct perf_cpu_context *cpuctx,
2400 enum event_type_t event_type,
2401 struct task_struct *task)
2404 int is_active = ctx->is_active;
2406 ctx->is_active |= event_type;
2407 if (likely(!ctx->nr_events))
2411 ctx->timestamp = now;
2412 perf_cgroup_set_timestamp(task, ctx);
2414 * First go through the list and put on any pinned groups
2415 * in order to give them the best chance of going on.
2417 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2418 ctx_pinned_sched_in(ctx, cpuctx);
2420 /* Then walk through the lower prio flexible groups */
2421 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2422 ctx_flexible_sched_in(ctx, cpuctx);
2425 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2426 enum event_type_t event_type,
2427 struct task_struct *task)
2429 struct perf_event_context *ctx = &cpuctx->ctx;
2431 ctx_sched_in(ctx, cpuctx, event_type, task);
2434 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2435 struct task_struct *task)
2437 struct perf_cpu_context *cpuctx;
2439 cpuctx = __get_cpu_context(ctx);
2440 if (cpuctx->task_ctx == ctx)
2443 perf_ctx_lock(cpuctx, ctx);
2444 perf_pmu_disable(ctx->pmu);
2446 * We want to keep the following priority order:
2447 * cpu pinned (that don't need to move), task pinned,
2448 * cpu flexible, task flexible.
2450 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2453 cpuctx->task_ctx = ctx;
2455 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2457 perf_pmu_enable(ctx->pmu);
2458 perf_ctx_unlock(cpuctx, ctx);
2461 * Since these rotations are per-cpu, we need to ensure the
2462 * cpu-context we got scheduled on is actually rotating.
2464 perf_pmu_rotate_start(ctx->pmu);
2468 * When sampling the branck stack in system-wide, it may be necessary
2469 * to flush the stack on context switch. This happens when the branch
2470 * stack does not tag its entries with the pid of the current task.
2471 * Otherwise it becomes impossible to associate a branch entry with a
2472 * task. This ambiguity is more likely to appear when the branch stack
2473 * supports priv level filtering and the user sets it to monitor only
2474 * at the user level (which could be a useful measurement in system-wide
2475 * mode). In that case, the risk is high of having a branch stack with
2476 * branch from multiple tasks. Flushing may mean dropping the existing
2477 * entries or stashing them somewhere in the PMU specific code layer.
2479 * This function provides the context switch callback to the lower code
2480 * layer. It is invoked ONLY when there is at least one system-wide context
2481 * with at least one active event using taken branch sampling.
2483 static void perf_branch_stack_sched_in(struct task_struct *prev,
2484 struct task_struct *task)
2486 struct perf_cpu_context *cpuctx;
2488 unsigned long flags;
2490 /* no need to flush branch stack if not changing task */
2494 local_irq_save(flags);
2498 list_for_each_entry_rcu(pmu, &pmus, entry) {
2499 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2502 * check if the context has at least one
2503 * event using PERF_SAMPLE_BRANCH_STACK
2505 if (cpuctx->ctx.nr_branch_stack > 0
2506 && pmu->flush_branch_stack) {
2508 pmu = cpuctx->ctx.pmu;
2510 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2512 perf_pmu_disable(pmu);
2514 pmu->flush_branch_stack();
2516 perf_pmu_enable(pmu);
2518 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2524 local_irq_restore(flags);
2528 * Called from scheduler to add the events of the current task
2529 * with interrupts disabled.
2531 * We restore the event value and then enable it.
2533 * This does not protect us against NMI, but enable()
2534 * sets the enabled bit in the control field of event _before_
2535 * accessing the event control register. If a NMI hits, then it will
2536 * keep the event running.
2538 void __perf_event_task_sched_in(struct task_struct *prev,
2539 struct task_struct *task)
2541 struct perf_event_context *ctx;
2544 for_each_task_context_nr(ctxn) {
2545 ctx = task->perf_event_ctxp[ctxn];
2549 perf_event_context_sched_in(ctx, task);
2552 * if cgroup events exist on this CPU, then we need
2553 * to check if we have to switch in PMU state.
2554 * cgroup event are system-wide mode only
2556 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2557 perf_cgroup_sched_in(prev, task);
2559 /* check for system-wide branch_stack events */
2560 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2561 perf_branch_stack_sched_in(prev, task);
2564 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2566 u64 frequency = event->attr.sample_freq;
2567 u64 sec = NSEC_PER_SEC;
2568 u64 divisor, dividend;
2570 int count_fls, nsec_fls, frequency_fls, sec_fls;
2572 count_fls = fls64(count);
2573 nsec_fls = fls64(nsec);
2574 frequency_fls = fls64(frequency);
2578 * We got @count in @nsec, with a target of sample_freq HZ
2579 * the target period becomes:
2582 * period = -------------------
2583 * @nsec * sample_freq
2588 * Reduce accuracy by one bit such that @a and @b converge
2589 * to a similar magnitude.
2591 #define REDUCE_FLS(a, b) \
2593 if (a##_fls > b##_fls) { \
2603 * Reduce accuracy until either term fits in a u64, then proceed with
2604 * the other, so that finally we can do a u64/u64 division.
2606 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2607 REDUCE_FLS(nsec, frequency);
2608 REDUCE_FLS(sec, count);
2611 if (count_fls + sec_fls > 64) {
2612 divisor = nsec * frequency;
2614 while (count_fls + sec_fls > 64) {
2615 REDUCE_FLS(count, sec);
2619 dividend = count * sec;
2621 dividend = count * sec;
2623 while (nsec_fls + frequency_fls > 64) {
2624 REDUCE_FLS(nsec, frequency);
2628 divisor = nsec * frequency;
2634 return div64_u64(dividend, divisor);
2637 static DEFINE_PER_CPU(int, perf_throttled_count);
2638 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2640 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2642 struct hw_perf_event *hwc = &event->hw;
2643 s64 period, sample_period;
2646 period = perf_calculate_period(event, nsec, count);
2648 delta = (s64)(period - hwc->sample_period);
2649 delta = (delta + 7) / 8; /* low pass filter */
2651 sample_period = hwc->sample_period + delta;
2656 hwc->sample_period = sample_period;
2658 if (local64_read(&hwc->period_left) > 8*sample_period) {
2660 event->pmu->stop(event, PERF_EF_UPDATE);
2662 local64_set(&hwc->period_left, 0);
2665 event->pmu->start(event, PERF_EF_RELOAD);
2670 * combine freq adjustment with unthrottling to avoid two passes over the
2671 * events. At the same time, make sure, having freq events does not change
2672 * the rate of unthrottling as that would introduce bias.
2674 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2677 struct perf_event *event;
2678 struct hw_perf_event *hwc;
2679 u64 now, period = TICK_NSEC;
2683 * only need to iterate over all events iff:
2684 * - context have events in frequency mode (needs freq adjust)
2685 * - there are events to unthrottle on this cpu
2687 if (!(ctx->nr_freq || needs_unthr))
2690 raw_spin_lock(&ctx->lock);
2691 perf_pmu_disable(ctx->pmu);
2693 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2694 if (event->state != PERF_EVENT_STATE_ACTIVE)
2697 if (!event_filter_match(event))
2702 if (needs_unthr && hwc->interrupts == MAX_INTERRUPTS) {
2703 hwc->interrupts = 0;
2704 perf_log_throttle(event, 1);
2705 event->pmu->start(event, 0);
2708 if (!event->attr.freq || !event->attr.sample_freq)
2712 * stop the event and update event->count
2714 event->pmu->stop(event, PERF_EF_UPDATE);
2716 now = local64_read(&event->count);
2717 delta = now - hwc->freq_count_stamp;
2718 hwc->freq_count_stamp = now;
2722 * reload only if value has changed
2723 * we have stopped the event so tell that
2724 * to perf_adjust_period() to avoid stopping it
2728 perf_adjust_period(event, period, delta, false);
2730 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2733 perf_pmu_enable(ctx->pmu);
2734 raw_spin_unlock(&ctx->lock);
2738 * Round-robin a context's events:
2740 static void rotate_ctx(struct perf_event_context *ctx)
2743 * Rotate the first entry last of non-pinned groups. Rotation might be
2744 * disabled by the inheritance code.
2746 if (!ctx->rotate_disable)
2747 list_rotate_left(&ctx->flexible_groups);
2751 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2752 * because they're strictly cpu affine and rotate_start is called with IRQs
2753 * disabled, while rotate_context is called from IRQ context.
2755 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2757 struct perf_event_context *ctx = NULL;
2758 int rotate = 0, remove = 1;
2760 if (cpuctx->ctx.nr_events) {
2762 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2766 ctx = cpuctx->task_ctx;
2767 if (ctx && ctx->nr_events) {
2769 if (ctx->nr_events != ctx->nr_active)
2776 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2777 perf_pmu_disable(cpuctx->ctx.pmu);
2779 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2781 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2783 rotate_ctx(&cpuctx->ctx);
2787 perf_event_sched_in(cpuctx, ctx, current);
2789 perf_pmu_enable(cpuctx->ctx.pmu);
2790 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2793 list_del_init(&cpuctx->rotation_list);
2798 #ifdef CONFIG_NO_HZ_FULL
2799 bool perf_event_can_stop_tick(void)
2801 if (list_empty(&__get_cpu_var(rotation_list)))
2808 void perf_event_task_tick(void)
2810 struct list_head *head = &__get_cpu_var(rotation_list);
2811 struct perf_cpu_context *cpuctx, *tmp;
2812 struct perf_event_context *ctx;
2815 WARN_ON(!irqs_disabled());
2817 __this_cpu_inc(perf_throttled_seq);
2818 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2820 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2822 perf_adjust_freq_unthr_context(ctx, throttled);
2824 ctx = cpuctx->task_ctx;
2826 perf_adjust_freq_unthr_context(ctx, throttled);
2830 static int event_enable_on_exec(struct perf_event *event,
2831 struct perf_event_context *ctx)
2833 if (!event->attr.enable_on_exec)
2836 event->attr.enable_on_exec = 0;
2837 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2840 __perf_event_mark_enabled(event);
2846 * Enable all of a task's events that have been marked enable-on-exec.
2847 * This expects task == current.
2849 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2851 struct perf_event *event;
2852 unsigned long flags;
2856 local_irq_save(flags);
2857 if (!ctx || !ctx->nr_events)
2861 * We must ctxsw out cgroup events to avoid conflict
2862 * when invoking perf_task_event_sched_in() later on
2863 * in this function. Otherwise we end up trying to
2864 * ctxswin cgroup events which are already scheduled
2867 perf_cgroup_sched_out(current, NULL);
2869 raw_spin_lock(&ctx->lock);
2870 task_ctx_sched_out(ctx);
2872 list_for_each_entry(event, &ctx->event_list, event_entry) {
2873 ret = event_enable_on_exec(event, ctx);
2879 * Unclone this context if we enabled any event.
2884 raw_spin_unlock(&ctx->lock);
2887 * Also calls ctxswin for cgroup events, if any:
2889 perf_event_context_sched_in(ctx, ctx->task);
2891 local_irq_restore(flags);
2895 * Cross CPU call to read the hardware event
2897 static void __perf_event_read(void *info)
2899 struct perf_event *event = info;
2900 struct perf_event_context *ctx = event->ctx;
2901 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2904 * If this is a task context, we need to check whether it is
2905 * the current task context of this cpu. If not it has been
2906 * scheduled out before the smp call arrived. In that case
2907 * event->count would have been updated to a recent sample
2908 * when the event was scheduled out.
2910 if (ctx->task && cpuctx->task_ctx != ctx)
2913 raw_spin_lock(&ctx->lock);
2914 if (ctx->is_active) {
2915 update_context_time(ctx);
2916 update_cgrp_time_from_event(event);
2918 update_event_times(event);
2919 if (event->state == PERF_EVENT_STATE_ACTIVE)
2920 event->pmu->read(event);
2921 raw_spin_unlock(&ctx->lock);
2924 static inline u64 perf_event_count(struct perf_event *event)
2926 return local64_read(&event->count) + atomic64_read(&event->child_count);
2929 static u64 perf_event_read(struct perf_event *event)
2932 * If event is enabled and currently active on a CPU, update the
2933 * value in the event structure:
2935 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2936 smp_call_function_single(event->oncpu,
2937 __perf_event_read, event, 1);
2938 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2939 struct perf_event_context *ctx = event->ctx;
2940 unsigned long flags;
2942 raw_spin_lock_irqsave(&ctx->lock, flags);
2944 * may read while context is not active
2945 * (e.g., thread is blocked), in that case
2946 * we cannot update context time
2948 if (ctx->is_active) {
2949 update_context_time(ctx);
2950 update_cgrp_time_from_event(event);
2952 update_event_times(event);
2953 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2956 return perf_event_count(event);
2960 * Initialize the perf_event context in a task_struct:
2962 static void __perf_event_init_context(struct perf_event_context *ctx)
2964 raw_spin_lock_init(&ctx->lock);
2965 mutex_init(&ctx->mutex);
2966 INIT_LIST_HEAD(&ctx->pinned_groups);
2967 INIT_LIST_HEAD(&ctx->flexible_groups);
2968 INIT_LIST_HEAD(&ctx->event_list);
2969 atomic_set(&ctx->refcount, 1);
2972 static struct perf_event_context *
2973 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2975 struct perf_event_context *ctx;
2977 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2981 __perf_event_init_context(ctx);
2984 get_task_struct(task);
2991 static struct task_struct *
2992 find_lively_task_by_vpid(pid_t vpid)
2994 struct task_struct *task;
3001 task = find_task_by_vpid(vpid);
3003 get_task_struct(task);
3007 return ERR_PTR(-ESRCH);
3009 /* Reuse ptrace permission checks for now. */
3011 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3016 put_task_struct(task);
3017 return ERR_PTR(err);
3022 * Returns a matching context with refcount and pincount.
3024 static struct perf_event_context *
3025 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3027 struct perf_event_context *ctx;
3028 struct perf_cpu_context *cpuctx;
3029 unsigned long flags;
3033 /* Must be root to operate on a CPU event: */
3034 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3035 return ERR_PTR(-EACCES);
3038 * We could be clever and allow to attach a event to an
3039 * offline CPU and activate it when the CPU comes up, but
3042 if (!cpu_online(cpu))
3043 return ERR_PTR(-ENODEV);
3045 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3054 ctxn = pmu->task_ctx_nr;
3059 ctx = perf_lock_task_context(task, ctxn, &flags);
3063 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3065 ctx = alloc_perf_context(pmu, task);
3071 mutex_lock(&task->perf_event_mutex);
3073 * If it has already passed perf_event_exit_task().
