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)              (struct perf_event *event, 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          * Initial, PMU driver specific configuration.
 398          */
 399         int (*get_drv_configs)          (struct perf_event *event,
 400                                          void __user *arg); /* optional */
 401         void (*free_drv_configs)        (struct perf_event *event);
 402                                         /* optional */
 403 };
 404
 405 /**
 406  * enum perf_event_active_state - the states of a event
 407  */
 408 enum perf_event_active_state {
 409         PERF_EVENT_STATE_EXIT           = -3,
 410         PERF_EVENT_STATE_ERROR          = -2,
 411         PERF_EVENT_STATE_OFF            = -1,
 412         PERF_EVENT_STATE_INACTIVE       =  0,
 413         PERF_EVENT_STATE_ACTIVE         =  1,
 414 };
 415
 416 struct file;
 417 struct perf_sample_data;
 418
 419 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 420                                         struct perf_sample_data *,
 421                                         struct pt_regs *regs);
 422
 423 enum perf_group_flag {
 424         PERF_GROUP_SOFTWARE             = 0x1,
 425 };
 426
 427 #define SWEVENT_HLIST_BITS              8
 428 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 429
 430 struct swevent_hlist {
 431         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 432         struct rcu_head                 rcu_head;
 433 };
 434
 435 #define PERF_ATTACH_CONTEXT     0x01
 436 #define PERF_ATTACH_GROUP       0x02
 437 #define PERF_ATTACH_TASK        0x04
 438 #define PERF_ATTACH_TASK_DATA   0x08
 439
 440 struct perf_cgroup;
 441 struct ring_buffer;
 442
 443 /**
 444  * struct perf_event - performance event kernel representation:
 445  */
 446 struct perf_event {
 447 #ifdef CONFIG_PERF_EVENTS
 448         /*
 449          * entry onto perf_event_context::event_list;
 450          *   modifications require ctx->lock
 451          *   RCU safe iterations.
 452          */
 453         struct list_head                event_entry;
 454
 455         /*
 456          * XXX: group_entry and sibling_list should be mutually exclusive;
 457          * either you're a sibling on a group, or you're the group leader.
 458          * Rework the code to always use the same list element.
 459          *
 460          * Locked for modification by both ctx->mutex and ctx->lock; holding
 461          * either sufficies for read.
 462          */
 463         struct list_head                group_entry;
 464         struct list_head                sibling_list;
 465
 466         /*
 467          * We need storage to track the entries in perf_pmu_migrate_context; we
 468          * cannot use the event_entry because of RCU and we want to keep the
 469          * group in tact which avoids us using the other two entries.
 470          */
 471         struct list_head                migrate_entry;
 472
 473         struct hlist_node               hlist_entry;
 474         struct list_head                active_entry;
 475         int                             nr_siblings;
 476         int                             group_flags;
 477         struct perf_event               *group_leader;
 478         struct pmu                      *pmu;
 479
 480         enum perf_event_active_state    state;
 481         unsigned int                    attach_state;
 482         local64_t                       count;
 483         atomic64_t                      child_count;
 484
 485         /*
 486          * These are the total time in nanoseconds that the event
 487          * has been enabled (i.e. eligible to run, and the task has
 488          * been scheduled in, if this is a per-task event)
 489          * and running (scheduled onto the CPU), respectively.
 490          *
 491          * They are computed from tstamp_enabled, tstamp_running and
 492          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 493          */
 494         u64                             total_time_enabled;
 495         u64                             total_time_running;
 496
 497         /*
 498          * These are timestamps used for computing total_time_enabled
 499          * and total_time_running when the event is in INACTIVE or
 500          * ACTIVE state, measured in nanoseconds from an arbitrary point
 501          * in time.
 502          * tstamp_enabled: the notional time when the event was enabled
 503          * tstamp_running: the notional time when the event was scheduled on
 504          * tstamp_stopped: in INACTIVE state, the notional time when the
 505          *      event was scheduled off.
