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         /*
 206          * Notification that the event was mapped or unmapped.  Called
 207          * in the context of the mapping task.
 208          */
 209         void (*event_mapped)            (struct perf_event *event); /*optional*/
 210         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 211
 212 #define PERF_EF_START   0x01            /* start the counter when adding    */
 213 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 214 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 215
 216         /*
 217          * Adds/Removes a counter to/from the PMU, can be done inside
 218          * a transaction, see the ->*_txn() methods.
 219          */
 220         int  (*add)                     (struct perf_event *event, int flags);
 221         void (*del)                     (struct perf_event *event, int flags);
 222
 223         /*
 224          * Starts/Stops a counter present on the PMU. The PMI handler
 225          * should stop the counter when perf_event_overflow() returns
 226          * !0. ->start() will be used to continue.
 227          */
 228         void (*start)                   (struct perf_event *event, int flags);
 229         void (*stop)                    (struct perf_event *event, int flags);
 230
 231         /*
 232          * Updates the counter value of the event.
 233          */
 234         void (*read)                    (struct perf_event *event);
 235
 236         /*
 237          * Group events scheduling is treated as a transaction, add
 238          * group events as a whole and perform one schedulability test.
 239          * If the test fails, roll back the whole group
 240          *
 241          * Start the transaction, after this ->add() doesn't need to
 242          * do schedulability tests.
 243          */
 244         void (*start_txn)               (struct pmu *pmu); /* optional */
 245         /*
 246          * If ->start_txn() disabled the ->add() schedulability test
 247          * then ->commit_txn() is required to perform one. On success
 248          * the transaction is closed. On error the transaction is kept
 249          * open until ->cancel_txn() is called.
 250          */
 251         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 252         /*
 253          * Will cancel the transaction, assumes ->del() is called
 254          * for each successful ->add() during the transaction.
 255          */
 256         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 257
 258         /*
 259          * Will return the value for perf_event_mmap_page::index for this event,
 260          * if no implementation is provided it will default to: event->hw.idx + 1.
 261          */
 262         int (*event_idx)                (struct perf_event *event); /*optional */
 263
 264         /*
 265          * context-switches callback
 266          */
 267         void (*sched_task)              (struct perf_event_context *ctx,
 268                                         bool sched_in);
 269         /*
 270          * PMU specific data size
 271          */
 272         size_t                          task_ctx_size;
 273
 274 };
 275
 276 /**
 277  * enum perf_event_active_state - the states of a event
 278  */
 279 enum perf_event_active_state {
 280         PERF_EVENT_STATE_EXIT           = -3,
 281         PERF_EVENT_STATE_ERROR          = -2,
 282         PERF_EVENT_STATE_OFF            = -1,
 283         PERF_EVENT_STATE_INACTIVE       =  0,
 284         PERF_EVENT_STATE_ACTIVE         =  1,
 285 };
 286
 287 struct file;
 288 struct perf_sample_data;
 289
 290 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 291                                         struct perf_sample_data *,
 292                                         struct pt_regs *regs);
 293
 294 enum perf_group_flag {
 295         PERF_GROUP_SOFTWARE             = 0x1,
 296 };
 297
 298 #define SWEVENT_HLIST_BITS              8
 299 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 300
 301 struct swevent_hlist {
 302         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 303         struct rcu_head                 rcu_head;
 304 };
 305
 306 #define PERF_ATTACH_CONTEXT     0x01
 307 #define PERF_ATTACH_GROUP       0x02
 308 #define PERF_ATTACH_TASK        0x04
 309 #define PERF_ATTACH_TASK_DATA   0x08
 310
 311 struct perf_cgroup;
 312 struct ring_buffer;
 313
 314 /**
 315  * struct perf_event - performance event kernel representation:
 316  */
 317 struct perf_event {
 318 #ifdef CONFIG_PERF_EVENTS
 319         /*
 320          * entry onto perf_event_context::event_list;
 321          *   modifications require ctx->lock
 322          *   RCU safe iterations.
 323          */
 324         struct list_head                event_entry;
 325
 326         /*
 327          * XXX: group_entry and sibling_list should be mutually exclusive;
 328          * either you're a sibling on a group, or you're the group leader.
 329          * Rework the code to always use the same list element.
