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                         int                     is_group_event;
 125                         struct list_head        cqm_events_entry;
 126                         struct list_head        cqm_groups_entry;
 127                         struct list_head        cqm_group_entry;
 128                 };
 129                 struct { /* itrace */
 130                         int                     itrace_started;
 131                 };
 132 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 133                 struct { /* breakpoint */
 134                         /*
 135                          * Crufty hack to avoid the chicken and egg
 136                          * problem hw_breakpoint has with context
 137                          * creation and event initalization.
 138                          */
 139                         struct arch_hw_breakpoint       info;
 140                         struct list_head                bp_list;
 141                 };
 142 #endif
 143         };
 144         /*
 145          * If the event is a per task event, this will point to the task in
 146          * question. See the comment in perf_event_alloc().
 147          */
 148         struct task_struct              *target;
 149
 150 /*
 151  * hw_perf_event::state flags; used to track the PERF_EF_* state.
 152  */
 153 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 154 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 155 #define PERF_HES_ARCH           0x04
 156
 157         int                             state;
 158
 159         /*
 160          * The last observed hardware counter value, updated with a
 161          * local64_cmpxchg() such that pmu::read() can be called nested.
 162          */
 163         local64_t                       prev_count;
 164
 165         /*
 166          * The period to start the next sample with.
 167          */
 168         u64                             sample_period;
 169
 170         /*
 171          * The period we started this sample with.
 172          */
 173         u64                             last_period;
 174
 175         /*
 176          * However much is left of the current period; note that this is
 177          * a full 64bit value and allows for generation of periods longer
 178          * than hardware might allow.
 179          */
 180         local64_t                       period_left;
 181
 182         /*
 183          * State for throttling the event, see __perf_event_overflow() and
 184          * perf_adjust_freq_unthr_context().
 185          */
 186         u64                             interrupts_seq;
 187         u64                             interrupts;
 188
 189         /*
 190          * State for freq target events, see __perf_event_overflow() and
 191          * perf_adjust_freq_unthr_context().
 192          */
 193         u64                             freq_time_stamp;
 194         u64                             freq_count_stamp;
 195 #endif
 196 };
 197
 198 struct perf_event;
 199
 200 /*
 201  * Common implementation detail of pmu::{start,commit,cancel}_txn
 202  */
 203 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
 204 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
 205
 206 /**
 207  * pmu::capabilities flags
 208  */
 209 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 210 #define PERF_PMU_CAP_NO_NMI                     0x02
 211 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
 212 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
 213 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
 214 #define PERF_PMU_CAP_ITRACE                     0x20
 215
 216 /**
 217  * struct pmu - generic performance monitoring unit
 218  */
 219 struct pmu {
 220         struct list_head                entry;
 221
 222         struct module                   *module;
 223         struct device                   *dev;
 224         const struct attribute_group    **attr_groups;
 225         const char                      *name;
 226         int                             type;
 227
 228         /*
 229          * various common per-pmu feature flags
 230          */
 231         int                             capabilities;
 232
 233         int * __percpu                  pmu_disable_count;
 234         struct perf_cpu_context * __percpu pmu_cpu_context;
 235         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 236         int                             task_ctx_nr;
 237         int                             hrtimer_interval_ms;
 238
 239         /*
 240          * Fully disable/enable this PMU, can be used to protect from the PMI
 241          * as well as for lazy/batch writing of the MSRs.
 242          */
 243         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 244         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 245
 246         /*
 247          * Try and initialize the event for this PMU.
 248          *
 249          * Returns:
 250          *  -ENOENT     -- @event is not for this PMU
 251          *
 252          *  -ENODEV     -- @event is for this PMU but PMU not present
 253          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
 254          *  -EINVAL     -- @event is for this PMU but @event is not valid
 255          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
 256          *  -EACCESS    -- @event is for this PMU, @event is valid, but no privilidges
 257          *
 258          *  0           -- @event is for this PMU and valid
 259          *
 260          * Other error return values are allowed.
 261          */
 262         int (*event_init)               (struct perf_event *event);
 263
 264         /*
 265          * Notification that the event was mapped or unmapped.  Called
 266          * in the context of the mapping task.
 267          */
 268         void (*event_mapped)            (struct perf_event *event); /*optional*/
 269         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 270
 271         /*
 272          * Flags for ->add()/->del()/ ->start()/->stop(). There are
 273          * matching hw_perf_event::state flags.
