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         /*
 144          * If the event is a per task event, this will point to the task in
 145          * question. See the comment in perf_event_alloc().
 146          */
 147         struct task_struct              *target;
 148
 149 /*
 150  * hw_perf_event::state flags; used to track the PERF_EF_* state.
 151  */
 152 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 153 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 154 #define PERF_HES_ARCH           0x04
 155
 156         int                             state;
 157
 158         /*
 159          * The last observed hardware counter value, updated with a
 160          * local64_cmpxchg() such that pmu::read() can be called nested.
 161          */
 162         local64_t                       prev_count;
 163
 164         /*
 165          * The period to start the next sample with.
 166          */
 167         u64                             sample_period;
 168
 169         /*
 170          * The period we started this sample with.
 171          */
 172         u64                             last_period;
 173
 174         /*
 175          * However much is left of the current period; note that this is
 176          * a full 64bit value and allows for generation of periods longer
 177          * than hardware might allow.
 178          */
 179         local64_t                       period_left;
 180
 181         /*
 182          * State for throttling the event, see __perf_event_overflow() and
 183          * perf_adjust_freq_unthr_context().
 184          */
 185         u64                             interrupts_seq;
 186         u64                             interrupts;
 187
 188         /*
 189          * State for freq target events, see __perf_event_overflow() and
 190          * perf_adjust_freq_unthr_context().
 191          */
 192         u64                             freq_time_stamp;
 193         u64                             freq_count_stamp;
 194 #endif
 195 };
 196
 197 struct perf_event;
 198
 199 /*
 200  * Common implementation detail of pmu::{start,commit,cancel}_txn
 201  */
 202 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
 203 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
 204
 205 /**
 206  * pmu::capabilities flags
 207  */
 208 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 209 #define PERF_PMU_CAP_NO_NMI                     0x02
 210 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
 211 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
 212 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
 213 #define PERF_PMU_CAP_ITRACE                     0x20
 214
 215 /**
 216  * struct pmu - generic performance monitoring unit
 217  */
 218 struct pmu {
 219         struct list_head                entry;
 220
 221         struct module                   *module;
 222         struct device                   *dev;
 223         const struct attribute_group    **attr_groups;
 224         const char                      *name;
 225         int                             type;
 226
 227         /*
 228          * various common per-pmu feature flags
 229          */
 230         int                             capabilities;
 231
 232         int * __percpu                  pmu_disable_count;
 233         struct perf_cpu_context * __percpu pmu_cpu_context;
 234         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 235         int                             task_ctx_nr;
 236         int                             hrtimer_interval_ms;
 237
 238         /*
 239          * Fully disable/enable this PMU, can be used to protect from the PMI
 240          * as well as for lazy/batch writing of the MSRs.
 241          */
 242         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 243         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 244
 245         /*
 246          * Try and initialize the event for this PMU.
 247          *
 248          * Returns:
 249          *  -ENOENT     -- @event is not for this PMU
 250          *
 251          *  -ENODEV     -- @event is for this PMU but PMU not present
 252          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
 253          *  -EINVAL     -- @event is for this PMU but @event is not valid
 254          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
 255          *  -EACCESS    -- @event is for this PMU, @event is valid, but no privilidges
 256          *
 257          *  0           -- @event is for this PMU and valid
 258          *
 259          * Other error return values are allowed.
 260          */
 261         int (*event_init)               (struct perf_event *event);
 262
 263         /*
 264          * Notification that the event was mapped or unmapped.  Called
 265          * in the context of the mapping task.
 266          */
 267         void (*event_mapped)            (struct perf_event *event); /*optional*/
 268         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 269
 270         /*
 271          * Flags for ->add()/->del()/ ->start()/->stop(). There are
 272          * matching hw_perf_event::state flags.
 273          */
 274 #define PERF_EF_START   0x01            /* start the counter when adding    */
 275 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 276 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 277
 278         /*
 279          * Adds/Removes a counter to/from the PMU, can be done inside a
 280          * transaction, see the ->*_txn() methods.
 281          *
 282          * The add/del callbacks will reserve all hardware resources required
 283          * to service the event, this includes any counter constraint
 284          * scheduling etc.
 285          *
 286          * Called with IRQs disabled and the PMU disabled on the CPU the event
 287          * is on.
 288          *
 289          * ->add() called without PERF_EF_START should result in the same state
 290          *  as ->add() followed by ->stop().
