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 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
  73                                      unsigned long len);
  74
  75 struct perf_raw_frag {
  76         union {
  77                 struct perf_raw_frag    *next;
  78                 unsigned long           pad;
  79         };
  80         perf_copy_f                     copy;
  81         void                            *data;
  82         u32                             size;
  83 } __packed;
  84
  85 struct perf_raw_record {
  86         struct perf_raw_frag            frag;
  87         u32                             size;
  88 };
  89
  90 /*
  91  * branch stack layout:
  92  *  nr: number of taken branches stored in entries[]
  93  *
  94  * Note that nr can vary from sample to sample
  95  * branches (to, from) are stored from most recent
  96  * to least recent, i.e., entries[0] contains the most
  97  * recent branch.
  98  */
  99 struct perf_branch_stack {
 100         __u64                           nr;
 101         struct perf_branch_entry        entries[0];
 102 };
 103
 104 struct task_struct;
 105
 106 /*
 107  * extra PMU register associated with an event
 108  */
 109 struct hw_perf_event_extra {
 110         u64             config; /* register value */
 111         unsigned int    reg;    /* register address or index */
 112         int             alloc;  /* extra register already allocated */
 113         int             idx;    /* index in shared_regs->regs[] */
 114 };
 115
 116 /**
 117  * struct hw_perf_event - performance event hardware details:
 118  */
 119 struct hw_perf_event {
 120 #ifdef CONFIG_PERF_EVENTS
 121         union {
 122                 struct { /* hardware */
 123                         u64             config;
 124                         u64             last_tag;
 125                         unsigned long   config_base;
 126                         unsigned long   event_base;
 127                         int             event_base_rdpmc;
 128                         int             idx;
 129                         int             last_cpu;
 130                         int             flags;
 131
 132                         struct hw_perf_event_extra extra_reg;
 133                         struct hw_perf_event_extra branch_reg;
 134                 };
 135                 struct { /* software */
 136                         struct hrtimer  hrtimer;
 137                 };
 138                 struct { /* tracepoint */
 139                         /* for tp_event->class */
 140                         struct list_head        tp_list;
 141                 };
 142                 struct { /* intel_cqm */
 143                         int                     cqm_state;
 144                         u32                     cqm_rmid;
 145                         int                     is_group_event;
 146                         struct list_head        cqm_events_entry;
 147                         struct list_head        cqm_groups_entry;
 148                         struct list_head        cqm_group_entry;
 149                 };
 150                 struct { /* itrace */
 151                         int                     itrace_started;
 152                 };
 153                 struct { /* amd_power */
 154                         u64     pwr_acc;
 155                         u64     ptsc;
 156                 };
 157 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 158                 struct { /* breakpoint */
 159                         /*
 160                          * Crufty hack to avoid the chicken and egg
 161                          * problem hw_breakpoint has with context
 162                          * creation and event initalization.
 163                          */
 164                         struct arch_hw_breakpoint       info;
 165                         struct list_head                bp_list;
 166                 };
 167 #endif
 168         };
 169         /*
 170          * If the event is a per task event, this will point to the task in
 171          * question. See the comment in perf_event_alloc().
 172          */
 173         struct task_struct              *target;
 174
 175         /*
 176          * PMU would store hardware filter configuration
 177          * here.
 178          */
 179         void                            *addr_filters;
 180
 181         /* Last sync'ed generation of filters */
 182         unsigned long                   addr_filters_gen;
 183
 184 /*
 185  * hw_perf_event::state flags; used to track the PERF_EF_* state.
 186  */
 187 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 188 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 189 #define PERF_HES_ARCH           0x04
 190
 191         int                             state;
 192
 193         /*
 194          * The last observed hardware counter value, updated with a
 195          * local64_cmpxchg() such that pmu::read() can be called nested.
 196          */
 197         local64_t                       prev_count;
 198
 199         /*
 200          * The period to start the next sample with.
 201          */
 202         u64                             sample_period;
 203
 204         /*
 205          * The period we started this sample with.
 206          */
 207         u64                             last_period;
 208
 209         /*
 210          * However much is left of the current period; note that this is
 211          * a full 64bit value and allows for generation of periods longer
 212          * than hardware might allow.
 213          */
 214         local64_t                       period_left;
 215
 216         /*
 217          * State for throttling the event, see __perf_event_overflow() and
 218          * perf_adjust_freq_unthr_context().
 219          */
 220         u64                             interrupts_seq;
 221         u64                             interrupts;
 222
 223         /*
 224          * State for freq target events, see __perf_event_overflow() and
 225          * perf_adjust_freq_unthr_context().
