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 struct pmu_event_list {
 521         raw_spinlock_t          lock;
 522         struct list_head        list;
 523 };
 524
 525 /**
 526  * struct perf_event - performance event kernel representation:
 527  */
 528 struct perf_event {
 529 #ifdef CONFIG_PERF_EVENTS
 530         /*
 531          * entry onto perf_event_context::event_list;
 532          *   modifications require ctx->lock
 533          *   RCU safe iterations.
 534          */
 535         struct list_head                event_entry;
 536
 537         /*
 538          * XXX: group_entry and sibling_list should be mutually exclusive;
 539          * either you're a sibling on a group, or you're the group leader.
 540          * Rework the code to always use the same list element.
 541          *
 542          * Locked for modification by both ctx->mutex and ctx->lock; holding
 543          * either sufficies for read.
 544          */
 545         struct list_head                group_entry;
 546         struct list_head                sibling_list;
 547
 548         /*
 549          * We need storage to track the entries in perf_pmu_migrate_context; we
 550          * cannot use the event_entry because of RCU and we want to keep the
 551          * group in tact which avoids us using the other two entries.
 552          */
 553         struct list_head                migrate_entry;
 554
 555         struct hlist_node               hlist_entry;
 556         struct list_head                active_entry;
 557         int                             nr_siblings;
 558         int                             group_flags;
 559         struct perf_event               *group_leader;
 560         struct pmu                      *pmu;
 561         void                            *pmu_private;
 562
 563         enum perf_event_active_state    state;
 564         unsigned int                    attach_state;
 565         local64_t                       count;
 566         atomic64_t                      child_count;
 567
 568         /*
 569          * These are the total time in nanoseconds that the event
 570          * has been enabled (i.e. eligible to run, and the task has
 571          * been scheduled in, if this is a per-task event)
 572          * and running (scheduled onto the CPU), respectively.
 573          *
 574          * They are computed from tstamp_enabled, tstamp_running and
 575          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 576          */
 577         u64                             total_time_enabled;
 578         u64                             total_time_running;
 579
 580         /*
 581          * These are timestamps used for computing total_time_enabled
 582          * and total_time_running when the event is in INACTIVE or
 583          * ACTIVE state, measured in nanoseconds from an arbitrary point
 584          * in time.
 585          * tstamp_enabled: the notional time when the event was enabled
 586          * tstamp_running: the notional time when the event was scheduled on
 587          * tstamp_stopped: in INACTIVE state, the notional time when the
 588          *      event was scheduled off.
 589          */
 590         u64                             tstamp_enabled;
 591         u64                             tstamp_running;
 592         u64                             tstamp_stopped;
 593
 594         /*
 595          * timestamp shadows the actual context timing but it can
 596          * be safely used in NMI interrupt context. It reflects the
 597          * context time as it was when the event was last scheduled in.
 598          *
 599          * ctx_time already accounts for ctx->timestamp. Therefore to
 600          * compute ctx_time for a sample, simply add perf_clock().
 601          */
 602         u64                             shadow_ctx_time;
 603
 604         struct perf_event_attr          attr;
 605         u16                             header_size;
 606         u16                             id_header_size;
 607         u16                             read_size;
 608         struct hw_perf_event            hw;
 609
 610         struct perf_event_context       *ctx;
 611         atomic_long_t                   refcount;
 612
 613         /*
 614          * These accumulate total time (in nanoseconds) that children
 615          * events have been enabled and running, respectively.
