include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   7  *
   8  * Data type definitions, declarations, prototypes.
   9  *
  10  *    Started by: Thomas Gleixner and Ingo Molnar
  11  *
  12  * For licencing details see kernel-base/COPYING
  13  */
  14 #ifndef _LINUX_PERF_EVENT_H
  15 #define _LINUX_PERF_EVENT_H
  16
  17 #include <uapi/linux/perf_event.h>
  18
  19 /*
  20  * Kernel-internal data types and definitions:
  21  */
  22
  23 #ifdef CONFIG_PERF_EVENTS
  24 # include <asm/perf_event.h>
  25 # include <asm/local64.h>
  26 #endif
  27
  28 struct perf_guest_info_callbacks {
  29         int                             (*is_in_guest)(void);
  30         int                             (*is_user_mode)(void);
  31         unsigned long                   (*get_guest_ip)(void);
  32 };
  33
  34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
  35 #include <asm/hw_breakpoint.h>
  36 #endif
  37
  38 #include <linux/list.h>
  39 #include <linux/mutex.h>
  40 #include <linux/rculist.h>
  41 #include <linux/rcupdate.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/hrtimer.h>
  44 #include <linux/fs.h>
  45 #include <linux/pid_namespace.h>
  46 #include <linux/workqueue.h>
  47 #include <linux/ftrace.h>
  48 #include <linux/cpu.h>
  49 #include <linux/irq_work.h>
  50 #include <linux/static_key.h>
  51 #include <linux/jump_label_ratelimit.h>
  52 #include <linux/atomic.h>
  53 #include <linux/sysfs.h>
  54 #include <linux/perf_regs.h>
  55 #include <linux/workqueue.h>
  56 #include <asm/local.h>
  57
  58 struct perf_callchain_entry {
  59         __u64                           nr;
  60         __u64                           ip[PERF_MAX_STACK_DEPTH];
  61 };
  62
  63 struct perf_raw_record {
  64         u32                             size;
  65         void                            *data;
  66 };
  67
  68 /*
  69  * branch stack layout:
  70  *  nr: number of taken branches stored in entries[]
  71  *
  72  * Note that nr can vary from sample to sample
  73  * branches (to, from) are stored from most recent
  74  * to least recent, i.e., entries[0] contains the most
  75  * recent branch.
  76  */
  77 struct perf_branch_stack {
  78         __u64                           nr;
  79         struct perf_branch_entry        entries[0];
  80 };
  81
  82 struct perf_regs {
  83         __u64           abi;
  84         struct pt_regs  *regs;
  85 };
  86
  87 struct task_struct;
  88
  89 /*
  90  * extra PMU register associated with an event
  91  */
  92 struct hw_perf_event_extra {
  93         u64             config; /* register value */
  94         unsigned int    reg;    /* register address or index */
  95         int             alloc;  /* extra register already allocated */
  96         int             idx;    /* index in shared_regs->regs[] */
  97 };
  98
  99 struct event_constraint;
 100
 101 /**
 102  * struct hw_perf_event - performance event hardware details:
 103  */
 104 struct hw_perf_event {
 105 #ifdef CONFIG_PERF_EVENTS
 106         union {
 107                 struct { /* hardware */
 108                         u64             config;
 109                         u64             last_tag;
 110                         unsigned long   config_base;
 111                         unsigned long   event_base;
 112                         int             event_base_rdpmc;
 113                         int             idx;
 114                         int             last_cpu;
 115                         int             flags;
 116
 117                         struct hw_perf_event_extra extra_reg;
 118                         struct hw_perf_event_extra branch_reg;
 119
 120                         struct event_constraint *constraint;
 121                 };
 122                 struct { /* software */
 123                         struct hrtimer  hrtimer;
 124                 };
 125                 struct { /* tracepoint */
 126                         struct task_struct      *tp_target;
 127                         /* for tp_event->class */
 128                         struct list_head        tp_list;
 129                 };
 130 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 131                 struct { /* breakpoint */
 132                         /*
 133                          * Crufty hack to avoid the chicken and egg
 134                          * problem hw_breakpoint has with context
 135                          * creation and event initalization.