 291          *
 292          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
 293          *  ->stop() that must deal with already being stopped without
 294          *  PERF_EF_UPDATE.
 295          */
 296         int  (*add)                     (struct perf_event *event, int flags);
 297         void (*del)                     (struct perf_event *event, int flags);
 298
 299         /*
 300          * Starts/Stops a counter present on the PMU.
 301          *
 302          * The PMI handler should stop the counter when perf_event_overflow()
 303          * returns !0. ->start() will be used to continue.
 304          *
 305          * Also used to change the sample period.
 306          *
 307          * Called with IRQs disabled and the PMU disabled on the CPU the event
 308          * is on -- will be called from NMI context with the PMU generates
 309          * NMIs.
 310          *
 311          * ->stop() with PERF_EF_UPDATE will read the counter and update
 312          *  period/count values like ->read() would.
 313          *
 314          * ->start() with PERF_EF_RELOAD will reprogram the the counter
 315          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
 316          */
 317         void (*start)                   (struct perf_event *event, int flags);
 318         void (*stop)                    (struct perf_event *event, int flags);
 319
 320         /*
 321          * Updates the counter value of the event.
 322          *
 323          * For sampling capable PMUs this will also update the software period
 324          * hw_perf_event::period_left field.
 325          */
 326         void (*read)                    (struct perf_event *event);
 327
 328         /*
 329          * Group events scheduling is treated as a transaction, add
 330          * group events as a whole and perform one schedulability test.
 331          * If the test fails, roll back the whole group
 332          *
 333          * Start the transaction, after this ->add() doesn't need to
 334          * do schedulability tests.
 335          *
 336          * Optional.
 337          */
 338         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
 339         /*
 340          * If ->start_txn() disabled the ->add() schedulability test
 341          * then ->commit_txn() is required to perform one. On success
 342          * the transaction is closed. On error the transaction is kept
 343          * open until ->cancel_txn() is called.
 344          *
 345          * Optional.
 346          */
 347         int  (*commit_txn)              (struct pmu *pmu);
 348         /*
 349          * Will cancel the transaction, assumes ->del() is called
 350          * for each successful ->add() during the transaction.
 351          *
 352          * Optional.
 353          */
 354         void (*cancel_txn)              (struct pmu *pmu);
 355
 356         /*
 357          * Will return the value for perf_event_mmap_page::index for this event,
 358          * if no implementation is provided it will default to: event->hw.idx + 1.
 359          */
 360         int (*event_idx)                (struct perf_event *event); /*optional */
 361
 362         /*
 363          * context-switches callback
 364          */
 365         void (*sched_task)              (struct perf_event_context *ctx,
 366                                         bool sched_in);
 367         /*
 368          * PMU specific data size
 369          */
 370         size_t                          task_ctx_size;
 371
 372
 373         /*
 374          * Return the count value for a counter.
 375          */
 376         u64 (*count)                    (struct perf_event *event); /*optional*/
 377
 378         /*
 379          * Set up pmu-private data structures for an AUX area
 380          */
 381         void *(*setup_aux)              (struct perf_event *event, void **pages,
 382                                          int nr_pages, bool overwrite);
 383                                         /* optional */
 384
 385         /*
 386          * Free pmu-private AUX data structures
 387          */
 388         void (*free_aux)                (void *aux); /* optional */
 389
 390         /*
 391          * Filter events for PMU-specific reasons.
 392          */
 393         int (*filter_match)             (struct perf_event *event); /* optional */
 394
 395         /*
 396          * Initial, PMU driver specific configuration.