 226          */
 227         u64                             freq_time_stamp;
 228         u64                             freq_count_stamp;
 229 #endif
 230 };
 231
 232 struct perf_event;
 233
 234 /*
 235  * Common implementation detail of pmu::{start,commit,cancel}_txn
 236  */
 237 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
 238 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
 239
 240 /**
 241  * pmu::capabilities flags
 242  */
 243 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 244 #define PERF_PMU_CAP_NO_NMI                     0x02
 245 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
 246 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
 247 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
 248 #define PERF_PMU_CAP_ITRACE                     0x20
 249 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS         0x40
 250
 251 /**
 252  * struct pmu - generic performance monitoring unit
 253  */
 254 struct pmu {
 255         struct list_head                entry;
 256
 257         struct module                   *module;
 258         struct device                   *dev;
 259         const struct attribute_group    **attr_groups;
 260         const char                      *name;
 261         int                             type;
 262
 263         /*
 264          * various common per-pmu feature flags
 265          */
 266         int                             capabilities;
 267
 268         int * __percpu                  pmu_disable_count;
 269         struct perf_cpu_context * __percpu pmu_cpu_context;
 270         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 271         int                             task_ctx_nr;
 272         int                             hrtimer_interval_ms;
 273
 274         /* number of address filters this PMU can do */
 275         unsigned int                    nr_addr_filters;
 276
 277         /*
 278          * Fully disable/enable this PMU, can be used to protect from the PMI
 279          * as well as for lazy/batch writing of the MSRs.
 280          */
 281         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 282         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 283
 284         /*
 285          * Try and initialize the event for this PMU.
 286          *
 287          * Returns:
 288          *  -ENOENT     -- @event is not for this PMU
 289          *
 290          *  -ENODEV     -- @event is for this PMU but PMU not present
 291          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
 292          *  -EINVAL     -- @event is for this PMU but @event is not valid
 293          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
 294          *  -EACCESS    -- @event is for this PMU, @event is valid, but no privilidges
 295          *
 296          *  0           -- @event is for this PMU and valid
 297          *
 298          * Other error return values are allowed.
 299          */
 300         int (*event_init)               (struct perf_event *event);
 301
 302         /*
 303          * Notification that the event was mapped or unmapped.  Called
 304          * in the context of the mapping task.
 305          */
 306         void (*event_mapped)            (struct perf_event *event); /*optional*/
 307         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 308
 309         /*
 310          * Flags for ->add()/->del()/ ->start()/->stop(). There are
 311          * matching hw_perf_event::state flags.
 312          */
 313 #define PERF_EF_START   0x01            /* start the counter when adding    */
 314 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 315 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 316
 317         /*
 318          * Adds/Removes a counter to/from the PMU, can be done inside a
 319          * transaction, see the ->*_txn() methods.
 320          *
 321          * The add/del callbacks will reserve all hardware resources required
 322          * to service the event, this includes any counter constraint
 323          * scheduling etc.
 324          *
 325          * Called with IRQs disabled and the PMU disabled on the CPU the event
 326          * is on.
 327          *
 328          * ->add() called without PERF_EF_START should result in the same state
 329          *  as ->add() followed by ->stop().
 330          *
 331          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
 332          *  ->stop() that must deal with already being stopped without
 333          *  PERF_EF_UPDATE.
 334          */
 335         int  (*add)                     (struct perf_event *event, int flags);
 336         void (*del)                     (struct perf_event *event, int flags);
 337
 338         /*
 339          * Starts/Stops a counter present on the PMU.
 340          *
 341          * The PMI handler should stop the counter when perf_event_overflow()
 342          * returns !0. ->start() will be used to continue.
 343          *
 344          * Also used to change the sample period.
 345          *
 346          * Called with IRQs disabled and the PMU disabled on the CPU the event
 347          * is on -- will be called from NMI context with the PMU generates
 348          * NMIs.
 349          *
 350          * ->stop() with PERF_EF_UPDATE will read the counter and update
 351          *  period/count values like ->read() would.
 352          *
 353          * ->start() with PERF_EF_RELOAD will reprogram the the counter
 354          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
 355          */
 356         void (*start)                   (struct perf_event *event, int flags);
 357         void (*stop)                    (struct perf_event *event, int flags);
 358
 359         /*
 360          * Updates the counter value of the event.
 361          *
 362          * For sampling capable PMUs this will also update the software period
 363          * hw_perf_event::period_left field.
 364          */
 365         void (*read)                    (struct perf_event *event);
 366
 367         /*
 368          * Group events scheduling is treated as a transaction, add
 369          * group events as a whole and perform one schedulability test.
 370          * If the test fails, roll back the whole group
 371          *
 372          * Start the transaction, after this ->add() doesn't need to
 373          * do schedulability tests.
 374          *
 375          * Optional.
 376          */
 377         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
 378         /*
 379          * If ->start_txn() disabled the ->add() schedulability test
 380          * then ->commit_txn() is required to perform one. On success
 381          * the transaction is closed. On error the transaction is kept
 382          * open until ->cancel_txn() is called.
 383          *
 384          * Optional.
 385          */
 386         int  (*commit_txn)              (struct pmu *pmu);
 387         /*
 388          * Will cancel the transaction, assumes ->del() is called
 389          * for each successful ->add() during the transaction.
 390          *
 391          * Optional.
 392          */
 393         void (*cancel_txn)              (struct pmu *pmu);
 394
 395         /*
 396          * Will return the value for perf_event_mmap_page::index for this event,
 397          * if no implementation is provided it will default to: event->hw.idx + 1.
 398          */
 399         int (*event_idx)                (struct perf_event *event); /*optional */
 400
 401         /*
 402          * context-switches callback
 403          */
 404         void (*sched_task)              (struct perf_event_context *ctx,
 405                                         bool sched_in);
 406         /*
 407          * PMU specific data size
 408          */
 409         size_t                          task_ctx_size;
 410
 411
 412         /*
 413          * Return the count value for a counter.