 616          */
 617         atomic64_t                      child_total_time_enabled;
 618         atomic64_t                      child_total_time_running;
 619
 620         /*
 621          * Protect attach/detach and child_list:
 622          */
 623         struct mutex                    child_mutex;
 624         struct list_head                child_list;
 625         struct perf_event               *parent;
 626
 627         int                             oncpu;
 628         int                             cpu;
 629
 630         struct list_head                owner_entry;
 631         struct task_struct              *owner;
 632
 633         /* mmap bits */
 634         struct mutex                    mmap_mutex;
 635         atomic_t                        mmap_count;
 636
 637         struct ring_buffer              *rb;
 638         struct list_head                rb_entry;
 639         unsigned long                   rcu_batches;
 640         int                             rcu_pending;
 641
 642         /* poll related */
 643         wait_queue_head_t               waitq;
 644         struct fasync_struct            *fasync;
 645
 646         /* delayed work for NMIs and such */
 647         int                             pending_wakeup;
 648         int                             pending_kill;
 649         int                             pending_disable;
 650         struct irq_work                 pending;
 651
 652         atomic_t                        event_limit;
 653
 654         /* address range filters */
 655         struct perf_addr_filters_head   addr_filters;
 656         /* vma address array for file-based filders */
 657         unsigned long                   *addr_filters_offs;
 658         unsigned long                   addr_filters_gen;
 659
 660         void (*destroy)(struct perf_event *);
 661         struct rcu_head                 rcu_head;
 662
 663         struct pid_namespace            *ns;
 664         u64                             id;
 665
 666         u64                             (*clock)(void);
 667         perf_overflow_handler_t         overflow_handler;
 668         void                            *overflow_handler_context;
 669
 670 #ifdef CONFIG_EVENT_TRACING
 671         struct trace_event_call         *tp_event;
 672         struct event_filter             *filter;
 673 #ifdef CONFIG_FUNCTION_TRACER
 674         struct ftrace_ops               ftrace_ops;
 675 #endif
 676 #endif
 677
 678 #ifdef CONFIG_CGROUP_PERF
 679         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 680         int                             cgrp_defer_enabled;
 681 #endif
 682
 683         struct list_head                sb_list;
 684 #endif /* CONFIG_PERF_EVENTS */
 685 };
 686
 687 /**
 688  * struct perf_event_context - event context structure
 689  *
 690  * Used as a container for task events and CPU events as well:
 691  */
 692 struct perf_event_context {
 693         struct pmu                      *pmu;
 694         /*
 695          * Protect the states of the events in the list,
 696          * nr_active, and the list:
 697          */
 698         raw_spinlock_t                  lock;
 699         /*
 700          * Protect the list of events.  Locking either mutex or lock
 701          * is sufficient to ensure the list doesn't change; to change
 702          * the list you need to lock both the mutex and the spinlock.
 703          */
 704         struct mutex                    mutex;
 705
 706         struct list_head                active_ctx_list;
 707         struct list_head                pinned_groups;
 708         struct list_head                flexible_groups;
 709         struct list_head                event_list;
 710         int                             nr_events;
 711         int                             nr_active;
 712         int                             is_active;
 713         int                             nr_stat;
 714         int                             nr_freq;
 715         int                             rotate_disable;
 716         atomic_t                        refcount;
 717         struct task_struct              *task;
 718
 719         /*
 720          * Context clock, runs when context enabled.
 721          */
 722         u64                             time;
 723         u64                             timestamp;
 724
 725         /*
 726          * These fields let us detect when two contexts have both
 727          * been cloned (inherited) from a common ancestor.
 728          */
 729         struct perf_event_context       *parent_ctx;
 730         u64                             parent_gen;
 731         u64                             generation;
 732         int                             pin_count;
 733         int                             nr_cgroups;      /* cgroup evts */
 734         void                            *task_ctx_data; /* pmu specific data */
 735         struct rcu_head                 rcu_head;
 736 };
 737
 738 /*
 739  * Number of contexts where an event can trigger:
 740  *      task, softirq, hardirq, nmi.