 136                          */
 137                         struct task_struct              *bp_target;
 138                         struct arch_hw_breakpoint       info;
 139                         struct list_head                bp_list;
 140                 };
 141 #endif
 142         };
 143         int                             state;
 144         local64_t                       prev_count;
 145         u64                             sample_period;
 146         u64                             last_period;
 147         local64_t                       period_left;
 148         u64                             interrupts_seq;
 149         u64                             interrupts;
 150
 151         u64                             freq_time_stamp;
 152         u64                             freq_count_stamp;
 153 #endif
 154 };
 155
 156 /*
 157  * hw_perf_event::state flags
 158  */
 159 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 160 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 161 #define PERF_HES_ARCH           0x04
 162
 163 struct perf_event;
 164
 165 /*
 166  * Common implementation detail of pmu::{start,commit,cancel}_txn
 167  */
 168 #define PERF_EVENT_TXN 0x1
 169
 170 /**
 171  * pmu::capabilities flags
 172  */
 173 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 174
 175 /**
 176  * struct pmu - generic performance monitoring unit
 177  */
 178 struct pmu {
 179         struct list_head                entry;
 180
 181         struct module                   *module;
 182         struct device                   *dev;
 183         const struct attribute_group    **attr_groups;
 184         const char                      *name;
 185         int                             type;
 186
 187         /*
 188          * various common per-pmu feature flags
 189          */
 190         int                             capabilities;
 191
 192         int * __percpu                  pmu_disable_count;
 193         struct perf_cpu_context * __percpu pmu_cpu_context;
 194         int                             task_ctx_nr;
 195         int                             hrtimer_interval_ms;
 196
 197         /*
 198          * Fully disable/enable this PMU, can be used to protect from the PMI
 199          * as well as for lazy/batch writing of the MSRs.
 200          */
 201         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 202         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 203
 204         /*
 205          * Try and initialize the event for this PMU.
 206          * Should return -ENOENT when the @event doesn't match this PMU.
 207          */
 208         int (*event_init)               (struct perf_event *event);
 209
 210 #define PERF_EF_START   0x01            /* start the counter when adding    */
 211 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 212 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 213
 214         /*
 215          * Adds/Removes a counter to/from the PMU, can be done inside
 216          * a transaction, see the ->*_txn() methods.
 217          */
 218         int  (*add)                     (struct perf_event *event, int flags);
 219         void (*del)                     (struct perf_event *event, int flags);
 220
 221         /*
 222          * Starts/Stops a counter present on the PMU. The PMI handler
 223          * should stop the counter when perf_event_overflow() returns
 224          * !0. ->start() will be used to continue.
 225          */
 226         void (*start)                   (struct perf_event *event, int flags);
 227         void (*stop)                    (struct perf_event *event, int flags);
 228
 229         /*
 230          * Updates the counter value of the event.
 231          */
 232         void (*read)                    (struct perf_event *event);
 233
 234         /*
 235          * Group events scheduling is treated as a transaction, add
 236          * group events as a whole and perform one schedulability test.
 237          * If the test fails, roll back the whole group
 238          *
 239          * Start the transaction, after this ->add() doesn't need to
 240          * do schedulability tests.
 241          */
 242         void (*start_txn)               (struct pmu *pmu); /* optional */
 243         /*
 244          * If ->start_txn() disabled the ->add() schedulability test
 245          * then ->commit_txn() is required to perform one. On success
 246          * the transaction is closed. On error the transaction is kept
 247          * open until ->cancel_txn() is called.
 248          */
 249         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 250         /*
 251          * Will cancel the transaction, assumes ->del() is called
 252          * for each successful ->add() during the transaction.
 253          */
 254         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 255
 256         /*
 257          * Will return the value for perf_event_mmap_page::index for this event,
 258          * if no implementation is provided it will default to: event->hw.idx + 1.
