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