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 struct event_constraint;
  96
  97 /**
  98  * struct hw_perf_event - performance event hardware details:
  99  */
 100 struct hw_perf_event {
 101 #ifdef CONFIG_PERF_EVENTS
 102         union {
 103                 struct { /* hardware */
 104                         u64             config;
 105                         u64             last_tag;
 106                         unsigned long   config_base;
 107                         unsigned long   event_base;
 108                         int             event_base_rdpmc;
 109                         int             idx;
 110                         int             last_cpu;
 111                         int             flags;
 112
 113                         struct hw_perf_event_extra extra_reg;
 114                         struct hw_perf_event_extra branch_reg;
 115
 116                         struct event_constraint *constraint;
 117                 };
 118                 struct { /* software */
 119                         struct hrtimer  hrtimer;
 120                 };
 121                 struct { /* tracepoint */
 122                         struct task_struct      *tp_target;
 123                         /* for tp_event->class */
 124                         struct list_head        tp_list;
 125                 };
 126 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 127                 struct { /* breakpoint */
 128                         /*
 129                          * Crufty hack to avoid the chicken and egg
 130                          * problem hw_breakpoint has with context
 131                          * creation and event initalization.
 132                          */
 133                         struct task_struct              *bp_target;
 134                         struct arch_hw_breakpoint       info;
 135                         struct list_head                bp_list;
 136                 };
 137 #endif
 138         };
 139         int                             state;
 140         local64_t                       prev_count;
 141         u64                             sample_period;
 142         u64                             last_period;
 143         local64_t                       period_left;
 144         u64                             interrupts_seq;
 145         u64                             interrupts;
 146
 147         u64                             freq_time_stamp;
 148         u64                             freq_count_stamp;
 149 #endif
 150 };
 151
 152 /*
 153  * hw_perf_event::state flags
 154  */
 155 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 156 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 157 #define PERF_HES_ARCH           0x04
 158
 159 struct perf_event;
 160
 161 /*
 162  * Common implementation detail of pmu::{start,commit,cancel}_txn
 163  */
 164 #define PERF_EVENT_TXN 0x1
 165
 166 /**
 167  * pmu::capabilities flags
 168  */
 169 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 170
 171 /**
 172  * struct pmu - generic performance monitoring unit
 173  */
 174 struct pmu {
 175         struct list_head                entry;
 176
 177         struct module                   *module;
 178         struct device                   *dev;
 179         const struct attribute_group    **attr_groups;
 180         const char                      *name;
 181         int                             type;
 182
 183         /*
 184          * various common per-pmu feature flags
 185          */
 186         int                             capabilities;
 187
 188         int * __percpu                  pmu_disable_count;
 189         struct perf_cpu_context * __percpu pmu_cpu_context;
 190         int                             task_ctx_nr;
 191         int                             hrtimer_interval_ms;
 192
 193         /*
 194          * Fully disable/enable this PMU, can be used to protect from the PMI
 195          * as well as for lazy/batch writing of the MSRs.
 196          */
 197         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 198         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 199
 200         /*
 201          * Try and initialize the event for this PMU.
 202          * Should return -ENOENT when the @event doesn't match this PMU.
 203          */
 204         int (*event_init)               (struct perf_event *event);
 205
 206         /*
 207          * Notification that the event was mapped or unmapped.  Called
 208          * in the context of the mapping task.
 209          */
 210         void (*event_mapped)            (struct perf_event *event); /*optional*/
 211         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 212
 213 #define PERF_EF_START   0x01            /* start the counter when adding    */
 214 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 215 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 216
 217         /*
 218          * Adds/Removes a counter to/from the PMU, can be done inside
 219          * a transaction, see the ->*_txn() methods.
 220          */
 221         int  (*add)                     (struct perf_event *event, int flags);
 222         void (*del)                     (struct perf_event *event, int flags);
 223
 224         /*
 225          * Starts/Stops a counter present on the PMU. The PMI handler
 226          * should stop the counter when perf_event_overflow() returns
 227          * !0. ->start() will be used to continue.
 228          */
 229         void (*start)                   (struct perf_event *event, int flags);
 230         void (*stop)                    (struct perf_event *event, int flags);
 231
 232         /*
 233          * Updates the counter value of the event.
