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
29#define PERF_GUEST_ACTIVE 0x01
30#define PERF_GUEST_USER 0x02
31
32struct perf_guest_info_callbacks {
33 unsigned int (*state)(void);
34 unsigned long (*get_ip)(void);
35 unsigned int (*handle_intel_pt_intr)(void);
36};
37
38#ifdef CONFIG_HAVE_HW_BREAKPOINT
39#include <linux/rhashtable-types.h>
40#include <asm/hw_breakpoint.h>
41#endif
42
43#include <linux/list.h>
44#include <linux/mutex.h>
45#include <linux/rculist.h>
46#include <linux/rcupdate.h>
47#include <linux/spinlock.h>
48#include <linux/hrtimer.h>
49#include <linux/fs.h>
50#include <linux/pid_namespace.h>
51#include <linux/workqueue.h>
52#include <linux/ftrace.h>
53#include <linux/cpu.h>
54#include <linux/irq_work.h>
55#include <linux/static_key.h>
56#include <linux/jump_label_ratelimit.h>
57#include <linux/atomic.h>
58#include <linux/sysfs.h>
59#include <linux/perf_regs.h>
60#include <linux/cgroup.h>
61#include <linux/refcount.h>
62#include <linux/security.h>
63#include <linux/static_call.h>
64#include <linux/lockdep.h>
65#include <asm/local.h>
66
67struct perf_callchain_entry {
68 __u64 nr;
69 __u64 ip[]; /* /proc/sys/kernel/perf_event_max_stack */
70};
71
72struct perf_callchain_entry_ctx {
73 struct perf_callchain_entry *entry;
74 u32 max_stack;
75 u32 nr;
76 short contexts;
77 bool contexts_maxed;
78};
79
80typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
81 unsigned long off, unsigned long len);
82
83struct perf_raw_frag {
84 union {
85 struct perf_raw_frag *next;
86 unsigned long pad;
87 };
88 perf_copy_f copy;
89 void *data;
90 u32 size;
91} __packed;
92
93struct perf_raw_record {
94 struct perf_raw_frag frag;
95 u32 size;
96};
97
98static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
99{
100 return frag->pad < sizeof(u64);
101}
102
103/*
104 * branch stack layout:
105 * nr: number of taken branches stored in entries[]
106 * hw_idx: The low level index of raw branch records
107 * for the most recent branch.
108 * -1ULL means invalid/unknown.
109 *
110 * Note that nr can vary from sample to sample
111 * branches (to, from) are stored from most recent
112 * to least recent, i.e., entries[0] contains the most
113 * recent branch.
114 * The entries[] is an abstraction of raw branch records,
115 * which may not be stored in age order in HW, e.g. Intel LBR.
116 * The hw_idx is to expose the low level index of raw
117 * branch record for the most recent branch aka entries[0].
118 * The hw_idx index is between -1 (unknown) and max depth,
119 * which can be retrieved in /sys/devices/cpu/caps/branches.
120 * For the architectures whose raw branch records are
121 * already stored in age order, the hw_idx should be 0.
122 */
123struct perf_branch_stack {
124 __u64 nr;
125 __u64 hw_idx;
126 struct perf_branch_entry entries[];
127};
128
129struct task_struct;
130
131/*
132 * extra PMU register associated with an event
133 */
134struct hw_perf_event_extra {
135 u64 config; /* register value */
136 unsigned int reg; /* register address or index */
137 int alloc; /* extra register already allocated */
138 int idx; /* index in shared_regs->regs[] */
139};
140
141/**
142 * hw_perf_event::flag values
143 *
144 * PERF_EVENT_FLAG_ARCH bits are reserved for architecture-specific
145 * usage.
146 */
147#define PERF_EVENT_FLAG_ARCH 0x000fffff
148#define PERF_EVENT_FLAG_USER_READ_CNT 0x80000000
149
150static_assert((PERF_EVENT_FLAG_USER_READ_CNT & PERF_EVENT_FLAG_ARCH) == 0);
151
152/**
153 * struct hw_perf_event - performance event hardware details:
154 */
155struct hw_perf_event {
156#ifdef CONFIG_PERF_EVENTS
157 union {
158 struct { /* hardware */
159 u64 config;
160 u64 last_tag;
161 unsigned long config_base;
162 unsigned long event_base;
163 int event_base_rdpmc;
164 int idx;
165 int last_cpu;
166 int flags;
167
168 struct hw_perf_event_extra extra_reg;
169 struct hw_perf_event_extra branch_reg;
170 };
171 struct { /* software */
172 struct hrtimer hrtimer;
173 };
174 struct { /* tracepoint */
175 /* for tp_event->class */
176 struct list_head tp_list;
177 };
178 struct { /* amd_power */
179 u64 pwr_acc;
180 u64 ptsc;
181 };
182#ifdef CONFIG_HAVE_HW_BREAKPOINT
183 struct { /* breakpoint */
184 /*
185 * Crufty hack to avoid the chicken and egg
186 * problem hw_breakpoint has with context
187 * creation and event initalization.
188 */
189 struct arch_hw_breakpoint info;
190 struct rhlist_head bp_list;
191 };
192#endif
193 struct { /* amd_iommu */
194 u8 iommu_bank;
195 u8 iommu_cntr;
196 u16 padding;
197 u64 conf;
198 u64 conf1;
199 };
200 };
201 /*
202 * If the event is a per task event, this will point to the task in
203 * question. See the comment in perf_event_alloc().
204 */
205 struct task_struct *target;
206
207 /*
208 * PMU would store hardware filter configuration
209 * here.
210 */
211 void *addr_filters;
212
213 /* Last sync'ed generation of filters */
214 unsigned long addr_filters_gen;
215
216/*
217 * hw_perf_event::state flags; used to track the PERF_EF_* state.
218 */
219#define PERF_HES_STOPPED 0x01 /* the counter is stopped */
220#define PERF_HES_UPTODATE 0x02 /* event->count up-to-date */
221#define PERF_HES_ARCH 0x04
222
223 int state;
224
225 /*
226 * The last observed hardware counter value, updated with a
227 * local64_cmpxchg() such that pmu::read() can be called nested.
228 */
229 local64_t prev_count;
230
231 /*
232 * The period to start the next sample with.
233 */
234 u64 sample_period;
235
236 union {
237 struct { /* Sampling */
238 /*
239 * The period we started this sample with.
240 */
241 u64 last_period;
242
243 /*
244 * However much is left of the current period;
245 * note that this is a full 64bit value and
246 * allows for generation of periods longer
247 * than hardware might allow.
248 */
249 local64_t period_left;
250 };
251 struct { /* Topdown events counting for context switch */
252 u64 saved_metric;
253 u64 saved_slots;
254 };
255 };
256
257 /*
258 * State for throttling the event, see __perf_event_overflow() and
259 * perf_adjust_freq_unthr_context().
260 */
261 u64 interrupts_seq;
262 u64 interrupts;
263
264 /*
265 * State for freq target events, see __perf_event_overflow() and
266 * perf_adjust_freq_unthr_context().