3074 * we must see PF_EXITING, it takes this mutex too.
3076 if (task->flags & PF_EXITING)
3078 else if (task->perf_event_ctxp[ctxn])
3083 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3085 mutex_unlock(&task->perf_event_mutex);
3087 if (unlikely(err)) {
3099 return ERR_PTR(err);
3102 static void perf_event_free_filter(struct perf_event *event);
3104 static void free_event_rcu(struct rcu_head *head)
3106 struct perf_event *event;
3108 event = container_of(head, struct perf_event, rcu_head);
3110 put_pid_ns(event->ns);
3111 perf_event_free_filter(event);
3115 static void ring_buffer_put(struct ring_buffer *rb);
3116 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3118 static void free_event(struct perf_event *event)
3120 irq_work_sync(&event->pending);
3122 if (!event->parent) {
3123 if (event->attach_state & PERF_ATTACH_TASK)
3124 static_key_slow_dec_deferred(&perf_sched_events);
3125 if (event->attr.mmap || event->attr.mmap_data)
3126 atomic_dec(&nr_mmap_events);
3127 if (event->attr.comm)
3128 atomic_dec(&nr_comm_events);
3129 if (event->attr.task)
3130 atomic_dec(&nr_task_events);
3131 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3132 put_callchain_buffers();
3133 if (is_cgroup_event(event)) {
3134 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
3135 static_key_slow_dec_deferred(&perf_sched_events);
3138 if (has_branch_stack(event)) {
3139 static_key_slow_dec_deferred(&perf_sched_events);
3140 /* is system-wide event */
3141 if (!(event->attach_state & PERF_ATTACH_TASK)) {
3142 atomic_dec(&per_cpu(perf_branch_stack_events,
3149 struct ring_buffer *rb;
3152 * Can happen when we close an event with re-directed output.
3154 * Since we have a 0 refcount, perf_mmap_close() will skip
3155 * over us; possibly making our ring_buffer_put() the last.
3157 mutex_lock(&event->mmap_mutex);
3160 rcu_assign_pointer(event->rb, NULL);
3161 ring_buffer_detach(event, rb);
3162 ring_buffer_put(rb); /* could be last */
3164 mutex_unlock(&event->mmap_mutex);
3167 if (is_cgroup_event(event))
3168 perf_detach_cgroup(event);
3171 event->destroy(event);
3174 put_ctx(event->ctx);
3176 call_rcu(&event->rcu_head, free_event_rcu);
3179 int perf_event_release_kernel(struct perf_event *event)
3181 struct perf_event_context *ctx = event->ctx;
3183 WARN_ON_ONCE(ctx->parent_ctx);
3185 * There are two ways this annotation is useful:
3187 * 1) there is a lock recursion from perf_event_exit_task
3188 * see the comment there.
3190 * 2) there is a lock-inversion with mmap_sem through
3191 * perf_event_read_group(), which takes faults while
3192 * holding ctx->mutex, however this is called after
3193 * the last filedesc died, so there is no possibility
3194 * to trigger the AB-BA case.
3196 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3197 raw_spin_lock_irq(&ctx->lock);
3198 perf_group_detach(event);
3199 raw_spin_unlock_irq(&ctx->lock);
3200 perf_remove_from_context(event);
3201 mutex_unlock(&ctx->mutex);
3207 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3210 * Called when the last reference to the file is gone.
3212 static void put_event(struct perf_event *event)
3214 struct task_struct *owner;
3216 if (!atomic_long_dec_and_test(&event->refcount))
3220 owner = ACCESS_ONCE(event->owner);
3222 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3223 * !owner it means the list deletion is complete and we can indeed
3224 * free this event, otherwise we need to serialize on
3225 * owner->perf_event_mutex.
3227 smp_read_barrier_depends();
3230 * Since delayed_put_task_struct() also drops the last
3231 * task reference we can safely take a new reference
3232 * while holding the rcu_read_lock().
3234 get_task_struct(owner);
3239 mutex_lock(&owner->perf_event_mutex);
3241 * We have to re-check the event->owner field, if it is cleared
3242 * we raced with perf_event_exit_task(), acquiring the mutex
3243 * ensured they're done, and we can proceed with freeing the
3247 list_del_init(&event->owner_entry);
3248 mutex_unlock(&owner->perf_event_mutex);
3249 put_task_struct(owner);
3252 perf_event_release_kernel(event);
3255 static int perf_release(struct inode *inode, struct file *file)
3257 put_event(file->private_data);
3261 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3263 struct perf_event *child;
3269 mutex_lock(&event->child_mutex);
3270 total += perf_event_read(event);
3271 *enabled += event->total_time_enabled +
3272 atomic64_read(&event->child_total_time_enabled);
3273 *running += event->total_time_running +
3274 atomic64_read(&event->child_total_time_running);
3276 list_for_each_entry(child, &event->child_list, child_list) {
3277 total += perf_event_read(child);
3278 *enabled += child->total_time_enabled;
3279 *running += child->total_time_running;
3281 mutex_unlock(&event->child_mutex);
3285 EXPORT_SYMBOL_GPL(perf_event_read_value);
3287 static int perf_event_read_group(struct perf_event *event,
3288 u64 read_format, char __user *buf)
3290 struct perf_event *leader = event->group_leader, *sub;
3291 int n = 0, size = 0, ret = -EFAULT;
3292 struct perf_event_context *ctx = leader->ctx;
3294 u64 count, enabled, running;
3296 mutex_lock(&ctx->mutex);
3297 count = perf_event_read_value(leader, &enabled, &running);
3299 values[n++] = 1 + leader->nr_siblings;
3300 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3301 values[n++] = enabled;
3302 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3303 values[n++] = running;
3304 values[n++] = count;
3305 if (read_format & PERF_FORMAT_ID)
3306 values[n++] = primary_event_id(leader);
3308 size = n * sizeof(u64);
3310 if (copy_to_user(buf, values, size))
3315 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3318 values[n++] = perf_event_read_value(sub, &enabled, &running);
3319 if (read_format & PERF_FORMAT_ID)
3320 values[n++] = primary_event_id(sub);
3322 size = n * sizeof(u64);
3324 if (copy_to_user(buf + ret, values, size)) {
3332 mutex_unlock(&ctx->mutex);
3337 static int perf_event_read_one(struct perf_event *event,
3338 u64 read_format, char __user *buf)
3340 u64 enabled, running;
3344 values[n++] = perf_event_read_value(event, &enabled, &running);
3345 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3346 values[n++] = enabled;
3347 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3348 values[n++] = running;
3349 if (read_format & PERF_FORMAT_ID)
3350 values[n++] = primary_event_id(event);
3352 if (copy_to_user(buf, values, n * sizeof(u64)))
3355 return n * sizeof(u64);
3359 * Read the performance event - simple non blocking version for now
3362 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3364 u64 read_format = event->attr.read_format;
3368 * Return end-of-file for a read on a event that is in
3369 * error state (i.e. because it was pinned but it couldn't be
3370 * scheduled on to the CPU at some point).
3372 if (event->state == PERF_EVENT_STATE_ERROR)
3375 if (count < event->read_size)
3378 WARN_ON_ONCE(event->ctx->parent_ctx);
3379 if (read_format & PERF_FORMAT_GROUP)
3380 ret = perf_event_read_group(event, read_format, buf);
3382 ret = perf_event_read_one(event, read_format, buf);
3388 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3390 struct perf_event *event = file->private_data;
3392 return perf_read_hw(event, buf, count);
3395 static unsigned int perf_poll(struct file *file, poll_table *wait)
3397 struct perf_event *event = file->private_data;
3398 struct ring_buffer *rb;
3399 unsigned int events = POLL_HUP;
3402 * Pin the event->rb by taking event->mmap_mutex; otherwise
3403 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3405 mutex_lock(&event->mmap_mutex);
3408 events = atomic_xchg(&rb->poll, 0);
3409 mutex_unlock(&event->mmap_mutex);
3411 poll_wait(file, &event->waitq, wait);
3416 static void perf_event_reset(struct perf_event *event)
3418 (void)perf_event_read(event);
3419 local64_set(&event->count, 0);
3420 perf_event_update_userpage(event);
3424 * Holding the top-level event's child_mutex means that any
3425 * descendant process that has inherited this event will block
3426 * in sync_child_event if it goes to exit, thus satisfying the
3427 * task existence requirements of perf_event_enable/disable.
3429 static void perf_event_for_each_child(struct perf_event *event,
3430 void (*func)(struct perf_event *))
3432 struct perf_event *child;
3434 WARN_ON_ONCE(event->ctx->parent_ctx);
3435 mutex_lock(&event->child_mutex);
3437 list_for_each_entry(child, &event->child_list, child_list)
3439 mutex_unlock(&event->child_mutex);
3442 static void perf_event_for_each(struct perf_event *event,
3443 void (*func)(struct perf_event *))
3445 struct perf_event_context *ctx = event->ctx;
3446 struct perf_event *sibling;
3448 WARN_ON_ONCE(ctx->parent_ctx);
3449 mutex_lock(&ctx->mutex);
3450 event = event->group_leader;
3452 perf_event_for_each_child(event, func);
3453 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3454 perf_event_for_each_child(sibling, func);
3455 mutex_unlock(&ctx->mutex);
3458 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3460 struct perf_event_context *ctx = event->ctx;
3464 if (!is_sampling_event(event))
3467 if (copy_from_user(&value, arg, sizeof(value)))
3473 raw_spin_lock_irq(&ctx->lock);
3474 if (event->attr.freq) {
3475 if (value > sysctl_perf_event_sample_rate) {
3480 event->attr.sample_freq = value;
3482 event->attr.sample_period = value;
3483 event->hw.sample_period = value;
3486 raw_spin_unlock_irq(&ctx->lock);
3491 static const struct file_operations perf_fops;
3493 static inline int perf_fget_light(int fd, struct fd *p)
3495 struct fd f = fdget(fd);
3499 if (f.file->f_op != &perf_fops) {
3507 static int perf_event_set_output(struct perf_event *event,
3508 struct perf_event *output_event);
3509 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3511 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3513 struct perf_event *event = file->private_data;
3514 void (*func)(struct perf_event *);
3518 case PERF_EVENT_IOC_ENABLE:
3519 func = perf_event_enable;
3521 case PERF_EVENT_IOC_DISABLE:
3522 func = perf_event_disable;
3524 case PERF_EVENT_IOC_RESET:
3525 func = perf_event_reset;
3528 case PERF_EVENT_IOC_REFRESH:
3529 return perf_event_refresh(event, arg);
3531 case PERF_EVENT_IOC_PERIOD:
3532 return perf_event_period(event, (u64 __user *)arg);
3534 case PERF_EVENT_IOC_SET_OUTPUT:
3538 struct perf_event *output_event;
3540 ret = perf_fget_light(arg, &output);
3543 output_event = output.file->private_data;
3544 ret = perf_event_set_output(event, output_event);
3547 ret = perf_event_set_output(event, NULL);
3552 case PERF_EVENT_IOC_SET_FILTER:
3553 return perf_event_set_filter(event, (void __user *)arg);
3559 if (flags & PERF_IOC_FLAG_GROUP)
3560 perf_event_for_each(event, func);
3562 perf_event_for_each_child(event, func);
3567 int perf_event_task_enable(void)
3569 struct perf_event *event;
3571 mutex_lock(¤t->perf_event_mutex);
3572 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3573 perf_event_for_each_child(event, perf_event_enable);
3574 mutex_unlock(¤t->perf_event_mutex);
3579 int perf_event_task_disable(void)
3581 struct perf_event *event;
3583 mutex_lock(¤t->perf_event_mutex);
3584 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3585 perf_event_for_each_child(event, perf_event_disable);
3586 mutex_unlock(¤t->perf_event_mutex);
3591 static int perf_event_index(struct perf_event *event)
3593 if (event->hw.state & PERF_HES_STOPPED)
3596 if (event->state != PERF_EVENT_STATE_ACTIVE)
3599 return event->pmu->event_idx(event);
3602 static void calc_timer_values(struct perf_event *event,
3609 *now = perf_clock();
3610 ctx_time = event->shadow_ctx_time + *now;
3611 *enabled = ctx_time - event->tstamp_enabled;
3612 *running = ctx_time - event->tstamp_running;
3615 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3620 * Callers need to ensure there can be no nesting of this function, otherwise
3621 * the seqlock logic goes bad. We can not serialize this because the arch
3622 * code calls this from NMI context.
3624 void perf_event_update_userpage(struct perf_event *event)
3626 struct perf_event_mmap_page *userpg;
3627 struct ring_buffer *rb;
3628 u64 enabled, running, now;
3632 * compute total_time_enabled, total_time_running
3633 * based on snapshot values taken when the event
3634 * was last scheduled in.
3636 * we cannot simply called update_context_time()
3637 * because of locking issue as we can be called in
3640 calc_timer_values(event, &now, &enabled, &running);
3641 rb = rcu_dereference(event->rb);
3645 userpg = rb->user_page;
3648 * Disable preemption so as to not let the corresponding user-space
3649 * spin too long if we get preempted.
3654 userpg->index = perf_event_index(event);
3655 userpg->offset = perf_event_count(event);
3657 userpg->offset -= local64_read(&event->hw.prev_count);
3659 userpg->time_enabled = enabled +
3660 atomic64_read(&event->child_total_time_enabled);
3662 userpg->time_running = running +
3663 atomic64_read(&event->child_total_time_running);
3665 arch_perf_update_userpage(userpg, now);
3674 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3676 struct perf_event *event = vma->vm_file->private_data;
3677 struct ring_buffer *rb;
3678 int ret = VM_FAULT_SIGBUS;
3680 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3681 if (vmf->pgoff == 0)
3687 rb = rcu_dereference(event->rb);
3691 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3694 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3698 get_page(vmf->page);
3699 vmf->page->mapping = vma->vm_file->f_mapping;
3700 vmf->page->index = vmf->pgoff;
3709 static void ring_buffer_attach(struct perf_event *event,
3710 struct ring_buffer *rb)
3712 unsigned long flags;
3714 if (!list_empty(&event->rb_entry))
3717 spin_lock_irqsave(&rb->event_lock, flags);
3718 if (list_empty(&event->rb_entry))
3719 list_add(&event->rb_entry, &rb->event_list);
3720 spin_unlock_irqrestore(&rb->event_lock, flags);
3723 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3725 unsigned long flags;
3727 if (list_empty(&event->rb_entry))
3730 spin_lock_irqsave(&rb->event_lock, flags);
3731 list_del_init(&event->rb_entry);
3732 wake_up_all(&event->waitq);
3733 spin_unlock_irqrestore(&rb->event_lock, flags);
3736 static void ring_buffer_wakeup(struct perf_event *event)
3738 struct ring_buffer *rb;
3741 rb = rcu_dereference(event->rb);
3743 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3744 wake_up_all(&event->waitq);
3749 static void rb_free_rcu(struct rcu_head *rcu_head)
3751 struct ring_buffer *rb;
3753 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3757 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3759 struct ring_buffer *rb;
3762 rb = rcu_dereference(event->rb);
3764 if (!atomic_inc_not_zero(&rb->refcount))
3772 static void ring_buffer_put(struct ring_buffer *rb)
3774 if (!atomic_dec_and_test(&rb->refcount))
3777 WARN_ON_ONCE(!list_empty(&rb->event_list));
3779 call_rcu(&rb->rcu_head, rb_free_rcu);
3782 static void perf_mmap_open(struct vm_area_struct *vma)
3784 struct perf_event *event = vma->vm_file->private_data;
3786 atomic_inc(&event->mmap_count);
3787 atomic_inc(&event->rb->mmap_count);
3791 * A buffer can be mmap()ed multiple times; either directly through the same
3792 * event, or through other events by use of perf_event_set_output().
