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 <linux/cgroup.h>
  57 #include <asm/local.h>
  58
  59 struct perf_callchain_entry {
  60         __u64                           nr;
  61         __u64                           ip[PERF_MAX_STACK_DEPTH];
  62 };
  63
  64 struct perf_raw_record {
  65         u32                             size;
  66         void                            *data;
  67 };
  68
  69 /*
  70  * branch stack layout:
  71  *  nr: number of taken branches stored in entries[]
  72  *
  73  * Note that nr can vary from sample to sample
  74  * branches (to, from) are stored from most recent
  75  * to least recent, i.e., entries[0] contains the most
  76  * recent branch.
  77  */
  78 struct perf_branch_stack {
  79         __u64                           nr;
  80         struct perf_branch_entry        entries[0];
  81 };
  82
  83 struct task_struct;
  84
  85 /*
  86  * extra PMU register associated with an event
  87  */
  88 struct hw_perf_event_extra {
  89         u64             config; /* register value */
  90         unsigned int    reg;    /* register address or index */
  91         int             alloc;  /* extra register already allocated */
  92         int             idx;    /* index in shared_regs->regs[] */
  93 };
  94
  95 /**
  96  * struct hw_perf_event - performance event hardware details:
  97  */
  98 struct hw_perf_event {
  99 #ifdef CONFIG_PERF_EVENTS
 100         union {
 101                 struct { /* hardware */
 102                         u64             config;
 103                         u64             last_tag;
 104                         unsigned long   config_base;
 105                         unsigned long   event_base;
 106                         int             event_base_rdpmc;
 107                         int             idx;
 108                         int             last_cpu;
 109                         int             flags;
 110
 111                         struct hw_perf_event_extra extra_reg;
 112                         struct hw_perf_event_extra branch_reg;
 113                 };
 114                 struct { /* software */
 115                         struct hrtimer  hrtimer;
 116                 };
 117                 struct { /* tracepoint */
 118                         /* for tp_event->class */
 119                         struct list_head        tp_list;
 120                 };
 121                 struct { /* intel_cqm */
 122                         int                     cqm_state;
 123                         u32                     cqm_rmid;
 124                         struct list_head        cqm_events_entry;
 125                         struct list_head        cqm_groups_entry;
 126                         struct list_head        cqm_group_entry;
 127                 };
 128                 struct { /* itrace */
 129                         int                     itrace_started;
 130                 };
 131 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 132                 struct { /* breakpoint */
 133                         /*
 134                          * Crufty hack to avoid the chicken and egg
 135                          * problem hw_breakpoint has with context
 136                          * creation and event initalization.
 137                          */
 138                         struct arch_hw_breakpoint       info;
 139                         struct list_head                bp_list;
 140                 };
 141 #endif
 142         };
 143         struct task_struct              *target;
 144         int                             state;
 145         local64_t                       prev_count;
 146         u64                             sample_period;
 147         u64                             last_period;
 148         local64_t                       period_left;
 149         u64                             interrupts_seq;
 150         u64                             interrupts;
 151
 152         u64                             freq_time_stamp;
 153         u64                             freq_count_stamp;
 154 #endif
 155 };
 156
 157 /*
 158  * hw_perf_event::state flags
 159  */
 160 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 161 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 162 #define PERF_HES_ARCH           0x04
 163
 164 struct perf_event;
 165
 166 /*
 167  * Common implementation detail of pmu::{start,commit,cancel}_txn
 168  */
 169 #define PERF_EVENT_TXN 0x1
 170
 171 /**
 172  * pmu::capabilities flags
 173  */
 174 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 175 #define PERF_PMU_CAP_NO_NMI                     0x02
 176 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
 177 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
 178 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
 179 #define PERF_PMU_CAP_ITRACE                     0x20
 180
 181 /**
 182  * struct pmu - generic performance monitoring unit
 183  */
 184 struct pmu {
 185         struct list_head                entry;
 186
 187         struct module                   *module;
 188         struct device                   *dev;
 189         const struct attribute_group    **attr_groups;
 190         const char                      *name;
 191         int                             type;
 192
 193         /*
 194          * various common per-pmu feature flags
 195          */
 196         int                             capabilities;
 197
 198         int * __percpu                  pmu_disable_count;
 199         struct perf_cpu_context * __percpu pmu_cpu_context;
 200         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 201         int                             task_ctx_nr;
 202         int                             hrtimer_interval_ms;
 203
 204         /*
 205          * Fully disable/enable this PMU, can be used to protect from the PMI
 206          * as well as for lazy/batch writing of the MSRs.
