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[0]; /* /proc/sys/kernel/perf_event_max_stack */
  62 };
  63
  64 struct perf_callchain_entry_ctx {
  65         struct perf_callchain_entry *entry;
  66         u32                         max_stack;
  67         u32                         nr;
  68         short                       contexts;
  69         bool                        contexts_maxed;
  70 };
  71
  72 struct perf_raw_record {
  73         u32                             size;
  74         void                            *data;
  75 };
  76
  77 /*
  78  * branch stack layout:
  79  *  nr: number of taken branches stored in entries[]
  80  *
  81  * Note that nr can vary from sample to sample
  82  * branches (to, from) are stored from most recent
  83  * to least recent, i.e., entries[0] contains the most
  84  * recent branch.
  85  */
  86 struct perf_branch_stack {
  87         __u64                           nr;
  88         struct perf_branch_entry        entries[0];
  89 };
  90
  91 struct task_struct;
  92
  93 /*
  94  * extra PMU register associated with an event
  95  */
  96 struct hw_perf_event_extra {
  97         u64             config; /* register value */
  98         unsigned int    reg;    /* register address or index */
  99         int             alloc;  /* extra register already allocated */
 100         int             idx;    /* index in shared_regs->regs[] */
 101 };
 102
 103 /**
 104  * struct hw_perf_event - performance event hardware details:
 105  */
 106 struct hw_perf_event {
 107 #ifdef CONFIG_PERF_EVENTS
 108         union {
 109                 struct { /* hardware */
 110                         u64             config;
 111                         u64             last_tag;
 112                         unsigned long   config_base;
 113                         unsigned long   event_base;
 114                         int             event_base_rdpmc;
 115                         int             idx;
 116                         int             last_cpu;
 117                         int             flags;
 118
 119                         struct hw_perf_event_extra extra_reg;
 120                         struct hw_perf_event_extra branch_reg;
 121                 };
 122                 struct { /* software */
 123                         struct hrtimer  hrtimer;
 124                 };
 125                 struct { /* tracepoint */
 126                         /* for tp_event->class */
 127                         struct list_head        tp_list;
 128                 };
 129                 struct { /* intel_cqm */
 130                         int                     cqm_state;
 131                         u32                     cqm_rmid;
 132                         int                     is_group_event;
 133                         struct list_head        cqm_events_entry;
 134                         struct list_head        cqm_groups_entry;
 135                         struct list_head        cqm_group_entry;
 136                 };
 137                 struct { /* itrace */
 138                         int                     itrace_started;
 139                 };
 140                 struct { /* amd_power */
 141                         u64     pwr_acc;
 142                         u64     ptsc;
 143                 };
 144 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 145                 struct { /* breakpoint */
 146                         /*
 147                          * Crufty hack to avoid the chicken and egg
 148                          * problem hw_breakpoint has with context
 149                          * creation and event initalization.
 150                          */
 151                         struct arch_hw_breakpoint       info;
 152                         struct list_head                bp_list;
 153                 };
 154 #endif
 155         };
 156         /*
 157          * If the event is a per task event, this will point to the task in
 158          * question. See the comment in perf_event_alloc().
 159          */
 160         struct task_struct              *target;
 161
 162         /*
 163          * PMU would store hardware filter configuration
 164          * here.
 165          */
 166         void                            *addr_filters;
 167
 168         /* Last sync'ed generation of filters */
 169         unsigned long                   addr_filters_gen;
 170
 171 /*
 172  * hw_perf_event::state flags; used to track the PERF_EF_* state.
 173  */
 174 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 175 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 176 #define PERF_HES_ARCH           0x04
 177
 178         int                             state;
 179
 180         /*
 181          * The last observed hardware counter value, updated with a
 182          * local64_cmpxchg() such that pmu::read() can be called nested.
 183          */
 184         local64_t                       prev_count;
 185
 186         /*
 187          * The period to start the next sample with.
 188          */
 189         u64                             sample_period;
 190
 191         /*
 192          * The period we started this sample with.
 193          */
 194         u64                             last_period;
 195
 196         /*
 197          * However much is left of the current period; note that this is
 198          * a full 64bit value and allows for generation of periods longer
 199          * than hardware might allow.
 200          */
 201         local64_t                       period_left;
 202
 203         /*
 204          * State for throttling the event, see __perf_event_overflow() and
 205          * perf_adjust_freq_unthr_context().
 206          */
 207         u64                             interrupts_seq;
 208         u64                             interrupts;
 209
 210         /*
 211          * State for freq target events, see __perf_event_overflow() and
 212          * perf_adjust_freq_unthr_context().
