include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   7  *
   8  * Data type definitions, declarations, prototypes.
   9  *
  10  *    Started by: Thomas Gleixner and Ingo Molnar
  11  *
  12  * For licencing details see kernel-base/COPYING
  13  */
  14 #ifndef _LINUX_PERF_EVENT_H
  15 #define _LINUX_PERF_EVENT_H
  16
  17 #include <uapi/linux/perf_event.h>
  18
  19 /*
  20  * Kernel-internal data types and definitions:
  21  */
  22
  23 #ifdef CONFIG_PERF_EVENTS
  24 # include <asm/perf_event.h>
  25 # include <asm/local64.h>
  26 #endif
  27
  28 struct perf_guest_info_callbacks {
  29         int                             (*is_in_guest)(void);
  30         int                             (*is_user_mode)(void);
  31         unsigned long                   (*get_guest_ip)(void);
  32 };
  33
  34 #ifdef CONFIG_HAVE_HW_BREAKPOINT
  35 #include <asm/hw_breakpoint.h>
  36 #endif
  37
  38 #include <linux/list.h>
  39 #include <linux/mutex.h>
  40 #include <linux/rculist.h>
  41 #include <linux/rcupdate.h>
  42 #include <linux/spinlock.h>
  43 #include <linux/hrtimer.h>
  44 #include <linux/fs.h>
  45 #include <linux/pid_namespace.h>
  46 #include <linux/workqueue.h>
  47 #include <linux/ftrace.h>
  48 #include <linux/cpu.h>
  49 #include <linux/irq_work.h>
  50 #include <linux/static_key.h>
  51 #include <linux/jump_label_ratelimit.h>
  52 #include <linux/atomic.h>
  53 #include <linux/sysfs.h>
  54 #include <linux/perf_regs.h>
  55 #include <linux/workqueue.h>
  56 #include <linux/cgroup.h>
  57 #include <asm/local.h>
  58
  59 struct perf_callchain_entry {
  60         __u64                           nr;
  61         __u64                           ip[PERF_MAX_STACK_DEPTH];
  62 };
  63
  64 struct perf_raw_record {
  65         u32                             size;
  66         void                            *data;
  67 };
  68
  69 /*
  70  * branch stack layout:
  71  *  nr: number of taken branches stored in entries[]
  72  *
  73  * Note that nr can vary from sample to sample
  74  * branches (to, from) are stored from most recent
  75  * to least recent, i.e., entries[0] contains the most
  76  * recent branch.
  77  */
  78 struct perf_branch_stack {
  79         __u64                           nr;
  80         struct perf_branch_entry        entries[0];
  81 };
  82
  83 struct task_struct;
  84
  85 /*
  86  * extra PMU register associated with an event
  87  */
  88 struct hw_perf_event_extra {
  89         u64             config; /* register value */
  90         unsigned int    reg;    /* register address or index */
  91         int             alloc;  /* extra register already allocated */
  92         int             idx;    /* index in shared_regs->regs[] */
  93 };
  94
  95 /**
  96  * struct hw_perf_event - performance event hardware details:
  97  */
  98 struct hw_perf_event {
  99 #ifdef CONFIG_PERF_EVENTS
 100         union {
 101                 struct { /* hardware */
 102                         u64             config;
 103                         u64             last_tag;
 104                         unsigned long   config_base;
 105                         unsigned long   event_base;
 106                         int             event_base_rdpmc;
 107                         int             idx;
 108                         int             last_cpu;
 109                         int             flags;
 110
 111                         struct hw_perf_event_extra extra_reg;
 112                         struct hw_perf_event_extra branch_reg;
 113                 };
 114                 struct { /* software */
 115                         struct hrtimer  hrtimer;
 116                 };
 117                 struct { /* tracepoint */
 118                         /* for tp_event->class */
 119                         struct list_head        tp_list;
 120                 };
 121                 struct { /* intel_cqm */
 122                         int                     cqm_state;
 123                         u32                     cqm_rmid;
 124                         int                     is_group_event;
 125                         struct list_head        cqm_events_entry;
 126                         struct list_head        cqm_groups_entry;
 127                         struct list_head        cqm_group_entry;
 128                 };
 129                 struct { /* itrace */
 130                         int                     itrace_started;
 131                 };
 132 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 133                 struct { /* breakpoint */
 134                         /*
 135                          * Crufty hack to avoid the chicken and egg
 136                          * problem hw_breakpoint has with context
 137                          * creation and event initalization.
 138                          */
 139                         struct arch_hw_breakpoint       info;
 140                         struct list_head                bp_list;
 141                 };
 142 #endif
 143         };
 144         /*
 145          * If the event is a per task event, this will point to the task in
 146          * question. See the comment in perf_event_alloc().
 147          */
 148         struct task_struct              *target;
 149
 150 /*
 151  * hw_perf_event::state flags; used to track the PERF_EF_* state.
 152  */
 153 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 154 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 155 #define PERF_HES_ARCH           0x04
 156
 157         int                             state;
 158
 159         /*
 160          * The last observed hardware counter value, updated with a
 161          * local64_cmpxchg() such that pmu::read() can be called nested.