 207          */
 208         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 209         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 210
 211         /*
 212          * Try and initialize the event for this PMU.
 213          * Should return -ENOENT when the @event doesn't match this PMU.
 214          */
 215         int (*event_init)               (struct perf_event *event);
 216
 217         /*
 218          * Notification that the event was mapped or unmapped.  Called
 219          * in the context of the mapping task.
 220          */
 221         void (*event_mapped)            (struct perf_event *event); /*optional*/
 222         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 223
 224 #define PERF_EF_START   0x01            /* start the counter when adding    */
 225 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 226 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 227
 228         /*
 229          * Adds/Removes a counter to/from the PMU, can be done inside
 230          * a transaction, see the ->*_txn() methods.
 231          */
 232         int  (*add)                     (struct perf_event *event, int flags);
 233         void (*del)                     (struct perf_event *event, int flags);
 234
 235         /*
 236          * Starts/Stops a counter present on the PMU. The PMI handler
 237          * should stop the counter when perf_event_overflow() returns
 238          * !0. ->start() will be used to continue.
 239          */
 240         void (*start)                   (struct perf_event *event, int flags);
 241         void (*stop)                    (struct perf_event *event, int flags);
 242
 243         /*
 244          * Updates the counter value of the event.
 245          */
 246         void (*read)                    (struct perf_event *event);
 247
 248         /*
 249          * Group events scheduling is treated as a transaction, add
 250          * group events as a whole and perform one schedulability test.
 251          * If the test fails, roll back the whole group
 252          *
 253          * Start the transaction, after this ->add() doesn't need to
 254          * do schedulability tests.
 255          */
 256         void (*start_txn)               (struct pmu *pmu); /* optional */
 257         /*
 258          * If ->start_txn() disabled the ->add() schedulability test
 259          * then ->commit_txn() is required to perform one. On success
 260          * the transaction is closed. On error the transaction is kept
 261          * open until ->cancel_txn() is called.
 262          */
 263         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 264         /*
 265          * Will cancel the transaction, assumes ->del() is called
 266          * for each successful ->add() during the transaction.
 267          */
 268         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 269
 270         /*
 271          * Will return the value for perf_event_mmap_page::index for this event,
 272          * if no implementation is provided it will default to: event->hw.idx + 1.
 273          */
 274         int (*event_idx)                (struct perf_event *event); /*optional */
 275
 276         /*
 277          * context-switches callback
 278          */
 279         void (*sched_task)              (struct perf_event_context *ctx,
 280                                         bool sched_in);
 281         /*
 282          * PMU specific data size
 283          */
 284         size_t                          task_ctx_size;
 285
 286
 287         /*
 288          * Return the count value for a counter.
 289          */
 290         u64 (*count)                    (struct perf_event *event); /*optional*/
 291
 292         /*
 293          * Set up pmu-private data structures for an AUX area
 294          */
 295         void *(*setup_aux)              (int cpu, void **pages,
 296                                          int nr_pages, bool overwrite);
 297                                         /* optional */
 298
 299         /*
 300          * Free pmu-private AUX data structures
 301          */
 302         void (*free_aux)                (void *aux); /* optional */
 303
 304         /*
 305          * Filter events for PMU-specific reasons.
 306          */
 307         int (*filter_match)             (struct perf_event *event); /* optional */
 308 };
 309
 310 /**
 311  * enum perf_event_active_state - the states of a event
 312  */
 313 enum perf_event_active_state {
 314         PERF_EVENT_STATE_EXIT           = -3,
 315         PERF_EVENT_STATE_ERROR          = -2,
 316         PERF_EVENT_STATE_OFF            = -1,
 317         PERF_EVENT_STATE_INACTIVE       =  0,
 318         PERF_EVENT_STATE_ACTIVE         =  1,
 319 };
 320
 321 struct file;
 322 struct perf_sample_data;
 323
 324 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 325                                         struct perf_sample_data *,
 326                                         struct pt_regs *regs);
 327
 328 enum perf_group_flag {
 329         PERF_GROUP_SOFTWARE             = 0x1,
 330 };
 331
 332 #define SWEVENT_HLIST_BITS              8
 333 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 334
 335 struct swevent_hlist {
 336         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 337         struct rcu_head                 rcu_head;
 338 };
 339
 340 #define PERF_ATTACH_CONTEXT     0x01
 341 #define PERF_ATTACH_GROUP       0x02
 342 #define PERF_ATTACH_TASK        0x04
 343 #define PERF_ATTACH_TASK_DATA   0x08
 344
 345 struct perf_cgroup;
 346 struct ring_buffer;
 347
 348 /**
 349  * struct perf_event - performance event kernel representation:
 350  */
 351 struct perf_event {
 352 #ifdef CONFIG_PERF_EVENTS
 353         /*
 354          * entry onto perf_event_context::event_list;
 355          *   modifications require ctx->lock
 356          *   RCU safe iterations.