 274          */
 275 #define PERF_EF_START   0x01            /* start the counter when adding    */
 276 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 277 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 278
 279         /*
 280          * Adds/Removes a counter to/from the PMU, can be done inside a
 281          * transaction, see the ->*_txn() methods.
 282          *
 283          * The add/del callbacks will reserve all hardware resources required
 284          * to service the event, this includes any counter constraint
 285          * scheduling etc.
 286          *
 287          * Called with IRQs disabled and the PMU disabled on the CPU the event
 288          * is on.
 289          *
 290          * ->add() called without PERF_EF_START should result in the same state
 291          *  as ->add() followed by ->stop().
 292          *
 293          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
 294          *  ->stop() that must deal with already being stopped without
 295          *  PERF_EF_UPDATE.
 296          */
 297         int  (*add)                     (struct perf_event *event, int flags);
 298         void (*del)                     (struct perf_event *event, int flags);
 299
 300         /*
 301          * Starts/Stops a counter present on the PMU.
 302          *
 303          * The PMI handler should stop the counter when perf_event_overflow()
 304          * returns !0. ->start() will be used to continue.
 305          *
 306          * Also used to change the sample period.
 307          *
 308          * Called with IRQs disabled and the PMU disabled on the CPU the event
 309          * is on -- will be called from NMI context with the PMU generates
 310          * NMIs.
 311          *
 312          * ->stop() with PERF_EF_UPDATE will read the counter and update
 313          *  period/count values like ->read() would.
 314          *
 315          * ->start() with PERF_EF_RELOAD will reprogram the the counter
 316          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
 317          */
 318         void (*start)                   (struct perf_event *event, int flags);
 319         void (*stop)                    (struct perf_event *event, int flags);
 320
 321         /*
 322          * Updates the counter value of the event.
 323          *
 324          * For sampling capable PMUs this will also update the software period
 325          * hw_perf_event::period_left field.
 326          */
 327         void (*read)                    (struct perf_event *event);
 328
 329         /*
 330          * Group events scheduling is treated as a transaction, add
 331          * group events as a whole and perform one schedulability test.
 332          * If the test fails, roll back the whole group
 333          *
 334          * Start the transaction, after this ->add() doesn't need to
 335          * do schedulability tests.
 336          *
 337          * Optional.
 338          */
 339         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
 340         /*
 341          * If ->start_txn() disabled the ->add() schedulability test
 342          * then ->commit_txn() is required to perform one. On success
 343          * the transaction is closed. On error the transaction is kept
 344          * open until ->cancel_txn() is called.
 345          *
 346          * Optional.
 347          */
 348         int  (*commit_txn)              (struct pmu *pmu);
 349         /*
 350          * Will cancel the transaction, assumes ->del() is called
 351          * for each successful ->add() during the transaction.
 352          *
 353          * Optional.
 354          */
 355         void (*cancel_txn)              (struct pmu *pmu);
 356
 357         /*
 358          * Will return the value for perf_event_mmap_page::index for this event,
 359          * if no implementation is provided it will default to: event->hw.idx + 1.
 360          */
 361         int (*event_idx)                (struct perf_event *event); /*optional */
 362
 363         /*
 364          * context-switches callback
 365          */
 366         void (*sched_task)              (struct perf_event_context *ctx,
 367                                         bool sched_in);
 368         /*
 369          * PMU specific data size
 370          */
 371         size_t                          task_ctx_size;
 372
 373
 374         /*
 375          * Return the count value for a counter.
 376          */
 377         u64 (*count)                    (struct perf_event *event); /*optional*/
 378
 379         /*
 380          * Set up pmu-private data structures for an AUX area
 381          */
 382         void *(*setup_aux)              (int cpu, void **pages,
 383                                          int nr_pages, bool overwrite);
 384                                         /* optional */
 385
 386         /*
 387          * Free pmu-private AUX data structures
 388          */
 389         void (*free_aux)                (void *aux); /* optional */
 390
 391         /*
 392          * Filter events for PMU-specific reasons.