 414          */
 415         u64 (*count)                    (struct perf_event *event); /*optional*/
 416
 417         /*
 418          * Set up pmu-private data structures for an AUX area
 419          */
 420         void *(*setup_aux)              (int cpu, void **pages,
 421                                          int nr_pages, bool overwrite);
 422                                         /* optional */
 423
 424         /*
 425          * Free pmu-private AUX data structures
 426          */
 427         void (*free_aux)                (void *aux); /* optional */
 428
 429         /*
 430          * Validate address range filters: make sure the HW supports the
 431          * requested configuration and number of filters; return 0 if the
 432          * supplied filters are valid, -errno otherwise.
 433          *
 434          * Runs in the context of the ioctl()ing process and is not serialized
 435          * with the rest of the PMU callbacks.
 436          */
 437         int (*addr_filters_validate)    (struct list_head *filters);
 438                                         /* optional */
 439
 440         /*
 441          * Synchronize address range filter configuration:
 442          * translate hw-agnostic filters into hardware configuration in
 443          * event::hw::addr_filters.
 444          *
 445          * Runs as a part of filter sync sequence that is done in ->start()
 446          * callback by calling perf_event_addr_filters_sync().
 447          *
 448          * May (and should) traverse event::addr_filters::list, for which its
 449          * caller provides necessary serialization.
 450          */
 451         void (*addr_filters_sync)       (struct perf_event *event);
 452                                         /* optional */
 453
 454         /*
 455          * Filter events for PMU-specific reasons.
 456          */
 457         int (*filter_match)             (struct perf_event *event); /* optional */
 458 };
 459
 460 /**
 461  * struct perf_addr_filter - address range filter definition
 462  * @entry:      event's filter list linkage
 463  * @inode:      object file's inode for file-based filters
 464  * @offset:     filter range offset
 465  * @size:       filter range size
 466  * @range:      1: range, 0: address
 467  * @filter:     1: filter/start, 0: stop
 468  *
 469  * This is a hardware-agnostic filter configuration as specified by the user.
 470  */
 471 struct perf_addr_filter {
 472         struct list_head        entry;
 473         struct inode            *inode;
 474         unsigned long           offset;
 475         unsigned long           size;
 476         unsigned int            range   : 1,
 477                                 filter  : 1;
 478 };
 479
 480 /**
 481  * struct perf_addr_filters_head - container for address range filters
 482  * @list:       list of filters for this event
 483  * @lock:       spinlock that serializes accesses to the @list and event's
 484  *              (and its children's) filter generations.
 485  *
 486  * A child event will use parent's @list (and therefore @lock), so they are
 487  * bundled together; see perf_event_addr_filters().
 488  */
 489 struct perf_addr_filters_head {
 490         struct list_head        list;
 491         raw_spinlock_t          lock;
 492 };
 493
 494 /**
 495  * enum perf_event_active_state - the states of a event
 496  */
 497 enum perf_event_active_state {
 498         PERF_EVENT_STATE_DEAD           = -4,
 499         PERF_EVENT_STATE_EXIT           = -3,
 500         PERF_EVENT_STATE_ERROR          = -2,
 501         PERF_EVENT_STATE_OFF            = -1,
 502         PERF_EVENT_STATE_INACTIVE       =  0,
 503         PERF_EVENT_STATE_ACTIVE         =  1,
 504 };
 505
 506 struct file;
 507 struct perf_sample_data;
 508
 509 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 510                                         struct perf_sample_data *,
 511                                         struct pt_regs *regs);
 512
 513 enum perf_group_flag {
 514         PERF_GROUP_SOFTWARE             = 0x1,
 515 };
 516
 517 #define SWEVENT_HLIST_BITS              8
 518 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 519
 520 struct swevent_hlist {
 521         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 522         struct rcu_head                 rcu_head;
 523 };
 524
 525 #define PERF_ATTACH_CONTEXT     0x01
 526 #define PERF_ATTACH_GROUP       0x02
 527 #define PERF_ATTACH_TASK        0x04
 528 #define PERF_ATTACH_TASK_DATA   0x08
 529
 530 struct perf_cgroup;
 531 struct ring_buffer;
 532
 533 /**
 534  * struct perf_event - performance event kernel representation:
 535  */
 536 struct perf_event {
 537 #ifdef CONFIG_PERF_EVENTS
 538         /*
 539          * entry onto perf_event_context::event_list;
 540          *   modifications require ctx->lock
 541          *   RCU safe iterations.
 542          */
 543         struct list_head                event_entry;
 544
 545         /*
 546          * XXX: group_entry and sibling_list should be mutually exclusive;
 547          * either you're a sibling on a group, or you're the group leader.
 548          * Rework the code to always use the same list element.
 549          *
 550          * Locked for modification by both ctx->mutex and ctx->lock; holding
 551          * either sufficies for read.
 552          */
 553         struct list_head                group_entry;
 554         struct list_head                sibling_list;
 555
 556         /*
 557          * We need storage to track the entries in perf_pmu_migrate_context; we
 558          * cannot use the event_entry because of RCU and we want to keep the
 559          * group in tact which avoids us using the other two entries.