 357          */
 358         struct list_head                event_entry;
 359
 360         /*
 361          * XXX: group_entry and sibling_list should be mutually exclusive;
 362          * either you're a sibling on a group, or you're the group leader.
 363          * Rework the code to always use the same list element.
 364          *
 365          * Locked for modification by both ctx->mutex and ctx->lock; holding
 366          * either sufficies for read.
 367          */
 368         struct list_head                group_entry;
 369         struct list_head                sibling_list;
 370
 371         /*
 372          * We need storage to track the entries in perf_pmu_migrate_context; we
 373          * cannot use the event_entry because of RCU and we want to keep the
 374          * group in tact which avoids us using the other two entries.
 375          */
 376         struct list_head                migrate_entry;
 377
 378         struct hlist_node               hlist_entry;
 379         struct list_head                active_entry;
 380         int                             nr_siblings;
 381         int                             group_flags;
 382         struct perf_event               *group_leader;
 383         struct pmu                      *pmu;
 384
 385         enum perf_event_active_state    state;
 386         unsigned int                    attach_state;
 387         local64_t                       count;
 388         atomic64_t                      child_count;
 389
 390         /*
 391          * These are the total time in nanoseconds that the event
 392          * has been enabled (i.e. eligible to run, and the task has
 393          * been scheduled in, if this is a per-task event)
 394          * and running (scheduled onto the CPU), respectively.
 395          *
 396          * They are computed from tstamp_enabled, tstamp_running and
 397          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 398          */
 399         u64                             total_time_enabled;
 400         u64                             total_time_running;
 401
 402         /*
 403          * These are timestamps used for computing total_time_enabled
 404          * and total_time_running when the event is in INACTIVE or
 405          * ACTIVE state, measured in nanoseconds from an arbitrary point
 406          * in time.
 407          * tstamp_enabled: the notional time when the event was enabled
 408          * tstamp_running: the notional time when the event was scheduled on
 409          * tstamp_stopped: in INACTIVE state, the notional time when the
 410          *      event was scheduled off.
 411          */
 412         u64                             tstamp_enabled;
 413         u64                             tstamp_running;
 414         u64                             tstamp_stopped;
 415
 416         /*
 417          * timestamp shadows the actual context timing but it can
 418          * be safely used in NMI interrupt context. It reflects the
 419          * context time as it was when the event was last scheduled in.
 420          *
 421          * ctx_time already accounts for ctx->timestamp. Therefore to
 422          * compute ctx_time for a sample, simply add perf_clock().
 423          */
 424         u64                             shadow_ctx_time;
 425
 426         struct perf_event_attr          attr;
 427         u16                             header_size;
 428         u16                             id_header_size;
 429         u16                             read_size;
 430         struct hw_perf_event            hw;
 431
 432         struct perf_event_context       *ctx;
 433         atomic_long_t                   refcount;
 434
 435         /*
 436          * These accumulate total time (in nanoseconds) that children
 437          * events have been enabled and running, respectively.