 330          *
 331          * Locked for modification by both ctx->mutex and ctx->lock; holding
 332          * either sufficies for read.
 333          */
 334         struct list_head                group_entry;
 335         struct list_head                sibling_list;
 336
 337         /*
 338          * We need storage to track the entries in perf_pmu_migrate_context; we
 339          * cannot use the event_entry because of RCU and we want to keep the
 340          * group in tact which avoids us using the other two entries.
 341          */
 342         struct list_head                migrate_entry;
 343
 344         struct hlist_node               hlist_entry;
 345         struct list_head                active_entry;
 346         int                             nr_siblings;
 347         int                             group_flags;
 348         struct perf_event               *group_leader;
 349         struct pmu                      *pmu;
 350
 351         enum perf_event_active_state    state;
 352         unsigned int                    attach_state;
 353         local64_t                       count;
 354         atomic64_t                      child_count;
 355
 356         /*
 357          * These are the total time in nanoseconds that the event
 358          * has been enabled (i.e. eligible to run, and the task has
 359          * been scheduled in, if this is a per-task event)
 360          * and running (scheduled onto the CPU), respectively.
 361          *
 362          * They are computed from tstamp_enabled, tstamp_running and
 363          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 364          */
 365         u64                             total_time_enabled;
 366         u64                             total_time_running;
 367
 368         /*
 369          * These are timestamps used for computing total_time_enabled
 370          * and total_time_running when the event is in INACTIVE or
 371          * ACTIVE state, measured in nanoseconds from an arbitrary point
 372          * in time.
 373          * tstamp_enabled: the notional time when the event was enabled
 374          * tstamp_running: the notional time when the event was scheduled on
 375          * tstamp_stopped: in INACTIVE state, the notional time when the
 376          *      event was scheduled off.
 377          */
 378         u64                             tstamp_enabled;
 379         u64                             tstamp_running;
 380         u64                             tstamp_stopped;
 381
 382         /*
 383          * timestamp shadows the actual context timing but it can
 384          * be safely used in NMI interrupt context. It reflects the
 385          * context time as it was when the event was last scheduled in.
 386          *
 387          * ctx_time already accounts for ctx->timestamp. Therefore to
 388          * compute ctx_time for a sample, simply add perf_clock().
 389          */
 390         u64                             shadow_ctx_time;
 391
 392         struct perf_event_attr          attr;
 393         u16                             header_size;
 394         u16                             id_header_size;
 395         u16                             read_size;
 396         struct hw_perf_event            hw;
 397
 398         struct perf_event_context       *ctx;
 399         atomic_long_t                   refcount;
 400
 401         /*
 402          * These accumulate total time (in nanoseconds) that children
 403          * events have been enabled and running, respectively.
 404          */
 405         atomic64_t                      child_total_time_enabled;
 406         atomic64_t                      child_total_time_running;
 407
 408         /*
 409          * Protect attach/detach and child_list:
 410          */
 411         struct mutex                    child_mutex;
 412         struct list_head                child_list;
 413         struct perf_event               *parent;
 414
 415         int                             oncpu;
 416         int                             cpu;
 417
 418         struct list_head                owner_entry;
 419         struct task_struct              *owner;
 420
 421         /* mmap bits */
 422         struct mutex                    mmap_mutex;
 423         atomic_t                        mmap_count;
 424
 425         struct ring_buffer              *rb;
 426         struct list_head                rb_entry;
 427         unsigned long                   rcu_batches;
 428         int                             rcu_pending;
 429
 430         /* poll related */
 431         wait_queue_head_t               waitq;
 432         struct fasync_struct            *fasync;
 433
 434         /* delayed work for NMIs and such */
 435         int                             pending_wakeup;
 436         int                             pending_kill;
 437         int                             pending_disable;
 438         struct irq_work                 pending;
 439
 440         atomic_t                        event_limit;
 441
 442         void (*destroy)(struct perf_event *);
 443         struct rcu_head                 rcu_head;
 444
 445         struct pid_namespace            *ns;
 446         u64                             id;
 447
 448         perf_overflow_handler_t         overflow_handler;
 449         void                            *overflow_handler_context;
 450
 451 #ifdef CONFIG_EVENT_TRACING
 452         struct ftrace_event_call        *tp_event;
 453         struct event_filter             *filter;
 454 #ifdef CONFIG_FUNCTION_TRACER
 455         struct ftrace_ops               ftrace_ops;
 456 #endif
 457 #endif
 458
 459 #ifdef CONFIG_CGROUP_PERF
 460         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 461         int                             cgrp_defer_enabled;
 462 #endif
 463
 464 #endif /* CONFIG_PERF_EVENTS */
 465 };
 466
 467 /**
 468  * struct perf_event_context - event context structure
 469  *
 470  * Used as a container for task events and CPU events as well:
 471  */
 472 struct perf_event_context {
 473         struct pmu                      *pmu;
 474         /*
 475          * Protect the states of the events in the list,
 476          * nr_active, and the list:
 477          */
 478         raw_spinlock_t                  lock;
 479         /*
 480          * Protect the list of events.  Locking either mutex or lock
 481          * is sufficient to ensure the list doesn't change; to change
 482          * the list you need to lock both the mutex and the spinlock.