3794 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3795 * the buffer here, where we still have a VM context. This means we need
3796 * to detach all events redirecting to us.
3798 static void perf_mmap_close(struct vm_area_struct *vma)
3800 struct perf_event *event = vma->vm_file->private_data;
3802 struct ring_buffer *rb = event->rb;
3803 struct user_struct *mmap_user = rb->mmap_user;
3804 int mmap_locked = rb->mmap_locked;
3805 unsigned long size = perf_data_size(rb);
3807 atomic_dec(&rb->mmap_count);
3809 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3812 /* Detach current event from the buffer. */
3813 rcu_assign_pointer(event->rb, NULL);
3814 ring_buffer_detach(event, rb);
3815 mutex_unlock(&event->mmap_mutex);
3817 /* If there's still other mmap()s of this buffer, we're done. */
3818 if (atomic_read(&rb->mmap_count)) {
3819 ring_buffer_put(rb); /* can't be last */
3824 * No other mmap()s, detach from all other events that might redirect
3825 * into the now unreachable buffer. Somewhat complicated by the
3826 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3830 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3831 if (!atomic_long_inc_not_zero(&event->refcount)) {
3833 * This event is en-route to free_event() which will
3834 * detach it and remove it from the list.
3840 mutex_lock(&event->mmap_mutex);
3842 * Check we didn't race with perf_event_set_output() which can
3843 * swizzle the rb from under us while we were waiting to
3844 * acquire mmap_mutex.
3846 * If we find a different rb; ignore this event, a next
3847 * iteration will no longer find it on the list. We have to
3848 * still restart the iteration to make sure we're not now
3849 * iterating the wrong list.
3851 if (event->rb == rb) {
3852 rcu_assign_pointer(event->rb, NULL);
3853 ring_buffer_detach(event, rb);
3854 ring_buffer_put(rb); /* can't be last, we still have one */
3856 mutex_unlock(&event->mmap_mutex);
3860 * Restart the iteration; either we're on the wrong list or
3861 * destroyed its integrity by doing a deletion.
3868 * It could be there's still a few 0-ref events on the list; they'll
3869 * get cleaned up by free_event() -- they'll also still have their
3870 * ref on the rb and will free it whenever they are done with it.
3872 * Aside from that, this buffer is 'fully' detached and unmapped,
3873 * undo the VM accounting.
3876 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3877 vma->vm_mm->pinned_vm -= mmap_locked;
3878 free_uid(mmap_user);
3880 ring_buffer_put(rb); /* could be last */
3883 static const struct vm_operations_struct perf_mmap_vmops = {
3884 .open = perf_mmap_open,
3885 .close = perf_mmap_close,
3886 .fault = perf_mmap_fault,
3887 .page_mkwrite = perf_mmap_fault,
3890 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3892 struct perf_event *event = file->private_data;
3893 unsigned long user_locked, user_lock_limit;
3894 struct user_struct *user = current_user();
3895 unsigned long locked, lock_limit;
3896 struct ring_buffer *rb;
3897 unsigned long vma_size;
3898 unsigned long nr_pages;
3899 long user_extra, extra;
3900 int ret = 0, flags = 0;
3903 * Don't allow mmap() of inherited per-task counters. This would
3904 * create a performance issue due to all children writing to the
3907 if (event->cpu == -1 && event->attr.inherit)
3910 if (!(vma->vm_flags & VM_SHARED))
3913 vma_size = vma->vm_end - vma->vm_start;
3914 nr_pages = (vma_size / PAGE_SIZE) - 1;
3917 * If we have rb pages ensure they're a power-of-two number, so we
3918 * can do bitmasks instead of modulo.
3920 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3923 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3926 if (vma->vm_pgoff != 0)
3929 WARN_ON_ONCE(event->ctx->parent_ctx);
3931 mutex_lock(&event->mmap_mutex);
3933 if (event->rb->nr_pages != nr_pages) {
3938 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
3940 * Raced against perf_mmap_close() through
3941 * perf_event_set_output(). Try again, hope for better
3944 mutex_unlock(&event->mmap_mutex);
3951 user_extra = nr_pages + 1;
3952 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3955 * Increase the limit linearly with more CPUs:
3957 user_lock_limit *= num_online_cpus();
3959 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3962 if (user_locked > user_lock_limit)
3963 extra = user_locked - user_lock_limit;
3965 lock_limit = rlimit(RLIMIT_MEMLOCK);
3966 lock_limit >>= PAGE_SHIFT;
3967 locked = vma->vm_mm->pinned_vm + extra;
3969 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3970 !capable(CAP_IPC_LOCK)) {
3977 if (vma->vm_flags & VM_WRITE)
3978 flags |= RING_BUFFER_WRITABLE;
3980 rb = rb_alloc(nr_pages,
3981 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3989 atomic_set(&rb->mmap_count, 1);
3990 rb->mmap_locked = extra;
3991 rb->mmap_user = get_current_user();
3993 atomic_long_add(user_extra, &user->locked_vm);
3994 vma->vm_mm->pinned_vm += extra;
3996 ring_buffer_attach(event, rb);
3997 rcu_assign_pointer(event->rb, rb);
3999 perf_event_update_userpage(event);
4003 atomic_inc(&event->mmap_count);
4004 mutex_unlock(&event->mmap_mutex);
4007 * Since pinned accounting is per vm we cannot allow fork() to copy our
4010 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4011 vma->vm_ops = &perf_mmap_vmops;
4016 static int perf_fasync(int fd, struct file *filp, int on)
4018 struct inode *inode = file_inode(filp);
4019 struct perf_event *event = filp->private_data;
4022 mutex_lock(&inode->i_mutex);
4023 retval = fasync_helper(fd, filp, on, &event->fasync);
4024 mutex_unlock(&inode->i_mutex);
4032 static const struct file_operations perf_fops = {
4033 .llseek = no_llseek,
4034 .release = perf_release,
4037 .unlocked_ioctl = perf_ioctl,
4038 .compat_ioctl = perf_ioctl,
4040 .fasync = perf_fasync,
4046 * If there's data, ensure we set the poll() state and publish everything
4047 * to user-space before waking everybody up.
4050 void perf_event_wakeup(struct perf_event *event)
4052 ring_buffer_wakeup(event);
4054 if (event->pending_kill) {
4055 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4056 event->pending_kill = 0;
4060 static void perf_pending_event(struct irq_work *entry)
4062 struct perf_event *event = container_of(entry,
4063 struct perf_event, pending);
4065 if (event->pending_disable) {
4066 event->pending_disable = 0;
4067 __perf_event_disable(event);
4070 if (event->pending_wakeup) {
4071 event->pending_wakeup = 0;
4072 perf_event_wakeup(event);
4077 * We assume there is only KVM supporting the callbacks.
4078 * Later on, we might change it to a list if there is
4079 * another virtualization implementation supporting the callbacks.
4081 struct perf_guest_info_callbacks *perf_guest_cbs;
4083 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4085 perf_guest_cbs = cbs;
4088 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4090 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4092 perf_guest_cbs = NULL;
4095 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4098 perf_output_sample_regs(struct perf_output_handle *handle,
4099 struct pt_regs *regs, u64 mask)
4103 for_each_set_bit(bit, (const unsigned long *) &mask,
4104 sizeof(mask) * BITS_PER_BYTE) {
4107 val = perf_reg_value(regs, bit);
4108 perf_output_put(handle, val);
4112 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4113 struct pt_regs *regs)
4115 if (!user_mode(regs)) {
4117 regs = task_pt_regs(current);
4123 regs_user->regs = regs;
4124 regs_user->abi = perf_reg_abi(current);
4129 * Get remaining task size from user stack pointer.
4131 * It'd be better to take stack vma map and limit this more
4132 * precisly, but there's no way to get it safely under interrupt,
4133 * so using TASK_SIZE as limit.
4135 static u64 perf_ustack_task_size(struct pt_regs *regs)
4137 unsigned long addr = perf_user_stack_pointer(regs);
4139 if (!addr || addr >= TASK_SIZE)
4142 return TASK_SIZE - addr;
4146 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4147 struct pt_regs *regs)
4151 /* No regs, no stack pointer, no dump. */
4156 * Check if we fit in with the requested stack size into the:
4158 * If we don't, we limit the size to the TASK_SIZE.
4160 * - remaining sample size
4161 * If we don't, we customize the stack size to
4162 * fit in to the remaining sample size.
4165 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4166 stack_size = min(stack_size, (u16) task_size);
4168 /* Current header size plus static size and dynamic size. */
4169 header_size += 2 * sizeof(u64);
4171 /* Do we fit in with the current stack dump size? */
4172 if ((u16) (header_size + stack_size) < header_size) {
4174 * If we overflow the maximum size for the sample,
4175 * we customize the stack dump size to fit in.
4177 stack_size = USHRT_MAX - header_size - sizeof(u64);
4178 stack_size = round_up(stack_size, sizeof(u64));
4185 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4186 struct pt_regs *regs)
4188 /* Case of a kernel thread, nothing to dump */
4191 perf_output_put(handle, size);
4200 * - the size requested by user or the best one we can fit
4201 * in to the sample max size
4203 * - user stack dump data
4205 * - the actual dumped size
4209 perf_output_put(handle, dump_size);
4212 sp = perf_user_stack_pointer(regs);
4213 rem = __output_copy_user(handle, (void *) sp, dump_size);
4214 dyn_size = dump_size - rem;
4216 perf_output_skip(handle, rem);
4219 perf_output_put(handle, dyn_size);
4223 static void __perf_event_header__init_id(struct perf_event_header *header,
4224 struct perf_sample_data *data,
4225 struct perf_event *event)
4227 u64 sample_type = event->attr.sample_type;
4229 data->type = sample_type;
4230 header->size += event->id_header_size;
4232 if (sample_type & PERF_SAMPLE_TID) {
4233 /* namespace issues */
4234 data->tid_entry.pid = perf_event_pid(event, current);
4235 data->tid_entry.tid = perf_event_tid(event, current);
4238 if (sample_type & PERF_SAMPLE_TIME)
4239 data->time = perf_clock();
4241 if (sample_type & PERF_SAMPLE_ID)
4242 data->id = primary_event_id(event);
4244 if (sample_type & PERF_SAMPLE_STREAM_ID)
4245 data->stream_id = event->id;
4247 if (sample_type & PERF_SAMPLE_CPU) {
4248 data->cpu_entry.cpu = raw_smp_processor_id();
4249 data->cpu_entry.reserved = 0;
4253 void perf_event_header__init_id(struct perf_event_header *header,
4254 struct perf_sample_data *data,
4255 struct perf_event *event)
4257 if (event->attr.sample_id_all)
4258 __perf_event_header__init_id(header, data, event);
4261 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4262 struct perf_sample_data *data)
4264 u64 sample_type = data->type;
4266 if (sample_type & PERF_SAMPLE_TID)
4267 perf_output_put(handle, data->tid_entry);
4269 if (sample_type & PERF_SAMPLE_TIME)
4270 perf_output_put(handle, data->time);
4272 if (sample_type & PERF_SAMPLE_ID)
4273 perf_output_put(handle, data->id);
4275 if (sample_type & PERF_SAMPLE_STREAM_ID)
4276 perf_output_put(handle, data->stream_id);
4278 if (sample_type & PERF_SAMPLE_CPU)
4279 perf_output_put(handle, data->cpu_entry);
4282 void perf_event__output_id_sample(struct perf_event *event,
4283 struct perf_output_handle *handle,
4284 struct perf_sample_data *sample)
4286 if (event->attr.sample_id_all)
4287 __perf_event__output_id_sample(handle, sample);
4290 static void perf_output_read_one(struct perf_output_handle *handle,
4291 struct perf_event *event,
4292 u64 enabled, u64 running)
4294 u64 read_format = event->attr.read_format;
4298 values[n++] = perf_event_count(event);
4299 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4300 values[n++] = enabled +
4301 atomic64_read(&event->child_total_time_enabled);
4303 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4304 values[n++] = running +
4305 atomic64_read(&event->child_total_time_running);
4307 if (read_format & PERF_FORMAT_ID)
4308 values[n++] = primary_event_id(event);
4310 __output_copy(handle, values, n * sizeof(u64));
4314 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4316 static void perf_output_read_group(struct perf_output_handle *handle,
4317 struct perf_event *event,
4318 u64 enabled, u64 running)
4320 struct perf_event *leader = event->group_leader, *sub;
4321 u64 read_format = event->attr.read_format;
4325 values[n++] = 1 + leader->nr_siblings;
4327 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4328 values[n++] = enabled;
4330 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4331 values[n++] = running;
4333 if (leader != event)
4334 leader->pmu->read(leader);
4336 values[n++] = perf_event_count(leader);
4337 if (read_format & PERF_FORMAT_ID)
4338 values[n++] = primary_event_id(leader);
4340 __output_copy(handle, values, n * sizeof(u64));
4342 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4346 sub->pmu->read(sub);
4348 values[n++] = perf_event_count(sub);
4349 if (read_format & PERF_FORMAT_ID)
4350 values[n++] = primary_event_id(sub);
4352 __output_copy(handle, values, n * sizeof(u64));
4356 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4357 PERF_FORMAT_TOTAL_TIME_RUNNING)
4359 static void perf_output_read(struct perf_output_handle *handle,
4360 struct perf_event *event)
4362 u64 enabled = 0, running = 0, now;
4363 u64 read_format = event->attr.read_format;
4366 * compute total_time_enabled, total_time_running
4367 * based on snapshot values taken when the event
4368 * was last scheduled in.