 438          */
 439         atomic64_t                      child_total_time_enabled;
 440         atomic64_t                      child_total_time_running;
 441
 442         /*
 443          * Protect attach/detach and child_list:
 444          */
 445         struct mutex                    child_mutex;
 446         struct list_head                child_list;
 447         struct perf_event               *parent;
 448
 449         int                             oncpu;
 450         int                             cpu;
 451
 452         struct list_head                owner_entry;
 453         struct task_struct              *owner;
 454
 455         /* mmap bits */
 456         struct mutex                    mmap_mutex;
 457         atomic_t                        mmap_count;
 458
 459         struct ring_buffer              *rb;
 460         struct list_head                rb_entry;
 461         unsigned long                   rcu_batches;
 462         int                             rcu_pending;
 463
 464         /* poll related */
 465         wait_queue_head_t               waitq;
 466         struct fasync_struct            *fasync;
 467
 468         /* delayed work for NMIs and such */
 469         int                             pending_wakeup;
 470         int                             pending_kill;
 471         int                             pending_disable;
 472         struct irq_work                 pending;
 473
 474         atomic_t                        event_limit;
 475
 476         void (*destroy)(struct perf_event *);
 477         struct rcu_head                 rcu_head;
 478
 479         struct pid_namespace            *ns;
 480         u64                             id;
 481
 482         u64                             (*clock)(void);
 483         perf_overflow_handler_t         overflow_handler;
 484         void                            *overflow_handler_context;
 485
 486 #ifdef CONFIG_EVENT_TRACING
 487         struct trace_event_call         *tp_event;
 488         struct event_filter             *filter;
 489 #ifdef CONFIG_FUNCTION_TRACER
 490         struct ftrace_ops               ftrace_ops;
 491 #endif
 492 #endif
 493
 494 #ifdef CONFIG_CGROUP_PERF
 495         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 496         int                             cgrp_defer_enabled;
 497 #endif
 498
 499 #endif /* CONFIG_PERF_EVENTS */
 500 };
 501
 502 /**
 503  * struct perf_event_context - event context structure
 504  *
 505  * Used as a container for task events and CPU events as well:
 506  */
 507 struct perf_event_context {
 508         struct pmu                      *pmu;
 509         /*
 510          * Protect the states of the events in the list,
 511          * nr_active, and the list:
 512          */
 513         raw_spinlock_t                  lock;
 514         /*
 515          * Protect the list of events.  Locking either mutex or lock
 516          * is sufficient to ensure the list doesn't change; to change
 517          * the list you need to lock both the mutex and the spinlock.
 518          */
 519         struct mutex                    mutex;
 520
 521         struct list_head                active_ctx_list;
 522         struct list_head                pinned_groups;
 523         struct list_head                flexible_groups;
 524         struct list_head                event_list;
 525         int                             nr_events;
 526         int                             nr_active;
 527         int                             is_active;
 528         int                             nr_stat;
 529         int                             nr_freq;
 530         int                             rotate_disable;
 531         atomic_t                        refcount;
 532         struct task_struct              *task;
 533
 534         /*
 535          * Context clock, runs when context enabled.
 536          */
 537         u64                             time;
 538         u64                             timestamp;
 539
 540         /*
 541          * These fields let us detect when two contexts have both
 542          * been cloned (inherited) from a common ancestor.
 543          */
 544         struct perf_event_context       *parent_ctx;
 545         u64                             parent_gen;
 546         u64                             generation;
 547         int                             pin_count;
 548         int                             nr_cgroups;      /* cgroup evts */
 549         void                            *task_ctx_data; /* pmu specific data */
 550         struct rcu_head                 rcu_head;
 551
 552         struct delayed_work             orphans_remove;
 553         bool                            orphans_remove_sched;
 554 };
 555
 556 /*
 557  * Number of contexts where an event can trigger:
 558  *      task, softirq, hardirq, nmi.
 559  */
 560 #define PERF_NR_CONTEXTS        4
 561
 562 /**
 563  * struct perf_event_cpu_context - per cpu event context structure
 564  */
 565 struct perf_cpu_context {
 566         struct perf_event_context       ctx;
 567         struct perf_event_context       *task_ctx;
 568         int                             active_oncpu;
 569         int                             exclusive;
 570
 571         raw_spinlock_t                  hrtimer_lock;
 572         struct hrtimer                  hrtimer;
 573         ktime_t                         hrtimer_interval;
 574         unsigned int                    hrtimer_active;
 575
 576         struct pmu                      *unique_pmu;
 577         struct perf_cgroup              *cgrp;
 578 };
 579
 580 struct perf_output_handle {
 581         struct perf_event               *event;
 582         struct ring_buffer              *rb;
 583         unsigned long                   wakeup;
 584         unsigned long                   size;
 585         union {
 586                 void                    *addr;
 587                 unsigned long           head;
 588         };
 589         int                             page;
 590 };
 591
 592 #ifdef CONFIG_CGROUP_PERF
 593
 594 /*
 595  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 596  * This is a per-cpu dynamically allocated data structure.
