2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
42 #include <linux/module.h>
43 #include <linux/mman.h>
44 #include <linux/compat.h>
48 #include <asm/irq_regs.h>
50 static struct workqueue_struct *perf_wq;
52 struct remote_function_call {
53 struct task_struct *p;
54 int (*func)(void *info);
59 static void remote_function(void *data)
61 struct remote_function_call *tfc = data;
62 struct task_struct *p = tfc->p;
66 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
70 tfc->ret = tfc->func(tfc->info);
74 * task_function_call - call a function on the cpu on which a task runs
75 * @p: the task to evaluate
76 * @func: the function to be called
77 * @info: the function call argument
79 * Calls the function @func when the task is currently running. This might
80 * be on the current CPU, which just calls the function directly
82 * returns: @func return value, or
83 * -ESRCH - when the process isn't running
84 * -EAGAIN - when the process moved away
87 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
89 struct remote_function_call data = {
93 .ret = -ESRCH, /* No such (running) process */
97 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
103 * cpu_function_call - call a function on the cpu
104 * @func: the function to be called
105 * @info: the function call argument
107 * Calls the function @func on the remote cpu.
109 * returns: @func return value or -ENXIO when the cpu is offline
111 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
113 struct remote_function_call data = {
117 .ret = -ENXIO, /* No such CPU */
120 smp_call_function_single(cpu, remote_function, &data, 1);
125 #define EVENT_OWNER_KERNEL ((void *) -1)
127 static bool is_kernel_event(struct perf_event *event)
129 return event->owner == EVENT_OWNER_KERNEL;
132 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
133 PERF_FLAG_FD_OUTPUT |\
134 PERF_FLAG_PID_CGROUP |\
135 PERF_FLAG_FD_CLOEXEC)
138 * branch priv levels that need permission checks
140 #define PERF_SAMPLE_BRANCH_PERM_PLM \
141 (PERF_SAMPLE_BRANCH_KERNEL |\
142 PERF_SAMPLE_BRANCH_HV)
145 EVENT_FLEXIBLE = 0x1,
147 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
151 * perf_sched_events : >0 events exist
152 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
154 struct static_key_deferred perf_sched_events __read_mostly;
155 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
156 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
158 static atomic_t nr_mmap_events __read_mostly;
159 static atomic_t nr_comm_events __read_mostly;
160 static atomic_t nr_task_events __read_mostly;
161 static atomic_t nr_freq_events __read_mostly;
163 static LIST_HEAD(pmus);
164 static DEFINE_MUTEX(pmus_lock);
165 static struct srcu_struct pmus_srcu;
168 * perf event paranoia level:
169 * -1 - not paranoid at all
170 * 0 - disallow raw tracepoint access for unpriv
171 * 1 - disallow cpu events for unpriv
172 * 2 - disallow kernel profiling for unpriv
174 int sysctl_perf_event_paranoid __read_mostly = 1;
176 /* Minimum for 512 kiB + 1 user control page */
177 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
180 * max perf event sample rate
182 #define DEFAULT_MAX_SAMPLE_RATE 100000
183 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
184 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
186 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
188 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
189 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
191 static int perf_sample_allowed_ns __read_mostly =
192 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
194 void update_perf_cpu_limits(void)
196 u64 tmp = perf_sample_period_ns;
198 tmp *= sysctl_perf_cpu_time_max_percent;
200 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
203 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
205 int perf_proc_update_handler(struct ctl_table *table, int write,
206 void __user *buffer, size_t *lenp,
209 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
214 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
215 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
216 update_perf_cpu_limits();
221 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
223 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
224 void __user *buffer, size_t *lenp,
227 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
232 update_perf_cpu_limits();
238 * perf samples are done in some very critical code paths (NMIs).
239 * If they take too much CPU time, the system can lock up and not
240 * get any real work done. This will drop the sample rate when
241 * we detect that events are taking too long.
243 #define NR_ACCUMULATED_SAMPLES 128
244 static DEFINE_PER_CPU(u64, running_sample_length);
246 static void perf_duration_warn(struct irq_work *w)
248 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
249 u64 avg_local_sample_len;
250 u64 local_samples_len;
252 local_samples_len = __this_cpu_read(running_sample_length);
253 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
255 printk_ratelimited(KERN_WARNING
256 "perf interrupt took too long (%lld > %lld), lowering "
257 "kernel.perf_event_max_sample_rate to %d\n",
258 avg_local_sample_len, allowed_ns >> 1,
259 sysctl_perf_event_sample_rate);
262 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
264 void perf_sample_event_took(u64 sample_len_ns)
266 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
267 u64 avg_local_sample_len;
268 u64 local_samples_len;
273 /* decay the counter by 1 average sample */
274 local_samples_len = __this_cpu_read(running_sample_length);
275 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
276 local_samples_len += sample_len_ns;
277 __this_cpu_write(running_sample_length, local_samples_len);
280 * note: this will be biased artifically low until we have
281 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
282 * from having to maintain a count.
284 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
286 if (avg_local_sample_len <= allowed_ns)
289 if (max_samples_per_tick <= 1)
292 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
293 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
294 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
296 update_perf_cpu_limits();
298 if (!irq_work_queue(&perf_duration_work)) {
299 early_printk("perf interrupt took too long (%lld > %lld), lowering "
300 "kernel.perf_event_max_sample_rate to %d\n",
301 avg_local_sample_len, allowed_ns >> 1,
302 sysctl_perf_event_sample_rate);
306 static atomic64_t perf_event_id;
308 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
309 enum event_type_t event_type);
311 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
312 enum event_type_t event_type,
313 struct task_struct *task);
315 static void update_context_time(struct perf_event_context *ctx);
316 static u64 perf_event_time(struct perf_event *event);
318 void __weak perf_event_print_debug(void) { }
320 extern __weak const char *perf_pmu_name(void)
325 static inline u64 perf_clock(void)
327 return local_clock();
330 static inline struct perf_cpu_context *
331 __get_cpu_context(struct perf_event_context *ctx)
333 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
336 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
337 struct perf_event_context *ctx)
339 raw_spin_lock(&cpuctx->ctx.lock);
341 raw_spin_lock(&ctx->lock);
344 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
345 struct perf_event_context *ctx)
348 raw_spin_unlock(&ctx->lock);
349 raw_spin_unlock(&cpuctx->ctx.lock);
352 #ifdef CONFIG_CGROUP_PERF
355 * perf_cgroup_info keeps track of time_enabled for a cgroup.
356 * This is a per-cpu dynamically allocated data structure.
358 struct perf_cgroup_info {
364 struct cgroup_subsys_state css;
365 struct perf_cgroup_info __percpu *info;
369 * Must ensure cgroup is pinned (css_get) before calling
370 * this function. In other words, we cannot call this function
371 * if there is no cgroup event for the current CPU context.
373 static inline struct perf_cgroup *
374 perf_cgroup_from_task(struct task_struct *task)
376 return container_of(task_css(task, perf_event_cgrp_id),
377 struct perf_cgroup, css);
381 perf_cgroup_match(struct perf_event *event)
383 struct perf_event_context *ctx = event->ctx;
384 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
386 /* @event doesn't care about cgroup */
390 /* wants specific cgroup scope but @cpuctx isn't associated with any */
395 * Cgroup scoping is recursive. An event enabled for a cgroup is
396 * also enabled for all its descendant cgroups. If @cpuctx's
397 * cgroup is a descendant of @event's (the test covers identity
398 * case), it's a match.
400 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
401 event->cgrp->css.cgroup);
404 static inline void perf_detach_cgroup(struct perf_event *event)
406 css_put(&event->cgrp->css);
410 static inline int is_cgroup_event(struct perf_event *event)
412 return event->cgrp != NULL;
415 static inline u64 perf_cgroup_event_time(struct perf_event *event)
417 struct perf_cgroup_info *t;
419 t = per_cpu_ptr(event->cgrp->info, event->cpu);
423 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
425 struct perf_cgroup_info *info;
430 info = this_cpu_ptr(cgrp->info);
432 info->time += now - info->timestamp;
433 info->timestamp = now;
436 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
438 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
440 __update_cgrp_time(cgrp_out);
443 static inline void update_cgrp_time_from_event(struct perf_event *event)
445 struct perf_cgroup *cgrp;
448 * ensure we access cgroup data only when needed and
449 * when we know the cgroup is pinned (css_get)
451 if (!is_cgroup_event(event))
454 cgrp = perf_cgroup_from_task(current);
456 * Do not update time when cgroup is not active
458 if (cgrp == event->cgrp)
459 __update_cgrp_time(event->cgrp);
463 perf_cgroup_set_timestamp(struct task_struct *task,
464 struct perf_event_context *ctx)
466 struct perf_cgroup *cgrp;
467 struct perf_cgroup_info *info;
470 * ctx->lock held by caller
471 * ensure we do not access cgroup data
472 * unless we have the cgroup pinned (css_get)
474 if (!task || !ctx->nr_cgroups)
477 cgrp = perf_cgroup_from_task(task);
478 info = this_cpu_ptr(cgrp->info);
479 info->timestamp = ctx->timestamp;
482 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
483 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
486 * reschedule events based on the cgroup constraint of task.
488 * mode SWOUT : schedule out everything
489 * mode SWIN : schedule in based on cgroup for next
491 void perf_cgroup_switch(struct task_struct *task, int mode)
493 struct perf_cpu_context *cpuctx;
498 * disable interrupts to avoid geting nr_cgroup
499 * changes via __perf_event_disable(). Also
502 local_irq_save(flags);
505 * we reschedule only in the presence of cgroup
506 * constrained events.
510 list_for_each_entry_rcu(pmu, &pmus, entry) {
511 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
512 if (cpuctx->unique_pmu != pmu)
513 continue; /* ensure we process each cpuctx once */
516 * perf_cgroup_events says at least one
517 * context on this CPU has cgroup events.
519 * ctx->nr_cgroups reports the number of cgroup
520 * events for a context.
522 if (cpuctx->ctx.nr_cgroups > 0) {
523 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
524 perf_pmu_disable(cpuctx->ctx.pmu);
526 if (mode & PERF_CGROUP_SWOUT) {
527 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
529 * must not be done before ctxswout due
530 * to event_filter_match() in event_sched_out()
535 if (mode & PERF_CGROUP_SWIN) {
536 WARN_ON_ONCE(cpuctx->cgrp);
538 * set cgrp before ctxsw in to allow
539 * event_filter_match() to not have to pass
542 cpuctx->cgrp = perf_cgroup_from_task(task);
543 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
545 perf_pmu_enable(cpuctx->ctx.pmu);
546 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
552 local_irq_restore(flags);
555 static inline void perf_cgroup_sched_out(struct task_struct *task,
556 struct task_struct *next)
558 struct perf_cgroup *cgrp1;
559 struct perf_cgroup *cgrp2 = NULL;
562 * we come here when we know perf_cgroup_events > 0
564 cgrp1 = perf_cgroup_from_task(task);
567 * next is NULL when called from perf_event_enable_on_exec()
568 * that will systematically cause a cgroup_switch()
571 cgrp2 = perf_cgroup_from_task(next);
574 * only schedule out current cgroup events if we know
575 * that we are switching to a different cgroup. Otherwise,
576 * do no touch the cgroup events.
579 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
582 static inline void perf_cgroup_sched_in(struct task_struct *prev,
583 struct task_struct *task)
585 struct perf_cgroup *cgrp1;
586 struct perf_cgroup *cgrp2 = NULL;
589 * we come here when we know perf_cgroup_events > 0
591 cgrp1 = perf_cgroup_from_task(task);
593 /* prev can never be NULL */
594 cgrp2 = perf_cgroup_from_task(prev);
597 * only need to schedule in cgroup events if we are changing
598 * cgroup during ctxsw. Cgroup events were not scheduled
599 * out of ctxsw out if that was not the case.
602 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
605 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
606 struct perf_event_attr *attr,
607 struct perf_event *group_leader)
609 struct perf_cgroup *cgrp;
610 struct cgroup_subsys_state *css;
611 struct fd f = fdget(fd);
617 css = css_tryget_online_from_dir(f.file->f_path.dentry,
618 &perf_event_cgrp_subsys);
624 cgrp = container_of(css, struct perf_cgroup, css);
628 * all events in a group must monitor
629 * the same cgroup because a task belongs
630 * to only one perf cgroup at a time
632 if (group_leader && group_leader->cgrp != cgrp) {
633 perf_detach_cgroup(event);
642 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
644 struct perf_cgroup_info *t;
645 t = per_cpu_ptr(event->cgrp->info, event->cpu);
646 event->shadow_ctx_time = now - t->timestamp;
650 perf_cgroup_defer_enabled(struct perf_event *event)
653 * when the current task's perf cgroup does not match
654 * the event's, we need to remember to call the
655 * perf_mark_enable() function the first time a task with
656 * a matching perf cgroup is scheduled in.
658 if (is_cgroup_event(event) && !perf_cgroup_match(event))
659 event->cgrp_defer_enabled = 1;
663 perf_cgroup_mark_enabled(struct perf_event *event,
664 struct perf_event_context *ctx)
666 struct perf_event *sub;
667 u64 tstamp = perf_event_time(event);
669 if (!event->cgrp_defer_enabled)
672 event->cgrp_defer_enabled = 0;
674 event->tstamp_enabled = tstamp - event->total_time_enabled;
675 list_for_each_entry(sub, &event->sibling_list, group_entry) {
676 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
677 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
678 sub->cgrp_defer_enabled = 0;
682 #else /* !CONFIG_CGROUP_PERF */
685 perf_cgroup_match(struct perf_event *event)
690 static inline void perf_detach_cgroup(struct perf_event *event)
693 static inline int is_cgroup_event(struct perf_event *event)
698 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
703 static inline void update_cgrp_time_from_event(struct perf_event *event)
707 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
711 static inline void perf_cgroup_sched_out(struct task_struct *task,
712 struct task_struct *next)
716 static inline void perf_cgroup_sched_in(struct task_struct *prev,
717 struct task_struct *task)
721 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
722 struct perf_event_attr *attr,
723 struct perf_event *group_leader)
729 perf_cgroup_set_timestamp(struct task_struct *task,
730 struct perf_event_context *ctx)
735 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
740 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
744 static inline u64 perf_cgroup_event_time(struct perf_event *event)
750 perf_cgroup_defer_enabled(struct perf_event *event)
755 perf_cgroup_mark_enabled(struct perf_event *event,
756 struct perf_event_context *ctx)
762 * set default to be dependent on timer tick just
765 #define PERF_CPU_HRTIMER (1000 / HZ)
767 * function must be called with interrupts disbled
769 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
771 struct perf_cpu_context *cpuctx;
772 enum hrtimer_restart ret = HRTIMER_NORESTART;
775 WARN_ON(!irqs_disabled());
777 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
779 rotations = perf_rotate_context(cpuctx);
782 * arm timer if needed
785 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
786 ret = HRTIMER_RESTART;
792 /* CPU is going down */
793 void perf_cpu_hrtimer_cancel(int cpu)
795 struct perf_cpu_context *cpuctx;
799 if (WARN_ON(cpu != smp_processor_id()))
802 local_irq_save(flags);
806 list_for_each_entry_rcu(pmu, &pmus, entry) {
807 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
809 if (pmu->task_ctx_nr == perf_sw_context)
812 hrtimer_cancel(&cpuctx->hrtimer);
817 local_irq_restore(flags);
820 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
822 struct hrtimer *hr = &cpuctx->hrtimer;
823 struct pmu *pmu = cpuctx->ctx.pmu;
826 /* no multiplexing needed for SW PMU */
827 if (pmu->task_ctx_nr == perf_sw_context)
831 * check default is sane, if not set then force to
832 * default interval (1/tick)
834 timer = pmu->hrtimer_interval_ms;
836 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
838 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
840 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
841 hr->function = perf_cpu_hrtimer_handler;
844 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
846 struct hrtimer *hr = &cpuctx->hrtimer;
847 struct pmu *pmu = cpuctx->ctx.pmu;
850 if (pmu->task_ctx_nr == perf_sw_context)
853 if (hrtimer_active(hr))
856 if (!hrtimer_callback_running(hr))
857 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
858 0, HRTIMER_MODE_REL_PINNED, 0);
861 void perf_pmu_disable(struct pmu *pmu)
863 int *count = this_cpu_ptr(pmu->pmu_disable_count);
865 pmu->pmu_disable(pmu);
868 void perf_pmu_enable(struct pmu *pmu)
870 int *count = this_cpu_ptr(pmu->pmu_disable_count);
872 pmu->pmu_enable(pmu);
875 static DEFINE_PER_CPU(struct list_head, rotation_list);
878 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
879 * because they're strictly cpu affine and rotate_start is called with IRQs
880 * disabled, while rotate_context is called from IRQ context.
882 static void perf_pmu_rotate_start(struct pmu *pmu)
884 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
885 struct list_head *head = this_cpu_ptr(&rotation_list);
887 WARN_ON(!irqs_disabled());
889 if (list_empty(&cpuctx->rotation_list))
890 list_add(&cpuctx->rotation_list, head);
893 static void get_ctx(struct perf_event_context *ctx)
895 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
898 static void put_ctx(struct perf_event_context *ctx)
900 if (atomic_dec_and_test(&ctx->refcount)) {
902 put_ctx(ctx->parent_ctx);
904 put_task_struct(ctx->task);
905 kfree_rcu(ctx, rcu_head);
910 * This must be done under the ctx->lock, such as to serialize against
911 * context_equiv(), therefore we cannot call put_ctx() since that might end up
912 * calling scheduler related locks and ctx->lock nests inside those.
914 static __must_check struct perf_event_context *
915 unclone_ctx(struct perf_event_context *ctx)
917 struct perf_event_context *parent_ctx = ctx->parent_ctx;
919 lockdep_assert_held(&ctx->lock);
922 ctx->parent_ctx = NULL;
928 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
931 * only top level events have the pid namespace they were created in
934 event = event->parent;
936 return task_tgid_nr_ns(p, event->ns);
939 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
942 * only top level events have the pid namespace they were created in
945 event = event->parent;
947 return task_pid_nr_ns(p, event->ns);
951 * If we inherit events we want to return the parent event id
954 static u64 primary_event_id(struct perf_event *event)
959 id = event->parent->id;
965 * Get the perf_event_context for a task and lock it.
966 * This has to cope with with the fact that until it is locked,
967 * the context could get moved to another task.
969 static struct perf_event_context *
970 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
972 struct perf_event_context *ctx;
976 * One of the few rules of preemptible RCU is that one cannot do
977 * rcu_read_unlock() while holding a scheduler (or nested) lock when
978 * part of the read side critical section was preemptible -- see
979 * rcu_read_unlock_special().
981 * Since ctx->lock nests under rq->lock we must ensure the entire read
982 * side critical section is non-preemptible.
986 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
989 * If this context is a clone of another, it might
990 * get swapped for another underneath us by
991 * perf_event_task_sched_out, though the
992 * rcu_read_lock() protects us from any context
993 * getting freed. Lock the context and check if it
994 * got swapped before we could get the lock, and retry
995 * if so. If we locked the right context, then it
996 * can't get swapped on us any more.
998 raw_spin_lock_irqsave(&ctx->lock, *flags);
999 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1000 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1006 if (!atomic_inc_not_zero(&ctx->refcount)) {
1007 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1017 * Get the context for a task and increment its pin_count so it
1018 * can't get swapped to another task. This also increments its
1019 * reference count so that the context can't get freed.
1021 static struct perf_event_context *
1022 perf_pin_task_context(struct task_struct *task, int ctxn)
1024 struct perf_event_context *ctx;
1025 unsigned long flags;
1027 ctx = perf_lock_task_context(task, ctxn, &flags);
1030 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1035 static void perf_unpin_context(struct perf_event_context *ctx)
1037 unsigned long flags;
1039 raw_spin_lock_irqsave(&ctx->lock, flags);
1041 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1045 * Update the record of the current time in a context.
1047 static void update_context_time(struct perf_event_context *ctx)
1049 u64 now = perf_clock();
1051 ctx->time += now - ctx->timestamp;
1052 ctx->timestamp = now;
1055 static u64 perf_event_time(struct perf_event *event)
1057 struct perf_event_context *ctx = event->ctx;
1059 if (is_cgroup_event(event))
1060 return perf_cgroup_event_time(event);
1062 return ctx ? ctx->time : 0;
1066 * Update the total_time_enabled and total_time_running fields for a event.
1067 * The caller of this function needs to hold the ctx->lock.
1069 static void update_event_times(struct perf_event *event)
1071 struct perf_event_context *ctx = event->ctx;
1074 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1075 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1078 * in cgroup mode, time_enabled represents
1079 * the time the event was enabled AND active
1080 * tasks were in the monitored cgroup. This is
1081 * independent of the activity of the context as
1082 * there may be a mix of cgroup and non-cgroup events.
1084 * That is why we treat cgroup events differently
1087 if (is_cgroup_event(event))
1088 run_end = perf_cgroup_event_time(event);
1089 else if (ctx->is_active)
1090 run_end = ctx->time;
1092 run_end = event->tstamp_stopped;
1094 event->total_time_enabled = run_end - event->tstamp_enabled;
1096 if (event->state == PERF_EVENT_STATE_INACTIVE)
1097 run_end = event->tstamp_stopped;
1099 run_end = perf_event_time(event);
1101 event->total_time_running = run_end - event->tstamp_running;
1106 * Update total_time_enabled and total_time_running for all events in a group.
1108 static void update_group_times(struct perf_event *leader)
1110 struct perf_event *event;
1112 update_event_times(leader);
1113 list_for_each_entry(event, &leader->sibling_list, group_entry)
1114 update_event_times(event);
1117 static struct list_head *
1118 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1120 if (event->attr.pinned)
1121 return &ctx->pinned_groups;
1123 return &ctx->flexible_groups;
1127 * Add a event from the lists for its context.
1128 * Must be called with ctx->mutex and ctx->lock held.
1131 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1133 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1134 event->attach_state |= PERF_ATTACH_CONTEXT;
1137 * If we're a stand alone event or group leader, we go to the context
1138 * list, group events are kept attached to the group so that
1139 * perf_group_detach can, at all times, locate all siblings.
