include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   7  *
   8  * Data type definitions, declarations, prototypes.
   9  *
  10  *    Started by: Thomas Gleixner and Ingo Molnar
  11  *
  12  * For licencing details see kernel-base/COPYING
  13  */
  14 #ifndef _LINUX_PERF_EVENT_H
  15 #define _LINUX_PERF_EVENT_H
  16
  17 #include <uapi/linux/perf_event.h>
  18
  19 /*
  20  * Kernel-internal data types and definitions:
  21  */
  22
  23 #ifdef CONFIG_PERF_EVENTS
  24 # include <asm/perf_event.h>
  25 # include <asm/local64.h>
  26 #endif
  27
  28 struct perf_guest_info_callbacks {
  29         int                             (*is_in_guest)(void);
  30         int                             (*is_user_mode)(void);
  31         unsigned long                   (*get_guest_ip)(void);
  32 };
  33
  34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
  35 #include <asm/hw_breakpoint.h>
  36 #endif
  37
  38 #include <linux/list.h>
  39 #include <linux/mutex.h>
  40 #include <linux/rculist.h>
  41 #include <linux/rcupdate.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/hrtimer.h>
  44 #include <linux/fs.h>
  45 #include <linux/pid_namespace.h>
  46 #include <linux/workqueue.h>
  47 #include <linux/ftrace.h>
  48 #include <linux/cpu.h>
  49 #include <linux/irq_work.h>
  50 #include <linux/static_key.h>
  51 #include <linux/jump_label_ratelimit.h>
  52 #include <linux/atomic.h>
  53 #include <linux/sysfs.h>
  54 #include <linux/perf_regs.h>
  55 #include <linux/workqueue.h>
  56 #include <asm/local.h>
  57
  58 struct perf_callchain_entry {
  59         __u64                           nr;
  60         __u64                           ip[PERF_MAX_STACK_DEPTH];
  61 };
  62
  63 struct perf_raw_record {
  64         u32                             size;
  65         void                            *data;
  66 };
  67
  68 /*
  69  * branch stack layout:
  70  *  nr: number of taken branches stored in entries[]
  71  *
  72  * Note that nr can vary from sample to sample
  73  * branches (to, from) are stored from most recent
  74  * to least recent, i.e., entries[0] contains the most
  75  * recent branch.
  76  */
  77 struct perf_branch_stack {
  78         __u64                           nr;
  79         struct perf_branch_entry        entries[0];
  80 };
  81
  82 struct task_struct;
  83
  84 /*
  85  * extra PMU register associated with an event
  86  */
  87 struct hw_perf_event_extra {
  88         u64             config; /* register value */
  89         unsigned int    reg;    /* register address or index */
  90         int             alloc;  /* extra register already allocated */
  91         int             idx;    /* index in shared_regs->regs[] */
  92 };
  93
  94 struct event_constraint;
  95
  96 /**
  97  * struct hw_perf_event - performance event hardware details:
  98  */
  99 struct hw_perf_event {
 100 #ifdef CONFIG_PERF_EVENTS
 101         union {
 102                 struct { /* hardware */
 103                         u64             config;
 104                         u64             last_tag;
 105                         unsigned long   config_base;
 106                         unsigned long   event_base;
 107                         int             event_base_rdpmc;
 108                         int             idx;
 109                         int             last_cpu;
 110                         int             flags;
 111
 112                         struct hw_perf_event_extra extra_reg;
 113                         struct hw_perf_event_extra branch_reg;
 114
 115                         struct event_constraint *constraint;
 116                 };
 117                 struct { /* software */
 118                         struct hrtimer  hrtimer;
 119                 };
 120                 struct { /* tracepoint */
 121                         struct task_struct      *tp_target;
 122                         /* for tp_event->class */
 123                         struct list_head        tp_list;
 124                 };
 125 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 126                 struct { /* breakpoint */
 127                         /*
 128                          * Crufty hack to avoid the chicken and egg
 129                          * problem hw_breakpoint has with context
 130                          * creation and event initalization.