 234          */
 235         void (*read)                    (struct perf_event *event);
 236
 237         /*
 238          * Group events scheduling is treated as a transaction, add
 239          * group events as a whole and perform one schedulability test.
 240          * If the test fails, roll back the whole group
 241          *
 242          * Start the transaction, after this ->add() doesn't need to
 243          * do schedulability tests.
 244          */
 245         void (*start_txn)               (struct pmu *pmu); /* optional */
 246         /*
 247          * If ->start_txn() disabled the ->add() schedulability test
 248          * then ->commit_txn() is required to perform one. On success
 249          * the transaction is closed. On error the transaction is kept
 250          * open until ->cancel_txn() is called.
 251          */
 252         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 253         /*
 254          * Will cancel the transaction, assumes ->del() is called
 255          * for each successful ->add() during the transaction.
 256          */
 257         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 258
 259         /*
 260          * Will return the value for perf_event_mmap_page::index for this event,
 261          * if no implementation is provided it will default to: event->hw.idx + 1.
 262          */
 263         int (*event_idx)                (struct perf_event *event); /*optional */
 264
 265         /*
 266          * context-switches callback
 267          */
 268         void (*sched_task)              (struct perf_event_context *ctx,
 269                                         bool sched_in);
 270         /*
 271          * PMU specific data size
 272          */
 273         size_t                          task_ctx_size;
 274
 275
 276         /*
 277          * Return the count value for a counter.
 278          */
 279         u64 (*count)                    (struct perf_event *event); /*optional*/
 280 };
 281
 282 /**
 283  * enum perf_event_active_state - the states of a event
 284  */
 285 enum perf_event_active_state {
 286         PERF_EVENT_STATE_EXIT           = -3,
 287         PERF_EVENT_STATE_ERROR          = -2,
 288         PERF_EVENT_STATE_OFF            = -1,
 289         PERF_EVENT_STATE_INACTIVE       =  0,
 290         PERF_EVENT_STATE_ACTIVE         =  1,
 291 };
 292
 293 struct file;
 294 struct perf_sample_data;
 295
 296 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 297                                         struct perf_sample_data *,
 298                                         struct pt_regs *regs);
 299
 300 enum perf_group_flag {
 301         PERF_GROUP_SOFTWARE             = 0x1,
 302 };
 303
 304 #define SWEVENT_HLIST_BITS              8
 305 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 306
 307 struct swevent_hlist {
 308         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 309         struct rcu_head                 rcu_head;
 310 };
 311
 312 #define PERF_ATTACH_CONTEXT     0x01
 313 #define PERF_ATTACH_GROUP       0x02
 314 #define PERF_ATTACH_TASK        0x04
 315 #define PERF_ATTACH_TASK_DATA   0x08
 316
 317 struct perf_cgroup;
 318 struct ring_buffer;
 319
 320 /**
 321  * struct perf_event - performance event kernel representation:
 322  */
 323 struct perf_event {
 324 #ifdef CONFIG_PERF_EVENTS
 325         /*
 326          * entry onto perf_event_context::event_list;
 327          *   modifications require ctx->lock
 328          *   RCU safe iterations.
 329          */
 330         struct list_head                event_entry;
 331
 332         /*
 333          * XXX: group_entry and sibling_list should be mutually exclusive;
 334          * either you're a sibling on a group, or you're the group leader.
 335          * Rework the code to always use the same list element.
 336          *
 337          * Locked for modification by both ctx->mutex and ctx->lock; holding
 338          * either sufficies for read.
 339          */
 340         struct list_head                group_entry;
 341         struct list_head                sibling_list;
 342
 343         /*
 344          * We need storage to track the entries in perf_pmu_migrate_context; we
 345          * cannot use the event_entry because of RCU and we want to keep the
 346          * group in tact which avoids us using the other two entries.
 347          */
 348         struct list_head                migrate_entry;
 349
 350         struct hlist_node               hlist_entry;
 351         struct list_head                active_entry;
 352         int                             nr_siblings;
 353         int                             group_flags;
 354         struct perf_event               *group_leader;
 355         struct pmu                      *pmu;
 356
 357         enum perf_event_active_state    state;
 358         unsigned int                    attach_state;
 359         local64_t                       count;
 360         atomic64_t                      child_count;
 361
 362         /*
 363          * These are the total time in nanoseconds that the event
 364          * has been enabled (i.e. eligible to run, and the task has
 365          * been scheduled in, if this is a per-task event)
 366          * and running (scheduled onto the CPU), respectively.