 393          */
 394         int (*filter_match)             (struct perf_event *event); /* optional */
 395 };
 396
 397 /**
 398  * enum perf_event_active_state - the states of a event
 399  */
 400 enum perf_event_active_state {
 401         PERF_EVENT_STATE_DEAD           = -4,
 402         PERF_EVENT_STATE_EXIT           = -3,
 403         PERF_EVENT_STATE_ERROR          = -2,
 404         PERF_EVENT_STATE_OFF            = -1,
 405         PERF_EVENT_STATE_INACTIVE       =  0,
 406         PERF_EVENT_STATE_ACTIVE         =  1,
 407 };
 408
 409 struct file;
 410 struct perf_sample_data;
 411
 412 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 413                                         struct perf_sample_data *,
 414                                         struct pt_regs *regs);
 415
 416 enum perf_group_flag {
 417         PERF_GROUP_SOFTWARE             = 0x1,
 418 };
 419
 420 #define SWEVENT_HLIST_BITS              8
 421 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 422
 423 struct swevent_hlist {
 424         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 425         struct rcu_head                 rcu_head;
 426 };
 427
 428 #define PERF_ATTACH_CONTEXT     0x01
 429 #define PERF_ATTACH_GROUP       0x02
 430 #define PERF_ATTACH_TASK        0x04
 431 #define PERF_ATTACH_TASK_DATA   0x08
 432
 433 struct perf_cgroup;
 434 struct ring_buffer;
 435
 436 /**
 437  * struct perf_event - performance event kernel representation:
 438  */
 439 struct perf_event {
 440 #ifdef CONFIG_PERF_EVENTS
 441         /*
 442          * entry onto perf_event_context::event_list;
 443          *   modifications require ctx->lock
 444          *   RCU safe iterations.
 445          */
 446         struct list_head                event_entry;
 447
 448         /*
 449          * XXX: group_entry and sibling_list should be mutually exclusive;
 450          * either you're a sibling on a group, or you're the group leader.
 451          * Rework the code to always use the same list element.
 452          *
 453          * Locked for modification by both ctx->mutex and ctx->lock; holding
 454          * either sufficies for read.
 455          */
 456         struct list_head                group_entry;
 457         struct list_head                sibling_list;
 458
 459         /*
 460          * We need storage to track the entries in perf_pmu_migrate_context; we
 461          * cannot use the event_entry because of RCU and we want to keep the
 462          * group in tact which avoids us using the other two entries.
 463          */
 464         struct list_head                migrate_entry;
 465
 466         struct hlist_node               hlist_entry;
 467         struct list_head                active_entry;
 468         int                             nr_siblings;
 469         int                             group_flags;
 470         struct perf_event               *group_leader;
 471         struct pmu                      *pmu;
 472         void                            *pmu_private;
 473
 474         enum perf_event_active_state    state;
 475         unsigned int                    attach_state;
 476         local64_t                       count;
 477         atomic64_t                      child_count;
 478
 479         /*
 480          * These are the total time in nanoseconds that the event
 481          * has been enabled (i.e. eligible to run, and the task has
 482          * been scheduled in, if this is a per-task event)
 483          * and running (scheduled onto the CPU), respectively.
 484          *
 485          * They are computed from tstamp_enabled, tstamp_running and
 486          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 487          */
 488         u64                             total_time_enabled;
 489         u64                             total_time_running;
 490
 491         /*
 492          * These are timestamps used for computing total_time_enabled
 493          * and total_time_running when the event is in INACTIVE or
 494          * ACTIVE state, measured in nanoseconds from an arbitrary point
 495          * in time.
 496          * tstamp_enabled: the notional time when the event was enabled
 497          * tstamp_running: the notional time when the event was scheduled on
 498          * tstamp_stopped: in INACTIVE state, the notional time when the
 499          *      event was scheduled off.
 500          */
 501         u64                             tstamp_enabled;
 502         u64                             tstamp_running;
 503         u64                             tstamp_stopped;
 504
 505         /*
 506          * timestamp shadows the actual context timing but it can
 507          * be safely used in NMI interrupt context. It reflects the
 508          * context time as it was when the event was last scheduled in.
 509          *
 510          * ctx_time already accounts for ctx->timestamp. Therefore to
 511          * compute ctx_time for a sample, simply add perf_clock().
 512          */
 513         u64                             shadow_ctx_time;
 514
 515         struct perf_event_attr          attr;
 516         u16                             header_size;
 517         u16                             id_header_size;
 518         u16                             read_size;
 519         struct hw_perf_event            hw;
 520
 521         struct perf_event_context       *ctx;
 522         atomic_long_t                   refcount;
 523
 524         /*
 525          * These accumulate total time (in nanoseconds) that children
 526          * events have been enabled and running, respectively.