 597  */
 598 struct perf_cgroup_info {
 599         u64                             time;
 600         u64                             timestamp;
 601 };
 602
 603 struct perf_cgroup {
 604         struct cgroup_subsys_state      css;
 605         struct perf_cgroup_info __percpu *info;
 606 };
 607
 608 /*
 609  * Must ensure cgroup is pinned (css_get) before calling
 610  * this function. In other words, we cannot call this function
 611  * if there is no cgroup event for the current CPU context.
 612  */
 613 static inline struct perf_cgroup *
 614 perf_cgroup_from_task(struct task_struct *task)
 615 {
 616         return container_of(task_css(task, perf_event_cgrp_id),
 617                             struct perf_cgroup, css);
 618 }
 619 #endif /* CONFIG_CGROUP_PERF */
 620
 621 #ifdef CONFIG_PERF_EVENTS
 622
 623 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 624                                    struct perf_event *event);
 625 extern void perf_aux_output_end(struct perf_output_handle *handle,
 626                                 unsigned long size, bool truncated);
 627 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 628                                 unsigned long size);
 629 extern void *perf_get_aux(struct perf_output_handle *handle);
 630
 631 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 632 extern void perf_pmu_unregister(struct pmu *pmu);
 633
 634 extern int perf_num_counters(void);
 635 extern const char *perf_pmu_name(void);
 636 extern void __perf_event_task_sched_in(struct task_struct *prev,
 637                                        struct task_struct *task);
 638 extern void __perf_event_task_sched_out(struct task_struct *prev,
 639                                         struct task_struct *next);
 640 extern int perf_event_init_task(struct task_struct *child);
 641 extern void perf_event_exit_task(struct task_struct *child);
 642 extern void perf_event_free_task(struct task_struct *task);
 643 extern void perf_event_delayed_put(struct task_struct *task);
 644 extern struct perf_event *perf_event_get(unsigned int fd);
 645 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 646 extern void perf_event_print_debug(void);
 647 extern void perf_pmu_disable(struct pmu *pmu);
 648 extern void perf_pmu_enable(struct pmu *pmu);
 649 extern void perf_sched_cb_dec(struct pmu *pmu);
 650 extern void perf_sched_cb_inc(struct pmu *pmu);
 651 extern int perf_event_task_disable(void);
 652 extern int perf_event_task_enable(void);
 653 extern int perf_event_refresh(struct perf_event *event, int refresh);
 654 extern void perf_event_update_userpage(struct perf_event *event);
 655 extern int perf_event_release_kernel(struct perf_event *event);
 656 extern struct perf_event *
 657 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 658                                 int cpu,
 659                                 struct task_struct *task,
 660                                 perf_overflow_handler_t callback,
 661                                 void *context);
 662 extern void perf_pmu_migrate_context(struct pmu *pmu,
 663                                 int src_cpu, int dst_cpu);
 664 extern u64 perf_event_read_local(struct perf_event *event);
 665 extern u64 perf_event_read_value(struct perf_event *event,
 666                                  u64 *enabled, u64 *running);
 667
 668
 669 struct perf_sample_data {
 670         /*
 671          * Fields set by perf_sample_data_init(), group so as to
 672          * minimize the cachelines touched.
 673          */
 674         u64                             addr;
 675         struct perf_raw_record          *raw;
 676         struct perf_branch_stack        *br_stack;
 677         u64                             period;
 678         u64                             weight;
 679         u64                             txn;
 680         union  perf_mem_data_src        data_src;
 681
 682         /*
 683          * The other fields, optionally {set,used} by
 684          * perf_{prepare,output}_sample().
 685          */
 686         u64                             type;
 687         u64                             ip;
 688         struct {
 689                 u32     pid;
 690                 u32     tid;
 691         }                               tid_entry;
 692         u64                             time;
 693         u64                             id;
 694         u64                             stream_id;
 695         struct {
 696                 u32     cpu;
 697                 u32     reserved;
 698         }                               cpu_entry;
 699         struct perf_callchain_entry     *callchain;
 700
 701         /*
 702          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 703          * on arch details.