 259          */
 260         int (*event_idx)                (struct perf_event *event); /*optional */
 261
 262         /*
 263          * flush branch stack on context-switches (needed in cpu-wide mode)
 264          */
 265         void (*flush_branch_stack)      (void);
 266 };
 267
 268 /**
 269  * enum perf_event_active_state - the states of a event
 270  */
 271 enum perf_event_active_state {
 272         PERF_EVENT_STATE_EXIT           = -3,
 273         PERF_EVENT_STATE_ERROR          = -2,
 274         PERF_EVENT_STATE_OFF            = -1,
 275         PERF_EVENT_STATE_INACTIVE       =  0,
 276         PERF_EVENT_STATE_ACTIVE         =  1,
 277 };
 278
 279 struct file;
 280 struct perf_sample_data;
 281
 282 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 283                                         struct perf_sample_data *,
 284                                         struct pt_regs *regs);
 285
 286 enum perf_group_flag {
 287         PERF_GROUP_SOFTWARE             = 0x1,
 288 };
 289
 290 #define SWEVENT_HLIST_BITS              8
 291 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 292
 293 struct swevent_hlist {
 294         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 295         struct rcu_head                 rcu_head;
 296 };
 297
 298 #define PERF_ATTACH_CONTEXT     0x01
 299 #define PERF_ATTACH_GROUP       0x02
 300 #define PERF_ATTACH_TASK        0x04
 301
 302 struct perf_cgroup;
 303 struct ring_buffer;
 304
 305 /**
 306  * struct perf_event - performance event kernel representation:
 307  */
 308 struct perf_event {
 309 #ifdef CONFIG_PERF_EVENTS
 310         /*
 311          * entry onto perf_event_context::event_list;
 312          *   modifications require ctx->lock
 313          *   RCU safe iterations.
 314          */
 315         struct list_head                event_entry;
 316
 317         /*
 318          * XXX: group_entry and sibling_list should be mutually exclusive;
 319          * either you're a sibling on a group, or you're the group leader.
 320          * Rework the code to always use the same list element.
 321          *
 322          * Locked for modification by both ctx->mutex and ctx->lock; holding
 323          * either sufficies for read.
 324          */
 325         struct list_head                group_entry;
 326         struct list_head                sibling_list;
 327
 328         /*
 329          * We need storage to track the entries in perf_pmu_migrate_context; we
 330          * cannot use the event_entry because of RCU and we want to keep the
 331          * group in tact which avoids us using the other two entries.
 332          */
 333         struct list_head                migrate_entry;
 334
 335         struct hlist_node               hlist_entry;
 336         struct list_head                active_entry;
 337         int                             nr_siblings;
 338         int                             group_flags;
 339         struct perf_event               *group_leader;
 340         struct pmu                      *pmu;
 341
 342         enum perf_event_active_state    state;
 343         unsigned int                    attach_state;
 344         local64_t                       count;
 345         atomic64_t                      child_count;
 346
 347         /*
 348          * These are the total time in nanoseconds that the event
 349          * has been enabled (i.e. eligible to run, and the task has
 350          * been scheduled in, if this is a per-task event)
 351          * and running (scheduled onto the CPU), respectively.
 352          *
 353          * They are computed from tstamp_enabled, tstamp_running and
 354          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 355          */
 356         u64                             total_time_enabled;
 357         u64                             total_time_running;
 358
 359         /*
 360          * These are timestamps used for computing total_time_enabled
 361          * and total_time_running when the event is in INACTIVE or
 362          * ACTIVE state, measured in nanoseconds from an arbitrary point
 363          * in time.
 364          * tstamp_enabled: the notional time when the event was enabled
 365          * tstamp_running: the notional time when the event was scheduled on
 366          * tstamp_stopped: in INACTIVE state, the notional time when the
 367          *      event was scheduled off.
 368          */
 369         u64                             tstamp_enabled;
 370         u64                             tstamp_running;
 371         u64                             tstamp_stopped;
 372
 373         /*
 374          * timestamp shadows the actual context timing but it can
 375          * be safely used in NMI interrupt context. It reflects the
 376          * context time as it was when the event was last scheduled in.
 377          *
 378          * ctx_time already accounts for ctx->timestamp. Therefore to
 379          * compute ctx_time for a sample, simply add perf_clock().
 380          */
 381         u64                             shadow_ctx_time;
 382
 383         struct perf_event_attr          attr;
 384         u16                             header_size;
 385         u16                             id_header_size;
 386         u16                             read_size;
 387         struct hw_perf_event            hw;
 388
 389         struct perf_event_context       *ctx;
 390         atomic_long_t                   refcount;
 391
 392         /*
 393          * These accumulate total time (in nanoseconds) that children
 394          * events have been enabled and running, respectively.
