projects / deliverable / linux.git / blob
? search:


dac4c2831d821038a337a9450f7f0dfc73cef8a0
[deliverable/linux.git] / include / linux / perf_event.h
1 /*
2 * Performance events:
3 *
4 * Copyright (C) 2008-2009, 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
7 *
8 * Data type definitions, declarations, prototypes.
9 *
10 * Started by: Thomas Gleixner and Ingo Molnar
11 *
12 * For licencing details see kernel-base/COPYING
13 */
14 #ifndef _LINUX_PERF_EVENT_H
15 #define _LINUX_PERF_EVENT_H
16
17 #include <uapi/linux/perf_event.h>
18
19 /*
20 * Kernel-internal data types and definitions:
21 */
22
23 #ifdef CONFIG_PERF_EVENTS
24 # include <asm/perf_event.h>
25 # include <asm/local64.h>
26 #endif
27
28 struct perf_guest_info_callbacks {
29 int (*is_in_guest)(void);
30 int (*is_user_mode)(void);
31 unsigned long (*get_guest_ip)(void);
32 };
33
34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
35 #include <asm/hw_breakpoint.h>
36 #endif
37
38 #include <linux/list.h>
39 #include <linux/mutex.h>
40 #include <linux/rculist.h>
41 #include <linux/rcupdate.h>
42 #include <linux/spinlock.h>
43 #include <linux/hrtimer.h>
44 #include <linux/fs.h>
45 #include <linux/pid_namespace.h>
46 #include <linux/workqueue.h>
47 #include <linux/ftrace.h>
48 #include <linux/cpu.h>
49 #include <linux/irq_work.h>
50 #include <linux/static_key.h>
51 #include <linux/jump_label_ratelimit.h>
52 #include <linux/atomic.h>
53 #include <linux/sysfs.h>
54 #include <linux/perf_regs.h>
55 #include <linux/workqueue.h>
56 #include <linux/cgroup.h>
57 #include <asm/local.h>
58
59 struct perf_callchain_entry {
60 __u64 nr;
61 __u64 ip[PERF_MAX_STACK_DEPTH];
62 };
63
64 struct perf_raw_record {
65 u32 size;
66 void *data;
67 };
68
69 /*
70 * branch stack layout:
71 * nr: number of taken branches stored in entries[]
72 *
73 * Note that nr can vary from sample to sample
74 * branches (to, from) are stored from most recent
75 * to least recent, i.e., entries[0] contains the most
76 * recent branch.
77 */
78 struct perf_branch_stack {
79 __u64 nr;
80 struct perf_branch_entry entries[0];
81 };
82
83 struct task_struct;
84
85 /*
86 * extra PMU register associated with an event
87 */
88 struct hw_perf_event_extra {
89 u64 config; /* register value */
90 unsigned int reg; /* register address or index */
91 int alloc; /* extra register already allocated */
92 int idx; /* index in shared_regs->regs[] */
93 };
94
95 struct event_constraint;
96
97 /**
98 * struct hw_perf_event - performance event hardware details:
99 */
100 struct hw_perf_event {
101 #ifdef CONFIG_PERF_EVENTS
102 union {
103 struct { /* hardware */
104 u64 config;
105 u64 last_tag;
106 unsigned long config_base;
107 unsigned long event_base;
108 int event_base_rdpmc;
109 int idx;
110 int last_cpu;
111 int flags;
112
113 struct hw_perf_event_extra extra_reg;
114 struct hw_perf_event_extra branch_reg;
115
116 struct event_constraint *constraint;
117 };
118 struct { /* software */
119 struct hrtimer hrtimer;
120 };
121 struct { /* tracepoint */
122 struct task_struct *tp_target;
123 /* for tp_event->class */
124 struct list_head tp_list;
125 };
126 struct { /* intel_cqm */
127 int cqm_state;
128 int cqm_rmid;
129 struct list_head cqm_events_entry;
130 struct list_head cqm_groups_entry;
131 struct list_head cqm_group_entry;
132 struct task_struct *cqm_target;
133 };
134 #ifdef CONFIG_HAVE_HW_BREAKPOINT
135 struct { /* breakpoint */
136 /*
137 * Crufty hack to avoid the chicken and egg
138 * problem hw_breakpoint has with context
139 * creation and event initalization.
