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                 struct { /* intel_cqm */
 127                         int                     cqm_state;
 128                         int                     cqm_rmid;
 129                         struct list_head        cqm_events_entry;
 130                         struct list_head        cqm_groups_entry;
 131                         struct list_head        cqm_group_entry;
 132                 };
 133 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 134                 struct { /* breakpoint */
 135                         /*
 136                          * Crufty hack to avoid the chicken and egg
 137                          * problem hw_breakpoint has with context
 138                          * creation and event initalization.
 139                          */
 140                         struct task_struct              *bp_target;
 141                         struct arch_hw_breakpoint       info;
 142                         struct list_head                bp_list;
 143                 };
 144 #endif
 145         };
 146         int                             state;
 147         local64_t                       prev_count;
 148         u64                             sample_period;
 149         u64                             last_period;
 150         local64_t                       period_left;
 151         u64                             interrupts_seq;
 152         u64                             interrupts;
 153
 154         u64                             freq_time_stamp;
 155         u64                             freq_count_stamp;
 156 #endif
 157 };
 158
 159 /*
 160  * hw_perf_event::state flags
 161  */
 162 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 163 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 164 #define PERF_HES_ARCH           0x04
 165
 166 struct perf_event;
 167
 168 /*
 169  * Common implementation detail of pmu::{start,commit,cancel}_txn
 170  */
 171 #define PERF_EVENT_TXN 0x1
 172
 173 /**
 174  * pmu::capabilities flags
 175  */
 176 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 177
 178 /**
 179  * struct pmu - generic performance monitoring unit
 180  */
 181 struct pmu {
 182         struct list_head                entry;
 183
 184         struct module                   *module;
 185         struct device                   *dev;
 186         const struct attribute_group    **attr_groups;
 187         const char                      *name;
 188         int                             type;
 189
 190         /*
 191          * various common per-pmu feature flags
 192          */
 193         int                             capabilities;
 194
 195         int * __percpu                  pmu_disable_count;
 196         struct perf_cpu_context * __percpu pmu_cpu_context;
 197         int                             task_ctx_nr;
 198         int                             hrtimer_interval_ms;
 199
 200         /*
 201          * Fully disable/enable this PMU, can be used to protect from the PMI
 202          * as well as for lazy/batch writing of the MSRs.
 203          */
 204         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 205         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 206
 207         /*
 208          * Try and initialize the event for this PMU.
 209          * Should return -ENOENT when the @event doesn't match this PMU.
 210          */
 211         int (*event_init)               (struct perf_event *event);
 212
 213         /*
 214          * Notification that the event was mapped or unmapped.  Called
 215          * in the context of the mapping task.
 216          */
 217         void (*event_mapped)            (struct perf_event *event); /*optional*/
 218         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 219
 220 #define PERF_EF_START   0x01            /* start the counter when adding    */
 221 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 222 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 223
 224         /*
 225          * Adds/Removes a counter to/from the PMU, can be done inside
 226          * a transaction, see the ->*_txn() methods.
 227          */
 228         int  (*add)                     (struct perf_event *event, int flags);
 229         void (*del)                     (struct perf_event *event, int flags);
 230
 231         /*
 232          * Starts/Stops a counter present on the PMU. The PMI handler
 233          * should stop the counter when perf_event_overflow() returns
 234          * !0. ->start() will be used to continue.
 235          */
 236         void (*start)                   (struct perf_event *event, int flags);
 237         void (*stop)                    (struct perf_event *event, int flags);
 238
 239         /*
 240          * Updates the counter value of the event.
 241          */
 242         void (*read)                    (struct perf_event *event);
 243
 244         /*
 245          * Group events scheduling is treated as a transaction, add
 246          * group events as a whole and perform one schedulability test.
 247          * If the test fails, roll back the whole group
 248          *
 249          * Start the transaction, after this ->add() doesn't need to
 250          * do schedulability tests.
 251          */
 252         void (*start_txn)               (struct pmu *pmu); /* optional */
 253         /*
 254          * If ->start_txn() disabled the ->add() schedulability test
 255          * then ->commit_txn() is required to perform one. On success
 256          * the transaction is closed. On error the transaction is kept
 257          * open until ->cancel_txn() is called.
 258          */
 259         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 260         /*
 261          * Will cancel the transaction, assumes ->del() is called
 262          * for each successful ->add() during the transaction.
 263          */
 264         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 265
 266         /*
 267          * Will return the value for perf_event_mmap_page::index for this event,
 268          * if no implementation is provided it will default to: event->hw.idx + 1.
