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#include <uapi/linux/bpf_perf_event.h>
19
20/*
21 * Kernel-internal data types and definitions:
22 */
23
24#ifdef CONFIG_PERF_EVENTS
25# include <asm/perf_event.h>
26# include <asm/local64.h>
27#endif
28
29struct perf_guest_info_callbacks {
30 int (*is_in_guest)(void);
31 int (*is_user_mode)(void);
32 unsigned long (*get_guest_ip)(void);
33};
34
35#ifdef CONFIG_HAVE_HW_BREAKPOINT
36#include <asm/hw_breakpoint.h>
37#endif
38
39#include <linux/list.h>
40#include <linux/mutex.h>
41#include <linux/rculist.h>
42#include <linux/rcupdate.h>
43#include <linux/spinlock.h>
44#include <linux/hrtimer.h>
45#include <linux/fs.h>
46#include <linux/pid_namespace.h>
47#include <linux/workqueue.h>
48#include <linux/ftrace.h>
49#include <linux/cpu.h>
50#include <linux/irq_work.h>
51#include <linux/static_key.h>
52#include <linux/jump_label_ratelimit.h>
53#include <linux/atomic.h>
54#include <linux/sysfs.h>
55#include <linux/perf_regs.h>
56#include <linux/workqueue.h>
57#include <linux/cgroup.h>
58#include <asm/local.h>
59
60struct perf_callchain_entry {
61 __u64 nr;
62 __u64 ip[0]; /* /proc/sys/kernel/perf_event_max_stack */
63};
64
65struct perf_callchain_entry_ctx {
66 struct perf_callchain_entry *entry;
67 u32 max_stack;
68 u32 nr;
69 short contexts;
70 bool contexts_maxed;
71};
72
73typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
74 unsigned long off, unsigned long len);
75
76struct perf_raw_frag {
77 union {
78 struct perf_raw_frag *next;
79 unsigned long pad;
80 };
81 perf_copy_f copy;
82 void *data;
83 u32 size;
84} __packed;
85
86struct perf_raw_record {
87 struct perf_raw_frag frag;
88 u32 size;
89};
90
91/*
92 * branch stack layout:
93 * nr: number of taken branches stored in entries[]
94 *
95 * Note that nr can vary from sample to sample
96 * branches (to, from) are stored from most recent
97 * to least recent, i.e., entries[0] contains the most
98 * recent branch.
99 */
100struct perf_branch_stack {
101 __u64 nr;
102 struct perf_branch_entry entries[0];
103};
104
105struct task_struct;
106
107/*
108 * extra PMU register associated with an event
109 */
110struct hw_perf_event_extra {
111 u64 config; /* register value */
112 unsigned int reg; /* register address or index */
113 int alloc; /* extra register already allocated */
114 int idx; /* index in shared_regs->regs[] */
115};
116
117/**
118 * struct hw_perf_event - performance event hardware details:
119 */
120struct hw_perf_event {
121#ifdef CONFIG_PERF_EVENTS
122 union {
123 struct { /* hardware */
124 u64 config;
125 u64 last_tag;
126 unsigned long config_base;
127 unsigned long event_base;
128 int event_base_rdpmc;
129 int idx;
130 int last_cpu;
131 int flags;
132
133 struct hw_perf_event_extra extra_reg;
134 struct hw_perf_event_extra branch_reg;
135 };
136 struct { /* software */
137 struct hrtimer hrtimer;
138 };
139 struct { /* tracepoint */
140 /* for tp_event->class */
141 struct list_head tp_list;
142 };
143 struct { /* amd_power */
144 u64 pwr_acc;
145 u64 ptsc;
146 };
147#ifdef CONFIG_HAVE_HW_BREAKPOINT
148 struct { /* breakpoint */
149 /*
150 * Crufty hack to avoid the chicken and egg
151 * problem hw_breakpoint has with context
152 * creation and event initalization.
153 */
154 struct arch_hw_breakpoint info;
155 struct list_head bp_list;
156 };
157#endif
158 struct { /* amd_iommu */
159 u8 iommu_bank;
160 u8 iommu_cntr;
161 u16 padding;
162 u64 conf;
163 u64 conf1;
164 };
165 };
166 /*
167 * If the event is a per task event, this will point to the task in
168 * question. See the comment in perf_event_alloc().
169 */
170 struct task_struct *target;
171
172 /*
173 * PMU would store hardware filter configuration
174 * here.
175 */
176 void *addr_filters;
177
178 /* Last sync'ed generation of filters */
179 unsigned long addr_filters_gen;
180
181/*
182 * hw_perf_event::state flags; used to track the PERF_EF_* state.
183 */
184#define PERF_HES_STOPPED 0x01 /* the counter is stopped */
185#define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
186#define PERF_HES_ARCH 0x04
187
188 int state;
189
190 /*
191 * The last observed hardware counter value, updated with a
192 * local64_cmpxchg() such that pmu::read() can be called nested.
193 */
194 local64_t prev_count;
195
196 /*
197 * The period to start the next sample with.
198 */
199 u64 sample_period;
200
201 /*
202 * The period we started this sample with.
203 */
204 u64 last_period;
205
206 /*
207 * However much is left of the current period; note that this is
208 * a full 64bit value and allows for generation of periods longer
209 * than hardware might allow.
210 */
211 local64_t period_left;
212
213 /*
214 * State for throttling the event, see __perf_event_overflow() and
215 * perf_adjust_freq_unthr_context().