 162          */
 163         local64_t                       prev_count;
 164
 165         /*
 166          * The period to start the next sample with.
 167          */
 168         u64                             sample_period;
 169
 170         /*
 171          * The period we started this sample with.
 172          */
 173         u64                             last_period;
 174
 175         /*
 176          * However much is left of the current period; note that this is
 177          * a full 64bit value and allows for generation of periods longer
 178          * than hardware might allow.
 179          */
 180         local64_t                       period_left;
 181
 182         /*
 183          * State for throttling the event, see __perf_event_overflow() and
 184          * perf_adjust_freq_unthr_context().
 185          */
 186         u64                             interrupts_seq;
 187         u64                             interrupts;
 188
 189         /*
 190          * State for freq target events, see __perf_event_overflow() and
 191          * perf_adjust_freq_unthr_context().
 192          */
 193         u64                             freq_time_stamp;
 194         u64                             freq_count_stamp;
 195 #endif
 196 };
 197
 198 struct perf_event;
 199
 200 /*
 201  * Common implementation detail of pmu::{start,commit,cancel}_txn
 202  */
 203 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
 204 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
 205
 206 /**
 207  * pmu::capabilities flags
 208  */
 209 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 210 #define PERF_PMU_CAP_NO_NMI                     0x02
 211 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
 212 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
 213 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
 214 #define PERF_PMU_CAP_ITRACE                     0x20
 215
 216 /**
 217  * struct pmu - generic performance monitoring unit
 218  */
 219 struct pmu {
 220         struct list_head                entry;
 221
 222         struct module                   *module;
 223         struct device                   *dev;
 224         const struct attribute_group    **attr_groups;
 225         const char                      *name;
 226         int                             type;
 227
 228         /*
 229          * various common per-pmu feature flags
 230          */
 231         int                             capabilities;
 232
 233         int * __percpu                  pmu_disable_count;
 234         struct perf_cpu_context __percpu *pmu_cpu_context;
 235         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 236         int                             task_ctx_nr;
 237         int                             hrtimer_interval_ms;
 238         u32                             events_across_hotplug:1,
 239                                         reserved:31;
 240
 241         /*
 242          * Fully disable/enable this PMU, can be used to protect from the PMI
 243          * as well as for lazy/batch writing of the MSRs.
 244          */
 245         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 246         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 247
 248         /*
 249          * Try and initialize the event for this PMU.
 250          *
 251          * Returns:
 252          *  -ENOENT     -- @event is not for this PMU
 253          *
 254          *  -ENODEV     -- @event is for this PMU but PMU not present
 255          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
 256          *  -EINVAL     -- @event is for this PMU but @event is not valid
 257          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
 258          *  -EACCESS    -- @event is for this PMU, @event is valid, but no privilidges
 259          *
 260          *  0           -- @event is for this PMU and valid
 261          *
 262          * Other error return values are allowed.
 263          */
 264         int (*event_init)               (struct perf_event *event);
 265
 266         /*
 267          * Notification that the event was mapped or unmapped.  Called
 268          * in the context of the mapping task.
 269          */
 270         void (*event_mapped)            (struct perf_event *event); /*optional*/
 271         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 272
 273         /*
 274          * Flags for ->add()/->del()/ ->start()/->stop(). There are
 275          * matching hw_perf_event::state flags.
 276          */
 277 #define PERF_EF_START   0x01            /* start the counter when adding    */
 278 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 279 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 280
 281         /*
 282          * Adds/Removes a counter to/from the PMU, can be done inside a
 283          * transaction, see the ->*_txn() methods.
 284          *
 285          * The add/del callbacks will reserve all hardware resources required
 286          * to service the event, this includes any counter constraint
 287          * scheduling etc.
 288          *
 289          * Called with IRQs disabled and the PMU disabled on the CPU the event
 290          * is on.
 291          *
 292          * ->add() called without PERF_EF_START should result in the same state
 293          *  as ->add() followed by ->stop().
 294          *
 295          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
 296          *  ->stop() that must deal with already being stopped without
 297          *  PERF_EF_UPDATE.
 298          */
 299         int  (*add)                     (struct perf_event *event, int flags);
 300         void (*del)                     (struct perf_event *event, int flags);
 301
 302         /*
 303          * Starts/Stops a counter present on the PMU.
 304          *
 305          * The PMI handler should stop the counter when perf_event_overflow()
 306          * returns !0. ->start() will be used to continue.
 307          *
 308          * Also used to change the sample period.
 309          *
 310          * Called with IRQs disabled and the PMU disabled on the CPU the event
 311          * is on -- will be called from NMI context with the PMU generates
 312          * NMIs.
 313          *
 314          * ->stop() with PERF_EF_UPDATE will read the counter and update
 315          *  period/count values like ->read() would.