267 */
268 u64 freq_time_stamp;
269 u64 freq_count_stamp;
270#endif
271};
272
273struct perf_event;
274struct perf_event_pmu_context;
275
276/*
277 * Common implementation detail of pmu::{start,commit,cancel}_txn
278 */
279#define PERF_PMU_TXN_ADD 0x1 /* txn to add/schedule event on PMU */
280#define PERF_PMU_TXN_READ 0x2 /* txn to read event group from PMU */
281
282/**
283 * pmu::capabilities flags
284 */
285#define PERF_PMU_CAP_NO_INTERRUPT 0x0001
286#define PERF_PMU_CAP_NO_NMI 0x0002
287#define PERF_PMU_CAP_AUX_NO_SG 0x0004
288#define PERF_PMU_CAP_EXTENDED_REGS 0x0008
289#define PERF_PMU_CAP_EXCLUSIVE 0x0010
290#define PERF_PMU_CAP_ITRACE 0x0020
291#define PERF_PMU_CAP_NO_EXCLUDE 0x0040
292#define PERF_PMU_CAP_AUX_OUTPUT 0x0080
293#define PERF_PMU_CAP_EXTENDED_HW_TYPE 0x0100
294
295struct perf_output_handle;
296
297#define PMU_NULL_DEV ((void *)(~0UL))
298
299/**
300 * struct pmu - generic performance monitoring unit
301 */
302struct pmu {
303 struct list_head entry;
304
305 struct module *module;
306 struct device *dev;
307 struct device *parent;
308 const struct attribute_group **attr_groups;
309 const struct attribute_group **attr_update;
310 const char *name;
311 int type;
312
313 /*
314 * various common per-pmu feature flags
315 */
316 int capabilities;
317
318 int __percpu *pmu_disable_count;
319 struct perf_cpu_pmu_context __percpu *cpu_pmu_context;
320 atomic_t exclusive_cnt; /* < 0: cpu; > 0: tsk */
321 int task_ctx_nr;
322 int hrtimer_interval_ms;
323
324 /* number of address filters this PMU can do */
325 unsigned int nr_addr_filters;
326
327 /*
328 * Fully disable/enable this PMU, can be used to protect from the PMI
329 * as well as for lazy/batch writing of the MSRs.
330 */
331 void (*pmu_enable) (struct pmu *pmu); /* optional */
332 void (*pmu_disable) (struct pmu *pmu); /* optional */
333
334 /*
335 * Try and initialize the event for this PMU.
336 *
337 * Returns:
338 * -ENOENT -- @event is not for this PMU
339 *
340 * -ENODEV -- @event is for this PMU but PMU not present
341 * -EBUSY -- @event is for this PMU but PMU temporarily unavailable
342 * -EINVAL -- @event is for this PMU but @event is not valid
343 * -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
344 * -EACCES -- @event is for this PMU, @event is valid, but no privileges
345 *
346 * 0 -- @event is for this PMU and valid
347 *
348 * Other error return values are allowed.
349 */
350 int (*event_init) (struct perf_event *event);
351
352 /*
353 * Notification that the event was mapped or unmapped. Called
354 * in the context of the mapping task.
355 */
356 void (*event_mapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
357 void (*event_unmapped) (struct perf_event *event, struct mm_struct *mm); /* optional */
358
359 /*
360 * Flags for ->add()/->del()/ ->start()/->stop(). There are
361 * matching hw_perf_event::state flags.
362 */
363#define PERF_EF_START 0x01 /* start the counter when adding */
364#define PERF_EF_RELOAD 0x02 /* reload the counter when starting */
365#define PERF_EF_UPDATE 0x04 /* update the counter when stopping */
366
367 /*
368 * Adds/Removes a counter to/from the PMU, can be done inside a
369 * transaction, see the ->*_txn() methods.
370 *
371 * The add/del callbacks will reserve all hardware resources required
372 * to service the event, this includes any counter constraint
373 * scheduling etc.
374 *
375 * Called with IRQs disabled and the PMU disabled on the CPU the event
376 * is on.
377 *
378 * ->add() called without PERF_EF_START should result in the same state
379 * as ->add() followed by ->stop().
380 *
381 * ->del() must always PERF_EF_UPDATE stop an event. If it calls
382 * ->stop() that must deal with already being stopped without
383 * PERF_EF_UPDATE.
384 */
385 int (*add) (struct perf_event *event, int flags);
386 void (*del) (struct perf_event *event, int flags);
387
388 /*
389 * Starts/Stops a counter present on the PMU.
390 *
391 * The PMI handler should stop the counter when perf_event_overflow()
392 * returns !0. ->start() will be used to continue.
393 *
394 * Also used to change the sample period.
395 *
396 * Called with IRQs disabled and the PMU disabled on the CPU the event
397 * is on -- will be called from NMI context with the PMU generates
398 * NMIs.
399 *
400 * ->stop() with PERF_EF_UPDATE will read the counter and update
401 * period/count values like ->read() would.
402 *
403 * ->start() with PERF_EF_RELOAD will reprogram the counter
404 * value, must be preceded by a ->stop() with PERF_EF_UPDATE.
405 */
406 void (*start) (struct perf_event *event, int flags);
407 void (*stop) (struct perf_event *event, int flags);
408
409 /*
410 * Updates the counter value of the event.
411 *
412 * For sampling capable PMUs this will also update the software period
413 * hw_perf_event::period_left field.
414 */
415 void (*read) (struct perf_event *event);
416
417 /*
418 * Group events scheduling is treated as a transaction, add
419 * group events as a whole and perform one schedulability test.
420 * If the test fails, roll back the whole group
421 *
422 * Start the transaction, after this ->add() doesn't need to
423 * do schedulability tests.
424 *
425 * Optional.
426 */
427 void (*start_txn) (struct pmu *pmu, unsigned int txn_flags);
428 /*
429 * If ->start_txn() disabled the ->add() schedulability test
430 * then ->commit_txn() is required to perform one. On success
431 * the transaction is closed. On error the transaction is kept
432 * open until ->cancel_txn() is called.
433 *
434 * Optional.
435 */
436 int (*commit_txn) (struct pmu *pmu);
437 /*
438 * Will cancel the transaction, assumes ->del() is called
439 * for each successful ->add() during the transaction.
440 *
441 * Optional.
442 */
443 void (*cancel_txn) (struct pmu *pmu);
444
445 /*
446 * Will return the value for perf_event_mmap_page::index for this event,
447 * if no implementation is provided it will default to 0 (see
448 * perf_event_idx_default).
449 */
450 int (*event_idx) (struct perf_event *event); /*optional */
451
452 /*
453 * context-switches callback
454 */
455 void (*sched_task) (struct perf_event_pmu_context *pmu_ctx,
456 bool sched_in);
457
458 /*
459 * Kmem cache of PMU specific data
460 */
461 struct kmem_cache *task_ctx_cache;
462
463 /*
464 * PMU specific parts of task perf event context (i.e. ctx->task_ctx_data)
465 * can be synchronized using this function. See Intel LBR callstack support
466 * implementation and Perf core context switch handling callbacks for usage
467 * examples.
468 */
469 void (*swap_task_ctx) (struct perf_event_pmu_context *prev_epc,
470 struct perf_event_pmu_context *next_epc);
471 /* optional */
472
473 /*
474 * Set up pmu-private data structures for an AUX area
475 */
476 void *(*setup_aux) (struct perf_event *event, void **pages,
477 int nr_pages, bool overwrite);
478 /* optional */
479
480 /*
481 * Free pmu-private AUX data structures
482 */
483 void (*free_aux) (void *aux); /* optional */
484
485 /*
486 * Take a snapshot of the AUX buffer without touching the event
487 * state, so that preempting ->start()/->stop() callbacks does
488 * not interfere with their logic. Called in PMI context.
489 *
490 * Returns the size of AUX data copied to the output handle.
491 *
492 * Optional.
493 */
494 long (*snapshot_aux) (struct perf_event *event,
495 struct perf_output_handle *handle,
496 unsigned long size);
497
498 /*
499 * Validate address range filters: make sure the HW supports the
500 * requested configuration and number of filters; return 0 if the
501 * supplied filters are valid, -errno otherwise.
502 *
503 * Runs in the context of the ioctl()ing process and is not serialized
504 * with the rest of the PMU callbacks.
505 */
506 int (*addr_filters_validate) (struct list_head *filters);
507 /* optional */
508
509 /*
510 * Synchronize address range filter configuration:
511 * translate hw-agnostic filters into hardware configuration in
512 * event::hw::addr_filters.
513 *
514 * Runs as a part of filter sync sequence that is done in ->start()
515 * callback by calling perf_event_addr_filters_sync().
516 *
517 * May (and should) traverse event::addr_filters::list, for which its
518 * caller provides necessary serialization.
519 */
520 void (*addr_filters_sync) (struct perf_event *event);
521 /* optional */
522
523 /*
524 * Check if event can be used for aux_output purposes for
525 * events of this PMU.
526 *
527 * Runs from perf_event_open(). Should return 0 for "no match"
528 * or non-zero for "match".