1141 if (event->group_leader == event) {
1142 struct list_head *list;
1144 if (is_software_event(event))
1145 event->group_flags |= PERF_GROUP_SOFTWARE;
1147 list = ctx_group_list(event, ctx);
1148 list_add_tail(&event->group_entry, list);
1151 if (is_cgroup_event(event))
1154 if (has_branch_stack(event))
1155 ctx->nr_branch_stack++;
1157 list_add_rcu(&event->event_entry, &ctx->event_list);
1158 if (!ctx->nr_events)
1159 perf_pmu_rotate_start(ctx->pmu);
1161 if (event->attr.inherit_stat)
1168 * Initialize event state based on the perf_event_attr::disabled.
1170 static inline void perf_event__state_init(struct perf_event *event)
1172 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1173 PERF_EVENT_STATE_INACTIVE;
1177 * Called at perf_event creation and when events are attached/detached from a
1180 static void perf_event__read_size(struct perf_event *event)
1182 int entry = sizeof(u64); /* value */
1186 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1187 size += sizeof(u64);
1189 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1190 size += sizeof(u64);
1192 if (event->attr.read_format & PERF_FORMAT_ID)
1193 entry += sizeof(u64);
1195 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1196 nr += event->group_leader->nr_siblings;
1197 size += sizeof(u64);
1201 event->read_size = size;
1204 static void perf_event__header_size(struct perf_event *event)
1206 struct perf_sample_data *data;
1207 u64 sample_type = event->attr.sample_type;
1210 perf_event__read_size(event);
1212 if (sample_type & PERF_SAMPLE_IP)
1213 size += sizeof(data->ip);
1215 if (sample_type & PERF_SAMPLE_ADDR)
1216 size += sizeof(data->addr);
1218 if (sample_type & PERF_SAMPLE_PERIOD)
1219 size += sizeof(data->period);
1221 if (sample_type & PERF_SAMPLE_WEIGHT)
1222 size += sizeof(data->weight);
1224 if (sample_type & PERF_SAMPLE_READ)
1225 size += event->read_size;
1227 if (sample_type & PERF_SAMPLE_DATA_SRC)
1228 size += sizeof(data->data_src.val);
1230 if (sample_type & PERF_SAMPLE_TRANSACTION)
1231 size += sizeof(data->txn);
1233 event->header_size = size;
1236 static void perf_event__id_header_size(struct perf_event *event)
1238 struct perf_sample_data *data;
1239 u64 sample_type = event->attr.sample_type;
1242 if (sample_type & PERF_SAMPLE_TID)
1243 size += sizeof(data->tid_entry);
1245 if (sample_type & PERF_SAMPLE_TIME)
1246 size += sizeof(data->time);
1248 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1249 size += sizeof(data->id);
1251 if (sample_type & PERF_SAMPLE_ID)
1252 size += sizeof(data->id);
1254 if (sample_type & PERF_SAMPLE_STREAM_ID)
1255 size += sizeof(data->stream_id);
1257 if (sample_type & PERF_SAMPLE_CPU)
1258 size += sizeof(data->cpu_entry);
1260 event->id_header_size = size;
1263 static void perf_group_attach(struct perf_event *event)
1265 struct perf_event *group_leader = event->group_leader, *pos;
1268 * We can have double attach due to group movement in perf_event_open.
1270 if (event->attach_state & PERF_ATTACH_GROUP)
1273 event->attach_state |= PERF_ATTACH_GROUP;
1275 if (group_leader == event)
1278 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1279 !is_software_event(event))
1280 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1282 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1283 group_leader->nr_siblings++;
1285 perf_event__header_size(group_leader);
1287 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1288 perf_event__header_size(pos);
1292 * Remove a event from the lists for its context.
1293 * Must be called with ctx->mutex and ctx->lock held.
1296 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1298 struct perf_cpu_context *cpuctx;
1300 * We can have double detach due to exit/hot-unplug + close.
1302 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1305 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1307 if (is_cgroup_event(event)) {
1309 cpuctx = __get_cpu_context(ctx);
1311 * if there are no more cgroup events
1312 * then cler cgrp to avoid stale pointer
1313 * in update_cgrp_time_from_cpuctx()
1315 if (!ctx->nr_cgroups)
1316 cpuctx->cgrp = NULL;
1319 if (has_branch_stack(event))
1320 ctx->nr_branch_stack--;
1323 if (event->attr.inherit_stat)
1326 list_del_rcu(&event->event_entry);
1328 if (event->group_leader == event)
1329 list_del_init(&event->group_entry);
1331 update_group_times(event);
1334 * If event was in error state, then keep it
1335 * that way, otherwise bogus counts will be
1336 * returned on read(). The only way to get out
1337 * of error state is by explicit re-enabling
1340 if (event->state > PERF_EVENT_STATE_OFF)
1341 event->state = PERF_EVENT_STATE_OFF;
1346 static void perf_group_detach(struct perf_event *event)
1348 struct perf_event *sibling, *tmp;
1349 struct list_head *list = NULL;
1352 * We can have double detach due to exit/hot-unplug + close.
1354 if (!(event->attach_state & PERF_ATTACH_GROUP))
1357 event->attach_state &= ~PERF_ATTACH_GROUP;
1360 * If this is a sibling, remove it from its group.
1362 if (event->group_leader != event) {
1363 list_del_init(&event->group_entry);
1364 event->group_leader->nr_siblings--;
1368 if (!list_empty(&event->group_entry))
1369 list = &event->group_entry;
1372 * If this was a group event with sibling events then
1373 * upgrade the siblings to singleton events by adding them
1374 * to whatever list we are on.
1376 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1378 list_move_tail(&sibling->group_entry, list);
1379 sibling->group_leader = sibling;
1381 /* Inherit group flags from the previous leader */
1382 sibling->group_flags = event->group_flags;
1386 perf_event__header_size(event->group_leader);
1388 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1389 perf_event__header_size(tmp);
1393 * User event without the task.
1395 static bool is_orphaned_event(struct perf_event *event)
1397 return event && !is_kernel_event(event) && !event->owner;
1401 * Event has a parent but parent's task finished and it's
1402 * alive only because of children holding refference.
1404 static bool is_orphaned_child(struct perf_event *event)
1406 return is_orphaned_event(event->parent);
1409 static void orphans_remove_work(struct work_struct *work);
1411 static void schedule_orphans_remove(struct perf_event_context *ctx)
1413 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1416 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1418 ctx->orphans_remove_sched = true;
1422 static int __init perf_workqueue_init(void)
1424 perf_wq = create_singlethread_workqueue("perf");
1425 WARN(!perf_wq, "failed to create perf workqueue\n");
1426 return perf_wq ? 0 : -1;
1429 core_initcall(perf_workqueue_init);
1432 event_filter_match(struct perf_event *event)
1434 return (event->cpu == -1 || event->cpu == smp_processor_id())
1435 && perf_cgroup_match(event);
1439 event_sched_out(struct perf_event *event,
1440 struct perf_cpu_context *cpuctx,
1441 struct perf_event_context *ctx)
1443 u64 tstamp = perf_event_time(event);
1446 * An event which could not be activated because of
1447 * filter mismatch still needs to have its timings
1448 * maintained, otherwise bogus information is return
1449 * via read() for time_enabled, time_running:
1451 if (event->state == PERF_EVENT_STATE_INACTIVE
1452 && !event_filter_match(event)) {
1453 delta = tstamp - event->tstamp_stopped;
1454 event->tstamp_running += delta;
1455 event->tstamp_stopped = tstamp;
1458 if (event->state != PERF_EVENT_STATE_ACTIVE)
1461 perf_pmu_disable(event->pmu);
1463 event->state = PERF_EVENT_STATE_INACTIVE;
1464 if (event->pending_disable) {
1465 event->pending_disable = 0;
1466 event->state = PERF_EVENT_STATE_OFF;
1468 event->tstamp_stopped = tstamp;
1469 event->pmu->del(event, 0);
1472 if (!is_software_event(event))
1473 cpuctx->active_oncpu--;
1475 if (event->attr.freq && event->attr.sample_freq)
1477 if (event->attr.exclusive || !cpuctx->active_oncpu)
1478 cpuctx->exclusive = 0;
1480 if (is_orphaned_child(event))
1481 schedule_orphans_remove(ctx);
1483 perf_pmu_enable(event->pmu);
1487 group_sched_out(struct perf_event *group_event,
1488 struct perf_cpu_context *cpuctx,
1489 struct perf_event_context *ctx)
1491 struct perf_event *event;
1492 int state = group_event->state;
1494 event_sched_out(group_event, cpuctx, ctx);
1497 * Schedule out siblings (if any):
1499 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1500 event_sched_out(event, cpuctx, ctx);
1502 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1503 cpuctx->exclusive = 0;
1506 struct remove_event {
1507 struct perf_event *event;
1512 * Cross CPU call to remove a performance event
1514 * We disable the event on the hardware level first. After that we
1515 * remove it from the context list.
1517 static int __perf_remove_from_context(void *info)
1519 struct remove_event *re = info;
1520 struct perf_event *event = re->event;
1521 struct perf_event_context *ctx = event->ctx;
1522 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1524 raw_spin_lock(&ctx->lock);
1525 event_sched_out(event, cpuctx, ctx);
1526 if (re->detach_group)
1527 perf_group_detach(event);
1528 list_del_event(event, ctx);
1529 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1531 cpuctx->task_ctx = NULL;
1533 raw_spin_unlock(&ctx->lock);
1540 * Remove the event from a task's (or a CPU's) list of events.
1542 * CPU events are removed with a smp call. For task events we only
1543 * call when the task is on a CPU.
1545 * If event->ctx is a cloned context, callers must make sure that
1546 * every task struct that event->ctx->task could possibly point to
1547 * remains valid. This is OK when called from perf_release since
1548 * that only calls us on the top-level context, which can't be a clone.
1549 * When called from perf_event_exit_task, it's OK because the
1550 * context has been detached from its task.
1552 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1554 struct perf_event_context *ctx = event->ctx;
1555 struct task_struct *task = ctx->task;
1556 struct remove_event re = {
1558 .detach_group = detach_group,
1561 lockdep_assert_held(&ctx->mutex);
1565 * Per cpu events are removed via an smp call. The removal can
1566 * fail if the CPU is currently offline, but in that case we
1567 * already called __perf_remove_from_context from
1568 * perf_event_exit_cpu.
1570 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1575 if (!task_function_call(task, __perf_remove_from_context, &re))
1578 raw_spin_lock_irq(&ctx->lock);
1580 * If we failed to find a running task, but find the context active now
1581 * that we've acquired the ctx->lock, retry.
1583 if (ctx->is_active) {
1584 raw_spin_unlock_irq(&ctx->lock);
1586 * Reload the task pointer, it might have been changed by
1587 * a concurrent perf_event_context_sched_out().
1594 * Since the task isn't running, its safe to remove the event, us
1595 * holding the ctx->lock ensures the task won't get scheduled in.
1598 perf_group_detach(event);
1599 list_del_event(event, ctx);
1600 raw_spin_unlock_irq(&ctx->lock);
1604 * Cross CPU call to disable a performance event
1606 int __perf_event_disable(void *info)
1608 struct perf_event *event = info;
1609 struct perf_event_context *ctx = event->ctx;
1610 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1613 * If this is a per-task event, need to check whether this
1614 * event's task is the current task on this cpu.
1616 * Can trigger due to concurrent perf_event_context_sched_out()
1617 * flipping contexts around.
1619 if (ctx->task && cpuctx->task_ctx != ctx)
1622 raw_spin_lock(&ctx->lock);
1625 * If the event is on, turn it off.
1626 * If it is in error state, leave it in error state.
1628 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1629 update_context_time(ctx);
1630 update_cgrp_time_from_event(event);
1631 update_group_times(event);
1632 if (event == event->group_leader)
1633 group_sched_out(event, cpuctx, ctx);
1635 event_sched_out(event, cpuctx, ctx);
1636 event->state = PERF_EVENT_STATE_OFF;
1639 raw_spin_unlock(&ctx->lock);
1647 * If event->ctx is a cloned context, callers must make sure that
1648 * every task struct that event->ctx->task could possibly point to
1649 * remains valid. This condition is satisifed when called through
1650 * perf_event_for_each_child or perf_event_for_each because they
1651 * hold the top-level event's child_mutex, so any descendant that
1652 * goes to exit will block in sync_child_event.
1653 * When called from perf_pending_event it's OK because event->ctx
1654 * is the current context on this CPU and preemption is disabled,
1655 * hence we can't get into perf_event_task_sched_out for this context.
1657 void perf_event_disable(struct perf_event *event)
1659 struct perf_event_context *ctx = event->ctx;
1660 struct task_struct *task = ctx->task;
1664 * Disable the event on the cpu that it's on
1666 cpu_function_call(event->cpu, __perf_event_disable, event);
1671 if (!task_function_call(task, __perf_event_disable, event))
1674 raw_spin_lock_irq(&ctx->lock);
1676 * If the event is still active, we need to retry the cross-call.
1678 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1679 raw_spin_unlock_irq(&ctx->lock);
1681 * Reload the task pointer, it might have been changed by
1682 * a concurrent perf_event_context_sched_out().
1689 * Since we have the lock this context can't be scheduled
1690 * in, so we can change the state safely.
1692 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1693 update_group_times(event);
1694 event->state = PERF_EVENT_STATE_OFF;
1696 raw_spin_unlock_irq(&ctx->lock);
1698 EXPORT_SYMBOL_GPL(perf_event_disable);
1700 static void perf_set_shadow_time(struct perf_event *event,
1701 struct perf_event_context *ctx,
1705 * use the correct time source for the time snapshot
1707 * We could get by without this by leveraging the
1708 * fact that to get to this function, the caller
1709 * has most likely already called update_context_time()
1710 * and update_cgrp_time_xx() and thus both timestamp
1711 * are identical (or very close). Given that tstamp is,
1712 * already adjusted for cgroup, we could say that:
1713 * tstamp - ctx->timestamp
1715 * tstamp - cgrp->timestamp.
1717 * Then, in perf_output_read(), the calculation would
1718 * work with no changes because:
1719 * - event is guaranteed scheduled in
1720 * - no scheduled out in between
1721 * - thus the timestamp would be the same
1723 * But this is a bit hairy.
1725 * So instead, we have an explicit cgroup call to remain
1726 * within the time time source all along. We believe it
1727 * is cleaner and simpler to understand.
1729 if (is_cgroup_event(event))
1730 perf_cgroup_set_shadow_time(event, tstamp);
1732 event->shadow_ctx_time = tstamp - ctx->timestamp;
1735 #define MAX_INTERRUPTS (~0ULL)
1737 static void perf_log_throttle(struct perf_event *event, int enable);
1740 event_sched_in(struct perf_event *event,
1741 struct perf_cpu_context *cpuctx,
1742 struct perf_event_context *ctx)
1744 u64 tstamp = perf_event_time(event);
1747 lockdep_assert_held(&ctx->lock);
1749 if (event->state <= PERF_EVENT_STATE_OFF)
1752 event->state = PERF_EVENT_STATE_ACTIVE;
1753 event->oncpu = smp_processor_id();
1756 * Unthrottle events, since we scheduled we might have missed several
1757 * ticks already, also for a heavily scheduling task there is little
1758 * guarantee it'll get a tick in a timely manner.
1760 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1761 perf_log_throttle(event, 1);
1762 event->hw.interrupts = 0;
1766 * The new state must be visible before we turn it on in the hardware:
1770 perf_pmu_disable(event->pmu);
1772 if (event->pmu->add(event, PERF_EF_START)) {
1773 event->state = PERF_EVENT_STATE_INACTIVE;
1779 event->tstamp_running += tstamp - event->tstamp_stopped;
1781 perf_set_shadow_time(event, ctx, tstamp);
1783 if (!is_software_event(event))
1784 cpuctx->active_oncpu++;
1786 if (event->attr.freq && event->attr.sample_freq)
1789 if (event->attr.exclusive)
1790 cpuctx->exclusive = 1;
1792 if (is_orphaned_child(event))
1793 schedule_orphans_remove(ctx);
1796 perf_pmu_enable(event->pmu);
1802 group_sched_in(struct perf_event *group_event,
1803 struct perf_cpu_context *cpuctx,
1804 struct perf_event_context *ctx)
1806 struct perf_event *event, *partial_group = NULL;
1807 struct pmu *pmu = ctx->pmu;
1808 u64 now = ctx->time;
1809 bool simulate = false;
1811 if (group_event->state == PERF_EVENT_STATE_OFF)
1814 pmu->start_txn(pmu);
1816 if (event_sched_in(group_event, cpuctx, ctx)) {
1817 pmu->cancel_txn(pmu);
1818 perf_cpu_hrtimer_restart(cpuctx);
1823 * Schedule in siblings as one group (if any):
1825 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1826 if (event_sched_in(event, cpuctx, ctx)) {
1827 partial_group = event;
1832 if (!pmu->commit_txn(pmu))
1837 * Groups can be scheduled in as one unit only, so undo any
1838 * partial group before returning:
1839 * The events up to the failed event are scheduled out normally,
1840 * tstamp_stopped will be updated.
1842 * The failed events and the remaining siblings need to have
1843 * their timings updated as if they had gone thru event_sched_in()
1844 * and event_sched_out(). This is required to get consistent timings
1845 * across the group. This also takes care of the case where the group
1846 * could never be scheduled by ensuring tstamp_stopped is set to mark
1847 * the time the event was actually stopped, such that time delta
1848 * calculation in update_event_times() is correct.
1850 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1851 if (event == partial_group)
1855 event->tstamp_running += now - event->tstamp_stopped;
1856 event->tstamp_stopped = now;
1858 event_sched_out(event, cpuctx, ctx);
1861 event_sched_out(group_event, cpuctx, ctx);
1863 pmu->cancel_txn(pmu);
1865 perf_cpu_hrtimer_restart(cpuctx);
1871 * Work out whether we can put this event group on the CPU now.
1873 static int group_can_go_on(struct perf_event *event,
1874 struct perf_cpu_context *cpuctx,
1878 * Groups consisting entirely of software events can always go on.
1880 if (event->group_flags & PERF_GROUP_SOFTWARE)
1883 * If an exclusive group is already on, no other hardware
1886 if (cpuctx->exclusive)
1889 * If this group is exclusive and there are already
1890 * events on the CPU, it can't go on.
1892 if (event->attr.exclusive && cpuctx->active_oncpu)
1895 * Otherwise, try to add it if all previous groups were able
1901 static void add_event_to_ctx(struct perf_event *event,
1902 struct perf_event_context *ctx)
1904 u64 tstamp = perf_event_time(event);
1906 list_add_event(event, ctx);
1907 perf_group_attach(event);
1908 event->tstamp_enabled = tstamp;
1909 event->tstamp_running = tstamp;
1910 event->tstamp_stopped = tstamp;
1913 static void task_ctx_sched_out(struct perf_event_context *ctx);
1915 ctx_sched_in(struct perf_event_context *ctx,
1916 struct perf_cpu_context *cpuctx,
1917 enum event_type_t event_type,
1918 struct task_struct *task);
1920 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1921 struct perf_event_context *ctx,
1922 struct task_struct *task)
1924 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1926 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1927 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1929 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1933 * Cross CPU call to install and enable a performance event
1935 * Must be called with ctx->mutex held
1937 static int __perf_install_in_context(void *info)
1939 struct perf_event *event = info;
1940 struct perf_event_context *ctx = event->ctx;
1941 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1942 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1943 struct task_struct *task = current;
1945 perf_ctx_lock(cpuctx, task_ctx);
1946 perf_pmu_disable(cpuctx->ctx.pmu);
1949 * If there was an active task_ctx schedule it out.
1952 task_ctx_sched_out(task_ctx);
1955 * If the context we're installing events in is not the
1956 * active task_ctx, flip them.
1958 if (ctx->task && task_ctx != ctx) {
1960 raw_spin_unlock(&task_ctx->lock);
1961 raw_spin_lock(&ctx->lock);
1966 cpuctx->task_ctx = task_ctx;
1967 task = task_ctx->task;
1970 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1972 update_context_time(ctx);
1974 * update cgrp time only if current cgrp
1975 * matches event->cgrp. Must be done before
1976 * calling add_event_to_ctx()
1978 update_cgrp_time_from_event(event);
1980 add_event_to_ctx(event, ctx);
1983 * Schedule everything back in
1985 perf_event_sched_in(cpuctx, task_ctx, task);
1987 perf_pmu_enable(cpuctx->ctx.pmu);
1988 perf_ctx_unlock(cpuctx, task_ctx);
1994 * Attach a performance event to a context
1996 * First we add the event to the list with the hardware enable bit
1997 * in event->hw_config cleared.
1999 * If the event is attached to a task which is on a CPU we use a smp
2000 * call to enable it in the task context. The task might have been
2001 * scheduled away, but we check this in the smp call again.
2004 perf_install_in_context(struct perf_event_context *ctx,
2005 struct perf_event *event,
2008 struct task_struct *task = ctx->task;
2010 lockdep_assert_held(&ctx->mutex);
2013 if (event->cpu != -1)
2018 * Per cpu events are installed via an smp call and
2019 * the install is always successful.
2021 cpu_function_call(cpu, __perf_install_in_context, event);
2026 if (!task_function_call(task, __perf_install_in_context, event))
2029 raw_spin_lock_irq(&ctx->lock);
2031 * If we failed to find a running task, but find the context active now
2032 * that we've acquired the ctx->lock, retry.
2034 if (ctx->is_active) {
2035 raw_spin_unlock_irq(&ctx->lock);
2037 * Reload the task pointer, it might have been changed by
2038 * a concurrent perf_event_context_sched_out().
2045 * Since the task isn't running, its safe to add the event, us holding
2046 * the ctx->lock ensures the task won't get scheduled in.
2048 add_event_to_ctx(event, ctx);
2049 raw_spin_unlock_irq(&ctx->lock);
2053 * Put a event into inactive state and update time fields.
2054 * Enabling the leader of a group effectively enables all
2055 * the group members that aren't explicitly disabled, so we
2056 * have to update their ->tstamp_enabled also.
2057 * Note: this works for group members as well as group leaders
2058 * since the non-leader members' sibling_lists will be empty.