 395          */
 396         atomic64_t                      child_total_time_enabled;
 397         atomic64_t                      child_total_time_running;
 398
 399         /*
 400          * Protect attach/detach and child_list:
 401          */
 402         struct mutex                    child_mutex;
 403         struct list_head                child_list;
 404         struct perf_event               *parent;
 405
 406         int                             oncpu;
 407         int                             cpu;
 408
 409         struct list_head                owner_entry;
 410         struct task_struct              *owner;
 411
 412         /* mmap bits */
 413         struct mutex                    mmap_mutex;
 414         atomic_t                        mmap_count;
 415
 416         struct ring_buffer              *rb;
 417         struct list_head                rb_entry;
 418         unsigned long                   rcu_batches;
 419         int                             rcu_pending;
 420
 421         /* poll related */
 422         wait_queue_head_t               waitq;
 423         struct fasync_struct            *fasync;
 424
 425         /* delayed work for NMIs and such */
 426         int                             pending_wakeup;
 427         int                             pending_kill;
 428         int                             pending_disable;
 429         struct irq_work                 pending;
 430
 431         atomic_t                        event_limit;
 432
 433         void (*destroy)(struct perf_event *);
 434         struct rcu_head                 rcu_head;
 435
 436         struct pid_namespace            *ns;
 437         u64                             id;
 438
 439         perf_overflow_handler_t         overflow_handler;
 440         void                            *overflow_handler_context;
 441
 442 #ifdef CONFIG_EVENT_TRACING
 443         struct ftrace_event_call        *tp_event;
 444         struct event_filter             *filter;
 445 #ifdef CONFIG_FUNCTION_TRACER
 446         struct ftrace_ops               ftrace_ops;
 447 #endif
 448 #endif
 449
 450 #ifdef CONFIG_CGROUP_PERF
 451         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 452         int                             cgrp_defer_enabled;
 453 #endif
 454
 455 #endif /* CONFIG_PERF_EVENTS */
 456 };
 457
 458 enum perf_event_context_type {
 459         task_context,
 460         cpu_context,
 461 };
 462
 463 /**
 464  * struct perf_event_context - event context structure
 465  *
 466  * Used as a container for task events and CPU events as well:
 467  */
 468 struct perf_event_context {
 469         struct pmu                      *pmu;
 470         enum perf_event_context_type    type;
 471         /*
 472          * Protect the states of the events in the list,
 473          * nr_active, and the list:
 474          */
 475         raw_spinlock_t                  lock;
 476         /*
 477          * Protect the list of events.  Locking either mutex or lock
 478          * is sufficient to ensure the list doesn't change; to change
 479          * the list you need to lock both the mutex and the spinlock.
 480          */
 481         struct mutex                    mutex;
 482
 483         struct list_head                pinned_groups;
 484         struct list_head                flexible_groups;
 485         struct list_head                event_list;
 486         int                             nr_events;
 487         int                             nr_active;
 488         int                             is_active;
 489         int                             nr_stat;
 490         int                             nr_freq;
 491         int                             rotate_disable;
 492         atomic_t                        refcount;
 493         struct task_struct              *task;
 494
 495         /*
 496          * Context clock, runs when context enabled.
 497          */
 498         u64                             time;
 499         u64                             timestamp;
 500
 501         /*
 502          * These fields let us detect when two contexts have both
 503          * been cloned (inherited) from a common ancestor.
 504          */
 505         struct perf_event_context       *parent_ctx;
 506         u64                             parent_gen;
 507         u64                             generation;
 508         int                             pin_count;
 509         int                             nr_cgroups;      /* cgroup evts */
 510         int                             nr_branch_stack; /* branch_stack evt */
 511         struct rcu_head                 rcu_head;
 512
 513         struct delayed_work             orphans_remove;
 514         bool                            orphans_remove_sched;
 515 };
 516
 517 /*
 518  * Number of contexts where an event can trigger:
 519  *      task, softirq, hardirq, nmi.