 316          *
 317          * ->start() with PERF_EF_RELOAD will reprogram the the counter
 318          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
 319          */
 320         void (*start)                   (struct perf_event *event, int flags);
 321         void (*stop)                    (struct perf_event *event, int flags);
 322
 323         /*
 324          * Updates the counter value of the event.
 325          *
 326          * For sampling capable PMUs this will also update the software period
 327          * hw_perf_event::period_left field.
 328          */
 329         void (*read)                    (struct perf_event *event);
 330
 331         /*
 332          * Group events scheduling is treated as a transaction, add
 333          * group events as a whole and perform one schedulability test.
 334          * If the test fails, roll back the whole group
 335          *
 336          * Start the transaction, after this ->add() doesn't need to
 337          * do schedulability tests.
 338          *
 339          * Optional.
 340          */
 341         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
 342         /*
 343          * If ->start_txn() disabled the ->add() schedulability test
 344          * then ->commit_txn() is required to perform one. On success
 345          * the transaction is closed. On error the transaction is kept
 346          * open until ->cancel_txn() is called.
 347          *
 348          * Optional.
 349          */
 350         int  (*commit_txn)              (struct pmu *pmu);
 351         /*
 352          * Will cancel the transaction, assumes ->del() is called
 353          * for each successful ->add() during the transaction.
 354          *
 355          * Optional.
 356          */
 357         void (*cancel_txn)              (struct pmu *pmu);
 358
 359         /*
 360          * Will return the value for perf_event_mmap_page::index for this event,
 361          * if no implementation is provided it will default to: event->hw.idx + 1.
 362          */
 363         int (*event_idx)                (struct perf_event *event); /*optional */
 364
 365         /*
 366          * context-switches callback
 367          */
 368         void (*sched_task)              (struct perf_event_context *ctx,
 369                                         bool sched_in);
 370         /*
 371          * PMU specific data size
 372          */
 373         size_t                          task_ctx_size;
 374
 375
 376         /*
 377          * Return the count value for a counter.
 378          */
 379         u64 (*count)                    (struct perf_event *event); /*optional*/
 380
 381         /*
 382          * Set up pmu-private data structures for an AUX area
 383          */
 384         void *(*setup_aux)              (struct perf_event *event, void **pages,
 385                                          int nr_pages, bool overwrite);
 386                                         /* optional */
 387
 388         /*
 389          * Free pmu-private AUX data structures
 390          */
 391         void (*free_aux)                (void *aux); /* optional */
 392
 393         /*
 394          * Filter events for PMU-specific reasons.
 395          */
 396         int (*filter_match)             (struct perf_event *event); /* optional */
 397
 398         /*
 399          * Initial, PMU driver specific configuration.
 400          */
 401         int (*get_drv_configs)          (struct perf_event *event,
 402                                          void __user *arg); /* optional */
 403         void (*free_drv_configs)        (struct perf_event *event);
 404                                         /* optional */
 405 };
 406
 407 /**
 408  * enum perf_event_active_state - the states of a event
 409  */
 410 enum perf_event_active_state {
 411         PERF_EVENT_STATE_EXIT           = -3,
 412         PERF_EVENT_STATE_ERROR          = -2,
 413         PERF_EVENT_STATE_OFF            = -1,
 414         PERF_EVENT_STATE_INACTIVE       =  0,
 415         PERF_EVENT_STATE_ACTIVE         =  1,
 416 };
 417
 418 struct file;
 419 struct perf_sample_data;
 420
 421 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 422                                         struct perf_sample_data *,
 423                                         struct pt_regs *regs);
 424
 425 enum perf_group_flag {
 426         PERF_GROUP_SOFTWARE             = 0x1,
 427 };
 428
 429 #define SWEVENT_HLIST_BITS              8
 430 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 431
 432 struct swevent_hlist {
 433         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 434         struct rcu_head                 rcu_head;
 435 };
 436
 437 #define PERF_ATTACH_CONTEXT     0x01
 438 #define PERF_ATTACH_GROUP       0x02
 439 #define PERF_ATTACH_TASK        0x04
 440 #define PERF_ATTACH_TASK_DATA   0x08
 441
 442 struct perf_cgroup;
 443 struct ring_buffer;
 444
 445 /**
 446  * struct perf_event - performance event kernel representation:
 447  */
 448 struct perf_event {
 449 #ifdef CONFIG_PERF_EVENTS
 450         /*
 451          * entry onto perf_event_context::event_list;
 452          *   modifications require ctx->lock
 453          *   RCU safe iterations.
 454          */
 455         struct list_head                event_entry;
 456
 457         /*
 458          * XXX: group_entry and sibling_list should be mutually exclusive;
 459          * either you're a sibling on a group, or you're the group leader.
 460          * Rework the code to always use the same list element.
 461          *
 462          * Locked for modification by both ctx->mutex and ctx->lock; holding
 463          * either sufficies for read.
 464          */
 465         struct list_head                group_entry;
 466         struct list_head                sibling_list;
 467
 468         /*
 469          * We need storage to track the entries in perf_pmu_migrate_context; we
 470          * cannot use the event_entry because of RCU and we want to keep the
 471          * group in tact which avoids us using the other two entries.