216 */
217 u64 interrupts_seq;
218 u64 interrupts;
219
220 /*
221 * State for freq target events, see __perf_event_overflow() and
222 * perf_adjust_freq_unthr_context().
223 */
224 u64 freq_time_stamp;
225 u64 freq_count_stamp;
226#endif
227};
228
229struct perf_event;
230
231/*
232 * Common implementation detail of pmu::{start,commit,cancel}_txn
233 */
234#define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
235#define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
236
237/**
238 * pmu::capabilities flags
239 */
240#define PERF_PMU_CAP_NO_INTERRUPT 0x01
241#define PERF_PMU_CAP_NO_NMI 0x02
242#define PERF_PMU_CAP_AUX_NO_SG 0x04
243#define PERF_PMU_CAP_AUX_SW_DOUBLEBUF 0x08
244#define PERF_PMU_CAP_EXCLUSIVE 0x10
245#define PERF_PMU_CAP_ITRACE 0x20
246#define PERF_PMU_CAP_HETEROGENEOUS_CPUS 0x40
247
248/**
249 * struct pmu - generic performance monitoring unit
250 */
251struct pmu {
252 struct list_head entry;
253
254 struct module *module;
255 struct device *dev;
256 const struct attribute_group **attr_groups;
257 const char *name;
258 int type;
259
260 /*
261 * various common per-pmu feature flags
262 */
263 int capabilities;
264
265 int * __percpu pmu_disable_count;
266 struct perf_cpu_context * __percpu pmu_cpu_context;
267 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
268 int task_ctx_nr;
269 int hrtimer_interval_ms;
270
271 /* number of address filters this PMU can do */
272 unsigned int nr_addr_filters;
273
274 /*
275 * Fully disable/enable this PMU, can be used to protect from the PMI
276 * as well as for lazy/batch writing of the MSRs.
277 */
278 void (*pmu_enable) (struct pmu *pmu); /* optional */
279 void (*pmu_disable) (struct pmu *pmu); /* optional */
280
281 /*
282 * Try and initialize the event for this PMU.
283 *
284 * Returns:
285 * -ENOENT -- @event is not for this PMU
286 *
287 * -ENODEV -- @event is for this PMU but PMU not present
288 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
289 * -EINVAL -- @event is for this PMU but @event is not valid
290 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
291 * -EACCESS -- @event is for this PMU, @event is valid, but no privilidges
292 *
293 * 0 -- @event is for this PMU and valid
294 *
295 * Other error return values are allowed.
296 */
297 int (*event_init) (struct perf_event *event);
298
299 /*
300 * Notification that the event was mapped or unmapped. Called
301 * in the context of the mapping task.
302 */
303 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
304 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
305
306 /*
307 * Flags for ->add()/->del()/ ->start()/->stop(). There are
308 * matching hw_perf_event::state flags.
309 */
310#define PERF_EF_START 0x01 /* start the counter when adding */
311#define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
312#define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
313
314 /*
315 * Adds/Removes a counter to/from the PMU, can be done inside a
316 * transaction, see the ->*_txn() methods.
317 *
318 * The add/del callbacks will reserve all hardware resources required
319 * to service the event, this includes any counter constraint
320 * scheduling etc.
321 *
322 * Called with IRQs disabled and the PMU disabled on the CPU the event
323 * is on.
324 *
325 * ->add() called without PERF_EF_START should result in the same state
326 * as ->add() followed by ->stop().
327 *
328 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
329 * ->stop() that must deal with already being stopped without
330 * PERF_EF_UPDATE.
331 */
332 int (*add) (struct perf_event *event, int flags);
333 void (*del) (struct perf_event *event, int flags);
334
335 /*
336 * Starts/Stops a counter present on the PMU.
337 *
338 * The PMI handler should stop the counter when perf_event_overflow()
339 * returns !0. ->start() will be used to continue.
340 *
341 * Also used to change the sample period.
342 *
343 * Called with IRQs disabled and the PMU disabled on the CPU the event
344 * is on -- will be called from NMI context with the PMU generates
345 * NMIs.
346 *
347 * ->stop() with PERF_EF_UPDATE will read the counter and update
348 * period/count values like ->read() would.
349 *
350 * ->start() with PERF_EF_RELOAD will reprogram the the counter
351 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
352 */
353 void (*start) (struct perf_event *event, int flags);
354 void (*stop) (struct perf_event *event, int flags);
355
356 /*
357 * Updates the counter value of the event.
358 *
359 * For sampling capable PMUs this will also update the software period
360 * hw_perf_event::period_left field.
361 */
362 void (*read) (struct perf_event *event);
363
364 /*
365 * Group events scheduling is treated as a transaction, add
366 * group events as a whole and perform one schedulability test.
367 * If the test fails, roll back the whole group
368 *
369 * Start the transaction, after this ->add() doesn't need to
370 * do schedulability tests.
371 *
372 * Optional.
373 */
374 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
375 /*
376 * If ->start_txn() disabled the ->add() schedulability test
377 * then ->commit_txn() is required to perform one. On success
378 * the transaction is closed. On error the transaction is kept
379 * open until ->cancel_txn() is called.
380 *
381 * Optional.
382 */
383 int (*commit_txn) (struct pmu *pmu);
384 /*
385 * Will cancel the transaction, assumes ->del() is called
386 * for each successful ->add() during the transaction.