 741  */
 742 #define PERF_NR_CONTEXTS        4
 743
 744 /**
 745  * struct perf_event_cpu_context - per cpu event context structure
 746  */
 747 struct perf_cpu_context {
 748         struct perf_event_context       ctx;
 749         struct perf_event_context       *task_ctx;
 750         int                             active_oncpu;
 751         int                             exclusive;
 752
 753         raw_spinlock_t                  hrtimer_lock;
 754         struct hrtimer                  hrtimer;
 755         ktime_t                         hrtimer_interval;
 756         unsigned int                    hrtimer_active;
 757
 758         struct pmu                      *unique_pmu;
 759         struct perf_cgroup              *cgrp;
 760 };
 761
 762 struct perf_output_handle {
 763         struct perf_event               *event;
 764         struct ring_buffer              *rb;
 765         unsigned long                   wakeup;
 766         unsigned long                   size;
 767         union {
 768                 void                    *addr;
 769                 unsigned long           head;
 770         };
 771         int                             page;
 772 };
 773
 774 #ifdef CONFIG_CGROUP_PERF
 775
 776 /*
 777  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 778  * This is a per-cpu dynamically allocated data structure.
 779  */
 780 struct perf_cgroup_info {
 781         u64                             time;
 782         u64                             timestamp;
 783 };
 784
 785 struct perf_cgroup {
 786         struct cgroup_subsys_state      css;
 787         struct perf_cgroup_info __percpu *info;
 788 };
 789
 790 /*
 791  * Must ensure cgroup is pinned (css_get) before calling
 792  * this function. In other words, we cannot call this function
 793  * if there is no cgroup event for the current CPU context.
 794  */
 795 static inline struct perf_cgroup *
 796 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 797 {
 798         return container_of(task_css_check(task, perf_event_cgrp_id,
 799                                            ctx ? lockdep_is_held(&ctx->lock)
 800                                                : true),
 801                             struct perf_cgroup, css);
 802 }
 803 #endif /* CONFIG_CGROUP_PERF */
 804
 805 #ifdef CONFIG_PERF_EVENTS
 806
 807 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 808                                    struct perf_event *event);
 809 extern void perf_aux_output_end(struct perf_output_handle *handle,
 810                                 unsigned long size, bool truncated);
 811 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 812                                 unsigned long size);
 813 extern void *perf_get_aux(struct perf_output_handle *handle);
 814
 815 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 816 extern void perf_pmu_unregister(struct pmu *pmu);
 817
 818 extern int perf_num_counters(void);
 819 extern const char *perf_pmu_name(void);
 820 extern void __perf_event_task_sched_in(struct task_struct *prev,
 821                                        struct task_struct *task);
 822 extern void __perf_event_task_sched_out(struct task_struct *prev,
 823                                         struct task_struct *next);
 824 extern int perf_event_init_task(struct task_struct *child);
 825 extern void perf_event_exit_task(struct task_struct *child);
 826 extern void perf_event_free_task(struct task_struct *task);
 827 extern void perf_event_delayed_put(struct task_struct *task);
 828 extern struct file *perf_event_get(unsigned int fd);
 829 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 830 extern void perf_event_print_debug(void);
 831 extern void perf_pmu_disable(struct pmu *pmu);
 832 extern void perf_pmu_enable(struct pmu *pmu);
 833 extern void perf_sched_cb_dec(struct pmu *pmu);
 834 extern void perf_sched_cb_inc(struct pmu *pmu);
 835 extern int perf_event_task_disable(void);
 836 extern int perf_event_task_enable(void);
 837 extern int perf_event_refresh(struct perf_event *event, int refresh);
 838 extern void perf_event_update_userpage(struct perf_event *event);
 839 extern int perf_event_release_kernel(struct perf_event *event);
 840 extern struct perf_event *
 841 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 842                                 int cpu,
 843                                 struct task_struct *task,
 844                                 perf_overflow_handler_t callback,
 845                                 void *context);
 846 extern void perf_pmu_migrate_context(struct pmu *pmu,
 847                                 int src_cpu, int dst_cpu);
 848 extern u64 perf_event_read_local(struct perf_event *event);
 849 extern u64 perf_event_read_value(struct perf_event *event,
 850                                  u64 *enabled, u64 *running);
 851
 852
 853 struct perf_sample_data {
 854         /*
 855          * Fields set by perf_sample_data_init(), group so as to
 856          * minimize the cachelines touched.