 213          */
 214         u64                             freq_time_stamp;
 215         u64                             freq_count_stamp;
 216 #endif
 217 };
 218
 219 struct perf_event;
 220
 221 /*
 222  * Common implementation detail of pmu::{start,commit,cancel}_txn
 223  */
 224 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
 225 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
 226
 227 /**
 228  * pmu::capabilities flags
 229  */
 230 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 231 #define PERF_PMU_CAP_NO_NMI                     0x02
 232 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
 233 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
 234 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
 235 #define PERF_PMU_CAP_ITRACE                     0x20
 236 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS         0x40
 237
 238 /**
 239  * struct pmu - generic performance monitoring unit
 240  */
 241 struct pmu {
 242         struct list_head                entry;
 243
 244         struct module                   *module;
 245         struct device                   *dev;
 246         const struct attribute_group    **attr_groups;
 247         const char                      *name;
 248         int                             type;
 249
 250         /*
 251          * various common per-pmu feature flags
 252          */
 253         int                             capabilities;
 254
 255         int * __percpu                  pmu_disable_count;
 256         struct perf_cpu_context * __percpu pmu_cpu_context;
 257         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 258         int                             task_ctx_nr;
 259         int                             hrtimer_interval_ms;
 260
 261         /* number of address filters this PMU can do */
 262         unsigned int                    nr_addr_filters;
 263
 264         /*
 265          * Fully disable/enable this PMU, can be used to protect from the PMI
 266          * as well as for lazy/batch writing of the MSRs.
 267          */
 268         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 269         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 270
 271         /*
 272          * Try and initialize the event for this PMU.
 273          *
 274          * Returns:
 275          *  -ENOENT     -- @event is not for this PMU
 276          *
 277          *  -ENODEV     -- @event is for this PMU but PMU not present
 278          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
 279          *  -EINVAL     -- @event is for this PMU but @event is not valid
 280          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
 281          *  -EACCESS    -- @event is for this PMU, @event is valid, but no privilidges
 282          *
 283          *  0           -- @event is for this PMU and valid
 284          *
 285          * Other error return values are allowed.
 286          */
 287         int (*event_init)               (struct perf_event *event);
 288
 289         /*
 290          * Notification that the event was mapped or unmapped.  Called
 291          * in the context of the mapping task.
 292          */
 293         void (*event_mapped)            (struct perf_event *event); /*optional*/
 294         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 295
 296         /*
 297          * Flags for ->add()/->del()/ ->start()/->stop(). There are
 298          * matching hw_perf_event::state flags.
 299          */
 300 #define PERF_EF_START   0x01            /* start the counter when adding    */
 301 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 302 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 303
 304         /*
 305          * Adds/Removes a counter to/from the PMU, can be done inside a
 306          * transaction, see the ->*_txn() methods.
 307          *
 308          * The add/del callbacks will reserve all hardware resources required
 309          * to service the event, this includes any counter constraint
 310          * scheduling etc.
 311          *
 312          * Called with IRQs disabled and the PMU disabled on the CPU the event
 313          * is on.
 314          *
 315          * ->add() called without PERF_EF_START should result in the same state
 316          *  as ->add() followed by ->stop().
 317          *
 318          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
 319          *  ->stop() that must deal with already being stopped without
 320          *  PERF_EF_UPDATE.
 321          */
 322         int  (*add)                     (struct perf_event *event, int flags);
 323         void (*del)                     (struct perf_event *event, int flags);
 324
 325         /*
 326          * Starts/Stops a counter present on the PMU.
 327          *
 328          * The PMI handler should stop the counter when perf_event_overflow()
 329          * returns !0. ->start() will be used to continue.
 330          *
 331          * Also used to change the sample period.
 332          *
 333          * Called with IRQs disabled and the PMU disabled on the CPU the event
 334          * is on -- will be called from NMI context with the PMU generates
 335          * NMIs.
 336          *
 337          * ->stop() with PERF_EF_UPDATE will read the counter and update
 338          *  period/count values like ->read() would.
 339          *
 340          * ->start() with PERF_EF_RELOAD will reprogram the the counter
 341          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
 342          */
 343         void (*start)                   (struct perf_event *event, int flags);
 344         void (*stop)                    (struct perf_event *event, int flags);
 345
 346         /*
 347          * Updates the counter value of the event.
 348          *
 349          * For sampling capable PMUs this will also update the software period
 350          * hw_perf_event::period_left field.
 351          */
 352         void (*read)                    (struct perf_event *event);
 353
 354         /*
 355          * Group events scheduling is treated as a transaction, add
 356          * group events as a whole and perform one schedulability test.
 357          * If the test fails, roll back the whole group
 358          *
 359          * Start the transaction, after this ->add() doesn't need to
 360          * do schedulability tests.
 361          *
 362          * Optional.
 363          */
 364         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
 365         /*
 366          * If ->start_txn() disabled the ->add() schedulability test
 367          * then ->commit_txn() is required to perform one. On success
 368          * the transaction is closed. On error the transaction is kept
 369          * open until ->cancel_txn() is called.
 370          *
 371          * Optional.
 372          */
 373         int  (*commit_txn)              (struct pmu *pmu);
 374         /*
 375          * Will cancel the transaction, assumes ->del() is called
 376          * for each successful ->add() during the transaction.
 377          *
 378          * Optional.
 379          */
 380         void (*cancel_txn)              (struct pmu *pmu);
 381
 382         /*
 383          * Will return the value for perf_event_mmap_page::index for this event,
 384          * if no implementation is provided it will default to: event->hw.idx + 1.
 385          */
 386         int (*event_idx)                (struct perf_event *event); /*optional */
 387
 388         /*
 389          * context-switches callback
 390          */
 391         void (*sched_task)              (struct perf_event_context *ctx,
 392                                         bool sched_in);
 393         /*
 394          * PMU specific data size
 395          */
 396         size_t                          task_ctx_size;
 397
 398
 399         /*
 400          * Return the count value for a counter.