387 *
388 * Optional.
389 */
390 void (*cancel_txn) (struct pmu *pmu);
391
392 /*
393 * Will return the value for perf_event_mmap_page::index for this event,
394 * if no implementation is provided it will default to: event->hw.idx + 1.
395 */
396 int (*event_idx) (struct perf_event *event); /*optional */
397
398 /*
399 * context-switches callback
400 */
401 void (*sched_task) (struct perf_event_context *ctx,
402 bool sched_in);
403 /*
404 * PMU specific data size
405 */
406 size_t task_ctx_size;
407
408
409 /*
410 * Set up pmu-private data structures for an AUX area
411 */
412 void *(*setup_aux) (int cpu, void **pages,
413 int nr_pages, bool overwrite);
414 /* optional */
415
416 /*
417 * Free pmu-private AUX data structures
418 */
419 void (*free_aux) (void *aux); /* optional */
420
421 /*
422 * Validate address range filters: make sure the HW supports the
423 * requested configuration and number of filters; return 0 if the
424 * supplied filters are valid, -errno otherwise.
425 *
426 * Runs in the context of the ioctl()ing process and is not serialized
427 * with the rest of the PMU callbacks.
428 */
429 int (*addr_filters_validate) (struct list_head *filters);
430 /* optional */
431
432 /*
433 * Synchronize address range filter configuration:
434 * translate hw-agnostic filters into hardware configuration in
435 * event::hw::addr_filters.
436 *
437 * Runs as a part of filter sync sequence that is done in ->start()
438 * callback by calling perf_event_addr_filters_sync().
439 *
440 * May (and should) traverse event::addr_filters::list, for which its
441 * caller provides necessary serialization.
442 */
443 void (*addr_filters_sync) (struct perf_event *event);
444 /* optional */
445
446 /*
447 * Filter events for PMU-specific reasons.
448 */
449 int (*filter_match) (struct perf_event *event); /* optional */
450};
451
452enum perf_addr_filter_action_t {
453 PERF_ADDR_FILTER_ACTION_STOP = 0,
454 PERF_ADDR_FILTER_ACTION_START,
455 PERF_ADDR_FILTER_ACTION_FILTER,
456};
457
458/**
459 * struct perf_addr_filter - address range filter definition
460 * @entry: event's filter list linkage
461 * @inode: object file's inode for file-based filters
462 * @offset: filter range offset
463 * @size: filter range size (size==0 means single address trigger)
464 * @action: filter/start/stop
465 *
466 * This is a hardware-agnostic filter configuration as specified by the user.
467 */
468struct perf_addr_filter {
469 struct list_head entry;
470 struct path path;
471 unsigned long offset;
472 unsigned long size;
473 enum perf_addr_filter_action_t action;
474};
475
476/**
477 * struct perf_addr_filters_head - container for address range filters
478 * @list: list of filters for this event
479 * @lock: spinlock that serializes accesses to the @list and event's
480 * (and its children's) filter generations.
481 * @nr_file_filters: number of file-based filters
482 *
483 * A child event will use parent's @list (and therefore @lock), so they are
484 * bundled together; see perf_event_addr_filters().
485 */
486struct perf_addr_filters_head {
487 struct list_head list;
488 raw_spinlock_t lock;
489 unsigned int nr_file_filters;
490};
491
492/**
493 * enum perf_event_state - the states of an event:
494 */
495enum perf_event_state {
496 PERF_EVENT_STATE_DEAD = -4,
497 PERF_EVENT_STATE_EXIT = -3,
498 PERF_EVENT_STATE_ERROR = -2,
499 PERF_EVENT_STATE_OFF = -1,
500 PERF_EVENT_STATE_INACTIVE = 0,
501 PERF_EVENT_STATE_ACTIVE = 1,
502};
503
504struct file;
505struct perf_sample_data;
506
507typedef void (*perf_overflow_handler_t)(struct perf_event *,
508 struct perf_sample_data *,
509 struct pt_regs *regs);
510
511/*
512 * Event capabilities. For event_caps and groups caps.
513 *
514 * PERF_EV_CAP_SOFTWARE: Is a software event.
515 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
516 * from any CPU in the package where it is active.
517 */
518#define PERF_EV_CAP_SOFTWARE BIT(0)
519#define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
520
521#define SWEVENT_HLIST_BITS 8
522#define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
523
524struct swevent_hlist {
525 struct hlist_head heads[SWEVENT_HLIST_SIZE];
526 struct rcu_head rcu_head;
527};
528
529#define PERF_ATTACH_CONTEXT 0x01
530#define PERF_ATTACH_GROUP 0x02
531#define PERF_ATTACH_TASK 0x04
532#define PERF_ATTACH_TASK_DATA 0x08
533#define PERF_ATTACH_ITRACE 0x10
534
535struct perf_cgroup;
536struct ring_buffer;
537
538struct pmu_event_list {
539 raw_spinlock_t lock;
540 struct list_head list;
541};
542
543#define for_each_sibling_event(sibling, event) \
544 if ((event)->group_leader == (event)) \
545 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
546
547/**
548 * struct perf_event - performance event kernel representation:
549 */
550struct perf_event {
551#ifdef CONFIG_PERF_EVENTS
552 /*
553 * entry onto perf_event_context::event_list;
554 * modifications require ctx->lock
555 * RCU safe iterations.