 560          */
 561         struct list_head                migrate_entry;
 562
 563         struct hlist_node               hlist_entry;
 564         struct list_head                active_entry;
 565         int                             nr_siblings;
 566         int                             group_flags;
 567         struct perf_event               *group_leader;
 568         struct pmu                      *pmu;
 569         void                            *pmu_private;
 570
 571         enum perf_event_active_state    state;
 572         unsigned int                    attach_state;
 573         local64_t                       count;
 574         atomic64_t                      child_count;
 575
 576         /*
 577          * These are the total time in nanoseconds that the event
 578          * has been enabled (i.e. eligible to run, and the task has
 579          * been scheduled in, if this is a per-task event)
 580          * and running (scheduled onto the CPU), respectively.
 581          *
 582          * They are computed from tstamp_enabled, tstamp_running and
 583          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 584          */
 585         u64                             total_time_enabled;
 586         u64                             total_time_running;
 587
 588         /*
 589          * These are timestamps used for computing total_time_enabled
 590          * and total_time_running when the event is in INACTIVE or
 591          * ACTIVE state, measured in nanoseconds from an arbitrary point
 592          * in time.
 593          * tstamp_enabled: the notional time when the event was enabled
 594          * tstamp_running: the notional time when the event was scheduled on
 595          * tstamp_stopped: in INACTIVE state, the notional time when the
 596          *      event was scheduled off.
 597          */
 598         u64                             tstamp_enabled;
 599         u64                             tstamp_running;
 600         u64                             tstamp_stopped;
 601
 602         /*
 603          * timestamp shadows the actual context timing but it can
 604          * be safely used in NMI interrupt context. It reflects the
 605          * context time as it was when the event was last scheduled in.
 606          *
 607          * ctx_time already accounts for ctx->timestamp. Therefore to
 608          * compute ctx_time for a sample, simply add perf_clock().
 609          */
 610         u64                             shadow_ctx_time;
 611
 612         struct perf_event_attr          attr;
 613         u16                             header_size;
 614         u16                             id_header_size;
 615         u16                             read_size;
 616         struct hw_perf_event            hw;
 617
 618         struct perf_event_context       *ctx;
 619         atomic_long_t                   refcount;
 620
 621         /*
 622          * These accumulate total time (in nanoseconds) that children
 623          * events have been enabled and running, respectively.
 624          */
 625         atomic64_t                      child_total_time_enabled;
 626         atomic64_t                      child_total_time_running;
 627
 628         /*
 629          * Protect attach/detach and child_list:
 630          */
 631         struct mutex                    child_mutex;
 632         struct list_head                child_list;
 633         struct perf_event               *parent;
 634
 635         int                             oncpu;
 636         int                             cpu;
 637
 638         struct list_head                owner_entry;
 639         struct task_struct              *owner;
 640
 641         /* mmap bits */
 642         struct mutex                    mmap_mutex;
 643         atomic_t                        mmap_count;
 644
 645         struct ring_buffer              *rb;
 646         struct list_head                rb_entry;
 647         unsigned long                   rcu_batches;
 648         int                             rcu_pending;
 649
 650         /* poll related */
 651         wait_queue_head_t               waitq;
 652         struct fasync_struct            *fasync;
 653
 654         /* delayed work for NMIs and such */
 655         int                             pending_wakeup;
 656         int                             pending_kill;
 657         int                             pending_disable;
 658         struct irq_work                 pending;
 659
 660         atomic_t                        event_limit;
 661
 662         /* address range filters */
 663         struct perf_addr_filters_head   addr_filters;
 664         /* vma address array for file-based filders */
 665         unsigned long                   *addr_filters_offs;
 666         unsigned long                   addr_filters_gen;
 667
 668         void (*destroy)(struct perf_event *);
 669         struct rcu_head                 rcu_head;
 670
 671         struct pid_namespace            *ns;
 672         u64                             id;
 673
 674         u64                             (*clock)(void);
 675         perf_overflow_handler_t         overflow_handler;
 676         void                            *overflow_handler_context;
 677
 678 #ifdef CONFIG_EVENT_TRACING
 679         struct trace_event_call         *tp_event;
 680         struct event_filter             *filter;
 681 #ifdef CONFIG_FUNCTION_TRACER
 682         struct ftrace_ops               ftrace_ops;
 683 #endif
 684 #endif
 685
 686 #ifdef CONFIG_CGROUP_PERF
 687         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 688         int                             cgrp_defer_enabled;
 689 #endif
 690
 691 #endif /* CONFIG_PERF_EVENTS */
 692 };
 693
 694 /**
 695  * struct perf_event_context - event context structure
 696  *
 697  * Used as a container for task events and CPU events as well:
 698  */
 699 struct perf_event_context {
 700         struct pmu                      *pmu;
 701         /*
 702          * Protect the states of the events in the list,
 703          * nr_active, and the list:
 704          */
 705         raw_spinlock_t                  lock;
 706         /*
 707          * Protect the list of events.  Locking either mutex or lock
 708          * is sufficient to ensure the list doesn't change; to change
 709          * the list you need to lock both the mutex and the spinlock.
 710          */
 711         struct mutex                    mutex;
 712
 713         struct list_head                active_ctx_list;
 714         struct list_head                pinned_groups;
 715         struct list_head                flexible_groups;
 716         struct list_head                event_list;
 717         int                             nr_events;
 718         int                             nr_active;
 719         int                             is_active;
 720         int                             nr_stat;
 721         int                             nr_freq;
 722         int                             rotate_disable;
 723         atomic_t                        refcount;
 724         struct task_struct              *task;
 725
 726         /*
 727          * Context clock, runs when context enabled.