 401          */
 402         u64 (*count)                    (struct perf_event *event); /*optional*/
 403
 404         /*
 405          * Set up pmu-private data structures for an AUX area
 406          */
 407         void *(*setup_aux)              (int cpu, void **pages,
 408                                          int nr_pages, bool overwrite);
 409                                         /* optional */
 410
 411         /*
 412          * Free pmu-private AUX data structures
 413          */
 414         void (*free_aux)                (void *aux); /* optional */
 415
 416         /*
 417          * Validate address range filters: make sure the HW supports the
 418          * requested configuration and number of filters; return 0 if the
 419          * supplied filters are valid, -errno otherwise.
 420          *
 421          * Runs in the context of the ioctl()ing process and is not serialized
 422          * with the rest of the PMU callbacks.
 423          */
 424         int (*addr_filters_validate)    (struct list_head *filters);
 425                                         /* optional */
 426
 427         /*
 428          * Synchronize address range filter configuration:
 429          * translate hw-agnostic filters into hardware configuration in
 430          * event::hw::addr_filters.
 431          *
 432          * Runs as a part of filter sync sequence that is done in ->start()
 433          * callback by calling perf_event_addr_filters_sync().
 434          *
 435          * May (and should) traverse event::addr_filters::list, for which its
 436          * caller provides necessary serialization.
 437          */
 438         void (*addr_filters_sync)       (struct perf_event *event);
 439                                         /* optional */
 440
 441         /*
 442          * Filter events for PMU-specific reasons.
 443          */
 444         int (*filter_match)             (struct perf_event *event); /* optional */
 445 };
 446
 447 /**
 448  * struct perf_addr_filter - address range filter definition
 449  * @entry:      event's filter list linkage
 450  * @inode:      object file's inode for file-based filters
 451  * @offset:     filter range offset
 452  * @size:       filter range size
 453  * @range:      1: range, 0: address
 454  * @filter:     1: filter/start, 0: stop
 455  *
 456  * This is a hardware-agnostic filter configuration as specified by the user.
 457  */
 458 struct perf_addr_filter {
 459         struct list_head        entry;
 460         struct inode            *inode;
 461         unsigned long           offset;
 462         unsigned long           size;
 463         unsigned int            range   : 1,
 464                                 filter  : 1;
 465 };
 466
 467 /**
 468  * struct perf_addr_filters_head - container for address range filters
 469  * @list:       list of filters for this event
 470  * @lock:       spinlock that serializes accesses to the @list and event's
 471  *              (and its children's) filter generations.
 472  *
 473  * A child event will use parent's @list (and therefore @lock), so they are
 474  * bundled together; see perf_event_addr_filters().
 475  */
 476 struct perf_addr_filters_head {
 477         struct list_head        list;
 478         raw_spinlock_t          lock;
 479 };
 480
 481 /**
 482  * enum perf_event_active_state - the states of a event
 483  */
 484 enum perf_event_active_state {
 485         PERF_EVENT_STATE_DEAD           = -4,
 486         PERF_EVENT_STATE_EXIT           = -3,
 487         PERF_EVENT_STATE_ERROR          = -2,
 488         PERF_EVENT_STATE_OFF            = -1,
 489         PERF_EVENT_STATE_INACTIVE       =  0,
 490         PERF_EVENT_STATE_ACTIVE         =  1,
 491 };
 492
 493 struct file;
 494 struct perf_sample_data;
 495
 496 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 497                                         struct perf_sample_data *,
 498                                         struct pt_regs *regs);
 499
 500 enum perf_group_flag {
 501         PERF_GROUP_SOFTWARE             = 0x1,
 502 };
 503
 504 #define SWEVENT_HLIST_BITS              8
 505 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 506
 507 struct swevent_hlist {
 508         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 509         struct rcu_head                 rcu_head;
 510 };
 511
 512 #define PERF_ATTACH_CONTEXT     0x01
 513 #define PERF_ATTACH_GROUP       0x02
 514 #define PERF_ATTACH_TASK        0x04
 515 #define PERF_ATTACH_TASK_DATA   0x08
 516
 517 struct perf_cgroup;
 518 struct ring_buffer;
 519
 520 /**
 521  * struct perf_event - performance event kernel representation:
 522  */
 523 struct perf_event {
 524 #ifdef CONFIG_PERF_EVENTS
 525         /*
 526          * entry onto perf_event_context::event_list;
 527          *   modifications require ctx->lock
 528          *   RCU safe iterations.
 529          */
 530         struct list_head                event_entry;
 531
 532         /*
 533          * XXX: group_entry and sibling_list should be mutually exclusive;
 534          * either you're a sibling on a group, or you're the group leader.
 535          * Rework the code to always use the same list element.
 536          *
 537          * Locked for modification by both ctx->mutex and ctx->lock; holding
 538          * either sufficies for read.
 539          */
 540         struct list_head                group_entry;
 541         struct list_head                sibling_list;
 542
 543         /*
 544          * We need storage to track the entries in perf_pmu_migrate_context; we
 545          * cannot use the event_entry because of RCU and we want to keep the
 546          * group in tact which avoids us using the other two entries.