4370 * we cannot simply called update_context_time()
4371 * because of locking issue as we are called in
4374 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4375 calc_timer_values(event, &now, &enabled, &running);
4377 if (event->attr.read_format & PERF_FORMAT_GROUP)
4378 perf_output_read_group(handle, event, enabled, running);
4380 perf_output_read_one(handle, event, enabled, running);
4383 void perf_output_sample(struct perf_output_handle *handle,
4384 struct perf_event_header *header,
4385 struct perf_sample_data *data,
4386 struct perf_event *event)
4388 u64 sample_type = data->type;
4390 perf_output_put(handle, *header);
4392 if (sample_type & PERF_SAMPLE_IP)
4393 perf_output_put(handle, data->ip);
4395 if (sample_type & PERF_SAMPLE_TID)
4396 perf_output_put(handle, data->tid_entry);
4398 if (sample_type & PERF_SAMPLE_TIME)
4399 perf_output_put(handle, data->time);
4401 if (sample_type & PERF_SAMPLE_ADDR)
4402 perf_output_put(handle, data->addr);
4404 if (sample_type & PERF_SAMPLE_ID)
4405 perf_output_put(handle, data->id);
4407 if (sample_type & PERF_SAMPLE_STREAM_ID)
4408 perf_output_put(handle, data->stream_id);
4410 if (sample_type & PERF_SAMPLE_CPU)
4411 perf_output_put(handle, data->cpu_entry);
4413 if (sample_type & PERF_SAMPLE_PERIOD)
4414 perf_output_put(handle, data->period);
4416 if (sample_type & PERF_SAMPLE_READ)
4417 perf_output_read(handle, event);
4419 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4420 if (data->callchain) {
4423 if (data->callchain)
4424 size += data->callchain->nr;
4426 size *= sizeof(u64);
4428 __output_copy(handle, data->callchain, size);
4431 perf_output_put(handle, nr);
4435 if (sample_type & PERF_SAMPLE_RAW) {
4437 perf_output_put(handle, data->raw->size);
4438 __output_copy(handle, data->raw->data,
4445 .size = sizeof(u32),
4448 perf_output_put(handle, raw);
4452 if (!event->attr.watermark) {
4453 int wakeup_events = event->attr.wakeup_events;
4455 if (wakeup_events) {
4456 struct ring_buffer *rb = handle->rb;
4457 int events = local_inc_return(&rb->events);
4459 if (events >= wakeup_events) {
4460 local_sub(wakeup_events, &rb->events);
4461 local_inc(&rb->wakeup);
4466 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4467 if (data->br_stack) {
4470 size = data->br_stack->nr
4471 * sizeof(struct perf_branch_entry);
4473 perf_output_put(handle, data->br_stack->nr);
4474 perf_output_copy(handle, data->br_stack->entries, size);
4477 * we always store at least the value of nr
4480 perf_output_put(handle, nr);
4484 if (sample_type & PERF_SAMPLE_REGS_USER) {
4485 u64 abi = data->regs_user.abi;
4488 * If there are no regs to dump, notice it through
4489 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4491 perf_output_put(handle, abi);
4494 u64 mask = event->attr.sample_regs_user;
4495 perf_output_sample_regs(handle,
4496 data->regs_user.regs,
4501 if (sample_type & PERF_SAMPLE_STACK_USER)
4502 perf_output_sample_ustack(handle,
4503 data->stack_user_size,
4504 data->regs_user.regs);
4506 if (sample_type & PERF_SAMPLE_WEIGHT)
4507 perf_output_put(handle, data->weight);
4509 if (sample_type & PERF_SAMPLE_DATA_SRC)
4510 perf_output_put(handle, data->data_src.val);
4513 void perf_prepare_sample(struct perf_event_header *header,
4514 struct perf_sample_data *data,
4515 struct perf_event *event,
4516 struct pt_regs *regs)
4518 u64 sample_type = event->attr.sample_type;
4520 header->type = PERF_RECORD_SAMPLE;
4521 header->size = sizeof(*header) + event->header_size;
4524 header->misc |= perf_misc_flags(regs);
4526 __perf_event_header__init_id(header, data, event);
4528 if (sample_type & PERF_SAMPLE_IP)
4529 data->ip = perf_instruction_pointer(regs);
4531 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4534 data->callchain = perf_callchain(event, regs);
4536 if (data->callchain)
4537 size += data->callchain->nr;
4539 header->size += size * sizeof(u64);
4542 if (sample_type & PERF_SAMPLE_RAW) {
4543 int size = sizeof(u32);
4546 size += data->raw->size;
4548 size += sizeof(u32);
4550 WARN_ON_ONCE(size & (sizeof(u64)-1));
4551 header->size += size;
4554 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4555 int size = sizeof(u64); /* nr */
4556 if (data->br_stack) {
4557 size += data->br_stack->nr
4558 * sizeof(struct perf_branch_entry);
4560 header->size += size;
4563 if (sample_type & PERF_SAMPLE_REGS_USER) {
4564 /* regs dump ABI info */
4565 int size = sizeof(u64);
4567 perf_sample_regs_user(&data->regs_user, regs);
4569 if (data->regs_user.regs) {
4570 u64 mask = event->attr.sample_regs_user;
4571 size += hweight64(mask) * sizeof(u64);
4574 header->size += size;
4577 if (sample_type & PERF_SAMPLE_STACK_USER) {
4579 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4580 * processed as the last one or have additional check added
4581 * in case new sample type is added, because we could eat
4582 * up the rest of the sample size.
4584 struct perf_regs_user *uregs = &data->regs_user;
4585 u16 stack_size = event->attr.sample_stack_user;
4586 u16 size = sizeof(u64);
4589 perf_sample_regs_user(uregs, regs);
4591 stack_size = perf_sample_ustack_size(stack_size, header->size,
4595 * If there is something to dump, add space for the dump
4596 * itself and for the field that tells the dynamic size,
4597 * which is how many have been actually dumped.
4600 size += sizeof(u64) + stack_size;
4602 data->stack_user_size = stack_size;
4603 header->size += size;
4607 static void perf_event_output(struct perf_event *event,
4608 struct perf_sample_data *data,
4609 struct pt_regs *regs)
4611 struct perf_output_handle handle;
4612 struct perf_event_header header;
4614 /* protect the callchain buffers */
4617 perf_prepare_sample(&header, data, event, regs);
4619 if (perf_output_begin(&handle, event, header.size))
4622 perf_output_sample(&handle, &header, data, event);
4624 perf_output_end(&handle);
4634 struct perf_read_event {
4635 struct perf_event_header header;
4642 perf_event_read_event(struct perf_event *event,
4643 struct task_struct *task)
4645 struct perf_output_handle handle;
4646 struct perf_sample_data sample;
4647 struct perf_read_event read_event = {
4649 .type = PERF_RECORD_READ,
4651 .size = sizeof(read_event) + event->read_size,
4653 .pid = perf_event_pid(event, task),
4654 .tid = perf_event_tid(event, task),
4658 perf_event_header__init_id(&read_event.header, &sample, event);
4659 ret = perf_output_begin(&handle, event, read_event.header.size);
4663 perf_output_put(&handle, read_event);
4664 perf_output_read(&handle, event);
4665 perf_event__output_id_sample(event, &handle, &sample);
4667 perf_output_end(&handle);
4670 typedef int (perf_event_aux_match_cb)(struct perf_event *event, void *data);
4671 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4674 perf_event_aux_ctx(struct perf_event_context *ctx,
4675 perf_event_aux_match_cb match,
4676 perf_event_aux_output_cb output,
4679 struct perf_event *event;
4681 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4682 if (event->state < PERF_EVENT_STATE_INACTIVE)
4684 if (!event_filter_match(event))
4686 if (match(event, data))
4687 output(event, data);
4692 perf_event_aux(perf_event_aux_match_cb match,
4693 perf_event_aux_output_cb output,
4695 struct perf_event_context *task_ctx)
4697 struct perf_cpu_context *cpuctx;
4698 struct perf_event_context *ctx;
4703 list_for_each_entry_rcu(pmu, &pmus, entry) {
4704 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4705 if (cpuctx->unique_pmu != pmu)
4707 perf_event_aux_ctx(&cpuctx->ctx, match, output, data);
4710 ctxn = pmu->task_ctx_nr;
4713 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4715 perf_event_aux_ctx(ctx, match, output, data);
4717 put_cpu_ptr(pmu->pmu_cpu_context);
4722 perf_event_aux_ctx(task_ctx, match, output, data);
4729 * task tracking -- fork/exit
4731 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4734 struct perf_task_event {
4735 struct task_struct *task;
4736 struct perf_event_context *task_ctx;
4739 struct perf_event_header header;
4749 static void perf_event_task_output(struct perf_event *event,
4752 struct perf_task_event *task_event = data;
4753 struct perf_output_handle handle;
4754 struct perf_sample_data sample;
4755 struct task_struct *task = task_event->task;
4756 int ret, size = task_event->event_id.header.size;
4758 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4760 ret = perf_output_begin(&handle, event,
4761 task_event->event_id.header.size);
4765 task_event->event_id.pid = perf_event_pid(event, task);
4766 task_event->event_id.ppid = perf_event_pid(event, current);
4768 task_event->event_id.tid = perf_event_tid(event, task);
4769 task_event->event_id.ptid = perf_event_tid(event, current);
4771 perf_output_put(&handle, task_event->event_id);
4773 perf_event__output_id_sample(event, &handle, &sample);
4775 perf_output_end(&handle);
4777 task_event->event_id.header.size = size;
4780 static int perf_event_task_match(struct perf_event *event,
4781 void *data __maybe_unused)
4783 return event->attr.comm || event->attr.mmap ||
4784 event->attr.mmap_data || event->attr.task;
4787 static void perf_event_task(struct task_struct *task,
4788 struct perf_event_context *task_ctx,
4791 struct perf_task_event task_event;
4793 if (!atomic_read(&nr_comm_events) &&
4794 !atomic_read(&nr_mmap_events) &&
4795 !atomic_read(&nr_task_events))
4798 task_event = (struct perf_task_event){
4800 .task_ctx = task_ctx,
4803 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4805 .size = sizeof(task_event.event_id),
4811 .time = perf_clock(),
4815 perf_event_aux(perf_event_task_match,
4816 perf_event_task_output,
4821 void perf_event_fork(struct task_struct *task)
4823 perf_event_task(task, NULL, 1);
4830 struct perf_comm_event {
4831 struct task_struct *task;
4836 struct perf_event_header header;
4843 static void perf_event_comm_output(struct perf_event *event,
4846 struct perf_comm_event *comm_event = data;
4847 struct perf_output_handle handle;
4848 struct perf_sample_data sample;
4849 int size = comm_event->event_id.header.size;
4852 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4853 ret = perf_output_begin(&handle, event,
4854 comm_event->event_id.header.size);
4859 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4860 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4862 perf_output_put(&handle, comm_event->event_id);
4863 __output_copy(&handle, comm_event->comm,
4864 comm_event->comm_size);
4866 perf_event__output_id_sample(event, &handle, &sample);
4868 perf_output_end(&handle);
4870 comm_event->event_id.header.size = size;
4873 static int perf_event_comm_match(struct perf_event *event,
4874 void *data __maybe_unused)
4876 return event->attr.comm;
4879 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4881 char comm[TASK_COMM_LEN];
4884 memset(comm, 0, sizeof(comm));
4885 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4886 size = ALIGN(strlen(comm)+1, sizeof(u64));
4888 comm_event->comm = comm;
4889 comm_event->comm_size = size;
4891 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4893 perf_event_aux(perf_event_comm_match,
4894 perf_event_comm_output,
4899 void perf_event_comm(struct task_struct *task)
4901 struct perf_comm_event comm_event;
4902 struct perf_event_context *ctx;
4906 for_each_task_context_nr(ctxn) {
4907 ctx = task->perf_event_ctxp[ctxn];
4911 perf_event_enable_on_exec(ctx);
4915 if (!atomic_read(&nr_comm_events))
4918 comm_event = (struct perf_comm_event){
4924 .type = PERF_RECORD_COMM,
4933 perf_event_comm_event(&comm_event);
4940 struct perf_mmap_event {
4941 struct vm_area_struct *vma;
4943 const char *file_name;
4947 struct perf_event_header header;
4957 static void perf_event_mmap_output(struct perf_event *event,
4960 struct perf_mmap_event *mmap_event = data;
4961 struct perf_output_handle handle;
4962 struct perf_sample_data sample;
4963 int size = mmap_event->event_id.header.size;
4966 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4967 ret = perf_output_begin(&handle, event,
4968 mmap_event->event_id.header.size);
4972 mmap_event->event_id.pid = perf_event_pid(event, current);
4973 mmap_event->event_id.tid = perf_event_tid(event, current);
4975 perf_output_put(&handle, mmap_event->event_id);
4976 __output_copy(&handle, mmap_event->file_name,
4977 mmap_event->file_size);
4979 perf_event__output_id_sample(event, &handle, &sample);
4981 perf_output_end(&handle);
4983 mmap_event->event_id.header.size = size;
4986 static int perf_event_mmap_match(struct perf_event *event,
4989 struct perf_mmap_event *mmap_event = data;
4990 struct vm_area_struct *vma = mmap_event->vma;
4991 int executable = vma->vm_flags & VM_EXEC;
4993 return (!executable && event->attr.mmap_data) ||
4994 (executable && event->attr.mmap);
4997 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4999 struct vm_area_struct *vma = mmap_event->vma;
5000 struct file *file = vma->vm_file;
5006 memset(tmp, 0, sizeof(tmp));
5010 * d_path works from the end of the rb backwards, so we
5011 * need to add enough zero bytes after the string to handle
5012 * the 64bit alignment we do later.