 527          */
 528         atomic64_t                      child_total_time_enabled;
 529         atomic64_t                      child_total_time_running;
 530
 531         /*
 532          * Protect attach/detach and child_list:
 533          */
 534         struct mutex                    child_mutex;
 535         struct list_head                child_list;
 536         struct perf_event               *parent;
 537
 538         int                             oncpu;
 539         int                             cpu;
 540
 541         struct list_head                owner_entry;
 542         struct task_struct              *owner;
 543
 544         /* mmap bits */
 545         struct mutex                    mmap_mutex;
 546         atomic_t                        mmap_count;
 547
 548         struct ring_buffer              *rb;
 549         struct list_head                rb_entry;
 550         unsigned long                   rcu_batches;
 551         int                             rcu_pending;
 552
 553         /* poll related */
 554         wait_queue_head_t               waitq;
 555         struct fasync_struct            *fasync;
 556
 557         /* delayed work for NMIs and such */
 558         int                             pending_wakeup;
 559         int                             pending_kill;
 560         int                             pending_disable;
 561         struct irq_work                 pending;
 562
 563         atomic_t                        event_limit;
 564
 565         void (*destroy)(struct perf_event *);
 566         struct rcu_head                 rcu_head;
 567
 568         struct pid_namespace            *ns;
 569         u64                             id;
 570
 571         u64                             (*clock)(void);
 572         perf_overflow_handler_t         overflow_handler;
 573         void                            *overflow_handler_context;
 574
 575 #ifdef CONFIG_EVENT_TRACING
 576         struct trace_event_call         *tp_event;
 577         struct event_filter             *filter;
 578 #ifdef CONFIG_FUNCTION_TRACER
 579         struct ftrace_ops               ftrace_ops;
 580 #endif
 581 #endif
 582
 583 #ifdef CONFIG_CGROUP_PERF
 584         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 585         int                             cgrp_defer_enabled;
 586 #endif
 587
 588 #endif /* CONFIG_PERF_EVENTS */
 589 };
 590
 591 /**
 592  * struct perf_event_context - event context structure
 593  *
 594  * Used as a container for task events and CPU events as well:
 595  */
 596 struct perf_event_context {
 597         struct pmu                      *pmu;
 598         /*
 599          * Protect the states of the events in the list,
 600          * nr_active, and the list:
 601          */
 602         raw_spinlock_t                  lock;
 603         /*
 604          * Protect the list of events.  Locking either mutex or lock
 605          * is sufficient to ensure the list doesn't change; to change
 606          * the list you need to lock both the mutex and the spinlock.
 607          */
 608         struct mutex                    mutex;
 609
 610         struct list_head                active_ctx_list;
 611         struct list_head                pinned_groups;
 612         struct list_head                flexible_groups;
 613         struct list_head                event_list;
 614         int                             nr_events;
 615         int                             nr_active;
 616         int                             is_active;
 617         int                             nr_stat;
 618         int                             nr_freq;
 619         int                             rotate_disable;
 620         atomic_t                        refcount;
 621         struct task_struct              *task;
 622
 623         /*
 624          * Context clock, runs when context enabled.
 625          */
 626         u64                             time;
 627         u64                             timestamp;
 628
 629         /*
 630          * These fields let us detect when two contexts have both
 631          * been cloned (inherited) from a common ancestor.
 632          */
 633         struct perf_event_context       *parent_ctx;
 634         u64                             parent_gen;
 635         u64                             generation;
 636         int                             pin_count;
 637         int                             nr_cgroups;      /* cgroup evts */
 638         void                            *task_ctx_data; /* pmu specific data */
 639         struct rcu_head                 rcu_head;
 640 };
 641
 642 /*
 643  * Number of contexts where an event can trigger:
 644  *      task, softirq, hardirq, nmi.