 506          */
 507         u64                             tstamp_enabled;
 508         u64                             tstamp_running;
 509         u64                             tstamp_stopped;
 510
 511         /*
 512          * timestamp shadows the actual context timing but it can
 513          * be safely used in NMI interrupt context. It reflects the
 514          * context time as it was when the event was last scheduled in.
 515          *
 516          * ctx_time already accounts for ctx->timestamp. Therefore to
 517          * compute ctx_time for a sample, simply add perf_clock().
 518          */
 519         u64                             shadow_ctx_time;
 520
 521         struct perf_event_attr          attr;
 522         u16                             header_size;
 523         u16                             id_header_size;
 524         u16                             read_size;
 525         struct hw_perf_event            hw;
 526
 527         struct perf_event_context       *ctx;
 528         atomic_long_t                   refcount;
 529
 530         /*
 531          * These accumulate total time (in nanoseconds) that children
 532          * events have been enabled and running, respectively.
 533          */
 534         atomic64_t                      child_total_time_enabled;
 535         atomic64_t                      child_total_time_running;
 536
 537         /*
 538          * Protect attach/detach and child_list:
 539          */
 540         struct mutex                    child_mutex;
 541         struct list_head                child_list;
 542         struct perf_event               *parent;
 543
 544         int                             oncpu;
 545         int                             cpu;
 546
 547         struct list_head                owner_entry;
 548         struct task_struct              *owner;
 549
 550         /* mmap bits */
 551         struct mutex                    mmap_mutex;
 552         atomic_t                        mmap_count;
 553
 554         struct ring_buffer              *rb;
 555         struct list_head                rb_entry;
 556         unsigned long                   rcu_batches;
 557         int                             rcu_pending;
 558
 559         /* poll related */
 560         wait_queue_head_t               waitq;
 561         struct fasync_struct            *fasync;
 562
 563         /* delayed work for NMIs and such */
 564         int                             pending_wakeup;
 565         int                             pending_kill;
 566         int                             pending_disable;
 567         struct irq_work                 pending;
 568
 569         atomic_t                        event_limit;
 570         struct list_head                drv_configs;
 571
 572         void (*destroy)(struct perf_event *);
 573         struct rcu_head                 rcu_head;
 574
 575         struct pid_namespace            *ns;
 576         u64                             id;
 577
 578         u64                             (*clock)(void);
 579         perf_overflow_handler_t         overflow_handler;
 580         void                            *overflow_handler_context;
 581
 582 #ifdef CONFIG_EVENT_TRACING
 583         struct trace_event_call         *tp_event;
 584         struct event_filter             *filter;
 585 #ifdef CONFIG_FUNCTION_TRACER
 586         struct ftrace_ops               ftrace_ops;
 587 #endif
 588 #endif
 589
 590 #ifdef CONFIG_CGROUP_PERF
 591         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 592         int                             cgrp_defer_enabled;
 593 #endif
 594
 595 #endif /* CONFIG_PERF_EVENTS */
 596 };
 597
 598 /**
 599  * struct perf_event_context - event context structure
 600  *
 601  * Used as a container for task events and CPU events as well:
 602  */
 603 struct perf_event_context {
 604         struct pmu                      *pmu;
 605         /*
 606          * Protect the states of the events in the list,
 607          * nr_active, and the list:
 608          */
 609         raw_spinlock_t                  lock;
 610         /*
 611          * Protect the list of events.  Locking either mutex or lock
 612          * is sufficient to ensure the list doesn't change; to change
 613          * the list you need to lock both the mutex and the spinlock.