 367          *
 368          * They are computed from tstamp_enabled, tstamp_running and
 369          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 370          */
 371         u64                             total_time_enabled;
 372         u64                             total_time_running;
 373
 374         /*
 375          * These are timestamps used for computing total_time_enabled
 376          * and total_time_running when the event is in INACTIVE or
 377          * ACTIVE state, measured in nanoseconds from an arbitrary point
 378          * in time.
 379          * tstamp_enabled: the notional time when the event was enabled
 380          * tstamp_running: the notional time when the event was scheduled on
 381          * tstamp_stopped: in INACTIVE state, the notional time when the
 382          *      event was scheduled off.
 383          */
 384         u64                             tstamp_enabled;
 385         u64                             tstamp_running;
 386         u64                             tstamp_stopped;
 387
 388         /*
 389          * timestamp shadows the actual context timing but it can
 390          * be safely used in NMI interrupt context. It reflects the
 391          * context time as it was when the event was last scheduled in.
 392          *
 393          * ctx_time already accounts for ctx->timestamp. Therefore to
 394          * compute ctx_time for a sample, simply add perf_clock().
 395          */
 396         u64                             shadow_ctx_time;
 397
 398         struct perf_event_attr          attr;
 399         u16                             header_size;
 400         u16                             id_header_size;
 401         u16                             read_size;
 402         struct hw_perf_event            hw;
 403
 404         struct perf_event_context       *ctx;
 405         atomic_long_t                   refcount;
 406
 407         /*
 408          * These accumulate total time (in nanoseconds) that children
 409          * events have been enabled and running, respectively.
 410          */
 411         atomic64_t                      child_total_time_enabled;
 412         atomic64_t                      child_total_time_running;
 413
 414         /*
 415          * Protect attach/detach and child_list:
 416          */
 417         struct mutex                    child_mutex;
 418         struct list_head                child_list;
 419         struct perf_event               *parent;
 420
 421         int                             oncpu;
 422         int                             cpu;
 423
 424         struct list_head                owner_entry;
 425         struct task_struct              *owner;
 426
 427         /* mmap bits */
 428         struct mutex                    mmap_mutex;
 429         atomic_t                        mmap_count;
 430
 431         struct ring_buffer              *rb;
 432         struct list_head                rb_entry;
 433         unsigned long                   rcu_batches;
 434         int                             rcu_pending;
 435
 436         /* poll related */
 437         wait_queue_head_t               waitq;
 438         struct fasync_struct            *fasync;
 439
 440         /* delayed work for NMIs and such */
 441         int                             pending_wakeup;
 442         int                             pending_kill;
 443         int                             pending_disable;
 444         struct irq_work                 pending;
 445
 446         atomic_t                        event_limit;
 447
 448         void (*destroy)(struct perf_event *);
 449         struct rcu_head                 rcu_head;
 450
 451         struct pid_namespace            *ns;
 452         u64                             id;
 453
 454         perf_overflow_handler_t         overflow_handler;
 455         void                            *overflow_handler_context;
 456
 457 #ifdef CONFIG_EVENT_TRACING
 458         struct ftrace_event_call        *tp_event;
 459         struct event_filter             *filter;
 460 #ifdef CONFIG_FUNCTION_TRACER
 461         struct ftrace_ops               ftrace_ops;
 462 #endif
 463 #endif
 464
 465 #ifdef CONFIG_CGROUP_PERF
 466         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 467         int                             cgrp_defer_enabled;
 468 #endif
 469
 470 #endif /* CONFIG_PERF_EVENTS */
 471 };
 472
 473 /**
 474  * struct perf_event_context - event context structure
 475  *
 476  * Used as a container for task events and CPU events as well:
 477  */
 478 struct perf_event_context {
 479         struct pmu                      *pmu;
 480         /*
 481          * Protect the states of the events in the list,
 482          * nr_active, and the list:
 483          */
 484         raw_spinlock_t                  lock;
 485         /*
 486          * Protect the list of events.  Locking either mutex or lock
 487          * is sufficient to ensure the list doesn't change; to change
 488          * the list you need to lock both the mutex and the spinlock.