 547          */
 548         struct list_head                migrate_entry;
 549
 550         struct hlist_node               hlist_entry;
 551         struct list_head                active_entry;
 552         int                             nr_siblings;
 553         int                             group_flags;
 554         struct perf_event               *group_leader;
 555         struct pmu                      *pmu;
 556         void                            *pmu_private;
 557
 558         enum perf_event_active_state    state;
 559         unsigned int                    attach_state;
 560         local64_t                       count;
 561         atomic64_t                      child_count;
 562
 563         /*
 564          * These are the total time in nanoseconds that the event
 565          * has been enabled (i.e. eligible to run, and the task has
 566          * been scheduled in, if this is a per-task event)
 567          * and running (scheduled onto the CPU), respectively.
 568          *
 569          * They are computed from tstamp_enabled, tstamp_running and
 570          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 571          */
 572         u64                             total_time_enabled;
 573         u64                             total_time_running;
 574
 575         /*
 576          * These are timestamps used for computing total_time_enabled
 577          * and total_time_running when the event is in INACTIVE or
 578          * ACTIVE state, measured in nanoseconds from an arbitrary point
 579          * in time.
 580          * tstamp_enabled: the notional time when the event was enabled
 581          * tstamp_running: the notional time when the event was scheduled on
 582          * tstamp_stopped: in INACTIVE state, the notional time when the
 583          *      event was scheduled off.
 584          */
 585         u64                             tstamp_enabled;
 586         u64                             tstamp_running;
 587         u64                             tstamp_stopped;
 588
 589         /*
 590          * timestamp shadows the actual context timing but it can
 591          * be safely used in NMI interrupt context. It reflects the
 592          * context time as it was when the event was last scheduled in.
 593          *
 594          * ctx_time already accounts for ctx->timestamp. Therefore to
 595          * compute ctx_time for a sample, simply add perf_clock().
 596          */
 597         u64                             shadow_ctx_time;
 598
 599         struct perf_event_attr          attr;
 600         u16                             header_size;
 601         u16                             id_header_size;
 602         u16                             read_size;
 603         struct hw_perf_event            hw;
 604
 605         struct perf_event_context       *ctx;
 606         atomic_long_t                   refcount;
 607
 608         /*
 609          * These accumulate total time (in nanoseconds) that children
 610          * events have been enabled and running, respectively.
 611          */
 612         atomic64_t                      child_total_time_enabled;
 613         atomic64_t                      child_total_time_running;
 614
 615         /*
 616          * Protect attach/detach and child_list:
 617          */
 618         struct mutex                    child_mutex;
 619         struct list_head                child_list;
 620         struct perf_event               *parent;
 621
 622         int                             oncpu;
 623         int                             cpu;
 624
 625         struct list_head                owner_entry;
 626         struct task_struct              *owner;
 627
 628         /* mmap bits */
 629         struct mutex                    mmap_mutex;
 630         atomic_t                        mmap_count;
 631
 632         struct ring_buffer              *rb;
 633         struct list_head                rb_entry;
 634         unsigned long                   rcu_batches;
 635         int                             rcu_pending;
 636
 637         /* poll related */
 638         wait_queue_head_t               waitq;
 639         struct fasync_struct            *fasync;
 640
 641         /* delayed work for NMIs and such */
 642         int                             pending_wakeup;
 643         int                             pending_kill;
 644         int                             pending_disable;
 645         struct irq_work                 pending;
 646
 647         atomic_t                        event_limit;
 648
 649         /* address range filters */
 650         struct perf_addr_filters_head   addr_filters;
 651         /* vma address array for file-based filders */
 652         unsigned long                   *addr_filters_offs;
 653         unsigned long                   addr_filters_gen;
 654
 655         void (*destroy)(struct perf_event *);
 656         struct rcu_head                 rcu_head;
 657
 658         struct pid_namespace            *ns;
 659         u64                             id;
 660
 661         u64                             (*clock)(void);
 662         perf_overflow_handler_t         overflow_handler;
 663         void                            *overflow_handler_context;
 664
 665 #ifdef CONFIG_EVENT_TRACING
 666         struct trace_event_call         *tp_event;
 667         struct event_filter             *filter;
 668 #ifdef CONFIG_FUNCTION_TRACER
 669         struct ftrace_ops               ftrace_ops;
 670 #endif
 671 #endif
 672
 673 #ifdef CONFIG_CGROUP_PERF
 674         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 675         int                             cgrp_defer_enabled;
 676 #endif
 677
 678 #endif /* CONFIG_PERF_EVENTS */
 679 };
 680
 681 /**
 682  * struct perf_event_context - event context structure
 683  *
 684  * Used as a container for task events and CPU events as well:
 685  */
 686 struct perf_event_context {
 687         struct pmu                      *pmu;
 688         /*
 689          * Protect the states of the events in the list,
 690          * nr_active, and the list:
 691          */
 692         raw_spinlock_t                  lock;
 693         /*
 694          * Protect the list of events.  Locking either mutex or lock
 695          * is sufficient to ensure the list doesn't change; to change
 696          * the list you need to lock both the mutex and the spinlock.
 697          */
 698         struct mutex                    mutex;
 699
 700         struct list_head                active_ctx_list;
 701         struct list_head                pinned_groups;
 702         struct list_head                flexible_groups;
 703         struct list_head                event_list;
 704         int                             nr_events;
 705         int                             nr_active;
 706         int                             is_active;
 707         int                             nr_stat;
 708         int                             nr_freq;
 709         int                             rotate_disable;
 710         atomic_t                        refcount;
 711         struct task_struct              *task;
 712
 713         /*
 714          * Context clock, runs when context enabled.