5014 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
5016 name = strncpy(tmp, "//enomem", sizeof(tmp));
5019 name = d_path(&file->f_path, buf, PATH_MAX);
5021 name = strncpy(tmp, "//toolong", sizeof(tmp));
5025 if (arch_vma_name(mmap_event->vma)) {
5026 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
5028 tmp[sizeof(tmp) - 1] = '\0';
5033 name = strncpy(tmp, "[vdso]", sizeof(tmp));
5035 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
5036 vma->vm_end >= vma->vm_mm->brk) {
5037 name = strncpy(tmp, "[heap]", sizeof(tmp));
5039 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
5040 vma->vm_end >= vma->vm_mm->start_stack) {
5041 name = strncpy(tmp, "[stack]", sizeof(tmp));
5045 name = strncpy(tmp, "//anon", sizeof(tmp));
5050 size = ALIGN(strlen(name)+1, sizeof(u64));
5052 mmap_event->file_name = name;
5053 mmap_event->file_size = size;
5055 if (!(vma->vm_flags & VM_EXEC))
5056 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5058 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5060 perf_event_aux(perf_event_mmap_match,
5061 perf_event_mmap_output,
5068 void perf_event_mmap(struct vm_area_struct *vma)
5070 struct perf_mmap_event mmap_event;
5072 if (!atomic_read(&nr_mmap_events))
5075 mmap_event = (struct perf_mmap_event){
5081 .type = PERF_RECORD_MMAP,
5082 .misc = PERF_RECORD_MISC_USER,
5087 .start = vma->vm_start,
5088 .len = vma->vm_end - vma->vm_start,
5089 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5093 perf_event_mmap_event(&mmap_event);
5097 * IRQ throttle logging
5100 static void perf_log_throttle(struct perf_event *event, int enable)
5102 struct perf_output_handle handle;
5103 struct perf_sample_data sample;
5107 struct perf_event_header header;
5111 } throttle_event = {
5113 .type = PERF_RECORD_THROTTLE,
5115 .size = sizeof(throttle_event),
5117 .time = perf_clock(),
5118 .id = primary_event_id(event),
5119 .stream_id = event->id,
5123 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5125 perf_event_header__init_id(&throttle_event.header, &sample, event);
5127 ret = perf_output_begin(&handle, event,
5128 throttle_event.header.size);
5132 perf_output_put(&handle, throttle_event);
5133 perf_event__output_id_sample(event, &handle, &sample);
5134 perf_output_end(&handle);
5138 * Generic event overflow handling, sampling.
5141 static int __perf_event_overflow(struct perf_event *event,
5142 int throttle, struct perf_sample_data *data,
5143 struct pt_regs *regs)
5145 int events = atomic_read(&event->event_limit);
5146 struct hw_perf_event *hwc = &event->hw;
5151 * Non-sampling counters might still use the PMI to fold short
5152 * hardware counters, ignore those.
5154 if (unlikely(!is_sampling_event(event)))
5157 seq = __this_cpu_read(perf_throttled_seq);
5158 if (seq != hwc->interrupts_seq) {
5159 hwc->interrupts_seq = seq;
5160 hwc->interrupts = 1;
5163 if (unlikely(throttle
5164 && hwc->interrupts >= max_samples_per_tick)) {
5165 __this_cpu_inc(perf_throttled_count);
5166 hwc->interrupts = MAX_INTERRUPTS;
5167 perf_log_throttle(event, 0);
5172 if (event->attr.freq) {
5173 u64 now = perf_clock();
5174 s64 delta = now - hwc->freq_time_stamp;
5176 hwc->freq_time_stamp = now;
5178 if (delta > 0 && delta < 2*TICK_NSEC)
5179 perf_adjust_period(event, delta, hwc->last_period, true);
5183 * XXX event_limit might not quite work as expected on inherited
5187 event->pending_kill = POLL_IN;
5188 if (events && atomic_dec_and_test(&event->event_limit)) {
5190 event->pending_kill = POLL_HUP;
5191 event->pending_disable = 1;
5192 irq_work_queue(&event->pending);
5195 if (event->overflow_handler)
5196 event->overflow_handler(event, data, regs);
5198 perf_event_output(event, data, regs);
5200 if (event->fasync && event->pending_kill) {
5201 event->pending_wakeup = 1;
5202 irq_work_queue(&event->pending);
5208 int perf_event_overflow(struct perf_event *event,
5209 struct perf_sample_data *data,
5210 struct pt_regs *regs)
5212 return __perf_event_overflow(event, 1, data, regs);
5216 * Generic software event infrastructure
5219 struct swevent_htable {
5220 struct swevent_hlist *swevent_hlist;
5221 struct mutex hlist_mutex;
5224 /* Recursion avoidance in each contexts */
5225 int recursion[PERF_NR_CONTEXTS];
5228 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5231 * We directly increment event->count and keep a second value in
5232 * event->hw.period_left to count intervals. This period event
5233 * is kept in the range [-sample_period, 0] so that we can use the
5237 u64 perf_swevent_set_period(struct perf_event *event)
5239 struct hw_perf_event *hwc = &event->hw;
5240 u64 period = hwc->last_period;
5244 hwc->last_period = hwc->sample_period;
5247 old = val = local64_read(&hwc->period_left);
5251 nr = div64_u64(period + val, period);
5252 offset = nr * period;
5254 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5260 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5261 struct perf_sample_data *data,
5262 struct pt_regs *regs)
5264 struct hw_perf_event *hwc = &event->hw;
5268 overflow = perf_swevent_set_period(event);
5270 if (hwc->interrupts == MAX_INTERRUPTS)
5273 for (; overflow; overflow--) {
5274 if (__perf_event_overflow(event, throttle,
5277 * We inhibit the overflow from happening when
5278 * hwc->interrupts == MAX_INTERRUPTS.
5286 static void perf_swevent_event(struct perf_event *event, u64 nr,
5287 struct perf_sample_data *data,
5288 struct pt_regs *regs)
5290 struct hw_perf_event *hwc = &event->hw;
5292 local64_add(nr, &event->count);
5297 if (!is_sampling_event(event))
5300 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5302 return perf_swevent_overflow(event, 1, data, regs);
5304 data->period = event->hw.last_period;
5306 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5307 return perf_swevent_overflow(event, 1, data, regs);
5309 if (local64_add_negative(nr, &hwc->period_left))
5312 perf_swevent_overflow(event, 0, data, regs);
5315 static int perf_exclude_event(struct perf_event *event,
5316 struct pt_regs *regs)
5318 if (event->hw.state & PERF_HES_STOPPED)
5322 if (event->attr.exclude_user && user_mode(regs))
5325 if (event->attr.exclude_kernel && !user_mode(regs))
5332 static int perf_swevent_match(struct perf_event *event,
5333 enum perf_type_id type,
5335 struct perf_sample_data *data,
5336 struct pt_regs *regs)
5338 if (event->attr.type != type)
5341 if (event->attr.config != event_id)
5344 if (perf_exclude_event(event, regs))
5350 static inline u64 swevent_hash(u64 type, u32 event_id)
5352 u64 val = event_id | (type << 32);
5354 return hash_64(val, SWEVENT_HLIST_BITS);
5357 static inline struct hlist_head *
5358 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5360 u64 hash = swevent_hash(type, event_id);
5362 return &hlist->heads[hash];
5365 /* For the read side: events when they trigger */
5366 static inline struct hlist_head *
5367 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5369 struct swevent_hlist *hlist;
5371 hlist = rcu_dereference(swhash->swevent_hlist);
5375 return __find_swevent_head(hlist, type, event_id);
5378 /* For the event head insertion and removal in the hlist */
5379 static inline struct hlist_head *
5380 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5382 struct swevent_hlist *hlist;
5383 u32 event_id = event->attr.config;
5384 u64 type = event->attr.type;
5387 * Event scheduling is always serialized against hlist allocation
5388 * and release. Which makes the protected version suitable here.
5389 * The context lock guarantees that.
5391 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5392 lockdep_is_held(&event->ctx->lock));
5396 return __find_swevent_head(hlist, type, event_id);
5399 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5401 struct perf_sample_data *data,
5402 struct pt_regs *regs)
5404 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5405 struct perf_event *event;
5406 struct hlist_head *head;
5409 head = find_swevent_head_rcu(swhash, type, event_id);
5413 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5414 if (perf_swevent_match(event, type, event_id, data, regs))
5415 perf_swevent_event(event, nr, data, regs);
5421 int perf_swevent_get_recursion_context(void)
5423 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5425 return get_recursion_context(swhash->recursion);
5427 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5429 inline void perf_swevent_put_recursion_context(int rctx)
5431 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5433 put_recursion_context(swhash->recursion, rctx);
5436 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5438 struct perf_sample_data data;
5441 preempt_disable_notrace();
5442 rctx = perf_swevent_get_recursion_context();
5446 perf_sample_data_init(&data, addr, 0);
5448 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5450 perf_swevent_put_recursion_context(rctx);
5451 preempt_enable_notrace();
5454 static void perf_swevent_read(struct perf_event *event)
5458 static int perf_swevent_add(struct perf_event *event, int flags)
5460 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5461 struct hw_perf_event *hwc = &event->hw;
5462 struct hlist_head *head;
5464 if (is_sampling_event(event)) {
5465 hwc->last_period = hwc->sample_period;
5466 perf_swevent_set_period(event);
5469 hwc->state = !(flags & PERF_EF_START);
5471 head = find_swevent_head(swhash, event);
5472 if (WARN_ON_ONCE(!head))
5475 hlist_add_head_rcu(&event->hlist_entry, head);
5480 static void perf_swevent_del(struct perf_event *event, int flags)
5482 hlist_del_rcu(&event->hlist_entry);
5485 static void perf_swevent_start(struct perf_event *event, int flags)
5487 event->hw.state = 0;
5490 static void perf_swevent_stop(struct perf_event *event, int flags)
5492 event->hw.state = PERF_HES_STOPPED;
5495 /* Deref the hlist from the update side */
5496 static inline struct swevent_hlist *
5497 swevent_hlist_deref(struct swevent_htable *swhash)
5499 return rcu_dereference_protected(swhash->swevent_hlist,
5500 lockdep_is_held(&swhash->hlist_mutex));
5503 static void swevent_hlist_release(struct swevent_htable *swhash)
5505 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5510 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5511 kfree_rcu(hlist, rcu_head);
5514 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5516 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5518 mutex_lock(&swhash->hlist_mutex);
5520 if (!--swhash->hlist_refcount)
5521 swevent_hlist_release(swhash);
5523 mutex_unlock(&swhash->hlist_mutex);
5526 static void swevent_hlist_put(struct perf_event *event)
5530 if (event->cpu != -1) {
5531 swevent_hlist_put_cpu(event, event->cpu);
5535 for_each_possible_cpu(cpu)
5536 swevent_hlist_put_cpu(event, cpu);
5539 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5541 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5544 mutex_lock(&swhash->hlist_mutex);
5546 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5547 struct swevent_hlist *hlist;
5549 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5554 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5556 swhash->hlist_refcount++;
5558 mutex_unlock(&swhash->hlist_mutex);
5563 static int swevent_hlist_get(struct perf_event *event)
5566 int cpu, failed_cpu;
5568 if (event->cpu != -1)
5569 return swevent_hlist_get_cpu(event, event->cpu);
5572 for_each_possible_cpu(cpu) {
5573 err = swevent_hlist_get_cpu(event, cpu);
5583 for_each_possible_cpu(cpu) {
5584 if (cpu == failed_cpu)
5586 swevent_hlist_put_cpu(event, cpu);
5593 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5595 static void sw_perf_event_destroy(struct perf_event *event)
5597 u64 event_id = event->attr.config;
5599 WARN_ON(event->parent);
5601 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5602 swevent_hlist_put(event);
5605 static int perf_swevent_init(struct perf_event *event)
5607 u64 event_id = event->attr.config;
5609 if (event->attr.type != PERF_TYPE_SOFTWARE)
5613 * no branch sampling for software events
5615 if (has_branch_stack(event))
5619 case PERF_COUNT_SW_CPU_CLOCK:
5620 case PERF_COUNT_SW_TASK_CLOCK:
5627 if (event_id >= PERF_COUNT_SW_MAX)
5630 if (!event->parent) {
5633 err = swevent_hlist_get(event);
5637 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5638 event->destroy = sw_perf_event_destroy;
5644 static int perf_swevent_event_idx(struct perf_event *event)
5649 static struct pmu perf_swevent = {
5650 .task_ctx_nr = perf_sw_context,
5652 .event_init = perf_swevent_init,
5653 .add = perf_swevent_add,
5654 .del = perf_swevent_del,
5655 .start = perf_swevent_start,
5656 .stop = perf_swevent_stop,
5657 .read = perf_swevent_read,
5659 .event_idx = perf_swevent_event_idx,
5662 #ifdef CONFIG_EVENT_TRACING
5664 static int perf_tp_filter_match(struct perf_event *event,
5665 struct perf_sample_data *data)
5667 void *record = data->raw->data;
5669 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5674 static int perf_tp_event_match(struct perf_event *event,
5675 struct perf_sample_data *data,
5676 struct pt_regs *regs)
5678 if (event->hw.state & PERF_HES_STOPPED)
5681 * All tracepoints are from kernel-space.
5683 if (event->attr.exclude_kernel)
5686 if (!perf_tp_filter_match(event, data))
5692 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5693 struct pt_regs *regs, struct hlist_head *head, int rctx,
5694 struct task_struct *task)
5696 struct perf_sample_data data;
5697 struct perf_event *event;
5699 struct perf_raw_record raw = {
5704 perf_sample_data_init(&data, addr, 0);
5707 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5708 if (perf_tp_event_match(event, &data, regs))
5709 perf_swevent_event(event, count, &data, regs);
5713 * If we got specified a target task, also iterate its context and
5714 * deliver this event there too.