2060 static void __perf_event_mark_enabled(struct perf_event *event)
2062 struct perf_event *sub;
2063 u64 tstamp = perf_event_time(event);
2065 event->state = PERF_EVENT_STATE_INACTIVE;
2066 event->tstamp_enabled = tstamp - event->total_time_enabled;
2067 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2068 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2069 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2074 * Cross CPU call to enable a performance event
2076 static int __perf_event_enable(void *info)
2078 struct perf_event *event = info;
2079 struct perf_event_context *ctx = event->ctx;
2080 struct perf_event *leader = event->group_leader;
2081 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2085 * There's a time window between 'ctx->is_active' check
2086 * in perf_event_enable function and this place having:
2088 * - ctx->lock unlocked
2090 * where the task could be killed and 'ctx' deactivated
2091 * by perf_event_exit_task.
2093 if (!ctx->is_active)
2096 raw_spin_lock(&ctx->lock);
2097 update_context_time(ctx);
2099 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2103 * set current task's cgroup time reference point
2105 perf_cgroup_set_timestamp(current, ctx);
2107 __perf_event_mark_enabled(event);
2109 if (!event_filter_match(event)) {
2110 if (is_cgroup_event(event))
2111 perf_cgroup_defer_enabled(event);
2116 * If the event is in a group and isn't the group leader,
2117 * then don't put it on unless the group is on.
2119 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2122 if (!group_can_go_on(event, cpuctx, 1)) {
2125 if (event == leader)
2126 err = group_sched_in(event, cpuctx, ctx);
2128 err = event_sched_in(event, cpuctx, ctx);
2133 * If this event can't go on and it's part of a
2134 * group, then the whole group has to come off.
2136 if (leader != event) {
2137 group_sched_out(leader, cpuctx, ctx);
2138 perf_cpu_hrtimer_restart(cpuctx);
2140 if (leader->attr.pinned) {
2141 update_group_times(leader);
2142 leader->state = PERF_EVENT_STATE_ERROR;
2147 raw_spin_unlock(&ctx->lock);
2155 * If event->ctx is a cloned context, callers must make sure that
2156 * every task struct that event->ctx->task could possibly point to
2157 * remains valid. This condition is satisfied when called through
2158 * perf_event_for_each_child or perf_event_for_each as described
2159 * for perf_event_disable.
2161 void perf_event_enable(struct perf_event *event)
2163 struct perf_event_context *ctx = event->ctx;
2164 struct task_struct *task = ctx->task;
2168 * Enable the event on the cpu that it's on
2170 cpu_function_call(event->cpu, __perf_event_enable, event);
2174 raw_spin_lock_irq(&ctx->lock);
2175 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2179 * If the event is in error state, clear that first.
2180 * That way, if we see the event in error state below, we
2181 * know that it has gone back into error state, as distinct
2182 * from the task having been scheduled away before the
2183 * cross-call arrived.
2185 if (event->state == PERF_EVENT_STATE_ERROR)
2186 event->state = PERF_EVENT_STATE_OFF;
2189 if (!ctx->is_active) {
2190 __perf_event_mark_enabled(event);
2194 raw_spin_unlock_irq(&ctx->lock);
2196 if (!task_function_call(task, __perf_event_enable, event))
2199 raw_spin_lock_irq(&ctx->lock);
2202 * If the context is active and the event is still off,
2203 * we need to retry the cross-call.
2205 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2207 * task could have been flipped by a concurrent
2208 * perf_event_context_sched_out()
2215 raw_spin_unlock_irq(&ctx->lock);
2217 EXPORT_SYMBOL_GPL(perf_event_enable);
2219 int perf_event_refresh(struct perf_event *event, int refresh)
2222 * not supported on inherited events
2224 if (event->attr.inherit || !is_sampling_event(event))
2227 atomic_add(refresh, &event->event_limit);
2228 perf_event_enable(event);
2232 EXPORT_SYMBOL_GPL(perf_event_refresh);
2234 static void ctx_sched_out(struct perf_event_context *ctx,
2235 struct perf_cpu_context *cpuctx,
2236 enum event_type_t event_type)
2238 struct perf_event *event;
2239 int is_active = ctx->is_active;
2241 ctx->is_active &= ~event_type;
2242 if (likely(!ctx->nr_events))
2245 update_context_time(ctx);
2246 update_cgrp_time_from_cpuctx(cpuctx);
2247 if (!ctx->nr_active)
2250 perf_pmu_disable(ctx->pmu);
2251 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2252 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2253 group_sched_out(event, cpuctx, ctx);
2256 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2257 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2258 group_sched_out(event, cpuctx, ctx);
2260 perf_pmu_enable(ctx->pmu);
2264 * Test whether two contexts are equivalent, i.e. whether they have both been
2265 * cloned from the same version of the same context.
2267 * Equivalence is measured using a generation number in the context that is
2268 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2269 * and list_del_event().
2271 static int context_equiv(struct perf_event_context *ctx1,
2272 struct perf_event_context *ctx2)
2274 lockdep_assert_held(&ctx1->lock);
2275 lockdep_assert_held(&ctx2->lock);
2277 /* Pinning disables the swap optimization */
2278 if (ctx1->pin_count || ctx2->pin_count)
2281 /* If ctx1 is the parent of ctx2 */
2282 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2285 /* If ctx2 is the parent of ctx1 */
2286 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2290 * If ctx1 and ctx2 have the same parent; we flatten the parent
2291 * hierarchy, see perf_event_init_context().
2293 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2294 ctx1->parent_gen == ctx2->parent_gen)
2301 static void __perf_event_sync_stat(struct perf_event *event,
2302 struct perf_event *next_event)
2306 if (!event->attr.inherit_stat)
2310 * Update the event value, we cannot use perf_event_read()
2311 * because we're in the middle of a context switch and have IRQs
2312 * disabled, which upsets smp_call_function_single(), however
2313 * we know the event must be on the current CPU, therefore we
2314 * don't need to use it.
2316 switch (event->state) {
2317 case PERF_EVENT_STATE_ACTIVE:
2318 event->pmu->read(event);
2321 case PERF_EVENT_STATE_INACTIVE:
2322 update_event_times(event);
2330 * In order to keep per-task stats reliable we need to flip the event
2331 * values when we flip the contexts.
2333 value = local64_read(&next_event->count);
2334 value = local64_xchg(&event->count, value);
2335 local64_set(&next_event->count, value);
2337 swap(event->total_time_enabled, next_event->total_time_enabled);
2338 swap(event->total_time_running, next_event->total_time_running);
2341 * Since we swizzled the values, update the user visible data too.
2343 perf_event_update_userpage(event);
2344 perf_event_update_userpage(next_event);
2347 static void perf_event_sync_stat(struct perf_event_context *ctx,
2348 struct perf_event_context *next_ctx)
2350 struct perf_event *event, *next_event;
2355 update_context_time(ctx);
2357 event = list_first_entry(&ctx->event_list,
2358 struct perf_event, event_entry);
2360 next_event = list_first_entry(&next_ctx->event_list,
2361 struct perf_event, event_entry);
2363 while (&event->event_entry != &ctx->event_list &&
2364 &next_event->event_entry != &next_ctx->event_list) {
2366 __perf_event_sync_stat(event, next_event);
2368 event = list_next_entry(event, event_entry);
2369 next_event = list_next_entry(next_event, event_entry);
2373 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2374 struct task_struct *next)
2376 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2377 struct perf_event_context *next_ctx;
2378 struct perf_event_context *parent, *next_parent;
2379 struct perf_cpu_context *cpuctx;
2385 cpuctx = __get_cpu_context(ctx);
2386 if (!cpuctx->task_ctx)
2390 next_ctx = next->perf_event_ctxp[ctxn];
2394 parent = rcu_dereference(ctx->parent_ctx);
2395 next_parent = rcu_dereference(next_ctx->parent_ctx);
2397 /* If neither context have a parent context; they cannot be clones. */
2398 if (!parent && !next_parent)
2401 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2403 * Looks like the two contexts are clones, so we might be
2404 * able to optimize the context switch. We lock both
2405 * contexts and check that they are clones under the
2406 * lock (including re-checking that neither has been
2407 * uncloned in the meantime). It doesn't matter which
2408 * order we take the locks because no other cpu could
2409 * be trying to lock both of these tasks.
2411 raw_spin_lock(&ctx->lock);
2412 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2413 if (context_equiv(ctx, next_ctx)) {
2415 * XXX do we need a memory barrier of sorts
2416 * wrt to rcu_dereference() of perf_event_ctxp
2418 task->perf_event_ctxp[ctxn] = next_ctx;
2419 next->perf_event_ctxp[ctxn] = ctx;
2421 next_ctx->task = task;
2424 perf_event_sync_stat(ctx, next_ctx);
2426 raw_spin_unlock(&next_ctx->lock);
2427 raw_spin_unlock(&ctx->lock);
2433 raw_spin_lock(&ctx->lock);
2434 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2435 cpuctx->task_ctx = NULL;
2436 raw_spin_unlock(&ctx->lock);
2440 #define for_each_task_context_nr(ctxn) \
2441 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2444 * Called from scheduler to remove the events of the current task,
2445 * with interrupts disabled.
2447 * We stop each event and update the event value in event->count.
2449 * This does not protect us against NMI, but disable()
2450 * sets the disabled bit in the control field of event _before_
2451 * accessing the event control register. If a NMI hits, then it will
2452 * not restart the event.
2454 void __perf_event_task_sched_out(struct task_struct *task,
2455 struct task_struct *next)
2459 for_each_task_context_nr(ctxn)
2460 perf_event_context_sched_out(task, ctxn, next);
2463 * if cgroup events exist on this CPU, then we need
2464 * to check if we have to switch out PMU state.
2465 * cgroup event are system-wide mode only
2467 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2468 perf_cgroup_sched_out(task, next);
2471 static void task_ctx_sched_out(struct perf_event_context *ctx)
2473 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2475 if (!cpuctx->task_ctx)
2478 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2481 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2482 cpuctx->task_ctx = NULL;
2486 * Called with IRQs disabled
2488 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2489 enum event_type_t event_type)
2491 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2495 ctx_pinned_sched_in(struct perf_event_context *ctx,
2496 struct perf_cpu_context *cpuctx)
2498 struct perf_event *event;
2500 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2501 if (event->state <= PERF_EVENT_STATE_OFF)
2503 if (!event_filter_match(event))
2506 /* may need to reset tstamp_enabled */
2507 if (is_cgroup_event(event))
2508 perf_cgroup_mark_enabled(event, ctx);
2510 if (group_can_go_on(event, cpuctx, 1))
2511 group_sched_in(event, cpuctx, ctx);
2514 * If this pinned group hasn't been scheduled,
2515 * put it in error state.
2517 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2518 update_group_times(event);
2519 event->state = PERF_EVENT_STATE_ERROR;
2525 ctx_flexible_sched_in(struct perf_event_context *ctx,
2526 struct perf_cpu_context *cpuctx)
2528 struct perf_event *event;
2531 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2532 /* Ignore events in OFF or ERROR state */
2533 if (event->state <= PERF_EVENT_STATE_OFF)
2536 * Listen to the 'cpu' scheduling filter constraint
2539 if (!event_filter_match(event))
2542 /* may need to reset tstamp_enabled */
2543 if (is_cgroup_event(event))
2544 perf_cgroup_mark_enabled(event, ctx);
2546 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2547 if (group_sched_in(event, cpuctx, ctx))
2554 ctx_sched_in(struct perf_event_context *ctx,
2555 struct perf_cpu_context *cpuctx,
2556 enum event_type_t event_type,
2557 struct task_struct *task)
2560 int is_active = ctx->is_active;
2562 ctx->is_active |= event_type;
2563 if (likely(!ctx->nr_events))
2567 ctx->timestamp = now;
2568 perf_cgroup_set_timestamp(task, ctx);
2570 * First go through the list and put on any pinned groups
2571 * in order to give them the best chance of going on.
2573 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2574 ctx_pinned_sched_in(ctx, cpuctx);
2576 /* Then walk through the lower prio flexible groups */
2577 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2578 ctx_flexible_sched_in(ctx, cpuctx);
2581 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2582 enum event_type_t event_type,
2583 struct task_struct *task)
2585 struct perf_event_context *ctx = &cpuctx->ctx;
2587 ctx_sched_in(ctx, cpuctx, event_type, task);
2590 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2591 struct task_struct *task)
2593 struct perf_cpu_context *cpuctx;
2595 cpuctx = __get_cpu_context(ctx);
2596 if (cpuctx->task_ctx == ctx)
2599 perf_ctx_lock(cpuctx, ctx);
2600 perf_pmu_disable(ctx->pmu);
2602 * We want to keep the following priority order:
2603 * cpu pinned (that don't need to move), task pinned,
2604 * cpu flexible, task flexible.
2606 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2609 cpuctx->task_ctx = ctx;
2611 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2613 perf_pmu_enable(ctx->pmu);
2614 perf_ctx_unlock(cpuctx, ctx);
2617 * Since these rotations are per-cpu, we need to ensure the
2618 * cpu-context we got scheduled on is actually rotating.
2620 perf_pmu_rotate_start(ctx->pmu);
2624 * When sampling the branck stack in system-wide, it may be necessary
2625 * to flush the stack on context switch. This happens when the branch
2626 * stack does not tag its entries with the pid of the current task.
2627 * Otherwise it becomes impossible to associate a branch entry with a
2628 * task. This ambiguity is more likely to appear when the branch stack
2629 * supports priv level filtering and the user sets it to monitor only
2630 * at the user level (which could be a useful measurement in system-wide
2631 * mode). In that case, the risk is high of having a branch stack with
2632 * branch from multiple tasks. Flushing may mean dropping the existing
2633 * entries or stashing them somewhere in the PMU specific code layer.
2635 * This function provides the context switch callback to the lower code
2636 * layer. It is invoked ONLY when there is at least one system-wide context
2637 * with at least one active event using taken branch sampling.
2639 static void perf_branch_stack_sched_in(struct task_struct *prev,
2640 struct task_struct *task)
2642 struct perf_cpu_context *cpuctx;
2644 unsigned long flags;
2646 /* no need to flush branch stack if not changing task */
2650 local_irq_save(flags);
2654 list_for_each_entry_rcu(pmu, &pmus, entry) {
2655 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2658 * check if the context has at least one
2659 * event using PERF_SAMPLE_BRANCH_STACK
2661 if (cpuctx->ctx.nr_branch_stack > 0
2662 && pmu->flush_branch_stack) {
2664 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2666 perf_pmu_disable(pmu);
2668 pmu->flush_branch_stack();
2670 perf_pmu_enable(pmu);
2672 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2678 local_irq_restore(flags);
2682 * Called from scheduler to add the events of the current task
2683 * with interrupts disabled.
2685 * We restore the event value and then enable it.
2687 * This does not protect us against NMI, but enable()
2688 * sets the enabled bit in the control field of event _before_
2689 * accessing the event control register. If a NMI hits, then it will
2690 * keep the event running.
2692 void __perf_event_task_sched_in(struct task_struct *prev,
2693 struct task_struct *task)
2695 struct perf_event_context *ctx;
2698 for_each_task_context_nr(ctxn) {
2699 ctx = task->perf_event_ctxp[ctxn];
2703 perf_event_context_sched_in(ctx, task);
2706 * if cgroup events exist on this CPU, then we need
2707 * to check if we have to switch in PMU state.
2708 * cgroup event are system-wide mode only
2710 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2711 perf_cgroup_sched_in(prev, task);
2713 /* check for system-wide branch_stack events */
2714 if (atomic_read(this_cpu_ptr(&perf_branch_stack_events)))
2715 perf_branch_stack_sched_in(prev, task);
2718 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2720 u64 frequency = event->attr.sample_freq;
2721 u64 sec = NSEC_PER_SEC;
2722 u64 divisor, dividend;
2724 int count_fls, nsec_fls, frequency_fls, sec_fls;
2726 count_fls = fls64(count);
2727 nsec_fls = fls64(nsec);
2728 frequency_fls = fls64(frequency);
2732 * We got @count in @nsec, with a target of sample_freq HZ
2733 * the target period becomes:
2736 * period = -------------------
2737 * @nsec * sample_freq
2742 * Reduce accuracy by one bit such that @a and @b converge
2743 * to a similar magnitude.
2745 #define REDUCE_FLS(a, b) \
2747 if (a##_fls > b##_fls) { \
2757 * Reduce accuracy until either term fits in a u64, then proceed with
2758 * the other, so that finally we can do a u64/u64 division.
2760 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2761 REDUCE_FLS(nsec, frequency);
2762 REDUCE_FLS(sec, count);
2765 if (count_fls + sec_fls > 64) {
2766 divisor = nsec * frequency;
2768 while (count_fls + sec_fls > 64) {
2769 REDUCE_FLS(count, sec);
2773 dividend = count * sec;
2775 dividend = count * sec;
2777 while (nsec_fls + frequency_fls > 64) {
2778 REDUCE_FLS(nsec, frequency);
2782 divisor = nsec * frequency;
2788 return div64_u64(dividend, divisor);
2791 static DEFINE_PER_CPU(int, perf_throttled_count);
2792 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2794 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2796 struct hw_perf_event *hwc = &event->hw;
2797 s64 period, sample_period;
2800 period = perf_calculate_period(event, nsec, count);
2802 delta = (s64)(period - hwc->sample_period);
2803 delta = (delta + 7) / 8; /* low pass filter */
2805 sample_period = hwc->sample_period + delta;
2810 hwc->sample_period = sample_period;
2812 if (local64_read(&hwc->period_left) > 8*sample_period) {
2814 event->pmu->stop(event, PERF_EF_UPDATE);
2816 local64_set(&hwc->period_left, 0);
2819 event->pmu->start(event, PERF_EF_RELOAD);
2824 * combine freq adjustment with unthrottling to avoid two passes over the
2825 * events. At the same time, make sure, having freq events does not change
2826 * the rate of unthrottling as that would introduce bias.
2828 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2831 struct perf_event *event;
2832 struct hw_perf_event *hwc;
2833 u64 now, period = TICK_NSEC;
2837 * only need to iterate over all events iff:
2838 * - context have events in frequency mode (needs freq adjust)
2839 * - there are events to unthrottle on this cpu
2841 if (!(ctx->nr_freq || needs_unthr))
2844 raw_spin_lock(&ctx->lock);
2845 perf_pmu_disable(ctx->pmu);
2847 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2848 if (event->state != PERF_EVENT_STATE_ACTIVE)
2851 if (!event_filter_match(event))
2854 perf_pmu_disable(event->pmu);
2858 if (hwc->interrupts == MAX_INTERRUPTS) {
2859 hwc->interrupts = 0;
2860 perf_log_throttle(event, 1);
2861 event->pmu->start(event, 0);
2864 if (!event->attr.freq || !event->attr.sample_freq)
2868 * stop the event and update event->count
2870 event->pmu->stop(event, PERF_EF_UPDATE);
2872 now = local64_read(&event->count);
2873 delta = now - hwc->freq_count_stamp;
2874 hwc->freq_count_stamp = now;
2878 * reload only if value has changed
2879 * we have stopped the event so tell that
2880 * to perf_adjust_period() to avoid stopping it
2884 perf_adjust_period(event, period, delta, false);
2886 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2888 perf_pmu_enable(event->pmu);
2891 perf_pmu_enable(ctx->pmu);
2892 raw_spin_unlock(&ctx->lock);
2896 * Round-robin a context's events:
2898 static void rotate_ctx(struct perf_event_context *ctx)
2901 * Rotate the first entry last of non-pinned groups. Rotation might be
2902 * disabled by the inheritance code.
2904 if (!ctx->rotate_disable)
2905 list_rotate_left(&ctx->flexible_groups);
2909 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2910 * because they're strictly cpu affine and rotate_start is called with IRQs
2911 * disabled, while rotate_context is called from IRQ context.
2913 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2915 struct perf_event_context *ctx = NULL;
2916 int rotate = 0, remove = 1;
2918 if (cpuctx->ctx.nr_events) {
2920 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2924 ctx = cpuctx->task_ctx;
2925 if (ctx && ctx->nr_events) {
2927 if (ctx->nr_events != ctx->nr_active)
2934 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2935 perf_pmu_disable(cpuctx->ctx.pmu);
2937 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2939 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2941 rotate_ctx(&cpuctx->ctx);
2945 perf_event_sched_in(cpuctx, ctx, current);
2947 perf_pmu_enable(cpuctx->ctx.pmu);
2948 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2951 list_del_init(&cpuctx->rotation_list);
2956 #ifdef CONFIG_NO_HZ_FULL
2957 bool perf_event_can_stop_tick(void)
2959 if (atomic_read(&nr_freq_events) ||
2960 __this_cpu_read(perf_throttled_count))
2967 void perf_event_task_tick(void)
2969 struct list_head *head = this_cpu_ptr(&rotation_list);
2970 struct perf_cpu_context *cpuctx, *tmp;
2971 struct perf_event_context *ctx;
2974 WARN_ON(!irqs_disabled());
2976 __this_cpu_inc(perf_throttled_seq);
2977 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2979 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2981 perf_adjust_freq_unthr_context(ctx, throttled);
2983 ctx = cpuctx->task_ctx;
2985 perf_adjust_freq_unthr_context(ctx, throttled);
2989 static int event_enable_on_exec(struct perf_event *event,
2990 struct perf_event_context *ctx)
2992 if (!event->attr.enable_on_exec)
2995 event->attr.enable_on_exec = 0;
2996 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2999 __perf_event_mark_enabled(event);
3005 * Enable all of a task's events that have been marked enable-on-exec.
3006 * This expects task == current.
3008 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3010 struct perf_event_context *clone_ctx = NULL;
3011 struct perf_event *event;
3012 unsigned long flags;
3016 local_irq_save(flags);
3017 if (!ctx || !ctx->nr_events)
3021 * We must ctxsw out cgroup events to avoid conflict
3022 * when invoking perf_task_event_sched_in() later on
3023 * in this function. Otherwise we end up trying to
3024 * ctxswin cgroup events which are already scheduled
3027 perf_cgroup_sched_out(current, NULL);
3029 raw_spin_lock(&ctx->lock);
3030 task_ctx_sched_out(ctx);
3032 list_for_each_entry(event, &ctx->event_list, event_entry) {
3033 ret = event_enable_on_exec(event, ctx);
3039 * Unclone this context if we enabled any event.