556 */
557 struct list_head event_entry;
558
559 /*
560 * Locked for modification by both ctx->mutex and ctx->lock; holding
561 * either sufficies for read.
562 */
563 struct list_head sibling_list;
564 struct list_head active_list;
565 /*
566 * Node on the pinned or flexible tree located at the event context;
567 */
568 struct rb_node group_node;
569 u64 group_index;
570 /*
571 * We need storage to track the entries in perf_pmu_migrate_context; we
572 * cannot use the event_entry because of RCU and we want to keep the
573 * group in tact which avoids us using the other two entries.
574 */
575 struct list_head migrate_entry;
576
577 struct hlist_node hlist_entry;
578 struct list_head active_entry;
579 int nr_siblings;
580
581 /* Not serialized. Only written during event initialization. */
582 int event_caps;
583 /* The cumulative AND of all event_caps for events in this group. */
584 int group_caps;
585
586 struct perf_event *group_leader;
587 struct pmu *pmu;
588 void *pmu_private;
589
590 enum perf_event_state state;
591 unsigned int attach_state;
592 local64_t count;
593 atomic64_t child_count;
594
595 /*
596 * These are the total time in nanoseconds that the event
597 * has been enabled (i.e. eligible to run, and the task has
598 * been scheduled in, if this is a per-task event)
599 * and running (scheduled onto the CPU), respectively.
600 */
601 u64 total_time_enabled;
602 u64 total_time_running;
603 u64 tstamp;
604
605 /*
606 * timestamp shadows the actual context timing but it can
607 * be safely used in NMI interrupt context. It reflects the
608 * context time as it was when the event was last scheduled in.
609 *
610 * ctx_time already accounts for ctx->timestamp. Therefore to
611 * compute ctx_time for a sample, simply add perf_clock().
612 */
613 u64 shadow_ctx_time;
614
615 struct perf_event_attr attr;
616 u16 header_size;
617 u16 id_header_size;
618 u16 read_size;
619 struct hw_perf_event hw;
620
621 struct perf_event_context *ctx;
622 atomic_long_t refcount;
623
624 /*
625 * These accumulate total time (in nanoseconds) that children
626 * events have been enabled and running, respectively.
627 */
628 atomic64_t child_total_time_enabled;
629 atomic64_t child_total_time_running;
630
631 /*
632 * Protect attach/detach and child_list:
633 */
634 struct mutex child_mutex;
635 struct list_head child_list;
636 struct perf_event *parent;
637
638 int oncpu;
639 int cpu;
640
641 struct list_head owner_entry;
642 struct task_struct *owner;
643
644 /* mmap bits */
645 struct mutex mmap_mutex;
646 atomic_t mmap_count;
647
648 struct ring_buffer *rb;
649 struct list_head rb_entry;
650 unsigned long rcu_batches;
651 int rcu_pending;
652
653 /* poll related */
654 wait_queue_head_t waitq;
655 struct fasync_struct *fasync;
656
657 /* delayed work for NMIs and such */
658 int pending_wakeup;
659 int pending_kill;
660 int pending_disable;
661 struct irq_work pending;
662
663 atomic_t event_limit;
664
665 /* address range filters */
666 struct perf_addr_filters_head addr_filters;
667 /* vma address array for file-based filders */
668 unsigned long *addr_filters_offs;
669 unsigned long addr_filters_gen;
670
671 void (*destroy)(struct perf_event *);
672 struct rcu_head rcu_head;
673
674 struct pid_namespace *ns;
675 u64 id;
676
677 u64 (*clock)(void);
678 perf_overflow_handler_t overflow_handler;
679 void *overflow_handler_context;
680#ifdef CONFIG_BPF_SYSCALL
681 perf_overflow_handler_t orig_overflow_handler;
682 struct bpf_prog *prog;
683#endif
684
685#ifdef CONFIG_EVENT_TRACING
686 struct trace_event_call *tp_event;
687 struct event_filter *filter;
688#ifdef CONFIG_FUNCTION_TRACER
689 struct ftrace_ops ftrace_ops;
690#endif
691#endif
692
693#ifdef CONFIG_CGROUP_PERF
694 struct perf_cgroup *cgrp; /* cgroup event is attach to */
695#endif
696
697 struct list_head sb_list;
698#endif /* CONFIG_PERF_EVENTS */
699};
700
701
702struct perf_event_groups {
703 struct rb_root tree;
704 u64 index;
705};
706
707/**
708 * struct perf_event_context - event context structure
709 *
710 * Used as a container for task events and CPU events as well:
711 */
712struct perf_event_context {
713 struct pmu *pmu;
714 /*
715 * Protect the states of the events in the list,
716 * nr_active, and the list:
717 */
718 raw_spinlock_t lock;
719 /*
720 * Protect the list of events. Locking either mutex or lock
721 * is sufficient to ensure the list doesn't change; to change
722 * the list you need to lock both the mutex and the spinlock.
723 */
724 struct mutex mutex;
725
726 struct list_head active_ctx_list;
727 struct perf_event_groups pinned_groups;
728 struct perf_event_groups flexible_groups;
729 struct list_head event_list;
730
731 struct list_head pinned_active;
732 struct list_head flexible_active;
733
734 int nr_events;
735 int nr_active;
736 int is_active;
737 int nr_stat;
738 int nr_freq;
739 int rotate_disable;
740 atomic_t refcount;
741 struct task_struct *task;
742
743 /*
744 * Context clock, runs when context enabled.