5716 if (task && task != current) {
5717 struct perf_event_context *ctx;
5718 struct trace_entry *entry = record;
5721 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5725 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5726 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5728 if (event->attr.config != entry->type)
5730 if (perf_tp_event_match(event, &data, regs))
5731 perf_swevent_event(event, count, &data, regs);
5737 perf_swevent_put_recursion_context(rctx);
5739 EXPORT_SYMBOL_GPL(perf_tp_event);
5741 static void tp_perf_event_destroy(struct perf_event *event)
5743 perf_trace_destroy(event);
5746 static int perf_tp_event_init(struct perf_event *event)
5750 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5754 * no branch sampling for tracepoint events
5756 if (has_branch_stack(event))
5759 err = perf_trace_init(event);
5763 event->destroy = tp_perf_event_destroy;
5768 static struct pmu perf_tracepoint = {
5769 .task_ctx_nr = perf_sw_context,
5771 .event_init = perf_tp_event_init,
5772 .add = perf_trace_add,
5773 .del = perf_trace_del,
5774 .start = perf_swevent_start,
5775 .stop = perf_swevent_stop,
5776 .read = perf_swevent_read,
5778 .event_idx = perf_swevent_event_idx,
5781 static inline void perf_tp_register(void)
5783 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5786 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5791 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5794 filter_str = strndup_user(arg, PAGE_SIZE);
5795 if (IS_ERR(filter_str))
5796 return PTR_ERR(filter_str);
5798 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5804 static void perf_event_free_filter(struct perf_event *event)
5806 ftrace_profile_free_filter(event);
5811 static inline void perf_tp_register(void)
5815 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5820 static void perf_event_free_filter(struct perf_event *event)
5824 #endif /* CONFIG_EVENT_TRACING */
5826 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5827 void perf_bp_event(struct perf_event *bp, void *data)
5829 struct perf_sample_data sample;
5830 struct pt_regs *regs = data;
5832 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5834 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5835 perf_swevent_event(bp, 1, &sample, regs);
5840 * hrtimer based swevent callback
5843 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5845 enum hrtimer_restart ret = HRTIMER_RESTART;
5846 struct perf_sample_data data;
5847 struct pt_regs *regs;
5848 struct perf_event *event;
5851 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5853 if (event->state != PERF_EVENT_STATE_ACTIVE)
5854 return HRTIMER_NORESTART;
5856 event->pmu->read(event);
5858 perf_sample_data_init(&data, 0, event->hw.last_period);
5859 regs = get_irq_regs();
5861 if (regs && !perf_exclude_event(event, regs)) {
5862 if (!(event->attr.exclude_idle && is_idle_task(current)))
5863 if (__perf_event_overflow(event, 1, &data, regs))
5864 ret = HRTIMER_NORESTART;
5867 period = max_t(u64, 10000, event->hw.sample_period);
5868 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5873 static void perf_swevent_start_hrtimer(struct perf_event *event)
5875 struct hw_perf_event *hwc = &event->hw;
5878 if (!is_sampling_event(event))
5881 period = local64_read(&hwc->period_left);
5886 local64_set(&hwc->period_left, 0);
5888 period = max_t(u64, 10000, hwc->sample_period);
5890 __hrtimer_start_range_ns(&hwc->hrtimer,
5891 ns_to_ktime(period), 0,
5892 HRTIMER_MODE_REL_PINNED, 0);
5895 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5897 struct hw_perf_event *hwc = &event->hw;
5899 if (is_sampling_event(event)) {
5900 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5901 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5903 hrtimer_cancel(&hwc->hrtimer);
5907 static void perf_swevent_init_hrtimer(struct perf_event *event)
5909 struct hw_perf_event *hwc = &event->hw;
5911 if (!is_sampling_event(event))
5914 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5915 hwc->hrtimer.function = perf_swevent_hrtimer;
5918 * Since hrtimers have a fixed rate, we can do a static freq->period
5919 * mapping and avoid the whole period adjust feedback stuff.
5921 if (event->attr.freq) {
5922 long freq = event->attr.sample_freq;
5924 event->attr.sample_period = NSEC_PER_SEC / freq;
5925 hwc->sample_period = event->attr.sample_period;
5926 local64_set(&hwc->period_left, hwc->sample_period);
5927 hwc->last_period = hwc->sample_period;
5928 event->attr.freq = 0;
5933 * Software event: cpu wall time clock
5936 static void cpu_clock_event_update(struct perf_event *event)
5941 now = local_clock();
5942 prev = local64_xchg(&event->hw.prev_count, now);
5943 local64_add(now - prev, &event->count);
5946 static void cpu_clock_event_start(struct perf_event *event, int flags)
5948 local64_set(&event->hw.prev_count, local_clock());
5949 perf_swevent_start_hrtimer(event);
5952 static void cpu_clock_event_stop(struct perf_event *event, int flags)
5954 perf_swevent_cancel_hrtimer(event);
5955 cpu_clock_event_update(event);
5958 static int cpu_clock_event_add(struct perf_event *event, int flags)
5960 if (flags & PERF_EF_START)
5961 cpu_clock_event_start(event, flags);
5966 static void cpu_clock_event_del(struct perf_event *event, int flags)
5968 cpu_clock_event_stop(event, flags);
5971 static void cpu_clock_event_read(struct perf_event *event)
5973 cpu_clock_event_update(event);
5976 static int cpu_clock_event_init(struct perf_event *event)
5978 if (event->attr.type != PERF_TYPE_SOFTWARE)
5981 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5985 * no branch sampling for software events
5987 if (has_branch_stack(event))
5990 perf_swevent_init_hrtimer(event);
5995 static struct pmu perf_cpu_clock = {
5996 .task_ctx_nr = perf_sw_context,
5998 .event_init = cpu_clock_event_init,
5999 .add = cpu_clock_event_add,
6000 .del = cpu_clock_event_del,
6001 .start = cpu_clock_event_start,
6002 .stop = cpu_clock_event_stop,
6003 .read = cpu_clock_event_read,
6005 .event_idx = perf_swevent_event_idx,
6009 * Software event: task time clock
6012 static void task_clock_event_update(struct perf_event *event, u64 now)
6017 prev = local64_xchg(&event->hw.prev_count, now);
6019 local64_add(delta, &event->count);
6022 static void task_clock_event_start(struct perf_event *event, int flags)
6024 local64_set(&event->hw.prev_count, event->ctx->time);
6025 perf_swevent_start_hrtimer(event);
6028 static void task_clock_event_stop(struct perf_event *event, int flags)
6030 perf_swevent_cancel_hrtimer(event);
6031 task_clock_event_update(event, event->ctx->time);
6034 static int task_clock_event_add(struct perf_event *event, int flags)
6036 if (flags & PERF_EF_START)
6037 task_clock_event_start(event, flags);
6042 static void task_clock_event_del(struct perf_event *event, int flags)
6044 task_clock_event_stop(event, PERF_EF_UPDATE);
6047 static void task_clock_event_read(struct perf_event *event)
6049 u64 now = perf_clock();
6050 u64 delta = now - event->ctx->timestamp;
6051 u64 time = event->ctx->time + delta;
6053 task_clock_event_update(event, time);
6056 static int task_clock_event_init(struct perf_event *event)
6058 if (event->attr.type != PERF_TYPE_SOFTWARE)
6061 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6065 * no branch sampling for software events
6067 if (has_branch_stack(event))
6070 perf_swevent_init_hrtimer(event);
6075 static struct pmu perf_task_clock = {
6076 .task_ctx_nr = perf_sw_context,
6078 .event_init = task_clock_event_init,
6079 .add = task_clock_event_add,
6080 .del = task_clock_event_del,
6081 .start = task_clock_event_start,
6082 .stop = task_clock_event_stop,
6083 .read = task_clock_event_read,
6085 .event_idx = perf_swevent_event_idx,
6088 static void perf_pmu_nop_void(struct pmu *pmu)
6092 static int perf_pmu_nop_int(struct pmu *pmu)
6097 static void perf_pmu_start_txn(struct pmu *pmu)
6099 perf_pmu_disable(pmu);
6102 static int perf_pmu_commit_txn(struct pmu *pmu)
6104 perf_pmu_enable(pmu);
6108 static void perf_pmu_cancel_txn(struct pmu *pmu)
6110 perf_pmu_enable(pmu);
6113 static int perf_event_idx_default(struct perf_event *event)
6115 return event->hw.idx + 1;
6119 * Ensures all contexts with the same task_ctx_nr have the same
6120 * pmu_cpu_context too.
6122 static void *find_pmu_context(int ctxn)
6129 list_for_each_entry(pmu, &pmus, entry) {
6130 if (pmu->task_ctx_nr == ctxn)
6131 return pmu->pmu_cpu_context;
6137 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6141 for_each_possible_cpu(cpu) {
6142 struct perf_cpu_context *cpuctx;
6144 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6146 if (cpuctx->unique_pmu == old_pmu)
6147 cpuctx->unique_pmu = pmu;
6151 static void free_pmu_context(struct pmu *pmu)
6155 mutex_lock(&pmus_lock);
6157 * Like a real lame refcount.
6159 list_for_each_entry(i, &pmus, entry) {
6160 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6161 update_pmu_context(i, pmu);
6166 free_percpu(pmu->pmu_cpu_context);
6168 mutex_unlock(&pmus_lock);
6170 static struct idr pmu_idr;
6173 type_show(struct device *dev, struct device_attribute *attr, char *page)
6175 struct pmu *pmu = dev_get_drvdata(dev);
6177 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6181 perf_event_mux_interval_ms_show(struct device *dev,
6182 struct device_attribute *attr,
6185 struct pmu *pmu = dev_get_drvdata(dev);
6187 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6191 perf_event_mux_interval_ms_store(struct device *dev,
6192 struct device_attribute *attr,
6193 const char *buf, size_t count)
6195 struct pmu *pmu = dev_get_drvdata(dev);
6196 int timer, cpu, ret;
6198 ret = kstrtoint(buf, 0, &timer);
6205 /* same value, noting to do */
6206 if (timer == pmu->hrtimer_interval_ms)
6209 pmu->hrtimer_interval_ms = timer;
6211 /* update all cpuctx for this PMU */
6212 for_each_possible_cpu(cpu) {
6213 struct perf_cpu_context *cpuctx;
6214 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6215 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6217 if (hrtimer_active(&cpuctx->hrtimer))
6218 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6224 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
6226 static struct device_attribute pmu_dev_attrs[] = {
6228 __ATTR_RW(perf_event_mux_interval_ms),
6232 static int pmu_bus_running;
6233 static struct bus_type pmu_bus = {
6234 .name = "event_source",
6235 .dev_attrs = pmu_dev_attrs,
6238 static void pmu_dev_release(struct device *dev)
6243 static int pmu_dev_alloc(struct pmu *pmu)
6247 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6251 pmu->dev->groups = pmu->attr_groups;
6252 device_initialize(pmu->dev);
6253 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6257 dev_set_drvdata(pmu->dev, pmu);
6258 pmu->dev->bus = &pmu_bus;
6259 pmu->dev->release = pmu_dev_release;
6260 ret = device_add(pmu->dev);
6268 put_device(pmu->dev);
6272 static struct lock_class_key cpuctx_mutex;
6273 static struct lock_class_key cpuctx_lock;
6275 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6279 mutex_lock(&pmus_lock);
6281 pmu->pmu_disable_count = alloc_percpu(int);
6282 if (!pmu->pmu_disable_count)
6291 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6299 if (pmu_bus_running) {
6300 ret = pmu_dev_alloc(pmu);
6306 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6307 if (pmu->pmu_cpu_context)
6308 goto got_cpu_context;
6311 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6312 if (!pmu->pmu_cpu_context)
6315 for_each_possible_cpu(cpu) {
6316 struct perf_cpu_context *cpuctx;
6318 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6319 __perf_event_init_context(&cpuctx->ctx);
6320 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6321 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6322 cpuctx->ctx.type = cpu_context;
6323 cpuctx->ctx.pmu = pmu;
6325 __perf_cpu_hrtimer_init(cpuctx, cpu);
6327 INIT_LIST_HEAD(&cpuctx->rotation_list);
6328 cpuctx->unique_pmu = pmu;
6332 if (!pmu->start_txn) {
6333 if (pmu->pmu_enable) {
6335 * If we have pmu_enable/pmu_disable calls, install
6336 * transaction stubs that use that to try and batch
6337 * hardware accesses.
6339 pmu->start_txn = perf_pmu_start_txn;
6340 pmu->commit_txn = perf_pmu_commit_txn;
6341 pmu->cancel_txn = perf_pmu_cancel_txn;
6343 pmu->start_txn = perf_pmu_nop_void;
6344 pmu->commit_txn = perf_pmu_nop_int;
6345 pmu->cancel_txn = perf_pmu_nop_void;
6349 if (!pmu->pmu_enable) {
6350 pmu->pmu_enable = perf_pmu_nop_void;
6351 pmu->pmu_disable = perf_pmu_nop_void;
6354 if (!pmu->event_idx)
6355 pmu->event_idx = perf_event_idx_default;
6357 list_add_rcu(&pmu->entry, &pmus);
6360 mutex_unlock(&pmus_lock);
6365 device_del(pmu->dev);
6366 put_device(pmu->dev);
6369 if (pmu->type >= PERF_TYPE_MAX)
6370 idr_remove(&pmu_idr, pmu->type);
6373 free_percpu(pmu->pmu_disable_count);
6377 void perf_pmu_unregister(struct pmu *pmu)
6379 mutex_lock(&pmus_lock);
6380 list_del_rcu(&pmu->entry);
6381 mutex_unlock(&pmus_lock);
6384 * We dereference the pmu list under both SRCU and regular RCU, so
6385 * synchronize against both of those.
6387 synchronize_srcu(&pmus_srcu);
6390 free_percpu(pmu->pmu_disable_count);
6391 if (pmu->type >= PERF_TYPE_MAX)
6392 idr_remove(&pmu_idr, pmu->type);
6393 device_del(pmu->dev);
6394 put_device(pmu->dev);
6395 free_pmu_context(pmu);
6398 struct pmu *perf_init_event(struct perf_event *event)
6400 struct pmu *pmu = NULL;
6404 idx = srcu_read_lock(&pmus_srcu);
6407 pmu = idr_find(&pmu_idr, event->attr.type);
6411 ret = pmu->event_init(event);
6417 list_for_each_entry_rcu(pmu, &pmus, entry) {
6419 ret = pmu->event_init(event);
6423 if (ret != -ENOENT) {
6428 pmu = ERR_PTR(-ENOENT);
6430 srcu_read_unlock(&pmus_srcu, idx);
6436 * Allocate and initialize a event structure
6438 static struct perf_event *
6439 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6440 struct task_struct *task,
6441 struct perf_event *group_leader,
6442 struct perf_event *parent_event,
6443 perf_overflow_handler_t overflow_handler,
6447 struct perf_event *event;
6448 struct hw_perf_event *hwc;
6451 if ((unsigned)cpu >= nr_cpu_ids) {
6452 if (!task || cpu != -1)
6453 return ERR_PTR(-EINVAL);
6456 event = kzalloc(sizeof(*event), GFP_KERNEL);
6458 return ERR_PTR(-ENOMEM);
6461 * Single events are their own group leaders, with an
6462 * empty sibling list:
6465 group_leader = event;
6467 mutex_init(&event->child_mutex);
6468 INIT_LIST_HEAD(&event->child_list);
6470 INIT_LIST_HEAD(&event->group_entry);
6471 INIT_LIST_HEAD(&event->event_entry);
6472 INIT_LIST_HEAD(&event->sibling_list);
6473 INIT_LIST_HEAD(&event->rb_entry);
6475 init_waitqueue_head(&event->waitq);
6476 init_irq_work(&event->pending, perf_pending_event);
6478 mutex_init(&event->mmap_mutex);
6480 atomic_long_set(&event->refcount, 1);
6482 event->attr = *attr;
6483 event->group_leader = group_leader;
6487 event->parent = parent_event;
6489 event->ns = get_pid_ns(task_active_pid_ns(current));
6490 event->id = atomic64_inc_return(&perf_event_id);
6492 event->state = PERF_EVENT_STATE_INACTIVE;
6495 event->attach_state = PERF_ATTACH_TASK;
6497 if (attr->type == PERF_TYPE_TRACEPOINT)
6498 event->hw.tp_target = task;
6499 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6501 * hw_breakpoint is a bit difficult here..