 131                          */
 132                         struct task_struct              *bp_target;
 133                         struct arch_hw_breakpoint       info;
 134                         struct list_head                bp_list;
 135                 };
 136 #endif
 137         };
 138         int                             state;
 139         local64_t                       prev_count;
 140         u64                             sample_period;
 141         u64                             last_period;
 142         local64_t                       period_left;
 143         u64                             interrupts_seq;
 144         u64                             interrupts;
 145
 146         u64                             freq_time_stamp;
 147         u64                             freq_count_stamp;
 148 #endif
 149 };
 150
 151 /*
 152  * hw_perf_event::state flags
 153  */
 154 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 155 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 156 #define PERF_HES_ARCH           0x04
 157
 158 struct perf_event;
 159
 160 /*
 161  * Common implementation detail of pmu::{start,commit,cancel}_txn
 162  */
 163 #define PERF_EVENT_TXN 0x1
 164
 165 /**
 166  * pmu::capabilities flags
 167  */
 168 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 169
 170 /**
 171  * struct pmu - generic performance monitoring unit
 172  */
 173 struct pmu {
 174         struct list_head                entry;
 175
 176         struct module                   *module;
 177         struct device                   *dev;
 178         const struct attribute_group    **attr_groups;
 179         const char                      *name;
 180         int                             type;
 181
 182         /*
 183          * various common per-pmu feature flags
 184          */
 185         int                             capabilities;
 186
 187         int * __percpu                  pmu_disable_count;
 188         struct perf_cpu_context * __percpu pmu_cpu_context;
 189         int                             task_ctx_nr;
 190         int                             hrtimer_interval_ms;
 191
 192         /*
 193          * Fully disable/enable this PMU, can be used to protect from the PMI
 194          * as well as for lazy/batch writing of the MSRs.
 195          */
 196         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 197         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 198
 199         /*
 200          * Try and initialize the event for this PMU.
 201          * Should return -ENOENT when the @event doesn't match this PMU.
 202          */
 203         int (*event_init)               (struct perf_event *event);
 204
 205 #define PERF_EF_START   0x01            /* start the counter when adding    */
 206 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 207 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 208
 209         /*
 210          * Adds/Removes a counter to/from the PMU, can be done inside
 211          * a transaction, see the ->*_txn() methods.
 212          */
 213         int  (*add)                     (struct perf_event *event, int flags);
 214         void (*del)                     (struct perf_event *event, int flags);
 215
 216         /*
 217          * Starts/Stops a counter present on the PMU. The PMI handler
 218          * should stop the counter when perf_event_overflow() returns
 219          * !0. ->start() will be used to continue.
 220          */
 221         void (*start)                   (struct perf_event *event, int flags);
 222         void (*stop)                    (struct perf_event *event, int flags);
 223
 224         /*
 225          * Updates the counter value of the event.
 226          */
 227         void (*read)                    (struct perf_event *event);
 228
 229         /*
 230          * Group events scheduling is treated as a transaction, add
 231          * group events as a whole and perform one schedulability test.
 232          * If the test fails, roll back the whole group
 233          *
 234          * Start the transaction, after this ->add() doesn't need to
 235          * do schedulability tests.
 236          */
 237         void (*start_txn)               (struct pmu *pmu); /* optional */
 238         /*
 239          * If ->start_txn() disabled the ->add() schedulability test
 240          * then ->commit_txn() is required to perform one. On success
 241          * the transaction is closed. On error the transaction is kept
 242          * open until ->cancel_txn() is called.
 243          */
 244         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 245         /*
 246          * Will cancel the transaction, assumes ->del() is called
 247          * for each successful ->add() during the transaction.
 248          */
 249         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 250
 251         /*
 252          * Will return the value for perf_event_mmap_page::index for this event,
 253          * if no implementation is provided it will default to: event->hw.idx + 1.