 715          */
 716         u64                             time;
 717         u64                             timestamp;
 718
 719         /*
 720          * These fields let us detect when two contexts have both
 721          * been cloned (inherited) from a common ancestor.
 722          */
 723         struct perf_event_context       *parent_ctx;
 724         u64                             parent_gen;
 725         u64                             generation;
 726         int                             pin_count;
 727         int                             nr_cgroups;      /* cgroup evts */
 728         void                            *task_ctx_data; /* pmu specific data */
 729         struct rcu_head                 rcu_head;
 730 };
 731
 732 /*
 733  * Number of contexts where an event can trigger:
 734  *      task, softirq, hardirq, nmi.
 735  */
 736 #define PERF_NR_CONTEXTS        4
 737
 738 /**
 739  * struct perf_event_cpu_context - per cpu event context structure
 740  */
 741 struct perf_cpu_context {
 742         struct perf_event_context       ctx;
 743         struct perf_event_context       *task_ctx;
 744         int                             active_oncpu;
 745         int                             exclusive;
 746
 747         raw_spinlock_t                  hrtimer_lock;
 748         struct hrtimer                  hrtimer;
 749         ktime_t                         hrtimer_interval;
 750         unsigned int                    hrtimer_active;
 751
 752         struct pmu                      *unique_pmu;
 753         struct perf_cgroup              *cgrp;
 754 };
 755
 756 struct perf_output_handle {
 757         struct perf_event               *event;
 758         struct ring_buffer              *rb;
 759         unsigned long                   wakeup;
 760         unsigned long                   size;
 761         union {
 762                 void                    *addr;
 763                 unsigned long           head;
 764         };
 765         int                             page;
 766 };
 767
 768 #ifdef CONFIG_CGROUP_PERF
 769
 770 /*
 771  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 772  * This is a per-cpu dynamically allocated data structure.
 773  */
 774 struct perf_cgroup_info {
 775         u64                             time;
 776         u64                             timestamp;
 777 };
 778
 779 struct perf_cgroup {
 780         struct cgroup_subsys_state      css;
 781         struct perf_cgroup_info __percpu *info;
 782 };
 783
 784 /*
 785  * Must ensure cgroup is pinned (css_get) before calling
 786  * this function. In other words, we cannot call this function
 787  * if there is no cgroup event for the current CPU context.
 788  */
 789 static inline struct perf_cgroup *
 790 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 791 {
 792         return container_of(task_css_check(task, perf_event_cgrp_id,
 793                                            ctx ? lockdep_is_held(&ctx->lock)
 794                                                : true),
 795                             struct perf_cgroup, css);
 796 }
 797 #endif /* CONFIG_CGROUP_PERF */
 798
 799 #ifdef CONFIG_PERF_EVENTS
 800
 801 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 802                                    struct perf_event *event);
 803 extern void perf_aux_output_end(struct perf_output_handle *handle,
 804                                 unsigned long size, bool truncated);
 805 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 806                                 unsigned long size);
 807 extern void *perf_get_aux(struct perf_output_handle *handle);
 808
 809 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 810 extern void perf_pmu_unregister(struct pmu *pmu);
 811
 812 extern int perf_num_counters(void);
 813 extern const char *perf_pmu_name(void);
 814 extern void __perf_event_task_sched_in(struct task_struct *prev,
 815                                        struct task_struct *task);
 816 extern void __perf_event_task_sched_out(struct task_struct *prev,
 817                                         struct task_struct *next);
 818 extern int perf_event_init_task(struct task_struct *child);
 819 extern void perf_event_exit_task(struct task_struct *child);
 820 extern void perf_event_free_task(struct task_struct *task);
 821 extern void perf_event_delayed_put(struct task_struct *task);
 822 extern struct file *perf_event_get(unsigned int fd);
 823 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 824 extern void perf_event_print_debug(void);
 825 extern void perf_pmu_disable(struct pmu *pmu);
 826 extern void perf_pmu_enable(struct pmu *pmu);
 827 extern void perf_sched_cb_dec(struct pmu *pmu);
 828 extern void perf_sched_cb_inc(struct pmu *pmu);
 829 extern int perf_event_task_disable(void);
 830 extern int perf_event_task_enable(void);
 831 extern int perf_event_refresh(struct perf_event *event, int refresh);
 832 extern void perf_event_update_userpage(struct perf_event *event);
 833 extern int perf_event_release_kernel(struct perf_event *event);
 834 extern struct perf_event *
 835 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 836                                 int cpu,
 837                                 struct task_struct *task,
 838                                 perf_overflow_handler_t callback,
 839                                 void *context);
 840 extern void perf_pmu_migrate_context(struct pmu *pmu,
 841                                 int src_cpu, int dst_cpu);
 842 extern u64 perf_event_read_local(struct perf_event *event);
 843 extern u64 perf_event_read_value(struct perf_event *event,
 844                                  u64 *enabled, u64 *running);
 845
 846
 847 struct perf_sample_data {
 848         /*
 849          * Fields set by perf_sample_data_init(), group so as to
 850          * minimize the cachelines touched.