 397          */
 398         int (*get_drv_configs)          (struct perf_event *event,
 399                                          void __user *arg); /* optional */
 400         void (*free_drv_configs)        (struct perf_event *event);
 401                                         /* optional */
 402 };
 403
 404 /**
 405  * enum perf_event_active_state - the states of a event
 406  */
 407 enum perf_event_active_state {
 408         PERF_EVENT_STATE_EXIT           = -3,
 409         PERF_EVENT_STATE_ERROR          = -2,
 410         PERF_EVENT_STATE_OFF            = -1,
 411         PERF_EVENT_STATE_INACTIVE       =  0,
 412         PERF_EVENT_STATE_ACTIVE         =  1,
 413 };
 414
 415 struct file;
 416 struct perf_sample_data;
 417
 418 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 419                                         struct perf_sample_data *,
 420                                         struct pt_regs *regs);
 421
 422 enum perf_group_flag {
 423         PERF_GROUP_SOFTWARE             = 0x1,
 424 };
 425
 426 #define SWEVENT_HLIST_BITS              8
 427 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 428
 429 struct swevent_hlist {
 430         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 431         struct rcu_head                 rcu_head;
 432 };
 433
 434 #define PERF_ATTACH_CONTEXT     0x01
 435 #define PERF_ATTACH_GROUP       0x02
 436 #define PERF_ATTACH_TASK        0x04
 437 #define PERF_ATTACH_TASK_DATA   0x08
 438
 439 struct perf_cgroup;
 440 struct ring_buffer;
 441
 442 /**
 443  * struct perf_event - performance event kernel representation:
 444  */
 445 struct perf_event {
 446 #ifdef CONFIG_PERF_EVENTS
 447         /*
 448          * entry onto perf_event_context::event_list;
 449          *   modifications require ctx->lock
 450          *   RCU safe iterations.
 451          */
 452         struct list_head                event_entry;
 453
 454         /*
 455          * XXX: group_entry and sibling_list should be mutually exclusive;
 456          * either you're a sibling on a group, or you're the group leader.
 457          * Rework the code to always use the same list element.
 458          *
 459          * Locked for modification by both ctx->mutex and ctx->lock; holding
 460          * either sufficies for read.
 461          */
 462         struct list_head                group_entry;
 463         struct list_head                sibling_list;
 464
 465         /*
 466          * We need storage to track the entries in perf_pmu_migrate_context; we
 467          * cannot use the event_entry because of RCU and we want to keep the
 468          * group in tact which avoids us using the other two entries.
 469          */
 470         struct list_head                migrate_entry;
 471
 472         struct hlist_node               hlist_entry;
 473         struct list_head                active_entry;
 474         int                             nr_siblings;
 475         int                             group_flags;
 476         struct perf_event               *group_leader;
 477         struct pmu                      *pmu;
 478
 479         enum perf_event_active_state    state;
 480         unsigned int                    attach_state;
 481         local64_t                       count;
 482         atomic64_t                      child_count;
 483
 484         /*
 485          * These are the total time in nanoseconds that the event
 486          * has been enabled (i.e. eligible to run, and the task has
 487          * been scheduled in, if this is a per-task event)
 488          * and running (scheduled onto the CPU), respectively.
 489          *
 490          * They are computed from tstamp_enabled, tstamp_running and
 491          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 492          */
 493         u64                             total_time_enabled;
 494         u64                             total_time_running;
 495
 496         /*
 497          * These are timestamps used for computing total_time_enabled
 498          * and total_time_running when the event is in INACTIVE or
 499          * ACTIVE state, measured in nanoseconds from an arbitrary point
 500          * in time.
 501          * tstamp_enabled: the notional time when the event was enabled
 502          * tstamp_running: the notional time when the event was scheduled on
 503          * tstamp_stopped: in INACTIVE state, the notional time when the
 504          *      event was scheduled off.
 505          */
 506         u64                             tstamp_enabled;
 507         u64                             tstamp_running;
 508         u64                             tstamp_stopped;
 509
 510         /*
 511          * timestamp shadows the actual context timing but it can
 512          * be safely used in NMI interrupt context. It reflects the
 513          * context time as it was when the event was last scheduled in.
 514          *
 515          * ctx_time already accounts for ctx->timestamp. Therefore to
 516          * compute ctx_time for a sample, simply add perf_clock().
 517          */
 518         u64                             shadow_ctx_time;
 519
 520         struct perf_event_attr          attr;
 521         u16                             header_size;
 522         u16                             id_header_size;
 523         u16                             read_size;
 524         struct hw_perf_event            hw;
 525
 526         struct perf_event_context       *ctx;
 527         atomic_long_t                   refcount;
 528
 529         /*
 530          * These accumulate total time (in nanoseconds) that children
 531          * events have been enabled and running, respectively.