 483          */
 484         struct mutex                    mutex;
 485
 486         struct list_head                active_ctx_list;
 487         struct list_head                pinned_groups;
 488         struct list_head                flexible_groups;
 489         struct list_head                event_list;
 490         int                             nr_events;
 491         int                             nr_active;
 492         int                             is_active;
 493         int                             nr_stat;
 494         int                             nr_freq;
 495         int                             rotate_disable;
 496         atomic_t                        refcount;
 497         struct task_struct              *task;
 498
 499         /*
 500          * Context clock, runs when context enabled.
 501          */
 502         u64                             time;
 503         u64                             timestamp;
 504
 505         /*
 506          * These fields let us detect when two contexts have both
 507          * been cloned (inherited) from a common ancestor.
 508          */
 509         struct perf_event_context       *parent_ctx;
 510         u64                             parent_gen;
 511         u64                             generation;
 512         int                             pin_count;
 513         int                             nr_cgroups;      /* cgroup evts */
 514         void                            *task_ctx_data; /* pmu specific data */
 515         struct rcu_head                 rcu_head;
 516
 517         struct delayed_work             orphans_remove;
 518         bool                            orphans_remove_sched;
 519 };
 520
 521 /*
 522  * Number of contexts where an event can trigger:
 523  *      task, softirq, hardirq, nmi.
 524  */
 525 #define PERF_NR_CONTEXTS        4
 526
 527 /**
 528  * struct perf_event_cpu_context - per cpu event context structure
 529  */
 530 struct perf_cpu_context {
 531         struct perf_event_context       ctx;
 532         struct perf_event_context       *task_ctx;
 533         int                             active_oncpu;
 534         int                             exclusive;
 535         struct hrtimer                  hrtimer;
 536         ktime_t                         hrtimer_interval;
 537         struct pmu                      *unique_pmu;
 538         struct perf_cgroup              *cgrp;
 539 };
 540
 541 struct perf_output_handle {
 542         struct perf_event               *event;
 543         struct ring_buffer              *rb;
 544         unsigned long                   wakeup;
 545         unsigned long                   size;
 546         void                            *addr;
 547         int                             page;
 548 };
 549
 550 #ifdef CONFIG_PERF_EVENTS
 551
 552 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 553 extern void perf_pmu_unregister(struct pmu *pmu);
 554
 555 extern int perf_num_counters(void);
 556 extern const char *perf_pmu_name(void);
 557 extern void __perf_event_task_sched_in(struct task_struct *prev,
 558                                        struct task_struct *task);
 559 extern void __perf_event_task_sched_out(struct task_struct *prev,
 560                                         struct task_struct *next);
 561 extern int perf_event_init_task(struct task_struct *child);
 562 extern void perf_event_exit_task(struct task_struct *child);
 563 extern void perf_event_free_task(struct task_struct *task);
 564 extern void perf_event_delayed_put(struct task_struct *task);
 565 extern void perf_event_print_debug(void);
 566 extern void perf_pmu_disable(struct pmu *pmu);
 567 extern void perf_pmu_enable(struct pmu *pmu);
 568 extern void perf_sched_cb_dec(struct pmu *pmu);
 569 extern void perf_sched_cb_inc(struct pmu *pmu);
 570 extern int perf_event_task_disable(void);
 571 extern int perf_event_task_enable(void);
 572 extern int perf_event_refresh(struct perf_event *event, int refresh);
 573 extern void perf_event_update_userpage(struct perf_event *event);
 574 extern int perf_event_release_kernel(struct perf_event *event);
 575 extern struct perf_event *
 576 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 577                                 int cpu,
 578                                 struct task_struct *task,
 579                                 perf_overflow_handler_t callback,
 580                                 void *context);
 581 extern void perf_pmu_migrate_context(struct pmu *pmu,
 582                                 int src_cpu, int dst_cpu);
 583 extern u64 perf_event_read_value(struct perf_event *event,
 584                                  u64 *enabled, u64 *running);
 585
 586
 587 struct perf_sample_data {
 588         /*
 589          * Fields set by perf_sample_data_init(), group so as to
 590          * minimize the cachelines touched.