 728          */
 729         u64                             time;
 730         u64                             timestamp;
 731
 732         /*
 733          * These fields let us detect when two contexts have both
 734          * been cloned (inherited) from a common ancestor.
 735          */
 736         struct perf_event_context       *parent_ctx;
 737         u64                             parent_gen;
 738         u64                             generation;
 739         int                             pin_count;
 740         int                             nr_cgroups;      /* cgroup evts */
 741         void                            *task_ctx_data; /* pmu specific data */
 742         struct rcu_head                 rcu_head;
 743 };
 744
 745 /*
 746  * Number of contexts where an event can trigger:
 747  *      task, softirq, hardirq, nmi.
 748  */
 749 #define PERF_NR_CONTEXTS        4
 750
 751 /**
 752  * struct perf_event_cpu_context - per cpu event context structure
 753  */
 754 struct perf_cpu_context {
 755         struct perf_event_context       ctx;
 756         struct perf_event_context       *task_ctx;
 757         int                             active_oncpu;
 758         int                             exclusive;
 759
 760         raw_spinlock_t                  hrtimer_lock;
 761         struct hrtimer                  hrtimer;
 762         ktime_t                         hrtimer_interval;
 763         unsigned int                    hrtimer_active;
 764
 765         struct pmu                      *unique_pmu;
 766         struct perf_cgroup              *cgrp;
 767 };
 768
 769 struct perf_output_handle {
 770         struct perf_event               *event;
 771         struct ring_buffer              *rb;
 772         unsigned long                   wakeup;
 773         unsigned long                   size;
 774         union {
 775                 void                    *addr;
 776                 unsigned long           head;
 777         };
 778         int                             page;
 779 };
 780
 781 #ifdef CONFIG_CGROUP_PERF
 782
 783 /*
 784  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 785  * This is a per-cpu dynamically allocated data structure.
 786  */
 787 struct perf_cgroup_info {
 788         u64                             time;
 789         u64                             timestamp;
 790 };
 791
 792 struct perf_cgroup {
 793         struct cgroup_subsys_state      css;
 794         struct perf_cgroup_info __percpu *info;
 795 };
 796
 797 /*
 798  * Must ensure cgroup is pinned (css_get) before calling
 799  * this function. In other words, we cannot call this function
 800  * if there is no cgroup event for the current CPU context.
 801  */
 802 static inline struct perf_cgroup *
 803 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 804 {
 805         return container_of(task_css_check(task, perf_event_cgrp_id,
 806                                            ctx ? lockdep_is_held(&ctx->lock)
 807                                                : true),
 808                             struct perf_cgroup, css);
 809 }
 810 #endif /* CONFIG_CGROUP_PERF */
 811
 812 #ifdef CONFIG_PERF_EVENTS
 813
 814 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 815                                    struct perf_event *event);
 816 extern void perf_aux_output_end(struct perf_output_handle *handle,
 817                                 unsigned long size, bool truncated);
 818 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 819                                 unsigned long size);
 820 extern void *perf_get_aux(struct perf_output_handle *handle);
 821
 822 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 823 extern void perf_pmu_unregister(struct pmu *pmu);
 824
 825 extern int perf_num_counters(void);
 826 extern const char *perf_pmu_name(void);
 827 extern void __perf_event_task_sched_in(struct task_struct *prev,
 828                                        struct task_struct *task);
 829 extern void __perf_event_task_sched_out(struct task_struct *prev,
 830                                         struct task_struct *next);
 831 extern int perf_event_init_task(struct task_struct *child);
 832 extern void perf_event_exit_task(struct task_struct *child);
 833 extern void perf_event_free_task(struct task_struct *task);
 834 extern void perf_event_delayed_put(struct task_struct *task);
 835 extern struct file *perf_event_get(unsigned int fd);
 836 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 837 extern void perf_event_print_debug(void);
 838 extern void perf_pmu_disable(struct pmu *pmu);
 839 extern void perf_pmu_enable(struct pmu *pmu);
 840 extern void perf_sched_cb_dec(struct pmu *pmu);
 841 extern void perf_sched_cb_inc(struct pmu *pmu);
 842 extern int perf_event_task_disable(void);
 843 extern int perf_event_task_enable(void);
 844 extern int perf_event_refresh(struct perf_event *event, int refresh);
 845 extern void perf_event_update_userpage(struct perf_event *event);
 846 extern int perf_event_release_kernel(struct perf_event *event);
 847 extern struct perf_event *
 848 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 849                                 int cpu,
 850                                 struct task_struct *task,
 851                                 perf_overflow_handler_t callback,
 852                                 void *context);
 853 extern void perf_pmu_migrate_context(struct pmu *pmu,
 854                                 int src_cpu, int dst_cpu);
 855 extern u64 perf_event_read_local(struct perf_event *event);
 856 extern u64 perf_event_read_value(struct perf_event *event,
 857                                  u64 *enabled, u64 *running);
 858
 859
 860 struct perf_sample_data {
 861         /*
 862          * Fields set by perf_sample_data_init(), group so as to
 863          * minimize the cachelines touched.