 472          */
 473         struct list_head                migrate_entry;
 474
 475         struct hlist_node               hlist_entry;
 476         struct list_head                active_entry;
 477         int                             nr_siblings;
 478         int                             group_flags;
 479         struct perf_event               *group_leader;
 480
 481         /*
 482          * Protect the pmu, attributes and context of a group leader.
 483          * Note: does not protect the pointer to the group_leader.
 484          */
 485         struct mutex                    group_leader_mutex;
 486         struct pmu                      *pmu;
 487
 488         enum perf_event_active_state    state;
 489         unsigned int                    attach_state;
 490         local64_t                       count;
 491         atomic64_t                      child_count;
 492
 493         /*
 494          * These are the total time in nanoseconds that the event
 495          * has been enabled (i.e. eligible to run, and the task has
 496          * been scheduled in, if this is a per-task event)
 497          * and running (scheduled onto the CPU), respectively.
 498          *
 499          * They are computed from tstamp_enabled, tstamp_running and
 500          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 501          */
 502         u64                             total_time_enabled;
 503         u64                             total_time_running;
 504
 505         /*
 506          * These are timestamps used for computing total_time_enabled
 507          * and total_time_running when the event is in INACTIVE or
 508          * ACTIVE state, measured in nanoseconds from an arbitrary point
 509          * in time.
 510          * tstamp_enabled: the notional time when the event was enabled
 511          * tstamp_running: the notional time when the event was scheduled on
 512          * tstamp_stopped: in INACTIVE state, the notional time when the
 513          *      event was scheduled off.
 514          */
 515         u64                             tstamp_enabled;
 516         u64                             tstamp_running;
 517         u64                             tstamp_stopped;
 518
 519         /*
 520          * timestamp shadows the actual context timing but it can
 521          * be safely used in NMI interrupt context. It reflects the
 522          * context time as it was when the event was last scheduled in.
 523          *
 524          * ctx_time already accounts for ctx->timestamp. Therefore to
 525          * compute ctx_time for a sample, simply add perf_clock().
 526          */
 527         u64                             shadow_ctx_time;
 528
 529         struct perf_event_attr          attr;
 530         u16                             header_size;
 531         u16                             id_header_size;
 532         u16                             read_size;
 533         struct hw_perf_event            hw;
 534
 535         struct perf_event_context       *ctx;
 536         atomic_long_t                   refcount;
 537
 538         /*
 539          * These accumulate total time (in nanoseconds) that children
 540          * events have been enabled and running, respectively.
 541          */
 542         atomic64_t                      child_total_time_enabled;
 543         atomic64_t                      child_total_time_running;
 544
 545         /*
 546          * Protect attach/detach and child_list:
 547          */
 548         struct mutex                    child_mutex;
 549         struct list_head                child_list;
 550         struct perf_event               *parent;
 551
 552         int                             oncpu;
 553         int                             cpu;
 554
 555         struct list_head                owner_entry;
 556         struct task_struct              *owner;
 557
 558         /* mmap bits */
 559         struct mutex                    mmap_mutex;
 560         atomic_t                        mmap_count;
 561
 562         struct ring_buffer              *rb;
 563         struct list_head                rb_entry;
 564         unsigned long                   rcu_batches;
 565         int                             rcu_pending;
 566
 567         /* poll related */
 568         wait_queue_head_t               waitq;
 569         struct fasync_struct            *fasync;
 570
 571         /* delayed work for NMIs and such */
 572         int                             pending_wakeup;
 573         int                             pending_kill;
 574         int                             pending_disable;
 575         struct irq_work                 pending;
 576
 577         atomic_t                        event_limit;
 578         struct list_head                drv_configs;
 579
 580         void (*destroy)(struct perf_event *);
 581         struct rcu_head                 rcu_head;
 582
 583         struct pid_namespace            *ns;
 584         u64                             id;
 585
 586         u64                             (*clock)(void);
 587         perf_overflow_handler_t         overflow_handler;
 588         void                            *overflow_handler_context;
 589
 590 #ifdef CONFIG_EVENT_TRACING
 591         struct trace_event_call         *tp_event;
 592         struct event_filter             *filter;
 593 #ifdef CONFIG_FUNCTION_TRACER
 594         struct ftrace_ops               ftrace_ops;
 595 #endif
 596 #endif
 597
 598 #ifdef CONFIG_CGROUP_PERF
 599         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 600         int                             cgrp_defer_enabled;
 601 #endif
 602
 603 #endif /* CONFIG_PERF_EVENTS */
 604 };
 605
 606 /**
 607  * struct perf_event_context - event context structure
 608  *
 609  * Used as a container for task events and CPU events as well:
 610  */
 611 struct perf_event_context {
 612         struct pmu                      *pmu;
 613         /*
 614          * Protect the states of the events in the list,
 615          * nr_active, and the list:
 616          */
 617         raw_spinlock_t                  lock;
 618         /*
 619          * Protect the list of events.  Locking either mutex or lock
 620          * is sufficient to ensure the list doesn't change; to change
 621          * the list you need to lock both the mutex and the spinlock.