140 */
141 struct task_struct *bp_target;
142 struct arch_hw_breakpoint info;
143 struct list_head bp_list;
144 };
145 #endif
146 };
147 int state;
148 local64_t prev_count;
149 u64 sample_period;
150 u64 last_period;
151 local64_t period_left;
152 u64 interrupts_seq;
153 u64 interrupts;
154
155 u64 freq_time_stamp;
156 u64 freq_count_stamp;
157 #endif
158 };
159
160 /*
161 * hw_perf_event::state flags
162 */
163 #define PERF_HES_STOPPED 0x01 /* the counter is stopped */
164 #define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
165 #define PERF_HES_ARCH 0x04
166
167 struct perf_event;
168
169 /*
170 * Common implementation detail of pmu::{start,commit,cancel}_txn
171 */
172 #define PERF_EVENT_TXN 0x1
173
174 /**
175 * pmu::capabilities flags
176 */
177 #define PERF_PMU_CAP_NO_INTERRUPT 0x01
178
179 /**
180 * struct pmu - generic performance monitoring unit
181 */
182 struct pmu {
183 struct list_head entry;
184
185 struct module *module;
186 struct device *dev;
187 const struct attribute_group **attr_groups;
188 const char *name;
189 int type;
190
191 /*
192 * various common per-pmu feature flags
193 */
194 int capabilities;
195
196 int * __percpu pmu_disable_count;
197 struct perf_cpu_context * __percpu pmu_cpu_context;
198 int task_ctx_nr;
199 int hrtimer_interval_ms;
200
201 /*
202 * Fully disable/enable this PMU, can be used to protect from the PMI
203 * as well as for lazy/batch writing of the MSRs.
204 */
205 void (*pmu_enable) (struct pmu *pmu); /* optional */
206 void (*pmu_disable) (struct pmu *pmu); /* optional */
207
208 /*
209 * Try and initialize the event for this PMU.
210 * Should return -ENOENT when the @event doesn't match this PMU.
211 */
212 int (*event_init) (struct perf_event *event);
213
214 /*
215 * Notification that the event was mapped or unmapped. Called
216 * in the context of the mapping task.
217 */
218 void (*event_mapped) (struct perf_event *event); /*optional*/
219 void (*event_unmapped) (struct perf_event *event); /*optional*/
220
221 #define PERF_EF_START 0x01 /* start the counter when adding */
222 #define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
223 #define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
224
225 /*
226 * Adds/Removes a counter to/from the PMU, can be done inside
227 * a transaction, see the ->*_txn() methods.
228 */
229 int (*add) (struct perf_event *event, int flags);
230 void (*del) (struct perf_event *event, int flags);
231
232 /*
233 * Starts/Stops a counter present on the PMU. The PMI handler
234 * should stop the counter when perf_event_overflow() returns
235 * !0. ->start() will be used to continue.
236 */
237 void (*start) (struct perf_event *event, int flags);
238 void (*stop) (struct perf_event *event, int flags);
239
240 /*
241 * Updates the counter value of the event.
242 */
243 void (*read) (struct perf_event *event);
244
245 /*
246 * Group events scheduling is treated as a transaction, add
247 * group events as a whole and perform one schedulability test.
248 * If the test fails, roll back the whole group
249 *
250 * Start the transaction, after this ->add() doesn't need to
251 * do schedulability tests.
252 */
253 void (*start_txn) (struct pmu *pmu); /* optional */
254 /*
255 * If ->start_txn() disabled the ->add() schedulability test
256 * then ->commit_txn() is required to perform one. On success
257 * the transaction is closed. On error the transaction is kept
258 * open until ->cancel_txn() is called.
