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/atomic.h>
  52 #include <linux/sysfs.h>
  53 #include <linux/perf_regs.h>
  54 #include <asm/local.h>
  55
  56 struct perf_callchain_entry {
  57         __u64                           nr;
  58         __u64                           ip[PERF_MAX_STACK_DEPTH];
  59 };
  60
  61 struct perf_raw_record {
  62         u32                             size;
  63         void                            *data;
  64 };
  65
  66 /*
  67  * single taken branch record layout:
  68  *
  69  *      from: source instruction (may not always be a branch insn)
  70  *        to: branch target
  71  *   mispred: branch target was mispredicted
  72  * predicted: branch target was predicted
  73  *
  74  * support for mispred, predicted is optional. In case it
  75  * is not supported mispred = predicted = 0.
  76  */
  77 struct perf_branch_entry {
  78         __u64   from;
  79         __u64   to;
  80         __u64   mispred:1,  /* target mispredicted */
  81                 predicted:1,/* target predicted */
  82                 reserved:62;
  83 };
  84
  85 /*
  86  * branch stack layout:
  87  *  nr: number of taken branches stored in entries[]
  88  *
  89  * Note that nr can vary from sample to sample
  90  * branches (to, from) are stored from most recent
  91  * to least recent, i.e., entries[0] contains the most
  92  * recent branch.
  93  */
  94 struct perf_branch_stack {
  95         __u64                           nr;
  96         struct perf_branch_entry        entries[0];
  97 };
  98
  99 struct perf_regs_user {
 100         __u64           abi;
 101         struct pt_regs  *regs;
 102 };
 103
 104 struct task_struct;
 105
 106 /*
 107  * extra PMU register associated with an event
 108  */
 109 struct hw_perf_event_extra {
 110         u64             config; /* register value */
 111         unsigned int    reg;    /* register address or index */
 112         int             alloc;  /* extra register already allocated */
 113         int             idx;    /* index in shared_regs->regs[] */
 114 };
 115
 116 /**
 117  * struct hw_perf_event - performance event hardware details:
 118  */
 119 struct hw_perf_event {
 120 #ifdef CONFIG_PERF_EVENTS
 121         union {
 122                 struct { /* hardware */
 123                         u64             config;
 124                         u64             last_tag;
 125                         unsigned long   config_base;
 126                         unsigned long   event_base;
 127                         int             event_base_rdpmc;
 128                         int             idx;
 129                         int             last_cpu;
 130                         int             flags;
 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 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         struct list_head                group_entry;
 311         struct list_head                event_entry;
 312         struct list_head                sibling_list;
 313         struct hlist_node               hlist_entry;
 314         int                             nr_siblings;
 315         int                             group_flags;
 316         struct perf_event               *group_leader;
 317         struct pmu                      *pmu;
 318
 319         enum perf_event_active_state    state;
 320         unsigned int                    attach_state;
 321         local64_t                       count;
 322         atomic64_t                      child_count;
 323
 324         /*
 325          * These are the total time in nanoseconds that the event
 326          * has been enabled (i.e. eligible to run, and the task has
 327          * been scheduled in, if this is a per-task event)
 328          * and running (scheduled onto the CPU), respectively.
 329          *
 330          * They are computed from tstamp_enabled, tstamp_running and
 331          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 332          */
 333         u64                             total_time_enabled;
 334         u64                             total_time_running;
 335
 336         /*
 337          * These are timestamps used for computing total_time_enabled
 338          * and total_time_running when the event is in INACTIVE or
 339          * ACTIVE state, measured in nanoseconds from an arbitrary point
 340          * in time.
 341          * tstamp_enabled: the notional time when the event was enabled
 342          * tstamp_running: the notional time when the event was scheduled on
 343          * tstamp_stopped: in INACTIVE state, the notional time when the
 344          *      event was scheduled off.
 345          */
 346         u64                             tstamp_enabled;
 347         u64                             tstamp_running;
 348         u64                             tstamp_stopped;
 349
 350         /*
 351          * timestamp shadows the actual context timing but it can
 352          * be safely used in NMI interrupt context. It reflects the
 353          * context time as it was when the event was last scheduled in.