 591          */
 592         u64                             addr;
 593         struct perf_raw_record          *raw;
 594         struct perf_branch_stack        *br_stack;
 595         u64                             period;
 596         u64                             weight;
 597         u64                             txn;
 598         union  perf_mem_data_src        data_src;
 599
 600         /*
 601          * The other fields, optionally {set,used} by
 602          * perf_{prepare,output}_sample().
 603          */
 604         u64                             type;
 605         u64                             ip;
 606         struct {
 607                 u32     pid;
 608                 u32     tid;
 609         }                               tid_entry;
 610         u64                             time;
 611         u64                             id;
 612         u64                             stream_id;
 613         struct {
 614                 u32     cpu;
 615                 u32     reserved;
 616         }                               cpu_entry;
 617         struct perf_callchain_entry     *callchain;
 618
 619         /*
 620          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 621          * on arch details.
 622          */
 623         struct perf_regs                regs_user;
 624         struct pt_regs                  regs_user_copy;
 625
 626         struct perf_regs                regs_intr;
 627         u64                             stack_user_size;
 628 } ____cacheline_aligned;
 629
 630 /* default value for data source */
 631 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 632                     PERF_MEM_S(LVL, NA)   |\
 633                     PERF_MEM_S(SNOOP, NA) |\
 634                     PERF_MEM_S(LOCK, NA)  |\
 635                     PERF_MEM_S(TLB, NA))
 636
 637 static inline void perf_sample_data_init(struct perf_sample_data *data,
 638                                          u64 addr, u64 period)
 639 {
 640         /* remaining struct members initialized in perf_prepare_sample() */
 641         data->addr = addr;
 642         data->raw  = NULL;
 643         data->br_stack = NULL;
 644         data->period = period;
 645         data->weight = 0;
 646         data->data_src.val = PERF_MEM_NA;
 647         data->txn = 0;
 648 }
 649
 650 extern void perf_output_sample(struct perf_output_handle *handle,
 651                                struct perf_event_header *header,
 652                                struct perf_sample_data *data,
 653                                struct perf_event *event);
 654 extern void perf_prepare_sample(struct perf_event_header *header,
 655                                 struct perf_sample_data *data,
 656                                 struct perf_event *event,
 657                                 struct pt_regs *regs);
 658
 659 extern int perf_event_overflow(struct perf_event *event,
 660                                  struct perf_sample_data *data,
 661                                  struct pt_regs *regs);
 662
 663 static inline bool is_sampling_event(struct perf_event *event)
 664 {
 665         return event->attr.sample_period != 0;
 666 }
 667
 668 /*
 669  * Return 1 for a software event, 0 for a hardware event
 670  */
 671 static inline int is_software_event(struct perf_event *event)
 672 {
 673         return event->pmu->task_ctx_nr == perf_sw_context;
 674 }
 675
 676 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 677
 678 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 679 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 680
 681 #ifndef perf_arch_fetch_caller_regs
 682 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 683 #endif
 684
 685 /*
 686  * Take a snapshot of the regs. Skip ip and frame pointer to
 687  * the nth caller. We only need a few of the regs:
 688  * - ip for PERF_SAMPLE_IP
 689  * - cs for user_mode() tests
 690  * - bp for callchains
 691  * - eflags, for future purposes, just in case
 692  */
 693 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 694 {
 695         memset(regs, 0, sizeof(*regs));
 696
 697         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 698 }
 699
 700 static __always_inline void
 701 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 702 {
 703         if (static_key_false(&perf_swevent_enabled[event_id]))
 704                 __perf_sw_event(event_id, nr, regs, addr);
 705 }
 706
 707 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 708
 709 /*
 710  * 'Special' version for the scheduler, it hard assumes no recursion,
 711  * which is guaranteed by us not actually scheduling inside other swevents
 712  * because those disable preemption.