 622          */
 623         struct mutex                    mutex;
 624
 625         struct list_head                active_ctx_list;
 626         struct list_head                pinned_groups;
 627         struct list_head                flexible_groups;
 628         struct list_head                event_list;
 629         int                             nr_events;
 630         int                             nr_active;
 631         int                             is_active;
 632         int                             nr_stat;
 633         int                             nr_freq;
 634         int                             rotate_disable;
 635         atomic_t                        refcount;
 636         struct task_struct              *task;
 637
 638         /*
 639          * Context clock, runs when context enabled.
 640          */
 641         u64                             time;
 642         u64                             timestamp;
 643
 644         /*
 645          * These fields let us detect when two contexts have both
 646          * been cloned (inherited) from a common ancestor.
 647          */
 648         struct perf_event_context       *parent_ctx;
 649         u64                             parent_gen;
 650         u64                             generation;
 651         int                             pin_count;
 652         int                             nr_cgroups;      /* cgroup evts */
 653         void                            *task_ctx_data; /* pmu specific data */
 654         struct rcu_head                 rcu_head;
 655
 656         struct delayed_work             orphans_remove;
 657         bool                            orphans_remove_sched;
 658 };
 659
 660 /*
 661  * Number of contexts where an event can trigger:
 662  *      task, softirq, hardirq, nmi.
 663  */
 664 #define PERF_NR_CONTEXTS        4
 665
 666 /**
 667  * struct perf_event_cpu_context - per cpu event context structure
 668  */
 669 struct perf_cpu_context {
 670         struct perf_event_context       ctx;
 671         struct perf_event_context       *task_ctx;
 672         int                             active_oncpu;
 673         int                             exclusive;
 674
 675         raw_spinlock_t                  hrtimer_lock;
 676         struct hrtimer                  hrtimer;
 677         ktime_t                         hrtimer_interval;
 678         unsigned int                    hrtimer_active;
 679
 680         struct pmu                      *unique_pmu;
 681         struct perf_cgroup              *cgrp;
 682 };
 683
 684 struct perf_output_handle {
 685         struct perf_event               *event;
 686         struct ring_buffer              *rb;
 687         unsigned long                   wakeup;
 688         unsigned long                   size;
 689         union {
 690                 void                    *addr;
 691                 unsigned long           head;
 692         };
 693         int                             page;
 694 };
 695
 696 #ifdef CONFIG_CGROUP_PERF
 697
 698 /*
 699  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 700  * This is a per-cpu dynamically allocated data structure.
 701  */
 702 struct perf_cgroup_info {
 703         u64                             time;
 704         u64                             timestamp;
 705 };
 706
 707 struct perf_cgroup {
 708         struct cgroup_subsys_state      css;
 709         struct perf_cgroup_info __percpu *info;
 710 };
 711
 712 /*
 713  * Must ensure cgroup is pinned (css_get) before calling
 714  * this function. In other words, we cannot call this function
 715  * if there is no cgroup event for the current CPU context.
 716  */
 717 static inline struct perf_cgroup *
 718 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
 719 {
 720         return container_of(task_css_check(task, perf_event_cgrp_id,
 721                                            ctx ? lockdep_is_held(&ctx->lock)
 722                                                : true),
 723                             struct perf_cgroup, css);
 724 }
 725 #endif /* CONFIG_CGROUP_PERF */
 726
 727 #ifdef CONFIG_PERF_EVENTS
 728
 729 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 730                                    struct perf_event *event);
 731 extern void perf_aux_output_end(struct perf_output_handle *handle,
 732                                 unsigned long size, bool truncated);
 733 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 734                                 unsigned long size);
 735 extern void *perf_get_aux(struct perf_output_handle *handle);
 736
 737 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 738 extern void perf_pmu_unregister(struct pmu *pmu);
 739
 740 extern int perf_num_counters(void);
 741 extern const char *perf_pmu_name(void);
 742 extern void __perf_event_task_sched_in(struct task_struct *prev,
 743                                        struct task_struct *task);
 744 extern void __perf_event_task_sched_out(struct task_struct *prev,
 745                                         struct task_struct *next);
 746 extern int perf_event_init_task(struct task_struct *child);
 747 extern void perf_event_exit_task(struct task_struct *child);
 748 extern void perf_event_free_task(struct task_struct *task);
 749 extern void perf_event_delayed_put(struct task_struct *task);
 750 extern struct perf_event *perf_event_get(unsigned int fd);
 751 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
 752 extern void perf_event_print_debug(void);
 753 extern void perf_pmu_disable(struct pmu *pmu);
 754 extern void perf_pmu_enable(struct pmu *pmu);
 755 extern void perf_sched_cb_dec(struct pmu *pmu);
 756 extern void perf_sched_cb_inc(struct pmu *pmu);
 757 extern int perf_event_task_disable(void);
 758 extern int perf_event_task_enable(void);
 759 extern int perf_event_refresh(struct perf_event *event, int refresh);
 760 extern void perf_event_update_userpage(struct perf_event *event);
 761 extern int perf_event_release_kernel(struct perf_event *event);
 762 extern struct perf_event *
 763 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 764                                 int cpu,
 765                                 struct task_struct *task,
 766                                 perf_overflow_handler_t callback,
 767                                 void *context);
 768 extern void perf_pmu_migrate_context(struct pmu *pmu,
 769                                 int src_cpu, int dst_cpu);
 770 extern u64 perf_event_read_local(struct perf_event *event);
 771 extern u64 perf_event_read_value(struct perf_event *event,
 772                                  u64 *enabled, u64 *running);
 773
 774 extern struct dentry *perf_create_debug_dir(void);
 775
 776 struct perf_sample_data {
 777         /*
 778          * Fields set by perf_sample_data_init(), group so as to
 779          * minimize the cachelines touched.