 520  */
 521 #define PERF_NR_CONTEXTS        4
 522
 523 /**
 524  * struct perf_event_cpu_context - per cpu event context structure
 525  */
 526 struct perf_cpu_context {
 527         struct perf_event_context       ctx;
 528         struct perf_event_context       *task_ctx;
 529         int                             active_oncpu;
 530         int                             exclusive;
 531         struct hrtimer                  hrtimer;
 532         ktime_t                         hrtimer_interval;
 533         struct list_head                rotation_list;
 534         struct pmu                      *unique_pmu;
 535         struct perf_cgroup              *cgrp;
 536 };
 537
 538 struct perf_output_handle {
 539         struct perf_event               *event;
 540         struct ring_buffer              *rb;
 541         unsigned long                   wakeup;
 542         unsigned long                   size;
 543         void                            *addr;
 544         int                             page;
 545 };
 546
 547 #ifdef CONFIG_PERF_EVENTS
 548
 549 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 550 extern void perf_pmu_unregister(struct pmu *pmu);
 551
 552 extern int perf_num_counters(void);
 553 extern const char *perf_pmu_name(void);
 554 extern void __perf_event_task_sched_in(struct task_struct *prev,
 555                                        struct task_struct *task);
 556 extern void __perf_event_task_sched_out(struct task_struct *prev,
 557                                         struct task_struct *next);
 558 extern int perf_event_init_task(struct task_struct *child);
 559 extern void perf_event_exit_task(struct task_struct *child);
 560 extern void perf_event_free_task(struct task_struct *task);
 561 extern void perf_event_delayed_put(struct task_struct *task);
 562 extern void perf_event_print_debug(void);
 563 extern void perf_pmu_disable(struct pmu *pmu);
 564 extern void perf_pmu_enable(struct pmu *pmu);
 565 extern int perf_event_task_disable(void);
 566 extern int perf_event_task_enable(void);
 567 extern int perf_event_refresh(struct perf_event *event, int refresh);
 568 extern void perf_event_update_userpage(struct perf_event *event);
 569 extern int perf_event_release_kernel(struct perf_event *event);
 570 extern struct perf_event *
 571 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 572                                 int cpu,
 573                                 struct task_struct *task,
 574                                 perf_overflow_handler_t callback,
 575                                 void *context);
 576 extern void perf_pmu_migrate_context(struct pmu *pmu,
 577                                 int src_cpu, int dst_cpu);
 578 extern u64 perf_event_read_value(struct perf_event *event,
 579                                  u64 *enabled, u64 *running);
 580
 581
 582 struct perf_sample_data {
 583         /*
 584          * Fields set by perf_sample_data_init(), group so as to
 585          * minimize the cachelines touched.
 586          */
 587         u64                             addr;
 588         struct perf_raw_record          *raw;
 589         struct perf_branch_stack        *br_stack;
 590         u64                             period;
 591         u64                             weight;
 592         u64                             txn;
 593         union  perf_mem_data_src        data_src;
 594
 595         /*
 596          * The other fields, optionally {set,used} by
 597          * perf_{prepare,output}_sample().
 598          */
 599         u64                             type;
 600         u64                             ip;
 601         struct {
 602                 u32     pid;
 603                 u32     tid;
 604         }                               tid_entry;
 605         u64                             time;
 606         u64                             id;
 607         u64                             stream_id;
 608         struct {
 609                 u32     cpu;
 610                 u32     reserved;
 611         }                               cpu_entry;
 612         struct perf_callchain_entry     *callchain;
 613         struct perf_regs                regs_user;
 614         struct perf_regs                regs_intr;
 615         u64                             stack_user_size;
 616 } ____cacheline_aligned;
 617
 618 /* default value for data source */
 619 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 620                     PERF_MEM_S(LVL, NA)   |\
 621                     PERF_MEM_S(SNOOP, NA) |\
 622                     PERF_MEM_S(LOCK, NA)  |\
 623                     PERF_MEM_S(TLB, NA))
 624
 625 static inline void perf_sample_data_init(struct perf_sample_data *data,
 626                                          u64 addr, u64 period)
 627 {
 628         /* remaining struct members initialized in perf_prepare_sample() */
 629         data->addr = addr;
 630         data->raw  = NULL;
 631         data->br_stack = NULL;
 632         data->period = period;
 633         data->weight = 0;
 634         data->data_src.val = PERF_MEM_NA;
 635         data->txn = 0;
 636 }
 637
 638 extern void perf_output_sample(struct perf_output_handle *handle,
 639                                struct perf_event_header *header,
 640                                struct perf_sample_data *data,
 641                                struct perf_event *event);
 642 extern void perf_prepare_sample(struct perf_event_header *header,
 643                                 struct perf_sample_data *data,
 644                                 struct perf_event *event,
 645                                 struct pt_regs *regs);
 646
 647 extern int perf_event_overflow(struct perf_event *event,
 648                                  struct perf_sample_data *data,
 649                                  struct pt_regs *regs);