259 */
260 int (*commit_txn) (struct pmu *pmu); /* optional */
261 /*
262 * Will cancel the transaction, assumes ->del() is called
263 * for each successful ->add() during the transaction.
264 */
265 void (*cancel_txn) (struct pmu *pmu); /* optional */
266
267 /*
268 * Will return the value for perf_event_mmap_page::index for this event,
269 * if no implementation is provided it will default to: event->hw.idx + 1.
270 */
271 int (*event_idx) (struct perf_event *event); /*optional */
272
273 /*
274 * context-switches callback
275 */
276 void (*sched_task) (struct perf_event_context *ctx,
277 bool sched_in);
278 /*
279 * PMU specific data size
280 */
281 size_t task_ctx_size;
282
283
284 /*
285 * Return the count value for a counter.
286 */
287 u64 (*count) (struct perf_event *event); /*optional*/
288 };
289
290 /**
291 * enum perf_event_active_state - the states of a event
292 */
293 enum perf_event_active_state {
294 PERF_EVENT_STATE_EXIT = -3,
295 PERF_EVENT_STATE_ERROR = -2,
296 PERF_EVENT_STATE_OFF = -1,
297 PERF_EVENT_STATE_INACTIVE = 0,
298 PERF_EVENT_STATE_ACTIVE = 1,
299 };
300
301 struct file;
302 struct perf_sample_data;
303
304 typedef void (*perf_overflow_handler_t)(struct perf_event *,
305 struct perf_sample_data *,
306 struct pt_regs *regs);
307
308 enum perf_group_flag {
309 PERF_GROUP_SOFTWARE = 0x1,
310 };
311
312 #define SWEVENT_HLIST_BITS 8
313 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
314
315 struct swevent_hlist {
316 struct hlist_head heads[SWEVENT_HLIST_SIZE];
317 struct rcu_head rcu_head;
318 };
319
320 #define PERF_ATTACH_CONTEXT 0x01
321 #define PERF_ATTACH_GROUP 0x02
322 #define PERF_ATTACH_TASK 0x04
323 #define PERF_ATTACH_TASK_DATA 0x08
324
325 struct perf_cgroup;
326 struct ring_buffer;
327
328 /**
329 * struct perf_event - performance event kernel representation:
330 */
331 struct perf_event {
332 #ifdef CONFIG_PERF_EVENTS
333 /*
334 * entry onto perf_event_context::event_list;
335 * modifications require ctx->lock
336 * RCU safe iterations.
337 */
338 struct list_head event_entry;
339
340 /*
341 * XXX: group_entry and sibling_list should be mutually exclusive;
342 * either you're a sibling on a group, or you're the group leader.
343 * Rework the code to always use the same list element.
344 *
345 * Locked for modification by both ctx->mutex and ctx->lock; holding
346 * either sufficies for read.
347 */
348 struct list_head group_entry;
349 struct list_head sibling_list;
350
351 /*
352 * We need storage to track the entries in perf_pmu_migrate_context; we
353 * cannot use the event_entry because of RCU and we want to keep the
354 * group in tact which avoids us using the other two entries.
355 */
356 struct list_head migrate_entry;
357
358 struct hlist_node hlist_entry;
359 struct list_head active_entry;
360 int nr_siblings;
361 int group_flags;
362 struct perf_event *group_leader;
363 struct pmu *pmu;
364
365 enum perf_event_active_state state;
366 unsigned int attach_state;
367 local64_t count;
368 atomic64_t child_count;
369
370 /*
371 * These are the total time in nanoseconds that the event
372 * has been enabled (i.e. eligible to run, and the task has
373 * been scheduled in, if this is a per-task event)
374 * and running (scheduled onto the CPU), respectively.
375 *
376 * They are computed from tstamp_enabled, tstamp_running and
377 * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
378 */
379 u64 total_time_enabled;
380 u64 total_time_running;
381
382 /*
383 * These are timestamps used for computing total_time_enabled
384 * and total_time_running when the event is in INACTIVE or
385 * ACTIVE state, measured in nanoseconds from an arbitrary point
386 * in time.