529 */
530 int (*aux_output_match) (struct perf_event *event);
531 /* optional */
532
533 /*
534 * Skip programming this PMU on the given CPU. Typically needed for
535 * big.LITTLE things.
536 */
537 bool (*filter) (struct pmu *pmu, int cpu); /* optional */
538
539 /*
540 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
541 */
542 int (*check_period) (struct perf_event *event, u64 value); /* optional */
543};
544
545enum perf_addr_filter_action_t {
546 PERF_ADDR_FILTER_ACTION_STOP = 0,
547 PERF_ADDR_FILTER_ACTION_START,
548 PERF_ADDR_FILTER_ACTION_FILTER,
549};
550
551/**
552 * struct perf_addr_filter - address range filter definition
553 * @entry: event's filter list linkage
554 * @path: object file's path for file-based filters
555 * @offset: filter range offset
556 * @size: filter range size (size==0 means single address trigger)
557 * @action: filter/start/stop
558 *
559 * This is a hardware-agnostic filter configuration as specified by the user.
560 */
561struct perf_addr_filter {
562 struct list_head entry;
563 struct path path;
564 unsigned long offset;
565 unsigned long size;
566 enum perf_addr_filter_action_t action;
567};
568
569/**
570 * struct perf_addr_filters_head - container for address range filters
571 * @list: list of filters for this event
572 * @lock: spinlock that serializes accesses to the @list and event's
573 * (and its children's) filter generations.
574 * @nr_file_filters: number of file-based filters
575 *
576 * A child event will use parent's @list (and therefore @lock), so they are
577 * bundled together; see perf_event_addr_filters().
578 */
579struct perf_addr_filters_head {
580 struct list_head list;
581 raw_spinlock_t lock;
582 unsigned int nr_file_filters;
583};
584
585struct perf_addr_filter_range {
586 unsigned long start;
587 unsigned long size;
588};
589
590/**
591 * enum perf_event_state - the states of an event:
592 */
593enum perf_event_state {
594 PERF_EVENT_STATE_DEAD = -4,
595 PERF_EVENT_STATE_EXIT = -3,
596 PERF_EVENT_STATE_ERROR = -2,
597 PERF_EVENT_STATE_OFF = -1,
598 PERF_EVENT_STATE_INACTIVE = 0,
599 PERF_EVENT_STATE_ACTIVE = 1,
600};
601
602struct file;
603struct perf_sample_data;
604
605typedef void (*perf_overflow_handler_t)(struct perf_event *,
606 struct perf_sample_data *,
607 struct pt_regs *regs);
608
609/*
610 * Event capabilities. For event_caps and groups caps.
611 *
612 * PERF_EV_CAP_SOFTWARE: Is a software event.
613 * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
614 * from any CPU in the package where it is active.
615 * PERF_EV_CAP_SIBLING: An event with this flag must be a group sibling and
616 * cannot be a group leader. If an event with this flag is detached from the
617 * group it is scheduled out and moved into an unrecoverable ERROR state.
618 */
619#define PERF_EV_CAP_SOFTWARE BIT(0)
620#define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1)
621#define PERF_EV_CAP_SIBLING BIT(2)
622
623#define SWEVENT_HLIST_BITS 8
624#define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS)
625
626struct swevent_hlist {
627 struct hlist_head heads[SWEVENT_HLIST_SIZE];
628 struct rcu_head rcu_head;
629};
630
631#define PERF_ATTACH_CONTEXT 0x01
632#define PERF_ATTACH_GROUP 0x02
633#define PERF_ATTACH_TASK 0x04
634#define PERF_ATTACH_TASK_DATA 0x08
635#define PERF_ATTACH_ITRACE 0x10
636#define PERF_ATTACH_SCHED_CB 0x20
637#define PERF_ATTACH_CHILD 0x40
638
639struct bpf_prog;
640struct perf_cgroup;
641struct perf_buffer;
642
643struct pmu_event_list {
644 raw_spinlock_t lock;
645 struct list_head list;
646};
647
648/*
649 * event->sibling_list is modified whole holding both ctx->lock and ctx->mutex
650 * as such iteration must hold either lock. However, since ctx->lock is an IRQ
651 * safe lock, and is only held by the CPU doing the modification, having IRQs
652 * disabled is sufficient since it will hold-off the IPIs.
653 */
654#ifdef CONFIG_PROVE_LOCKING
655#define lockdep_assert_event_ctx(event) \
656 WARN_ON_ONCE(__lockdep_enabled && \
657 (this_cpu_read(hardirqs_enabled) && \
658 lockdep_is_held(&(event)->ctx->mutex) != LOCK_STATE_HELD))
659#else
660#define lockdep_assert_event_ctx(event)
661#endif
662
663#define for_each_sibling_event(sibling, event) \
664 lockdep_assert_event_ctx(event); \
665 if ((event)->group_leader == (event)) \
666 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
667
668/**
669 * struct perf_event - performance event kernel representation:
670 */
671struct perf_event {
672#ifdef CONFIG_PERF_EVENTS
673 /*
674 * entry onto perf_event_context::event_list;
675 * modifications require ctx->lock
676 * RCU safe iterations.
677 */
678 struct list_head event_entry;
679
680 /*
681 * Locked for modification by both ctx->mutex and ctx->lock; holding
682 * either sufficies for read.
683 */
684 struct list_head sibling_list;
685 struct list_head active_list;
686 /*
687 * Node on the pinned or flexible tree located at the event context;
688 */
689 struct rb_node group_node;
690 u64 group_index;
691 /*
692 * We need storage to track the entries in perf_pmu_migrate_context; we
693 * cannot use the event_entry because of RCU and we want to keep the
694 * group in tact which avoids us using the other two entries.
695 */
696 struct list_head migrate_entry;
697
698 struct hlist_node hlist_entry;
699 struct list_head active_entry;
700 int nr_siblings;
701
702 /* Not serialized. Only written during event initialization. */
703 int event_caps;
704 /* The cumulative AND of all event_caps for events in this group. */
705 int group_caps;
706
707 unsigned int group_generation;
708 struct perf_event *group_leader;
709 /*
710 * event->pmu will always point to pmu in which this event belongs.
711 * Whereas event->pmu_ctx->pmu may point to other pmu when group of
712 * different pmu events is created.
713 */
714 struct pmu *pmu;
715 void *pmu_private;
716
717 enum perf_event_state state;
718 unsigned int attach_state;
719 local64_t count;
720 atomic64_t child_count;
721
722 /*
723 * These are the total time in nanoseconds that the event
724 * has been enabled (i.e. eligible to run, and the task has
725 * been scheduled in, if this is a per-task event)
726 * and running (scheduled onto the CPU), respectively.
727 */
728 u64 total_time_enabled;
729 u64 total_time_running;
730 u64 tstamp;
731
732 struct perf_event_attr attr;
733 u16 header_size;
734 u16 id_header_size;
735 u16 read_size;
736 struct hw_perf_event hw;
737
738 struct perf_event_context *ctx;
739 /*
740 * event->pmu_ctx points to perf_event_pmu_context in which the event
741 * is added. This pmu_ctx can be of other pmu for sw event when that
742 * sw event is part of a group which also contains non-sw events.
743 */
744 struct perf_event_pmu_context *pmu_ctx;
745 atomic_long_t refcount;
746
747 /*
748 * These accumulate total time (in nanoseconds) that children
749 * events have been enabled and running, respectively.