 857          */
 858         u64                             addr;
 859         struct perf_raw_record          *raw;
 860         struct perf_branch_stack        *br_stack;
 861         u64                             period;
 862         u64                             weight;
 863         u64                             txn;
 864         union  perf_mem_data_src        data_src;
 865
 866         /*
 867          * The other fields, optionally {set,used} by
 868          * perf_{prepare,output}_sample().
 869          */
 870         u64                             type;
 871         u64                             ip;
 872         struct {
 873                 u32     pid;
 874                 u32     tid;
 875         }                               tid_entry;
 876         u64                             time;
 877         u64                             id;
 878         u64                             stream_id;
 879         struct {
 880                 u32     cpu;
 881                 u32     reserved;
 882         }                               cpu_entry;
 883         struct perf_callchain_entry     *callchain;
 884
 885         /*
 886          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 887          * on arch details.
 888          */
 889         struct perf_regs                regs_user;
 890         struct pt_regs                  regs_user_copy;
 891
 892         struct perf_regs                regs_intr;
 893         u64                             stack_user_size;
 894 } ____cacheline_aligned;
 895
 896 /* default value for data source */
 897 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 898                     PERF_MEM_S(LVL, NA)   |\
 899                     PERF_MEM_S(SNOOP, NA) |\
 900                     PERF_MEM_S(LOCK, NA)  |\
 901                     PERF_MEM_S(TLB, NA))
 902
 903 static inline void perf_sample_data_init(struct perf_sample_data *data,
 904                                          u64 addr, u64 period)
 905 {
 906         /* remaining struct members initialized in perf_prepare_sample() */
 907         data->addr = addr;
 908         data->raw  = NULL;
 909         data->br_stack = NULL;
 910         data->period = period;
 911         data->weight = 0;
 912         data->data_src.val = PERF_MEM_NA;
 913         data->txn = 0;
 914 }
 915
 916 extern void perf_output_sample(struct perf_output_handle *handle,
 917                                struct perf_event_header *header,
 918                                struct perf_sample_data *data,
 919                                struct perf_event *event);
 920 extern void perf_prepare_sample(struct perf_event_header *header,
 921                                 struct perf_sample_data *data,
 922                                 struct perf_event *event,
 923                                 struct pt_regs *regs);
 924
 925 extern int perf_event_overflow(struct perf_event *event,
 926                                  struct perf_sample_data *data,
 927                                  struct pt_regs *regs);
 928
 929 extern void perf_event_output_forward(struct perf_event *event,
 930                                      struct perf_sample_data *data,
 931                                      struct pt_regs *regs);
 932 extern void perf_event_output_backward(struct perf_event *event,
 933                                        struct perf_sample_data *data,
 934                                        struct pt_regs *regs);
 935 extern void perf_event_output(struct perf_event *event,
 936                               struct perf_sample_data *data,
 937                               struct pt_regs *regs);
 938
 939 static inline bool
 940 is_default_overflow_handler(struct perf_event *event)
 941 {
 942         if (likely(event->overflow_handler == perf_event_output_forward))
 943                 return true;
 944         if (unlikely(event->overflow_handler == perf_event_output_backward))
 945                 return true;
 946         return false;
 947 }
 948
 949 extern void
 950 perf_event_header__init_id(struct perf_event_header *header,
 951                            struct perf_sample_data *data,
 952                            struct perf_event *event);
 953 extern void
 954 perf_event__output_id_sample(struct perf_event *event,
 955                              struct perf_output_handle *handle,
 956                              struct perf_sample_data *sample);
 957
 958 extern void
 959 perf_log_lost_samples(struct perf_event *event, u64 lost);
 960
 961 static inline bool is_sampling_event(struct perf_event *event)
 962 {
 963         return event->attr.sample_period != 0;
 964 }
 965
 966 /*
 967  * Return 1 for a software event, 0 for a hardware event
 968  */
 969 static inline int is_software_event(struct perf_event *event)
 970 {
 971         return event->pmu->task_ctx_nr == perf_sw_context;
 972 }
 973
 974 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 975
 976 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 977 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 978
 979 #ifndef perf_arch_fetch_caller_regs
 980 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 981 #endif
 982
 983 /*
 984  * Take a snapshot of the regs. Skip ip and frame pointer to
 985  * the nth caller. We only need a few of the regs:
 986  * - ip for PERF_SAMPLE_IP
 987  * - cs for user_mode() tests
 988  * - bp for callchains
 989  * - eflags, for future purposes, just in case
 990  */
 991 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 992 {
 993         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 994 }
 995
 996 static __always_inline void
 997 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 998 {
 999         if (static_key_false(&perf_swevent_enabled[event_id]))
1000                 __perf_sw_event(event_id, nr, regs, addr);
1001 }
1002
1003 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1004
1005 /*
1006  * 'Special' version for the scheduler, it hard assumes no recursion,
1007  * which is guaranteed by us not actually scheduling inside other swevents
1008  * because those disable preemption.