387 * tstamp_enabled: the notional time when the event was enabled
388 * tstamp_running: the notional time when the event was scheduled on
389 * tstamp_stopped: in INACTIVE state, the notional time when the
390 * event was scheduled off.
391 */
392 u64 tstamp_enabled;
393 u64 tstamp_running;
394 u64 tstamp_stopped;
395
396 /*
397 * timestamp shadows the actual context timing but it can
398 * be safely used in NMI interrupt context. It reflects the
399 * context time as it was when the event was last scheduled in.
400 *
401 * ctx_time already accounts for ctx->timestamp. Therefore to
402 * compute ctx_time for a sample, simply add perf_clock().
403 */
404 u64 shadow_ctx_time;
405
406 struct perf_event_attr attr;
407 u16 header_size;
408 u16 id_header_size;
409 u16 read_size;
410 struct hw_perf_event hw;
411
412 struct perf_event_context *ctx;
413 atomic_long_t refcount;
414
415 /*
416 * These accumulate total time (in nanoseconds) that children
417 * events have been enabled and running, respectively.
418 */
419 atomic64_t child_total_time_enabled;
420 atomic64_t child_total_time_running;
421
422 /*
423 * Protect attach/detach and child_list:
424 */
425 struct mutex child_mutex;
426 struct list_head child_list;
427 struct perf_event *parent;
428
429 int oncpu;
430 int cpu;
431
432 struct list_head owner_entry;
433 struct task_struct *owner;
434
435 /* mmap bits */
436 struct mutex mmap_mutex;
437 atomic_t mmap_count;
438
439 struct ring_buffer *rb;
440 struct list_head rb_entry;
441 unsigned long rcu_batches;
442 int rcu_pending;
443
444 /* poll related */
445 wait_queue_head_t waitq;
446 struct fasync_struct *fasync;
447
448 /* delayed work for NMIs and such */
449 int pending_wakeup;
450 int pending_kill;
451 int pending_disable;
452 struct irq_work pending;
453
454 atomic_t event_limit;
455
456 void (*destroy)(struct perf_event *);
457 struct rcu_head rcu_head;
458
459 struct pid_namespace *ns;
460 u64 id;
461
462 perf_overflow_handler_t overflow_handler;
463 void *overflow_handler_context;
464
465 #ifdef CONFIG_EVENT_TRACING
466 struct ftrace_event_call *tp_event;
467 struct event_filter *filter;
468 #ifdef CONFIG_FUNCTION_TRACER
469 struct ftrace_ops ftrace_ops;
470 #endif
471 #endif
472
473 #ifdef CONFIG_CGROUP_PERF
474 struct perf_cgroup *cgrp; /* cgroup event is attach to */
475 int cgrp_defer_enabled;
476 #endif
477
478 #endif /* CONFIG_PERF_EVENTS */
479 };
480
481 /**
482 * struct perf_event_context - event context structure
483 *
484 * Used as a container for task events and CPU events as well:
485 */
486 struct perf_event_context {
487 struct pmu *pmu;
488 /*
489 * Protect the states of the events in the list,
490 * nr_active, and the list:
491 */
492 raw_spinlock_t lock;
493 /*
494 * Protect the list of events. Locking either mutex or lock
495 * is sufficient to ensure the list doesn't change; to change
496 * the list you need to lock both the mutex and the spinlock.
497 */
498 struct mutex mutex;
499
500 struct list_head active_ctx_list;
501 struct list_head pinned_groups;
502 struct list_head flexible_groups;
503 struct list_head event_list;
504 int nr_events;
505 int nr_active;
506 int is_active;
507 int nr_stat;
508 int nr_freq;
509 int rotate_disable;
510 atomic_t refcount;
511 struct task_struct *task;
512
513 /*
514 * Context clock, runs when context enabled.