750 */
751 atomic64_t child_total_time_enabled;
752 atomic64_t child_total_time_running;
753
754 /*
755 * Protect attach/detach and child_list:
756 */
757 struct mutex child_mutex;
758 struct list_head child_list;
759 struct perf_event *parent;
760
761 int oncpu;
762 int cpu;
763
764 struct list_head owner_entry;
765 struct task_struct *owner;
766
767 /* mmap bits */
768 struct mutex mmap_mutex;
769 atomic_t mmap_count;
770
771 struct perf_buffer *rb;
772 struct list_head rb_entry;
773 unsigned long rcu_batches;
774 int rcu_pending;
775
776 /* poll related */
777 wait_queue_head_t waitq;
778 struct fasync_struct *fasync;
779
780 /* delayed work for NMIs and such */
781 unsigned int pending_wakeup;
782 unsigned int pending_kill;
783 unsigned int pending_disable;
784 unsigned int pending_sigtrap;
785 unsigned long pending_addr; /* SIGTRAP */
786 struct irq_work pending_irq;
787 struct callback_head pending_task;
788 unsigned int pending_work;
789
790 atomic_t event_limit;
791
792 /* address range filters */
793 struct perf_addr_filters_head addr_filters;
794 /* vma address array for file-based filders */
795 struct perf_addr_filter_range *addr_filter_ranges;
796 unsigned long addr_filters_gen;
797
798 /* for aux_output events */
799 struct perf_event *aux_event;
800
801 void (*destroy)(struct perf_event *);
802 struct rcu_head rcu_head;
803
804 struct pid_namespace *ns;
805 u64 id;
806
807 atomic64_t lost_samples;
808
809 u64 (*clock)(void);
810 perf_overflow_handler_t overflow_handler;
811 void *overflow_handler_context;
812#ifdef CONFIG_BPF_SYSCALL
813 perf_overflow_handler_t orig_overflow_handler;
814 struct bpf_prog *prog;
815 u64 bpf_cookie;
816#endif
817
818#ifdef CONFIG_EVENT_TRACING
819 struct trace_event_call *tp_event;
820 struct event_filter *filter;
821#ifdef CONFIG_FUNCTION_TRACER
822 struct ftrace_ops ftrace_ops;
823#endif
824#endif
825
826#ifdef CONFIG_CGROUP_PERF
827 struct perf_cgroup *cgrp; /* cgroup event is attach to */
828#endif
829
830#ifdef CONFIG_SECURITY
831 void *security;
832#endif
833 struct list_head sb_list;
834
835 /*
836 * Certain events gets forwarded to another pmu internally by over-
837 * writing kernel copy of event->attr.type without user being aware
838 * of it. event->orig_type contains original 'type' requested by
839 * user.
840 */
841 __u32 orig_type;
842#endif /* CONFIG_PERF_EVENTS */
843};
844
845/*
846 * ,-----------------------[1:n]------------------------.
847 * V V
848 * perf_event_context <-[1:n]-> perf_event_pmu_context <-[1:n]- perf_event
849 * | |
850 * `--[n:1]-> pmu <-[1:n]--'
851 *
852 *
853 * struct perf_event_pmu_context lifetime is refcount based and RCU freed
854 * (similar to perf_event_context). Locking is as if it were a member of
855 * perf_event_context; specifically:
856 *
857 * modification, both: ctx->mutex && ctx->lock
858 * reading, either: ctx->mutex || ctx->lock
859 *
860 * There is one exception to this; namely put_pmu_ctx() isn't always called
861 * with ctx->mutex held; this means that as long as we can guarantee the epc
862 * has events the above rules hold.
863 *
864 * Specificially, sys_perf_event_open()'s group_leader case depends on
865 * ctx->mutex pinning the configuration. Since we hold a reference on
866 * group_leader (through the filedesc) it can't go away, therefore it's
867 * associated pmu_ctx must exist and cannot change due to ctx->mutex.
868 *
869 * perf_event holds a refcount on perf_event_context
870 * perf_event holds a refcount on perf_event_pmu_context
871 */
872struct perf_event_pmu_context {
873 struct pmu *pmu;
874 struct perf_event_context *ctx;
875
876 struct list_head pmu_ctx_entry;
877
878 struct list_head pinned_active;
879 struct list_head flexible_active;
880
881 /* Used to avoid freeing per-cpu perf_event_pmu_context */
882 unsigned int embedded : 1;
883
884 unsigned int nr_events;
885 unsigned int nr_cgroups;
886
887 atomic_t refcount; /* event <-> epc */
888 struct rcu_head rcu_head;
889
890 void *task_ctx_data; /* pmu specific data */
891 /*
892 * Set when one or more (plausibly active) event can't be scheduled
893 * due to pmu overcommit or pmu constraints, except tolerant to
894 * events not necessary to be active due to scheduling constraints,
895 * such as cgroups.
896 */
897 int rotate_necessary;
898};
899
900struct perf_event_groups {
901 struct rb_root tree;
902 u64 index;
903};
904
905
906/**
907 * struct perf_event_context - event context structure
908 *
909 * Used as a container for task events and CPU events as well:
910 */
911struct perf_event_context {
912 /*
913 * Protect the states of the events in the list,
914 * nr_active, and the list:
915 */
916 raw_spinlock_t lock;
917 /*
918 * Protect the list of events. Locking either mutex or lock
919 * is sufficient to ensure the list doesn't change; to change
920 * the list you need to lock both the mutex and the spinlock.
921 */
922 struct mutex mutex;
923
924 struct list_head pmu_ctx_list;
925 struct perf_event_groups pinned_groups;
926 struct perf_event_groups flexible_groups;
927 struct list_head event_list;
928
929 int nr_events;
930 int nr_user;
931 int is_active;
932
933 int nr_task_data;
934 int nr_stat;
935 int nr_freq;
936 int rotate_disable;
937
938 refcount_t refcount; /* event <-> ctx */
939 struct task_struct *task;
940
941 /*
942 * Context clock, runs when context enabled.
943 */
944 u64 time;
945 u64 timestamp;
946 u64 timeoffset;
947
948 /*
949 * These fields let us detect when two contexts have both
950 * been cloned (inherited) from a common ancestor.
951 */
952 struct perf_event_context *parent_ctx;
953 u64 parent_gen;
954 u64 generation;
955 int pin_count;
956#ifdef CONFIG_CGROUP_PERF
957 int nr_cgroups; /* cgroup evts */
958#endif
959 struct rcu_head rcu_head;
960
961 /*
962 * Sum (event->pending_sigtrap + event->pending_work)
963 *
964 * The SIGTRAP is targeted at ctx->task, as such it won't do changing
965 * that until the signal is delivered.
966 */
967 local_t nr_pending;
968};
969
970/*
971 * Number of contexts where an event can trigger:
972 * task, softirq, hardirq, nmi.
973 */
974#define PERF_NR_CONTEXTS 4
975
976struct perf_cpu_pmu_context {
977 struct perf_event_pmu_context epc;
978 struct perf_event_pmu_context *task_epc;
979
980 struct list_head sched_cb_entry;
981 int sched_cb_usage;
982
983 int active_oncpu;
984 int exclusive;
985
986 raw_spinlock_t hrtimer_lock;
987 struct hrtimer hrtimer;
988 ktime_t hrtimer_interval;
989 unsigned int hrtimer_active;
990};
991
992/**
993 * struct perf_event_cpu_context - per cpu event context structure
994 */
995struct perf_cpu_context {
996 struct perf_event_context ctx;
997 struct perf_event_context *task_ctx;
998 int online;
999
1000#ifdef CONFIG_CGROUP_PERF
1001 struct perf_cgroup *cgrp;
1002#endif
1003
1004 /*
1005 * Per-CPU storage for iterators used in visit_groups_merge. The default
1006 * storage is of size 2 to hold the CPU and any CPU event iterators.
1007 */
1008 int heap_size;
1009 struct perf_event **heap;
1010 struct perf_event *heap_default[2];
1011};
1012
1013struct perf_output_handle {
1014 struct perf_event *event;
1015 struct perf_buffer *rb;
1016 unsigned long wakeup;
1017 unsigned long size;
1018 u64 aux_flags;
1019 union {
1020 void *addr;
1021 unsigned long head;
1022 };
1023 int page;
1024};
1025
1026struct bpf_perf_event_data_kern {
1027 bpf_user_pt_regs_t *regs;
1028 struct perf_sample_data *data;
1029 struct perf_event *event;
1030};
1031
1032#ifdef CONFIG_CGROUP_PERF
1033
1034/*
1035 * perf_cgroup_info keeps track of time_enabled for a cgroup.
1036 * This is a per-cpu dynamically allocated data structure.