 864          */
 865         u64                             addr;
 866         struct perf_raw_record          *raw;
 867         struct perf_branch_stack        *br_stack;
 868         u64                             period;
 869         u64                             weight;
 870         u64                             txn;
 871         union  perf_mem_data_src        data_src;
 872
 873         /*
 874          * The other fields, optionally {set,used} by
 875          * perf_{prepare,output}_sample().
 876          */
 877         u64                             type;
 878         u64                             ip;
 879         struct {
 880                 u32     pid;
 881                 u32     tid;
 882         }                               tid_entry;
 883         u64                             time;
 884         u64                             id;
 885         u64                             stream_id;
 886         struct {
 887                 u32     cpu;
 888                 u32     reserved;
 889         }                               cpu_entry;
 890         struct perf_callchain_entry     *callchain;
 891
 892         /*
 893          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 894          * on arch details.
 895          */
 896         struct perf_regs                regs_user;
 897         struct pt_regs                  regs_user_copy;
 898
 899         struct perf_regs                regs_intr;
 900         u64                             stack_user_size;
 901 } ____cacheline_aligned;
 902
 903 /* default value for data source */
 904 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 905                     PERF_MEM_S(LVL, NA)   |\
 906                     PERF_MEM_S(SNOOP, NA) |\
 907                     PERF_MEM_S(LOCK, NA)  |\
 908                     PERF_MEM_S(TLB, NA))
 909
 910 static inline void perf_sample_data_init(struct perf_sample_data *data,
 911                                          u64 addr, u64 period)
 912 {
 913         /* remaining struct members initialized in perf_prepare_sample() */
 914         data->addr = addr;
 915         data->raw  = NULL;
 916         data->br_stack = NULL;
 917         data->period = period;
 918         data->weight = 0;
 919         data->data_src.val = PERF_MEM_NA;
 920         data->txn = 0;
 921 }
 922
 923 extern void perf_output_sample(struct perf_output_handle *handle,
 924                                struct perf_event_header *header,
 925                                struct perf_sample_data *data,
 926                                struct perf_event *event);
 927 extern void perf_prepare_sample(struct perf_event_header *header,
 928                                 struct perf_sample_data *data,
 929                                 struct perf_event *event,
 930                                 struct pt_regs *regs);
 931
 932 extern int perf_event_overflow(struct perf_event *event,
 933                                  struct perf_sample_data *data,
 934                                  struct pt_regs *regs);
 935
 936 extern void perf_event_output_forward(struct perf_event *event,
 937                                      struct perf_sample_data *data,
 938                                      struct pt_regs *regs);
 939 extern void perf_event_output_backward(struct perf_event *event,
 940                                        struct perf_sample_data *data,
 941                                        struct pt_regs *regs);
 942 extern void perf_event_output(struct perf_event *event,
 943                               struct perf_sample_data *data,
 944                               struct pt_regs *regs);
 945
 946 static inline bool
 947 is_default_overflow_handler(struct perf_event *event)
 948 {
 949         if (likely(event->overflow_handler == perf_event_output_forward))
 950                 return true;
 951         if (unlikely(event->overflow_handler == perf_event_output_backward))
 952                 return true;
 953         return false;
 954 }
 955
 956 extern void
 957 perf_event_header__init_id(struct perf_event_header *header,
 958                            struct perf_sample_data *data,
 959                            struct perf_event *event);
 960 extern void
 961 perf_event__output_id_sample(struct perf_event *event,
 962                              struct perf_output_handle *handle,
 963                              struct perf_sample_data *sample);
 964
 965 extern void
 966 perf_log_lost_samples(struct perf_event *event, u64 lost);
 967
 968 static inline bool is_sampling_event(struct perf_event *event)
 969 {
 970         return event->attr.sample_period != 0;
 971 }
 972
 973 /*
 974  * Return 1 for a software event, 0 for a hardware event
 975  */
 976 static inline int is_software_event(struct perf_event *event)
 977 {
 978         return event->pmu->task_ctx_nr == perf_sw_context;
 979 }
 980
 981 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 982
 983 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 984 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 985
 986 #ifndef perf_arch_fetch_caller_regs
 987 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 988 #endif
 989
 990 /*
 991  * Take a snapshot of the regs. Skip ip and frame pointer to
 992  * the nth caller. We only need a few of the regs:
 993  * - ip for PERF_SAMPLE_IP
 994  * - cs for user_mode() tests
 995  * - bp for callchains
 996  * - eflags, for future purposes, just in case
 997  */
 998 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 999 {
1000         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1001 }
1002
1003 static __always_inline void
1004 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1005 {
1006         if (static_key_false(&perf_swevent_enabled[event_id]))
1007                 __perf_sw_event(event_id, nr, regs, addr);
1008 }
1009
1010 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1011
1012 /*
1013  * 'Special' version for the scheduler, it hard assumes no recursion,
1014  * which is guaranteed by us not actually scheduling inside other swevents
1015  * because those disable preemption.