1009  */
1010 static __always_inline void
1011 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1012 {
1013         if (static_key_false(&perf_swevent_enabled[event_id])) {
1014                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1015
1016                 perf_fetch_caller_regs(regs);
1017                 ___perf_sw_event(event_id, nr, regs, addr);
1018         }
1019 }
1020
1021 extern struct static_key_false perf_sched_events;
1022
1023 static __always_inline bool
1024 perf_sw_migrate_enabled(void)
1025 {
1026         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1027                 return true;
1028         return false;
1029 }
1030
1031 static inline void perf_event_task_migrate(struct task_struct *task)
1032 {
1033         if (perf_sw_migrate_enabled())
1034                 task->sched_migrated = 1;
1035 }
1036
1037 static inline void perf_event_task_sched_in(struct task_struct *prev,
1038                                             struct task_struct *task)
1039 {
1040         if (static_branch_unlikely(&perf_sched_events))
1041                 __perf_event_task_sched_in(prev, task);
1042
1043         if (perf_sw_migrate_enabled() && task->sched_migrated) {
1044                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1045
1046                 perf_fetch_caller_regs(regs);
1047                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1048                 task->sched_migrated = 0;
1049         }
1050 }
1051
1052 static inline void perf_event_task_sched_out(struct task_struct *prev,
1053                                              struct task_struct *next)
1054 {
1055         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1056
1057         if (static_branch_unlikely(&perf_sched_events))
1058                 __perf_event_task_sched_out(prev, next);
1059 }
1060
1061 static inline u64 __perf_event_count(struct perf_event *event)
1062 {
1063         return local64_read(&event->count) + atomic64_read(&event->child_count);
1064 }
1065
1066 extern void perf_event_mmap(struct vm_area_struct *vma);
1067 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1068 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1069 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1070
1071 extern void perf_event_exec(void);
1072 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1073 extern void perf_event_fork(struct task_struct *tsk);
1074
1075 /* Callchains */
1076 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1077
1078 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1079 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1080 extern struct perf_callchain_entry *
1081 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1082                    u32 max_stack, bool crosstask, bool add_mark);
1083 extern int get_callchain_buffers(int max_stack);
1084 extern void put_callchain_buffers(void);
1085
1086 extern int sysctl_perf_event_max_stack;
1087 extern int sysctl_perf_event_max_contexts_per_stack;
1088
1089 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1090 {
1091         if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1092                 struct perf_callchain_entry *entry = ctx->entry;
1093                 entry->ip[entry->nr++] = ip;
1094                 ++ctx->contexts;
1095                 return 0;
1096         } else {
1097                 ctx->contexts_maxed = true;
1098                 return -1; /* no more room, stop walking the stack */
1099         }
1100 }
1101
1102 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1103 {
1104         if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1105                 struct perf_callchain_entry *entry = ctx->entry;
1106                 entry->ip[entry->nr++] = ip;
1107                 ++ctx->nr;
1108                 return 0;
1109         } else {
1110                 return -1; /* no more room, stop walking the stack */
1111         }
1112 }
1113
1114 extern int sysctl_perf_event_paranoid;
1115 extern int sysctl_perf_event_mlock;
1116 extern int sysctl_perf_event_sample_rate;
1117 extern int sysctl_perf_cpu_time_max_percent;
1118
1119 extern void perf_sample_event_took(u64 sample_len_ns);
1120
1121 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1122                 void __user *buffer, size_t *lenp,
1123                 loff_t *ppos);
1124 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1125                 void __user *buffer, size_t *lenp,
1126                 loff_t *ppos);
1127
1128 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1129                                  void __user *buffer, size_t *lenp, loff_t *ppos);
1130
1131 static inline bool perf_paranoid_tracepoint_raw(void)
1132 {
1133         return sysctl_perf_event_paranoid > -1;
1134 }
1135
1136 static inline bool perf_paranoid_cpu(void)
1137 {
1138         return sysctl_perf_event_paranoid > 0;
1139 }
1140
1141 static inline bool perf_paranoid_kernel(void)
1142 {
1143         return sysctl_perf_event_paranoid > 1;
1144 }
1145
1146 extern void perf_event_init(void);
1147 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1148                           int entry_size, struct pt_regs *regs,
1149                           struct hlist_head *head, int rctx,
1150                           struct task_struct *task);