515 */
516 u64 time;
517 u64 timestamp;
518
519 /*
520 * These fields let us detect when two contexts have both
521 * been cloned (inherited) from a common ancestor.
522 */
523 struct perf_event_context *parent_ctx;
524 u64 parent_gen;
525 u64 generation;
526 int pin_count;
527 int nr_cgroups; /* cgroup evts */
528 void *task_ctx_data; /* pmu specific data */
529 struct rcu_head rcu_head;
530
531 struct delayed_work orphans_remove;
532 bool orphans_remove_sched;
533 };
534
535 /*
536 * Number of contexts where an event can trigger:
537 * task, softirq, hardirq, nmi.
538 */
539 #define PERF_NR_CONTEXTS 4
540
541 /**
542 * struct perf_event_cpu_context - per cpu event context structure
543 */
544 struct perf_cpu_context {
545 struct perf_event_context ctx;
546 struct perf_event_context *task_ctx;
547 int active_oncpu;
548 int exclusive;
549 struct hrtimer hrtimer;
550 ktime_t hrtimer_interval;
551 struct pmu *unique_pmu;
552 struct perf_cgroup *cgrp;
553 };
554
555 struct perf_output_handle {
556 struct perf_event *event;
557 struct ring_buffer *rb;
558 unsigned long wakeup;
559 unsigned long size;
560 void *addr;
561 int page;
562 };
563
564 #ifdef CONFIG_CGROUP_PERF
565
566 /*
567 * perf_cgroup_info keeps track of time_enabled for a cgroup.
568 * This is a per-cpu dynamically allocated data structure.
569 */
570 struct perf_cgroup_info {
571 u64 time;
572 u64 timestamp;
573 };
574
575 struct perf_cgroup {
576 struct cgroup_subsys_state css;
577 struct perf_cgroup_info __percpu *info;
578 };
579
580 /*
581 * Must ensure cgroup is pinned (css_get) before calling
582 * this function. In other words, we cannot call this function
583 * if there is no cgroup event for the current CPU context.
584 */
585 static inline struct perf_cgroup *
586 perf_cgroup_from_task(struct task_struct *task)
587 {
588 return container_of(task_css(task, perf_event_cgrp_id),
589 struct perf_cgroup, css);
590 }
591 #endif /* CONFIG_CGROUP_PERF */
592
593 #ifdef CONFIG_PERF_EVENTS
594
595 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
596 extern void perf_pmu_unregister(struct pmu *pmu);
597
598 extern int perf_num_counters(void);
599 extern const char *perf_pmu_name(void);
600 extern void __perf_event_task_sched_in(struct task_struct *prev,
601 struct task_struct *task);
602 extern void __perf_event_task_sched_out(struct task_struct *prev,
603 struct task_struct *next);
604 extern int perf_event_init_task(struct task_struct *child);
605 extern void perf_event_exit_task(struct task_struct *child);
606 extern void perf_event_free_task(struct task_struct *task);
607 extern void perf_event_delayed_put(struct task_struct *task);
608 extern void perf_event_print_debug(void);
609 extern void perf_pmu_disable(struct pmu *pmu);
610 extern void perf_pmu_enable(struct pmu *pmu);
611 extern void perf_sched_cb_dec(struct pmu *pmu);
612 extern void perf_sched_cb_inc(struct pmu *pmu);
613 extern int perf_event_task_disable(void);
614 extern int perf_event_task_enable(void);
615 extern int perf_event_refresh(struct perf_event *event, int refresh);
616 extern void perf_event_update_userpage(struct perf_event *event);
617 extern int perf_event_release_kernel(struct perf_event *event);
618 extern struct perf_event *
619 perf_event_create_kernel_counter(struct perf_event_attr *attr,
620 int cpu,
621 struct task_struct *task,
622 perf_overflow_handler_t callback,
623 void *context);
624 extern void perf_pmu_migrate_context(struct pmu *pmu,
625 int src_cpu, int dst_cpu);
626 extern u64 perf_event_read_value(struct perf_event *event,
627 u64 *enabled, u64 *running);
628
629
630 struct perf_sample_data {
631 /*
632 * Fields set by perf_sample_data_init(), group so as to
633 * minimize the cachelines touched.