 851          */
 852         u64                             addr;
 853         struct perf_raw_record          *raw;
 854         struct perf_branch_stack        *br_stack;
 855         u64                             period;
 856         u64                             weight;
 857         u64                             txn;
 858         union  perf_mem_data_src        data_src;
 859
 860         /*
 861          * The other fields, optionally {set,used} by
 862          * perf_{prepare,output}_sample().
 863          */
 864         u64                             type;
 865         u64                             ip;
 866         struct {
 867                 u32     pid;
 868                 u32     tid;
 869         }                               tid_entry;
 870         u64                             time;
 871         u64                             id;
 872         u64                             stream_id;
 873         struct {
 874                 u32     cpu;
 875                 u32     reserved;
 876         }                               cpu_entry;
 877         struct perf_callchain_entry     *callchain;
 878
 879         /*
 880          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 881          * on arch details.
 882          */
 883         struct perf_regs                regs_user;
 884         struct pt_regs                  regs_user_copy;
 885
 886         struct perf_regs                regs_intr;
 887         u64                             stack_user_size;
 888 } ____cacheline_aligned;
 889
 890 /* default value for data source */
 891 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 892                     PERF_MEM_S(LVL, NA)   |\
 893                     PERF_MEM_S(SNOOP, NA) |\
 894                     PERF_MEM_S(LOCK, NA)  |\
 895                     PERF_MEM_S(TLB, NA))
 896
 897 static inline void perf_sample_data_init(struct perf_sample_data *data,
 898                                          u64 addr, u64 period)
 899 {
 900         /* remaining struct members initialized in perf_prepare_sample() */
 901         data->addr = addr;
 902         data->raw  = NULL;
 903         data->br_stack = NULL;
 904         data->period = period;
 905         data->weight = 0;
 906         data->data_src.val = PERF_MEM_NA;
 907         data->txn = 0;
 908 }
 909
 910 extern void perf_output_sample(struct perf_output_handle *handle,
 911                                struct perf_event_header *header,
 912                                struct perf_sample_data *data,
 913                                struct perf_event *event);
 914 extern void perf_prepare_sample(struct perf_event_header *header,
 915                                 struct perf_sample_data *data,
 916                                 struct perf_event *event,
 917                                 struct pt_regs *regs);
 918
 919 extern int perf_event_overflow(struct perf_event *event,
 920                                  struct perf_sample_data *data,
 921                                  struct pt_regs *regs);
 922
 923 extern void perf_event_output_forward(struct perf_event *event,
 924                                      struct perf_sample_data *data,
 925                                      struct pt_regs *regs);
 926 extern void perf_event_output_backward(struct perf_event *event,
 927                                        struct perf_sample_data *data,
 928                                        struct pt_regs *regs);
 929 extern void perf_event_output(struct perf_event *event,
 930                               struct perf_sample_data *data,
 931                               struct pt_regs *regs);
 932
 933 static inline bool
 934 is_default_overflow_handler(struct perf_event *event)
 935 {
 936         if (likely(event->overflow_handler == perf_event_output_forward))
 937                 return true;
 938         if (unlikely(event->overflow_handler == perf_event_output_backward))
 939                 return true;
 940         return false;
 941 }
 942
 943 extern void
 944 perf_event_header__init_id(struct perf_event_header *header,
 945                            struct perf_sample_data *data,
 946                            struct perf_event *event);
 947 extern void
 948 perf_event__output_id_sample(struct perf_event *event,
 949                              struct perf_output_handle *handle,
 950                              struct perf_sample_data *sample);
 951
 952 extern void
 953 perf_log_lost_samples(struct perf_event *event, u64 lost);
 954
 955 static inline bool is_sampling_event(struct perf_event *event)
 956 {
 957         return event->attr.sample_period != 0;
 958 }
 959
 960 /*
 961  * Return 1 for a software event, 0 for a hardware event
 962  */
 963 static inline int is_software_event(struct perf_event *event)
 964 {
 965         return event->pmu->task_ctx_nr == perf_sw_context;
 966 }
 967
 968 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 969
 970 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 971 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 972
 973 #ifndef perf_arch_fetch_caller_regs
 974 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 975 #endif
 976
 977 /*
 978  * Take a snapshot of the regs. Skip ip and frame pointer to
 979  * the nth caller. We only need a few of the regs:
 980  * - ip for PERF_SAMPLE_IP
 981  * - cs for user_mode() tests
 982  * - bp for callchains
 983  * - eflags, for future purposes, just in case
 984  */
 985 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 986 {
 987         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 988 }
 989
 990 static __always_inline void
 991 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 992 {
 993         if (static_key_false(&perf_swevent_enabled[event_id]))
 994                 __perf_sw_event(event_id, nr, regs, addr);
 995 }
 996
 997 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 998
 999 /*
1000  * 'Special' version for the scheduler, it hard assumes no recursion,
1001  * which is guaranteed by us not actually scheduling inside other swevents
1002  * because those disable preemption.