 614          */
 615         struct mutex                    mutex;
 616
 617         struct list_head                active_ctx_list;
 618         struct list_head                pinned_groups;
 619         struct list_head                flexible_groups;
 620         struct list_head                event_list;
 621         int                             nr_events;
 622         int                             nr_active;
 623         int                             is_active;
 624         int                             nr_stat;
 625         int                             nr_freq;
 626         int                             rotate_disable;
 627         atomic_t                        refcount;
 628         struct task_struct              *task;
 629
 630         /*
 631          * Context clock, runs when context enabled.
 632          */
 633         u64                             time;
 634         u64                             timestamp;
 635
 636         /*
 637          * These fields let us detect when two contexts have both
 638          * been cloned (inherited) from a common ancestor.
 639          */
 640         struct perf_event_context       *parent_ctx;
 641         u64                             parent_gen;
 642         u64                             generation;
 643         int                             pin_count;
 644         int                             nr_cgroups;      /* cgroup evts */
 645         void                            *task_ctx_data; /* pmu specific data */
 646         struct rcu_head                 rcu_head;
 647
 648         struct delayed_work             orphans_remove;
 649         bool                            orphans_remove_sched;
 650 };
 651
 652 /*
 653  * Number of contexts where an event can trigger:
 654  *      task, softirq, hardirq, nmi.
 655  */
 656 #define PERF_NR_CONTEXTS        4
 657
 658 /**
 659  * struct perf_event_cpu_context - per cpu event context structure
 660  */
 661 struct perf_cpu_context {
 662         struct perf_event_context       ctx;
 663         struct perf_event_context       *task_ctx;
 664         int                             active_oncpu;
 665         int                             exclusive;
 666
 667         raw_spinlock_t                  hrtimer_lock;
 668         struct hrtimer                  hrtimer;
 669         ktime_t                         hrtimer_interval;
 670         unsigned int                    hrtimer_active;
 671
 672         struct pmu                      *unique_pmu;
 673         struct perf_cgroup              *cgrp;
 674 };
 675
 676 struct perf_output_handle {
 677         struct perf_event               *event;
 678         struct ring_buffer              *rb;
 679         unsigned long                   wakeup;
 680         unsigned long                   size;
 681         union {
 682                 void                    *addr;
 683                 unsigned long           head;
 684         };
 685         int                             page;
 686 };
 687
 688 #ifdef CONFIG_CGROUP_PERF
 689
 690 /*
 691  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 692  * This is a per-cpu dynamically allocated data structure.
 693  */
 694 struct perf_cgroup_info {
 695         u64                             time;
 696         u64                             timestamp;
 697 };
 698
 699 struct perf_cgroup {
 700         struct cgroup_subsys_state      css;
 701         struct perf_cgroup_info __percpu *info;
 702 };
 703
 704 /*
 705  * Must ensure cgroup is pinned (css_get) before calling
 706  * this function. In other words, we cannot call this function
 707  * if there is no cgroup event for the current CPU context.