3042 clone_ctx = unclone_ctx(ctx);
3044 raw_spin_unlock(&ctx->lock);
3047 * Also calls ctxswin for cgroup events, if any:
3049 perf_event_context_sched_in(ctx, ctx->task);
3051 local_irq_restore(flags);
3057 void perf_event_exec(void)
3059 struct perf_event_context *ctx;
3063 for_each_task_context_nr(ctxn) {
3064 ctx = current->perf_event_ctxp[ctxn];
3068 perf_event_enable_on_exec(ctx);
3074 * Cross CPU call to read the hardware event
3076 static void __perf_event_read(void *info)
3078 struct perf_event *event = info;
3079 struct perf_event_context *ctx = event->ctx;
3080 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3083 * If this is a task context, we need to check whether it is
3084 * the current task context of this cpu. If not it has been
3085 * scheduled out before the smp call arrived. In that case
3086 * event->count would have been updated to a recent sample
3087 * when the event was scheduled out.
3089 if (ctx->task && cpuctx->task_ctx != ctx)
3092 raw_spin_lock(&ctx->lock);
3093 if (ctx->is_active) {
3094 update_context_time(ctx);
3095 update_cgrp_time_from_event(event);
3097 update_event_times(event);
3098 if (event->state == PERF_EVENT_STATE_ACTIVE)
3099 event->pmu->read(event);
3100 raw_spin_unlock(&ctx->lock);
3103 static inline u64 perf_event_count(struct perf_event *event)
3105 return local64_read(&event->count) + atomic64_read(&event->child_count);
3108 static u64 perf_event_read(struct perf_event *event)
3111 * If event is enabled and currently active on a CPU, update the
3112 * value in the event structure:
3114 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3115 smp_call_function_single(event->oncpu,
3116 __perf_event_read, event, 1);
3117 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3118 struct perf_event_context *ctx = event->ctx;
3119 unsigned long flags;
3121 raw_spin_lock_irqsave(&ctx->lock, flags);
3123 * may read while context is not active
3124 * (e.g., thread is blocked), in that case
3125 * we cannot update context time
3127 if (ctx->is_active) {
3128 update_context_time(ctx);
3129 update_cgrp_time_from_event(event);
3131 update_event_times(event);
3132 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3135 return perf_event_count(event);
3139 * Initialize the perf_event context in a task_struct:
3141 static void __perf_event_init_context(struct perf_event_context *ctx)
3143 raw_spin_lock_init(&ctx->lock);
3144 mutex_init(&ctx->mutex);
3145 INIT_LIST_HEAD(&ctx->pinned_groups);
3146 INIT_LIST_HEAD(&ctx->flexible_groups);
3147 INIT_LIST_HEAD(&ctx->event_list);
3148 atomic_set(&ctx->refcount, 1);
3149 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3152 static struct perf_event_context *
3153 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3155 struct perf_event_context *ctx;
3157 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3161 __perf_event_init_context(ctx);
3164 get_task_struct(task);
3171 static struct task_struct *
3172 find_lively_task_by_vpid(pid_t vpid)
3174 struct task_struct *task;
3181 task = find_task_by_vpid(vpid);
3183 get_task_struct(task);
3187 return ERR_PTR(-ESRCH);
3189 /* Reuse ptrace permission checks for now. */
3191 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3196 put_task_struct(task);
3197 return ERR_PTR(err);
3202 * Returns a matching context with refcount and pincount.
3204 static struct perf_event_context *
3205 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3207 struct perf_event_context *ctx, *clone_ctx = NULL;
3208 struct perf_cpu_context *cpuctx;
3209 unsigned long flags;
3213 /* Must be root to operate on a CPU event: */
3214 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3215 return ERR_PTR(-EACCES);
3218 * We could be clever and allow to attach a event to an
3219 * offline CPU and activate it when the CPU comes up, but
3222 if (!cpu_online(cpu))
3223 return ERR_PTR(-ENODEV);
3225 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3234 ctxn = pmu->task_ctx_nr;
3239 ctx = perf_lock_task_context(task, ctxn, &flags);
3241 clone_ctx = unclone_ctx(ctx);
3243 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3248 ctx = alloc_perf_context(pmu, task);
3254 mutex_lock(&task->perf_event_mutex);
3256 * If it has already passed perf_event_exit_task().
3257 * we must see PF_EXITING, it takes this mutex too.
3259 if (task->flags & PF_EXITING)
3261 else if (task->perf_event_ctxp[ctxn])
3266 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3268 mutex_unlock(&task->perf_event_mutex);
3270 if (unlikely(err)) {
3282 return ERR_PTR(err);
3285 static void perf_event_free_filter(struct perf_event *event);
3287 static void free_event_rcu(struct rcu_head *head)
3289 struct perf_event *event;
3291 event = container_of(head, struct perf_event, rcu_head);
3293 put_pid_ns(event->ns);
3294 perf_event_free_filter(event);
3298 static void ring_buffer_put(struct ring_buffer *rb);
3299 static void ring_buffer_attach(struct perf_event *event,
3300 struct ring_buffer *rb);
3302 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3307 if (has_branch_stack(event)) {
3308 if (!(event->attach_state & PERF_ATTACH_TASK))
3309 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3311 if (is_cgroup_event(event))
3312 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3315 static void unaccount_event(struct perf_event *event)
3320 if (event->attach_state & PERF_ATTACH_TASK)
3321 static_key_slow_dec_deferred(&perf_sched_events);
3322 if (event->attr.mmap || event->attr.mmap_data)
3323 atomic_dec(&nr_mmap_events);
3324 if (event->attr.comm)
3325 atomic_dec(&nr_comm_events);
3326 if (event->attr.task)
3327 atomic_dec(&nr_task_events);
3328 if (event->attr.freq)
3329 atomic_dec(&nr_freq_events);
3330 if (is_cgroup_event(event))
3331 static_key_slow_dec_deferred(&perf_sched_events);
3332 if (has_branch_stack(event))
3333 static_key_slow_dec_deferred(&perf_sched_events);
3335 unaccount_event_cpu(event, event->cpu);
3338 static void __free_event(struct perf_event *event)
3340 if (!event->parent) {
3341 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3342 put_callchain_buffers();
3346 event->destroy(event);
3349 put_ctx(event->ctx);
3352 module_put(event->pmu->module);
3354 call_rcu(&event->rcu_head, free_event_rcu);
3357 static void _free_event(struct perf_event *event)
3359 irq_work_sync(&event->pending);
3361 unaccount_event(event);
3365 * Can happen when we close an event with re-directed output.
3367 * Since we have a 0 refcount, perf_mmap_close() will skip
3368 * over us; possibly making our ring_buffer_put() the last.
3370 mutex_lock(&event->mmap_mutex);
3371 ring_buffer_attach(event, NULL);
3372 mutex_unlock(&event->mmap_mutex);
3375 if (is_cgroup_event(event))
3376 perf_detach_cgroup(event);
3378 __free_event(event);
3382 * Used to free events which have a known refcount of 1, such as in error paths
3383 * where the event isn't exposed yet and inherited events.
3385 static void free_event(struct perf_event *event)
3387 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3388 "unexpected event refcount: %ld; ptr=%p\n",
3389 atomic_long_read(&event->refcount), event)) {
3390 /* leak to avoid use-after-free */
3398 * Remove user event from the owner task.
3400 static void perf_remove_from_owner(struct perf_event *event)
3402 struct task_struct *owner;
3405 owner = ACCESS_ONCE(event->owner);
3407 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3408 * !owner it means the list deletion is complete and we can indeed
3409 * free this event, otherwise we need to serialize on
3410 * owner->perf_event_mutex.
3412 smp_read_barrier_depends();
3415 * Since delayed_put_task_struct() also drops the last
3416 * task reference we can safely take a new reference
3417 * while holding the rcu_read_lock().
3419 get_task_struct(owner);
3424 mutex_lock(&owner->perf_event_mutex);
3426 * We have to re-check the event->owner field, if it is cleared
3427 * we raced with perf_event_exit_task(), acquiring the mutex
3428 * ensured they're done, and we can proceed with freeing the
3432 list_del_init(&event->owner_entry);
3433 mutex_unlock(&owner->perf_event_mutex);
3434 put_task_struct(owner);
3439 * Called when the last reference to the file is gone.
3441 static void put_event(struct perf_event *event)
3443 struct perf_event_context *ctx = event->ctx;
3445 if (!atomic_long_dec_and_test(&event->refcount))
3448 if (!is_kernel_event(event))
3449 perf_remove_from_owner(event);
3451 WARN_ON_ONCE(ctx->parent_ctx);
3453 * There are two ways this annotation is useful:
3455 * 1) there is a lock recursion from perf_event_exit_task
3456 * see the comment there.
3458 * 2) there is a lock-inversion with mmap_sem through
3459 * perf_event_read_group(), which takes faults while
3460 * holding ctx->mutex, however this is called after
3461 * the last filedesc died, so there is no possibility
3462 * to trigger the AB-BA case.
3464 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3465 perf_remove_from_context(event, true);
3466 mutex_unlock(&ctx->mutex);
3471 int perf_event_release_kernel(struct perf_event *event)
3476 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3478 static int perf_release(struct inode *inode, struct file *file)
3480 put_event(file->private_data);
3485 * Remove all orphanes events from the context.
3487 static void orphans_remove_work(struct work_struct *work)
3489 struct perf_event_context *ctx;
3490 struct perf_event *event, *tmp;
3492 ctx = container_of(work, struct perf_event_context,
3493 orphans_remove.work);
3495 mutex_lock(&ctx->mutex);
3496 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3497 struct perf_event *parent_event = event->parent;
3499 if (!is_orphaned_child(event))
3502 perf_remove_from_context(event, true);
3504 mutex_lock(&parent_event->child_mutex);
3505 list_del_init(&event->child_list);
3506 mutex_unlock(&parent_event->child_mutex);
3509 put_event(parent_event);
3512 raw_spin_lock_irq(&ctx->lock);
3513 ctx->orphans_remove_sched = false;
3514 raw_spin_unlock_irq(&ctx->lock);
3515 mutex_unlock(&ctx->mutex);
3520 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3522 struct perf_event *child;
3528 mutex_lock(&event->child_mutex);
3529 total += perf_event_read(event);
3530 *enabled += event->total_time_enabled +
3531 atomic64_read(&event->child_total_time_enabled);
3532 *running += event->total_time_running +
3533 atomic64_read(&event->child_total_time_running);
3535 list_for_each_entry(child, &event->child_list, child_list) {
3536 total += perf_event_read(child);
3537 *enabled += child->total_time_enabled;
3538 *running += child->total_time_running;
3540 mutex_unlock(&event->child_mutex);
3544 EXPORT_SYMBOL_GPL(perf_event_read_value);
3546 static int perf_event_read_group(struct perf_event *event,
3547 u64 read_format, char __user *buf)
3549 struct perf_event *leader = event->group_leader, *sub;
3550 int n = 0, size = 0, ret = -EFAULT;
3551 struct perf_event_context *ctx = leader->ctx;
3553 u64 count, enabled, running;
3555 mutex_lock(&ctx->mutex);
3556 count = perf_event_read_value(leader, &enabled, &running);
3558 values[n++] = 1 + leader->nr_siblings;
3559 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3560 values[n++] = enabled;
3561 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3562 values[n++] = running;
3563 values[n++] = count;
3564 if (read_format & PERF_FORMAT_ID)
3565 values[n++] = primary_event_id(leader);
3567 size = n * sizeof(u64);
3569 if (copy_to_user(buf, values, size))
3574 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3577 values[n++] = perf_event_read_value(sub, &enabled, &running);
3578 if (read_format & PERF_FORMAT_ID)
3579 values[n++] = primary_event_id(sub);
3581 size = n * sizeof(u64);
3583 if (copy_to_user(buf + ret, values, size)) {
3591 mutex_unlock(&ctx->mutex);
3596 static int perf_event_read_one(struct perf_event *event,
3597 u64 read_format, char __user *buf)
3599 u64 enabled, running;
3603 values[n++] = perf_event_read_value(event, &enabled, &running);
3604 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3605 values[n++] = enabled;
3606 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3607 values[n++] = running;
3608 if (read_format & PERF_FORMAT_ID)
3609 values[n++] = primary_event_id(event);
3611 if (copy_to_user(buf, values, n * sizeof(u64)))
3614 return n * sizeof(u64);
3617 static bool is_event_hup(struct perf_event *event)
3621 if (event->state != PERF_EVENT_STATE_EXIT)
3624 mutex_lock(&event->child_mutex);
3625 no_children = list_empty(&event->child_list);
3626 mutex_unlock(&event->child_mutex);
3631 * Read the performance event - simple non blocking version for now
3634 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3636 u64 read_format = event->attr.read_format;
3640 * Return end-of-file for a read on a event that is in
3641 * error state (i.e. because it was pinned but it couldn't be
3642 * scheduled on to the CPU at some point).
3644 if (event->state == PERF_EVENT_STATE_ERROR)
3647 if (count < event->read_size)
3650 WARN_ON_ONCE(event->ctx->parent_ctx);
3651 if (read_format & PERF_FORMAT_GROUP)
3652 ret = perf_event_read_group(event, read_format, buf);
3654 ret = perf_event_read_one(event, read_format, buf);
3660 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3662 struct perf_event *event = file->private_data;
3664 return perf_read_hw(event, buf, count);
3667 static unsigned int perf_poll(struct file *file, poll_table *wait)
3669 struct perf_event *event = file->private_data;
3670 struct ring_buffer *rb;
3671 unsigned int events = POLLHUP;
3673 poll_wait(file, &event->waitq, wait);
3675 if (is_event_hup(event))
3679 * Pin the event->rb by taking event->mmap_mutex; otherwise
3680 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3682 mutex_lock(&event->mmap_mutex);
3685 events = atomic_xchg(&rb->poll, 0);
3686 mutex_unlock(&event->mmap_mutex);
3690 static void perf_event_reset(struct perf_event *event)
3692 (void)perf_event_read(event);
3693 local64_set(&event->count, 0);
3694 perf_event_update_userpage(event);
3698 * Holding the top-level event's child_mutex means that any
3699 * descendant process that has inherited this event will block
3700 * in sync_child_event if it goes to exit, thus satisfying the
3701 * task existence requirements of perf_event_enable/disable.
3703 static void perf_event_for_each_child(struct perf_event *event,
3704 void (*func)(struct perf_event *))
3706 struct perf_event *child;
3708 WARN_ON_ONCE(event->ctx->parent_ctx);
3709 mutex_lock(&event->child_mutex);
3711 list_for_each_entry(child, &event->child_list, child_list)
3713 mutex_unlock(&event->child_mutex);
3716 static void perf_event_for_each(struct perf_event *event,
3717 void (*func)(struct perf_event *))
3719 struct perf_event_context *ctx = event->ctx;
3720 struct perf_event *sibling;
3722 WARN_ON_ONCE(ctx->parent_ctx);
3723 mutex_lock(&ctx->mutex);
3724 event = event->group_leader;
3726 perf_event_for_each_child(event, func);
3727 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3728 perf_event_for_each_child(sibling, func);
3729 mutex_unlock(&ctx->mutex);
3732 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3734 struct perf_event_context *ctx = event->ctx;
3735 int ret = 0, active;
3738 if (!is_sampling_event(event))
3741 if (copy_from_user(&value, arg, sizeof(value)))
3747 raw_spin_lock_irq(&ctx->lock);
3748 if (event->attr.freq) {
3749 if (value > sysctl_perf_event_sample_rate) {
3754 event->attr.sample_freq = value;
3756 event->attr.sample_period = value;
3757 event->hw.sample_period = value;
3760 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3762 perf_pmu_disable(ctx->pmu);
3763 event->pmu->stop(event, PERF_EF_UPDATE);
3766 local64_set(&event->hw.period_left, 0);
3769 event->pmu->start(event, PERF_EF_RELOAD);
3770 perf_pmu_enable(ctx->pmu);
3774 raw_spin_unlock_irq(&ctx->lock);
3779 static const struct file_operations perf_fops;
3781 static inline int perf_fget_light(int fd, struct fd *p)
3783 struct fd f = fdget(fd);
3787 if (f.file->f_op != &perf_fops) {
3795 static int perf_event_set_output(struct perf_event *event,
3796 struct perf_event *output_event);
3797 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3799 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3801 struct perf_event *event = file->private_data;
3802 void (*func)(struct perf_event *);
3806 case PERF_EVENT_IOC_ENABLE:
3807 func = perf_event_enable;
3809 case PERF_EVENT_IOC_DISABLE:
3810 func = perf_event_disable;
3812 case PERF_EVENT_IOC_RESET:
3813 func = perf_event_reset;
3816 case PERF_EVENT_IOC_REFRESH:
3817 return perf_event_refresh(event, arg);
3819 case PERF_EVENT_IOC_PERIOD:
3820 return perf_event_period(event, (u64 __user *)arg);
3822 case PERF_EVENT_IOC_ID:
3824 u64 id = primary_event_id(event);
3826 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3831 case PERF_EVENT_IOC_SET_OUTPUT:
3835 struct perf_event *output_event;
3837 ret = perf_fget_light(arg, &output);
3840 output_event = output.file->private_data;
3841 ret = perf_event_set_output(event, output_event);
3844 ret = perf_event_set_output(event, NULL);
3849 case PERF_EVENT_IOC_SET_FILTER:
3850 return perf_event_set_filter(event, (void __user *)arg);
3856 if (flags & PERF_IOC_FLAG_GROUP)
3857 perf_event_for_each(event, func);
3859 perf_event_for_each_child(event, func);
3864 #ifdef CONFIG_COMPAT
3865 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
3868 switch (_IOC_NR(cmd)) {
3869 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
3870 case _IOC_NR(PERF_EVENT_IOC_ID):
3871 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
3872 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
3873 cmd &= ~IOCSIZE_MASK;
3874 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
3878 return perf_ioctl(file, cmd, arg);
3881 # define perf_compat_ioctl NULL
3884 int perf_event_task_enable(void)
3886 struct perf_event *event;
3888 mutex_lock(¤t->perf_event_mutex);
3889 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3890 perf_event_for_each_child(event, perf_event_enable);
3891 mutex_unlock(¤t->perf_event_mutex);
3896 int perf_event_task_disable(void)
3898 struct perf_event *event;
3900 mutex_lock(¤t->perf_event_mutex);
3901 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3902 perf_event_for_each_child(event, perf_event_disable);
3903 mutex_unlock(¤t->perf_event_mutex);
3908 static int perf_event_index(struct perf_event *event)
3910 if (event->hw.state & PERF_HES_STOPPED)
3913 if (event->state != PERF_EVENT_STATE_ACTIVE)
3916 return event->pmu->event_idx(event);
3919 static void calc_timer_values(struct perf_event *event,
3926 *now = perf_clock();
3927 ctx_time = event->shadow_ctx_time + *now;
3928 *enabled = ctx_time - event->tstamp_enabled;
3929 *running = ctx_time - event->tstamp_running;
3932 static void perf_event_init_userpage(struct perf_event *event)
3934 struct perf_event_mmap_page *userpg;
3935 struct ring_buffer *rb;
3938 rb = rcu_dereference(event->rb);
3942 userpg = rb->user_page;
3944 /* Allow new userspace to detect that bit 0 is deprecated */
3945 userpg->cap_bit0_is_deprecated = 1;
3946 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3952 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3957 * Callers need to ensure there can be no nesting of this function, otherwise
3958 * the seqlock logic goes bad. We can not serialize this because the arch
3959 * code calls this from NMI context.
3961 void perf_event_update_userpage(struct perf_event *event)
3963 struct perf_event_mmap_page *userpg;
3964 struct ring_buffer *rb;
3965 u64 enabled, running, now;
3968 rb = rcu_dereference(event->rb);
3973 * compute total_time_enabled, total_time_running
3974 * based on snapshot values taken when the event
3975 * was last scheduled in.
3977 * we cannot simply called update_context_time()
3978 * because of locking issue as we can be called in
3981 calc_timer_values(event, &now, &enabled, &running);
3983 userpg = rb->user_page;
3985 * Disable preemption so as to not let the corresponding user-space
3986 * spin too long if we get preempted.
3991 userpg->index = perf_event_index(event);
3992 userpg->offset = perf_event_count(event);
3994 userpg->offset -= local64_read(&event->hw.prev_count);
3996 userpg->time_enabled = enabled +
3997 atomic64_read(&event->child_total_time_enabled);
3999 userpg->time_running = running +
4000 atomic64_read(&event->child_total_time_running);
4002 arch_perf_update_userpage(userpg, now);
4011 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4013 struct perf_event *event = vma->vm_file->private_data;
4014 struct ring_buffer *rb;
4015 int ret = VM_FAULT_SIGBUS;
4017 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4018 if (vmf->pgoff == 0)
4024 rb = rcu_dereference(event->rb);
4028 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4031 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4035 get_page(vmf->page);
4036 vmf->page->mapping = vma->vm_file->f_mapping;
4037 vmf->page->index = vmf->pgoff;
4046 static void ring_buffer_attach(struct perf_event *event,
4047 struct ring_buffer *rb)
4049 struct ring_buffer *old_rb = NULL;
4050 unsigned long flags;
4054 * Should be impossible, we set this when removing
4055 * event->rb_entry and wait/clear when adding event->rb_entry.
4057 WARN_ON_ONCE(event->rcu_pending);
4060 event->rcu_batches = get_state_synchronize_rcu();
4061 event->rcu_pending = 1;
4063 spin_lock_irqsave(&old_rb->event_lock, flags);
4064 list_del_rcu(&event->rb_entry);
4065 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4068 if (event->rcu_pending && rb) {
4069 cond_synchronize_rcu(event->rcu_batches);
4070 event->rcu_pending = 0;
4074 spin_lock_irqsave(&rb->event_lock, flags);
4075 list_add_rcu(&event->rb_entry, &rb->event_list);
4076 spin_unlock_irqrestore(&rb->event_lock, flags);
4079 rcu_assign_pointer(event->rb, rb);
4082 ring_buffer_put(old_rb);
4084 * Since we detached before setting the new rb, so that we
4085 * could attach the new rb, we could have missed a wakeup.