 532          */
 533         atomic64_t                      child_total_time_enabled;
 534         atomic64_t                      child_total_time_running;
 535
 536         /*
 537          * Protect attach/detach and child_list:
 538          */
 539         struct mutex                    child_mutex;
 540         struct list_head                child_list;
 541         struct perf_event               *parent;
 542
 543         int                             oncpu;
 544         int                             cpu;
 545
 546         struct list_head                owner_entry;
 547         struct task_struct              *owner;
 548
 549         /* mmap bits */
 550         struct mutex                    mmap_mutex;
 551         atomic_t                        mmap_count;
 552
 553         struct ring_buffer              *rb;
 554         struct list_head                rb_entry;
 555         unsigned long                   rcu_batches;
 556         int                             rcu_pending;
 557
 558         /* poll related */
 559         wait_queue_head_t               waitq;
 560         struct fasync_struct            *fasync;
 561
 562         /* delayed work for NMIs and such */
 563         int                             pending_wakeup;
 564         int                             pending_kill;
 565         int                             pending_disable;
 566         struct irq_work                 pending;
 567
 568         atomic_t                        event_limit;
 569         struct list_head                drv_configs;
 570
 571         void (*destroy)(struct perf_event *);
 572         struct rcu_head                 rcu_head;
 573
 574         struct pid_namespace            *ns;
 575         u64                             id;
 576
 577         u64                             (*clock)(void);
 578         perf_overflow_handler_t         overflow_handler;
 579         void                            *overflow_handler_context;
 580
 581 #ifdef CONFIG_EVENT_TRACING
 582         struct trace_event_call         *tp_event;
 583         struct event_filter             *filter;
 584 #ifdef CONFIG_FUNCTION_TRACER
 585         struct ftrace_ops               ftrace_ops;
 586 #endif
 587 #endif
 588
 589 #ifdef CONFIG_CGROUP_PERF
 590         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 591         int                             cgrp_defer_enabled;
 592 #endif
 593
 594 #endif /* CONFIG_PERF_EVENTS */
 595 };
 596
 597 /**
 598  * struct perf_event_context - event context structure
 599  *
 600  * Used as a container for task events and CPU events as well:
 601  */
 602 struct perf_event_context {
 603         struct pmu                      *pmu;
 604         /*
 605          * Protect the states of the events in the list,
 606          * nr_active, and the list:
 607          */
 608         raw_spinlock_t                  lock;
 609         /*
 610          * Protect the list of events.  Locking either mutex or lock
 611          * is sufficient to ensure the list doesn't change; to change
 612          * the list you need to lock both the mutex and the spinlock.
 613          */
 614         struct mutex                    mutex;
 615
 616         struct list_head                active_ctx_list;
 617         struct list_head                pinned_groups;
 618         struct list_head                flexible_groups;
 619         struct list_head                event_list;
 620         int                             nr_events;
 621         int                             nr_active;
 622         int                             is_active;
 623         int                             nr_stat;
 624         int                             nr_freq;
 625         int                             rotate_disable;
 626         atomic_t                        refcount;
 627         struct task_struct              *task;
 628
 629         /*
 630          * Context clock, runs when context enabled.
 631          */
 632         u64                             time;
 633         u64                             timestamp;
 634
 635         /*
 636          * These fields let us detect when two contexts have both
 637          * been cloned (inherited) from a common ancestor.
 638          */
 639         struct perf_event_context       *parent_ctx;
 640         u64                             parent_gen;
 641         u64                             generation;
 642         int                             pin_count;
 643         int                             nr_cgroups;      /* cgroup evts */
 644         void                            *task_ctx_data; /* pmu specific data */
 645         struct rcu_head                 rcu_head;
 646
 647         struct delayed_work             orphans_remove;
 648         bool                            orphans_remove_sched;
 649 };
 650
 651 /*
 652  * Number of contexts where an event can trigger:
 653  *      task, softirq, hardirq, nmi.
 654  */
 655 #define PERF_NR_CONTEXTS        4
 656
 657 /**
 658  * struct perf_event_cpu_context - per cpu event context structure
 659  */
 660 struct perf_cpu_context {
 661         struct perf_event_context       ctx;
 662         struct perf_event_context       *task_ctx;
 663         int                             active_oncpu;
 664         int                             exclusive;
 665
 666         raw_spinlock_t                  hrtimer_lock;
 667         struct hrtimer                  hrtimer;
 668         ktime_t                         hrtimer_interval;
 669         unsigned int                    hrtimer_active;
 670
 671         struct pmu                      *unique_pmu;
 672         struct perf_cgroup              *cgrp;
 673 };
 674
 675 struct perf_output_handle {
 676         struct perf_event               *event;
 677         struct ring_buffer              *rb;
 678         unsigned long                   wakeup;
 679         unsigned long                   size;
 680         union {
 681                 void                    *addr;
 682                 unsigned long           head;
 683         };
 684         int                             page;
 685 };
 686
 687 #ifdef CONFIG_CGROUP_PERF
 688
 689 /*
 690  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 691  * This is a per-cpu dynamically allocated data structure.