745 */
746 u64 time;
747 u64 timestamp;
748
749 /*
750 * These fields let us detect when two contexts have both
751 * been cloned (inherited) from a common ancestor.
752 */
753 struct perf_event_context *parent_ctx;
754 u64 parent_gen;
755 u64 generation;
756 int pin_count;
757#ifdef CONFIG_CGROUP_PERF
758 int nr_cgroups; /* cgroup evts */
759#endif
760 void *task_ctx_data; /* pmu specific data */
761 struct rcu_head rcu_head;
762};
763
764/*
765 * Number of contexts where an event can trigger:
766 * task, softirq, hardirq, nmi.
767 */
768#define PERF_NR_CONTEXTS 4
769
770/**
771 * struct perf_event_cpu_context - per cpu event context structure
772 */
773struct perf_cpu_context {
774 struct perf_event_context ctx;
775 struct perf_event_context *task_ctx;
776 int active_oncpu;
777 int exclusive;
778
779 raw_spinlock_t hrtimer_lock;
780 struct hrtimer hrtimer;
781 ktime_t hrtimer_interval;
782 unsigned int hrtimer_active;
783
784#ifdef CONFIG_CGROUP_PERF
785 struct perf_cgroup *cgrp;
786 struct list_head cgrp_cpuctx_entry;
787#endif
788
789 struct list_head sched_cb_entry;
790 int sched_cb_usage;
791
792 int online;
793};
794
795struct perf_output_handle {
796 struct perf_event *event;
797 struct ring_buffer *rb;
798 unsigned long wakeup;
799 unsigned long size;
800 u64 aux_flags;
801 union {
802 void *addr;
803 unsigned long head;
804 };
805 int page;
806};
807
808struct bpf_perf_event_data_kern {
809 bpf_user_pt_regs_t *regs;
810 struct perf_sample_data *data;
811 struct perf_event *event;
812};
813
814#ifdef CONFIG_CGROUP_PERF
815
816/*
817 * perf_cgroup_info keeps track of time_enabled for a cgroup.
818 * This is a per-cpu dynamically allocated data structure.
819 */
820struct perf_cgroup_info {
821 u64 time;
822 u64 timestamp;
823};
824
825struct perf_cgroup {
826 struct cgroup_subsys_state css;
827 struct perf_cgroup_info __percpu *info;
828};
829
830/*
831 * Must ensure cgroup is pinned (css_get) before calling
832 * this function. In other words, we cannot call this function
833 * if there is no cgroup event for the current CPU context.
834 */
835static inline struct perf_cgroup *
836perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
837{
838 return container_of(task_css_check(task, perf_event_cgrp_id,
839 ctx ? lockdep_is_held(&ctx->lock)
840 : true),
841 struct perf_cgroup, css);
842}
843#endif /* CONFIG_CGROUP_PERF */
844
845#ifdef CONFIG_PERF_EVENTS
846
847extern void *perf_aux_output_begin(struct perf_output_handle *handle,
848 struct perf_event *event);
849extern void perf_aux_output_end(struct perf_output_handle *handle,
850 unsigned long size);
851extern int perf_aux_output_skip(struct perf_output_handle *handle,
852 unsigned long size);
853extern void *perf_get_aux(struct perf_output_handle *handle);
854extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
855extern void perf_event_itrace_started(struct perf_event *event);
856
857extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
858extern void perf_pmu_unregister(struct pmu *pmu);
859
860extern int perf_num_counters(void);
861extern const char *perf_pmu_name(void);
862extern void __perf_event_task_sched_in(struct task_struct *prev,
863 struct task_struct *task);
864extern void __perf_event_task_sched_out(struct task_struct *prev,
865 struct task_struct *next);
866extern int perf_event_init_task(struct task_struct *child);
867extern void perf_event_exit_task(struct task_struct *child);
868extern void perf_event_free_task(struct task_struct *task);
869extern void perf_event_delayed_put(struct task_struct *task);
870extern struct file *perf_event_get(unsigned int fd);
871extern const struct perf_event *perf_get_event(struct file *file);
872extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
873extern void perf_event_print_debug(void);
874extern void perf_pmu_disable(struct pmu *pmu);
875extern void perf_pmu_enable(struct pmu *pmu);
876extern void perf_sched_cb_dec(struct pmu *pmu);
877extern void perf_sched_cb_inc(struct pmu *pmu);
878extern int perf_event_task_disable(void);
879extern int perf_event_task_enable(void);
880extern int perf_event_refresh(struct perf_event *event, int refresh);
881extern void perf_event_update_userpage(struct perf_event *event);
882extern int perf_event_release_kernel(struct perf_event *event);
883extern struct perf_event *
884perf_event_create_kernel_counter(struct perf_event_attr *attr,
885 int cpu,
886 struct task_struct *task,
887 perf_overflow_handler_t callback,
888 void *context);
889extern void perf_pmu_migrate_context(struct pmu *pmu,
890 int src_cpu, int dst_cpu);
891int perf_event_read_local(struct perf_event *event, u64 *value,
892 u64 *enabled, u64 *running);
893extern u64 perf_event_read_value(struct perf_event *event,
894 u64 *enabled, u64 *running);
895
896
897struct perf_sample_data {
898 /*
899 * Fields set by perf_sample_data_init(), group so as to
900 * minimize the cachelines touched.