 713  */
 714 static __always_inline void
 715 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 716 {
 717         if (static_key_false(&perf_swevent_enabled[event_id])) {
 718                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 719
 720                 perf_fetch_caller_regs(regs);
 721                 ___perf_sw_event(event_id, nr, regs, addr);
 722         }
 723 }
 724
 725 extern struct static_key_deferred perf_sched_events;
 726
 727 static inline void perf_event_task_sched_in(struct task_struct *prev,
 728                                             struct task_struct *task)
 729 {
 730         if (static_key_false(&perf_sched_events.key))
 731                 __perf_event_task_sched_in(prev, task);
 732 }
 733
 734 static inline void perf_event_task_sched_out(struct task_struct *prev,
 735                                              struct task_struct *next)
 736 {
 737         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 738
 739         if (static_key_false(&perf_sched_events.key))
 740                 __perf_event_task_sched_out(prev, next);
 741 }
 742
 743 extern void perf_event_mmap(struct vm_area_struct *vma);
 744 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 745 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 746 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 747
 748 extern void perf_event_exec(void);
 749 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 750 extern void perf_event_fork(struct task_struct *tsk);
 751
 752 /* Callchains */
 753 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 754
 755 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 756 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 757
 758 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 759 {
 760         if (entry->nr < PERF_MAX_STACK_DEPTH)
 761                 entry->ip[entry->nr++] = ip;
 762 }
 763
 764 extern int sysctl_perf_event_paranoid;
 765 extern int sysctl_perf_event_mlock;
 766 extern int sysctl_perf_event_sample_rate;
 767 extern int sysctl_perf_cpu_time_max_percent;
 768
 769 extern void perf_sample_event_took(u64 sample_len_ns);
 770
 771 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 772                 void __user *buffer, size_t *lenp,
 773                 loff_t *ppos);
 774 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 775                 void __user *buffer, size_t *lenp,
 776                 loff_t *ppos);
 777
 778
 779 static inline bool perf_paranoid_tracepoint_raw(void)
 780 {
 781         return sysctl_perf_event_paranoid > -1;
 782 }
 783
 784 static inline bool perf_paranoid_cpu(void)
 785 {
 786         return sysctl_perf_event_paranoid > 0;
 787 }
 788
 789 static inline bool perf_paranoid_kernel(void)
 790 {
 791         return sysctl_perf_event_paranoid > 1;
 792 }
 793
 794 extern void perf_event_init(void);
 795 extern void perf_tp_event(u64 addr, u64 count, void *record,
 796                           int entry_size, struct pt_regs *regs,
 797                           struct hlist_head *head, int rctx,
 798                           struct task_struct *task);
 799 extern void perf_bp_event(struct perf_event *event, void *data);
 800
 801 #ifndef perf_misc_flags
 802 # define perf_misc_flags(regs) \
 803                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 804 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 805 #endif
 806
 807 static inline bool has_branch_stack(struct perf_event *event)
 808 {
 809         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 810 }
 811
 812 static inline bool needs_branch_stack(struct perf_event *event)
 813 {
 814         return event->attr.branch_sample_type != 0;
 815 }
 816
 817 extern int perf_output_begin(struct perf_output_handle *handle,
 818                              struct perf_event *event, unsigned int size);
 819 extern void perf_output_end(struct perf_output_handle *handle);
 820 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 821                              const void *buf, unsigned int len);
 822 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 823                                      unsigned int len);
 824 extern int perf_swevent_get_recursion_context(void);
 825 extern void perf_swevent_put_recursion_context(int rctx);
 826 extern u64 perf_swevent_set_period(struct perf_event *event);
 827 extern void perf_event_enable(struct perf_event *event);
 828 extern void perf_event_disable(struct perf_event *event);
 829 extern int __perf_event_disable(void *info);
 830 extern void perf_event_task_tick(void);
 831 #else /* !CONFIG_PERF_EVENTS: */
 832 static inline void
 833 perf_event_task_sched_in(struct task_struct *prev,
 834                          struct task_struct *task)                      { }
 835 static inline void
 836 perf_event_task_sched_out(struct task_struct *prev,
 837                           struct task_struct *next)                     { }
 838 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 839 static inline void perf_event_exit_task(struct task_struct *child)      { }
 840 static inline void perf_event_free_task(struct task_struct *task)       { }
 841 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 842 static inline void perf_event_print_debug(void)                         { }
 843 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 844 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 845 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 846 {
 847         return -EINVAL;
 848 }
 849
 850 static inline void
 851 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 852 static inline void
 853 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