 692  */
 693 struct perf_cgroup_info {
 694         u64                             time;
 695         u64                             timestamp;
 696 };
 697
 698 struct perf_cgroup {
 699         struct cgroup_subsys_state      css;
 700         struct perf_cgroup_info __percpu *info;
 701 };
 702
 703 /*
 704  * Must ensure cgroup is pinned (css_get) before calling
 705  * this function. In other words, we cannot call this function
 706  * if there is no cgroup event for the current CPU context.
 707  */
 708 static inline struct perf_cgroup *
 709 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 710 {
 711         return container_of(task_css_check(task, perf_event_cgrp_id,
 712                                            ctx ? lockdep_is_held(&ctx->lock)
 713                                                : true),
 714                             struct perf_cgroup, css);
 715 }
 716 #endif /* CONFIG_CGROUP_PERF */
 717
 718 #ifdef CONFIG_PERF_EVENTS
 719
 720 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 721                                    struct perf_event *event);
 722 extern void perf_aux_output_end(struct perf_output_handle *handle,
 723                                 unsigned long size, bool truncated);
 724 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 725                                 unsigned long size);
 726 extern void *perf_get_aux(struct perf_output_handle *handle);
 727
 728 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 729 extern void perf_pmu_unregister(struct pmu *pmu);
 730
 731 extern int perf_num_counters(void);
 732 extern const char *perf_pmu_name(void);
 733 extern void __perf_event_task_sched_in(struct task_struct *prev,
 734                                        struct task_struct *task);
 735 extern void __perf_event_task_sched_out(struct task_struct *prev,
 736                                         struct task_struct *next);
 737 extern int perf_event_init_task(struct task_struct *child);
 738 extern void perf_event_exit_task(struct task_struct *child);
 739 extern void perf_event_free_task(struct task_struct *task);
 740 extern void perf_event_delayed_put(struct task_struct *task);
 741 extern struct perf_event *perf_event_get(unsigned int fd);
 742 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 743 extern void perf_event_print_debug(void);
 744 extern void perf_pmu_disable(struct pmu *pmu);
 745 extern void perf_pmu_enable(struct pmu *pmu);
 746 extern void perf_sched_cb_dec(struct pmu *pmu);
 747 extern void perf_sched_cb_inc(struct pmu *pmu);
 748 extern int perf_event_task_disable(void);
 749 extern int perf_event_task_enable(void);
 750 extern int perf_event_refresh(struct perf_event *event, int refresh);
 751 extern void perf_event_update_userpage(struct perf_event *event);
 752 extern int perf_event_release_kernel(struct perf_event *event);
 753 extern struct perf_event *
 754 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 755                                 int cpu,
 756                                 struct task_struct *task,
 757                                 perf_overflow_handler_t callback,
 758                                 void *context);
 759 extern void perf_pmu_migrate_context(struct pmu *pmu,
 760                                 int src_cpu, int dst_cpu);
 761 extern u64 perf_event_read_local(struct perf_event *event);
 762 extern u64 perf_event_read_value(struct perf_event *event,
 763                                  u64 *enabled, u64 *running);
 764
 765
 766 struct perf_sample_data {
 767         /*
 768          * Fields set by perf_sample_data_init(), group so as to
 769          * minimize the cachelines touched.
 770          */
 771         u64                             addr;
 772         struct perf_raw_record          *raw;
 773         struct perf_branch_stack        *br_stack;
 774         u64                             period;
 775         u64                             weight;
 776         u64                             txn;
 777         union  perf_mem_data_src        data_src;
 778
 779         /*
 780          * The other fields, optionally {set,used} by
 781          * perf_{prepare,output}_sample().
 782          */
 783         u64                             type;
 784         u64                             ip;
 785         struct {
 786                 u32     pid;
 787                 u32     tid;
 788         }                               tid_entry;
 789         u64                             time;
 790         u64                             id;
 791         u64                             stream_id;
 792         struct {
 793                 u32     cpu;
 794                 u32     reserved;
 795         }                               cpu_entry;
 796         struct perf_callchain_entry     *callchain;
 797
 798         /*
 799          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 800          * on arch details.