 650
 651 static inline bool is_sampling_event(struct perf_event *event)
 652 {
 653         return event->attr.sample_period != 0;
 654 }
 655
 656 /*
 657  * Return 1 for a software event, 0 for a hardware event
 658  */
 659 static inline int is_software_event(struct perf_event *event)
 660 {
 661         return event->pmu->task_ctx_nr == perf_sw_context;
 662 }
 663
 664 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 665
 666 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 667
 668 #ifndef perf_arch_fetch_caller_regs
 669 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 670 #endif
 671
 672 /*
 673  * Take a snapshot of the regs. Skip ip and frame pointer to
 674  * the nth caller. We only need a few of the regs:
 675  * - ip for PERF_SAMPLE_IP
 676  * - cs for user_mode() tests
 677  * - bp for callchains
 678  * - eflags, for future purposes, just in case
 679  */
 680 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 681 {
 682         memset(regs, 0, sizeof(*regs));
 683
 684         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 685 }
 686
 687 static __always_inline void
 688 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 689 {
 690         struct pt_regs hot_regs;
 691
 692         if (static_key_false(&perf_swevent_enabled[event_id])) {
 693                 if (!regs) {
 694                         perf_fetch_caller_regs(&hot_regs);
 695                         regs = &hot_regs;
 696                 }
 697                 __perf_sw_event(event_id, nr, regs, addr);
 698         }
 699 }
 700
 701 extern struct static_key_deferred perf_sched_events;
 702
 703 static inline void perf_event_task_sched_in(struct task_struct *prev,
 704                                             struct task_struct *task)
 705 {
 706         if (static_key_false(&perf_sched_events.key))
 707                 __perf_event_task_sched_in(prev, task);
 708 }
 709
 710 static inline void perf_event_task_sched_out(struct task_struct *prev,
 711                                              struct task_struct *next)
 712 {
 713         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
 714
 715         if (static_key_false(&perf_sched_events.key))
 716                 __perf_event_task_sched_out(prev, next);
 717 }
 718
 719 extern void perf_event_mmap(struct vm_area_struct *vma);
 720 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 721 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 722 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 723
 724 extern void perf_event_exec(void);
 725 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 726 extern void perf_event_fork(struct task_struct *tsk);
 727
 728 /* Callchains */
 729 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 730
 731 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 732 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 733
 734 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 735 {
 736         if (entry->nr < PERF_MAX_STACK_DEPTH)
 737                 entry->ip[entry->nr++] = ip;
 738 }
 739
 740 extern int sysctl_perf_event_paranoid;
 741 extern int sysctl_perf_event_mlock;
 742 extern int sysctl_perf_event_sample_rate;
 743 extern int sysctl_perf_cpu_time_max_percent;
 744
 745 extern void perf_sample_event_took(u64 sample_len_ns);
 746
 747 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 748                 void __user *buffer, size_t *lenp,
 749                 loff_t *ppos);
 750 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 751                 void __user *buffer, size_t *lenp,
 752                 loff_t *ppos);
 753
 754
 755 static inline bool perf_paranoid_tracepoint_raw(void)
 756 {
 757         return sysctl_perf_event_paranoid > -1;
 758 }
 759
 760 static inline bool perf_paranoid_cpu(void)
 761 {
 762         return sysctl_perf_event_paranoid > 0;
 763 }
 764
 765 static inline bool perf_paranoid_kernel(void)
 766 {
 767         return sysctl_perf_event_paranoid > 1;
 768 }
 769
 770 extern void perf_event_init(void);
 771 extern void perf_tp_event(u64 addr, u64 count, void *record,
 772                           int entry_size, struct pt_regs *regs,
 773                           struct hlist_head *head, int rctx,
 774                           struct task_struct *task);
 775 extern void perf_bp_event(struct perf_event *event, void *data);
 776
 777 #ifndef perf_misc_flags
 778 # define perf_misc_flags(regs) \
 779                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 780 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 781 #endif
 782
 783 static inline bool has_branch_stack(struct perf_event *event)
 784 {
 785         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 786 }
 787
 788 extern int perf_output_begin(struct perf_output_handle *handle,
 789                              struct perf_event *event, unsigned int size);
 790 extern void perf_output_end(struct perf_output_handle *handle);
 791 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 792                              const void *buf, unsigned int len);
 793 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 794                                      unsigned int len);
 795 extern int perf_swevent_get_recursion_context(void);
 796 extern void perf_swevent_put_recursion_context(int rctx);
 797 extern u64 perf_swevent_set_period(struct perf_event *event);
 798 extern void perf_event_enable(struct perf_event *event);
 799 extern void perf_event_disable(struct perf_event *event);
 800 extern int __perf_event_disable(void *info);
 801 extern void perf_event_task_tick(void);
 802 #else /* !CONFIG_PERF_EVENTS: */
 803 static inline void
 804 perf_event_task_sched_in(struct task_struct *prev,
 805                          struct task_struct *task)                      { }
 806 static inline void
 807 perf_event_task_sched_out(struct task_struct *prev,