4088 wake_up_all(&event->waitq);
4092 static void ring_buffer_wakeup(struct perf_event *event)
4094 struct ring_buffer *rb;
4097 rb = rcu_dereference(event->rb);
4099 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4100 wake_up_all(&event->waitq);
4105 static void rb_free_rcu(struct rcu_head *rcu_head)
4107 struct ring_buffer *rb;
4109 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4113 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
4115 struct ring_buffer *rb;
4118 rb = rcu_dereference(event->rb);
4120 if (!atomic_inc_not_zero(&rb->refcount))
4128 static void ring_buffer_put(struct ring_buffer *rb)
4130 if (!atomic_dec_and_test(&rb->refcount))
4133 WARN_ON_ONCE(!list_empty(&rb->event_list));
4135 call_rcu(&rb->rcu_head, rb_free_rcu);
4138 static void perf_mmap_open(struct vm_area_struct *vma)
4140 struct perf_event *event = vma->vm_file->private_data;
4142 atomic_inc(&event->mmap_count);
4143 atomic_inc(&event->rb->mmap_count);
4147 * A buffer can be mmap()ed multiple times; either directly through the same
4148 * event, or through other events by use of perf_event_set_output().
4150 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4151 * the buffer here, where we still have a VM context. This means we need
4152 * to detach all events redirecting to us.
4154 static void perf_mmap_close(struct vm_area_struct *vma)
4156 struct perf_event *event = vma->vm_file->private_data;
4158 struct ring_buffer *rb = ring_buffer_get(event);
4159 struct user_struct *mmap_user = rb->mmap_user;
4160 int mmap_locked = rb->mmap_locked;
4161 unsigned long size = perf_data_size(rb);
4163 atomic_dec(&rb->mmap_count);
4165 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4168 ring_buffer_attach(event, NULL);
4169 mutex_unlock(&event->mmap_mutex);
4171 /* If there's still other mmap()s of this buffer, we're done. */
4172 if (atomic_read(&rb->mmap_count))
4176 * No other mmap()s, detach from all other events that might redirect
4177 * into the now unreachable buffer. Somewhat complicated by the
4178 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4182 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4183 if (!atomic_long_inc_not_zero(&event->refcount)) {
4185 * This event is en-route to free_event() which will
4186 * detach it and remove it from the list.
4192 mutex_lock(&event->mmap_mutex);
4194 * Check we didn't race with perf_event_set_output() which can
4195 * swizzle the rb from under us while we were waiting to
4196 * acquire mmap_mutex.
4198 * If we find a different rb; ignore this event, a next
4199 * iteration will no longer find it on the list. We have to
4200 * still restart the iteration to make sure we're not now
4201 * iterating the wrong list.
4203 if (event->rb == rb)
4204 ring_buffer_attach(event, NULL);
4206 mutex_unlock(&event->mmap_mutex);
4210 * Restart the iteration; either we're on the wrong list or
4211 * destroyed its integrity by doing a deletion.
4218 * It could be there's still a few 0-ref events on the list; they'll
4219 * get cleaned up by free_event() -- they'll also still have their
4220 * ref on the rb and will free it whenever they are done with it.
4222 * Aside from that, this buffer is 'fully' detached and unmapped,
4223 * undo the VM accounting.
4226 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4227 vma->vm_mm->pinned_vm -= mmap_locked;
4228 free_uid(mmap_user);
4231 ring_buffer_put(rb); /* could be last */
4234 static const struct vm_operations_struct perf_mmap_vmops = {
4235 .open = perf_mmap_open,
4236 .close = perf_mmap_close,
4237 .fault = perf_mmap_fault,
4238 .page_mkwrite = perf_mmap_fault,
4241 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4243 struct perf_event *event = file->private_data;
4244 unsigned long user_locked, user_lock_limit;
4245 struct user_struct *user = current_user();
4246 unsigned long locked, lock_limit;
4247 struct ring_buffer *rb;
4248 unsigned long vma_size;
4249 unsigned long nr_pages;
4250 long user_extra, extra;
4251 int ret = 0, flags = 0;
4254 * Don't allow mmap() of inherited per-task counters. This would
4255 * create a performance issue due to all children writing to the
4258 if (event->cpu == -1 && event->attr.inherit)
4261 if (!(vma->vm_flags & VM_SHARED))
4264 vma_size = vma->vm_end - vma->vm_start;
4265 nr_pages = (vma_size / PAGE_SIZE) - 1;
4268 * If we have rb pages ensure they're a power-of-two number, so we
4269 * can do bitmasks instead of modulo.
4271 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4274 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4277 if (vma->vm_pgoff != 0)
4280 WARN_ON_ONCE(event->ctx->parent_ctx);
4282 mutex_lock(&event->mmap_mutex);
4284 if (event->rb->nr_pages != nr_pages) {
4289 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4291 * Raced against perf_mmap_close() through
4292 * perf_event_set_output(). Try again, hope for better
4295 mutex_unlock(&event->mmap_mutex);
4302 user_extra = nr_pages + 1;
4303 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4306 * Increase the limit linearly with more CPUs:
4308 user_lock_limit *= num_online_cpus();
4310 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4313 if (user_locked > user_lock_limit)
4314 extra = user_locked - user_lock_limit;
4316 lock_limit = rlimit(RLIMIT_MEMLOCK);
4317 lock_limit >>= PAGE_SHIFT;
4318 locked = vma->vm_mm->pinned_vm + extra;
4320 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4321 !capable(CAP_IPC_LOCK)) {
4328 if (vma->vm_flags & VM_WRITE)
4329 flags |= RING_BUFFER_WRITABLE;
4331 rb = rb_alloc(nr_pages,
4332 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4340 atomic_set(&rb->mmap_count, 1);
4341 rb->mmap_locked = extra;
4342 rb->mmap_user = get_current_user();
4344 atomic_long_add(user_extra, &user->locked_vm);
4345 vma->vm_mm->pinned_vm += extra;
4347 ring_buffer_attach(event, rb);
4349 perf_event_init_userpage(event);
4350 perf_event_update_userpage(event);
4354 atomic_inc(&event->mmap_count);
4355 mutex_unlock(&event->mmap_mutex);
4358 * Since pinned accounting is per vm we cannot allow fork() to copy our
4361 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4362 vma->vm_ops = &perf_mmap_vmops;
4367 static int perf_fasync(int fd, struct file *filp, int on)
4369 struct inode *inode = file_inode(filp);
4370 struct perf_event *event = filp->private_data;
4373 mutex_lock(&inode->i_mutex);
4374 retval = fasync_helper(fd, filp, on, &event->fasync);
4375 mutex_unlock(&inode->i_mutex);
4383 static const struct file_operations perf_fops = {
4384 .llseek = no_llseek,
4385 .release = perf_release,
4388 .unlocked_ioctl = perf_ioctl,
4389 .compat_ioctl = perf_compat_ioctl,
4391 .fasync = perf_fasync,
4397 * If there's data, ensure we set the poll() state and publish everything
4398 * to user-space before waking everybody up.
4401 void perf_event_wakeup(struct perf_event *event)
4403 ring_buffer_wakeup(event);
4405 if (event->pending_kill) {
4406 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4407 event->pending_kill = 0;
4411 static void perf_pending_event(struct irq_work *entry)
4413 struct perf_event *event = container_of(entry,
4414 struct perf_event, pending);
4416 if (event->pending_disable) {
4417 event->pending_disable = 0;
4418 __perf_event_disable(event);
4421 if (event->pending_wakeup) {
4422 event->pending_wakeup = 0;
4423 perf_event_wakeup(event);
4428 * We assume there is only KVM supporting the callbacks.
4429 * Later on, we might change it to a list if there is
4430 * another virtualization implementation supporting the callbacks.
4432 struct perf_guest_info_callbacks *perf_guest_cbs;
4434 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4436 perf_guest_cbs = cbs;
4439 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4441 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4443 perf_guest_cbs = NULL;
4446 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4449 perf_output_sample_regs(struct perf_output_handle *handle,
4450 struct pt_regs *regs, u64 mask)
4454 for_each_set_bit(bit, (const unsigned long *) &mask,
4455 sizeof(mask) * BITS_PER_BYTE) {
4458 val = perf_reg_value(regs, bit);
4459 perf_output_put(handle, val);
4463 static void perf_sample_regs_user(struct perf_regs *regs_user,
4464 struct pt_regs *regs)
4466 if (!user_mode(regs)) {
4468 regs = task_pt_regs(current);
4474 regs_user->abi = perf_reg_abi(current);
4475 regs_user->regs = regs;
4477 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4478 regs_user->regs = NULL;
4482 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4483 struct pt_regs *regs)
4485 regs_intr->regs = regs;
4486 regs_intr->abi = perf_reg_abi(current);
4491 * Get remaining task size from user stack pointer.
4493 * It'd be better to take stack vma map and limit this more
4494 * precisly, but there's no way to get it safely under interrupt,
4495 * so using TASK_SIZE as limit.
4497 static u64 perf_ustack_task_size(struct pt_regs *regs)
4499 unsigned long addr = perf_user_stack_pointer(regs);
4501 if (!addr || addr >= TASK_SIZE)
4504 return TASK_SIZE - addr;
4508 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4509 struct pt_regs *regs)
4513 /* No regs, no stack pointer, no dump. */
4518 * Check if we fit in with the requested stack size into the:
4520 * If we don't, we limit the size to the TASK_SIZE.
4522 * - remaining sample size
4523 * If we don't, we customize the stack size to
4524 * fit in to the remaining sample size.
4527 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4528 stack_size = min(stack_size, (u16) task_size);
4530 /* Current header size plus static size and dynamic size. */
4531 header_size += 2 * sizeof(u64);
4533 /* Do we fit in with the current stack dump size? */
4534 if ((u16) (header_size + stack_size) < header_size) {
4536 * If we overflow the maximum size for the sample,
4537 * we customize the stack dump size to fit in.
4539 stack_size = USHRT_MAX - header_size - sizeof(u64);
4540 stack_size = round_up(stack_size, sizeof(u64));
4547 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4548 struct pt_regs *regs)
4550 /* Case of a kernel thread, nothing to dump */
4553 perf_output_put(handle, size);
4562 * - the size requested by user or the best one we can fit
4563 * in to the sample max size
4565 * - user stack dump data
4567 * - the actual dumped size
4571 perf_output_put(handle, dump_size);
4574 sp = perf_user_stack_pointer(regs);
4575 rem = __output_copy_user(handle, (void *) sp, dump_size);
4576 dyn_size = dump_size - rem;
4578 perf_output_skip(handle, rem);
4581 perf_output_put(handle, dyn_size);
4585 static void __perf_event_header__init_id(struct perf_event_header *header,
4586 struct perf_sample_data *data,
4587 struct perf_event *event)
4589 u64 sample_type = event->attr.sample_type;
4591 data->type = sample_type;
4592 header->size += event->id_header_size;
4594 if (sample_type & PERF_SAMPLE_TID) {
4595 /* namespace issues */
4596 data->tid_entry.pid = perf_event_pid(event, current);
4597 data->tid_entry.tid = perf_event_tid(event, current);
4600 if (sample_type & PERF_SAMPLE_TIME)
4601 data->time = perf_clock();
4603 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4604 data->id = primary_event_id(event);
4606 if (sample_type & PERF_SAMPLE_STREAM_ID)
4607 data->stream_id = event->id;
4609 if (sample_type & PERF_SAMPLE_CPU) {
4610 data->cpu_entry.cpu = raw_smp_processor_id();
4611 data->cpu_entry.reserved = 0;
4615 void perf_event_header__init_id(struct perf_event_header *header,
4616 struct perf_sample_data *data,
4617 struct perf_event *event)
4619 if (event->attr.sample_id_all)
4620 __perf_event_header__init_id(header, data, event);
4623 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4624 struct perf_sample_data *data)
4626 u64 sample_type = data->type;
4628 if (sample_type & PERF_SAMPLE_TID)
4629 perf_output_put(handle, data->tid_entry);
4631 if (sample_type & PERF_SAMPLE_TIME)
4632 perf_output_put(handle, data->time);
4634 if (sample_type & PERF_SAMPLE_ID)
4635 perf_output_put(handle, data->id);
4637 if (sample_type & PERF_SAMPLE_STREAM_ID)
4638 perf_output_put(handle, data->stream_id);
4640 if (sample_type & PERF_SAMPLE_CPU)
4641 perf_output_put(handle, data->cpu_entry);
4643 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4644 perf_output_put(handle, data->id);
4647 void perf_event__output_id_sample(struct perf_event *event,
4648 struct perf_output_handle *handle,
4649 struct perf_sample_data *sample)
4651 if (event->attr.sample_id_all)
4652 __perf_event__output_id_sample(handle, sample);
4655 static void perf_output_read_one(struct perf_output_handle *handle,
4656 struct perf_event *event,
4657 u64 enabled, u64 running)
4659 u64 read_format = event->attr.read_format;
4663 values[n++] = perf_event_count(event);
4664 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4665 values[n++] = enabled +
4666 atomic64_read(&event->child_total_time_enabled);
4668 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4669 values[n++] = running +
4670 atomic64_read(&event->child_total_time_running);
4672 if (read_format & PERF_FORMAT_ID)
4673 values[n++] = primary_event_id(event);
4675 __output_copy(handle, values, n * sizeof(u64));
4679 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4681 static void perf_output_read_group(struct perf_output_handle *handle,
4682 struct perf_event *event,
4683 u64 enabled, u64 running)
4685 struct perf_event *leader = event->group_leader, *sub;
4686 u64 read_format = event->attr.read_format;
4690 values[n++] = 1 + leader->nr_siblings;
4692 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4693 values[n++] = enabled;
4695 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4696 values[n++] = running;
4698 if (leader != event)
4699 leader->pmu->read(leader);
4701 values[n++] = perf_event_count(leader);
4702 if (read_format & PERF_FORMAT_ID)
4703 values[n++] = primary_event_id(leader);
4705 __output_copy(handle, values, n * sizeof(u64));
4707 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4710 if ((sub != event) &&
4711 (sub->state == PERF_EVENT_STATE_ACTIVE))
4712 sub->pmu->read(sub);
4714 values[n++] = perf_event_count(sub);
4715 if (read_format & PERF_FORMAT_ID)
4716 values[n++] = primary_event_id(sub);
4718 __output_copy(handle, values, n * sizeof(u64));
4722 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4723 PERF_FORMAT_TOTAL_TIME_RUNNING)
4725 static void perf_output_read(struct perf_output_handle *handle,
4726 struct perf_event *event)
4728 u64 enabled = 0, running = 0, now;
4729 u64 read_format = event->attr.read_format;
4732 * compute total_time_enabled, total_time_running
4733 * based on snapshot values taken when the event
4734 * was last scheduled in.
4736 * we cannot simply called update_context_time()
4737 * because of locking issue as we are called in
4740 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4741 calc_timer_values(event, &now, &enabled, &running);
4743 if (event->attr.read_format & PERF_FORMAT_GROUP)
4744 perf_output_read_group(handle, event, enabled, running);
4746 perf_output_read_one(handle, event, enabled, running);
4749 void perf_output_sample(struct perf_output_handle *handle,
4750 struct perf_event_header *header,
4751 struct perf_sample_data *data,
4752 struct perf_event *event)
4754 u64 sample_type = data->type;
4756 perf_output_put(handle, *header);
4758 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4759 perf_output_put(handle, data->id);
4761 if (sample_type & PERF_SAMPLE_IP)
4762 perf_output_put(handle, data->ip);
4764 if (sample_type & PERF_SAMPLE_TID)
4765 perf_output_put(handle, data->tid_entry);
4767 if (sample_type & PERF_SAMPLE_TIME)
4768 perf_output_put(handle, data->time);
4770 if (sample_type & PERF_SAMPLE_ADDR)
4771 perf_output_put(handle, data->addr);
4773 if (sample_type & PERF_SAMPLE_ID)
4774 perf_output_put(handle, data->id);
4776 if (sample_type & PERF_SAMPLE_STREAM_ID)
4777 perf_output_put(handle, data->stream_id);
4779 if (sample_type & PERF_SAMPLE_CPU)
4780 perf_output_put(handle, data->cpu_entry);
4782 if (sample_type & PERF_SAMPLE_PERIOD)
4783 perf_output_put(handle, data->period);
4785 if (sample_type & PERF_SAMPLE_READ)
4786 perf_output_read(handle, event);
4788 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4789 if (data->callchain) {
4792 if (data->callchain)
4793 size += data->callchain->nr;
4795 size *= sizeof(u64);
4797 __output_copy(handle, data->callchain, size);
4800 perf_output_put(handle, nr);
4804 if (sample_type & PERF_SAMPLE_RAW) {
4806 perf_output_put(handle, data->raw->size);
4807 __output_copy(handle, data->raw->data,
4814 .size = sizeof(u32),
4817 perf_output_put(handle, raw);
4821 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4822 if (data->br_stack) {
4825 size = data->br_stack->nr
4826 * sizeof(struct perf_branch_entry);
4828 perf_output_put(handle, data->br_stack->nr);
4829 perf_output_copy(handle, data->br_stack->entries, size);
4832 * we always store at least the value of nr
4835 perf_output_put(handle, nr);
4839 if (sample_type & PERF_SAMPLE_REGS_USER) {
4840 u64 abi = data->regs_user.abi;
4843 * If there are no regs to dump, notice it through
4844 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4846 perf_output_put(handle, abi);
4849 u64 mask = event->attr.sample_regs_user;
4850 perf_output_sample_regs(handle,
4851 data->regs_user.regs,
4856 if (sample_type & PERF_SAMPLE_STACK_USER) {
4857 perf_output_sample_ustack(handle,
4858 data->stack_user_size,
4859 data->regs_user.regs);
4862 if (sample_type & PERF_SAMPLE_WEIGHT)
4863 perf_output_put(handle, data->weight);
4865 if (sample_type & PERF_SAMPLE_DATA_SRC)
4866 perf_output_put(handle, data->data_src.val);
4868 if (sample_type & PERF_SAMPLE_TRANSACTION)
4869 perf_output_put(handle, data->txn);
4871 if (sample_type & PERF_SAMPLE_REGS_INTR) {
4872 u64 abi = data->regs_intr.abi;
4874 * If there are no regs to dump, notice it through
4875 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4877 perf_output_put(handle, abi);
4880 u64 mask = event->attr.sample_regs_intr;
4882 perf_output_sample_regs(handle,
4883 data->regs_intr.regs,
4888 if (!event->attr.watermark) {
4889 int wakeup_events = event->attr.wakeup_events;
4891 if (wakeup_events) {
4892 struct ring_buffer *rb = handle->rb;
4893 int events = local_inc_return(&rb->events);
4895 if (events >= wakeup_events) {
4896 local_sub(wakeup_events, &rb->events);
4897 local_inc(&rb->wakeup);
4903 void perf_prepare_sample(struct perf_event_header *header,
4904 struct perf_sample_data *data,
4905 struct perf_event *event,
4906 struct pt_regs *regs)
4908 u64 sample_type = event->attr.sample_type;
4910 header->type = PERF_RECORD_SAMPLE;
4911 header->size = sizeof(*header) + event->header_size;
4914 header->misc |= perf_misc_flags(regs);
4916 __perf_event_header__init_id(header, data, event);
4918 if (sample_type & PERF_SAMPLE_IP)
4919 data->ip = perf_instruction_pointer(regs);
4921 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4924 data->callchain = perf_callchain(event, regs);
4926 if (data->callchain)
4927 size += data->callchain->nr;
4929 header->size += size * sizeof(u64);
4932 if (sample_type & PERF_SAMPLE_RAW) {
4933 int size = sizeof(u32);
4936 size += data->raw->size;
4938 size += sizeof(u32);
4940 WARN_ON_ONCE(size & (sizeof(u64)-1));
4941 header->size += size;
4944 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4945 int size = sizeof(u64); /* nr */
4946 if (data->br_stack) {
4947 size += data->br_stack->nr
4948 * sizeof(struct perf_branch_entry);
4950 header->size += size;
4953 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
4954 perf_sample_regs_user(&data->regs_user, regs);
4956 if (sample_type & PERF_SAMPLE_REGS_USER) {
4957 /* regs dump ABI info */
4958 int size = sizeof(u64);
4960 if (data->regs_user.regs) {
4961 u64 mask = event->attr.sample_regs_user;
4962 size += hweight64(mask) * sizeof(u64);
4965 header->size += size;
4968 if (sample_type & PERF_SAMPLE_STACK_USER) {
4970 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4971 * processed as the last one or have additional check added
4972 * in case new sample type is added, because we could eat
4973 * up the rest of the sample size.
4975 u16 stack_size = event->attr.sample_stack_user;
4976 u16 size = sizeof(u64);
4978 stack_size = perf_sample_ustack_size(stack_size, header->size,
4979 data->regs_user.regs);
4982 * If there is something to dump, add space for the dump
4983 * itself and for the field that tells the dynamic size,
4984 * which is how many have been actually dumped.