1037 */
1038struct perf_cgroup_info {
1039 u64 time;
1040 u64 timestamp;
1041 u64 timeoffset;
1042 int active;
1043};
1044
1045struct perf_cgroup {
1046 struct cgroup_subsys_state css;
1047 struct perf_cgroup_info __percpu *info;
1048};
1049
1050/*
1051 * Must ensure cgroup is pinned (css_get) before calling
1052 * this function. In other words, we cannot call this function
1053 * if there is no cgroup event for the current CPU context.
1054 */
1055static inline struct perf_cgroup *
1056perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
1057{
1058 return container_of(task_css_check(task, perf_event_cgrp_id,
1059 ctx ? lockdep_is_held(&ctx->lock)
1060 : true),
1061 struct perf_cgroup, css);
1062}
1063#endif /* CONFIG_CGROUP_PERF */
1064
1065#ifdef CONFIG_PERF_EVENTS
1066
1067extern struct perf_event_context *perf_cpu_task_ctx(void);
1068
1069extern void *perf_aux_output_begin(struct perf_output_handle *handle,
1070 struct perf_event *event);
1071extern void perf_aux_output_end(struct perf_output_handle *handle,
1072 unsigned long size);
1073extern int perf_aux_output_skip(struct perf_output_handle *handle,
1074 unsigned long size);
1075extern void *perf_get_aux(struct perf_output_handle *handle);
1076extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
1077extern void perf_event_itrace_started(struct perf_event *event);
1078
1079extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
1080extern void perf_pmu_unregister(struct pmu *pmu);
1081
1082extern void __perf_event_task_sched_in(struct task_struct *prev,
1083 struct task_struct *task);
1084extern void __perf_event_task_sched_out(struct task_struct *prev,
1085 struct task_struct *next);
1086extern int perf_event_init_task(struct task_struct *child, u64 clone_flags);
1087extern void perf_event_exit_task(struct task_struct *child);
1088extern void perf_event_free_task(struct task_struct *task);
1089extern void perf_event_delayed_put(struct task_struct *task);
1090extern struct file *perf_event_get(unsigned int fd);
1091extern const struct perf_event *perf_get_event(struct file *file);
1092extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
1093extern void perf_event_print_debug(void);
1094extern void perf_pmu_disable(struct pmu *pmu);
1095extern void perf_pmu_enable(struct pmu *pmu);
1096extern void perf_sched_cb_dec(struct pmu *pmu);
1097extern void perf_sched_cb_inc(struct pmu *pmu);
1098extern int perf_event_task_disable(void);
1099extern int perf_event_task_enable(void);
1100
1101extern void perf_pmu_resched(struct pmu *pmu);
1102
1103extern int perf_event_refresh(struct perf_event *event, int refresh);
1104extern void perf_event_update_userpage(struct perf_event *event);
1105extern int perf_event_release_kernel(struct perf_event *event);
1106extern struct perf_event *
1107perf_event_create_kernel_counter(struct perf_event_attr *attr,
1108 int cpu,
1109 struct task_struct *task,
1110 perf_overflow_handler_t callback,
1111 void *context);
1112extern void perf_pmu_migrate_context(struct pmu *pmu,
1113 int src_cpu, int dst_cpu);
1114int perf_event_read_local(struct perf_event *event, u64 *value,
1115 u64 *enabled, u64 *running);
1116extern u64 perf_event_read_value(struct perf_event *event,
1117 u64 *enabled, u64 *running);
1118
1119extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1120
1121static inline bool branch_sample_no_flags(const struct perf_event *event)
1122{
1123 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_FLAGS;
1124}
1125
1126static inline bool branch_sample_no_cycles(const struct perf_event *event)
1127{
1128 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_CYCLES;
1129}
1130
1131static inline bool branch_sample_type(const struct perf_event *event)
1132{
1133 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_TYPE_SAVE;
1134}
1135
1136static inline bool branch_sample_hw_index(const struct perf_event *event)
1137{
1138 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX;
1139}
1140
1141static inline bool branch_sample_priv(const struct perf_event *event)
1142{
1143 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_PRIV_SAVE;
1144}
1145
1146static inline bool branch_sample_counters(const struct perf_event *event)
1147{
1148 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS;
1149}
1150
1151static inline bool branch_sample_call_stack(const struct perf_event *event)
1152{
1153 return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_CALL_STACK;
1154}
1155
1156struct perf_sample_data {
1157 /*
1158 * Fields set by perf_sample_data_init() unconditionally,
1159 * group so as to minimize the cachelines touched.
1160 */
1161 u64 sample_flags;
1162 u64 period;
1163 u64 dyn_size;
1164
1165 /*
1166 * Fields commonly set by __perf_event_header__init_id(),
1167 * group so as to minimize the cachelines touched.
1168 */
1169 u64 type;
1170 struct {
1171 u32 pid;
1172 u32 tid;
1173 } tid_entry;
1174 u64 time;
1175 u64 id;
1176 struct {
1177 u32 cpu;
1178 u32 reserved;
1179 } cpu_entry;
1180
1181 /*
1182 * The other fields, optionally {set,used} by
1183 * perf_{prepare,output}_sample().
1184 */
1185 u64 ip;
1186 struct perf_callchain_entry *callchain;
1187 struct perf_raw_record *raw;
1188 struct perf_branch_stack *br_stack;
1189 u64 *br_stack_cntr;
1190 union perf_sample_weight weight;
1191 union perf_mem_data_src data_src;
1192 u64 txn;
1193
1194 struct perf_regs regs_user;
1195 struct perf_regs regs_intr;
1196 u64 stack_user_size;
1197
1198 u64 stream_id;
1199 u64 cgroup;
1200 u64 addr;
1201 u64 phys_addr;
1202 u64 data_page_size;
1203 u64 code_page_size;
1204 u64 aux_size;
1205} ____cacheline_aligned;
1206
1207/* default value for data source */
1208#define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
1209 PERF_MEM_S(LVL, NA) |\
1210 PERF_MEM_S(SNOOP, NA) |\
1211 PERF_MEM_S(LOCK, NA) |\
1212 PERF_MEM_S(TLB, NA) |\
1213 PERF_MEM_S(LVLNUM, NA))
1214
1215static inline void perf_sample_data_init(struct perf_sample_data *data,
1216 u64 addr, u64 period)
1217{
1218 /* remaining struct members initialized in perf_prepare_sample() */
1219 data->sample_flags = PERF_SAMPLE_PERIOD;
1220 data->period = period;
1221 data->dyn_size = 0;
1222
1223 if (addr) {
1224 data->addr = addr;
1225 data->sample_flags |= PERF_SAMPLE_ADDR;
1226 }
1227}
1228
1229static inline void perf_sample_save_callchain(struct perf_sample_data *data,
1230 struct perf_event *event,
1231 struct pt_regs *regs)
1232{
1233 int size = 1;
1234
1235 data->callchain = perf_callchain(event, regs);
1236 size += data->callchain->nr;
1237
1238 data->dyn_size += size * sizeof(u64);
1239 data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
1240}
1241
1242static inline void perf_sample_save_raw_data(struct perf_sample_data *data,
1243 struct perf_raw_record *raw)
1244{
1245 struct perf_raw_frag *frag = &raw->frag;
1246 u32 sum = 0;
1247 int size;
1248
1249 do {
1250 sum += frag->size;
1251 if (perf_raw_frag_last(frag))
1252 break;
1253 frag = frag->next;
1254 } while (1);
1255
1256 size = round_up(sum + sizeof(u32), sizeof(u64));
1257 raw->size = size - sizeof(u32);
1258 frag->pad = raw->size - sum;
1259
1260 data->raw = raw;
1261 data->dyn_size += size;
1262 data->sample_flags |= PERF_SAMPLE_RAW;
1263}
1264
1265static inline void perf_sample_save_brstack(struct perf_sample_data *data,
1266 struct perf_event *event,
1267 struct perf_branch_stack *brs,
1268 u64 *brs_cntr)
1269{
1270 int size = sizeof(u64); /* nr */
1271
1272 if (branch_sample_hw_index(event))
1273 size += sizeof(u64);
1274 size += brs->nr * sizeof(struct perf_branch_entry);
1275
1276 /*
1277 * The extension space for counters is appended after the
1278 * struct perf_branch_stack. It is used to store the occurrences
1279 * of events of each branch.