6503 else if (attr->type == PERF_TYPE_BREAKPOINT)
6504 event->hw.bp_target = task;
6508 if (!overflow_handler && parent_event) {
6509 overflow_handler = parent_event->overflow_handler;
6510 context = parent_event->overflow_handler_context;
6513 event->overflow_handler = overflow_handler;
6514 event->overflow_handler_context = context;
6516 perf_event__state_init(event);
6521 hwc->sample_period = attr->sample_period;
6522 if (attr->freq && attr->sample_freq)
6523 hwc->sample_period = 1;
6524 hwc->last_period = hwc->sample_period;
6526 local64_set(&hwc->period_left, hwc->sample_period);
6529 * we currently do not support PERF_FORMAT_GROUP on inherited events
6531 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6534 pmu = perf_init_event(event);
6540 else if (IS_ERR(pmu))
6545 put_pid_ns(event->ns);
6547 return ERR_PTR(err);
6550 if (!event->parent) {
6551 if (event->attach_state & PERF_ATTACH_TASK)
6552 static_key_slow_inc(&perf_sched_events.key);
6553 if (event->attr.mmap || event->attr.mmap_data)
6554 atomic_inc(&nr_mmap_events);
6555 if (event->attr.comm)
6556 atomic_inc(&nr_comm_events);
6557 if (event->attr.task)
6558 atomic_inc(&nr_task_events);
6559 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6560 err = get_callchain_buffers();
6563 return ERR_PTR(err);
6566 if (has_branch_stack(event)) {
6567 static_key_slow_inc(&perf_sched_events.key);
6568 if (!(event->attach_state & PERF_ATTACH_TASK))
6569 atomic_inc(&per_cpu(perf_branch_stack_events,
6577 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6578 struct perf_event_attr *attr)
6583 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6587 * zero the full structure, so that a short copy will be nice.
6589 memset(attr, 0, sizeof(*attr));
6591 ret = get_user(size, &uattr->size);
6595 if (size > PAGE_SIZE) /* silly large */
6598 if (!size) /* abi compat */
6599 size = PERF_ATTR_SIZE_VER0;
6601 if (size < PERF_ATTR_SIZE_VER0)
6605 * If we're handed a bigger struct than we know of,
6606 * ensure all the unknown bits are 0 - i.e. new
6607 * user-space does not rely on any kernel feature
6608 * extensions we dont know about yet.
6610 if (size > sizeof(*attr)) {
6611 unsigned char __user *addr;
6612 unsigned char __user *end;
6615 addr = (void __user *)uattr + sizeof(*attr);
6616 end = (void __user *)uattr + size;
6618 for (; addr < end; addr++) {
6619 ret = get_user(val, addr);
6625 size = sizeof(*attr);
6628 ret = copy_from_user(attr, uattr, size);
6632 if (attr->__reserved_1)
6635 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6638 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6641 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6642 u64 mask = attr->branch_sample_type;
6644 /* only using defined bits */
6645 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6648 /* at least one branch bit must be set */
6649 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6652 /* propagate priv level, when not set for branch */
6653 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6655 /* exclude_kernel checked on syscall entry */
6656 if (!attr->exclude_kernel)
6657 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6659 if (!attr->exclude_user)
6660 mask |= PERF_SAMPLE_BRANCH_USER;
6662 if (!attr->exclude_hv)
6663 mask |= PERF_SAMPLE_BRANCH_HV;
6665 * adjust user setting (for HW filter setup)
6667 attr->branch_sample_type = mask;
6669 /* privileged levels capture (kernel, hv): check permissions */
6670 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6671 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6675 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6676 ret = perf_reg_validate(attr->sample_regs_user);
6681 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6682 if (!arch_perf_have_user_stack_dump())
6686 * We have __u32 type for the size, but so far
6687 * we can only use __u16 as maximum due to the
6688 * __u16 sample size limit.
6690 if (attr->sample_stack_user >= USHRT_MAX)
6692 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6700 put_user(sizeof(*attr), &uattr->size);
6706 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6708 struct ring_buffer *rb = NULL, *old_rb = NULL;
6714 /* don't allow circular references */
6715 if (event == output_event)
6719 * Don't allow cross-cpu buffers
6721 if (output_event->cpu != event->cpu)
6725 * If its not a per-cpu rb, it must be the same task.
6727 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6731 mutex_lock(&event->mmap_mutex);
6732 /* Can't redirect output if we've got an active mmap() */
6733 if (atomic_read(&event->mmap_count))
6739 /* get the rb we want to redirect to */
6740 rb = ring_buffer_get(output_event);
6746 ring_buffer_detach(event, old_rb);
6749 ring_buffer_attach(event, rb);
6751 rcu_assign_pointer(event->rb, rb);
6754 ring_buffer_put(old_rb);
6756 * Since we detached before setting the new rb, so that we
6757 * could attach the new rb, we could have missed a wakeup.
6760 wake_up_all(&event->waitq);
6765 mutex_unlock(&event->mmap_mutex);
6772 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6774 * @attr_uptr: event_id type attributes for monitoring/sampling
6777 * @group_fd: group leader event fd
6779 SYSCALL_DEFINE5(perf_event_open,
6780 struct perf_event_attr __user *, attr_uptr,
6781 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6783 struct perf_event *group_leader = NULL, *output_event = NULL;
6784 struct perf_event *event, *sibling;
6785 struct perf_event_attr attr;
6786 struct perf_event_context *ctx;
6787 struct file *event_file = NULL;
6788 struct fd group = {NULL, 0};
6789 struct task_struct *task = NULL;
6795 /* for future expandability... */
6796 if (flags & ~PERF_FLAG_ALL)
6799 err = perf_copy_attr(attr_uptr, &attr);
6803 if (!attr.exclude_kernel) {
6804 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6809 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6814 * In cgroup mode, the pid argument is used to pass the fd
6815 * opened to the cgroup directory in cgroupfs. The cpu argument
6816 * designates the cpu on which to monitor threads from that
6819 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6822 event_fd = get_unused_fd();
6826 if (group_fd != -1) {
6827 err = perf_fget_light(group_fd, &group);
6830 group_leader = group.file->private_data;
6831 if (flags & PERF_FLAG_FD_OUTPUT)
6832 output_event = group_leader;
6833 if (flags & PERF_FLAG_FD_NO_GROUP)
6834 group_leader = NULL;
6837 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6838 task = find_lively_task_by_vpid(pid);
6840 err = PTR_ERR(task);
6847 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6849 if (IS_ERR(event)) {
6850 err = PTR_ERR(event);
6854 if (flags & PERF_FLAG_PID_CGROUP) {
6855 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6860 * - that has cgroup constraint on event->cpu
6861 * - that may need work on context switch
6863 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6864 static_key_slow_inc(&perf_sched_events.key);
6868 * Special case software events and allow them to be part of
6869 * any hardware group.
6874 (is_software_event(event) != is_software_event(group_leader))) {
6875 if (is_software_event(event)) {
6877 * If event and group_leader are not both a software
6878 * event, and event is, then group leader is not.
6880 * Allow the addition of software events to !software
6881 * groups, this is safe because software events never
6884 pmu = group_leader->pmu;
6885 } else if (is_software_event(group_leader) &&
6886 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6888 * In case the group is a pure software group, and we
6889 * try to add a hardware event, move the whole group to
6890 * the hardware context.
6897 * Get the target context (task or percpu):
6899 ctx = find_get_context(pmu, task, event->cpu);
6906 put_task_struct(task);
6911 * Look up the group leader (we will attach this event to it):
6917 * Do not allow a recursive hierarchy (this new sibling
6918 * becoming part of another group-sibling):
6920 if (group_leader->group_leader != group_leader)
6923 * Do not allow to attach to a group in a different
6924 * task or CPU context:
6927 if (group_leader->ctx->type != ctx->type)
6930 if (group_leader->ctx != ctx)
6935 * Only a group leader can be exclusive or pinned
6937 if (attr.exclusive || attr.pinned)
6942 err = perf_event_set_output(event, output_event);
6947 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6948 if (IS_ERR(event_file)) {
6949 err = PTR_ERR(event_file);
6954 struct perf_event_context *gctx = group_leader->ctx;
6956 mutex_lock(&gctx->mutex);
6957 perf_remove_from_context(group_leader);
6960 * Removing from the context ends up with disabled
6961 * event. What we want here is event in the initial
6962 * startup state, ready to be add into new context.
6964 perf_event__state_init(group_leader);
6965 list_for_each_entry(sibling, &group_leader->sibling_list,
6967 perf_remove_from_context(sibling);
6968 perf_event__state_init(sibling);
6971 mutex_unlock(&gctx->mutex);
6975 WARN_ON_ONCE(ctx->parent_ctx);
6976 mutex_lock(&ctx->mutex);
6980 perf_install_in_context(ctx, group_leader, event->cpu);
6982 list_for_each_entry(sibling, &group_leader->sibling_list,
6984 perf_install_in_context(ctx, sibling, event->cpu);
6989 perf_install_in_context(ctx, event, event->cpu);
6991 perf_unpin_context(ctx);
6992 mutex_unlock(&ctx->mutex);
6996 event->owner = current;
6998 mutex_lock(¤t->perf_event_mutex);
6999 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7000 mutex_unlock(¤t->perf_event_mutex);
7003 * Precalculate sample_data sizes
7005 perf_event__header_size(event);
7006 perf_event__id_header_size(event);
7009 * Drop the reference on the group_event after placing the
7010 * new event on the sibling_list. This ensures destruction
7011 * of the group leader will find the pointer to itself in
7012 * perf_group_detach().
7015 fd_install(event_fd, event_file);
7019 perf_unpin_context(ctx);
7026 put_task_struct(task);
7030 put_unused_fd(event_fd);
7035 * perf_event_create_kernel_counter
7037 * @attr: attributes of the counter to create
7038 * @cpu: cpu in which the counter is bound
7039 * @task: task to profile (NULL for percpu)
7042 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7043 struct task_struct *task,
7044 perf_overflow_handler_t overflow_handler,
7047 struct perf_event_context *ctx;
7048 struct perf_event *event;
7052 * Get the target context (task or percpu):
7055 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7056 overflow_handler, context);
7057 if (IS_ERR(event)) {
7058 err = PTR_ERR(event);
7062 ctx = find_get_context(event->pmu, task, cpu);
7068 WARN_ON_ONCE(ctx->parent_ctx);
7069 mutex_lock(&ctx->mutex);
7070 perf_install_in_context(ctx, event, cpu);
7072 perf_unpin_context(ctx);
7073 mutex_unlock(&ctx->mutex);
7080 return ERR_PTR(err);
7082 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7084 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7086 struct perf_event_context *src_ctx;
7087 struct perf_event_context *dst_ctx;
7088 struct perf_event *event, *tmp;
7091 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7092 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7094 mutex_lock(&src_ctx->mutex);
7095 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7097 perf_remove_from_context(event);
7099 list_add(&event->event_entry, &events);
7101 mutex_unlock(&src_ctx->mutex);
7105 mutex_lock(&dst_ctx->mutex);
7106 list_for_each_entry_safe(event, tmp, &events, event_entry) {
7107 list_del(&event->event_entry);
7108 if (event->state >= PERF_EVENT_STATE_OFF)
7109 event->state = PERF_EVENT_STATE_INACTIVE;
7110 perf_install_in_context(dst_ctx, event, dst_cpu);
7113 mutex_unlock(&dst_ctx->mutex);
7115 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7117 static void sync_child_event(struct perf_event *child_event,
7118 struct task_struct *child)
7120 struct perf_event *parent_event = child_event->parent;
7123 if (child_event->attr.inherit_stat)
7124 perf_event_read_event(child_event, child);
7126 child_val = perf_event_count(child_event);
7129 * Add back the child's count to the parent's count:
7131 atomic64_add(child_val, &parent_event->child_count);
7132 atomic64_add(child_event->total_time_enabled,
7133 &parent_event->child_total_time_enabled);
7134 atomic64_add(child_event->total_time_running,
7135 &parent_event->child_total_time_running);
7138 * Remove this event from the parent's list
7140 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7141 mutex_lock(&parent_event->child_mutex);
7142 list_del_init(&child_event->child_list);
7143 mutex_unlock(&parent_event->child_mutex);
7146 * Release the parent event, if this was the last
7149 put_event(parent_event);
7153 __perf_event_exit_task(struct perf_event *child_event,
7154 struct perf_event_context *child_ctx,
7155 struct task_struct *child)
7157 if (child_event->parent) {
7158 raw_spin_lock_irq(&child_ctx->lock);
7159 perf_group_detach(child_event);
7160 raw_spin_unlock_irq(&child_ctx->lock);
7163 perf_remove_from_context(child_event);
7166 * It can happen that the parent exits first, and has events
7167 * that are still around due to the child reference. These
7168 * events need to be zapped.
7170 if (child_event->parent) {
7171 sync_child_event(child_event, child);
7172 free_event(child_event);
7176 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7178 struct perf_event *child_event, *tmp;
7179 struct perf_event_context *child_ctx;
7180 unsigned long flags;
7182 if (likely(!child->perf_event_ctxp[ctxn])) {
7183 perf_event_task(child, NULL, 0);
7187 local_irq_save(flags);
7189 * We can't reschedule here because interrupts are disabled,
7190 * and either child is current or it is a task that can't be
7191 * scheduled, so we are now safe from rescheduling changing
7194 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7197 * Take the context lock here so that if find_get_context is
7198 * reading child->perf_event_ctxp, we wait until it has
7199 * incremented the context's refcount before we do put_ctx below.
7201 raw_spin_lock(&child_ctx->lock);
7202 task_ctx_sched_out(child_ctx);
7203 child->perf_event_ctxp[ctxn] = NULL;
7205 * If this context is a clone; unclone it so it can't get
7206 * swapped to another process while we're removing all
7207 * the events from it.