 354          *
 355          * ctx_time already accounts for ctx->timestamp. Therefore to
 356          * compute ctx_time for a sample, simply add perf_clock().
 357          */
 358         u64                             shadow_ctx_time;
 359
 360         struct perf_event_attr          attr;
 361         u16                             header_size;
 362         u16                             id_header_size;
 363         u16                             read_size;
 364         struct hw_perf_event            hw;
 365
 366         struct perf_event_context       *ctx;
 367         atomic_long_t                   refcount;
 368
 369         /*
 370          * These accumulate total time (in nanoseconds) that children
 371          * events have been enabled and running, respectively.
 372          */
 373         atomic64_t                      child_total_time_enabled;
 374         atomic64_t                      child_total_time_running;
 375
 376         /*
 377          * Protect attach/detach and child_list:
 378          */
 379         struct mutex                    child_mutex;
 380         struct list_head                child_list;
 381         struct perf_event               *parent;
 382
 383         int                             oncpu;
 384         int                             cpu;
 385
 386         struct list_head                owner_entry;
 387         struct task_struct              *owner;
 388
 389         /* mmap bits */
 390         struct mutex                    mmap_mutex;
 391         atomic_t                        mmap_count;
 392
 393         struct ring_buffer              *rb;
 394         struct list_head                rb_entry;
 395
 396         /* poll related */
 397         wait_queue_head_t               waitq;
 398         struct fasync_struct            *fasync;
 399
 400         /* delayed work for NMIs and such */
 401         int                             pending_wakeup;
 402         int                             pending_kill;
 403         int                             pending_disable;
 404         struct irq_work                 pending;
 405
 406         atomic_t                        event_limit;
 407
 408         void (*destroy)(struct perf_event *);
 409         struct rcu_head                 rcu_head;
 410
 411         struct pid_namespace            *ns;
 412         u64                             id;
 413
 414         perf_overflow_handler_t         overflow_handler;
 415         void                            *overflow_handler_context;
 416
 417 #ifdef CONFIG_EVENT_TRACING
 418         struct ftrace_event_call        *tp_event;
 419         struct event_filter             *filter;
 420 #ifdef CONFIG_FUNCTION_TRACER
 421         struct ftrace_ops               ftrace_ops;
 422 #endif
 423 #endif
 424
 425 #ifdef CONFIG_CGROUP_PERF
 426         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 427         int                             cgrp_defer_enabled;
 428 #endif
 429
 430 #endif /* CONFIG_PERF_EVENTS */
 431 };
 432
 433 enum perf_event_context_type {
 434         task_context,
 435         cpu_context,
 436 };
 437
 438 /**
 439  * struct perf_event_context - event context structure
 440  *
 441  * Used as a container for task events and CPU events as well:
 442  */
 443 struct perf_event_context {
 444         struct pmu                      *pmu;
 445         enum perf_event_context_type    type;
 446         /*
 447          * Protect the states of the events in the list,
 448          * nr_active, and the list:
 449          */
 450         raw_spinlock_t                  lock;
 451         /*
 452          * Protect the list of events.  Locking either mutex or lock
 453          * is sufficient to ensure the list doesn't change; to change
 454          * the list you need to lock both the mutex and the spinlock.
 455          */
 456         struct mutex                    mutex;
 457
 458         struct list_head                pinned_groups;
 459         struct list_head                flexible_groups;
 460         struct list_head                event_list;
 461         int                             nr_events;
 462         int                             nr_active;
 463         int                             is_active;
 464         int                             nr_stat;
 465         int                             nr_freq;
 466         int                             rotate_disable;
 467         atomic_t                        refcount;
 468         struct task_struct              *task;
 469
 470         /*
 471          * Context clock, runs when context enabled.
 472          */
 473         u64                             time;
 474         u64                             timestamp;
 475
 476         /*
 477          * These fields let us detect when two contexts have both
 478          * been cloned (inherited) from a common ancestor.
 479          */
 480         struct perf_event_context       *parent_ctx;
 481         u64                             parent_gen;
 482         u64                             generation;
 483         int                             pin_count;
 484         int                             nr_cgroups;      /* cgroup evts */
 485         int                             nr_branch_stack; /* branch_stack evt */
 486         struct rcu_head                 rcu_head;
 487 };
 488
 489 /*
 490  * Number of contexts where an event can trigger:
 491  *      task, softirq, hardirq, nmi.