1003  */
1004 static __always_inline void
1005 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1006 {
1007         if (static_key_false(&perf_swevent_enabled[event_id])) {
1008                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1009
1010                 perf_fetch_caller_regs(regs);
1011                 ___perf_sw_event(event_id, nr, regs, addr);
1012         }
1013 }
1014
1015 extern struct static_key_false perf_sched_events;
1016
1017 static __always_inline bool
1018 perf_sw_migrate_enabled(void)
1019 {
1020         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1021                 return true;
1022         return false;
1023 }
1024
1025 static inline void perf_event_task_migrate(struct task_struct *task)
1026 {
1027         if (perf_sw_migrate_enabled())
1028                 task->sched_migrated = 1;
1029 }
1030
1031 static inline void perf_event_task_sched_in(struct task_struct *prev,
1032                                             struct task_struct *task)
1033 {
1034         if (static_branch_unlikely(&perf_sched_events))
1035                 __perf_event_task_sched_in(prev, task);
1036
1037         if (perf_sw_migrate_enabled() && task->sched_migrated) {
1038                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1039
1040                 perf_fetch_caller_regs(regs);
1041                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1042                 task->sched_migrated = 0;
1043         }
1044 }
1045
1046 static inline void perf_event_task_sched_out(struct task_struct *prev,
1047                                              struct task_struct *next)
1048 {
1049         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1050
1051         if (static_branch_unlikely(&perf_sched_events))
1052                 __perf_event_task_sched_out(prev, next);
1053 }
1054
1055 static inline u64 __perf_event_count(struct perf_event *event)
1056 {
1057         return local64_read(&event->count) + atomic64_read(&event->child_count);
1058 }
1059
1060 extern void perf_event_mmap(struct vm_area_struct *vma);
1061 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1062 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1063 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1064
1065 extern void perf_event_exec(void);
1066 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1067 extern void perf_event_fork(struct task_struct *tsk);
1068
1069 /* Callchains */
1070 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1071
1072 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1073 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1074 extern struct perf_callchain_entry *
1075 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1076                    u32 max_stack, bool crosstask, bool add_mark);
1077 extern int get_callchain_buffers(void);
1078 extern void put_callchain_buffers(void);
1079
1080 extern int sysctl_perf_event_max_stack;
1081 extern int sysctl_perf_event_max_contexts_per_stack;
1082
1083 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1084 {
1085         if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1086                 struct perf_callchain_entry *entry = ctx->entry;
1087                 entry->ip[entry->nr++] = ip;
1088                 ++ctx->contexts;
1089                 return 0;
1090         } else {
1091                 ctx->contexts_maxed = true;
1092                 return -1; /* no more room, stop walking the stack */
1093         }
1094 }
1095
1096 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1097 {
1098         if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1099                 struct perf_callchain_entry *entry = ctx->entry;
1100                 entry->ip[entry->nr++] = ip;
1101                 ++ctx->nr;
1102                 return 0;
1103         } else {
1104                 return -1; /* no more room, stop walking the stack */
1105         }
1106 }
1107
1108 extern int sysctl_perf_event_paranoid;
1109 extern int sysctl_perf_event_mlock;
1110 extern int sysctl_perf_event_sample_rate;
1111 extern int sysctl_perf_cpu_time_max_percent;
1112
1113 extern void perf_sample_event_took(u64 sample_len_ns);
1114
1115 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1116                 void __user *buffer, size_t *lenp,
1117                 loff_t *ppos);
1118 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1119                 void __user *buffer, size_t *lenp,
1120                 loff_t *ppos);
1121
1122 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1123                                  void __user *buffer, size_t *lenp, loff_t *ppos);
1124
1125 static inline bool perf_paranoid_tracepoint_raw(void)
1126 {
1127         return sysctl_perf_event_paranoid > -1;
1128 }
1129
1130 static inline bool perf_paranoid_cpu(void)
1131 {
1132         return sysctl_perf_event_paranoid > 0;
1133 }
1134
1135 static inline bool perf_paranoid_kernel(void)
1136 {
1137         return sysctl_perf_event_paranoid > 1;
1138 }
1139
1140 extern void perf_event_init(void);
1141 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1142                           int entry_size, struct pt_regs *regs,
1143                           struct hlist_head *head, int rctx,
1144                           struct task_struct *task);
1145 extern void perf_bp_event(struct perf_event *event, void *data);
1146
1147 #ifndef perf_misc_flags
1148 # define perf_misc_flags(regs) \
1149                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1150 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1151 #endif
1152
1153 static inline bool has_branch_stack(struct perf_event *event)
1154 {
1155         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1156 }
1157
1158 static inline bool needs_branch_stack(struct perf_event *event)
1159 {
1160         return event->attr.branch_sample_type != 0;
1161 }
1162
1163 static inline bool has_aux(struct perf_event *event)
1164 {
1165         return event->pmu->setup_aux;
1166 }
1167
1168 static inline bool is_write_backward(struct perf_event *event)
1169 {
1170         return !!event->attr.write_backward;
1171 }
1172
1173 static inline bool has_addr_filter(struct perf_event *event)
1174 {
1175         return event->pmu->nr_addr_filters;
1176 }
1177
1178 /*
1179  * An inherited event uses parent's filters
1180  */
1181 static inline struct perf_addr_filters_head *
1182 perf_event_addr_filters(struct perf_event *event)
1183 {
1184         struct perf_addr_filters_head *ifh = &event->addr_filters;
1185
1186         if (event->parent)
1187                 ifh = &event->parent->addr_filters;
1188
1189         return ifh;
1190 }
1191
1192 extern void perf_event_addr_filters_sync(struct perf_event *event);
1193
1194 extern int perf_output_begin(struct perf_output_handle *handle,
1195                              struct perf_event *event, unsigned int size);
1196 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1197                                     struct perf_event *event,
1198                                     unsigned int size);
1199 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1200                                       struct perf_event *event,
1201                                       unsigned int size);
1202
1203 extern void perf_output_end(struct perf_output_handle *handle);
1204 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1205                              const void *buf, unsigned int len);
1206 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1207                                      unsigned int len);
1208 extern int perf_swevent_get_recursion_context(void);
1209 extern void perf_swevent_put_recursion_context(int rctx);
1210 extern u64 perf_swevent_set_period(struct perf_event *event);