7209 unclone_ctx(child_ctx);
7210 update_context_time(child_ctx);
7211 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7214 * Report the task dead after unscheduling the events so that we
7215 * won't get any samples after PERF_RECORD_EXIT. We can however still
7216 * get a few PERF_RECORD_READ events.
7218 perf_event_task(child, child_ctx, 0);
7221 * We can recurse on the same lock type through:
7223 * __perf_event_exit_task()
7224 * sync_child_event()
7226 * mutex_lock(&ctx->mutex)
7228 * But since its the parent context it won't be the same instance.
7230 mutex_lock(&child_ctx->mutex);
7233 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7235 __perf_event_exit_task(child_event, child_ctx, child);
7237 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7239 __perf_event_exit_task(child_event, child_ctx, child);
7242 * If the last event was a group event, it will have appended all
7243 * its siblings to the list, but we obtained 'tmp' before that which
7244 * will still point to the list head terminating the iteration.
7246 if (!list_empty(&child_ctx->pinned_groups) ||
7247 !list_empty(&child_ctx->flexible_groups))
7250 mutex_unlock(&child_ctx->mutex);
7256 * When a child task exits, feed back event values to parent events.
7258 void perf_event_exit_task(struct task_struct *child)
7260 struct perf_event *event, *tmp;
7263 mutex_lock(&child->perf_event_mutex);
7264 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7266 list_del_init(&event->owner_entry);
7269 * Ensure the list deletion is visible before we clear
7270 * the owner, closes a race against perf_release() where
7271 * we need to serialize on the owner->perf_event_mutex.
7274 event->owner = NULL;
7276 mutex_unlock(&child->perf_event_mutex);
7278 for_each_task_context_nr(ctxn)
7279 perf_event_exit_task_context(child, ctxn);
7282 static void perf_free_event(struct perf_event *event,
7283 struct perf_event_context *ctx)
7285 struct perf_event *parent = event->parent;
7287 if (WARN_ON_ONCE(!parent))
7290 mutex_lock(&parent->child_mutex);
7291 list_del_init(&event->child_list);
7292 mutex_unlock(&parent->child_mutex);
7296 perf_group_detach(event);
7297 list_del_event(event, ctx);
7302 * free an unexposed, unused context as created by inheritance by
7303 * perf_event_init_task below, used by fork() in case of fail.
7305 void perf_event_free_task(struct task_struct *task)
7307 struct perf_event_context *ctx;
7308 struct perf_event *event, *tmp;
7311 for_each_task_context_nr(ctxn) {
7312 ctx = task->perf_event_ctxp[ctxn];
7316 mutex_lock(&ctx->mutex);
7318 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7320 perf_free_event(event, ctx);
7322 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7324 perf_free_event(event, ctx);
7326 if (!list_empty(&ctx->pinned_groups) ||
7327 !list_empty(&ctx->flexible_groups))
7330 mutex_unlock(&ctx->mutex);
7336 void perf_event_delayed_put(struct task_struct *task)
7340 for_each_task_context_nr(ctxn)
7341 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7345 * inherit a event from parent task to child task:
7347 static struct perf_event *
7348 inherit_event(struct perf_event *parent_event,
7349 struct task_struct *parent,
7350 struct perf_event_context *parent_ctx,
7351 struct task_struct *child,
7352 struct perf_event *group_leader,
7353 struct perf_event_context *child_ctx)
7355 struct perf_event *child_event;
7356 unsigned long flags;
7359 * Instead of creating recursive hierarchies of events,
7360 * we link inherited events back to the original parent,
7361 * which has a filp for sure, which we use as the reference
7364 if (parent_event->parent)
7365 parent_event = parent_event->parent;
7367 child_event = perf_event_alloc(&parent_event->attr,
7370 group_leader, parent_event,
7372 if (IS_ERR(child_event))
7375 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7376 free_event(child_event);
7383 * Make the child state follow the state of the parent event,
7384 * not its attr.disabled bit. We hold the parent's mutex,
7385 * so we won't race with perf_event_{en, dis}able_family.
7387 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7388 child_event->state = PERF_EVENT_STATE_INACTIVE;
7390 child_event->state = PERF_EVENT_STATE_OFF;
7392 if (parent_event->attr.freq) {
7393 u64 sample_period = parent_event->hw.sample_period;
7394 struct hw_perf_event *hwc = &child_event->hw;
7396 hwc->sample_period = sample_period;
7397 hwc->last_period = sample_period;
7399 local64_set(&hwc->period_left, sample_period);
7402 child_event->ctx = child_ctx;
7403 child_event->overflow_handler = parent_event->overflow_handler;
7404 child_event->overflow_handler_context
7405 = parent_event->overflow_handler_context;
7408 * Precalculate sample_data sizes
7410 perf_event__header_size(child_event);
7411 perf_event__id_header_size(child_event);
7414 * Link it up in the child's context:
7416 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7417 add_event_to_ctx(child_event, child_ctx);
7418 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7421 * Link this into the parent event's child list
7423 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7424 mutex_lock(&parent_event->child_mutex);
7425 list_add_tail(&child_event->child_list, &parent_event->child_list);
7426 mutex_unlock(&parent_event->child_mutex);
7431 static int inherit_group(struct perf_event *parent_event,
7432 struct task_struct *parent,
7433 struct perf_event_context *parent_ctx,
7434 struct task_struct *child,
7435 struct perf_event_context *child_ctx)
7437 struct perf_event *leader;
7438 struct perf_event *sub;
7439 struct perf_event *child_ctr;
7441 leader = inherit_event(parent_event, parent, parent_ctx,
7442 child, NULL, child_ctx);
7444 return PTR_ERR(leader);
7445 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7446 child_ctr = inherit_event(sub, parent, parent_ctx,
7447 child, leader, child_ctx);
7448 if (IS_ERR(child_ctr))
7449 return PTR_ERR(child_ctr);
7455 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7456 struct perf_event_context *parent_ctx,
7457 struct task_struct *child, int ctxn,
7461 struct perf_event_context *child_ctx;
7463 if (!event->attr.inherit) {
7468 child_ctx = child->perf_event_ctxp[ctxn];
7471 * This is executed from the parent task context, so
7472 * inherit events that have been marked for cloning.
7473 * First allocate and initialize a context for the
7477 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7481 child->perf_event_ctxp[ctxn] = child_ctx;
7484 ret = inherit_group(event, parent, parent_ctx,
7494 * Initialize the perf_event context in task_struct
7496 int perf_event_init_context(struct task_struct *child, int ctxn)
7498 struct perf_event_context *child_ctx, *parent_ctx;
7499 struct perf_event_context *cloned_ctx;
7500 struct perf_event *event;
7501 struct task_struct *parent = current;
7502 int inherited_all = 1;
7503 unsigned long flags;
7506 if (likely(!parent->perf_event_ctxp[ctxn]))
7510 * If the parent's context is a clone, pin it so it won't get
7513 parent_ctx = perf_pin_task_context(parent, ctxn);
7516 * No need to check if parent_ctx != NULL here; since we saw
7517 * it non-NULL earlier, the only reason for it to become NULL
7518 * is if we exit, and since we're currently in the middle of
7519 * a fork we can't be exiting at the same time.
7523 * Lock the parent list. No need to lock the child - not PID
7524 * hashed yet and not running, so nobody can access it.
7526 mutex_lock(&parent_ctx->mutex);
7529 * We dont have to disable NMIs - we are only looking at
7530 * the list, not manipulating it:
7532 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7533 ret = inherit_task_group(event, parent, parent_ctx,
7534 child, ctxn, &inherited_all);
7540 * We can't hold ctx->lock when iterating the ->flexible_group list due
7541 * to allocations, but we need to prevent rotation because
7542 * rotate_ctx() will change the list from interrupt context.
7544 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7545 parent_ctx->rotate_disable = 1;
7546 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7548 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7549 ret = inherit_task_group(event, parent, parent_ctx,
7550 child, ctxn, &inherited_all);
7555 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7556 parent_ctx->rotate_disable = 0;
7558 child_ctx = child->perf_event_ctxp[ctxn];
7560 if (child_ctx && inherited_all) {
7562 * Mark the child context as a clone of the parent
7563 * context, or of whatever the parent is a clone of.
7565 * Note that if the parent is a clone, the holding of
7566 * parent_ctx->lock avoids it from being uncloned.
7568 cloned_ctx = parent_ctx->parent_ctx;
7570 child_ctx->parent_ctx = cloned_ctx;
7571 child_ctx->parent_gen = parent_ctx->parent_gen;
7573 child_ctx->parent_ctx = parent_ctx;
7574 child_ctx->parent_gen = parent_ctx->generation;
7576 get_ctx(child_ctx->parent_ctx);
7579 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7580 mutex_unlock(&parent_ctx->mutex);
7582 perf_unpin_context(parent_ctx);
7583 put_ctx(parent_ctx);
7589 * Initialize the perf_event context in task_struct
7591 int perf_event_init_task(struct task_struct *child)
7595 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7596 mutex_init(&child->perf_event_mutex);
7597 INIT_LIST_HEAD(&child->perf_event_list);
7599 for_each_task_context_nr(ctxn) {
7600 ret = perf_event_init_context(child, ctxn);
7608 static void __init perf_event_init_all_cpus(void)
7610 struct swevent_htable *swhash;
7613 for_each_possible_cpu(cpu) {
7614 swhash = &per_cpu(swevent_htable, cpu);
7615 mutex_init(&swhash->hlist_mutex);
7616 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7620 static void __cpuinit perf_event_init_cpu(int cpu)
7622 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7624 mutex_lock(&swhash->hlist_mutex);
7625 if (swhash->hlist_refcount > 0) {
7626 struct swevent_hlist *hlist;
7628 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7630 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7632 mutex_unlock(&swhash->hlist_mutex);
7635 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7636 static void perf_pmu_rotate_stop(struct pmu *pmu)
7638 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7640 WARN_ON(!irqs_disabled());
7642 list_del_init(&cpuctx->rotation_list);
7645 static void __perf_event_exit_context(void *__info)
7647 struct perf_event_context *ctx = __info;
7648 struct perf_event *event, *tmp;
7650 perf_pmu_rotate_stop(ctx->pmu);
7652 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7653 __perf_remove_from_context(event);
7654 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7655 __perf_remove_from_context(event);
7658 static void perf_event_exit_cpu_context(int cpu)
7660 struct perf_event_context *ctx;
7664 idx = srcu_read_lock(&pmus_srcu);
7665 list_for_each_entry_rcu(pmu, &pmus, entry) {
7666 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7668 mutex_lock(&ctx->mutex);
7669 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7670 mutex_unlock(&ctx->mutex);
7672 srcu_read_unlock(&pmus_srcu, idx);
7675 static void perf_event_exit_cpu(int cpu)
7677 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7679 mutex_lock(&swhash->hlist_mutex);
7680 swevent_hlist_release(swhash);
7681 mutex_unlock(&swhash->hlist_mutex);
7683 perf_event_exit_cpu_context(cpu);
7686 static inline void perf_event_exit_cpu(int cpu) { }
7690 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7694 for_each_online_cpu(cpu)
7695 perf_event_exit_cpu(cpu);
7701 * Run the perf reboot notifier at the very last possible moment so that
7702 * the generic watchdog code runs as long as possible.
7704 static struct notifier_block perf_reboot_notifier = {
7705 .notifier_call = perf_reboot,
7706 .priority = INT_MIN,
7709 static int __cpuinit
7710 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7712 unsigned int cpu = (long)hcpu;
7714 switch (action & ~CPU_TASKS_FROZEN) {
7716 case CPU_UP_PREPARE:
7717 case CPU_DOWN_FAILED:
7718 perf_event_init_cpu(cpu);
7721 case CPU_UP_CANCELED:
7722 case CPU_DOWN_PREPARE:
7723 perf_event_exit_cpu(cpu);
7732 void __init perf_event_init(void)
7738 perf_event_init_all_cpus();
7739 init_srcu_struct(&pmus_srcu);
7740 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7741 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7742 perf_pmu_register(&perf_task_clock, NULL, -1);
7744 perf_cpu_notifier(perf_cpu_notify);
7745 register_reboot_notifier(&perf_reboot_notifier);
7747 ret = init_hw_breakpoint();
7748 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7750 /* do not patch jump label more than once per second */
7751 jump_label_rate_limit(&perf_sched_events, HZ);
7754 * Build time assertion that we keep the data_head at the intended
7755 * location. IOW, validation we got the __reserved[] size right.
7757 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7761 static int __init perf_event_sysfs_init(void)
7766 mutex_lock(&pmus_lock);
7768 ret = bus_register(&pmu_bus);
7772 list_for_each_entry(pmu, &pmus, entry) {
7773 if (!pmu->name || pmu->type < 0)
7776 ret = pmu_dev_alloc(pmu);
7777 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7779 pmu_bus_running = 1;
7783 mutex_unlock(&pmus_lock);
7787 device_initcall(perf_event_sysfs_init);
7789 #ifdef CONFIG_CGROUP_PERF
7790 static struct cgroup_subsys_state *perf_cgroup_css_alloc(struct cgroup *cont)
7792 struct perf_cgroup *jc;
7794 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7796 return ERR_PTR(-ENOMEM);
7798 jc->info = alloc_percpu(struct perf_cgroup_info);
7801 return ERR_PTR(-ENOMEM);
7807 static void perf_cgroup_css_free(struct cgroup *cont)
7809 struct perf_cgroup *jc;
7810 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7811 struct perf_cgroup, css);
7812 free_percpu(jc->info);
7816 static int __perf_cgroup_move(void *info)
7818 struct task_struct *task = info;
7819 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7823 static void perf_cgroup_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
7825 struct task_struct *task;
7827 cgroup_taskset_for_each(task, cgrp, tset)
7828 task_function_call(task, __perf_cgroup_move, task);
7831 static void perf_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7832 struct task_struct *task)
7835 * cgroup_exit() is called in the copy_process() failure path.
7836 * Ignore this case since the task hasn't ran yet, this avoids
7837 * trying to poke a half freed task state from generic code.
7839 if (!(task->flags & PF_EXITING))
7842 task_function_call(task, __perf_cgroup_move, task);
7845 struct cgroup_subsys perf_subsys = {
7846 .name = "perf_event",
7847 .subsys_id = perf_subsys_id,
7848 .css_alloc = perf_cgroup_css_alloc,
7849 .css_free = perf_cgroup_css_free,
7850 .exit = perf_cgroup_exit,
7851 .attach = perf_cgroup_attach,
7853 #endif /* CONFIG_CGROUP_PERF */