1151 extern void perf_bp_event(struct perf_event *event, void *data);
1152
1153 #ifndef perf_misc_flags
1154 # define perf_misc_flags(regs) \
1155                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1156 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1157 #endif
1158
1159 static inline bool has_branch_stack(struct perf_event *event)
1160 {
1161         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1162 }
1163
1164 static inline bool needs_branch_stack(struct perf_event *event)
1165 {
1166         return event->attr.branch_sample_type != 0;
1167 }
1168
1169 static inline bool has_aux(struct perf_event *event)
1170 {
1171         return event->pmu->setup_aux;
1172 }
1173
1174 static inline bool is_write_backward(struct perf_event *event)
1175 {
1176         return !!event->attr.write_backward;
1177 }
1178
1179 static inline bool has_addr_filter(struct perf_event *event)
1180 {
1181         return event->pmu->nr_addr_filters;
1182 }
1183
1184 /*
1185  * An inherited event uses parent's filters
1186  */
1187 static inline struct perf_addr_filters_head *
1188 perf_event_addr_filters(struct perf_event *event)
1189 {
1190         struct perf_addr_filters_head *ifh = &event->addr_filters;
1191
1192         if (event->parent)
1193                 ifh = &event->parent->addr_filters;
1194
1195         return ifh;
1196 }
1197
1198 extern void perf_event_addr_filters_sync(struct perf_event *event);
1199
1200 extern int perf_output_begin(struct perf_output_handle *handle,
1201                              struct perf_event *event, unsigned int size);
1202 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1203                                     struct perf_event *event,
1204                                     unsigned int size);
1205 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1206                                       struct perf_event *event,
1207                                       unsigned int size);
1208
1209 extern void perf_output_end(struct perf_output_handle *handle);
1210 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1211                              const void *buf, unsigned int len);
1212 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1213                                      unsigned int len);
1214 extern int perf_swevent_get_recursion_context(void);
1215 extern void perf_swevent_put_recursion_context(int rctx);
1216 extern u64 perf_swevent_set_period(struct perf_event *event);
1217 extern void perf_event_enable(struct perf_event *event);
1218 extern void perf_event_disable(struct perf_event *event);
1219 extern void perf_event_disable_local(struct perf_event *event);
1220 extern void perf_event_task_tick(void);
1221 #else /* !CONFIG_PERF_EVENTS: */
1222 static inline void *
1223 perf_aux_output_begin(struct perf_output_handle *handle,
1224                       struct perf_event *event)                         { return NULL; }
1225 static inline void
1226 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1227                     bool truncated)                                     { }
1228 static inline int
1229 perf_aux_output_skip(struct perf_output_handle *handle,
1230                      unsigned long size)                                { return -EINVAL; }
1231 static inline void *
1232 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1233 static inline void
1234 perf_event_task_migrate(struct task_struct *task)                       { }
1235 static inline void
1236 perf_event_task_sched_in(struct task_struct *prev,
1237                          struct task_struct *task)                      { }
1238 static inline void
1239 perf_event_task_sched_out(struct task_struct *prev,
1240                           struct task_struct *next)                     { }
1241 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1242 static inline void perf_event_exit_task(struct task_struct *child)      { }
1243 static inline void perf_event_free_task(struct task_struct *task)       { }
1244 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1245 static inline struct file *perf_event_get(unsigned int fd)      { return ERR_PTR(-EINVAL); }
1246 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1247 {
1248         return ERR_PTR(-EINVAL);
1249 }
1250 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1251 static inline void perf_event_print_debug(void)                         { }
1252 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1253 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1254 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1255 {
1256         return -EINVAL;
1257 }
1258
1259 static inline void
1260 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1261 static inline void
1262 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1263 static inline void
1264 perf_bp_event(struct perf_event *event, void *data)                     { }
1265
1266 static inline int perf_register_guest_info_callbacks
1267 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1268 static inline int perf_unregister_guest_info_callbacks
1269 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1270
1271 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1272 static inline void perf_event_exec(void)                                { }