1280 */
1281 if (brs_cntr)
1282 size += brs->nr * sizeof(u64);
1283
1284 data->br_stack = brs;
1285 data->br_stack_cntr = brs_cntr;
1286 data->dyn_size += size;
1287 data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
1288}
1289
1290static inline u32 perf_sample_data_size(struct perf_sample_data *data,
1291 struct perf_event *event)
1292{
1293 u32 size = sizeof(struct perf_event_header);
1294
1295 size += event->header_size + event->id_header_size;
1296 size += data->dyn_size;
1297
1298 return size;
1299}
1300
1301/*
1302 * Clear all bitfields in the perf_branch_entry.
1303 * The to and from fields are not cleared because they are
1304 * systematically modified by caller.
1305 */
1306static inline void perf_clear_branch_entry_bitfields(struct perf_branch_entry *br)
1307{
1308 br->mispred = 0;
1309 br->predicted = 0;
1310 br->in_tx = 0;
1311 br->abort = 0;
1312 br->cycles = 0;
1313 br->type = 0;
1314 br->spec = PERF_BR_SPEC_NA;
1315 br->reserved = 0;
1316}
1317
1318extern void perf_output_sample(struct perf_output_handle *handle,
1319 struct perf_event_header *header,
1320 struct perf_sample_data *data,
1321 struct perf_event *event);
1322extern void perf_prepare_sample(struct perf_sample_data *data,
1323 struct perf_event *event,
1324 struct pt_regs *regs);
1325extern void perf_prepare_header(struct perf_event_header *header,
1326 struct perf_sample_data *data,
1327 struct perf_event *event,
1328 struct pt_regs *regs);
1329
1330extern int perf_event_overflow(struct perf_event *event,
1331 struct perf_sample_data *data,
1332 struct pt_regs *regs);
1333
1334extern void perf_event_output_forward(struct perf_event *event,
1335 struct perf_sample_data *data,
1336 struct pt_regs *regs);
1337extern void perf_event_output_backward(struct perf_event *event,
1338 struct perf_sample_data *data,
1339 struct pt_regs *regs);
1340extern int perf_event_output(struct perf_event *event,
1341 struct perf_sample_data *data,
1342 struct pt_regs *regs);
1343
1344static inline bool
1345__is_default_overflow_handler(perf_overflow_handler_t overflow_handler)
1346{
1347 if (likely(overflow_handler == perf_event_output_forward))
1348 return true;
1349 if (unlikely(overflow_handler == perf_event_output_backward))
1350 return true;
1351 return false;
1352}
1353
1354#define is_default_overflow_handler(event) \
1355 __is_default_overflow_handler((event)->overflow_handler)
1356
1357#ifdef CONFIG_BPF_SYSCALL
1358static inline bool uses_default_overflow_handler(struct perf_event *event)
1359{
1360 if (likely(is_default_overflow_handler(event)))
1361 return true;
1362
1363 return __is_default_overflow_handler(overflow_handler: event->orig_overflow_handler);
1364}
1365#else
1366#define uses_default_overflow_handler(event) \
1367 is_default_overflow_handler(event)
1368#endif
1369
1370extern void
1371perf_event_header__init_id(struct perf_event_header *header,
1372 struct perf_sample_data *data,
1373 struct perf_event *event);
1374extern void
1375perf_event__output_id_sample(struct perf_event *event,
1376 struct perf_output_handle *handle,
1377 struct perf_sample_data *sample);
1378
1379extern void
1380perf_log_lost_samples(struct perf_event *event, u64 lost);
1381
1382static inline bool event_has_any_exclude_flag(struct perf_event *event)
1383{
1384 struct perf_event_attr *attr = &event->attr;
1385
1386 return attr->exclude_idle || attr->exclude_user ||
1387 attr->exclude_kernel || attr->exclude_hv ||
1388 attr->exclude_guest || attr->exclude_host;
1389}
1390
1391static inline bool is_sampling_event(struct perf_event *event)
1392{
1393 return event->attr.sample_period != 0;
1394}
1395
1396/*
1397 * Return 1 for a software event, 0 for a hardware event
1398 */
1399static inline int is_software_event(struct perf_event *event)
1400{
1401 return event->event_caps & PERF_EV_CAP_SOFTWARE;
1402}
1403
1404/*
1405 * Return 1 for event in sw context, 0 for event in hw context
1406 */
1407static inline int in_software_context(struct perf_event *event)
1408{
1409 return event->pmu_ctx->pmu->task_ctx_nr == perf_sw_context;
1410}
1411
1412static inline int is_exclusive_pmu(struct pmu *pmu)
1413{
1414 return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1415}
1416
1417extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1418
1419extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1420extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1421
1422#ifndef perf_arch_fetch_caller_regs
1423static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1424#endif
1425
1426/*
1427 * When generating a perf sample in-line, instead of from an interrupt /
1428 * exception, we lack a pt_regs. This is typically used from software events
1429 * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1430 *
1431 * We typically don't need a full set, but (for x86) do require:
1432 * - ip for PERF_SAMPLE_IP
1433 * - cs for user_mode() tests
1434 * - sp for PERF_SAMPLE_CALLCHAIN
1435 * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1436 *
1437 * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1438 * things like PERF_SAMPLE_REGS_INTR.
1439 */
1440static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1441{
1442 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1443}
1444
1445static __always_inline void
1446perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1447{
1448 if (static_key_false(key: &perf_swevent_enabled[event_id]))
1449 __perf_sw_event(event_id, nr, regs, addr);
1450}
1451
1452DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1453
1454/*
1455 * 'Special' version for the scheduler, it hard assumes no recursion,
1456 * which is guaranteed by us not actually scheduling inside other swevents
1457 * because those disable preemption.