 489          */
 490         struct mutex                    mutex;
 491
 492         struct list_head                active_ctx_list;
 493         struct list_head                pinned_groups;
 494         struct list_head                flexible_groups;
 495         struct list_head                event_list;
 496         int                             nr_events;
 497         int                             nr_active;
 498         int                             is_active;
 499         int                             nr_stat;
 500         int                             nr_freq;
 501         int                             rotate_disable;
 502         atomic_t                        refcount;
 503         struct task_struct              *task;
 504
 505         /*
 506          * Context clock, runs when context enabled.
 507          */
 508         u64                             time;
 509         u64                             timestamp;
 510
 511         /*
 512          * These fields let us detect when two contexts have both
 513          * been cloned (inherited) from a common ancestor.
 514          */
 515         struct perf_event_context       *parent_ctx;
 516         u64                             parent_gen;
 517         u64                             generation;
 518         int                             pin_count;
 519         int                             nr_cgroups;      /* cgroup evts */
 520         void                            *task_ctx_data; /* pmu specific data */
 521         struct rcu_head                 rcu_head;
 522
 523         struct delayed_work             orphans_remove;
 524         bool                            orphans_remove_sched;
 525 };
 526
 527 /*
 528  * Number of contexts where an event can trigger:
 529  *      task, softirq, hardirq, nmi.
 530  */
 531 #define PERF_NR_CONTEXTS        4
 532
 533 /**
 534  * struct perf_event_cpu_context - per cpu event context structure
 535  */
 536 struct perf_cpu_context {
 537         struct perf_event_context       ctx;
 538         struct perf_event_context       *task_ctx;
 539         int                             active_oncpu;
 540         int                             exclusive;
 541         struct hrtimer                  hrtimer;
 542         ktime_t                         hrtimer_interval;
 543         struct pmu                      *unique_pmu;
 544         struct perf_cgroup              *cgrp;
 545 };
 546
 547 struct perf_output_handle {
 548         struct perf_event               *event;
 549         struct ring_buffer              *rb;
 550         unsigned long                   wakeup;
 551         unsigned long                   size;
 552         void                            *addr;
 553         int                             page;
 554 };
 555
 556 #ifdef CONFIG_CGROUP_PERF
 557
 558 /*
 559  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 560  * This is a per-cpu dynamically allocated data structure.
 561  */
 562 struct perf_cgroup_info {
 563         u64                             time;
 564         u64                             timestamp;
 565 };
 566
 567 struct perf_cgroup {
 568         struct cgroup_subsys_state      css;
 569         struct perf_cgroup_info __percpu *info;
 570 };
 571
 572 /*
 573  * Must ensure cgroup is pinned (css_get) before calling
 574  * this function. In other words, we cannot call this function
 575  * if there is no cgroup event for the current CPU context.
 576  */
 577 static inline struct perf_cgroup *
 578 perf_cgroup_from_task(struct task_struct *task)
 579 {
 580         return container_of(task_css(task, perf_event_cgrp_id),
 581                             struct perf_cgroup, css);
 582 }
 583 #endif /* CONFIG_CGROUP_PERF */
 584
 585 #ifdef CONFIG_PERF_EVENTS
 586
 587 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 588 extern void perf_pmu_unregister(struct pmu *pmu);
 589
 590 extern int perf_num_counters(void);
 591 extern const char *perf_pmu_name(void);
 592 extern void __perf_event_task_sched_in(struct task_struct *prev,
 593                                        struct task_struct *task);
 594 extern void __perf_event_task_sched_out(struct task_struct *prev,
 595                                         struct task_struct *next);
 596 extern int perf_event_init_task(struct task_struct *child);
 597 extern void perf_event_exit_task(struct task_struct *child);
 598 extern void perf_event_free_task(struct task_struct *task);
 599 extern void perf_event_delayed_put(struct task_struct *task);
 600 extern void perf_event_print_debug(void);
 601 extern void perf_pmu_disable(struct pmu *pmu);
 602 extern void perf_pmu_enable(struct pmu *pmu);
 603 extern void perf_sched_cb_dec(struct pmu *pmu);
 604 extern void perf_sched_cb_inc(struct pmu *pmu);
 605 extern int perf_event_task_disable(void);
 606 extern int perf_event_task_enable(void);
 607 extern int perf_event_refresh(struct perf_event *event, int refresh);
 608 extern void perf_event_update_userpage(struct perf_event *event);
 609 extern int perf_event_release_kernel(struct perf_event *event);
 610 extern struct perf_event *
 611 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 612                                 int cpu,
 613                                 struct task_struct *task,
 614                                 perf_overflow_handler_t callback,
 615                                 void *context);
 616 extern void perf_pmu_migrate_context(struct pmu *pmu,
 617                                 int src_cpu, int dst_cpu);
 618 extern u64 perf_event_read_value(struct perf_event *event,
 619                                  u64 *enabled, u64 *running);
 620
 621
 622 struct perf_sample_data {
 623         /*
 624          * Fields set by perf_sample_data_init(), group so as to
 625          * minimize the cachelines touched.