 708  */
 709 static inline struct perf_cgroup *
 710 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 711 {
 712         return container_of(task_css_check(task, perf_event_cgrp_id,
 713                                            ctx ? lockdep_is_held(&ctx->lock)
 714                                                : true),
 715                             struct perf_cgroup, css);
 716 }
 717 #endif /* CONFIG_CGROUP_PERF */
 718
 719 #ifdef CONFIG_PERF_EVENTS
 720
 721 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 722                                    struct perf_event *event);
 723 extern void perf_aux_output_end(struct perf_output_handle *handle,
 724                                 unsigned long size, bool truncated);
 725 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 726                                 unsigned long size);
 727 extern void *perf_get_aux(struct perf_output_handle *handle);
 728
 729 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 730 extern void perf_pmu_unregister(struct pmu *pmu);
 731
 732 extern int perf_num_counters(void);
 733 extern const char *perf_pmu_name(void);
 734 extern void __perf_event_task_sched_in(struct task_struct *prev,
 735                                        struct task_struct *task);
 736 extern void __perf_event_task_sched_out(struct task_struct *prev,
 737                                         struct task_struct *next);
 738 extern int perf_event_init_task(struct task_struct *child);
 739 extern void perf_event_exit_task(struct task_struct *child);
 740 extern void perf_event_free_task(struct task_struct *task);
 741 extern void perf_event_delayed_put(struct task_struct *task);
 742 extern struct perf_event *perf_event_get(unsigned int fd);
 743 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 744 extern void perf_event_print_debug(void);
 745 extern void perf_pmu_disable(struct pmu *pmu);
 746 extern void perf_pmu_enable(struct pmu *pmu);
 747 extern void perf_sched_cb_dec(struct pmu *pmu);
 748 extern void perf_sched_cb_inc(struct pmu *pmu);
 749 extern int perf_event_task_disable(void);
 750 extern int perf_event_task_enable(void);
 751 extern int perf_event_refresh(struct perf_event *event, int refresh);
 752 extern void perf_event_update_userpage(struct perf_event *event);
 753 extern int perf_event_release_kernel(struct perf_event *event);
 754 extern struct perf_event *
 755 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 756                                 int cpu,
 757                                 struct task_struct *task,
 758                                 perf_overflow_handler_t callback,
 759                                 void *context);
 760 extern void perf_pmu_migrate_context(struct pmu *pmu,
 761                                 int src_cpu, int dst_cpu);
 762 extern u64 perf_event_read_local(struct perf_event *event);
 763 extern u64 perf_event_read_value(struct perf_event *event,
 764                                  u64 *enabled, u64 *running);
 765
 766
 767 struct perf_sample_data {
 768         /*
 769          * Fields set by perf_sample_data_init(), group so as to
 770          * minimize the cachelines touched.
 771          */
 772         u64                             addr;
 773         struct perf_raw_record          *raw;
 774         struct perf_branch_stack        *br_stack;
 775         u64                             period;
 776         u64                             weight;
 777         u64                             txn;
 778         union  perf_mem_data_src        data_src;
 779
 780         /*
 781          * The other fields, optionally {set,used} by
 782          * perf_{prepare,output}_sample().
 783          */
 784         u64                             type;
 785         u64                             ip;
 786         struct {
 787                 u32     pid;
 788                 u32     tid;
 789         }                               tid_entry;
 790         u64                             time;
 791         u64                             id;
 792         u64                             stream_id;
 793         struct {
 794                 u32     cpu;
 795                 u32     reserved;
 796         }                               cpu_entry;
 797         struct perf_callchain_entry     *callchain;
 798
 799         /*
 800          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 801          * on arch details.
 802          */
 803         struct perf_regs                regs_user;
 804         struct pt_regs                  regs_user_copy;
 805
 806         struct perf_regs                regs_intr;
 807         u64                             stack_user_size;
 808 } ____cacheline_aligned;
 809
 810 /* default value for data source */
 811 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 812                     PERF_MEM_S(LVL, NA)   |\
 813                     PERF_MEM_S(SNOOP, NA) |\
 814                     PERF_MEM_S(LOCK, NA)  |\
 815                     PERF_MEM_S(TLB, NA))
 816
 817 static inline void perf_sample_data_init(struct perf_sample_data *data,
 818                                          u64 addr, u64 period)
 819 {
 820         /* remaining struct members initialized in perf_prepare_sample() */
 821         data->addr = addr;
 822         data->raw  = NULL;
 823         data->br_stack = NULL;
 824         data->period = period;
 825         data->weight = 0;
 826         data->data_src.val = PERF_MEM_NA;
 827         data->txn = 0;