 492  */
 493 #define PERF_NR_CONTEXTS        4
 494
 495 /**
 496  * struct perf_event_cpu_context - per cpu event context structure
 497  */
 498 struct perf_cpu_context {
 499         struct perf_event_context       ctx;
 500         struct perf_event_context       *task_ctx;
 501         int                             active_oncpu;
 502         int                             exclusive;
 503         struct list_head                rotation_list;
 504         int                             jiffies_interval;
 505         struct pmu                      *unique_pmu;
 506         struct perf_cgroup              *cgrp;
 507 };
 508
 509 struct perf_output_handle {
 510         struct perf_event               *event;
 511         struct ring_buffer              *rb;
 512         unsigned long                   wakeup;
 513         unsigned long                   size;
 514         void                            *addr;
 515         int                             page;
 516 };
 517
 518 #ifdef CONFIG_PERF_EVENTS
 519
 520 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
 521 extern void perf_pmu_unregister(struct pmu *pmu);
 522
 523 extern int perf_num_counters(void);
 524 extern const char *perf_pmu_name(void);
 525 extern void __perf_event_task_sched_in(struct task_struct *prev,
 526                                        struct task_struct *task);
 527 extern void __perf_event_task_sched_out(struct task_struct *prev,
 528                                         struct task_struct *next);
 529 extern int perf_event_init_task(struct task_struct *child);
 530 extern void perf_event_exit_task(struct task_struct *child);
 531 extern void perf_event_free_task(struct task_struct *task);
 532 extern void perf_event_delayed_put(struct task_struct *task);
 533 extern void perf_event_print_debug(void);
 534 extern void perf_pmu_disable(struct pmu *pmu);
 535 extern void perf_pmu_enable(struct pmu *pmu);
 536 extern int perf_event_task_disable(void);
 537 extern int perf_event_task_enable(void);
 538 extern int perf_event_refresh(struct perf_event *event, int refresh);
 539 extern void perf_event_update_userpage(struct perf_event *event);
 540 extern int perf_event_release_kernel(struct perf_event *event);
 541 extern struct perf_event *
 542 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 543                                 int cpu,
 544                                 struct task_struct *task,
 545                                 perf_overflow_handler_t callback,
 546                                 void *context);
 547 extern void perf_pmu_migrate_context(struct pmu *pmu,
 548                                 int src_cpu, int dst_cpu);
 549 extern u64 perf_event_read_value(struct perf_event *event,
 550                                  u64 *enabled, u64 *running);
 551
 552
 553 struct perf_sample_data {
 554         u64                             type;
 555
 556         u64                             ip;
 557         struct {
 558                 u32     pid;
 559                 u32     tid;
 560         }                               tid_entry;
 561         u64                             time;
 562         u64                             addr;
 563         u64                             id;
 564         u64                             stream_id;
 565         struct {
 566                 u32     cpu;
 567                 u32     reserved;
 568         }                               cpu_entry;
 569         u64                             period;
 570         union  perf_mem_data_src        data_src;
 571         struct perf_callchain_entry     *callchain;
 572         struct perf_raw_record          *raw;
 573         struct perf_branch_stack        *br_stack;
 574         struct perf_regs_user           regs_user;
 575         u64                             stack_user_size;
 576         u64                             weight;
 577 };
 578
 579 static inline void perf_sample_data_init(struct perf_sample_data *data,
 580                                          u64 addr, u64 period)
 581 {
 582         /* remaining struct members initialized in perf_prepare_sample() */
 583         data->addr = addr;
 584         data->raw  = NULL;
 585         data->br_stack = NULL;
 586         data->period = period;
 587         data->regs_user.abi = PERF_SAMPLE_REGS_ABI_NONE;
 588         data->regs_user.regs = NULL;
 589         data->stack_user_size = 0;
 590         data->weight = 0;
 591         data->data_src.val = 0;
 592 }
 593
 594 extern void perf_output_sample(struct perf_output_handle *handle,
 595                                struct perf_event_header *header,
 596                                struct perf_sample_data *data,
 597                                struct perf_event *event);
 598 extern void perf_prepare_sample(struct perf_event_header *header,
 599                                 struct perf_sample_data *data,
 600                                 struct perf_event *event,
 601                                 struct pt_regs *regs);
 602
 603 extern int perf_event_overflow(struct perf_event *event,
 604                                  struct perf_sample_data *data,
 605                                  struct pt_regs *regs);
 606
 607 static inline bool is_sampling_event(struct perf_event *event)