 626          */
 627         u64                             addr;
 628         struct perf_raw_record          *raw;
 629         struct perf_branch_stack        *br_stack;
 630         u64                             period;
 631         u64                             weight;
 632         u64                             txn;
 633         union  perf_mem_data_src        data_src;
 634
 635         /*
 636          * The other fields, optionally {set,used} by
 637          * perf_{prepare,output}_sample().
 638          */
 639         u64                             type;
 640         u64                             ip;
 641         struct {
 642                 u32     pid;
 643                 u32     tid;
 644         }                               tid_entry;
 645         u64                             time;
 646         u64                             id;
 647         u64                             stream_id;
 648         struct {
 649                 u32     cpu;
 650                 u32     reserved;
 651         }                               cpu_entry;
 652         struct perf_callchain_entry     *callchain;
 653
 654         /*
 655          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 656          * on arch details.
 657          */
 658         struct perf_regs                regs_user;
 659         struct pt_regs                  regs_user_copy;
 660
 661         struct perf_regs                regs_intr;
 662         u64                             stack_user_size;
 663 } ____cacheline_aligned;
 664
 665 /* default value for data source */
 666 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 667                     PERF_MEM_S(LVL, NA)   |\
 668                     PERF_MEM_S(SNOOP, NA) |\
 669                     PERF_MEM_S(LOCK, NA)  |\
 670                     PERF_MEM_S(TLB, NA))
 671
 672 static inline void perf_sample_data_init(struct perf_sample_data *data,
 673                                          u64 addr, u64 period)
 674 {
 675         /* remaining struct members initialized in perf_prepare_sample() */
 676         data->addr = addr;
 677         data->raw  = NULL;
 678         data->br_stack = NULL;
 679         data->period = period;
 680         data->weight = 0;
 681         data->data_src.val = PERF_MEM_NA;
 682         data->txn = 0;
 683 }
 684
 685 extern void perf_output_sample(struct perf_output_handle *handle,
 686                                struct perf_event_header *header,
 687                                struct perf_sample_data *data,
 688                                struct perf_event *event);
 689 extern void perf_prepare_sample(struct perf_event_header *header,
 690                                 struct perf_sample_data *data,
 691                                 struct perf_event *event,
 692                                 struct pt_regs *regs);
 693
 694 extern int perf_event_overflow(struct perf_event *event,
 695                                  struct perf_sample_data *data,
 696                                  struct pt_regs *regs);
 697
 698 static inline bool is_sampling_event(struct perf_event *event)
 699 {
 700         return event->attr.sample_period != 0;
 701 }
 702
 703 /*
 704  * Return 1 for a software event, 0 for a hardware event
 705  */
 706 static inline int is_software_event(struct perf_event *event)
 707 {
 708         return event->pmu->task_ctx_nr == perf_sw_context;
 709 }
 710
 711 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 712
 713 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 714 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 715
 716 #ifndef perf_arch_fetch_caller_regs
 717 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 718 #endif
 719
 720 /*
 721  * Take a snapshot of the regs. Skip ip and frame pointer to
 722  * the nth caller. We only need a few of the regs:
 723  * - ip for PERF_SAMPLE_IP
 724  * - cs for user_mode() tests
 725  * - bp for callchains
 726  * - eflags, for future purposes, just in case
 727  */
 728 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 729 {
 730         memset(regs, 0, sizeof(*regs));
 731
 732         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 733 }
 734
 735 static __always_inline void
 736 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 737 {
 738         if (static_key_false(&perf_swevent_enabled[event_id]))
 739                 __perf_sw_event(event_id, nr, regs, addr);
 740 }
 741
 742 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 743
 744 /*
 745  * 'Special' version for the scheduler, it hard assumes no recursion,
 746  * which is guaranteed by us not actually scheduling inside other swevents
 747  * because those disable preemption.