 801          */
 802         struct perf_regs                regs_user;
 803         struct pt_regs                  regs_user_copy;
 804
 805         struct perf_regs                regs_intr;
 806         u64                             stack_user_size;
 807 } ____cacheline_aligned;
 808
 809 /* default value for data source */
 810 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 811                     PERF_MEM_S(LVL, NA)   |\
 812                     PERF_MEM_S(SNOOP, NA) |\
 813                     PERF_MEM_S(LOCK, NA)  |\
 814                     PERF_MEM_S(TLB, NA))
 815
 816 static inline void perf_sample_data_init(struct perf_sample_data *data,
 817                                          u64 addr, u64 period)
 818 {
 819         /* remaining struct members initialized in perf_prepare_sample() */
 820         data->addr = addr;
 821         data->raw  = NULL;
 822         data->br_stack = NULL;
 823         data->period = period;
 824         data->weight = 0;
 825         data->data_src.val = PERF_MEM_NA;
 826         data->txn = 0;
 827 }
 828
 829 extern void perf_output_sample(struct perf_output_handle *handle,
 830                                struct perf_event_header *header,
 831                                struct perf_sample_data *data,
 832                                struct perf_event *event);
 833 extern void perf_prepare_sample(struct perf_event_header *header,
 834                                 struct perf_sample_data *data,
 835                                 struct perf_event *event,
 836                                 struct pt_regs *regs);
 837
 838 extern int perf_event_overflow(struct perf_event *event,
 839                                  struct perf_sample_data *data,
 840                                  struct pt_regs *regs);
 841
 842 extern void perf_event_output(struct perf_event *event,
 843                                 struct perf_sample_data *data,
 844                                 struct pt_regs *regs);
 845
 846 extern void
 847 perf_event_header__init_id(struct perf_event_header *header,
 848                            struct perf_sample_data *data,
 849                            struct perf_event *event);
 850 extern void
 851 perf_event__output_id_sample(struct perf_event *event,
 852                              struct perf_output_handle *handle,
 853                              struct perf_sample_data *sample);
 854
 855 extern void
 856 perf_log_lost_samples(struct perf_event *event, u64 lost);
 857
 858 static inline bool is_sampling_event(struct perf_event *event)
 859 {
 860         return event->attr.sample_period != 0;
 861 }
 862
 863 /*
 864  * Return 1 for a software event, 0 for a hardware event
 865  */
 866 static inline int is_software_event(struct perf_event *event)
 867 {
 868         return event->pmu->task_ctx_nr == perf_sw_context;
 869 }
 870
 871 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 872
 873 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 874 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 875
 876 #ifndef perf_arch_fetch_caller_regs
 877 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 878 #endif
 879
 880 /*
 881  * Take a snapshot of the regs. Skip ip and frame pointer to
 882  * the nth caller. We only need a few of the regs:
 883  * - ip for PERF_SAMPLE_IP
 884  * - cs for user_mode() tests
 885  * - bp for callchains
 886  * - eflags, for future purposes, just in case
 887  */
 888 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 889 {
 890         memset(regs, 0, sizeof(*regs));
 891
 892         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 893 }
 894
 895 static __always_inline void
 896 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 897 {
 898         if (static_key_false(&perf_swevent_enabled[event_id]))
 899                 __perf_sw_event(event_id, nr, regs, addr);
 900 }
 901
 902 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 903
 904 /*
 905  * 'Special' version for the scheduler, it hard assumes no recursion,
 906  * which is guaranteed by us not actually scheduling inside other swevents
 907  * because those disable preemption.