901 */
902 u64 addr;
903 struct perf_raw_record *raw;
904 struct perf_branch_stack *br_stack;
905 u64 period;
906 u64 weight;
907 u64 txn;
908 union perf_mem_data_src data_src;
909
910 /*
911 * The other fields, optionally {set,used} by
912 * perf_{prepare,output}_sample().
913 */
914 u64 type;
915 u64 ip;
916 struct {
917 u32 pid;
918 u32 tid;
919 } tid_entry;
920 u64 time;
921 u64 id;
922 u64 stream_id;
923 struct {
924 u32 cpu;
925 u32 reserved;
926 } cpu_entry;
927 struct perf_callchain_entry *callchain;
928
929 /*
930 * regs_user may point to task_pt_regs or to regs_user_copy, depending
931 * on arch details.
932 */
933 struct perf_regs regs_user;
934 struct pt_regs regs_user_copy;
935
936 struct perf_regs regs_intr;
937 u64 stack_user_size;
938
939 u64 phys_addr;
940} ____cacheline_aligned;
941
942/* default value for data source */
943#define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
944 PERF_MEM_S(LVL, NA) |\
945 PERF_MEM_S(SNOOP, NA) |\
946 PERF_MEM_S(LOCK, NA) |\
947 PERF_MEM_S(TLB, NA))
948
949static inline void perf_sample_data_init(struct perf_sample_data *data,
950 u64 addr, u64 period)
951{
952 /* remaining struct members initialized in perf_prepare_sample() */
953 data->addr = addr;
954 data->raw = NULL;
955 data->br_stack = NULL;
956 data->period = period;
957 data->weight = 0;
958 data->data_src.val = PERF_MEM_NA;
959 data->txn = 0;
960}
961
962extern void perf_output_sample(struct perf_output_handle *handle,
963 struct perf_event_header *header,
964 struct perf_sample_data *data,
965 struct perf_event *event);
966extern void perf_prepare_sample(struct perf_event_header *header,
967 struct perf_sample_data *data,
968 struct perf_event *event,
969 struct pt_regs *regs);
970
971extern int perf_event_overflow(struct perf_event *event,
972 struct perf_sample_data *data,
973 struct pt_regs *regs);
974
975extern void perf_event_output_forward(struct perf_event *event,
976 struct perf_sample_data *data,
977 struct pt_regs *regs);
978extern void perf_event_output_backward(struct perf_event *event,
979 struct perf_sample_data *data,
980 struct pt_regs *regs);
981extern void perf_event_output(struct perf_event *event,
982 struct perf_sample_data *data,
983 struct pt_regs *regs);
984
985static inline bool
986is_default_overflow_handler(struct perf_event *event)
987{
988 if (likely(event->overflow_handler == perf_event_output_forward))
989 return true;
990 if (unlikely(event->overflow_handler == perf_event_output_backward))
991 return true;
992 return false;
993}
994
995extern void
996perf_event_header__init_id(struct perf_event_header *header,
997 struct perf_sample_data *data,
998 struct perf_event *event);
999extern void
1000perf_event__output_id_sample(struct perf_event *event,
1001 struct perf_output_handle *handle,
1002 struct perf_sample_data *sample);
1003
1004extern void
1005perf_log_lost_samples(struct perf_event *event, u64 lost);
1006
1007static inline bool is_sampling_event(struct perf_event *event)
1008{
1009 return event->attr.sample_period != 0;
1010}
1011
1012/*
1013 * Return 1 for a software event, 0 for a hardware event
1014 */
1015static inline int is_software_event(struct perf_event *event)
1016{
1017 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1018}
1019
1020/*
1021 * Return 1 for event in sw context, 0 for event in hw context
1022 */
1023static inline int in_software_context(struct perf_event *event)
1024{
1025 return event->ctx->pmu->task_ctx_nr == perf_sw_context;
1026}
1027
1028extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1029
1030extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1031extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1032
1033#ifndef perf_arch_fetch_caller_regs
1034static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1035#endif
1036
1037/*
1038 * Take a snapshot of the regs. Skip ip and frame pointer to
1039 * the nth caller. We only need a few of the regs:
1040 * - ip for PERF_SAMPLE_IP
1041 * - cs for user_mode() tests
1042 * - bp for callchains
1043 * - eflags, for future purposes, just in case
1044 */
1045static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1046{
1047 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1048}
1049
1050static __always_inline void
1051perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1052{
1053 if (static_key_false(&perf_swevent_enabled[event_id]))
1054 __perf_sw_event(event_id, nr, regs, addr);
1055}
1056
1057DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1058
1059/*
1060 * 'Special' version for the scheduler, it hard assumes no recursion,
1061 * which is guaranteed by us not actually scheduling inside other swevents
1062 * because those disable preemption.