 780          */
 781         u64                             addr;
 782         struct perf_raw_record          *raw;
 783         struct perf_branch_stack        *br_stack;
 784         u64                             period;
 785         u64                             weight;
 786         u64                             txn;
 787         union  perf_mem_data_src        data_src;
 788
 789         /*
 790          * The other fields, optionally {set,used} by
 791          * perf_{prepare,output}_sample().
 792          */
 793         u64                             type;
 794         u64                             ip;
 795         struct {
 796                 u32     pid;
 797                 u32     tid;
 798         }                               tid_entry;
 799         u64                             time;
 800         u64                             id;
 801         u64                             stream_id;
 802         struct {
 803                 u32     cpu;
 804                 u32     reserved;
 805         }                               cpu_entry;
 806         struct perf_callchain_entry     *callchain;
 807
 808         /*
 809          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 810          * on arch details.
 811          */
 812         struct perf_regs                regs_user;
 813         struct pt_regs                  regs_user_copy;
 814
 815         struct perf_regs                regs_intr;
 816         u64                             stack_user_size;
 817 } ____cacheline_aligned;
 818
 819 /* default value for data source */
 820 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 821                     PERF_MEM_S(LVL, NA)   |\
 822                     PERF_MEM_S(SNOOP, NA) |\
 823                     PERF_MEM_S(LOCK, NA)  |\
 824                     PERF_MEM_S(TLB, NA))
 825
 826 static inline void perf_sample_data_init(struct perf_sample_data *data,
 827                                          u64 addr, u64 period)
 828 {
 829         /* remaining struct members initialized in perf_prepare_sample() */
 830         data->addr = addr;
 831         data->raw  = NULL;
 832         data->br_stack = NULL;
 833         data->period = period;
 834         data->weight = 0;
 835         data->data_src.val = PERF_MEM_NA;
 836         data->txn = 0;
 837 }
 838
 839 extern void perf_output_sample(struct perf_output_handle *handle,
 840                                struct perf_event_header *header,
 841                                struct perf_sample_data *data,
 842                                struct perf_event *event);
 843 extern void perf_prepare_sample(struct perf_event_header *header,
 844                                 struct perf_sample_data *data,
 845                                 struct perf_event *event,
 846                                 struct pt_regs *regs);
 847
 848 extern int perf_event_overflow(struct perf_event *event,
 849                                  struct perf_sample_data *data,
 850                                  struct pt_regs *regs);
 851
 852 extern void perf_event_output(struct perf_event *event,
 853                                 struct perf_sample_data *data,
 854                                 struct pt_regs *regs);
 855
 856 extern void
 857 perf_event_header__init_id(struct perf_event_header *header,
 858                            struct perf_sample_data *data,
 859                            struct perf_event *event);
 860 extern void
 861 perf_event__output_id_sample(struct perf_event *event,
 862                              struct perf_output_handle *handle,
 863                              struct perf_sample_data *sample);
 864
 865 extern void
 866 perf_log_lost_samples(struct perf_event *event, u64 lost);
 867
 868 static inline bool is_sampling_event(struct perf_event *event)
 869 {
 870         return event->attr.sample_period != 0;
 871 }
 872
 873 /*
 874  * Return 1 for a software event, 0 for a hardware event
 875  */
 876 static inline int is_software_event(struct perf_event *event)
 877 {
 878         return event->pmu->task_ctx_nr == perf_sw_context;
 879 }
 880
 881 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 882
 883 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 884 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 885
 886 #ifndef perf_arch_fetch_caller_regs
 887 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 888 #endif
 889
 890 /*
 891  * Take a snapshot of the regs. Skip ip and frame pointer to
 892  * the nth caller. We only need a few of the regs:
 893  * - ip for PERF_SAMPLE_IP
 894  * - cs for user_mode() tests
 895  * - bp for callchains
 896  * - eflags, for future purposes, just in case
 897  */
 898 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 899 {
 900         memset(regs, 0, sizeof(*regs));
 901
 902         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 903 }
 904
 905 static __always_inline void
 906 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 907 {
 908         if (static_key_false(&perf_swevent_enabled[event_id]))
 909                 __perf_sw_event(event_id, nr, regs, addr);
 910 }
 911
 912 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 913
 914 /*
 915  * 'Special' version for the scheduler, it hard assumes no recursion,
 916  * which is guaranteed by us not actually scheduling inside other swevents
 917  * because those disable preemption.