 854 static inline void
 855 perf_bp_event(struct perf_event *event, void *data)                     { }
 856
 857 static inline int perf_register_guest_info_callbacks
 858 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 859 static inline int perf_unregister_guest_info_callbacks
 860 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 861
 862 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 863 static inline void perf_event_exec(void)                                { }
 864 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
 865 static inline void perf_event_fork(struct task_struct *tsk)             { }
 866 static inline void perf_event_init(void)                                { }
 867 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 868 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 869 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
 870 static inline void perf_event_enable(struct perf_event *event)          { }
 871 static inline void perf_event_disable(struct perf_event *event)         { }
 872 static inline int __perf_event_disable(void *info)                      { return -1; }
 873 static inline void perf_event_task_tick(void)                           { }
 874 #endif
 875
 876 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 877 extern bool perf_event_can_stop_tick(void);
 878 #else
 879 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 880 #endif
 881
 882 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 883 extern void perf_restore_debug_store(void);
 884 #else
 885 static inline void perf_restore_debug_store(void)                       { }
 886 #endif
 887
 888 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 889
 890 /*
 891  * This has to have a higher priority than migration_notifier in sched/core.c.
 892  */
 893 #define perf_cpu_notifier(fn)                                           \
 894 do {                                                                    \
 895         static struct notifier_block fn##_nb =                          \
 896                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 897         unsigned long cpu = smp_processor_id();                         \
 898         unsigned long flags;                                            \
 899                                                                         \
 900         cpu_notifier_register_begin();                                  \
 901         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 902                 (void *)(unsigned long)cpu);                            \
 903         local_irq_save(flags);                                          \
 904         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
 905                 (void *)(unsigned long)cpu);                            \
 906         local_irq_restore(flags);                                       \
 907         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
 908                 (void *)(unsigned long)cpu);                            \
 909         __register_cpu_notifier(&fn##_nb);                              \
 910         cpu_notifier_register_done();                                   \
 911 } while (0)
 912
 913 /*
 914  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
 915  * callback for already online CPUs.
 916  */
 917 #define __perf_cpu_notifier(fn)                                         \
 918 do {                                                                    \
 919         static struct notifier_block fn##_nb =                          \
 920                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 921                                                                         \
 922         __register_cpu_notifier(&fn##_nb);                              \
 923 } while (0)
 924
 925 struct perf_pmu_events_attr {
 926         struct device_attribute attr;
 927         u64 id;
 928         const char *event_str;
 929 };
 930
 931 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
 932                               char *page);
 933
 934 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
 935 static struct perf_pmu_events_attr _var = {                             \
 936         .attr = __ATTR(_name, 0444, _show, NULL),                       \
 937         .id   =  _id,                                                   \
 938 };
 939
 940 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
 941 static struct perf_pmu_events_attr _var = {                                 \
 942         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
 943         .id             = 0,                                                \
 944         .event_str      = _str,                                             \
 945 };
 946
 947 #define PMU_FORMAT_ATTR(_name, _format)                                 \
 948 static ssize_t                                                          \
 949 _name##_show(struct device *dev,                                        \
 950                                struct device_attribute *attr,           \
 951                                char *page)                              \
 952 {                                                                       \
 953         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 954         return sprintf(page, _format "\n");                             \
 955 }                                                                       \
 956                                                                         \
 957 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
 958
 959 #endif /* _LINUX_PERF_EVENT_H */