 645  */
 646 #define PERF_NR_CONTEXTS        4
 647
 648 /**
 649  * struct perf_event_cpu_context - per cpu event context structure
 650  */
 651 struct perf_cpu_context {
 652         struct perf_event_context       ctx;
 653         struct perf_event_context       *task_ctx;
 654         int                             active_oncpu;
 655         int                             exclusive;
 656
 657         raw_spinlock_t                  hrtimer_lock;
 658         struct hrtimer                  hrtimer;
 659         ktime_t                         hrtimer_interval;
 660         unsigned int                    hrtimer_active;
 661
 662         struct pmu                      *unique_pmu;
 663         struct perf_cgroup              *cgrp;
 664 };
 665
 666 struct perf_output_handle {
 667         struct perf_event               *event;
 668         struct ring_buffer              *rb;
 669         unsigned long                   wakeup;
 670         unsigned long                   size;
 671         union {
 672                 void                    *addr;
 673                 unsigned long           head;
 674         };
 675         int                             page;
 676 };
 677
 678 #ifdef CONFIG_CGROUP_PERF
 679
 680 /*
 681  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 682  * This is a per-cpu dynamically allocated data structure.
 683  */
 684 struct perf_cgroup_info {
 685         u64                             time;
 686         u64                             timestamp;
 687 };
 688
 689 struct perf_cgroup {
 690         struct cgroup_subsys_state      css;
 691         struct perf_cgroup_info __percpu *info;
 692 };
 693
 694 /*
 695  * Must ensure cgroup is pinned (css_get) before calling
 696  * this function. In other words, we cannot call this function
 697  * if there is no cgroup event for the current CPU context.
 698  */
 699 static inline struct perf_cgroup *
 700 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 701 {
 702         return container_of(task_css_check(task, perf_event_cgrp_id,
 703                                            ctx ? lockdep_is_held(&ctx->lock)
 704                                                : true),
 705                             struct perf_cgroup, css);
 706 }
 707 #endif /* CONFIG_CGROUP_PERF */
 708
 709 #ifdef CONFIG_PERF_EVENTS
 710
 711 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 712                                    struct perf_event *event);
 713 extern void perf_aux_output_end(struct perf_output_handle *handle,
 714                                 unsigned long size, bool truncated);
 715 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 716                                 unsigned long size);
 717 extern void *perf_get_aux(struct perf_output_handle *handle);
 718
 719 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 720 extern void perf_pmu_unregister(struct pmu *pmu);
 721
 722 extern int perf_num_counters(void);
 723 extern const char *perf_pmu_name(void);
 724 extern void __perf_event_task_sched_in(struct task_struct *prev,
 725                                        struct task_struct *task);
 726 extern void __perf_event_task_sched_out(struct task_struct *prev,
 727                                         struct task_struct *next);
 728 extern int perf_event_init_task(struct task_struct *child);
 729 extern void perf_event_exit_task(struct task_struct *child);
 730 extern void perf_event_free_task(struct task_struct *task);
 731 extern void perf_event_delayed_put(struct task_struct *task);
 732 extern struct file *perf_event_get(unsigned int fd);
 733 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 734 extern void perf_event_print_debug(void);
 735 extern void perf_pmu_disable(struct pmu *pmu);
 736 extern void perf_pmu_enable(struct pmu *pmu);
 737 extern void perf_sched_cb_dec(struct pmu *pmu);
 738 extern void perf_sched_cb_inc(struct pmu *pmu);
 739 extern int perf_event_task_disable(void);
 740 extern int perf_event_task_enable(void);
 741 extern int perf_event_refresh(struct perf_event *event, int refresh);
 742 extern void perf_event_update_userpage(struct perf_event *event);
 743 extern int perf_event_release_kernel(struct perf_event *event);
 744 extern struct perf_event *
 745 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 746                                 int cpu,
 747                                 struct task_struct *task,
 748                                 perf_overflow_handler_t callback,
 749                                 void *context);
 750 extern void perf_pmu_migrate_context(struct pmu *pmu,
 751                                 int src_cpu, int dst_cpu);
 752 extern u64 perf_event_read_local(struct perf_event *event);
 753 extern u64 perf_event_read_value(struct perf_event *event,
 754                                  u64 *enabled, u64 *running);
 755
 756
 757 struct perf_sample_data {
 758         /*
 759          * Fields set by perf_sample_data_init(), group so as to
 760          * minimize the cachelines touched.
 761          */
 762         u64                             addr;
 763         struct perf_raw_record          *raw;
 764         struct perf_branch_stack        *br_stack;
 765         u64                             period;
 766         u64                             weight;
 767         u64                             txn;
 768         union  perf_mem_data_src        data_src;
 769
 770         /*
 771          * The other fields, optionally {set,used} by
 772          * perf_{prepare,output}_sample().