 254          */
 255         int (*event_idx)                (struct perf_event *event); /*optional */
 256
 257         /*
 258          * flush branch stack on context-switches (needed in cpu-wide mode)
 259          */
 260         void (*flush_branch_stack)      (void);
 261 };
 262
 263 /**
 264  * enum perf_event_active_state - the states of a event
 265  */
 266 enum perf_event_active_state {
 267         PERF_EVENT_STATE_EXIT           = -3,
 268         PERF_EVENT_STATE_ERROR          = -2,
 269         PERF_EVENT_STATE_OFF            = -1,
 270         PERF_EVENT_STATE_INACTIVE       =  0,
 271         PERF_EVENT_STATE_ACTIVE         =  1,
 272 };
 273
 274 struct file;
 275 struct perf_sample_data;
 276
 277 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 278                                         struct perf_sample_data *,
 279                                         struct pt_regs *regs);
 280
 281 enum perf_group_flag {
 282         PERF_GROUP_SOFTWARE             = 0x1,
 283 };
 284
 285 #define SWEVENT_HLIST_BITS              8
 286 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 287
 288 struct swevent_hlist {
 289         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 290         struct rcu_head                 rcu_head;
 291 };
 292
 293 #define PERF_ATTACH_CONTEXT     0x01
 294 #define PERF_ATTACH_GROUP       0x02
 295 #define PERF_ATTACH_TASK        0x04
 296
 297 struct perf_cgroup;
 298 struct ring_buffer;
 299
 300 /**
 301  * struct perf_event - performance event kernel representation:
 302  */
 303 struct perf_event {
 304 #ifdef CONFIG_PERF_EVENTS
 305         /*
 306          * entry onto perf_event_context::event_list;
 307          *   modifications require ctx->lock
 308          *   RCU safe iterations.
 309          */
 310         struct list_head                event_entry;
 311
 312         /*
 313          * XXX: group_entry and sibling_list should be mutually exclusive;
 314          * either you're a sibling on a group, or you're the group leader.
 315          * Rework the code to always use the same list element.
 316          *
 317          * Locked for modification by both ctx->mutex and ctx->lock; holding
 318          * either sufficies for read.
 319          */
 320         struct list_head                group_entry;
 321         struct list_head                sibling_list;
 322
 323         /*
 324          * We need storage to track the entries in perf_pmu_migrate_context; we
 325          * cannot use the event_entry because of RCU and we want to keep the
 326          * group in tact which avoids us using the other two entries.
 327          */
 328         struct list_head                migrate_entry;
 329
 330         struct hlist_node               hlist_entry;
 331         struct list_head                active_entry;
 332         int                             nr_siblings;
 333         int                             group_flags;
 334         struct perf_event               *group_leader;
 335         struct pmu                      *pmu;
 336
 337         enum perf_event_active_state    state;
 338         unsigned int                    attach_state;
 339         local64_t                       count;
 340         atomic64_t                      child_count;
 341
 342         /*
 343          * These are the total time in nanoseconds that the event
 344          * has been enabled (i.e. eligible to run, and the task has
 345          * been scheduled in, if this is a per-task event)
 346          * and running (scheduled onto the CPU), respectively.
 347          *
 348          * They are computed from tstamp_enabled, tstamp_running and
 349          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 350          */
 351         u64                             total_time_enabled;
 352         u64                             total_time_running;
 353
 354         /*
 355          * These are timestamps used for computing total_time_enabled
 356          * and total_time_running when the event is in INACTIVE or
 357          * ACTIVE state, measured in nanoseconds from an arbitrary point
 358          * in time.
 359          * tstamp_enabled: the notional time when the event was enabled
 360          * tstamp_running: the notional time when the event was scheduled on
 361          * tstamp_stopped: in INACTIVE state, the notional time when the
 362          *      event was scheduled off.
 363          */
 364         u64                             tstamp_enabled;
 365         u64                             tstamp_running;
 366         u64                             tstamp_stopped;
 367
 368         /*
 369          * timestamp shadows the actual context timing but it can
 370          * be safely used in NMI interrupt context. It reflects the
 371          * context time as it was when the event was last scheduled in.
 372          *
 373          * ctx_time already accounts for ctx->timestamp. Therefore to
 374          * compute ctx_time for a sample, simply add perf_clock().
 375          */
 376         u64                             shadow_ctx_time;
 377
 378         struct perf_event_attr          attr;
 379         u16                             header_size;
 380         u16                             id_header_size;
 381         u16                             read_size;
 382         struct hw_perf_event            hw;
 383
 384         struct perf_event_context       *ctx;
 385         atomic_long_t                   refcount;
 386
 387         /*
 388          * These accumulate total time (in nanoseconds) that children
 389          * events have been enabled and running, respectively.