1273 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1274 static inline void perf_event_fork(struct task_struct *tsk)             { }
1275 static inline void perf_event_init(void)                                { }
1276 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1277 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1278 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1279 static inline void perf_event_enable(struct perf_event *event)          { }
1280 static inline void perf_event_disable(struct perf_event *event)         { }
1281 static inline int __perf_event_disable(void *info)                      { return -1; }
1282 static inline void perf_event_task_tick(void)                           { }
1283 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1284 #endif
1285
1286 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1287 extern void perf_restore_debug_store(void);
1288 #else
1289 static inline void perf_restore_debug_store(void)                       { }
1290 #endif
1291
1292 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1293
1294 /*
1295  * This has to have a higher priority than migration_notifier in sched/core.c.
1296  */
1297 #define perf_cpu_notifier(fn)                                           \
1298 do {                                                                    \
1299         static struct notifier_block fn##_nb =                          \
1300                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1301         unsigned long cpu = smp_processor_id();                         \
1302         unsigned long flags;                                            \
1303                                                                         \
1304         cpu_notifier_register_begin();                                  \
1305         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1306                 (void *)(unsigned long)cpu);                            \
1307         local_irq_save(flags);                                          \
1308         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1309                 (void *)(unsigned long)cpu);                            \
1310         local_irq_restore(flags);                                       \
1311         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1312                 (void *)(unsigned long)cpu);                            \
1313         __register_cpu_notifier(&fn##_nb);                              \
1314         cpu_notifier_register_done();                                   \
1315 } while (0)
1316
1317 /*
1318  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1319  * callback for already online CPUs.
1320  */
1321 #define __perf_cpu_notifier(fn)                                         \
1322 do {                                                                    \
1323         static struct notifier_block fn##_nb =                          \
1324                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1325                                                                         \
1326         __register_cpu_notifier(&fn##_nb);                              \
1327 } while (0)
1328
1329 struct perf_pmu_events_attr {
1330         struct device_attribute attr;
1331         u64 id;
1332         const char *event_str;
1333 };
1334
1335 struct perf_pmu_events_ht_attr {
1336         struct device_attribute                 attr;
1337         u64                                     id;
1338         const char                              *event_str_ht;
1339         const char                              *event_str_noht;
1340 };
1341
1342 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1343                               char *page);
1344
1345 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1346 static struct perf_pmu_events_attr _var = {                             \
1347         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1348         .id   =  _id,                                                   \
1349 };
1350
1351 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1352 static struct perf_pmu_events_attr _var = {                                 \
1353         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1354         .id             = 0,                                                \
1355         .event_str      = _str,                                             \
1356 };
1357
1358 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1359 static ssize_t                                                          \
1360 _name##_show(struct device *dev,                                        \
1361                                struct device_attribute *attr,           \
1362                                char *page)                              \
1363 {                                                                       \
1364         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1365         return sprintf(page, _format "\n");                             \
1366 }                                                                       \
1367                                                                         \
1368 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1369
1370 #endif /* _LINUX_PERF_EVENT_H */