 918  */
 919 static __always_inline void
 920 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 921 {
 922         if (static_key_false(&perf_swevent_enabled[event_id])) {
 923                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 924
 925                 perf_fetch_caller_regs(regs);
 926                 ___perf_sw_event(event_id, nr, regs, addr);
 927         }
 928 }
 929
 930 extern struct static_key_deferred perf_sched_events;
 931
 932 static __always_inline bool
 933 perf_sw_migrate_enabled(void)
 934 {
 935         if (static_key_false(&perf_swevent_enabled[PERF_COUNT_SW_CPU_MIGRATIONS]))
 936                 return true;
 937         return false;
 938 }
 939
 940 static inline void perf_event_task_migrate(struct task_struct *task)
 941 {
 942         if (perf_sw_migrate_enabled())
 943                 task->sched_migrated = 1;
 944 }
 945
 946 static inline void perf_event_task_sched_in(struct task_struct *prev,
 947                                             struct task_struct *task)
 948 {
 949         if (static_key_false(&perf_sched_events.key))
 950                 __perf_event_task_sched_in(prev, task);
 951
 952         if (perf_sw_migrate_enabled() && task->sched_migrated) {
 953                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 954
 955                 perf_fetch_caller_regs(regs);
 956                 ___perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, regs, 0);
 957                 task->sched_migrated = 0;
 958         }
 959 }
 960
 961 static inline void perf_event_task_sched_out(struct task_struct *prev,
 962                                              struct task_struct *next)
 963 {
 964         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 965
 966         if (static_key_false(&perf_sched_events.key))
 967                 __perf_event_task_sched_out(prev, next);
 968 }
 969
 970 static inline u64 __perf_event_count(struct perf_event *event)
 971 {
 972         return local64_read(&event->count) + atomic64_read(&event->child_count);
 973 }
 974
 975 extern void perf_event_mmap(struct vm_area_struct *vma);
 976 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 977 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 978 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 979
 980 extern void perf_event_exec(void);
 981 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 982 extern void perf_event_fork(struct task_struct *tsk);
 983
 984 /* Callchains */
 985 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 986
 987 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 988 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 989
 990 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 991 {
 992         if (entry->nr < PERF_MAX_STACK_DEPTH)
 993                 entry->ip[entry->nr++] = ip;
 994 }
 995
 996 extern int sysctl_perf_event_paranoid;
 997 extern int sysctl_perf_event_mlock;
 998 extern int sysctl_perf_event_sample_rate;
 999 extern int sysctl_perf_cpu_time_max_percent;
1000
1001 extern void perf_sample_event_took(u64 sample_len_ns);
1002
1003 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1004                 void __user *buffer, size_t *lenp,
1005                 loff_t *ppos);
1006 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1007                 void __user *buffer, size_t *lenp,
1008                 loff_t *ppos);
1009
1010
1011 static inline bool perf_paranoid_any(void)
1012 {
1013         return sysctl_perf_event_paranoid > 2;
1014 }
1015
1016 static inline bool perf_paranoid_tracepoint_raw(void)
1017 {
1018         return sysctl_perf_event_paranoid > -1;
1019 }
1020
1021 static inline bool perf_paranoid_cpu(void)
1022 {
1023         return sysctl_perf_event_paranoid > 0;
1024 }
1025
1026 static inline bool perf_paranoid_kernel(void)
1027 {
1028         return sysctl_perf_event_paranoid > 1;
1029 }
1030
1031 extern void perf_event_init(void);
1032 extern void perf_tp_event(u64 addr, u64 count, void *record,
1033                           int entry_size, struct pt_regs *regs,
1034                           struct hlist_head *head, int rctx,
1035                           struct task_struct *task);
1036 extern void perf_bp_event(struct perf_event *event, void *data);
1037
1038 #ifndef perf_misc_flags
1039 # define perf_misc_flags(regs) \
1040                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1041 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1042 #endif
1043
1044 static inline bool has_branch_stack(struct perf_event *event)
1045 {
1046         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1047 }
1048
1049 static inline bool needs_branch_stack(struct perf_event *event)
1050 {
1051         return event->attr.branch_sample_type != 0;
1052 }
1053
1054 static inline bool has_aux(struct perf_event *event)
1055 {
1056         return event->pmu->setup_aux;
1057 }
1058
1059 extern int perf_output_begin(struct perf_output_handle *handle,
1060                              struct perf_event *event, unsigned int size);
1061 extern void perf_output_end(struct perf_output_handle *handle);
1062 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1063                              const void *buf, unsigned int len);
1064 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1065                                      unsigned int len);
1066 extern int perf_swevent_get_recursion_context(void);
1067 extern void perf_swevent_put_recursion_context(int rctx);
1068 extern u64 perf_swevent_set_period(struct perf_event *event);
1069 extern void perf_event_enable(struct perf_event *event);
1070 extern void perf_event_disable(struct perf_event *event);
1071 extern int __perf_event_disable(void *info);
1072 extern void perf_event_task_tick(void);
1073 #else /* !CONFIG_PERF_EVENTS: */
1074 static inline void *
1075 perf_aux_output_begin(struct perf_output_handle *handle,
1076                       struct perf_event *event)                         { return NULL; }
1077 static inline void
1078 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
1079                     bool truncated)                                     { }
1080 static inline int
1081 perf_aux_output_skip(struct perf_output_handle *handle,
1082                      unsigned long size)                                { return -EINVAL; }
1083 static inline void *
1084 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1085 static inline void
1086 perf_event_task_migrate(struct task_struct *task)                       { }
1087 static inline void
1088 perf_event_task_sched_in(struct task_struct *prev,
1089                          struct task_struct *task)                      { }
1090 static inline void
1091 perf_event_task_sched_out(struct task_struct *prev,