634 */
635 u64 addr;
636 struct perf_raw_record *raw;
637 struct perf_branch_stack *br_stack;
638 u64 period;
639 u64 weight;
640 u64 txn;
641 union perf_mem_data_src data_src;
642
643 /*
644 * The other fields, optionally {set,used} by
645 * perf_{prepare,output}_sample().
646 */
647 u64 type;
648 u64 ip;
649 struct {
650 u32 pid;
651 u32 tid;
652 } tid_entry;
653 u64 time;
654 u64 id;
655 u64 stream_id;
656 struct {
657 u32 cpu;
658 u32 reserved;
659 } cpu_entry;
660 struct perf_callchain_entry *callchain;
661
662 /*
663 * regs_user may point to task_pt_regs or to regs_user_copy, depending
664 * on arch details.
665 */
666 struct perf_regs regs_user;
667 struct pt_regs regs_user_copy;
668
669 struct perf_regs regs_intr;
670 u64 stack_user_size;
671 } ____cacheline_aligned;
672
673 /* default value for data source */
674 #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
675 PERF_MEM_S(LVL, NA) |\
676 PERF_MEM_S(SNOOP, NA) |\
677 PERF_MEM_S(LOCK, NA) |\
678 PERF_MEM_S(TLB, NA))
679
680 static inline void perf_sample_data_init(struct perf_sample_data *data,
681 u64 addr, u64 period)
682 {
683 /* remaining struct members initialized in perf_prepare_sample() */
684 data->addr = addr;
685 data->raw = NULL;
686 data->br_stack = NULL;
687 data->period = period;
688 data->weight = 0;
689 data->data_src.val = PERF_MEM_NA;
690 data->txn = 0;
691 }
692
693 extern void perf_output_sample(struct perf_output_handle *handle,
694 struct perf_event_header *header,
695 struct perf_sample_data *data,
696 struct perf_event *event);
697 extern void perf_prepare_sample(struct perf_event_header *header,
698 struct perf_sample_data *data,
699 struct perf_event *event,
700 struct pt_regs *regs);
701
702 extern int perf_event_overflow(struct perf_event *event,
703 struct perf_sample_data *data,
704 struct pt_regs *regs);
705
706 static inline bool is_sampling_event(struct perf_event *event)
707 {
708 return event->attr.sample_period != 0;
709 }
710
711 /*
712 * Return 1 for a software event, 0 for a hardware event
713 */
714 static inline int is_software_event(struct perf_event *event)
715 {
716 return event->pmu->task_ctx_nr == perf_sw_context;
717 }
718
719 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
720
721 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
722 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
723
724 #ifndef perf_arch_fetch_caller_regs
725 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
726 #endif
727
728 /*
729 * Take a snapshot of the regs. Skip ip and frame pointer to
730 * the nth caller. We only need a few of the regs:
731 * - ip for PERF_SAMPLE_IP
732 * - cs for user_mode() tests
733 * - bp for callchains
734 * - eflags, for future purposes, just in case
735 */
736 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
737 {
738 memset(regs, 0, sizeof(*regs));
739
740 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
741 }
742
743 static __always_inline void
744 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
745 {
746 if (static_key_false(&perf_swevent_enabled[event_id]))
747 __perf_sw_event(event_id, nr, regs, addr);
748 }
749
750 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
751
752 /*
753 * 'Special' version for the scheduler, it hard assumes no recursion,
754 * which is guaranteed by us not actually scheduling inside other swevents
755 * because those disable preemption.