1458 */
1459static __always_inline void __perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1460{
1461 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1462
1463 perf_fetch_caller_regs(regs);
1464 ___perf_sw_event(event_id, nr, regs, addr);
1465}
1466
1467extern struct static_key_false perf_sched_events;
1468
1469static __always_inline bool __perf_sw_enabled(int swevt)
1470{
1471 return static_key_false(key: &perf_swevent_enabled[swevt]);
1472}
1473
1474static inline void perf_event_task_migrate(struct task_struct *task)
1475{
1476 if (__perf_sw_enabled(swevt: PERF_COUNT_SW_CPU_MIGRATIONS))
1477 task->sched_migrated = 1;
1478}
1479
1480static inline void perf_event_task_sched_in(struct task_struct *prev,
1481 struct task_struct *task)
1482{
1483 if (static_branch_unlikely(&perf_sched_events))
1484 __perf_event_task_sched_in(prev, task);
1485
1486 if (__perf_sw_enabled(swevt: PERF_COUNT_SW_CPU_MIGRATIONS) &&
1487 task->sched_migrated) {
1488 __perf_sw_event_sched(event_id: PERF_COUNT_SW_CPU_MIGRATIONS, nr: 1, addr: 0);
1489 task->sched_migrated = 0;
1490 }
1491}
1492
1493static inline void perf_event_task_sched_out(struct task_struct *prev,
1494 struct task_struct *next)
1495{
1496 if (__perf_sw_enabled(swevt: PERF_COUNT_SW_CONTEXT_SWITCHES))
1497 __perf_sw_event_sched(event_id: PERF_COUNT_SW_CONTEXT_SWITCHES, nr: 1, addr: 0);
1498
1499#ifdef CONFIG_CGROUP_PERF
1500 if (__perf_sw_enabled(swevt: PERF_COUNT_SW_CGROUP_SWITCHES) &&
1501 perf_cgroup_from_task(task: prev, NULL) !=
1502 perf_cgroup_from_task(task: next, NULL))
1503 __perf_sw_event_sched(event_id: PERF_COUNT_SW_CGROUP_SWITCHES, nr: 1, addr: 0);
1504#endif
1505
1506 if (static_branch_unlikely(&perf_sched_events))
1507 __perf_event_task_sched_out(prev, next);
1508}
1509
1510extern void perf_event_mmap(struct vm_area_struct *vma);
1511
1512extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1513 bool unregister, const char *sym);
1514extern void perf_event_bpf_event(struct bpf_prog *prog,
1515 enum perf_bpf_event_type type,
1516 u16 flags);
1517
1518#ifdef CONFIG_GUEST_PERF_EVENTS
1519extern struct perf_guest_info_callbacks __rcu *perf_guest_cbs;
1520
1521DECLARE_STATIC_CALL(__perf_guest_state, *perf_guest_cbs->state);
1522DECLARE_STATIC_CALL(__perf_guest_get_ip, *perf_guest_cbs->get_ip);
1523DECLARE_STATIC_CALL(__perf_guest_handle_intel_pt_intr, *perf_guest_cbs->handle_intel_pt_intr);
1524
1525static inline unsigned int perf_guest_state(void)
1526{
1527 return static_call(__perf_guest_state)();
1528}
1529static inline unsigned long perf_guest_get_ip(void)
1530{
1531 return static_call(__perf_guest_get_ip)();
1532}
1533static inline unsigned int perf_guest_handle_intel_pt_intr(void)
1534{
1535 return static_call(__perf_guest_handle_intel_pt_intr)();
1536}
1537extern void perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1538extern void perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1539#else
1540static inline unsigned int perf_guest_state(void) { return 0; }
1541static inline unsigned long perf_guest_get_ip(void) { return 0; }
1542static inline unsigned int perf_guest_handle_intel_pt_intr(void) { return 0; }
1543#endif /* CONFIG_GUEST_PERF_EVENTS */
1544
1545extern void perf_event_exec(void);
1546extern void perf_event_comm(struct task_struct *tsk, bool exec);
1547extern void perf_event_namespaces(struct task_struct *tsk);
1548extern void perf_event_fork(struct task_struct *tsk);
1549extern void perf_event_text_poke(const void *addr,
1550 const void *old_bytes, size_t old_len,
1551 const void *new_bytes, size_t new_len);
1552
1553/* Callchains */
1554DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1555
1556extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1557extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1558extern struct perf_callchain_entry *
1559get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1560 u32 max_stack, bool crosstask, bool add_mark);
1561extern int get_callchain_buffers(int max_stack);
1562extern void put_callchain_buffers(void);
1563extern struct perf_callchain_entry *get_callchain_entry(int *rctx);
1564extern void put_callchain_entry(int rctx);
1565
1566extern int sysctl_perf_event_max_stack;
1567extern int sysctl_perf_event_max_contexts_per_stack;
1568
1569static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1570{
1571 if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1572 struct perf_callchain_entry *entry = ctx->entry;
1573 entry->ip[entry->nr++] = ip;
1574 ++ctx->contexts;
1575 return 0;
1576 } else {
1577 ctx->contexts_maxed = true;
1578 return -1; /* no more room, stop walking the stack */
1579 }
1580}
1581
1582static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1583{
1584 if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1585 struct perf_callchain_entry *entry = ctx->entry;
1586 entry->ip[entry->nr++] = ip;
1587 ++ctx->nr;
1588 return 0;
1589 } else {
1590 return -1; /* no more room, stop walking the stack */
1591 }
1592}
1593
1594extern int sysctl_perf_event_paranoid;
1595extern int sysctl_perf_event_mlock;
1596extern int sysctl_perf_event_sample_rate;
1597extern int sysctl_perf_cpu_time_max_percent;
1598
1599extern void perf_sample_event_took(u64 sample_len_ns);
1600
1601int perf_event_max_sample_rate_handler(struct ctl_table *table, int write,
1602 void *buffer, size_t *lenp, loff_t *ppos);
1603int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1604 void *buffer, size_t *lenp, loff_t *ppos);
1605int perf_event_max_stack_handler(struct ctl_table *table, int write,
1606 void *buffer, size_t *lenp, loff_t *ppos);
1607
1608/* Access to perf_event_open(2) syscall. */
1609#define PERF_SECURITY_OPEN 0
1610
1611/* Finer grained perf_event_open(2) access control. */
1612#define PERF_SECURITY_CPU 1
1613#define PERF_SECURITY_KERNEL 2
1614#define PERF_SECURITY_TRACEPOINT 3
1615
1616static inline int perf_is_paranoid(void)
1617{
1618 return sysctl_perf_event_paranoid > -1;
1619}
1620
1621static inline int perf_allow_kernel(struct perf_event_attr *attr)
1622{
1623 if (sysctl_perf_event_paranoid > 1 && !perfmon_capable())
1624 return -EACCES;
1625
1626 return security_perf_event_open(attr, PERF_SECURITY_KERNEL);
1627}
1628
1629static inline int perf_allow_cpu(struct perf_event_attr *attr)
1630{
1631 if (sysctl_perf_event_paranoid > 0 && !perfmon_capable())
1632 return -EACCES;
1633
1634 return security_perf_event_open(attr, PERF_SECURITY_CPU);
1635}
1636
1637static inline int perf_allow_tracepoint(struct perf_event_attr *attr)
1638{
1639 if (sysctl_perf_event_paranoid > -1 && !perfmon_capable())
1640 return -EPERM;
1641
1642 return security_perf_event_open(attr, PERF_SECURITY_TRACEPOINT);
1643}
1644
1645extern void perf_event_init(void);
1646extern void perf_tp_event(u16 event_type, u64 count, void *record,
1647 int entry_size, struct pt_regs *regs,
1648 struct hlist_head *head, int rctx,
1649 struct task_struct *task);
1650extern void perf_bp_event(struct perf_event *event, void *data);
1651
1652#ifndef perf_misc_flags
1653# define perf_misc_flags(regs) \
1654 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1655# define perf_instruction_pointer(regs) instruction_pointer(regs)
1656#endif
1657#ifndef perf_arch_bpf_user_pt_regs
1658# define perf_arch_bpf_user_pt_regs(regs) regs
1659#endif
1660
1661static inline bool has_branch_stack(struct perf_event *event)
1662{
1663 return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1664}
1665
1666static inline bool needs_branch_stack(struct perf_event *event)
1667{
1668 return event->attr.branch_sample_type != 0;
1669}
1670
1671static inline bool has_aux(struct perf_event *event)
1672{
1673 return event->pmu->setup_aux;
1674}
1675
1676static inline bool is_write_backward(struct perf_event *event)
1677{
1678 return !!event->attr.write_backward;
1679}
1680
1681static inline bool has_addr_filter(struct perf_event *event)
1682{
1683 return event->pmu->nr_addr_filters;
1684}
1685
1686/*
1687 * An inherited event uses parent's filters
1688 */
1689static inline struct perf_addr_filters_head *
1690perf_event_addr_filters(struct perf_event *event)
1691{
1692 struct perf_addr_filters_head *ifh = &event->addr_filters;
1693
1694 if (event->parent)
1695 ifh = &event->parent->addr_filters;
1696
1697 return ifh;
1698}
1699
1700extern void perf_event_addr_filters_sync(struct perf_event *event);
1701extern void perf_report_aux_output_id(struct perf_event *event, u64 hw_id);
1702
1703extern int perf_output_begin(struct perf_output_handle *handle,
1704 struct perf_sample_data *data,
1705 struct perf_event *event, unsigned int size);
1706extern int perf_output_begin_forward(struct perf_output_handle *handle,