4987 size += sizeof(u64) + stack_size;
4989 data->stack_user_size = stack_size;
4990 header->size += size;
4993 if (sample_type & PERF_SAMPLE_REGS_INTR) {
4994 /* regs dump ABI info */
4995 int size = sizeof(u64);
4997 perf_sample_regs_intr(&data->regs_intr, regs);
4999 if (data->regs_intr.regs) {
5000 u64 mask = event->attr.sample_regs_intr;
5002 size += hweight64(mask) * sizeof(u64);
5005 header->size += size;
5009 static void perf_event_output(struct perf_event *event,
5010 struct perf_sample_data *data,
5011 struct pt_regs *regs)
5013 struct perf_output_handle handle;
5014 struct perf_event_header header;
5016 /* protect the callchain buffers */
5019 perf_prepare_sample(&header, data, event, regs);
5021 if (perf_output_begin(&handle, event, header.size))
5024 perf_output_sample(&handle, &header, data, event);
5026 perf_output_end(&handle);
5036 struct perf_read_event {
5037 struct perf_event_header header;
5044 perf_event_read_event(struct perf_event *event,
5045 struct task_struct *task)
5047 struct perf_output_handle handle;
5048 struct perf_sample_data sample;
5049 struct perf_read_event read_event = {
5051 .type = PERF_RECORD_READ,
5053 .size = sizeof(read_event) + event->read_size,
5055 .pid = perf_event_pid(event, task),
5056 .tid = perf_event_tid(event, task),
5060 perf_event_header__init_id(&read_event.header, &sample, event);
5061 ret = perf_output_begin(&handle, event, read_event.header.size);
5065 perf_output_put(&handle, read_event);
5066 perf_output_read(&handle, event);
5067 perf_event__output_id_sample(event, &handle, &sample);
5069 perf_output_end(&handle);
5072 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5075 perf_event_aux_ctx(struct perf_event_context *ctx,
5076 perf_event_aux_output_cb output,
5079 struct perf_event *event;
5081 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5082 if (event->state < PERF_EVENT_STATE_INACTIVE)
5084 if (!event_filter_match(event))
5086 output(event, data);
5091 perf_event_aux(perf_event_aux_output_cb output, void *data,
5092 struct perf_event_context *task_ctx)
5094 struct perf_cpu_context *cpuctx;
5095 struct perf_event_context *ctx;
5100 list_for_each_entry_rcu(pmu, &pmus, entry) {
5101 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5102 if (cpuctx->unique_pmu != pmu)
5104 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5107 ctxn = pmu->task_ctx_nr;
5110 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5112 perf_event_aux_ctx(ctx, output, data);
5114 put_cpu_ptr(pmu->pmu_cpu_context);
5119 perf_event_aux_ctx(task_ctx, output, data);
5126 * task tracking -- fork/exit
5128 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5131 struct perf_task_event {
5132 struct task_struct *task;
5133 struct perf_event_context *task_ctx;
5136 struct perf_event_header header;
5146 static int perf_event_task_match(struct perf_event *event)
5148 return event->attr.comm || event->attr.mmap ||
5149 event->attr.mmap2 || event->attr.mmap_data ||
5153 static void perf_event_task_output(struct perf_event *event,
5156 struct perf_task_event *task_event = data;
5157 struct perf_output_handle handle;
5158 struct perf_sample_data sample;
5159 struct task_struct *task = task_event->task;
5160 int ret, size = task_event->event_id.header.size;
5162 if (!perf_event_task_match(event))
5165 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5167 ret = perf_output_begin(&handle, event,
5168 task_event->event_id.header.size);
5172 task_event->event_id.pid = perf_event_pid(event, task);
5173 task_event->event_id.ppid = perf_event_pid(event, current);
5175 task_event->event_id.tid = perf_event_tid(event, task);
5176 task_event->event_id.ptid = perf_event_tid(event, current);
5178 perf_output_put(&handle, task_event->event_id);
5180 perf_event__output_id_sample(event, &handle, &sample);
5182 perf_output_end(&handle);
5184 task_event->event_id.header.size = size;
5187 static void perf_event_task(struct task_struct *task,
5188 struct perf_event_context *task_ctx,
5191 struct perf_task_event task_event;
5193 if (!atomic_read(&nr_comm_events) &&
5194 !atomic_read(&nr_mmap_events) &&
5195 !atomic_read(&nr_task_events))
5198 task_event = (struct perf_task_event){
5200 .task_ctx = task_ctx,
5203 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5205 .size = sizeof(task_event.event_id),
5211 .time = perf_clock(),
5215 perf_event_aux(perf_event_task_output,
5220 void perf_event_fork(struct task_struct *task)
5222 perf_event_task(task, NULL, 1);
5229 struct perf_comm_event {
5230 struct task_struct *task;
5235 struct perf_event_header header;
5242 static int perf_event_comm_match(struct perf_event *event)
5244 return event->attr.comm;
5247 static void perf_event_comm_output(struct perf_event *event,
5250 struct perf_comm_event *comm_event = data;
5251 struct perf_output_handle handle;
5252 struct perf_sample_data sample;
5253 int size = comm_event->event_id.header.size;
5256 if (!perf_event_comm_match(event))
5259 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5260 ret = perf_output_begin(&handle, event,
5261 comm_event->event_id.header.size);
5266 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5267 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5269 perf_output_put(&handle, comm_event->event_id);
5270 __output_copy(&handle, comm_event->comm,
5271 comm_event->comm_size);
5273 perf_event__output_id_sample(event, &handle, &sample);
5275 perf_output_end(&handle);
5277 comm_event->event_id.header.size = size;
5280 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5282 char comm[TASK_COMM_LEN];
5285 memset(comm, 0, sizeof(comm));
5286 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5287 size = ALIGN(strlen(comm)+1, sizeof(u64));
5289 comm_event->comm = comm;
5290 comm_event->comm_size = size;
5292 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5294 perf_event_aux(perf_event_comm_output,
5299 void perf_event_comm(struct task_struct *task, bool exec)
5301 struct perf_comm_event comm_event;
5303 if (!atomic_read(&nr_comm_events))
5306 comm_event = (struct perf_comm_event){
5312 .type = PERF_RECORD_COMM,
5313 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5321 perf_event_comm_event(&comm_event);
5328 struct perf_mmap_event {
5329 struct vm_area_struct *vma;
5331 const char *file_name;
5339 struct perf_event_header header;
5349 static int perf_event_mmap_match(struct perf_event *event,
5352 struct perf_mmap_event *mmap_event = data;
5353 struct vm_area_struct *vma = mmap_event->vma;
5354 int executable = vma->vm_flags & VM_EXEC;
5356 return (!executable && event->attr.mmap_data) ||
5357 (executable && (event->attr.mmap || event->attr.mmap2));
5360 static void perf_event_mmap_output(struct perf_event *event,
5363 struct perf_mmap_event *mmap_event = data;
5364 struct perf_output_handle handle;
5365 struct perf_sample_data sample;
5366 int size = mmap_event->event_id.header.size;
5369 if (!perf_event_mmap_match(event, data))
5372 if (event->attr.mmap2) {
5373 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5374 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5375 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5376 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5377 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5378 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5379 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5382 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5383 ret = perf_output_begin(&handle, event,
5384 mmap_event->event_id.header.size);
5388 mmap_event->event_id.pid = perf_event_pid(event, current);
5389 mmap_event->event_id.tid = perf_event_tid(event, current);
5391 perf_output_put(&handle, mmap_event->event_id);
5393 if (event->attr.mmap2) {
5394 perf_output_put(&handle, mmap_event->maj);
5395 perf_output_put(&handle, mmap_event->min);
5396 perf_output_put(&handle, mmap_event->ino);
5397 perf_output_put(&handle, mmap_event->ino_generation);
5398 perf_output_put(&handle, mmap_event->prot);
5399 perf_output_put(&handle, mmap_event->flags);
5402 __output_copy(&handle, mmap_event->file_name,
5403 mmap_event->file_size);
5405 perf_event__output_id_sample(event, &handle, &sample);
5407 perf_output_end(&handle);
5409 mmap_event->event_id.header.size = size;
5412 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5414 struct vm_area_struct *vma = mmap_event->vma;
5415 struct file *file = vma->vm_file;
5416 int maj = 0, min = 0;
5417 u64 ino = 0, gen = 0;
5418 u32 prot = 0, flags = 0;
5425 struct inode *inode;
5428 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5434 * d_path() works from the end of the rb backwards, so we
5435 * need to add enough zero bytes after the string to handle
5436 * the 64bit alignment we do later.
5438 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5443 inode = file_inode(vma->vm_file);
5444 dev = inode->i_sb->s_dev;
5446 gen = inode->i_generation;
5450 if (vma->vm_flags & VM_READ)
5452 if (vma->vm_flags & VM_WRITE)
5454 if (vma->vm_flags & VM_EXEC)
5457 if (vma->vm_flags & VM_MAYSHARE)
5460 flags = MAP_PRIVATE;
5462 if (vma->vm_flags & VM_DENYWRITE)
5463 flags |= MAP_DENYWRITE;
5464 if (vma->vm_flags & VM_MAYEXEC)
5465 flags |= MAP_EXECUTABLE;
5466 if (vma->vm_flags & VM_LOCKED)
5467 flags |= MAP_LOCKED;
5468 if (vma->vm_flags & VM_HUGETLB)
5469 flags |= MAP_HUGETLB;
5473 if (vma->vm_ops && vma->vm_ops->name) {
5474 name = (char *) vma->vm_ops->name(vma);
5479 name = (char *)arch_vma_name(vma);
5483 if (vma->vm_start <= vma->vm_mm->start_brk &&
5484 vma->vm_end >= vma->vm_mm->brk) {
5488 if (vma->vm_start <= vma->vm_mm->start_stack &&
5489 vma->vm_end >= vma->vm_mm->start_stack) {
5499 strlcpy(tmp, name, sizeof(tmp));
5503 * Since our buffer works in 8 byte units we need to align our string
5504 * size to a multiple of 8. However, we must guarantee the tail end is
5505 * zero'd out to avoid leaking random bits to userspace.
5507 size = strlen(name)+1;
5508 while (!IS_ALIGNED(size, sizeof(u64)))
5509 name[size++] = '\0';
5511 mmap_event->file_name = name;
5512 mmap_event->file_size = size;
5513 mmap_event->maj = maj;
5514 mmap_event->min = min;
5515 mmap_event->ino = ino;
5516 mmap_event->ino_generation = gen;
5517 mmap_event->prot = prot;
5518 mmap_event->flags = flags;
5520 if (!(vma->vm_flags & VM_EXEC))
5521 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5523 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5525 perf_event_aux(perf_event_mmap_output,
5532 void perf_event_mmap(struct vm_area_struct *vma)
5534 struct perf_mmap_event mmap_event;
5536 if (!atomic_read(&nr_mmap_events))
5539 mmap_event = (struct perf_mmap_event){
5545 .type = PERF_RECORD_MMAP,
5546 .misc = PERF_RECORD_MISC_USER,
5551 .start = vma->vm_start,
5552 .len = vma->vm_end - vma->vm_start,
5553 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5555 /* .maj (attr_mmap2 only) */
5556 /* .min (attr_mmap2 only) */
5557 /* .ino (attr_mmap2 only) */
5558 /* .ino_generation (attr_mmap2 only) */
5559 /* .prot (attr_mmap2 only) */
5560 /* .flags (attr_mmap2 only) */
5563 perf_event_mmap_event(&mmap_event);
5567 * IRQ throttle logging
5570 static void perf_log_throttle(struct perf_event *event, int enable)
5572 struct perf_output_handle handle;
5573 struct perf_sample_data sample;
5577 struct perf_event_header header;
5581 } throttle_event = {
5583 .type = PERF_RECORD_THROTTLE,
5585 .size = sizeof(throttle_event),
5587 .time = perf_clock(),
5588 .id = primary_event_id(event),
5589 .stream_id = event->id,
5593 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5595 perf_event_header__init_id(&throttle_event.header, &sample, event);
5597 ret = perf_output_begin(&handle, event,
5598 throttle_event.header.size);
5602 perf_output_put(&handle, throttle_event);
5603 perf_event__output_id_sample(event, &handle, &sample);
5604 perf_output_end(&handle);
5608 * Generic event overflow handling, sampling.
5611 static int __perf_event_overflow(struct perf_event *event,
5612 int throttle, struct perf_sample_data *data,
5613 struct pt_regs *regs)
5615 int events = atomic_read(&event->event_limit);
5616 struct hw_perf_event *hwc = &event->hw;
5621 * Non-sampling counters might still use the PMI to fold short
5622 * hardware counters, ignore those.
5624 if (unlikely(!is_sampling_event(event)))
5627 seq = __this_cpu_read(perf_throttled_seq);
5628 if (seq != hwc->interrupts_seq) {
5629 hwc->interrupts_seq = seq;
5630 hwc->interrupts = 1;
5633 if (unlikely(throttle
5634 && hwc->interrupts >= max_samples_per_tick)) {
5635 __this_cpu_inc(perf_throttled_count);
5636 hwc->interrupts = MAX_INTERRUPTS;
5637 perf_log_throttle(event, 0);
5638 tick_nohz_full_kick();
5643 if (event->attr.freq) {
5644 u64 now = perf_clock();
5645 s64 delta = now - hwc->freq_time_stamp;
5647 hwc->freq_time_stamp = now;
5649 if (delta > 0 && delta < 2*TICK_NSEC)
5650 perf_adjust_period(event, delta, hwc->last_period, true);
5654 * XXX event_limit might not quite work as expected on inherited
5658 event->pending_kill = POLL_IN;
5659 if (events && atomic_dec_and_test(&event->event_limit)) {
5661 event->pending_kill = POLL_HUP;
5662 event->pending_disable = 1;
5663 irq_work_queue(&event->pending);
5666 if (event->overflow_handler)
5667 event->overflow_handler(event, data, regs);
5669 perf_event_output(event, data, regs);
5671 if (event->fasync && event->pending_kill) {
5672 event->pending_wakeup = 1;
5673 irq_work_queue(&event->pending);
5679 int perf_event_overflow(struct perf_event *event,
5680 struct perf_sample_data *data,
5681 struct pt_regs *regs)
5683 return __perf_event_overflow(event, 1, data, regs);
5687 * Generic software event infrastructure
5690 struct swevent_htable {
5691 struct swevent_hlist *swevent_hlist;
5692 struct mutex hlist_mutex;
5695 /* Recursion avoidance in each contexts */
5696 int recursion[PERF_NR_CONTEXTS];
5698 /* Keeps track of cpu being initialized/exited */
5702 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5705 * We directly increment event->count and keep a second value in
5706 * event->hw.period_left to count intervals. This period event
5707 * is kept in the range [-sample_period, 0] so that we can use the
5711 u64 perf_swevent_set_period(struct perf_event *event)
5713 struct hw_perf_event *hwc = &event->hw;
5714 u64 period = hwc->last_period;
5718 hwc->last_period = hwc->sample_period;
5721 old = val = local64_read(&hwc->period_left);
5725 nr = div64_u64(period + val, period);
5726 offset = nr * period;
5728 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5734 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5735 struct perf_sample_data *data,
5736 struct pt_regs *regs)
5738 struct hw_perf_event *hwc = &event->hw;
5742 overflow = perf_swevent_set_period(event);
5744 if (hwc->interrupts == MAX_INTERRUPTS)
5747 for (; overflow; overflow--) {
5748 if (__perf_event_overflow(event, throttle,
5751 * We inhibit the overflow from happening when
5752 * hwc->interrupts == MAX_INTERRUPTS.
5760 static void perf_swevent_event(struct perf_event *event, u64 nr,
5761 struct perf_sample_data *data,
5762 struct pt_regs *regs)
5764 struct hw_perf_event *hwc = &event->hw;
5766 local64_add(nr, &event->count);
5771 if (!is_sampling_event(event))
5774 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5776 return perf_swevent_overflow(event, 1, data, regs);
5778 data->period = event->hw.last_period;
5780 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5781 return perf_swevent_overflow(event, 1, data, regs);
5783 if (local64_add_negative(nr, &hwc->period_left))
5786 perf_swevent_overflow(event, 0, data, regs);
5789 static int perf_exclude_event(struct perf_event *event,
5790 struct pt_regs *regs)
5792 if (event->hw.state & PERF_HES_STOPPED)
5796 if (event->attr.exclude_user && user_mode(regs))
5799 if (event->attr.exclude_kernel && !user_mode(regs))
5806 static int perf_swevent_match(struct perf_event *event,
5807 enum perf_type_id type,
5809 struct perf_sample_data *data,
5810 struct pt_regs *regs)
5812 if (event->attr.type != type)
5815 if (event->attr.config != event_id)
5818 if (perf_exclude_event(event, regs))
5824 static inline u64 swevent_hash(u64 type, u32 event_id)
5826 u64 val = event_id | (type << 32);
5828 return hash_64(val, SWEVENT_HLIST_BITS);
5831 static inline struct hlist_head *
5832 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5834 u64 hash = swevent_hash(type, event_id);
5836 return &hlist->heads[hash];
5839 /* For the read side: events when they trigger */
5840 static inline struct hlist_head *
5841 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5843 struct swevent_hlist *hlist;
5845 hlist = rcu_dereference(swhash->swevent_hlist);
5849 return __find_swevent_head(hlist, type, event_id);
5852 /* For the event head insertion and removal in the hlist */
5853 static inline struct hlist_head *
5854 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5856 struct swevent_hlist *hlist;
5857 u32 event_id = event->attr.config;
5858 u64 type = event->attr.type;
5861 * Event scheduling is always serialized against hlist allocation
5862 * and release. Which makes the protected version suitable here.
5863 * The context lock guarantees that.
5865 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5866 lockdep_is_held(&event->ctx->lock));
5870 return __find_swevent_head(hlist, type, event_id);
5873 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5875 struct perf_sample_data *data,
5876 struct pt_regs *regs)
5878 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
5879 struct perf_event *event;
5880 struct hlist_head *head;
5883 head = find_swevent_head_rcu(swhash, type, event_id);
5887 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5888 if (perf_swevent_match(event, type, event_id, data, regs))
5889 perf_swevent_event(event, nr, data, regs);
5895 int perf_swevent_get_recursion_context(void)
5897 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
5899 return get_recursion_context(swhash->recursion);
5901 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5903 inline void perf_swevent_put_recursion_context(int rctx)
5905 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
5907 put_recursion_context(swhash->recursion, rctx);
5910 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5912 struct perf_sample_data data;
5915 preempt_disable_notrace();
5916 rctx = perf_swevent_get_recursion_context();
5920 perf_sample_data_init(&data, addr, 0);
5922 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5924 perf_swevent_put_recursion_context(rctx);
5925 preempt_enable_notrace();
5928 static void perf_swevent_read(struct perf_event *event)
5932 static int perf_swevent_add(struct perf_event *event, int flags)
5934 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
5935 struct hw_perf_event *hwc = &event->hw;
5936 struct hlist_head *head;
5938 if (is_sampling_event(event)) {
5939 hwc->last_period = hwc->sample_period;
5940 perf_swevent_set_period(event);
5943 hwc->state = !(flags & PERF_EF_START);
5945 head = find_swevent_head(swhash, event);
5948 * We can race with cpu hotplug code. Do not
5949 * WARN if the cpu just got unplugged.
5951 WARN_ON_ONCE(swhash->online);
5955 hlist_add_head_rcu(&event->hlist_entry, head);
5960 static void perf_swevent_del(struct perf_event *event, int flags)
5962 hlist_del_rcu(&event->hlist_entry);
5965 static void perf_swevent_start(struct perf_event *event, int flags)
5967 event->hw.state = 0;
5970 static void perf_swevent_stop(struct perf_event *event, int flags)
5972 event->hw.state = PERF_HES_STOPPED;
5975 /* Deref the hlist from the update side */
5976 static inline struct swevent_hlist *
5977 swevent_hlist_deref(struct swevent_htable *swhash)
5979 return rcu_dereference_protected(swhash->swevent_hlist,
5980 lockdep_is_held(&swhash->hlist_mutex));
5983 static void swevent_hlist_release(struct swevent_htable *swhash)
5985 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5990 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
5991 kfree_rcu(hlist, rcu_head);
5994 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5996 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5998 mutex_lock(&swhash->hlist_mutex);
6000 if (!--swhash->hlist_refcount)
6001 swevent_hlist_release(swhash);
6003 mutex_unlock(&swhash->hlist_mutex);
6006 static void swevent_hlist_put(struct perf_event *event)
6010 for_each_possible_cpu(cpu)
6011 swevent_hlist_put_cpu(event, cpu);
6014 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6016 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6019 mutex_lock(&swhash->hlist_mutex);
6021 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6022 struct swevent_hlist *hlist;
6024 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6029 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6031 swhash->hlist_refcount++;
6033 mutex_unlock(&swhash->hlist_mutex);
6038 static int swevent_hlist_get(struct perf_event *event)
6041 int cpu, failed_cpu;
6044 for_each_possible_cpu(cpu) {
6045 err = swevent_hlist_get_cpu(event, cpu);
6055 for_each_possible_cpu(cpu) {
6056 if (cpu == failed_cpu)
6058 swevent_hlist_put_cpu(event, cpu);
6065 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6067 static void sw_perf_event_destroy(struct perf_event *event)
6069 u64 event_id = event->attr.config;
6071 WARN_ON(event->parent);
6073 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6074 swevent_hlist_put(event);
6077 static int perf_swevent_init(struct perf_event *event)
6079 u64 event_id = event->attr.config;
6081 if (event->attr.type != PERF_TYPE_SOFTWARE)
6085 * no branch sampling for software events
6087 if (has_branch_stack(event))
6091 case PERF_COUNT_SW_CPU_CLOCK:
6092 case PERF_COUNT_SW_TASK_CLOCK:
6099 if (event_id >= PERF_COUNT_SW_MAX)
6102 if (!event->parent) {
6105 err = swevent_hlist_get(event);
6109 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6110 event->destroy = sw_perf_event_destroy;
6116 static struct pmu perf_swevent = {
6117 .task_ctx_nr = perf_sw_context,
6119 .event_init = perf_swevent_init,
6120 .add = perf_swevent_add,
6121 .del = perf_swevent_del,
6122 .start = perf_swevent_start,
6123 .stop = perf_swevent_stop,
6124 .read = perf_swevent_read,
6127 #ifdef CONFIG_EVENT_TRACING
6129 static int perf_tp_filter_match(struct perf_event *event,
6130 struct perf_sample_data *data)
6132 void *record = data->raw->data;
6134 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6139 static int perf_tp_event_match(struct perf_event *event,
6140 struct perf_sample_data *data,
6141 struct pt_regs *regs)
6143 if (event->hw.state & PERF_HES_STOPPED)
6146 * All tracepoints are from kernel-space.
6148 if (event->attr.exclude_kernel)
6151 if (!perf_tp_filter_match(event, data))
6157 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6158 struct pt_regs *regs, struct hlist_head *head, int rctx,
6159 struct task_struct *task)
6161 struct perf_sample_data data;
6162 struct perf_event *event;
6164 struct perf_raw_record raw = {
6169 perf_sample_data_init(&data, addr, 0);
6172 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6173 if (perf_tp_event_match(event, &data, regs))
6174 perf_swevent_event(event, count, &data, regs);
6178 * If we got specified a target task, also iterate its context and
6179 * deliver this event there too.