 773          */
 774         u64                             type;
 775         u64                             ip;
 776         struct {
 777                 u32     pid;
 778                 u32     tid;
 779         }                               tid_entry;
 780         u64                             time;
 781         u64                             id;
 782         u64                             stream_id;
 783         struct {
 784                 u32     cpu;
 785                 u32     reserved;
 786         }                               cpu_entry;
 787         struct perf_callchain_entry     *callchain;
 788
 789         /*
 790          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 791          * on arch details.
 792          */
 793         struct perf_regs                regs_user;
 794         struct pt_regs                  regs_user_copy;
 795
 796         struct perf_regs                regs_intr;
 797         u64                             stack_user_size;
 798 } ____cacheline_aligned;
 799
 800 /* default value for data source */
 801 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 802                     PERF_MEM_S(LVL, NA)   |\
 803                     PERF_MEM_S(SNOOP, NA) |\
 804                     PERF_MEM_S(LOCK, NA)  |\
 805                     PERF_MEM_S(TLB, NA))
 806
 807 static inline void perf_sample_data_init(struct perf_sample_data *data,
 808                                          u64 addr, u64 period)
 809 {
 810         /* remaining struct members initialized in perf_prepare_sample() */
 811         data->addr = addr;
 812         data->raw  = NULL;
 813         data->br_stack = NULL;
 814         data->period = period;
 815         data->weight = 0;
 816         data->data_src.val = PERF_MEM_NA;
 817         data->txn = 0;
 818 }
 819
 820 extern void perf_output_sample(struct perf_output_handle *handle,
 821                                struct perf_event_header *header,
 822                                struct perf_sample_data *data,
 823                                struct perf_event *event);
 824 extern void perf_prepare_sample(struct perf_event_header *header,
 825                                 struct perf_sample_data *data,
 826                                 struct perf_event *event,
 827                                 struct pt_regs *regs);
 828
 829 extern int perf_event_overflow(struct perf_event *event,
 830                                  struct perf_sample_data *data,
 831                                  struct pt_regs *regs);
 832
 833 extern void perf_event_output(struct perf_event *event,
 834                                 struct perf_sample_data *data,
 835                                 struct pt_regs *regs);
 836
 837 extern void
 838 perf_event_header__init_id(struct perf_event_header *header,
 839                            struct perf_sample_data *data,
 840                            struct perf_event *event);
 841 extern void
 842 perf_event__output_id_sample(struct perf_event *event,
 843                              struct perf_output_handle *handle,
 844                              struct perf_sample_data *sample);
 845
 846 extern void
 847 perf_log_lost_samples(struct perf_event *event, u64 lost);
 848
 849 static inline bool is_sampling_event(struct perf_event *event)
 850 {
 851         return event->attr.sample_period != 0;
 852 }
 853
 854 /*
 855  * Return 1 for a software event, 0 for a hardware event
 856  */
 857 static inline int is_software_event(struct perf_event *event)
 858 {
 859         return event->pmu->task_ctx_nr == perf_sw_context;
 860 }
 861
 862 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 863
 864 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 865 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 866
 867 #ifndef perf_arch_fetch_caller_regs
 868 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 869 #endif
 870
 871 /*
 872  * Take a snapshot of the regs. Skip ip and frame pointer to
 873  * the nth caller. We only need a few of the regs:
 874  * - ip for PERF_SAMPLE_IP
 875  * - cs for user_mode() tests
 876  * - bp for callchains
 877  * - eflags, for future purposes, just in case
 878  */
 879 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 880 {
 881         memset(regs, 0, sizeof(*regs));
 882
 883         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 884 }
 885
 886 static __always_inline void
 887 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 888 {
 889         if (static_key_false(&perf_swevent_enabled[event_id]))
 890                 __perf_sw_event(event_id, nr, regs, addr);
 891 }
 892
 893 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 894
 895 /*
 896  * 'Special' version for the scheduler, it hard assumes no recursion,
 897  * which is guaranteed by us not actually scheduling inside other swevents
 898  * because those disable preemption.