 704          */
 705         struct perf_regs                regs_user;
 706         struct pt_regs                  regs_user_copy;
 707
 708         struct perf_regs                regs_intr;
 709         u64                             stack_user_size;
 710 } ____cacheline_aligned;
 711
 712 /* default value for data source */
 713 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 714                     PERF_MEM_S(LVL, NA)   |\
 715                     PERF_MEM_S(SNOOP, NA) |\
 716                     PERF_MEM_S(LOCK, NA)  |\
 717                     PERF_MEM_S(TLB, NA))
 718
 719 static inline void perf_sample_data_init(struct perf_sample_data *data,
 720                                          u64 addr, u64 period)
 721 {
 722         /* remaining struct members initialized in perf_prepare_sample() */
 723         data->addr = addr;
 724         data->raw  = NULL;
 725         data->br_stack = NULL;
 726         data->period = period;
 727         data->weight = 0;
 728         data->data_src.val = PERF_MEM_NA;
 729         data->txn = 0;
 730 }
 731
 732 extern void perf_output_sample(struct perf_output_handle *handle,
 733                                struct perf_event_header *header,
 734                                struct perf_sample_data *data,
 735                                struct perf_event *event);
 736 extern void perf_prepare_sample(struct perf_event_header *header,
 737                                 struct perf_sample_data *data,
 738                                 struct perf_event *event,
 739                                 struct pt_regs *regs);
 740
 741 extern int perf_event_overflow(struct perf_event *event,
 742                                  struct perf_sample_data *data,
 743                                  struct pt_regs *regs);
 744
 745 extern void perf_event_output(struct perf_event *event,
 746                                 struct perf_sample_data *data,
 747                                 struct pt_regs *regs);
 748
 749 extern void
 750 perf_event_header__init_id(struct perf_event_header *header,
 751                            struct perf_sample_data *data,
 752                            struct perf_event *event);
 753 extern void
 754 perf_event__output_id_sample(struct perf_event *event,
 755                              struct perf_output_handle *handle,
 756                              struct perf_sample_data *sample);
 757
 758 extern void
 759 perf_log_lost_samples(struct perf_event *event, u64 lost);
 760
 761 static inline bool is_sampling_event(struct perf_event *event)
 762 {
 763         return event->attr.sample_period != 0;
 764 }
 765
 766 /*
 767  * Return 1 for a software event, 0 for a hardware event
 768  */
 769 static inline int is_software_event(struct perf_event *event)
 770 {
 771         return event->pmu->task_ctx_nr == perf_sw_context;
 772 }
 773
 774 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 775
 776 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 777 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 778
 779 #ifndef perf_arch_fetch_caller_regs
 780 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 781 #endif
 782
 783 /*
 784  * Take a snapshot of the regs. Skip ip and frame pointer to
 785  * the nth caller. We only need a few of the regs:
 786  * - ip for PERF_SAMPLE_IP
 787  * - cs for user_mode() tests
 788  * - bp for callchains
 789  * - eflags, for future purposes, just in case
 790  */
 791 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 792 {
 793         memset(regs, 0, sizeof(*regs));
 794
 795         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 796 }
 797
 798 static __always_inline void
 799 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 800 {
 801         if (static_key_false(&perf_swevent_enabled[event_id]))
 802                 __perf_sw_event(event_id, nr, regs, addr);
 803 }
 804
 805 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 806
 807 /*
 808  * 'Special' version for the scheduler, it hard assumes no recursion,
 809  * which is guaranteed by us not actually scheduling inside other swevents
 810  * because those disable preemption.