 608 {
 609         return event->attr.sample_period != 0;
 610 }
 611
 612 /*
 613  * Return 1 for a software event, 0 for a hardware event
 614  */
 615 static inline int is_software_event(struct perf_event *event)
 616 {
 617         return event->pmu->task_ctx_nr == perf_sw_context;
 618 }
 619
 620 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 621
 622 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 623
 624 #ifndef perf_arch_fetch_caller_regs
 625 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 626 #endif
 627
 628 /*
 629  * Take a snapshot of the regs. Skip ip and frame pointer to
 630  * the nth caller. We only need a few of the regs:
 631  * - ip for PERF_SAMPLE_IP
 632  * - cs for user_mode() tests
 633  * - bp for callchains
 634  * - eflags, for future purposes, just in case
 635  */
 636 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 637 {
 638         memset(regs, 0, sizeof(*regs));
 639
 640         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 641 }
 642
 643 static __always_inline void
 644 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 645 {
 646         struct pt_regs hot_regs;
 647
 648         if (static_key_false(&perf_swevent_enabled[event_id])) {
 649                 if (!regs) {
 650                         perf_fetch_caller_regs(&hot_regs);
 651                         regs = &hot_regs;
 652                 }
 653                 __perf_sw_event(event_id, nr, regs, addr);
 654         }
 655 }
 656
 657 extern struct static_key_deferred perf_sched_events;
 658
 659 static inline void perf_event_task_sched_in(struct task_struct *prev,
 660                                             struct task_struct *task)
 661 {
 662         if (static_key_false(&perf_sched_events.key))
 663                 __perf_event_task_sched_in(prev, task);
 664 }
 665
 666 static inline void perf_event_task_sched_out(struct task_struct *prev,
 667                                              struct task_struct *next)
 668 {
 669         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
 670
 671         if (static_key_false(&perf_sched_events.key))
 672                 __perf_event_task_sched_out(prev, next);
 673 }
 674
 675 extern void perf_event_mmap(struct vm_area_struct *vma);
 676 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 677 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 678 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 679
 680 extern void perf_event_comm(struct task_struct *tsk);
 681 extern void perf_event_fork(struct task_struct *tsk);
 682
 683 /* Callchains */
 684 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 685
 686 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 687 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 688
 689 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 690 {
 691         if (entry->nr < PERF_MAX_STACK_DEPTH)
 692                 entry->ip[entry->nr++] = ip;
 693 }
 694
 695 extern int sysctl_perf_event_paranoid;
 696 extern int sysctl_perf_event_mlock;
 697 extern int sysctl_perf_event_sample_rate;
 698 extern int sysctl_perf_cpu_time_max_percent;
 699
 700 extern void perf_sample_event_took(u64 sample_len_ns);
 701
 702 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 703                 void __user *buffer, size_t *lenp,
 704                 loff_t *ppos);
 705 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 706                 void __user *buffer, size_t *lenp,
 707                 loff_t *ppos);
 708
 709
 710 static inline bool perf_paranoid_tracepoint_raw(void)
 711 {
 712         return sysctl_perf_event_paranoid > -1;
 713 }
 714
 715 static inline bool perf_paranoid_cpu(void)
 716 {
 717         return sysctl_perf_event_paranoid > 0;
 718 }
 719
 720 static inline bool perf_paranoid_kernel(void)
 721 {
 722         return sysctl_perf_event_paranoid > 1;
 723 }
 724
 725 extern void perf_event_init(void);
 726 extern void perf_tp_event(u64 addr, u64 count, void *record,
 727                           int entry_size, struct pt_regs *regs,
 728                           struct hlist_head *head, int rctx,
 729                           struct task_struct *task);
 730 extern void perf_bp_event(struct perf_event *event, void *data);
 731
 732 #ifndef perf_misc_flags
 733 # define perf_misc_flags(regs) \
 734                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 735 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 736 #endif
 737
 738 static inline bool has_branch_stack(struct perf_event *event)
 739 {
 740         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 741 }
 742
 743 extern int perf_output_begin(struct perf_output_handle *handle,
 744                              struct perf_event *event, unsigned int size);
 745 extern void perf_output_end(struct perf_output_handle *handle);
 746 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 747                              const void *buf, unsigned int len);
 748 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 749                                      unsigned int len);
 750 extern int perf_swevent_get_recursion_context(void);
 751 extern void perf_swevent_put_recursion_context(int rctx);
 752 extern void perf_event_enable(struct perf_event *event);