1016  */
1017 static __always_inline void
1018 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1019 {
1020         if (static_key_false(&perf_swevent_enabled[event_id])) {
1021                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1022
1023                 perf_fetch_caller_regs(regs);
1024                 ___perf_sw_event(event_id, nr, regs, addr);
1025         }
1026 }
1027
1028 extern struct static_key_false perf_sched_events;
1029
1030 static __always_inline bool
1031 perf_sw_migrate_enabled(void)
1032 {
1033         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
1034                 return true;
1035         return false;
1036 }
1037
1038 static inline void perf_event_task_migrate(struct task_struct *task)
1039 {
1040         if (perf_sw_migrate_enabled())
1041                 task->sched_migrated = 1;
1042 }
1043
1044 static inline void perf_event_task_sched_in(struct task_struct *prev,
1045                                             struct task_struct *task)
1046 {
1047         if (static_branch_unlikely(&perf_sched_events))
1048                 __perf_event_task_sched_in(prev, task);
1049
1050         if (perf_sw_migrate_enabled() && task->sched_migrated) {
1051                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1052
1053                 perf_fetch_caller_regs(regs);
1054                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
1055                 task->sched_migrated = 0;
1056         }
1057 }
1058
1059 static inline void perf_event_task_sched_out(struct task_struct *prev,
1060                                              struct task_struct *next)
1061 {
1062         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1063
1064         if (static_branch_unlikely(&perf_sched_events))
1065                 __perf_event_task_sched_out(prev, next);
1066 }
1067
1068 static inline u64 __perf_event_count(struct perf_event *event)
1069 {
1070         return local64_read(&event->count) + atomic64_read(&event->child_count);
1071 }
1072
1073 extern void perf_event_mmap(struct vm_area_struct *vma);
1074 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1075 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1076 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1077
1078 extern void perf_event_exec(void);
1079 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1080 extern void perf_event_fork(struct task_struct *tsk);
1081
1082 /* Callchains */
1083 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1084
1085 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1086 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1087 extern struct perf_callchain_entry *
1088 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1089                    u32 max_stack, bool crosstask, bool add_mark);
1090 extern int get_callchain_buffers(void);
1091 extern void put_callchain_buffers(void);
1092
1093 extern int sysctl_perf_event_max_stack;
1094 extern int sysctl_perf_event_max_contexts_per_stack;
1095
1096 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1097 {
1098         if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1099                 struct perf_callchain_entry *entry = ctx->entry;
1100                 entry->ip[entry->nr++] = ip;
1101                 ++ctx->contexts;
1102                 return 0;
1103         } else {
1104                 ctx->contexts_maxed = true;
1105                 return -1; /* no more room, stop walking the stack */
1106         }
1107 }
1108
1109 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1110 {
1111         if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1112                 struct perf_callchain_entry *entry = ctx->entry;
1113                 entry->ip[entry->nr++] = ip;
1114                 ++ctx->nr;
1115                 return 0;
1116         } else {
1117                 return -1; /* no more room, stop walking the stack */
1118         }
1119 }
1120
1121 extern int sysctl_perf_event_paranoid;
1122 extern int sysctl_perf_event_mlock;
1123 extern int sysctl_perf_event_sample_rate;
1124 extern int sysctl_perf_cpu_time_max_percent;
1125
1126 extern void perf_sample_event_took(u64 sample_len_ns);
1127
1128 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1129                 void __user *buffer, size_t *lenp,
1130                 loff_t *ppos);
1131 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1132                 void __user *buffer, size_t *lenp,
1133                 loff_t *ppos);
1134
1135 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1136                                  void __user *buffer, size_t *lenp, loff_t *ppos);
1137
1138 static inline bool perf_paranoid_tracepoint_raw(void)
1139 {
1140         return sysctl_perf_event_paranoid > -1;
1141 }
1142
1143 static inline bool perf_paranoid_cpu(void)
1144 {
1145         return sysctl_perf_event_paranoid > 0;
1146 }
1147
1148 static inline bool perf_paranoid_kernel(void)
1149 {
1150         return sysctl_perf_event_paranoid > 1;
1151 }
1152
1153 extern void perf_event_init(void);
1154 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1155                           int entry_size, struct pt_regs *regs,
1156                           struct hlist_head *head, int rctx,
1157                           struct task_struct *task);
1158 extern void perf_bp_event(struct perf_event *event, void *data);
1159
1160 #ifndef perf_misc_flags
1161 # define perf_misc_flags(regs) \
1162                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1163 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1164 #endif
1165
1166 static inline bool has_branch_stack(struct perf_event *event)
1167 {
1168         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1169 }
1170
1171 static inline bool needs_branch_stack(struct perf_event *event)
1172 {
1173         return event->attr.branch_sample_type != 0;
1174 }
1175
1176 static inline bool has_aux(struct perf_event *event)
1177 {
1178         return event->pmu->setup_aux;
1179 }
1180
1181 static inline bool is_write_backward(struct perf_event *event)
1182 {
1183         return !!event->attr.write_backward;
1184 }
1185
1186 static inline bool has_addr_filter(struct perf_event *event)
1187 {
1188         return event->pmu->nr_addr_filters;
1189 }
1190
1191 /*
1192  * An inherited event uses parent's filters
1193  */
1194 static inline struct perf_addr_filters_head *
1195 perf_event_addr_filters(struct perf_event *event)
1196 {
1197         struct perf_addr_filters_head *ifh = &event->addr_filters;
1198
1199         if (event->parent)
1200                 ifh = &event->parent->addr_filters;
1201
1202         return ifh;
1203 }
1204
1205 extern void perf_event_addr_filters_sync(struct perf_event *event);
1206
1207 extern int perf_output_begin(struct perf_output_handle *handle,
1208                              struct perf_event *event, unsigned int size);
1209 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1210                                     struct perf_event *event,
1211                                     unsigned int size);
1212 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1213                                       struct perf_event *event,
1214                                       unsigned int size);
1215
1216 extern void perf_output_end(struct perf_output_handle *handle);
1217 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1218                              const void *buf, unsigned int len);
1219 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1220                                      unsigned int len);
1221 extern int perf_swevent_get_recursion_context(void);
1222 extern void perf_swevent_put_recursion_context(int rctx);
1223 extern u64 perf_swevent_set_period(struct perf_event *event);
1224 extern void perf_event_enable(struct perf_event *event);