 748  */
 749 static __always_inline void
 750 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 751 {
 752         if (static_key_false(&perf_swevent_enabled[event_id])) {
 753                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 754
 755                 perf_fetch_caller_regs(regs);
 756                 ___perf_sw_event(event_id, nr, regs, addr);
 757         }
 758 }
 759
 760 extern struct static_key_deferred perf_sched_events;
 761
 762 static inline void perf_event_task_sched_in(struct task_struct *prev,
 763                                             struct task_struct *task)
 764 {
 765         if (static_key_false(&perf_sched_events.key))
 766                 __perf_event_task_sched_in(prev, task);
 767 }
 768
 769 static inline void perf_event_task_sched_out(struct task_struct *prev,
 770                                              struct task_struct *next)
 771 {
 772         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 773
 774         if (static_key_false(&perf_sched_events.key))
 775                 __perf_event_task_sched_out(prev, next);
 776 }
 777
 778 static inline u64 __perf_event_count(struct perf_event *event)
 779 {
 780         return local64_read(&event->count) + atomic64_read(&event->child_count);
 781 }
 782
 783 extern void perf_event_mmap(struct vm_area_struct *vma);
 784 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 785 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 786 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 787
 788 extern void perf_event_exec(void);
 789 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 790 extern void perf_event_fork(struct task_struct *tsk);
 791
 792 /* Callchains */
 793 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 794
 795 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 796 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 797
 798 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 799 {
 800         if (entry->nr < PERF_MAX_STACK_DEPTH)
 801                 entry->ip[entry->nr++] = ip;
 802 }
 803
 804 extern int sysctl_perf_event_paranoid;
 805 extern int sysctl_perf_event_mlock;
 806 extern int sysctl_perf_event_sample_rate;
 807 extern int sysctl_perf_cpu_time_max_percent;
 808
 809 extern void perf_sample_event_took(u64 sample_len_ns);
 810
 811 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 812                 void __user *buffer, size_t *lenp,
 813                 loff_t *ppos);
 814 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 815                 void __user *buffer, size_t *lenp,
 816                 loff_t *ppos);
 817
 818
 819 static inline bool perf_paranoid_tracepoint_raw(void)
 820 {
 821         return sysctl_perf_event_paranoid > -1;
 822 }
 823
 824 static inline bool perf_paranoid_cpu(void)
 825 {
 826         return sysctl_perf_event_paranoid > 0;
 827 }
 828
 829 static inline bool perf_paranoid_kernel(void)
 830 {
 831         return sysctl_perf_event_paranoid > 1;
 832 }
 833
 834 extern void perf_event_init(void);
 835 extern void perf_tp_event(u64 addr, u64 count, void *record,
 836                           int entry_size, struct pt_regs *regs,
 837                           struct hlist_head *head, int rctx,
 838                           struct task_struct *task);
 839 extern void perf_bp_event(struct perf_event *event, void *data);
 840
 841 #ifndef perf_misc_flags
 842 # define perf_misc_flags(regs) \
 843                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 844 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 845 #endif
 846
 847 static inline bool has_branch_stack(struct perf_event *event)
 848 {
 849         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 850 }
 851
 852 static inline bool needs_branch_stack(struct perf_event *event)
 853 {
 854         return event->attr.branch_sample_type != 0;
 855 }
 856
 857 extern int perf_output_begin(struct perf_output_handle *handle,
 858                              struct perf_event *event, unsigned int size);
 859 extern void perf_output_end(struct perf_output_handle *handle);
 860 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 861                              const void *buf, unsigned int len);
 862 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 863                                      unsigned int len);
 864 extern int perf_swevent_get_recursion_context(void);
 865 extern void perf_swevent_put_recursion_context(int rctx);
 866 extern u64 perf_swevent_set_period(struct perf_event *event);
 867 extern void perf_event_enable(struct perf_event *event);
 868 extern void perf_event_disable(struct perf_event *event);
 869 extern int __perf_event_disable(void *info);
 870 extern void perf_event_task_tick(void);
 871 #else /* !CONFIG_PERF_EVENTS: */
 872 static inline void
 873 perf_event_task_sched_in(struct task_struct *prev,
 874                          struct task_struct *task)                      { }
 875 static inline void
 876 perf_event_task_sched_out(struct task_struct *prev,
 877                           struct task_struct *next)                     { }
 878 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 879 static inline void perf_event_exit_task(struct task_struct *child)      { }
 880 static inline void perf_event_free_task(struct task_struct *task)       { }
 881 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 882 static inline void perf_event_print_debug(void)                         { }
