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