1707 struct perf_sample_data *data,
1708 struct perf_event *event,
1709 unsigned int size);
1710extern int perf_output_begin_backward(struct perf_output_handle *handle,
1711 struct perf_sample_data *data,
1712 struct perf_event *event,
1713 unsigned int size);
1714
1715extern void perf_output_end(struct perf_output_handle *handle);
1716extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1717 const void *buf, unsigned int len);
1718extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1719 unsigned int len);
1720extern long perf_output_copy_aux(struct perf_output_handle *aux_handle,
1721 struct perf_output_handle *handle,
1722 unsigned long from, unsigned long to);
1723extern int perf_swevent_get_recursion_context(void);
1724extern void perf_swevent_put_recursion_context(int rctx);
1725extern u64 perf_swevent_set_period(struct perf_event *event);
1726extern void perf_event_enable(struct perf_event *event);
1727extern void perf_event_disable(struct perf_event *event);
1728extern void perf_event_disable_local(struct perf_event *event);
1729extern void perf_event_disable_inatomic(struct perf_event *event);
1730extern void perf_event_task_tick(void);
1731extern int perf_event_account_interrupt(struct perf_event *event);
1732extern int perf_event_period(struct perf_event *event, u64 value);
1733extern u64 perf_event_pause(struct perf_event *event, bool reset);
1734#else /* !CONFIG_PERF_EVENTS: */
1735static inline void *
1736perf_aux_output_begin(struct perf_output_handle *handle,
1737 struct perf_event *event) { return NULL; }
1738static inline void
1739perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1740 { }
1741static inline int
1742perf_aux_output_skip(struct perf_output_handle *handle,
1743 unsigned long size) { return -EINVAL; }
1744static inline void *
1745perf_get_aux(struct perf_output_handle *handle) { return NULL; }
1746static inline void
1747perf_event_task_migrate(struct task_struct *task) { }
1748static inline void
1749perf_event_task_sched_in(struct task_struct *prev,
1750 struct task_struct *task) { }
1751static inline void
1752perf_event_task_sched_out(struct task_struct *prev,
1753 struct task_struct *next) { }
1754static inline int perf_event_init_task(struct task_struct *child,
1755 u64 clone_flags) { return 0; }
1756static inline void perf_event_exit_task(struct task_struct *child) { }
1757static inline void perf_event_free_task(struct task_struct *task) { }
1758static inline void perf_event_delayed_put(struct task_struct *task) { }
1759static inline struct file *perf_event_get(unsigned int fd) { return ERR_PTR(-EINVAL); }
1760static inline const struct perf_event *perf_get_event(struct file *file)
1761{
1762 return ERR_PTR(-EINVAL);
1763}
1764static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1765{
1766 return ERR_PTR(-EINVAL);
1767}
1768static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1769 u64 *enabled, u64 *running)
1770{
1771 return -EINVAL;
1772}
1773static inline void perf_event_print_debug(void) { }
1774static inline int perf_event_task_disable(void) { return -EINVAL; }
1775static inline int perf_event_task_enable(void) { return -EINVAL; }
1776static inline int perf_event_refresh(struct perf_event *event, int refresh)
1777{
1778 return -EINVAL;
1779}
1780
1781static inline void
1782perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { }
1783static inline void
1784perf_bp_event(struct perf_event *event, void *data) { }
1785
1786static inline void perf_event_mmap(struct vm_area_struct *vma) { }
1787
1788typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1789static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1790 bool unregister, const char *sym) { }
1791static inline void perf_event_bpf_event(struct bpf_prog *prog,
1792 enum perf_bpf_event_type type,
1793 u16 flags) { }
1794static inline void perf_event_exec(void) { }
1795static inline void perf_event_comm(struct task_struct *tsk, bool exec) { }
1796static inline void perf_event_namespaces(struct task_struct *tsk) { }
1797static inline void perf_event_fork(struct task_struct *tsk) { }
1798static inline void perf_event_text_poke(const void *addr,
1799 const void *old_bytes,
1800 size_t old_len,
1801 const void *new_bytes,
1802 size_t new_len) { }
1803static inline void perf_event_init(void) { }
1804static inline int perf_swevent_get_recursion_context(void) { return -1; }
1805static inline void perf_swevent_put_recursion_context(int rctx) { }
1806static inline u64 perf_swevent_set_period(struct perf_event *event) { return 0; }
1807static inline void perf_event_enable(struct perf_event *event) { }
1808static inline void perf_event_disable(struct perf_event *event) { }
1809static inline int __perf_event_disable(void *info) { return -1; }
1810static inline void perf_event_task_tick(void) { }
1811static inline int perf_event_release_kernel(struct perf_event *event) { return 0; }
1812static inline int perf_event_period(struct perf_event *event, u64 value)
1813{
1814 return -EINVAL;
1815}
1816static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1817{
1818 return 0;
1819}
1820#endif
1821
1822#if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1823extern void perf_restore_debug_store(void);
1824#else
1825static inline void perf_restore_debug_store(void) { }
1826#endif
1827
1828#define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1829
1830struct perf_pmu_events_attr {
1831 struct device_attribute attr;
1832 u64 id;
1833 const char *event_str;
1834};
1835
1836struct perf_pmu_events_ht_attr {
1837 struct device_attribute attr;
1838 u64 id;
1839 const char *event_str_ht;
1840 const char *event_str_noht;
1841};
1842
1843struct perf_pmu_events_hybrid_attr {
1844 struct device_attribute attr;
1845 u64 id;
1846 const char *event_str;
1847 u64 pmu_type;
1848};
1849
1850struct perf_pmu_format_hybrid_attr {
1851 struct device_attribute attr;
1852 u64 pmu_type;
1853};
1854
1855ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1856 char *page);
1857
1858#define PMU_EVENT_ATTR(_name, _var, _id, _show) \
1859static struct perf_pmu_events_attr _var = { \
1860 .attr = __ATTR(_name, 0444, _show, NULL), \
1861 .id = _id, \
1862};
1863
1864#define PMU_EVENT_ATTR_STRING(_name, _var, _str) \
1865static struct perf_pmu_events_attr _var = { \
1866 .attr = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1867 .id = 0, \
1868 .event_str = _str, \
1869};
1870
1871#define PMU_EVENT_ATTR_ID(_name, _show, _id) \
1872 (&((struct perf_pmu_events_attr[]) { \
1873 { .attr = __ATTR(_name, 0444, _show, NULL), \
1874 .id = _id, } \
1875 })[0].attr.attr)
1876
1877#define PMU_FORMAT_ATTR_SHOW(_name, _format) \
1878static ssize_t \
1879_name##_show(struct device *dev, \
1880 struct device_attribute *attr, \
1881 char *page) \
1882{ \
1883 BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE); \
1884 return sprintf(page, _format "\n"); \
1885} \
1886
1887#define PMU_FORMAT_ATTR(_name, _format) \
1888 PMU_FORMAT_ATTR_SHOW(_name, _format) \
1889 \
1890static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1891
1892/* Performance counter hotplug functions */
1893#ifdef CONFIG_PERF_EVENTS
1894int perf_event_init_cpu(unsigned int cpu);
1895int perf_event_exit_cpu(unsigned int cpu);
1896#else
1897#define perf_event_init_cpu NULL
1898#define perf_event_exit_cpu NULL
1899#endif
1900
1901extern void arch_perf_update_userpage(struct perf_event *event,
1902 struct perf_event_mmap_page *userpg,
1903 u64 now);
1904
1905/*
1906 * Snapshot branch stack on software events.
1907 *
1908 * Branch stack can be very useful in understanding software events. For
1909 * example, when a long function, e.g. sys_perf_event_open, returns an
1910 * errno, it is not obvious why the function failed. Branch stack could
1911 * provide very helpful information in this type of scenarios.
1912 *
1913 * On software event, it is necessary to stop the hardware branch recorder
1914 * fast. Otherwise, the hardware register/buffer will be flushed with
1915 * entries of the triggering event. Therefore, static call is used to
1916 * stop the hardware recorder.
1917 */
1918
1919/*
1920 * cnt is the number of entries allocated for entries.
1921 * Return number of entries copied to .
1922 */
1923typedef int (perf_snapshot_branch_stack_t)(struct perf_branch_entry *entries,
1924 unsigned int cnt);
1925DECLARE_STATIC_CALL(perf_snapshot_branch_stack, perf_snapshot_branch_stack_t);
1926
1927#ifndef PERF_NEEDS_LOPWR_CB
1928static inline void perf_lopwr_cb(bool mode)
1929{
1930}
1931#endif
1932
1933#endif /* _LINUX_PERF_EVENT_H */
1934

source code of linux/include/linux/perf_event.h