1063 */
1064static __always_inline void
1065perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1066{
1067 if (static_key_false(&perf_swevent_enabled[event_id])) {
1068 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1069
1070 perf_fetch_caller_regs(regs);
1071 ___perf_sw_event(event_id, nr, regs, addr);
1072 }
1073}
1074
1075extern struct static_key_false perf_sched_events;
1076
1077static __always_inline bool
1078perf_sw_migrate_enabled(void)
1079{
1080 if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1081 return true;
1082 return false;
1083}
1084
1085static inline void perf_event_task_migrate(struct task_struct *task)
1086{
1087 if (perf_sw_migrate_enabled())
1088 task->sched_migrated = 1;
1089}
1090
1091static inline void perf_event_task_sched_in(struct task_struct *prev,
1092 struct task_struct *task)
1093{
1094 if (static_branch_unlikely(&perf_sched_events))
1095 __perf_event_task_sched_in(prev, task);
1096
1097 if (perf_sw_migrate_enabled() && task->sched_migrated) {
1098 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1099
1100 perf_fetch_caller_regs(regs);
1101 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1102 task->sched_migrated = 0;
1103 }
1104}
1105
1106static inline void perf_event_task_sched_out(struct task_struct *prev,
1107 struct task_struct *next)
1108{
1109 perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1110
1111 if (static_branch_unlikely(&perf_sched_events))
1112 __perf_event_task_sched_out(prev, next);
1113}
1114
1115extern void perf_event_mmap(struct vm_area_struct *vma);
1116extern struct perf_guest_info_callbacks *perf_guest_cbs;
1117extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1118extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1119
1120extern void perf_event_exec(void);
1121extern void perf_event_comm(struct task_struct *tsk, bool exec);
1122extern void perf_event_namespaces(struct task_struct *tsk);
1123extern void perf_event_fork(struct task_struct *tsk);
1124
1125/* Callchains */
1126DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1127
1128extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1129extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1130extern struct perf_callchain_entry *
1131get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1132 u32 max_stack, bool crosstask, bool add_mark);
1133extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1134extern int get_callchain_buffers(int max_stack);
1135extern void put_callchain_buffers(void);
1136
1137extern int sysctl_perf_event_max_stack;
1138extern int sysctl_perf_event_max_contexts_per_stack;
1139
1140static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1141{
1142 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1143 struct perf_callchain_entry *entry = ctx->entry;
1144 entry->ip[entry->nr++] = ip;
1145 ++ctx->contexts;
1146 return 0;
1147 } else {
1148 ctx->contexts_maxed = true;
1149 return -1; /* no more room, stop walking the stack */
1150 }
1151}
1152
1153static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1154{
1155 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1156 struct perf_callchain_entry *entry = ctx->entry;
1157 entry->ip[entry->nr++] = ip;
1158 ++ctx->nr;
1159 return 0;
1160 } else {
1161 return -1; /* no more room, stop walking the stack */
1162 }
1163}
1164
1165extern int sysctl_perf_event_paranoid;
1166extern int sysctl_perf_event_mlock;
1167extern int sysctl_perf_event_sample_rate;
1168extern int sysctl_perf_cpu_time_max_percent;
1169
1170extern void perf_sample_event_took(u64 sample_len_ns);
1171
1172extern int perf_proc_update_handler(struct ctl_table *table, int write,
1173 void __user *buffer, size_t *lenp,
1174 loff_t *ppos);
1175extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1176 void __user *buffer, size_t *lenp,
1177 loff_t *ppos);
1178
1179int perf_event_max_stack_handler(struct ctl_table *table, int write,
1180 void __user *buffer, size_t *lenp, loff_t *ppos);
1181
1182static inline bool perf_paranoid_tracepoint_raw(void)
1183{
1184 return sysctl_perf_event_paranoid > -1;
1185}
1186
1187static inline bool perf_paranoid_cpu(void)
1188{
1189 return sysctl_perf_event_paranoid > 0;
1190}
1191
1192static inline bool perf_paranoid_kernel(void)
1193{
1194 return sysctl_perf_event_paranoid > 1;
1195}
1196
1197extern void perf_event_init(void);
1198extern void perf_tp_event(u16 event_type, u64 count, void *record,
1199 int entry_size, struct pt_regs *regs,
1200 struct hlist_head *head, int rctx,
1201 struct task_struct *task);
1202extern void perf_bp_event(struct perf_event *event, void *data);
1203
1204#ifndef perf_misc_flags
1205# define perf_misc_flags(regs) \
1206 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1207# define perf_instruction_pointer(regs) instruction_pointer(regs)
1208#endif
1209#ifndef perf_arch_bpf_user_pt_regs
1210# define perf_arch_bpf_user_pt_regs(regs) regs
1211#endif
1212
1213static inline bool has_branch_stack(struct perf_event *event)
1214{
1215 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1216}
1217
1218static inline bool needs_branch_stack(struct perf_event *event)
1219{
1220 return event->attr.branch_sample_type != 0;
1221}
1222
1223static inline bool has_aux(struct perf_event *event)
1224{
1225 return event->pmu->setup_aux;
1226}
1227
1228static inline bool is_write_backward(struct perf_event *event)
1229{
1230 return !!event->attr.write_backward;
1231}
1232
1233static inline bool has_addr_filter(struct perf_event *event)
1234{
1235 return event->pmu->nr_addr_filters;
1236}
1237
1238/*
1239 * An inherited event uses parent's filters
1240 */
1241static inline struct perf_addr_filters_head *
1242perf_event_addr_filters(struct perf_event *event)
1243{
1244 struct perf_addr_filters_head *ifh = &event->addr_filters;
1245
1246 if (event->parent)
1247 ifh = &event->parent->addr_filters;
1248
1249 return ifh;
1250}
1251
1252extern void perf_event_addr_filters_sync(struct perf_event *event);
1253
1254extern int perf_output_begin(struct perf_output_handle *handle,
1255 struct perf_event *event, unsigned int size);
1256extern int perf_output_begin_forward(struct perf_output_handle *handle,