 390          */
 391         atomic64_t                      child_total_time_enabled;
 392         atomic64_t                      child_total_time_running;
 393
 394         /*
 395          * Protect attach/detach and child_list:
 396          */
 397         struct mutex                    child_mutex;
 398         struct list_head                child_list;
 399         struct perf_event               *parent;
 400
 401         int                             oncpu;
 402         int                             cpu;
 403
 404         struct list_head                owner_entry;
 405         struct task_struct              *owner;
 406
 407         /* mmap bits */
 408         struct mutex                    mmap_mutex;
 409         atomic_t                        mmap_count;
 410
 411         struct ring_buffer              *rb;
 412         struct list_head                rb_entry;
 413         unsigned long                   rcu_batches;
 414         int                             rcu_pending;
 415
 416         /* poll related */
 417         wait_queue_head_t               waitq;
 418         struct fasync_struct            *fasync;
 419
 420         /* delayed work for NMIs and such */
 421         int                             pending_wakeup;
 422         int                             pending_kill;
 423         int                             pending_disable;
 424         struct irq_work                 pending;
 425
 426         atomic_t                        event_limit;
 427
 428         void (*destroy)(struct perf_event *);
 429         struct rcu_head                 rcu_head;
 430
 431         struct pid_namespace            *ns;
 432         u64                             id;
 433
 434         perf_overflow_handler_t         overflow_handler;
 435         void                            *overflow_handler_context;
 436
 437 #ifdef CONFIG_EVENT_TRACING
 438         struct ftrace_event_call        *tp_event;
 439         struct event_filter             *filter;
 440 #ifdef CONFIG_FUNCTION_TRACER
 441         struct ftrace_ops               ftrace_ops;
 442 #endif
 443 #endif
 444
 445 #ifdef CONFIG_CGROUP_PERF
 446         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 447         int                             cgrp_defer_enabled;
 448 #endif
 449
 450 #endif /* CONFIG_PERF_EVENTS */
 451 };
 452
 453 /**
 454  * struct perf_event_context - event context structure
 455  *
 456  * Used as a container for task events and CPU events as well:
 457  */
 458 struct perf_event_context {
 459         struct pmu                      *pmu;
 460         /*
 461          * Protect the states of the events in the list,
 462          * nr_active, and the list:
 463          */
 464         raw_spinlock_t                  lock;
 465         /*
 466          * Protect the list of events.  Locking either mutex or lock
 467          * is sufficient to ensure the list doesn't change; to change
 468          * the list you need to lock both the mutex and the spinlock.
 469          */
 470         struct mutex                    mutex;
 471
 472         struct list_head                active_ctx_list;
 473         struct list_head                pinned_groups;
 474         struct list_head                flexible_groups;
 475         struct list_head                event_list;
 476         int                             nr_events;
 477         int                             nr_active;
 478         int                             is_active;
 479         int                             nr_stat;
 480         int                             nr_freq;
 481         int                             rotate_disable;
 482         atomic_t                        refcount;
 483         struct task_struct              *task;
 484
 485         /*
 486          * Context clock, runs when context enabled.
 487          */
 488         u64                             time;
 489         u64                             timestamp;
 490
 491         /*
 492          * These fields let us detect when two contexts have both
 493          * been cloned (inherited) from a common ancestor.
 494          */
 495         struct perf_event_context       *parent_ctx;
 496         u64                             parent_gen;
 497         u64                             generation;
 498         int                             pin_count;
 499         int                             nr_cgroups;      /* cgroup evts */
 500         int                             nr_branch_stack; /* branch_stack evt */
 501         struct rcu_head                 rcu_head;
 502
 503         struct delayed_work             orphans_remove;
 504         bool                            orphans_remove_sched;
 505 };
 506
 507 /*
 508  * Number of contexts where an event can trigger:
 509  *      task, softirq, hardirq, nmi.