 808                           struct task_struct *next)                     { }
 809 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 810 static inline void perf_event_exit_task(struct task_struct *child)      { }
 811 static inline void perf_event_free_task(struct task_struct *task)       { }
 812 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 813 static inline void perf_event_print_debug(void)                         { }
 814 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 815 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 816 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 817 {
 818         return -EINVAL;
 819 }
 820
 821 static inline void
 822 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 823 static inline void
 824 perf_bp_event(struct perf_event *event, void *data)                     { }
 825
 826 static inline int perf_register_guest_info_callbacks
 827 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 828 static inline int perf_unregister_guest_info_callbacks
 829 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 830
 831 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 832 static inline void perf_event_exec(void)                                { }
 833 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
 834 static inline void perf_event_fork(struct task_struct *tsk)             { }
 835 static inline void perf_event_init(void)                                { }
 836 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 837 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 838 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
 839 static inline void perf_event_enable(struct perf_event *event)          { }
 840 static inline void perf_event_disable(struct perf_event *event)         { }
 841 static inline int __perf_event_disable(void *info)                      { return -1; }
 842 static inline void perf_event_task_tick(void)                           { }
 843 #endif
 844
 845 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 846 extern bool perf_event_can_stop_tick(void);
 847 #else
 848 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 849 #endif
 850
 851 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 852 extern void perf_restore_debug_store(void);
 853 #else
 854 static inline void perf_restore_debug_store(void)                       { }
 855 #endif
 856
 857 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 858
 859 /*
 860  * This has to have a higher priority than migration_notifier in sched/core.c.
 861  */
 862 #define perf_cpu_notifier(fn)                                           \
 863 do {                                                                    \
 864         static struct notifier_block fn##_nb =                          \
 865                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 866         unsigned long cpu = smp_processor_id();                         \
 867         unsigned long flags;                                            \
 868                                                                         \
 869         cpu_notifier_register_begin();                                  \
 870         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 871                 (void *)(unsigned long)cpu);                            \
 872         local_irq_save(flags);                                          \
 873         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
 874                 (void *)(unsigned long)cpu);                            \
 875         local_irq_restore(flags);                                       \
 876         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
 877                 (void *)(unsigned long)cpu);                            \
 878         __register_cpu_notifier(&fn##_nb);                              \
 879         cpu_notifier_register_done();                                   \
 880 } while (0)
 881
 882 /*
 883  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
 884  * callback for already online CPUs.
 885  */
 886 #define __perf_cpu_notifier(fn)                                         \
 887 do {                                                                    \
 888         static struct notifier_block fn##_nb =                          \
 889                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 890                                                                         \
 891         __register_cpu_notifier(&fn##_nb);                              \
 892 } while (0)
 893
 894 struct perf_pmu_events_attr {
 895         struct device_attribute attr;
 896         u64 id;
 897         const char *event_str;
 898 };
 899
 900 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
 901 static struct perf_pmu_events_attr _var = {                             \
 902         .attr = __ATTR(_name, 0444, _show, NULL),                       \
 903         .id   =  _id,                                                   \
 904 };
 905
 906 #define PMU_FORMAT_ATTR(_name, _format)                                 \
 907 static ssize_t                                                          \
 908 _name##_show(struct device *dev,                                        \
 909                                struct device_attribute *attr,           \
 910                                char *page)                              \
 911 {                                                                       \
 912         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 913         return sprintf(page, _format "\n");                             \
 914 }                                                                       \
 915                                                                         \
 916 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
 917
 918 #endif /* _LINUX_PERF_EVENT_H */