 828 }
 829
 830 extern void perf_output_sample(struct perf_output_handle *handle,
 831                                struct perf_event_header *header,
 832                                struct perf_sample_data *data,
 833                                struct perf_event *event);
 834 extern void perf_prepare_sample(struct perf_event_header *header,
 835                                 struct perf_sample_data *data,
 836                                 struct perf_event *event,
 837                                 struct pt_regs *regs);
 838
 839 extern int perf_event_overflow(struct perf_event *event,
 840                                  struct perf_sample_data *data,
 841                                  struct pt_regs *regs);
 842
 843 extern void perf_event_output(struct perf_event *event,
 844                                 struct perf_sample_data *data,
 845                                 struct pt_regs *regs);
 846
 847 extern void
 848 perf_event_header__init_id(struct perf_event_header *header,
 849                            struct perf_sample_data *data,
 850                            struct perf_event *event);
 851 extern void
 852 perf_event__output_id_sample(struct perf_event *event,
 853                              struct perf_output_handle *handle,
 854                              struct perf_sample_data *sample);
 855
 856 extern void
 857 perf_log_lost_samples(struct perf_event *event, u64 lost);
 858
 859 static inline bool is_sampling_event(struct perf_event *event)
 860 {
 861         return event->attr.sample_period != 0;
 862 }
 863
 864 /*
 865  * Return 1 for a software event, 0 for a hardware event
 866  */
 867 static inline int is_software_event(struct perf_event *event)
 868 {
 869         return event->pmu->task_ctx_nr == perf_sw_context;
 870 }
 871
 872 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 873
 874 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 875 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 876
 877 #ifndef perf_arch_fetch_caller_regs
 878 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 879 #endif
 880
 881 /*
 882  * Take a snapshot of the regs. Skip ip and frame pointer to
 883  * the nth caller. We only need a few of the regs:
 884  * - ip for PERF_SAMPLE_IP
 885  * - cs for user_mode() tests
 886  * - bp for callchains
 887  * - eflags, for future purposes, just in case
 888  */
 889 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 890 {
 891         memset(regs, 0, sizeof(*regs));
 892
 893         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 894 }
 895
 896 static __always_inline void
 897 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 898 {
 899         if (static_key_false(&perf_swevent_enabled[event_id]))
 900                 __perf_sw_event(event_id, nr, regs, addr);
 901 }
 902
 903 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 904
 905 /*
 906  * 'Special' version for the scheduler, it hard assumes no recursion,
 907  * which is guaranteed by us not actually scheduling inside other swevents
 908  * because those disable preemption.
 909  */
 910 static __always_inline void
 911 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 912 {
 913         if (static_key_false(&perf_swevent_enabled[event_id])) {
 914                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 915
 916                 perf_fetch_caller_regs(regs);
 917                 ___perf_sw_event(event_id, nr, regs, addr);
 918         }
 919 }
 920
 921 extern struct static_key_deferred perf_sched_events;
 922
 923 static __always_inline bool
 924 perf_sw_migrate_enabled(void)
 925 {
 926         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
 927                 return true;
 928         return false;
 929 }
 930
 931 static inline void perf_event_task_migrate(struct task_struct *task)
 932 {
 933         if (perf_sw_migrate_enabled())
 934                 task->sched_migrated = 1;
 935 }
 936
 937 static inline void perf_event_task_sched_in(struct task_struct *prev,
 938                                             struct task_struct *task)
 939 {
 940         if (static_key_false(&perf_sched_events.key))
 941                 __perf_event_task_sched_in(prev, task);
 942
 943         if (perf_sw_migrate_enabled() && task->sched_migrated) {
 944                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 945
 946                 perf_fetch_caller_regs(regs);
 947                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
 948                 task->sched_migrated = 0;
 949         }
 950 }
 951
 952 static inline void perf_event_task_sched_out(struct task_struct *prev,
 953                                              struct task_struct *next)
 954 {
 955         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 956
 957         if (static_key_false(&perf_sched_events.key))
 958                 __perf_event_task_sched_out(prev, next);
 959 }
 960
 961 static inline u64 __perf_event_count(struct perf_event *event)
 962 {
 963         return local64_read(&event->count) + atomic64_read(&event->child_count);
 964 }
 965
 966 extern void perf_event_mmap(struct vm_area_struct *vma);
 967 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 968 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 969 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 970
 971 extern void perf_event_exec(void);
 972 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 973 extern void perf_event_fork(struct task_struct *tsk);