 510  */
 511 #define PERF_NR_CONTEXTS        4
 512
 513 /**
 514  * struct perf_event_cpu_context - per cpu event context structure
 515  */
 516 struct perf_cpu_context {
 517         struct perf_event_context       ctx;
 518         struct perf_event_context       *task_ctx;
 519         int                             active_oncpu;
 520         int                             exclusive;
 521         struct hrtimer                  hrtimer;
 522         ktime_t                         hrtimer_interval;
 523         struct pmu                      *unique_pmu;
 524         struct perf_cgroup              *cgrp;
 525 };
 526
 527 struct perf_output_handle {
 528         struct perf_event               *event;
 529         struct ring_buffer              *rb;
 530         unsigned long                   wakeup;
 531         unsigned long                   size;
 532         void                            *addr;
 533         int                             page;
 534 };
 535
 536 #ifdef CONFIG_PERF_EVENTS
 537
 538 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 539 extern void perf_pmu_unregister(struct pmu *pmu);
 540
 541 extern int perf_num_counters(void);
 542 extern const char *perf_pmu_name(void);
 543 extern void __perf_event_task_sched_in(struct task_struct *prev,
 544                                        struct task_struct *task);
 545 extern void __perf_event_task_sched_out(struct task_struct *prev,
 546                                         struct task_struct *next);
 547 extern int perf_event_init_task(struct task_struct *child);
 548 extern void perf_event_exit_task(struct task_struct *child);
 549 extern void perf_event_free_task(struct task_struct *task);
 550 extern void perf_event_delayed_put(struct task_struct *task);
 551 extern void perf_event_print_debug(void);
 552 extern void perf_pmu_disable(struct pmu *pmu);
 553 extern void perf_pmu_enable(struct pmu *pmu);
 554 extern int perf_event_task_disable(void);
 555 extern int perf_event_task_enable(void);
 556 extern int perf_event_refresh(struct perf_event *event, int refresh);
 557 extern void perf_event_update_userpage(struct perf_event *event);
 558 extern int perf_event_release_kernel(struct perf_event *event);
 559 extern struct perf_event *
 560 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 561                                 int cpu,
 562                                 struct task_struct *task,
 563                                 perf_overflow_handler_t callback,
 564                                 void *context);
 565 extern void perf_pmu_migrate_context(struct pmu *pmu,
 566                                 int src_cpu, int dst_cpu);
 567 extern u64 perf_event_read_value(struct perf_event *event,
 568                                  u64 *enabled, u64 *running);
 569
 570
 571 struct perf_sample_data {
 572         /*
 573          * Fields set by perf_sample_data_init(), group so as to
 574          * minimize the cachelines touched.
 575          */
 576         u64                             addr;
 577         struct perf_raw_record          *raw;
 578         struct perf_branch_stack        *br_stack;
 579         u64                             period;
 580         u64                             weight;
 581         u64                             txn;
 582         union  perf_mem_data_src        data_src;
 583
 584         /*
 585          * The other fields, optionally {set,used} by
 586          * perf_{prepare,output}_sample().
 587          */
 588         u64                             type;
 589         u64                             ip;
 590         struct {
 591                 u32     pid;
 592                 u32     tid;
 593         }                               tid_entry;
 594         u64                             time;
 595         u64                             id;
 596         u64                             stream_id;
 597         struct {
 598                 u32     cpu;
 599                 u32     reserved;
 600         }                               cpu_entry;
 601         struct perf_callchain_entry     *callchain;
 602
 603         /*
 604          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 605          * on arch details.
 606          */
 607         struct perf_regs                regs_user;
 608         struct pt_regs                  regs_user_copy;
 609
 610         struct perf_regs                regs_intr;
 611         u64                             stack_user_size;
 612 } ____cacheline_aligned;
 613
 614 /* default value for data source */
 615 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 616                     PERF_MEM_S(LVL, NA)   |\
 617                     PERF_MEM_S(SNOOP, NA) |\
 618                     PERF_MEM_S(LOCK, NA)  |\
 619                     PERF_MEM_S(TLB, NA))
 620
 621 static inline void perf_sample_data_init(struct perf_sample_data *data,
 622                                          u64 addr, u64 period)
 623 {
 624         /* remaining struct members initialized in perf_prepare_sample() */
 625         data->addr = addr;
 626         data->raw  = NULL;
 627         data->br_stack = NULL;
 628         data->period = period;
 629         data->weight = 0;
 630         data->data_src.val = PERF_MEM_NA;
 631         data->txn = 0;
 632 }
 633
 634 extern void perf_output_sample(struct perf_output_handle *handle,
 635                                struct perf_event_header *header,
 636                                struct perf_sample_data *data,
 637                                struct perf_event *event);
 638 extern void perf_prepare_sample(struct perf_event_header *header,
 639                                 struct perf_sample_data *data,
 640                                 struct perf_event *event,
 641                                 struct pt_regs *regs);
 642
 643 extern int perf_event_overflow(struct perf_event *event,
 644                                  struct perf_sample_data *data,
 645                                  struct pt_regs *regs);
 646
 647 static inline bool is_sampling_event(struct perf_event *event)
 648 {
 649         return event->attr.sample_period != 0;
 650 }
 651
 652 /*
 653  * Return 1 for a software event, 0 for a hardware event
 654  */
 655 static inline int is_software_event(struct perf_event *event)
 656 {
 657         return event->pmu->task_ctx_nr == perf_sw_context;
 658 }
 659
 660 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 661
 662 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 663 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 664
 665 #ifndef perf_arch_fetch_caller_regs
 666 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 667 #endif
 668
 669 /*
 670  * Take a snapshot of the regs. Skip ip and frame pointer to
 671  * the nth caller. We only need a few of the regs:
 672  * - ip for PERF_SAMPLE_IP
 673  * - cs for user_mode() tests
 674  * - bp for callchains
 675  * - eflags, for future purposes, just in case
 676  */
 677 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 678 {
 679         memset(regs, 0, sizeof(*regs));
 680
 681         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 682 }
 683
 684 static __always_inline void
 685 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 686 {
 687         if (static_key_false(&perf_swevent_enabled[event_id]))
 688                 __perf_sw_event(event_id, nr, regs, addr);
 689 }
 690
 691 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 692
 693 /*
 694  * 'Special' version for the scheduler, it hard assumes no recursion,
 695  * which is guaranteed by us not actually scheduling inside other swevents
 696  * because those disable preemption.