 811  */
 812 static __always_inline void
 813 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 814 {
 815         if (static_key_false(&perf_swevent_enabled[event_id])) {
 816                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 817
 818                 perf_fetch_caller_regs(regs);
 819                 ___perf_sw_event(event_id, nr, regs, addr);
 820         }
 821 }
 822
 823 extern struct static_key_deferred perf_sched_events;
 824
 825 static __always_inline bool
 826 perf_sw_migrate_enabled(void)
 827 {
 828         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
 829                 return true;
 830         return false;
 831 }
 832
 833 static inline void perf_event_task_migrate(struct task_struct *task)
 834 {
 835         if (perf_sw_migrate_enabled())
 836                 task->sched_migrated = 1;
 837 }
 838
 839 static inline void perf_event_task_sched_in(struct task_struct *prev,
 840                                             struct task_struct *task)
 841 {
 842         if (static_key_false(&perf_sched_events.key))
 843                 __perf_event_task_sched_in(prev, task);
 844
 845         if (perf_sw_migrate_enabled() && task->sched_migrated) {
 846                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 847
 848                 perf_fetch_caller_regs(regs);
 849                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
 850                 task->sched_migrated = 0;
 851         }
 852 }
 853
 854 static inline void perf_event_task_sched_out(struct task_struct *prev,
 855                                              struct task_struct *next)
 856 {
 857         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 858
 859         if (static_key_false(&perf_sched_events.key))
 860                 __perf_event_task_sched_out(prev, next);
 861 }
 862
 863 static inline u64 __perf_event_count(struct perf_event *event)
 864 {
 865         return local64_read(&event->count) + atomic64_read(&event->child_count);
 866 }
 867
 868 extern void perf_event_mmap(struct vm_area_struct *vma);
 869 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 870 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 871 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 872
 873 extern void perf_event_exec(void);
 874 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 875 extern void perf_event_fork(struct task_struct *tsk);
 876
 877 /* Callchains */
 878 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 879
 880 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 881 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 882
 883 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 884 {
 885         if (entry->nr < PERF_MAX_STACK_DEPTH)
 886                 entry->ip[entry->nr++] = ip;
 887 }
 888
 889 extern int sysctl_perf_event_paranoid;
 890 extern int sysctl_perf_event_mlock;
 891 extern int sysctl_perf_event_sample_rate;
 892 extern int sysctl_perf_cpu_time_max_percent;
 893
 894 extern void perf_sample_event_took(u64 sample_len_ns);
 895
 896 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 897                 void __user *buffer, size_t *lenp,
 898                 loff_t *ppos);
 899 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 900                 void __user *buffer, size_t *lenp,
 901                 loff_t *ppos);
 902
 903
 904 static inline bool perf_paranoid_tracepoint_raw(void)
 905 {
 906         return sysctl_perf_event_paranoid > -1;
 907 }
 908
 909 static inline bool perf_paranoid_cpu(void)
 910 {
 911         return sysctl_perf_event_paranoid > 0;
 912 }
 913
 914 static inline bool perf_paranoid_kernel(void)
 915 {
 916         return sysctl_perf_event_paranoid > 1;
 917 }
 918
 919 extern void perf_event_init(void);
 920 extern void perf_tp_event(u64 addr, u64 count, void *record,
 921                           int entry_size, struct pt_regs *regs,
 922                           struct hlist_head *head, int rctx,
 923                           struct task_struct *task);
 924 extern void perf_bp_event(struct perf_event *event, void *data);
 925
 926 #ifndef perf_misc_flags
 927 # define perf_misc_flags(regs) \
 928                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 929 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 930 #endif
 931
 932 static inline bool has_branch_stack(struct perf_event *event)
 933 {
 934         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 935 }
 936
 937 static inline bool needs_branch_stack(struct perf_event *event)
 938 {
 939         return event->attr.branch_sample_type != 0;
 940 }
 941
 942 static inline bool has_aux(struct perf_event *event)
 943 {
 944         return event->pmu->setup_aux;
 945 }
 946
 947 extern int perf_output_begin(struct perf_output_handle *handle,
 948                              struct perf_event *event, unsigned int size);
 949 extern void perf_output_end(struct perf_output_handle *handle);
 950 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 951                              const void *buf, unsigned int len);
 952 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 953                                      unsigned int len);
 954 extern int perf_swevent_get_recursion_context(void);
 955 extern void perf_swevent_put_recursion_context(int rctx);
 956 extern u64 perf_swevent_set_period(struct perf_event *event);
 957 extern void perf_event_enable(struct perf_event *event);
 958 extern void perf_event_disable(struct perf_event *event);
 959 extern int __perf_event_disable(void *info);
 960 extern void perf_event_task_tick(void);
 961 #else /* !CONFIG_PERF_EVENTS: */
 962 static inline void *
 963 perf_aux_output_begin(struct perf_output_handle *handle,
 964                       struct perf_event *event)                         { return NULL; }
 965 static inline void
 966 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
 967                     bool truncated)                                     { }
 968 static inline int
 969 perf_aux_output_skip(struct perf_output_handle *handle,
 970                      unsigned long size)                                { return -EINVAL; }
 971 static inline void *
 972 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
 973 static inline void
 974 perf_event_task_migrate(struct task_struct *task)                       { }
 975 static inline void
 976 perf_event_task_sched_in(struct task_struct *prev,
 977                          struct task_struct *task)                      { }
 978 static inline void
 979 perf_event_task_sched_out(struct task_struct *prev,
 980                           struct task_struct *next)                     { }