1092                           struct task_struct *next)                     { }
1093 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1094 static inline void perf_event_exit_task(struct task_struct *child)      { }
1095 static inline void perf_event_free_task(struct task_struct *task)       { }
1096 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1097 static inline struct perf_event *perf_event_get(unsigned int fd)        { return ERR_PTR(-EINVAL); }
1098 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1099 {
1100         return ERR_PTR(-EINVAL);
1101 }
1102 static inline u64 perf_event_read_local(struct perf_event *event)       { return -EINVAL; }
1103 static inline void perf_event_print_debug(void)                         { }
1104 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1105 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1106 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1107 {
1108         return -EINVAL;
1109 }
1110
1111 static inline void
1112 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1113 static inline void
1114 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
1115 static inline void
1116 perf_bp_event(struct perf_event *event, void *data)                     { }
1117
1118 static inline int perf_register_guest_info_callbacks
1119 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1120 static inline int perf_unregister_guest_info_callbacks
1121 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1122
1123 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1124 static inline void perf_event_exec(void)                                { }
1125 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1126 static inline void perf_event_fork(struct task_struct *tsk)             { }
1127 static inline void perf_event_init(void)                                { }
1128 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1129 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1130 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1131 static inline void perf_event_enable(struct perf_event *event)          { }
1132 static inline void perf_event_disable(struct perf_event *event)         { }
1133 static inline int __perf_event_disable(void *info)                      { return -1; }
1134 static inline void perf_event_task_tick(void)                           { }
1135 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1136 #endif
1137
1138 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
1139 extern bool perf_event_can_stop_tick(void);
1140 #else
1141 static inline bool perf_event_can_stop_tick(void)                       { return true; }
1142 #endif
1143
1144 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1145 extern void perf_restore_debug_store(void);
1146 #else
1147 static inline void perf_restore_debug_store(void)                       { }
1148 #endif
1149
1150 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1151
1152 /*
1153  * This has to have a higher priority than migration_notifier in sched/core.c.
1154  */
1155 #define perf_cpu_notifier(fn)                                           \
1156 do {                                                                    \
1157         static struct notifier_block fn##_nb =                          \
1158                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1159         unsigned long cpu = smp_processor_id();                         \
1160         unsigned long flags;                                            \
1161                                                                         \
1162         cpu_notifier_register_begin();                                  \
1163         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1164                 (void *)(unsigned long)cpu);                            \
1165         local_irq_save(flags);                                          \
1166         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1167                 (void *)(unsigned long)cpu);                            \
1168         local_irq_restore(flags);                                       \
1169         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1170                 (void *)(unsigned long)cpu);                            \
1171         __register_cpu_notifier(&fn##_nb);                              \
1172         cpu_notifier_register_done();                                   \
1173 } while (0)
1174
1175 /*
1176  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1177  * callback for already online CPUs.
1178  */
1179 #define __perf_cpu_notifier(fn)                                         \
1180 do {                                                                    \
1181         static struct notifier_block fn##_nb =                          \
1182                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1183                                                                         \
1184         __register_cpu_notifier(&fn##_nb);                              \
1185 } while (0)
1186
1187 struct perf_pmu_events_attr {
1188         struct device_attribute attr;
1189         u64 id;
1190         const char *event_str;
1191 };
1192
1193 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1194                               char *page);
1195
1196 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1197 static struct perf_pmu_events_attr _var = {                             \
1198         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1199         .id   =  _id,                                                   \
1200 };
1201
1202 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1203 static struct perf_pmu_events_attr _var = {                                 \
1204         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1205         .id             = 0,                                                \
1206         .event_str      = _str,                                             \
1207 };
1208
1209 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1210 static ssize_t                                                          \
1211 _name##_show(struct device *dev,                                        \
1212                                struct device_attribute *attr,           \
1213                                char *page)                              \
1214 {                                                                       \
1215         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1216         return sprintf(page, _format "\n");                             \
1217 }                                                                       \
1218                                                                         \
1219 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1220
1221 #endif /* _LINUX_PERF_EVENT_H */