1257 struct perf_event *event,
1258 unsigned int size);
1259extern int perf_output_begin_backward(struct perf_output_handle *handle,
1260 struct perf_event *event,
1261 unsigned int size);
1262
1263extern void perf_output_end(struct perf_output_handle *handle);
1264extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1265 const void *buf, unsigned int len);
1266extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1267 unsigned int len);
1268extern int perf_swevent_get_recursion_context(void);
1269extern void perf_swevent_put_recursion_context(int rctx);
1270extern u64 perf_swevent_set_period(struct perf_event *event);
1271extern void perf_event_enable(struct perf_event *event);
1272extern void perf_event_disable(struct perf_event *event);
1273extern void perf_event_disable_local(struct perf_event *event);
1274extern void perf_event_disable_inatomic(struct perf_event *event);
1275extern void perf_event_task_tick(void);
1276extern int perf_event_account_interrupt(struct perf_event *event);
1277#else /* !CONFIG_PERF_EVENTS: */
1278static inline void *
1279perf_aux_output_begin(struct perf_output_handle *handle,
1280 struct perf_event *event) { return NULL; }
1281static inline void
1282perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1283 { }
1284static inline int
1285perf_aux_output_skip(struct perf_output_handle *handle,
1286 unsigned long size) { return -EINVAL; }
1287static inline void *
1288perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1289static inline void
1290perf_event_task_migrate(struct task_struct *task) { }
1291static inline void
1292perf_event_task_sched_in(struct task_struct *prev,
1293 struct task_struct *task) { }
1294static inline void
1295perf_event_task_sched_out(struct task_struct *prev,
1296 struct task_struct *next) { }
1297static inline int perf_event_init_task(struct task_struct *child) { return 0; }
1298static inline void perf_event_exit_task(struct task_struct *child) { }
1299static inline void perf_event_free_task(struct task_struct *task) { }
1300static inline void perf_event_delayed_put(struct task_struct *task) { }
1301static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1302static inline const struct perf_event *perf_get_event(struct file *file)
1303{
1304 return ERR_PTR(-EINVAL);
1305}
1306static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1307{
1308 return ERR_PTR(-EINVAL);
1309}
1310static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1311 u64 *enabled, u64 *running)
1312{
1313 return -EINVAL;
1314}
1315static inline void perf_event_print_debug(void) { }
1316static inline int perf_event_task_disable(void) { return -EINVAL; }
1317static inline int perf_event_task_enable(void) { return -EINVAL; }
1318static inline int perf_event_refresh(struct perf_event *event, int refresh)
1319{
1320 return -EINVAL;
1321}
1322
1323static inline void
1324perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1325static inline void
1326perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) { }
1327static inline void
1328perf_bp_event(struct perf_event *event, void *data) { }
1329
1330static inline int perf_register_guest_info_callbacks
1331(struct perf_guest_info_callbacks *callbacks) { return 0; }
1332static inline int perf_unregister_guest_info_callbacks
1333(struct perf_guest_info_callbacks *callbacks) { return 0; }
1334
1335static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1336static inline void perf_event_exec(void) { }
1337static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1338static inline void perf_event_namespaces(struct task_struct *tsk) { }
1339static inline void perf_event_fork(struct task_struct *tsk) { }
1340static inline void perf_event_init(void) { }
1341static inline int perf_swevent_get_recursion_context(void) { return -1; }
1342static inline void perf_swevent_put_recursion_context(int rctx) { }
1343static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1344static inline void perf_event_enable(struct perf_event *event) { }
1345static inline void perf_event_disable(struct perf_event *event) { }
1346static inline int __perf_event_disable(void *info) { return -1; }
1347static inline void perf_event_task_tick(void) { }
1348static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1349#endif
1350
1351#if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1352extern void perf_restore_debug_store(void);
1353#else
1354static inline void perf_restore_debug_store(void) { }
1355#endif
1356
1357static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1358{
1359 return frag->pad < sizeof(u64);
1360}
1361
1362#define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1363
1364struct perf_pmu_events_attr {
1365 struct device_attribute attr;
1366 u64 id;
1367 const char *event_str;
1368};
1369
1370struct perf_pmu_events_ht_attr {
1371 struct device_attribute attr;
1372 u64 id;
1373 const char *event_str_ht;
1374 const char *event_str_noht;
1375};
1376
1377ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1378 char *page);
1379
1380#define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1381static struct perf_pmu_events_attr _var = { \
1382 .attr = __ATTR(_name, 0444, _show, NULL), \
1383 .id = _id, \
1384};
1385
1386#define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1387static struct perf_pmu_events_attr _var = { \
1388 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1389 .id = 0, \
1390 .event_str = _str, \
1391};
1392
1393#define PMU_FORMAT_ATTR(_name, _format) \
1394static ssize_t \
1395_name##_show(struct device *dev, \
1396 struct device_attribute *attr, \
1397 char *page) \
1398{ \
1399 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1400 return sprintf(page, _format "\n"); \
1401} \
1402 \
1403static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1404
1405/* Performance counter hotplug functions */
1406#ifdef CONFIG_PERF_EVENTS
1407int perf_event_init_cpu(unsigned int cpu);
1408int perf_event_exit_cpu(unsigned int cpu);
1409#else
1410#define perf_event_init_cpu NULL
1411#define perf_event_exit_cpu NULL
1412#endif
1413
1414#endif /* _LINUX_PERF_EVENT_H */
1415