756 */
757 static __always_inline void
758 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
759 {
760 if (static_key_false(&perf_swevent_enabled[event_id])) {
761 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
762
763 perf_fetch_caller_regs(regs);
764 ___perf_sw_event(event_id, nr, regs, addr);
765 }
766 }
767
768 extern struct static_key_deferred perf_sched_events;
769
770 static inline void perf_event_task_sched_in(struct task_struct *prev,
771 struct task_struct *task)
772 {
773 if (static_key_false(&perf_sched_events.key))
774 __perf_event_task_sched_in(prev, task);
775 }
776
777 static inline void perf_event_task_sched_out(struct task_struct *prev,
778 struct task_struct *next)
779 {
780 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
781
782 if (static_key_false(&perf_sched_events.key))
783 __perf_event_task_sched_out(prev, next);
784 }
785
786 static inline u64 __perf_event_count(struct perf_event *event)
787 {
788 return local64_read(&event->count) + atomic64_read(&event->child_count);
789 }
790
791 extern void perf_event_mmap(struct vm_area_struct *vma);
792 extern struct perf_guest_info_callbacks *perf_guest_cbs;
793 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
794 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
795
796 extern void perf_event_exec(void);
797 extern void perf_event_comm(struct task_struct *tsk, bool exec);
798 extern void perf_event_fork(struct task_struct *tsk);
799
800 /* Callchains */
801 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
802
803 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
804 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
805
806 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
807 {
808 if (entry->nr < PERF_MAX_STACK_DEPTH)
809 entry->ip[entry->nr++] = ip;
810 }
811
812 extern int sysctl_perf_event_paranoid;
813 extern int sysctl_perf_event_mlock;
814 extern int sysctl_perf_event_sample_rate;
815 extern int sysctl_perf_cpu_time_max_percent;
816
817 extern void perf_sample_event_took(u64 sample_len_ns);
818
819 extern int perf_proc_update_handler(struct ctl_table *table, int write,
820 void __user *buffer, size_t *lenp,
821 loff_t *ppos);
822 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
823 void __user *buffer, size_t *lenp,
824 loff_t *ppos);
825
826
827 static inline bool perf_paranoid_tracepoint_raw(void)
828 {
829 return sysctl_perf_event_paranoid > -1;
830 }
831
832 static inline bool perf_paranoid_cpu(void)
833 {
834 return sysctl_perf_event_paranoid > 0;
835 }
836
837 static inline bool perf_paranoid_kernel(void)
838 {
839 return sysctl_perf_event_paranoid > 1;
840 }
841
842 extern void perf_event_init(void);
843 extern void perf_tp_event(u64 addr, u64 count, void *record,
844 int entry_size, struct pt_regs *regs,
845 struct hlist_head *head, int rctx,
846 struct task_struct *task);
847 extern void perf_bp_event(struct perf_event *event, void *data);
848
849 #ifndef perf_misc_flags
850 # define perf_misc_flags(regs) \
851 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
852 # define perf_instruction_pointer(regs) instruction_pointer(regs)
853 #endif
854
855 static inline bool has_branch_stack(struct perf_event *event)
856 {
857 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
858 }
859
860 static inline bool needs_branch_stack(struct perf_event *event)
861 {
862 return event->attr.branch_sample_type != 0;
863 }
864
865 extern int perf_output_begin(struct perf_output_handle *handle,
866 struct perf_event *event, unsigned int size);
867 extern void perf_output_end(struct perf_output_handle *handle);
868 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
869 const void *buf, unsigned int len);
870 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
871 unsigned int len);
872 extern int perf_swevent_get_recursion_context(void);
873 extern void perf_swevent_put_recursion_context(int rctx);
874 extern u64 perf_swevent_set_period(struct perf_event *event);
875 extern void perf_event_enable(struct perf_event *event);
876 extern void perf_event_disable(struct perf_event *event);
877 extern int __perf_event_disable(void *info);
878 extern void perf_event_task_tick(void);
879 #else /* !CONFIG_PERF_EVENTS: */
880 static inline void
881 perf_event_task_sched_in(struct task_struct *prev,
882 struct task_struct *task) { }
883 static inline void