 883 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 884 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 885 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 886 {
 887         return -EINVAL;
 888 }
 889
 890 static inline void
 891 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 892 static inline void
 893 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
 894 static inline void
 895 perf_bp_event(struct perf_event *event, void *data)                     { }
 896
 897 static inline int perf_register_guest_info_callbacks
 898 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 899 static inline int perf_unregister_guest_info_callbacks
 900 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 901
 902 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 903 static inline void perf_event_exec(void)                                { }
 904 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
 905 static inline void perf_event_fork(struct task_struct *tsk)             { }
 906 static inline void perf_event_init(void)                                { }
 907 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 908 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 909 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
 910 static inline void perf_event_enable(struct perf_event *event)          { }
 911 static inline void perf_event_disable(struct perf_event *event)         { }
 912 static inline int __perf_event_disable(void *info)                      { return -1; }
 913 static inline void perf_event_task_tick(void)                           { }
 914 #endif
 915
 916 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 917 extern bool perf_event_can_stop_tick(void);
 918 #else
 919 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 920 #endif
 921
 922 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 923 extern void perf_restore_debug_store(void);
 924 #else
 925 static inline void perf_restore_debug_store(void)                       { }
 926 #endif
 927
 928 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 929
 930 /*
 931  * This has to have a higher priority than migration_notifier in sched/core.c.
 932  */
 933 #define perf_cpu_notifier(fn)                                           \
 934 do {                                                                    \
 935         static struct notifier_block fn##_nb =                          \
 936                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 937         unsigned long cpu = smp_processor_id();                         \
 938         unsigned long flags;                                            \
 939                                                                         \
 940         cpu_notifier_register_begin();                                  \
 941         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 942                 (void *)(unsigned long)cpu);                            \
 943         local_irq_save(flags);                                          \
 944         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
 945                 (void *)(unsigned long)cpu);                            \
 946         local_irq_restore(flags);                                       \
 947         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
 948                 (void *)(unsigned long)cpu);                            \
 949         __register_cpu_notifier(&fn##_nb);                              \
 950         cpu_notifier_register_done();                                   \
 951 } while (0)
 952
 953 /*
 954  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
 955  * callback for already online CPUs.
 956  */
 957 #define __perf_cpu_notifier(fn)                                         \
 958 do {                                                                    \
 959         static struct notifier_block fn##_nb =                          \
 960                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 961                                                                         \
 962         __register_cpu_notifier(&fn##_nb);                              \
 963 } while (0)
 964
 965 struct perf_pmu_events_attr {
 966         struct device_attribute attr;
 967         u64 id;
 968         const char *event_str;
 969 };
 970
 971 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
 972 static struct perf_pmu_events_attr _var = {                             \
 973         .attr = __ATTR(_name, 0444, _show, NULL),                       \
 974         .id   =  _id,                                                   \
 975 };
 976
 977 #define PMU_FORMAT_ATTR(_name, _format)                                 \
 978 static ssize_t                                                          \
 979 _name##_show(struct device *dev,                                        \
 980                                struct device_attribute *attr,           \
 981                                char *page)                              \
 982 {                                                                       \
 983         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 984         return sprintf(page, _format "\n");                             \
 985 }                                                                       \
 986                                                                         \
 987 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
 988
 989 #endif /* _LINUX_PERF_EVENT_H */