 981 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 982 static inline void perf_event_exit_task(struct task_struct *child)      { }
 983 static inline void perf_event_free_task(struct task_struct *task)       { }
 984 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 985 static inline struct perf_event *perf_event_get(unsigned int fd)        { return ERR_PTR(-EINVAL); }
 986 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
 987 {
 988         return ERR_PTR(-EINVAL);
 989 }
 990 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
 991 static inline void perf_event_print_debug(void)                         { }
 992 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 993 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 994 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 995 {
 996         return -EINVAL;
 997 }
 998
 999 static inline void
1000 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1001 static inline void
1002 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1003 static inline void
1004 perf_bp_event(struct perf_event *event, void *data)                     { }
1005
1006 static inline int perf_register_guest_info_callbacks
1007 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1008 static inline int perf_unregister_guest_info_callbacks
1009 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1010
1011 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1012 static inline void perf_event_exec(void)                                { }
1013 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1014 static inline void perf_event_fork(struct task_struct *tsk)             { }
1015 static inline void perf_event_init(void)                                { }
1016 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1017 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1018 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1019 static inline void perf_event_enable(struct perf_event *event)          { }
1020 static inline void perf_event_disable(struct perf_event *event)         { }
1021 static inline int __perf_event_disable(void *info)                      { return -1; }
1022 static inline void perf_event_task_tick(void)                           { }
1023 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1024 #endif
1025
1026 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
1027 extern bool perf_event_can_stop_tick(void);
1028 #else
1029 static inline bool perf_event_can_stop_tick(void)                       { return true; }
1030 #endif
1031
1032 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1033 extern void perf_restore_debug_store(void);
1034 #else
1035 static inline void perf_restore_debug_store(void)                       { }
1036 #endif
1037
1038 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1039
1040 /*
1041  * This has to have a higher priority than migration_notifier in sched/core.c.
1042  */
1043 #define perf_cpu_notifier(fn)                                           \
1044 do {                                                                    \
1045         static struct notifier_block fn##_nb =                          \
1046                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1047         unsigned long cpu = smp_processor_id();                         \
1048         unsigned long flags;                                            \
1049                                                                         \
1050         cpu_notifier_register_begin();                                  \
1051         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1052                 (void *)(unsigned long)cpu);                            \
1053         local_irq_save(flags);                                          \
1054         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1055                 (void *)(unsigned long)cpu);                            \
1056         local_irq_restore(flags);                                       \
1057         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1058                 (void *)(unsigned long)cpu);                            \
1059         __register_cpu_notifier(&fn##_nb);                              \
1060         cpu_notifier_register_done();                                   \
1061 } while (0)
1062
1063 /*
1064  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1065  * callback for already online CPUs.
1066  */
1067 #define __perf_cpu_notifier(fn)                                         \
1068 do {                                                                    \
1069         static struct notifier_block fn##_nb =                          \
1070                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1071                                                                         \
1072         __register_cpu_notifier(&fn##_nb);                              \
1073 } while (0)
1074
1075 struct perf_pmu_events_attr {
1076         struct device_attribute attr;
1077         u64 id;
1078         const char *event_str;
1079 };
1080
1081 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1082                               char *page);
1083
1084 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1085 static struct perf_pmu_events_attr _var = {                             \
1086         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1087         .id   =  _id,                                                   \
1088 };
1089
1090 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1091 static struct perf_pmu_events_attr _var = {                                 \
1092         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1093         .id             = 0,                                                \
1094         .event_str      = _str,                                             \
1095 };
1096
1097 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1098 static ssize_t                                                          \
1099 _name##_show(struct device *dev,                                        \
1100                                struct device_attribute *attr,           \
1101                                char *page)                              \
1102 {                                                                       \
1103         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1104         return sprintf(page, _format "\n");                             \
1105 }                                                                       \
1106                                                                         \
1107 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1108
1109 #endif /* _LINUX_PERF_EVENT_H */