 899  */
 900 static __always_inline void
 901 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 902 {
 903         if (static_key_false(&perf_swevent_enabled[event_id])) {
 904                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 905
 906                 perf_fetch_caller_regs(regs);
 907                 ___perf_sw_event(event_id, nr, regs, addr);
 908         }
 909 }
 910
 911 extern struct static_key_false perf_sched_events;
 912
 913 static __always_inline bool
 914 perf_sw_migrate_enabled(void)
 915 {
 916         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
 917                 return true;
 918         return false;
 919 }
 920
 921 static inline void perf_event_task_migrate(struct task_struct *task)
 922 {
 923         if (perf_sw_migrate_enabled())
 924                 task->sched_migrated = 1;
 925 }
 926
 927 static inline void perf_event_task_sched_in(struct task_struct *prev,
 928                                             struct task_struct *task)
 929 {
 930         if (static_branch_unlikely(&perf_sched_events))
 931                 __perf_event_task_sched_in(prev, task);
 932
 933         if (perf_sw_migrate_enabled() && task->sched_migrated) {
 934                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 935
 936                 perf_fetch_caller_regs(regs);
 937                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
 938                 task->sched_migrated = 0;
 939         }
 940 }
 941
 942 static inline void perf_event_task_sched_out(struct task_struct *prev,
 943                                              struct task_struct *next)
 944 {
 945         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 946
 947         if (static_branch_unlikely(&perf_sched_events))
 948                 __perf_event_task_sched_out(prev, next);
 949 }
 950
 951 static inline u64 __perf_event_count(struct perf_event *event)
 952 {
 953         return local64_read(&event->count) + atomic64_read(&event->child_count);
 954 }
 955
 956 extern void perf_event_mmap(struct vm_area_struct *vma);
 957 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 958 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 959 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 960
 961 extern void perf_event_exec(void);
 962 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 963 extern void perf_event_fork(struct task_struct *tsk);
 964
 965 /* Callchains */
 966 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 967
 968 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 969 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 970
 971 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 972 {
 973         if (entry->nr < PERF_MAX_STACK_DEPTH)
 974                 entry->ip[entry->nr++] = ip;
 975 }
 976
 977 extern int sysctl_perf_event_paranoid;
 978 extern int sysctl_perf_event_mlock;
 979 extern int sysctl_perf_event_sample_rate;
 980 extern int sysctl_perf_cpu_time_max_percent;
 981
 982 extern void perf_sample_event_took(u64 sample_len_ns);
 983
 984 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 985                 void __user *buffer, size_t *lenp,
 986                 loff_t *ppos);
 987 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 988                 void __user *buffer, size_t *lenp,
 989                 loff_t *ppos);
 990
 991
 992 static inline bool perf_paranoid_tracepoint_raw(void)
 993 {
 994         return sysctl_perf_event_paranoid > -1;
 995 }
 996
 997 static inline bool perf_paranoid_cpu(void)
 998 {
 999         return sysctl_perf_event_paranoid > 0;
1000 }
1001
1002 static inline bool perf_paranoid_kernel(void)
1003 {
1004         return sysctl_perf_event_paranoid > 1;
1005 }
1006
1007 extern void perf_event_init(void);
1008 extern void perf_tp_event(u64 addr, u64 count, void *record,
1009                           int entry_size, struct pt_regs *regs,
1010                           struct hlist_head *head, int rctx,
1011                           struct task_struct *task);
1012 extern void perf_bp_event(struct perf_event *event, void *data);
1013
1014 #ifndef perf_misc_flags
1015 # define perf_misc_flags(regs) \
1016                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1017 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1018 #endif
1019
1020 static inline bool has_branch_stack(struct perf_event *event)
1021 {
1022         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1023 }
1024
1025 static inline bool needs_branch_stack(struct perf_event *event)
1026 {
1027         return event->attr.branch_sample_type != 0;
1028 }
1029
1030 static inline bool has_aux(struct perf_event *event)
1031 {
1032         return event->pmu->setup_aux;
1033 }
1034
1035 extern int perf_output_begin(struct perf_output_handle *handle,
1036                              struct perf_event *event, unsigned int size);
1037 extern void perf_output_end(struct perf_output_handle *handle);
1038 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1039                              const void *buf, unsigned int len);
1040 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1041                                      unsigned int len);
1042 extern int perf_swevent_get_recursion_context(void);
1043 extern void perf_swevent_put_recursion_context(int rctx);
1044 extern u64 perf_swevent_set_period(struct perf_event *event);
1045 extern void perf_event_enable(struct perf_event *event);
1046 extern void perf_event_disable(struct perf_event *event);
1047 extern void perf_event_disable_local(struct perf_event *event);
1048 extern void perf_event_task_tick(void);
1049 #else /* !CONFIG_PERF_EVENTS: */
1050 static inline void *
1051 perf_aux_output_begin(struct perf_output_handle *handle,
1052                       struct perf_event *event)                         { return NULL; }
1053 static inline void
1054 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1055                     bool truncated)                                     { }
1056 static inline int
1057 perf_aux_output_skip(struct perf_output_handle *handle,
1058                      unsigned long size)                                { return -EINVAL; }
1059 static inline void *
1060 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1061 static inline void
1062 perf_event_task_migrate(struct task_struct *task)                       { }
1063 static inline void
1064 perf_event_task_sched_in(struct task_struct *prev,
1065                          struct task_struct *task)                      { }
1066 static inline void