1225 extern void perf_event_disable(struct perf_event *event);
1226 extern void perf_event_disable_local(struct perf_event *event);
1227 extern void perf_event_task_tick(void);
1228 #else /* !CONFIG_PERF_EVENTS: */
1229 static inline void *
1230 perf_aux_output_begin(struct perf_output_handle *handle,
1231                       struct perf_event *event)                         { return NULL; }
1232 static inline void
1233 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1234                     bool truncated)                                     { }
1235 static inline int
1236 perf_aux_output_skip(struct perf_output_handle *handle,
1237                      unsigned long size)                                { return -EINVAL; }
1238 static inline void *
1239 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1240 static inline void
1241 perf_event_task_migrate(struct task_struct *task)                       { }
1242 static inline void
1243 perf_event_task_sched_in(struct task_struct *prev,
1244                          struct task_struct *task)                      { }
1245 static inline void
1246 perf_event_task_sched_out(struct task_struct *prev,
1247                           struct task_struct *next)                     { }
1248 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1249 static inline void perf_event_exit_task(struct task_struct *child)      { }
1250 static inline void perf_event_free_task(struct task_struct *task)       { }
1251 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1252 static inline struct file *perf_event_get(unsigned int fd)      { return ERR_PTR(-EINVAL); }
1253 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1254 {
1255         return ERR_PTR(-EINVAL);
1256 }
1257 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1258 static inline void perf_event_print_debug(void)                         { }
1259 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1260 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1261 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1262 {
1263         return -EINVAL;
1264 }
1265
1266 static inline void
1267 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1268 static inline void
1269 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1270 static inline void
1271 perf_bp_event(struct perf_event *event, void *data)                     { }
1272
1273 static inline int perf_register_guest_info_callbacks
1274 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1275 static inline int perf_unregister_guest_info_callbacks
1276 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1277
1278 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1279 static inline void perf_event_exec(void)                                { }
1280 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1281 static inline void perf_event_fork(struct task_struct *tsk)             { }
1282 static inline void perf_event_init(void)                                { }
1283 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1284 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1285 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1286 static inline void perf_event_enable(struct perf_event *event)          { }
1287 static inline void perf_event_disable(struct perf_event *event)         { }
1288 static inline int __perf_event_disable(void *info)                      { return -1; }
1289 static inline void perf_event_task_tick(void)                           { }
1290 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1291 #endif
1292
1293 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1294 extern void perf_restore_debug_store(void);
1295 #else
1296 static inline void perf_restore_debug_store(void)                       { }
1297 #endif
1298
1299 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
1300 {
1301         return frag->pad < sizeof(u64);
1302 }
1303
1304 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1305
1306 /*
1307  * This has to have a higher priority than migration_notifier in sched/core.c.
1308  */
1309 #define perf_cpu_notifier(fn)                                           \
1310 do {                                                                    \
1311         static struct notifier_block fn##_nb =                          \
1312                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1313         unsigned long cpu = smp_processor_id();                         \
1314         unsigned long flags;                                            \
1315                                                                         \
1316         cpu_notifier_register_begin();                                  \
1317         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1318                 (void *)(unsigned long)cpu);                            \
1319         local_irq_save(flags);                                          \
1320         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1321                 (void *)(unsigned long)cpu);                            \
1322         local_irq_restore(flags);                                       \
1323         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1324                 (void *)(unsigned long)cpu);                            \
1325         __register_cpu_notifier(&fn##_nb);                              \
1326         cpu_notifier_register_done();                                   \
1327 } while (0)
1328
1329 /*
1330  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1331  * callback for already online CPUs.
1332  */
1333 #define __perf_cpu_notifier(fn)                                         \
1334 do {                                                                    \
1335         static struct notifier_block fn##_nb =                          \
1336                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1337                                                                         \
1338         __register_cpu_notifier(&fn##_nb);                              \
1339 } while (0)
1340
1341 struct perf_pmu_events_attr {
1342         struct device_attribute attr;
1343         u64 id;
1344         const char *event_str;
1345 };
1346
1347 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1348                               char *page);
1349
1350 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1351 static struct perf_pmu_events_attr _var = {                             \
1352         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1353         .id   =  _id,                                                   \
1354 };
1355
1356 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1357 static struct perf_pmu_events_attr _var = {                                 \
1358         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1359         .id             = 0,                                                \
1360         .event_str      = _str,                                             \
1361 };
1362
1363 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1364 static ssize_t                                                          \
1365 _name##_show(struct device *dev,                                        \
1366                                struct device_attribute *attr,           \
1367                                char *page)                              \
1368 {                                                                       \
1369         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1370         return sprintf(page, _format "\n");                             \
1371 }                                                                       \
1372                                                                         \
1373 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1374
1375 #endif /* _LINUX_PERF_EVENT_H */