 974
 975 /* Callchains */
 976 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 977
 978 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 979 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 980
 981 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 982 {
 983         if (entry->nr < PERF_MAX_STACK_DEPTH)
 984                 entry->ip[entry->nr++] = ip;
 985 }
 986
 987 extern int sysctl_perf_event_paranoid;
 988 extern int sysctl_perf_event_mlock;
 989 extern int sysctl_perf_event_sample_rate;
 990 extern int sysctl_perf_cpu_time_max_percent;
 991
 992 extern void perf_sample_event_took(u64 sample_len_ns);
 993
 994 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 995                 void __user *buffer, size_t *lenp,
 996                 loff_t *ppos);
 997 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 998                 void __user *buffer, size_t *lenp,
 999                 loff_t *ppos);
1000
1001
1002 static inline bool perf_paranoid_tracepoint_raw(void)
1003 {
1004         return sysctl_perf_event_paranoid > -1;
1005 }
1006
1007 static inline bool perf_paranoid_cpu(void)
1008 {
1009         return sysctl_perf_event_paranoid > 0;
1010 }
1011
1012 static inline bool perf_paranoid_kernel(void)
1013 {
1014         return sysctl_perf_event_paranoid > 1;
1015 }
1016
1017 extern void perf_event_init(void);
1018 extern void perf_tp_event(u64 addr, u64 count, void *record,
1019                           int entry_size, struct pt_regs *regs,
1020                           struct hlist_head *head, int rctx,
1021                           struct task_struct *task);
1022 extern void perf_bp_event(struct perf_event *event, void *data);
1023
1024 #ifndef perf_misc_flags
1025 # define perf_misc_flags(regs) \
1026                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1027 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1028 #endif
1029
1030 static inline bool has_branch_stack(struct perf_event *event)
1031 {
1032         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1033 }
1034
1035 static inline bool needs_branch_stack(struct perf_event *event)
1036 {
1037         return event->attr.branch_sample_type != 0;
1038 }
1039
1040 static inline bool has_aux(struct perf_event *event)
1041 {
1042         return event->pmu->setup_aux;
1043 }
1044
1045 extern int perf_output_begin(struct perf_output_handle *handle,
1046                              struct perf_event *event, unsigned int size);
1047 extern void perf_output_end(struct perf_output_handle *handle);
1048 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1049                              const void *buf, unsigned int len);
1050 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1051                                      unsigned int len);
1052 extern int perf_swevent_get_recursion_context(void);
1053 extern void perf_swevent_put_recursion_context(int rctx);
1054 extern u64 perf_swevent_set_period(struct perf_event *event);
1055 extern void perf_event_enable(struct perf_event *event);
1056 extern void perf_event_disable(struct perf_event *event);
1057 extern int __perf_event_disable(void *info);
1058 extern void perf_event_task_tick(void);
1059 #else /* !CONFIG_PERF_EVENTS: */
1060 static inline void *
1061 perf_aux_output_begin(struct perf_output_handle *handle,
1062                       struct perf_event *event)                         { return NULL; }
1063 static inline void
1064 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1065                     bool truncated)                                     { }
1066 static inline int
1067 perf_aux_output_skip(struct perf_output_handle *handle,
1068                      unsigned long size)                                { return -EINVAL; }
1069 static inline void *
1070 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1071 static inline void
1072 perf_event_task_migrate(struct task_struct *task)                       { }
1073 static inline void
1074 perf_event_task_sched_in(struct task_struct *prev,
1075                          struct task_struct *task)                      { }
1076 static inline void
1077 perf_event_task_sched_out(struct task_struct *prev,
1078                           struct task_struct *next)                     { }
1079 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1080 static inline void perf_event_exit_task(struct task_struct *child)      { }
1081 static inline void perf_event_free_task(struct task_struct *task)       { }
1082 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1083 static inline struct perf_event *perf_event_get(unsigned int fd)        { return ERR_PTR(-EINVAL); }
1084 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1085 {
1086         return ERR_PTR(-EINVAL);
1087 }
1088 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1089 static inline void perf_event_print_debug(void)                         { }
1090 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1091 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1092 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1093 {
1094         return -EINVAL;
1095 }
1096
1097 static inline void
1098 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1099 static inline void
1100 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1101 static inline void
1102 perf_bp_event(struct perf_event *event, void *data)                     { }
1103
1104 static inline int perf_register_guest_info_callbacks