1067 perf_event_task_sched_out(struct task_struct *prev,
1068                           struct task_struct *next)                     { }
1069 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1070 static inline void perf_event_exit_task(struct task_struct *child)      { }
1071 static inline void perf_event_free_task(struct task_struct *task)       { }
1072 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1073 static inline struct file *perf_event_get(unsigned int fd)      { return ERR_PTR(-EINVAL); }
1074 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1075 {
1076         return ERR_PTR(-EINVAL);
1077 }
1078 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1079 static inline void perf_event_print_debug(void)                         { }
1080 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1081 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1082 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1083 {
1084         return -EINVAL;
1085 }
1086
1087 static inline void
1088 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1089 static inline void
1090 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1091 static inline void
1092 perf_bp_event(struct perf_event *event, void *data)                     { }
1093
1094 static inline int perf_register_guest_info_callbacks
1095 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1096 static inline int perf_unregister_guest_info_callbacks
1097 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1098
1099 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1100 static inline void perf_event_exec(void)                                { }
1101 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1102 static inline void perf_event_fork(struct task_struct *tsk)             { }
1103 static inline void perf_event_init(void)                                { }
1104 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1105 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1106 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1107 static inline void perf_event_enable(struct perf_event *event)          { }
1108 static inline void perf_event_disable(struct perf_event *event)         { }
1109 static inline int __perf_event_disable(void *info)                      { return -1; }
1110 static inline void perf_event_task_tick(void)                           { }
1111 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1112 #endif
1113
1114 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1115 extern void perf_restore_debug_store(void);
1116 #else
1117 static inline void perf_restore_debug_store(void)                       { }
1118 #endif
1119
1120 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1121
1122 /*
1123  * This has to have a higher priority than migration_notifier in sched/core.c.
1124  */
1125 #define perf_cpu_notifier(fn)                                           \
1126 do {                                                                    \
1127         static struct notifier_block fn##_nb =                          \
1128                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1129         unsigned long cpu = smp_processor_id();                         \
1130         unsigned long flags;                                            \
1131                                                                         \
1132         cpu_notifier_register_begin();                                  \
1133         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1134                 (void *)(unsigned long)cpu);                            \
1135         local_irq_save(flags);                                          \
1136         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1137                 (void *)(unsigned long)cpu);                            \
1138         local_irq_restore(flags);                                       \
1139         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1140                 (void *)(unsigned long)cpu);                            \
1141         __register_cpu_notifier(&fn##_nb);                              \
1142         cpu_notifier_register_done();                                   \
1143 } while (0)
1144
1145 /*
1146  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1147  * callback for already online CPUs.
1148  */
1149 #define __perf_cpu_notifier(fn)                                         \
1150 do {                                                                    \
1151         static struct notifier_block fn##_nb =                          \
1152                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1153                                                                         \
1154         __register_cpu_notifier(&fn##_nb);                              \
1155 } while (0)
1156
1157 struct perf_pmu_events_attr {
1158         struct device_attribute attr;
1159         u64 id;
1160         const char *event_str;
1161 };
1162
1163 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1164                               char *page);
1165
1166 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1167 static struct perf_pmu_events_attr _var = {                             \
1168         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1169         .id   =  _id,                                                   \
1170 };
1171
1172 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1173 static struct perf_pmu_events_attr _var = {                                 \
1174         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1175         .id             = 0,                                                \
1176         .event_str      = _str,                                             \
1177 };
1178
1179 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1180 static ssize_t                                                          \
1181 _name##_show(struct device *dev,                                        \
1182                                struct device_attribute *attr,           \
1183                                char *page)                              \
1184 {                                                                       \
1185         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1186         return sprintf(page, _format "\n");                             \
1187 }                                                                       \
1188                                                                         \
1189 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1190
1191 #endif /* _LINUX_PERF_EVENT_H */