 697  */
 698 static __always_inline void
 699 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 700 {
 701         if (static_key_false(&perf_swevent_enabled[event_id])) {
 702                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 703
 704                 perf_fetch_caller_regs(regs);
 705                 ___perf_sw_event(event_id, nr, regs, addr);
 706         }
 707 }
 708
 709 extern struct static_key_deferred perf_sched_events;
 710
 711 static inline void perf_event_task_sched_in(struct task_struct *prev,
 712                                             struct task_struct *task)
 713 {
 714         if (static_key_false(&perf_sched_events.key))
 715                 __perf_event_task_sched_in(prev, task);
 716 }
 717
 718 static inline void perf_event_task_sched_out(struct task_struct *prev,
 719                                              struct task_struct *next)
 720 {
 721         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 722
 723         if (static_key_false(&perf_sched_events.key))
 724                 __perf_event_task_sched_out(prev, next);
 725 }
 726
 727 extern void perf_event_mmap(struct vm_area_struct *vma);
 728 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 729 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 730 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 731
 732 extern void perf_event_exec(void);
 733 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 734 extern void perf_event_fork(struct task_struct *tsk);
 735
 736 /* Callchains */
 737 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 738
 739 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 740 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 741
 742 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 743 {
 744         if (entry->nr < PERF_MAX_STACK_DEPTH)
 745                 entry->ip[entry->nr++] = ip;
 746 }
 747
 748 extern int sysctl_perf_event_paranoid;
 749 extern int sysctl_perf_event_mlock;
 750 extern int sysctl_perf_event_sample_rate;
 751 extern int sysctl_perf_cpu_time_max_percent;
 752
 753 extern void perf_sample_event_took(u64 sample_len_ns);
 754
 755 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 756                 void __user *buffer, size_t *lenp,
 757                 loff_t *ppos);
 758 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 759                 void __user *buffer, size_t *lenp,
 760                 loff_t *ppos);
 761
 762
 763 static inline bool perf_paranoid_tracepoint_raw(void)
 764 {
 765         return sysctl_perf_event_paranoid > -1;
 766 }
 767
 768 static inline bool perf_paranoid_cpu(void)
 769 {
 770         return sysctl_perf_event_paranoid > 0;
 771 }
 772
 773 static inline bool perf_paranoid_kernel(void)
 774 {
 775         return sysctl_perf_event_paranoid > 1;
 776 }
 777
 778 extern void perf_event_init(void);
 779 extern void perf_tp_event(u64 addr, u64 count, void *record,
 780                           int entry_size, struct pt_regs *regs,
 781                           struct hlist_head *head, int rctx,
 782                           struct task_struct *task);
 783 extern void perf_bp_event(struct perf_event *event, void *data);
 784
 785 #ifndef perf_misc_flags
 786 # define perf_misc_flags(regs) \
 787                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 788 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 789 #endif
 790
 791 static inline bool has_branch_stack(struct perf_event *event)
 792 {
 793         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 794 }
 795
 796 extern int perf_output_begin(struct perf_output_handle *handle,
 797                              struct perf_event *event, unsigned int size);
 798 extern void perf_output_end(struct perf_output_handle *handle);
 799 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 800                              const void *buf, unsigned int len);
 801 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 802                                      unsigned int len);
 803 extern int perf_swevent_get_recursion_context(void);
 804 extern void perf_swevent_put_recursion_context(int rctx);
 805 extern u64 perf_swevent_set_period(struct perf_event *event);
 806 extern void perf_event_enable(struct perf_event *event);
 807 extern void perf_event_disable(struct perf_event *event);
 808 extern int __perf_event_disable(void *info);
 809 extern void perf_event_task_tick(void);
 810 #else /* !CONFIG_PERF_EVENTS: */
 811 static inline void
 812 perf_event_task_sched_in(struct task_struct *prev,
 813                          struct task_struct *task)                      { }
 814 static inline void
 815 perf_event_task_sched_out(struct task_struct *prev,
 816                           struct task_struct *next)                     { }
 817 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 818 static inline void perf_event_exit_task(struct task_struct *child)      { }
 819 static inline void perf_event_free_task(struct task_struct *task)       { }
 820 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 821 static inline void perf_event_print_debug(void)                         { }
 822 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 823 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 824 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 825 {