6181 if (task && task != current) {
6182 struct perf_event_context *ctx;
6183 struct trace_entry *entry = record;
6186 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6190 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6191 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6193 if (event->attr.config != entry->type)
6195 if (perf_tp_event_match(event, &data, regs))
6196 perf_swevent_event(event, count, &data, regs);
6202 perf_swevent_put_recursion_context(rctx);
6204 EXPORT_SYMBOL_GPL(perf_tp_event);
6206 static void tp_perf_event_destroy(struct perf_event *event)
6208 perf_trace_destroy(event);
6211 static int perf_tp_event_init(struct perf_event *event)
6215 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6219 * no branch sampling for tracepoint events
6221 if (has_branch_stack(event))
6224 err = perf_trace_init(event);
6228 event->destroy = tp_perf_event_destroy;
6233 static struct pmu perf_tracepoint = {
6234 .task_ctx_nr = perf_sw_context,
6236 .event_init = perf_tp_event_init,
6237 .add = perf_trace_add,
6238 .del = perf_trace_del,
6239 .start = perf_swevent_start,
6240 .stop = perf_swevent_stop,
6241 .read = perf_swevent_read,
6244 static inline void perf_tp_register(void)
6246 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6249 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6254 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6257 filter_str = strndup_user(arg, PAGE_SIZE);
6258 if (IS_ERR(filter_str))
6259 return PTR_ERR(filter_str);
6261 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6267 static void perf_event_free_filter(struct perf_event *event)
6269 ftrace_profile_free_filter(event);
6274 static inline void perf_tp_register(void)
6278 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6283 static void perf_event_free_filter(struct perf_event *event)
6287 #endif /* CONFIG_EVENT_TRACING */
6289 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6290 void perf_bp_event(struct perf_event *bp, void *data)
6292 struct perf_sample_data sample;
6293 struct pt_regs *regs = data;
6295 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6297 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6298 perf_swevent_event(bp, 1, &sample, regs);
6303 * hrtimer based swevent callback
6306 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6308 enum hrtimer_restart ret = HRTIMER_RESTART;
6309 struct perf_sample_data data;
6310 struct pt_regs *regs;
6311 struct perf_event *event;
6314 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6316 if (event->state != PERF_EVENT_STATE_ACTIVE)
6317 return HRTIMER_NORESTART;
6319 event->pmu->read(event);
6321 perf_sample_data_init(&data, 0, event->hw.last_period);
6322 regs = get_irq_regs();
6324 if (regs && !perf_exclude_event(event, regs)) {
6325 if (!(event->attr.exclude_idle && is_idle_task(current)))
6326 if (__perf_event_overflow(event, 1, &data, regs))
6327 ret = HRTIMER_NORESTART;
6330 period = max_t(u64, 10000, event->hw.sample_period);
6331 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6336 static void perf_swevent_start_hrtimer(struct perf_event *event)
6338 struct hw_perf_event *hwc = &event->hw;
6341 if (!is_sampling_event(event))
6344 period = local64_read(&hwc->period_left);
6349 local64_set(&hwc->period_left, 0);
6351 period = max_t(u64, 10000, hwc->sample_period);
6353 __hrtimer_start_range_ns(&hwc->hrtimer,
6354 ns_to_ktime(period), 0,
6355 HRTIMER_MODE_REL_PINNED, 0);
6358 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6360 struct hw_perf_event *hwc = &event->hw;
6362 if (is_sampling_event(event)) {
6363 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6364 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6366 hrtimer_cancel(&hwc->hrtimer);
6370 static void perf_swevent_init_hrtimer(struct perf_event *event)
6372 struct hw_perf_event *hwc = &event->hw;
6374 if (!is_sampling_event(event))
6377 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6378 hwc->hrtimer.function = perf_swevent_hrtimer;
6381 * Since hrtimers have a fixed rate, we can do a static freq->period
6382 * mapping and avoid the whole period adjust feedback stuff.
6384 if (event->attr.freq) {
6385 long freq = event->attr.sample_freq;
6387 event->attr.sample_period = NSEC_PER_SEC / freq;
6388 hwc->sample_period = event->attr.sample_period;
6389 local64_set(&hwc->period_left, hwc->sample_period);
6390 hwc->last_period = hwc->sample_period;
6391 event->attr.freq = 0;
6396 * Software event: cpu wall time clock
6399 static void cpu_clock_event_update(struct perf_event *event)
6404 now = local_clock();
6405 prev = local64_xchg(&event->hw.prev_count, now);
6406 local64_add(now - prev, &event->count);
6409 static void cpu_clock_event_start(struct perf_event *event, int flags)
6411 local64_set(&event->hw.prev_count, local_clock());
6412 perf_swevent_start_hrtimer(event);
6415 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6417 perf_swevent_cancel_hrtimer(event);
6418 cpu_clock_event_update(event);
6421 static int cpu_clock_event_add(struct perf_event *event, int flags)
6423 if (flags & PERF_EF_START)
6424 cpu_clock_event_start(event, flags);
6429 static void cpu_clock_event_del(struct perf_event *event, int flags)
6431 cpu_clock_event_stop(event, flags);
6434 static void cpu_clock_event_read(struct perf_event *event)
6436 cpu_clock_event_update(event);
6439 static int cpu_clock_event_init(struct perf_event *event)
6441 if (event->attr.type != PERF_TYPE_SOFTWARE)
6444 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6448 * no branch sampling for software events
6450 if (has_branch_stack(event))
6453 perf_swevent_init_hrtimer(event);
6458 static struct pmu perf_cpu_clock = {
6459 .task_ctx_nr = perf_sw_context,
6461 .event_init = cpu_clock_event_init,
6462 .add = cpu_clock_event_add,
6463 .del = cpu_clock_event_del,
6464 .start = cpu_clock_event_start,
6465 .stop = cpu_clock_event_stop,
6466 .read = cpu_clock_event_read,
6470 * Software event: task time clock
6473 static void task_clock_event_update(struct perf_event *event, u64 now)
6478 prev = local64_xchg(&event->hw.prev_count, now);
6480 local64_add(delta, &event->count);
6483 static void task_clock_event_start(struct perf_event *event, int flags)
6485 local64_set(&event->hw.prev_count, event->ctx->time);
6486 perf_swevent_start_hrtimer(event);
6489 static void task_clock_event_stop(struct perf_event *event, int flags)
6491 perf_swevent_cancel_hrtimer(event);
6492 task_clock_event_update(event, event->ctx->time);
6495 static int task_clock_event_add(struct perf_event *event, int flags)
6497 if (flags & PERF_EF_START)
6498 task_clock_event_start(event, flags);
6503 static void task_clock_event_del(struct perf_event *event, int flags)
6505 task_clock_event_stop(event, PERF_EF_UPDATE);
6508 static void task_clock_event_read(struct perf_event *event)
6510 u64 now = perf_clock();
6511 u64 delta = now - event->ctx->timestamp;
6512 u64 time = event->ctx->time + delta;
6514 task_clock_event_update(event, time);
6517 static int task_clock_event_init(struct perf_event *event)
6519 if (event->attr.type != PERF_TYPE_SOFTWARE)
6522 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6526 * no branch sampling for software events
6528 if (has_branch_stack(event))
6531 perf_swevent_init_hrtimer(event);
6536 static struct pmu perf_task_clock = {
6537 .task_ctx_nr = perf_sw_context,
6539 .event_init = task_clock_event_init,
6540 .add = task_clock_event_add,
6541 .del = task_clock_event_del,
6542 .start = task_clock_event_start,
6543 .stop = task_clock_event_stop,
6544 .read = task_clock_event_read,
6547 static void perf_pmu_nop_void(struct pmu *pmu)
6551 static int perf_pmu_nop_int(struct pmu *pmu)
6556 static void perf_pmu_start_txn(struct pmu *pmu)
6558 perf_pmu_disable(pmu);
6561 static int perf_pmu_commit_txn(struct pmu *pmu)
6563 perf_pmu_enable(pmu);
6567 static void perf_pmu_cancel_txn(struct pmu *pmu)
6569 perf_pmu_enable(pmu);
6572 static int perf_event_idx_default(struct perf_event *event)
6578 * Ensures all contexts with the same task_ctx_nr have the same
6579 * pmu_cpu_context too.
6581 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6588 list_for_each_entry(pmu, &pmus, entry) {
6589 if (pmu->task_ctx_nr == ctxn)
6590 return pmu->pmu_cpu_context;
6596 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6600 for_each_possible_cpu(cpu) {
6601 struct perf_cpu_context *cpuctx;
6603 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6605 if (cpuctx->unique_pmu == old_pmu)
6606 cpuctx->unique_pmu = pmu;
6610 static void free_pmu_context(struct pmu *pmu)
6614 mutex_lock(&pmus_lock);
6616 * Like a real lame refcount.
6618 list_for_each_entry(i, &pmus, entry) {
6619 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6620 update_pmu_context(i, pmu);
6625 free_percpu(pmu->pmu_cpu_context);
6627 mutex_unlock(&pmus_lock);
6629 static struct idr pmu_idr;
6632 type_show(struct device *dev, struct device_attribute *attr, char *page)
6634 struct pmu *pmu = dev_get_drvdata(dev);
6636 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6638 static DEVICE_ATTR_RO(type);
6641 perf_event_mux_interval_ms_show(struct device *dev,
6642 struct device_attribute *attr,
6645 struct pmu *pmu = dev_get_drvdata(dev);
6647 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6651 perf_event_mux_interval_ms_store(struct device *dev,
6652 struct device_attribute *attr,
6653 const char *buf, size_t count)
6655 struct pmu *pmu = dev_get_drvdata(dev);
6656 int timer, cpu, ret;
6658 ret = kstrtoint(buf, 0, &timer);
6665 /* same value, noting to do */
6666 if (timer == pmu->hrtimer_interval_ms)
6669 pmu->hrtimer_interval_ms = timer;
6671 /* update all cpuctx for this PMU */
6672 for_each_possible_cpu(cpu) {
6673 struct perf_cpu_context *cpuctx;
6674 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6675 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6677 if (hrtimer_active(&cpuctx->hrtimer))
6678 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6683 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6685 static struct attribute *pmu_dev_attrs[] = {
6686 &dev_attr_type.attr,
6687 &dev_attr_perf_event_mux_interval_ms.attr,
6690 ATTRIBUTE_GROUPS(pmu_dev);
6692 static int pmu_bus_running;
6693 static struct bus_type pmu_bus = {
6694 .name = "event_source",
6695 .dev_groups = pmu_dev_groups,
6698 static void pmu_dev_release(struct device *dev)
6703 static int pmu_dev_alloc(struct pmu *pmu)
6707 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6711 pmu->dev->groups = pmu->attr_groups;
6712 device_initialize(pmu->dev);
6713 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6717 dev_set_drvdata(pmu->dev, pmu);
6718 pmu->dev->bus = &pmu_bus;
6719 pmu->dev->release = pmu_dev_release;
6720 ret = device_add(pmu->dev);
6728 put_device(pmu->dev);
6732 static struct lock_class_key cpuctx_mutex;
6733 static struct lock_class_key cpuctx_lock;
6735 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6739 mutex_lock(&pmus_lock);
6741 pmu->pmu_disable_count = alloc_percpu(int);
6742 if (!pmu->pmu_disable_count)
6751 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6759 if (pmu_bus_running) {
6760 ret = pmu_dev_alloc(pmu);
6766 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6767 if (pmu->pmu_cpu_context)
6768 goto got_cpu_context;
6771 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6772 if (!pmu->pmu_cpu_context)
6775 for_each_possible_cpu(cpu) {
6776 struct perf_cpu_context *cpuctx;
6778 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6779 __perf_event_init_context(&cpuctx->ctx);
6780 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6781 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6782 cpuctx->ctx.type = cpu_context;
6783 cpuctx->ctx.pmu = pmu;
6785 __perf_cpu_hrtimer_init(cpuctx, cpu);
6787 INIT_LIST_HEAD(&cpuctx->rotation_list);
6788 cpuctx->unique_pmu = pmu;
6792 if (!pmu->start_txn) {
6793 if (pmu->pmu_enable) {
6795 * If we have pmu_enable/pmu_disable calls, install
6796 * transaction stubs that use that to try and batch
6797 * hardware accesses.
6799 pmu->start_txn = perf_pmu_start_txn;
6800 pmu->commit_txn = perf_pmu_commit_txn;
6801 pmu->cancel_txn = perf_pmu_cancel_txn;
6803 pmu->start_txn = perf_pmu_nop_void;
6804 pmu->commit_txn = perf_pmu_nop_int;
6805 pmu->cancel_txn = perf_pmu_nop_void;
6809 if (!pmu->pmu_enable) {
6810 pmu->pmu_enable = perf_pmu_nop_void;
6811 pmu->pmu_disable = perf_pmu_nop_void;
6814 if (!pmu->event_idx)
6815 pmu->event_idx = perf_event_idx_default;
6817 list_add_rcu(&pmu->entry, &pmus);
6820 mutex_unlock(&pmus_lock);
6825 device_del(pmu->dev);
6826 put_device(pmu->dev);
6829 if (pmu->type >= PERF_TYPE_MAX)
6830 idr_remove(&pmu_idr, pmu->type);
6833 free_percpu(pmu->pmu_disable_count);
6836 EXPORT_SYMBOL_GPL(perf_pmu_register);
6838 void perf_pmu_unregister(struct pmu *pmu)
6840 mutex_lock(&pmus_lock);
6841 list_del_rcu(&pmu->entry);
6842 mutex_unlock(&pmus_lock);
6845 * We dereference the pmu list under both SRCU and regular RCU, so
6846 * synchronize against both of those.
6848 synchronize_srcu(&pmus_srcu);
6851 free_percpu(pmu->pmu_disable_count);
6852 if (pmu->type >= PERF_TYPE_MAX)
6853 idr_remove(&pmu_idr, pmu->type);
6854 device_del(pmu->dev);
6855 put_device(pmu->dev);
6856 free_pmu_context(pmu);
6858 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6860 struct pmu *perf_init_event(struct perf_event *event)
6862 struct pmu *pmu = NULL;
6866 idx = srcu_read_lock(&pmus_srcu);
6869 pmu = idr_find(&pmu_idr, event->attr.type);
6872 if (!try_module_get(pmu->module)) {
6873 pmu = ERR_PTR(-ENODEV);
6877 ret = pmu->event_init(event);
6883 list_for_each_entry_rcu(pmu, &pmus, entry) {
6884 if (!try_module_get(pmu->module)) {
6885 pmu = ERR_PTR(-ENODEV);
6889 ret = pmu->event_init(event);
6893 if (ret != -ENOENT) {
6898 pmu = ERR_PTR(-ENOENT);
6900 srcu_read_unlock(&pmus_srcu, idx);
6905 static void account_event_cpu(struct perf_event *event, int cpu)
6910 if (has_branch_stack(event)) {
6911 if (!(event->attach_state & PERF_ATTACH_TASK))
6912 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6914 if (is_cgroup_event(event))
6915 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6918 static void account_event(struct perf_event *event)
6923 if (event->attach_state & PERF_ATTACH_TASK)
6924 static_key_slow_inc(&perf_sched_events.key);
6925 if (event->attr.mmap || event->attr.mmap_data)
6926 atomic_inc(&nr_mmap_events);
6927 if (event->attr.comm)
6928 atomic_inc(&nr_comm_events);
6929 if (event->attr.task)
6930 atomic_inc(&nr_task_events);
6931 if (event->attr.freq) {
6932 if (atomic_inc_return(&nr_freq_events) == 1)
6933 tick_nohz_full_kick_all();
6935 if (has_branch_stack(event))
6936 static_key_slow_inc(&perf_sched_events.key);
6937 if (is_cgroup_event(event))
6938 static_key_slow_inc(&perf_sched_events.key);
6940 account_event_cpu(event, event->cpu);
6944 * Allocate and initialize a event structure
6946 static struct perf_event *
6947 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6948 struct task_struct *task,
6949 struct perf_event *group_leader,
6950 struct perf_event *parent_event,
6951 perf_overflow_handler_t overflow_handler,
6955 struct perf_event *event;
6956 struct hw_perf_event *hwc;
6959 if ((unsigned)cpu >= nr_cpu_ids) {
6960 if (!task || cpu != -1)
6961 return ERR_PTR(-EINVAL);
6964 event = kzalloc(sizeof(*event), GFP_KERNEL);
6966 return ERR_PTR(-ENOMEM);
6969 * Single events are their own group leaders, with an
6970 * empty sibling list:
6973 group_leader = event;
6975 mutex_init(&event->child_mutex);
6976 INIT_LIST_HEAD(&event->child_list);
6978 INIT_LIST_HEAD(&event->group_entry);
6979 INIT_LIST_HEAD(&event->event_entry);
6980 INIT_LIST_HEAD(&event->sibling_list);
6981 INIT_LIST_HEAD(&event->rb_entry);
6982 INIT_LIST_HEAD(&event->active_entry);
6983 INIT_HLIST_NODE(&event->hlist_entry);
6986 init_waitqueue_head(&event->waitq);
6987 init_irq_work(&event->pending, perf_pending_event);
6989 mutex_init(&event->mmap_mutex);
6991 atomic_long_set(&event->refcount, 1);
6993 event->attr = *attr;
6994 event->group_leader = group_leader;
6998 event->parent = parent_event;
7000 event->ns = get_pid_ns(task_active_pid_ns(current));
7001 event->id = atomic64_inc_return(&perf_event_id);
7003 event->state = PERF_EVENT_STATE_INACTIVE;
7006 event->attach_state = PERF_ATTACH_TASK;
7008 if (attr->type == PERF_TYPE_TRACEPOINT)
7009 event->hw.tp_target = task;
7010 #ifdef CONFIG_HAVE_HW_BREAKPOINT
7012 * hw_breakpoint is a bit difficult here..
7014 else if (attr->type == PERF_TYPE_BREAKPOINT)
7015 event->hw.bp_target = task;
7019 if (!overflow_handler && parent_event) {
7020 overflow_handler = parent_event->overflow_handler;
7021 context = parent_event->overflow_handler_context;
7024 event->overflow_handler = overflow_handler;
7025 event->overflow_handler_context = context;
7027 perf_event__state_init(event);
7032 hwc->sample_period = attr->sample_period;
7033 if (attr->freq && attr->sample_freq)
7034 hwc->sample_period = 1;
7035 hwc->last_period = hwc->sample_period;
7037 local64_set(&hwc->period_left, hwc->sample_period);
7040 * we currently do not support PERF_FORMAT_GROUP on inherited events
7042 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7045 pmu = perf_init_event(event);
7048 else if (IS_ERR(pmu)) {
7053 if (!event->parent) {
7054 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7055 err = get_callchain_buffers();
7065 event->destroy(event);
7066 module_put(pmu->module);
7069 put_pid_ns(event->ns);
7072 return ERR_PTR(err);
7075 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7076 struct perf_event_attr *attr)
7081 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7085 * zero the full structure, so that a short copy will be nice.
7087 memset(attr, 0, sizeof(*attr));
7089 ret = get_user(size, &uattr->size);
7093 if (size > PAGE_SIZE) /* silly large */
7096 if (!size) /* abi compat */
7097 size = PERF_ATTR_SIZE_VER0;
7099 if (size < PERF_ATTR_SIZE_VER0)
7103 * If we're handed a bigger struct than we know of,
7104 * ensure all the unknown bits are 0 - i.e. new
7105 * user-space does not rely on any kernel feature
7106 * extensions we dont know about yet.
7108 if (size > sizeof(*attr)) {
7109 unsigned char __user *addr;
7110 unsigned char __user *end;
7113 addr = (void __user *)uattr + sizeof(*attr);
7114 end = (void __user *)uattr + size;
7116 for (; addr < end; addr++) {
7117 ret = get_user(val, addr);
7123 size = sizeof(*attr);
7126 ret = copy_from_user(attr, uattr, size);
7130 if (attr->__reserved_1)
7133 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7136 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7139 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7140 u64 mask = attr->branch_sample_type;
7142 /* only using defined bits */
7143 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7146 /* at least one branch bit must be set */
7147 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7150 /* propagate priv level, when not set for branch */
7151 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7153 /* exclude_kernel checked on syscall entry */
7154 if (!attr->exclude_kernel)
7155 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7157 if (!attr->exclude_user)
7158 mask |= PERF_SAMPLE_BRANCH_USER;
7160 if (!attr->exclude_hv)
7161 mask |= PERF_SAMPLE_BRANCH_HV;
7163 * adjust user setting (for HW filter setup)
7165 attr->branch_sample_type = mask;
7167 /* privileged levels capture (kernel, hv): check permissions */
7168 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7169 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7173 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7174 ret = perf_reg_validate(attr->sample_regs_user);
7179 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7180 if (!arch_perf_have_user_stack_dump())
7184 * We have __u32 type for the size, but so far
7185 * we can only use __u16 as maximum due to the
7186 * __u16 sample size limit.
7188 if (attr->sample_stack_user >= USHRT_MAX)
7190 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7194 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7195 ret = perf_reg_validate(attr->sample_regs_intr);
7200 put_user(sizeof(*attr), &uattr->size);
7206 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7208 struct ring_buffer *rb = NULL;
7214 /* don't allow circular references */
7215 if (event == output_event)
7219 * Don't allow cross-cpu buffers
7221 if (output_event->cpu != event->cpu)
7225 * If its not a per-cpu rb, it must be the same task.
7227 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7231 mutex_lock(&event->mmap_mutex);
7232 /* Can't redirect output if we've got an active mmap() */
7233 if (atomic_read(&event->mmap_count))
7237 /* get the rb we want to redirect to */
7238 rb = ring_buffer_get(output_event);
7243 ring_buffer_attach(event, rb);
7247 mutex_unlock(&event->mmap_mutex);
7254 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7256 * @attr_uptr: event_id type attributes for monitoring/sampling
7259 * @group_fd: group leader event fd
7261 SYSCALL_DEFINE5(perf_event_open,
7262 struct perf_event_attr __user *, attr_uptr,
7263 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7265 struct perf_event *group_leader = NULL, *output_event = NULL;
7266 struct perf_event *event, *sibling;
7267 struct perf_event_attr attr;
7268 struct perf_event_context *ctx;
7269 struct file *event_file = NULL;
7270 struct fd group = {NULL, 0};
7271 struct task_struct *task = NULL;
7276 int f_flags = O_RDWR;
7278 /* for future expandability... */
7279 if (flags & ~PERF_FLAG_ALL)
7282 err = perf_copy_attr(attr_uptr, &attr);
7286 if (!attr.exclude_kernel) {
7287 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7292 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7295 if (attr.sample_period & (1ULL << 63))
7300 * In cgroup mode, the pid argument is used to pass the fd
7301 * opened to the cgroup directory in cgroupfs. The cpu argument
7302 * designates the cpu on which to monitor threads from that
7305 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7308 if (flags & PERF_FLAG_FD_CLOEXEC)
7309 f_flags |= O_CLOEXEC;
7311 event_fd = get_unused_fd_flags(f_flags);
7315 if (group_fd != -1) {
7316 err = perf_fget_light(group_fd, &group);
7319 group_leader = group.file->private_data;
7320 if (flags & PERF_FLAG_FD_OUTPUT)
7321 output_event = group_leader;
7322 if (flags & PERF_FLAG_FD_NO_GROUP)
7323 group_leader = NULL;
7326 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7327 task = find_lively_task_by_vpid(pid);
7329 err = PTR_ERR(task);
7334 if (task && group_leader &&
7335 group_leader->attr.inherit != attr.inherit) {
7342 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7344 if (IS_ERR(event)) {
7345 err = PTR_ERR(event);
7349 if (flags & PERF_FLAG_PID_CGROUP) {
7350 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7352 __free_event(event);
7357 if (is_sampling_event(event)) {
7358 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7364 account_event(event);
7367 * Special case software events and allow them to be part of
7368 * any hardware group.