1105 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1106 static inline int perf_unregister_guest_info_callbacks
1107 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1108
1109 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1110 static inline void perf_event_exec(void)                                { }
1111 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1112 static inline void perf_event_fork(struct task_struct *tsk)             { }
1113 static inline void perf_event_init(void)                                { }
1114 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1115 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1116 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1117 static inline void perf_event_enable(struct perf_event *event)          { }
1118 static inline void perf_event_disable(struct perf_event *event)         { }
1119 static inline int __perf_event_disable(void *info)                      { return -1; }
1120 static inline void perf_event_task_tick(void)                           { }
1121 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1122 #endif
1123
1124 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
1125 extern bool perf_event_can_stop_tick(void);
1126 #else
1127 static inline bool perf_event_can_stop_tick(void)                       { return true; }
1128 #endif
1129
1130 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1131 extern void perf_restore_debug_store(void);
1132 #else
1133 static inline void perf_restore_debug_store(void)                       { }
1134 #endif
1135
1136 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1137
1138 /*
1139  * This has to have a higher priority than migration_notifier in sched/core.c.
1140  */
1141 #define perf_cpu_notifier(fn)                                           \
1142 do {                                                                    \
1143         static struct notifier_block fn##_nb =                          \
1144                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1145         unsigned long cpu = smp_processor_id();                         \
1146         unsigned long flags;                                            \
1147                                                                         \
1148         cpu_notifier_register_begin();                                  \
1149         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1150                 (void *)(unsigned long)cpu);                            \
1151         local_irq_save(flags);                                          \
1152         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1153                 (void *)(unsigned long)cpu);                            \
1154         local_irq_restore(flags);                                       \
1155         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1156                 (void *)(unsigned long)cpu);                            \
1157         __register_cpu_notifier(&fn##_nb);                              \
1158         cpu_notifier_register_done();                                   \
1159 } while (0)
1160
1161 /*
1162  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1163  * callback for already online CPUs.
1164  */
1165 #define __perf_cpu_notifier(fn)                                         \
1166 do {                                                                    \
1167         static struct notifier_block fn##_nb =                          \
1168                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1169                                                                         \
1170         __register_cpu_notifier(&fn##_nb);                              \
1171 } while (0)
1172
1173 struct perf_pmu_events_attr {
1174         struct device_attribute attr;
1175         u64 id;
1176         const char *event_str;
1177 };
1178
1179 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1180                               char *page);
1181
1182 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1183 static struct perf_pmu_events_attr _var = {                             \
1184         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1185         .id   =  _id,                                                   \
1186 };
1187
1188 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1189 static struct perf_pmu_events_attr _var = {                                 \
1190         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1191         .id             = 0,                                                \
1192         .event_str      = _str,                                             \
1193 };
1194
1195 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1196 static ssize_t                                                          \
1197 _name##_show(struct device *dev,                                        \
1198                                struct device_attribute *attr,           \
1199                                char *page)                              \
1200 {                                                                       \
1201         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1202         return sprintf(page, _format "\n");                             \
1203 }                                                                       \
1204                                                                         \
1205 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1206
1207 #endif /* _LINUX_PERF_EVENT_H */