1211 extern void perf_event_enable(struct perf_event *event);
1212 extern void perf_event_disable(struct perf_event *event);
1213 extern void perf_event_disable_local(struct perf_event *event);
1214 extern void perf_event_task_tick(void);
1215 #else /* !CONFIG_PERF_EVENTS: */
1216 static inline void *
1217 perf_aux_output_begin(struct perf_output_handle *handle,
1218                       struct perf_event *event)                         { return NULL; }
1219 static inline void
1220 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1221                     bool truncated)                                     { }
1222 static inline int
1223 perf_aux_output_skip(struct perf_output_handle *handle,
1224                      unsigned long size)                                { return -EINVAL; }
1225 static inline void *
1226 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1227 static inline void
1228 perf_event_task_migrate(struct task_struct *task)                       { }
1229 static inline void
1230 perf_event_task_sched_in(struct task_struct *prev,
1231                          struct task_struct *task)                      { }
1232 static inline void
1233 perf_event_task_sched_out(struct task_struct *prev,
1234                           struct task_struct *next)                     { }
1235 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1236 static inline void perf_event_exit_task(struct task_struct *child)      { }
1237 static inline void perf_event_free_task(struct task_struct *task)       { }
1238 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1239 static inline struct file *perf_event_get(unsigned int fd)      { return ERR_PTR(-EINVAL); }
1240 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1241 {
1242         return ERR_PTR(-EINVAL);
1243 }
1244 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1245 static inline void perf_event_print_debug(void)                         { }
1246 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1247 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1248 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1249 {
1250         return -EINVAL;
1251 }
1252
1253 static inline void
1254 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1255 static inline void
1256 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1257 static inline void
1258 perf_bp_event(struct perf_event *event, void *data)                     { }
1259
1260 static inline int perf_register_guest_info_callbacks
1261 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1262 static inline int perf_unregister_guest_info_callbacks
1263 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1264
1265 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1266 static inline void perf_event_exec(void)                                { }
1267 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1268 static inline void perf_event_fork(struct task_struct *tsk)             { }
1269 static inline void perf_event_init(void)                                { }
1270 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1271 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1272 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1273 static inline void perf_event_enable(struct perf_event *event)          { }
1274 static inline void perf_event_disable(struct perf_event *event)         { }
1275 static inline int __perf_event_disable(void *info)                      { return -1; }
1276 static inline void perf_event_task_tick(void)                           { }
1277 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1278 #endif
1279
1280 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1281 extern void perf_restore_debug_store(void);
1282 #else
1283 static inline void perf_restore_debug_store(void)                       { }
1284 #endif
1285
1286 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1287
1288 /*
1289  * This has to have a higher priority than migration_notifier in sched/core.c.
1290  */
1291 #define perf_cpu_notifier(fn)                                           \
1292 do {                                                                    \
1293         static struct notifier_block fn##_nb =                          \
1294                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1295         unsigned long cpu = smp_processor_id();                         \
1296         unsigned long flags;                                            \
1297                                                                         \
1298         cpu_notifier_register_begin();                                  \
1299         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1300                 (void *)(unsigned long)cpu);                            \
1301         local_irq_save(flags);                                          \
1302         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1303                 (void *)(unsigned long)cpu);                            \
1304         local_irq_restore(flags);                                       \
1305         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1306                 (void *)(unsigned long)cpu);                            \
1307         __register_cpu_notifier(&fn##_nb);                              \
1308         cpu_notifier_register_done();                                   \
1309 } while (0)
1310
1311 /*
1312  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1313  * callback for already online CPUs.
1314  */
1315 #define __perf_cpu_notifier(fn)                                         \
1316 do {                                                                    \
1317         static struct notifier_block fn##_nb =                          \
1318                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1319                                                                         \
1320         __register_cpu_notifier(&fn##_nb);                              \
1321 } while (0)
1322
1323 struct perf_pmu_events_attr {
1324         struct device_attribute attr;
1325         u64 id;
1326         const char *event_str;
1327 };
1328
1329 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1330                               char *page);
1331
1332 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1333 static struct perf_pmu_events_attr _var = {                             \
1334         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1335         .id   =  _id,                                                   \
1336 };
1337
1338 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1339 static struct perf_pmu_events_attr _var = {                                 \
1340         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1341         .id             = 0,                                                \
1342         .event_str      = _str,                                             \
1343 };
1344
1345 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1346 static ssize_t                                                          \
1347 _name##_show(struct device *dev,                                        \
1348                                struct device_attribute *attr,           \
1349                                char *page)                              \
1350 {                                                                       \
1351         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1352         return sprintf(page, _format "\n");                             \
1353 }                                                                       \
1354                                                                         \
1355 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1356
1357 #endif /* _LINUX_PERF_EVENT_H */