884 perf_event_task_sched_out(struct task_struct *prev,
885 struct task_struct *next) { }
886 static inline int perf_event_init_task(struct task_struct *child) { return 0; }
887 static inline void perf_event_exit_task(struct task_struct *child) { }
888 static inline void perf_event_free_task(struct task_struct *task) { }
889 static inline void perf_event_delayed_put(struct task_struct *task) { }
890 static inline void perf_event_print_debug(void) { }
891 static inline int perf_event_task_disable(void) { return -EINVAL; }
892 static inline int perf_event_task_enable(void) { return -EINVAL; }
893 static inline int perf_event_refresh(struct perf_event *event, int refresh)
894 {
895 return -EINVAL;
896 }
897
898 static inline void
899 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
900 static inline void
901 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
902 static inline void
903 perf_bp_event(struct perf_event *event, void *data) { }
904
905 static inline int perf_register_guest_info_callbacks
906 (struct perf_guest_info_callbacks *callbacks) { return 0; }
907 static inline int perf_unregister_guest_info_callbacks
908 (struct perf_guest_info_callbacks *callbacks) { return 0; }
909
910 static inline void perf_event_mmap(struct vm_area_struct *vma) { }
911 static inline void perf_event_exec(void) { }
912 static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
913 static inline void perf_event_fork(struct task_struct *tsk) { }
914 static inline void perf_event_init(void) { }
915 static inline int perf_swevent_get_recursion_context(void) { return -1; }
916 static inline void perf_swevent_put_recursion_context(int rctx) { }
917 static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
918 static inline void perf_event_enable(struct perf_event *event) { }
919 static inline void perf_event_disable(struct perf_event *event) { }
920 static inline int __perf_event_disable(void *info) { return -1; }
921 static inline void perf_event_task_tick(void) { }
922 #endif
923
924 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
925 extern bool perf_event_can_stop_tick(void);
926 #else
927 static inline bool perf_event_can_stop_tick(void) { return true; }
928 #endif
929
930 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
931 extern void perf_restore_debug_store(void);
932 #else
933 static inline void perf_restore_debug_store(void) { }
934 #endif
935
936 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
937
938 /*
939 * This has to have a higher priority than migration_notifier in sched/core.c.
940 */
941 #define perf_cpu_notifier(fn) \
942 do { \
943 static struct notifier_block fn##_nb = \
944 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
945 unsigned long cpu = smp_processor_id(); \
946 unsigned long flags; \
947 \
948 cpu_notifier_register_begin(); \
949 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \
950 (void *)(unsigned long)cpu); \
951 local_irq_save(flags); \
952 fn(&fn##_nb, (unsigned long)CPU_STARTING, \
953 (void *)(unsigned long)cpu); \
954 local_irq_restore(flags); \
955 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \
956 (void *)(unsigned long)cpu); \
957 __register_cpu_notifier(&fn##_nb); \
958 cpu_notifier_register_done(); \
959 } while (0)
960
961 /*
962 * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
963 * callback for already online CPUs.
964 */
965 #define __perf_cpu_notifier(fn) \
966 do { \
967 static struct notifier_block fn##_nb = \
968 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
969 \
970 __register_cpu_notifier(&fn##_nb); \
971 } while (0)
972
973 struct perf_pmu_events_attr {
974 struct device_attribute attr;
975 u64 id;
976 const char *event_str;
977 };
978
979 #define PMU_EVENT_ATTR(_name, _var, _id, _show) \
980 static struct perf_pmu_events_attr _var = { \
981 .attr = __ATTR(_name, 0444, _show, NULL), \
982 .id = _id, \
983 };
984
985 #define PMU_FORMAT_ATTR(_name, _format) \
986 static ssize_t \
987 _name##_show(struct device *dev, \
988 struct device_attribute *attr, \
989 char *page) \
990 { \
991 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
992 return sprintf(page, _format "\n"); \
993 } \
994 \
995 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
996
997 #endif /* _LINUX_PERF_EVENT_H */
Linux kernel - Deliverables
This page took 0.050777 seconds and 4 git commands to generate.