 908  */
 909 static __always_inline void
 910 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 911 {
 912         if (static_key_false(&perf_swevent_enabled[event_id])) {
 913                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 914
 915                 perf_fetch_caller_regs(regs);
 916                 ___perf_sw_event(event_id, nr, regs, addr);
 917         }
 918 }
 919
 920 extern struct static_key_deferred perf_sched_events;
 921
 922 static __always_inline bool
 923 perf_sw_migrate_enabled(void)
 924 {
 925         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
 926                 return true;
 927         return false;
 928 }
 929
 930 static inline void perf_event_task_migrate(struct task_struct *task)
 931 {
 932         if (perf_sw_migrate_enabled())
 933                 task->sched_migrated = 1;
 934 }
 935
 936 static inline void perf_event_task_sched_in(struct task_struct *prev,
 937                                             struct task_struct *task)
 938 {
 939         if (static_key_false(&perf_sched_events.key))
 940                 __perf_event_task_sched_in(prev, task);
 941
 942         if (perf_sw_migrate_enabled() && task->sched_migrated) {
 943                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 944
 945                 perf_fetch_caller_regs(regs);
 946                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
 947                 task->sched_migrated = 0;
 948         }
 949 }
 950
 951 static inline void perf_event_task_sched_out(struct task_struct *prev,
 952                                              struct task_struct *next)
 953 {
 954         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 955
 956         if (static_key_false(&perf_sched_events.key))
 957                 __perf_event_task_sched_out(prev, next);
 958 }
 959
 960 static inline u64 __perf_event_count(struct perf_event *event)
 961 {
 962         return local64_read(&event->count) + atomic64_read(&event->child_count);
 963 }
 964
 965 extern void perf_event_mmap(struct vm_area_struct *vma);
 966 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 967 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 968 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 969
 970 extern void perf_event_exec(void);
 971 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 972 extern void perf_event_fork(struct task_struct *tsk);
 973
 974 /* Callchains */
 975 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 976
 977 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 978 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 979
 980 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 981 {
 982         if (entry->nr < PERF_MAX_STACK_DEPTH)
 983                 entry->ip[entry->nr++] = ip;
 984 }
 985
 986 extern int sysctl_perf_event_paranoid;
 987 extern int sysctl_perf_event_mlock;
 988 extern int sysctl_perf_event_sample_rate;
 989 extern int sysctl_perf_cpu_time_max_percent;
 990
 991 extern void perf_sample_event_took(u64 sample_len_ns);
 992
 993 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 994                 void __user *buffer, size_t *lenp,
 995                 loff_t *ppos);
 996 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 997                 void __user *buffer, size_t *lenp,
 998                 loff_t *ppos);
 999
1000
1001 static inline bool perf_paranoid_any(void)
1002 {
1003         return sysctl_perf_event_paranoid > 2;
1004 }
1005
1006 static inline bool perf_paranoid_tracepoint_raw(void)
1007 {
1008         return sysctl_perf_event_paranoid > -1;
1009 }
1010
1011 static inline bool perf_paranoid_cpu(void)
1012 {
1013         return sysctl_perf_event_paranoid > 0;
1014 }
1015
1016 static inline bool perf_paranoid_kernel(void)
1017 {
1018         return sysctl_perf_event_paranoid > 1;
1019 }
1020
1021 extern void perf_event_init(void);
1022 extern void perf_tp_event(u64 addr, u64 count, void *record,
1023                           int entry_size, struct pt_regs *regs,
1024                           struct hlist_head *head, int rctx,
1025                           struct task_struct *task);
1026 extern void perf_bp_event(struct perf_event *event, void *data);
1027
1028 #ifndef perf_misc_flags
1029 # define perf_misc_flags(regs) \
1030                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1031 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1032 #endif
1033
1034 static inline bool has_branch_stack(struct perf_event *event)
1035 {
1036         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1037 }
1038
1039 static inline bool needs_branch_stack(struct perf_event *event)
1040 {
1041         return event->attr.branch_sample_type != 0;
1042 }
1043
1044 static inline bool has_aux(struct perf_event *event)
1045 {
1046         return event->pmu->setup_aux;
1047 }
1048
1049 extern int perf_output_begin(struct perf_output_handle *handle,
1050                              struct perf_event *event, unsigned int size);
1051 extern void perf_output_end(struct perf_output_handle *handle);
1052 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1053                              const void *buf, unsigned int len);
1054 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1055                                      unsigned int len);
1056 extern int perf_swevent_get_recursion_context(void);
1057 extern void perf_swevent_put_recursion_context(int rctx);
1058 extern u64 perf_swevent_set_period(struct perf_event *event);
1059 extern void perf_event_enable(struct perf_event *event);
1060 extern void perf_event_disable(struct perf_event *event);
1061 extern int __perf_event_disable(void *info);
1062 extern void perf_event_task_tick(void);
1063 #else /* !CONFIG_PERF_EVENTS: */
1064 static inline void *
1065 perf_aux_output_begin(struct perf_output_handle *handle,
1066                       struct perf_event *event)                         { return NULL; }
1067 static inline void
1068 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1069                     bool truncated)                                     { }
1070 static inline int
1071 perf_aux_output_skip(struct perf_output_handle *handle,
1072                      unsigned long size)                                { return -EINVAL; }
1073 static inline void *
1074 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1075 static inline void
1076 perf_event_task_migrate(struct task_struct *task)                       { }
1077 static inline void
1078 perf_event_task_sched_in(struct task_struct *prev,
1079                          struct task_struct *task)                      { }
1080 static inline void
1081 perf_event_task_sched_out(struct task_struct *prev,
