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