 753 extern void perf_event_disable(struct perf_event *event);
 754 extern int __perf_event_disable(void *info);
 755 extern void perf_event_task_tick(void);
 756 #else
 757 static inline void
 758 perf_event_task_sched_in(struct task_struct *prev,
 759                          struct task_struct *task)                      { }
 760 static inline void
 761 perf_event_task_sched_out(struct task_struct *prev,
 762                           struct task_struct *next)                     { }
 763 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 764 static inline void perf_event_exit_task(struct task_struct *child)      { }
 765 static inline void perf_event_free_task(struct task_struct *task)       { }
 766 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 767 static inline void perf_event_print_debug(void)                         { }
 768 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 769 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 770 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 771 {
 772         return -EINVAL;
 773 }
 774
 775 static inline void
 776 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 777 static inline void
 778 perf_bp_event(struct perf_event *event, void *data)                     { }
 779
 780 static inline int perf_register_guest_info_callbacks
 781 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 782 static inline int perf_unregister_guest_info_callbacks
 783 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 784
 785 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 786 static inline void perf_event_comm(struct task_struct *tsk)             { }
 787 static inline void perf_event_fork(struct task_struct *tsk)             { }
 788 static inline void perf_event_init(void)                                { }
 789 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 790 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 791 static inline void perf_event_enable(struct perf_event *event)          { }
 792 static inline void perf_event_disable(struct perf_event *event)         { }
 793 static inline int __perf_event_disable(void *info)                      { return -1; }
 794 static inline void perf_event_task_tick(void)                           { }
 795 #endif
 796
 797 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 798 extern bool perf_event_can_stop_tick(void);
 799 #else
 800 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 801 #endif
 802
 803 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 804 extern void perf_restore_debug_store(void);
 805 #else
 806 static inline void perf_restore_debug_store(void)                       { }
 807 #endif
 808
 809 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 810
 811 /*
 812  * This has to have a higher priority than migration_notifier in sched.c.
 813  */
 814 #define perf_cpu_notifier(fn)                                           \
 815 do {                                                                    \
 816         static struct notifier_block fn##_nb __cpuinitdata =            \
 817                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 818         unsigned long cpu = smp_processor_id();                         \
 819         unsigned long flags;                                            \
 820         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 821                 (void *)(unsigned long)cpu);                            \
 822         local_irq_save(flags);                                          \
 823         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
 824                 (void *)(unsigned long)cpu);                            \
 825         local_irq_restore(flags);                                       \
 826         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
 827                 (void *)(unsigned long)cpu);                            \
 828         register_cpu_notifier(&fn##_nb);                                \
 829 } while (0)
 830
 831
 832 struct perf_pmu_events_attr {
 833         struct device_attribute attr;
 834         u64 id;
 835         const char *event_str;
 836 };
 837
 838 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
 839 static struct perf_pmu_events_attr _var = {                             \
 840         .attr = __ATTR(_name, 0444, _show, NULL),                       \
 841         .id   =  _id,                                                   \
 842 };
 843
 844 #define PMU_FORMAT_ATTR(_name, _format)                                 \
 845 static ssize_t                                                          \
 846 _name##_show(struct device *dev,                                        \
 847                                struct device_attribute *attr,           \
 848                                char *page)                              \
 849 {                                                                       \
 850         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 851         return sprintf(page, _format "\n");                             \
 852 }                                                                       \
 853                                                                         \
 854 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
 855
 856 #endif /* _LINUX_PERF_EVENT_H */