1082                           struct task_struct *next)                     { }
1083 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1084 static inline void perf_event_exit_task(struct task_struct *child)      { }
1085 static inline void perf_event_free_task(struct task_struct *task)       { }
1086 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1087 static inline struct perf_event *perf_event_get(unsigned int fd)        { return ERR_PTR(-EINVAL); }
1088 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1089 {
1090         return ERR_PTR(-EINVAL);
1091 }
1092 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1093 static inline void perf_event_print_debug(void)                         { }
1094 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1095 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1096 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1097 {
1098         return -EINVAL;
1099 }
1100
1101 static inline void
1102 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1103 static inline void
1104 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1105 static inline void
1106 perf_bp_event(struct perf_event *event, void *data)                     { }
1107
1108 static inline int perf_register_guest_info_callbacks
1109 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1110 static inline int perf_unregister_guest_info_callbacks
1111 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1112
1113 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1114 static inline void perf_event_exec(void)                                { }
1115 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1116 static inline void perf_event_fork(struct task_struct *tsk)             { }
1117 static inline void perf_event_init(void)                                { }
1118 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1119 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1120 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1121 static inline void perf_event_enable(struct perf_event *event)          { }
1122 static inline void perf_event_disable(struct perf_event *event)         { }
1123 static inline int __perf_event_disable(void *info)                      { return -1; }
1124 static inline void perf_event_task_tick(void)                           { }
1125 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1126 #endif
1127
1128 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
1129 extern bool perf_event_can_stop_tick(void);
1130 #else
1131 static inline bool perf_event_can_stop_tick(void)                       { return true; }
1132 #endif
1133
1134 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1135 extern void perf_restore_debug_store(void);
1136 #else
1137 static inline void perf_restore_debug_store(void)                       { }
1138 #endif
1139
1140 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1141
1142 /*
1143  * This has to have a higher priority than migration_notifier in sched/core.c.
1144  */
1145 #define perf_cpu_notifier(fn)                                           \
1146 do {                                                                    \
1147         static struct notifier_block fn##_nb =                          \
1148                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1149         unsigned long cpu = smp_processor_id();                         \
1150         unsigned long flags;                                            \
1151                                                                         \
1152         cpu_notifier_register_begin();                                  \
1153         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1154                 (void *)(unsigned long)cpu);                            \
1155         local_irq_save(flags);                                          \
1156         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1157                 (void *)(unsigned long)cpu);                            \
1158         local_irq_restore(flags);                                       \
1159         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1160                 (void *)(unsigned long)cpu);                            \
1161         __register_cpu_notifier(&fn##_nb);                              \
1162         cpu_notifier_register_done();                                   \
1163 } while (0)
1164
1165 /*
1166  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1167  * callback for already online CPUs.
1168  */
1169 #define __perf_cpu_notifier(fn)                                         \
1170 do {                                                                    \
1171         static struct notifier_block fn##_nb =                          \
1172                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1173                                                                         \
1174         __register_cpu_notifier(&fn##_nb);                              \
1175 } while (0)
1176
1177 struct perf_pmu_events_attr {
1178         struct device_attribute attr;
1179         u64 id;
1180         const char *event_str;
1181 };
1182
1183 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1184                               char *page);
1185
1186 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1187 static struct perf_pmu_events_attr _var = {                             \
1188         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1189         .id   =  _id,                                                   \
1190 };
1191
1192 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1193 static struct perf_pmu_events_attr _var = {                                 \
1194         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1195         .id             = 0,                                                \
1196         .event_str      = _str,                                             \
1197 };
1198
1199 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1200 static ssize_t                                                          \
1201 _name##_show(struct device *dev,                                        \
1202                                struct device_attribute *attr,           \
1203                                char *page)                              \
1204 {                                                                       \
1205         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1206         return sprintf(page, _format "\n");                             \
1207 }                                                                       \
1208                                                                         \
1209 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1210
1211 #endif /* _LINUX_PERF_EVENT_H */