 826         return -EINVAL;
 827 }
 828
 829 static inline void
 830 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 831 static inline void
 832 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
 833 static inline void
 834 perf_bp_event(struct perf_event *event, void *data)                     { }
 835
 836 static inline int perf_register_guest_info_callbacks
 837 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 838 static inline int perf_unregister_guest_info_callbacks
 839 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 840
 841 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 842 static inline void perf_event_exec(void)                                { }
 843 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
 844 static inline void perf_event_fork(struct task_struct *tsk)             { }
 845 static inline void perf_event_init(void)                                { }
 846 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 847 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 848 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
 849 static inline void perf_event_enable(struct perf_event *event)          { }
 850 static inline void perf_event_disable(struct perf_event *event)         { }
 851 static inline int __perf_event_disable(void *info)                      { return -1; }
 852 static inline void perf_event_task_tick(void)                           { }
 853 #endif
 854
 855 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 856 extern bool perf_event_can_stop_tick(void);
 857 #else
 858 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 859 #endif
 860
 861 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 862 extern void perf_restore_debug_store(void);
 863 #else
 864 static inline void perf_restore_debug_store(void)                       { }
 865 #endif
 866
 867 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 868
 869 /*
 870  * This has to have a higher priority than migration_notifier in sched/core.c.
 871  */
 872 #define perf_cpu_notifier(fn)                                           \
 873 do {                                                                    \
 874         static struct notifier_block fn##_nb =                          \
 875                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 876         unsigned long cpu = smp_processor_id();                         \
 877         unsigned long flags;                                            \
 878                                                                         \
 879         cpu_notifier_register_begin();                                  \
 880         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 881                 (void *)(unsigned long)cpu);                            \
 882         local_irq_save(flags);                                          \
 883         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
 884                 (void *)(unsigned long)cpu);                            \
 885         local_irq_restore(flags);                                       \
 886         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
 887                 (void *)(unsigned long)cpu);                            \
 888         __register_cpu_notifier(&fn##_nb);                              \
 889         cpu_notifier_register_done();                                   \
 890 } while (0)
 891
 892 /*
 893  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
 894  * callback for already online CPUs.
 895  */
 896 #define __perf_cpu_notifier(fn)                                         \
 897 do {                                                                    \
 898         static struct notifier_block fn##_nb =                          \
 899                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 900                                                                         \
 901         __register_cpu_notifier(&fn##_nb);                              \
 902 } while (0)
 903
 904 struct perf_pmu_events_attr {
 905         struct device_attribute attr;
 906         u64 id;
 907         const char *event_str;
 908 };
 909
 910 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
 911 static struct perf_pmu_events_attr _var = {                             \
 912         .attr = __ATTR(_name, 0444, _show, NULL),                       \
 913         .id   =  _id,                                                   \
 914 };
 915
 916 #define PMU_FORMAT_ATTR(_name, _format)                                 \
 917 static ssize_t                                                          \
 918 _name##_show(struct device *dev,                                        \
 919                                struct device_attribute *attr,           \
 920                                char *page)                              \
 921 {                                                                       \
 922         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 923         return sprintf(page, _format "\n");                             \
 924 }                                                                       \
 925                                                                         \
 926 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
 927
 928 #endif /* _LINUX_PERF_EVENT_H */