7373 (is_software_event(event) != is_software_event(group_leader))) {
7374 if (is_software_event(event)) {
7376 * If event and group_leader are not both a software
7377 * event, and event is, then group leader is not.
7379 * Allow the addition of software events to !software
7380 * groups, this is safe because software events never
7383 pmu = group_leader->pmu;
7384 } else if (is_software_event(group_leader) &&
7385 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7387 * In case the group is a pure software group, and we
7388 * try to add a hardware event, move the whole group to
7389 * the hardware context.
7396 * Get the target context (task or percpu):
7398 ctx = find_get_context(pmu, task, event->cpu);
7405 put_task_struct(task);
7410 * Look up the group leader (we will attach this event to it):
7416 * Do not allow a recursive hierarchy (this new sibling
7417 * becoming part of another group-sibling):
7419 if (group_leader->group_leader != group_leader)
7422 * Do not allow to attach to a group in a different
7423 * task or CPU context:
7426 if (group_leader->ctx->type != ctx->type)
7429 if (group_leader->ctx != ctx)
7434 * Only a group leader can be exclusive or pinned
7436 if (attr.exclusive || attr.pinned)
7441 err = perf_event_set_output(event, output_event);
7446 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7448 if (IS_ERR(event_file)) {
7449 err = PTR_ERR(event_file);
7454 struct perf_event_context *gctx = group_leader->ctx;
7456 mutex_lock(&gctx->mutex);
7457 perf_remove_from_context(group_leader, false);
7460 * Removing from the context ends up with disabled
7461 * event. What we want here is event in the initial
7462 * startup state, ready to be add into new context.
7464 perf_event__state_init(group_leader);
7465 list_for_each_entry(sibling, &group_leader->sibling_list,
7467 perf_remove_from_context(sibling, false);
7468 perf_event__state_init(sibling);
7471 mutex_unlock(&gctx->mutex);
7475 WARN_ON_ONCE(ctx->parent_ctx);
7476 mutex_lock(&ctx->mutex);
7480 perf_install_in_context(ctx, group_leader, event->cpu);
7482 list_for_each_entry(sibling, &group_leader->sibling_list,
7484 perf_install_in_context(ctx, sibling, event->cpu);
7489 perf_install_in_context(ctx, event, event->cpu);
7490 perf_unpin_context(ctx);
7491 mutex_unlock(&ctx->mutex);
7495 event->owner = current;
7497 mutex_lock(¤t->perf_event_mutex);
7498 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7499 mutex_unlock(¤t->perf_event_mutex);
7502 * Precalculate sample_data sizes
7504 perf_event__header_size(event);
7505 perf_event__id_header_size(event);
7508 * Drop the reference on the group_event after placing the
7509 * new event on the sibling_list. This ensures destruction
7510 * of the group leader will find the pointer to itself in
7511 * perf_group_detach().
7514 fd_install(event_fd, event_file);
7518 perf_unpin_context(ctx);
7526 put_task_struct(task);
7530 put_unused_fd(event_fd);
7535 * perf_event_create_kernel_counter
7537 * @attr: attributes of the counter to create
7538 * @cpu: cpu in which the counter is bound
7539 * @task: task to profile (NULL for percpu)
7542 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7543 struct task_struct *task,
7544 perf_overflow_handler_t overflow_handler,
7547 struct perf_event_context *ctx;
7548 struct perf_event *event;
7552 * Get the target context (task or percpu):
7555 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7556 overflow_handler, context);
7557 if (IS_ERR(event)) {
7558 err = PTR_ERR(event);
7562 /* Mark owner so we could distinguish it from user events. */
7563 event->owner = EVENT_OWNER_KERNEL;
7565 account_event(event);
7567 ctx = find_get_context(event->pmu, task, cpu);
7573 WARN_ON_ONCE(ctx->parent_ctx);
7574 mutex_lock(&ctx->mutex);
7575 perf_install_in_context(ctx, event, cpu);
7576 perf_unpin_context(ctx);
7577 mutex_unlock(&ctx->mutex);
7584 return ERR_PTR(err);
7586 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7588 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7590 struct perf_event_context *src_ctx;
7591 struct perf_event_context *dst_ctx;
7592 struct perf_event *event, *tmp;
7595 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7596 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7598 mutex_lock(&src_ctx->mutex);
7599 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7601 perf_remove_from_context(event, false);
7602 unaccount_event_cpu(event, src_cpu);
7604 list_add(&event->migrate_entry, &events);
7606 mutex_unlock(&src_ctx->mutex);
7610 mutex_lock(&dst_ctx->mutex);
7611 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7612 list_del(&event->migrate_entry);
7613 if (event->state >= PERF_EVENT_STATE_OFF)
7614 event->state = PERF_EVENT_STATE_INACTIVE;
7615 account_event_cpu(event, dst_cpu);
7616 perf_install_in_context(dst_ctx, event, dst_cpu);
7619 mutex_unlock(&dst_ctx->mutex);
7621 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7623 static void sync_child_event(struct perf_event *child_event,
7624 struct task_struct *child)
7626 struct perf_event *parent_event = child_event->parent;
7629 if (child_event->attr.inherit_stat)
7630 perf_event_read_event(child_event, child);
7632 child_val = perf_event_count(child_event);
7635 * Add back the child's count to the parent's count:
7637 atomic64_add(child_val, &parent_event->child_count);
7638 atomic64_add(child_event->total_time_enabled,
7639 &parent_event->child_total_time_enabled);
7640 atomic64_add(child_event->total_time_running,
7641 &parent_event->child_total_time_running);
7644 * Remove this event from the parent's list
7646 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7647 mutex_lock(&parent_event->child_mutex);
7648 list_del_init(&child_event->child_list);
7649 mutex_unlock(&parent_event->child_mutex);
7652 * Make sure user/parent get notified, that we just
7655 perf_event_wakeup(parent_event);
7658 * Release the parent event, if this was the last
7661 put_event(parent_event);
7665 __perf_event_exit_task(struct perf_event *child_event,
7666 struct perf_event_context *child_ctx,
7667 struct task_struct *child)
7670 * Do not destroy the 'original' grouping; because of the context
7671 * switch optimization the original events could've ended up in a
7672 * random child task.
7674 * If we were to destroy the original group, all group related
7675 * operations would cease to function properly after this random
7678 * Do destroy all inherited groups, we don't care about those
7679 * and being thorough is better.
7681 perf_remove_from_context(child_event, !!child_event->parent);
7684 * It can happen that the parent exits first, and has events
7685 * that are still around due to the child reference. These
7686 * events need to be zapped.
7688 if (child_event->parent) {
7689 sync_child_event(child_event, child);
7690 free_event(child_event);
7692 child_event->state = PERF_EVENT_STATE_EXIT;
7693 perf_event_wakeup(child_event);
7697 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7699 struct perf_event *child_event, *next;
7700 struct perf_event_context *child_ctx, *clone_ctx = NULL;
7701 unsigned long flags;
7703 if (likely(!child->perf_event_ctxp[ctxn])) {
7704 perf_event_task(child, NULL, 0);
7708 local_irq_save(flags);
7710 * We can't reschedule here because interrupts are disabled,
7711 * and either child is current or it is a task that can't be
7712 * scheduled, so we are now safe from rescheduling changing
7715 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7718 * Take the context lock here so that if find_get_context is
7719 * reading child->perf_event_ctxp, we wait until it has
7720 * incremented the context's refcount before we do put_ctx below.
7722 raw_spin_lock(&child_ctx->lock);
7723 task_ctx_sched_out(child_ctx);
7724 child->perf_event_ctxp[ctxn] = NULL;
7727 * If this context is a clone; unclone it so it can't get
7728 * swapped to another process while we're removing all
7729 * the events from it.
7731 clone_ctx = unclone_ctx(child_ctx);
7732 update_context_time(child_ctx);
7733 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7739 * Report the task dead after unscheduling the events so that we
7740 * won't get any samples after PERF_RECORD_EXIT. We can however still
7741 * get a few PERF_RECORD_READ events.
7743 perf_event_task(child, child_ctx, 0);
7746 * We can recurse on the same lock type through:
7748 * __perf_event_exit_task()
7749 * sync_child_event()
7751 * mutex_lock(&ctx->mutex)
7753 * But since its the parent context it won't be the same instance.
7755 mutex_lock(&child_ctx->mutex);
7757 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7758 __perf_event_exit_task(child_event, child_ctx, child);
7760 mutex_unlock(&child_ctx->mutex);
7766 * When a child task exits, feed back event values to parent events.
7768 void perf_event_exit_task(struct task_struct *child)
7770 struct perf_event *event, *tmp;
7773 mutex_lock(&child->perf_event_mutex);
7774 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7776 list_del_init(&event->owner_entry);
7779 * Ensure the list deletion is visible before we clear
7780 * the owner, closes a race against perf_release() where
7781 * we need to serialize on the owner->perf_event_mutex.
7784 event->owner = NULL;
7786 mutex_unlock(&child->perf_event_mutex);
7788 for_each_task_context_nr(ctxn)
7789 perf_event_exit_task_context(child, ctxn);
7792 static void perf_free_event(struct perf_event *event,
7793 struct perf_event_context *ctx)
7795 struct perf_event *parent = event->parent;
7797 if (WARN_ON_ONCE(!parent))
7800 mutex_lock(&parent->child_mutex);
7801 list_del_init(&event->child_list);
7802 mutex_unlock(&parent->child_mutex);
7806 perf_group_detach(event);
7807 list_del_event(event, ctx);
7812 * free an unexposed, unused context as created by inheritance by
7813 * perf_event_init_task below, used by fork() in case of fail.
7815 void perf_event_free_task(struct task_struct *task)
7817 struct perf_event_context *ctx;
7818 struct perf_event *event, *tmp;
7821 for_each_task_context_nr(ctxn) {
7822 ctx = task->perf_event_ctxp[ctxn];
7826 mutex_lock(&ctx->mutex);
7828 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7830 perf_free_event(event, ctx);
7832 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7834 perf_free_event(event, ctx);
7836 if (!list_empty(&ctx->pinned_groups) ||
7837 !list_empty(&ctx->flexible_groups))
7840 mutex_unlock(&ctx->mutex);
7846 void perf_event_delayed_put(struct task_struct *task)
7850 for_each_task_context_nr(ctxn)
7851 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7855 * inherit a event from parent task to child task:
7857 static struct perf_event *
7858 inherit_event(struct perf_event *parent_event,
7859 struct task_struct *parent,
7860 struct perf_event_context *parent_ctx,
7861 struct task_struct *child,
7862 struct perf_event *group_leader,
7863 struct perf_event_context *child_ctx)
7865 enum perf_event_active_state parent_state = parent_event->state;
7866 struct perf_event *child_event;
7867 unsigned long flags;
7870 * Instead of creating recursive hierarchies of events,
7871 * we link inherited events back to the original parent,
7872 * which has a filp for sure, which we use as the reference
7875 if (parent_event->parent)
7876 parent_event = parent_event->parent;
7878 child_event = perf_event_alloc(&parent_event->attr,
7881 group_leader, parent_event,
7883 if (IS_ERR(child_event))
7886 if (is_orphaned_event(parent_event) ||
7887 !atomic_long_inc_not_zero(&parent_event->refcount)) {
7888 free_event(child_event);
7895 * Make the child state follow the state of the parent event,
7896 * not its attr.disabled bit. We hold the parent's mutex,
7897 * so we won't race with perf_event_{en, dis}able_family.
7899 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
7900 child_event->state = PERF_EVENT_STATE_INACTIVE;
7902 child_event->state = PERF_EVENT_STATE_OFF;
7904 if (parent_event->attr.freq) {
7905 u64 sample_period = parent_event->hw.sample_period;
7906 struct hw_perf_event *hwc = &child_event->hw;
7908 hwc->sample_period = sample_period;
7909 hwc->last_period = sample_period;
7911 local64_set(&hwc->period_left, sample_period);
7914 child_event->ctx = child_ctx;
7915 child_event->overflow_handler = parent_event->overflow_handler;
7916 child_event->overflow_handler_context
7917 = parent_event->overflow_handler_context;
7920 * Precalculate sample_data sizes
7922 perf_event__header_size(child_event);
7923 perf_event__id_header_size(child_event);
7926 * Link it up in the child's context:
7928 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7929 add_event_to_ctx(child_event, child_ctx);
7930 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7933 * Link this into the parent event's child list
7935 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7936 mutex_lock(&parent_event->child_mutex);
7937 list_add_tail(&child_event->child_list, &parent_event->child_list);
7938 mutex_unlock(&parent_event->child_mutex);
7943 static int inherit_group(struct perf_event *parent_event,
7944 struct task_struct *parent,
7945 struct perf_event_context *parent_ctx,
7946 struct task_struct *child,
7947 struct perf_event_context *child_ctx)
7949 struct perf_event *leader;
7950 struct perf_event *sub;
7951 struct perf_event *child_ctr;
7953 leader = inherit_event(parent_event, parent, parent_ctx,
7954 child, NULL, child_ctx);
7956 return PTR_ERR(leader);
7957 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7958 child_ctr = inherit_event(sub, parent, parent_ctx,
7959 child, leader, child_ctx);
7960 if (IS_ERR(child_ctr))
7961 return PTR_ERR(child_ctr);
7967 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7968 struct perf_event_context *parent_ctx,
7969 struct task_struct *child, int ctxn,
7973 struct perf_event_context *child_ctx;
7975 if (!event->attr.inherit) {
7980 child_ctx = child->perf_event_ctxp[ctxn];
7983 * This is executed from the parent task context, so
7984 * inherit events that have been marked for cloning.
7985 * First allocate and initialize a context for the
7989 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7993 child->perf_event_ctxp[ctxn] = child_ctx;
7996 ret = inherit_group(event, parent, parent_ctx,
8006 * Initialize the perf_event context in task_struct
8008 static int perf_event_init_context(struct task_struct *child, int ctxn)
8010 struct perf_event_context *child_ctx, *parent_ctx;
8011 struct perf_event_context *cloned_ctx;
8012 struct perf_event *event;
8013 struct task_struct *parent = current;
8014 int inherited_all = 1;
8015 unsigned long flags;
8018 if (likely(!parent->perf_event_ctxp[ctxn]))
8022 * If the parent's context is a clone, pin it so it won't get
8025 parent_ctx = perf_pin_task_context(parent, ctxn);
8030 * No need to check if parent_ctx != NULL here; since we saw
8031 * it non-NULL earlier, the only reason for it to become NULL
8032 * is if we exit, and since we're currently in the middle of
8033 * a fork we can't be exiting at the same time.
8037 * Lock the parent list. No need to lock the child - not PID
8038 * hashed yet and not running, so nobody can access it.
8040 mutex_lock(&parent_ctx->mutex);
8043 * We dont have to disable NMIs - we are only looking at
8044 * the list, not manipulating it:
8046 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8047 ret = inherit_task_group(event, parent, parent_ctx,
8048 child, ctxn, &inherited_all);
8054 * We can't hold ctx->lock when iterating the ->flexible_group list due
8055 * to allocations, but we need to prevent rotation because
8056 * rotate_ctx() will change the list from interrupt context.
8058 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8059 parent_ctx->rotate_disable = 1;
8060 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8062 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8063 ret = inherit_task_group(event, parent, parent_ctx,
8064 child, ctxn, &inherited_all);
8069 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8070 parent_ctx->rotate_disable = 0;
8072 child_ctx = child->perf_event_ctxp[ctxn];
8074 if (child_ctx && inherited_all) {
8076 * Mark the child context as a clone of the parent
8077 * context, or of whatever the parent is a clone of.
8079 * Note that if the parent is a clone, the holding of
8080 * parent_ctx->lock avoids it from being uncloned.
8082 cloned_ctx = parent_ctx->parent_ctx;
8084 child_ctx->parent_ctx = cloned_ctx;
8085 child_ctx->parent_gen = parent_ctx->parent_gen;
8087 child_ctx->parent_ctx = parent_ctx;
8088 child_ctx->parent_gen = parent_ctx->generation;
8090 get_ctx(child_ctx->parent_ctx);
8093 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8094 mutex_unlock(&parent_ctx->mutex);
8096 perf_unpin_context(parent_ctx);
8097 put_ctx(parent_ctx);
8103 * Initialize the perf_event context in task_struct
8105 int perf_event_init_task(struct task_struct *child)
8109 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8110 mutex_init(&child->perf_event_mutex);
8111 INIT_LIST_HEAD(&child->perf_event_list);
8113 for_each_task_context_nr(ctxn) {
8114 ret = perf_event_init_context(child, ctxn);
8116 perf_event_free_task(child);
8124 static void __init perf_event_init_all_cpus(void)
8126 struct swevent_htable *swhash;
8129 for_each_possible_cpu(cpu) {
8130 swhash = &per_cpu(swevent_htable, cpu);
8131 mutex_init(&swhash->hlist_mutex);
8132 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
8136 static void perf_event_init_cpu(int cpu)
8138 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8140 mutex_lock(&swhash->hlist_mutex);
8141 swhash->online = true;
8142 if (swhash->hlist_refcount > 0) {
8143 struct swevent_hlist *hlist;
8145 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8147 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8149 mutex_unlock(&swhash->hlist_mutex);
8152 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8153 static void perf_pmu_rotate_stop(struct pmu *pmu)
8155 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
8157 WARN_ON(!irqs_disabled());
8159 list_del_init(&cpuctx->rotation_list);
8162 static void __perf_event_exit_context(void *__info)
8164 struct remove_event re = { .detach_group = true };
8165 struct perf_event_context *ctx = __info;
8167 perf_pmu_rotate_stop(ctx->pmu);
8170 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8171 __perf_remove_from_context(&re);
8175 static void perf_event_exit_cpu_context(int cpu)
8177 struct perf_event_context *ctx;
8181 idx = srcu_read_lock(&pmus_srcu);
8182 list_for_each_entry_rcu(pmu, &pmus, entry) {
8183 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8185 mutex_lock(&ctx->mutex);
8186 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8187 mutex_unlock(&ctx->mutex);
8189 srcu_read_unlock(&pmus_srcu, idx);
8192 static void perf_event_exit_cpu(int cpu)
8194 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8196 perf_event_exit_cpu_context(cpu);
8198 mutex_lock(&swhash->hlist_mutex);
8199 swhash->online = false;
8200 swevent_hlist_release(swhash);
8201 mutex_unlock(&swhash->hlist_mutex);
8204 static inline void perf_event_exit_cpu(int cpu) { }
8208 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8212 for_each_online_cpu(cpu)
8213 perf_event_exit_cpu(cpu);
8219 * Run the perf reboot notifier at the very last possible moment so that
8220 * the generic watchdog code runs as long as possible.
8222 static struct notifier_block perf_reboot_notifier = {
8223 .notifier_call = perf_reboot,
8224 .priority = INT_MIN,
8228 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8230 unsigned int cpu = (long)hcpu;
8232 switch (action & ~CPU_TASKS_FROZEN) {
8234 case CPU_UP_PREPARE:
8235 case CPU_DOWN_FAILED:
8236 perf_event_init_cpu(cpu);
8239 case CPU_UP_CANCELED:
8240 case CPU_DOWN_PREPARE:
8241 perf_event_exit_cpu(cpu);
8250 void __init perf_event_init(void)
8256 perf_event_init_all_cpus();
8257 init_srcu_struct(&pmus_srcu);
8258 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8259 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8260 perf_pmu_register(&perf_task_clock, NULL, -1);
8262 perf_cpu_notifier(perf_cpu_notify);
8263 register_reboot_notifier(&perf_reboot_notifier);
8265 ret = init_hw_breakpoint();
8266 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8268 /* do not patch jump label more than once per second */
8269 jump_label_rate_limit(&perf_sched_events, HZ);
8272 * Build time assertion that we keep the data_head at the intended
8273 * location. IOW, validation we got the __reserved[] size right.
8275 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8279 static int __init perf_event_sysfs_init(void)
8284 mutex_lock(&pmus_lock);
8286 ret = bus_register(&pmu_bus);
8290 list_for_each_entry(pmu, &pmus, entry) {
8291 if (!pmu->name || pmu->type < 0)
8294 ret = pmu_dev_alloc(pmu);
8295 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8297 pmu_bus_running = 1;
8301 mutex_unlock(&pmus_lock);
8305 device_initcall(perf_event_sysfs_init);
8307 #ifdef CONFIG_CGROUP_PERF
8308 static struct cgroup_subsys_state *
8309 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8311 struct perf_cgroup *jc;
8313 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8315 return ERR_PTR(-ENOMEM);
8317 jc->info = alloc_percpu(struct perf_cgroup_info);
8320 return ERR_PTR(-ENOMEM);
8326 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8328 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8330 free_percpu(jc->info);
8334 static int __perf_cgroup_move(void *info)
8336 struct task_struct *task = info;
8337 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8341 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8342 struct cgroup_taskset *tset)
8344 struct task_struct *task;
8346 cgroup_taskset_for_each(task, tset)
8347 task_function_call(task, __perf_cgroup_move, task);
8350 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8351 struct cgroup_subsys_state *old_css,
8352 struct task_struct *task)
8355 * cgroup_exit() is called in the copy_process() failure path.
8356 * Ignore this case since the task hasn't ran yet, this avoids
8357 * trying to poke a half freed task state from generic code.
8359 if (!(task->flags & PF_EXITING))
8362 task_function_call(task, __perf_cgroup_move, task);
8365 struct cgroup_subsys perf_event_cgrp_subsys = {
8366 .css_alloc = perf_cgroup_css_alloc,
8367 .css_free = perf_cgroup_css_free,
8368 .exit = perf_cgroup_exit,
8369 .attach = perf_cgroup_attach,
8371 #endif /* CONFIG_CGROUP_PERF */