 269          */
 270         int (*event_idx)                (struct perf_event *event); /*optional */
 271
 272         /*
 273          * context-switches callback
 274          */
 275         void (*sched_task)              (struct perf_event_context *ctx,
 276                                         bool sched_in);
 277         /*
 278          * PMU specific data size
 279          */
 280         size_t                          task_ctx_size;
 281
 282
 283         /*
 284          * Return the count value for a counter.
 285          */
 286         u64 (*count)                    (struct perf_event *event); /*optional*/
 287 };
 288
 289 /**
 290  * enum perf_event_active_state - the states of a event
 291  */
 292 enum perf_event_active_state {
 293         PERF_EVENT_STATE_EXIT           = -3,
 294         PERF_EVENT_STATE_ERROR          = -2,
 295         PERF_EVENT_STATE_OFF            = -1,
 296         PERF_EVENT_STATE_INACTIVE       =  0,
 297         PERF_EVENT_STATE_ACTIVE         =  1,
 298 };
 299
 300 struct file;
 301 struct perf_sample_data;
 302
 303 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 304                                         struct perf_sample_data *,
 305                                         struct pt_regs *regs);
 306
 307 enum perf_group_flag {
 308         PERF_GROUP_SOFTWARE             = 0x1,
 309 };
 310
 311 #define SWEVENT_HLIST_BITS              8
 312 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 313
 314 struct swevent_hlist {
 315         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 316         struct rcu_head                 rcu_head;
 317 };
 318
 319 #define PERF_ATTACH_CONTEXT     0x01
 320 #define PERF_ATTACH_GROUP       0x02
 321 #define PERF_ATTACH_TASK        0x04
 322 #define PERF_ATTACH_TASK_DATA   0x08
 323
 324 struct perf_cgroup;
 325 struct ring_buffer;
 326
 327 /**
 328  * struct perf_event - performance event kernel representation:
 329  */
 330 struct perf_event {
 331 #ifdef CONFIG_PERF_EVENTS
 332         /*
 333          * entry onto perf_event_context::event_list;
 334          *   modifications require ctx->lock
 335          *   RCU safe iterations.
 336          */
 337         struct list_head                event_entry;
 338
 339         /*
 340          * XXX: group_entry and sibling_list should be mutually exclusive;
 341          * either you're a sibling on a group, or you're the group leader.
 342          * Rework the code to always use the same list element.
 343          *
 344          * Locked for modification by both ctx->mutex and ctx->lock; holding
 345          * either sufficies for read.
 346          */
 347         struct list_head                group_entry;
 348         struct list_head                sibling_list;
 349
 350         /*
 351          * We need storage to track the entries in perf_pmu_migrate_context; we
 352          * cannot use the event_entry because of RCU and we want to keep the
 353          * group in tact which avoids us using the other two entries.
 354          */
 355         struct list_head                migrate_entry;
 356
 357         struct hlist_node               hlist_entry;
 358         struct list_head                active_entry;
 359         int                             nr_siblings;
 360         int                             group_flags;
 361         struct perf_event               *group_leader;
 362         struct pmu                      *pmu;
 363
 364         enum perf_event_active_state    state;
 365         unsigned int                    attach_state;
 366         local64_t                       count;
 367         atomic64_t                      child_count;
 368
 369         /*
 370          * These are the total time in nanoseconds that the event
 371          * has been enabled (i.e. eligible to run, and the task has
 372          * been scheduled in, if this is a per-task event)
 373          * and running (scheduled onto the CPU), respectively.
 374          *
 375          * They are computed from tstamp_enabled, tstamp_running and
 376          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 377          */
 378         u64                             total_time_enabled;
 379         u64                             total_time_running;
 380
 381         /*
 382          * These are timestamps used for computing total_time_enabled
 383          * and total_time_running when the event is in INACTIVE or
 384          * ACTIVE state, measured in nanoseconds from an arbitrary point
 385          * in time.
 386          * tstamp_enabled: the notional time when the event was enabled
 387          * tstamp_running: the notional time when the event was scheduled on
 388          * tstamp_stopped: in INACTIVE state, the notional time when the
 389          *      event was scheduled off.
 390          */
 391         u64                             tstamp_enabled;
 392         u64                             tstamp_running;
 393         u64                             tstamp_stopped;
 394
 395         /*
 396          * timestamp shadows the actual context timing but it can
 397          * be safely used in NMI interrupt context. It reflects the
 398          * context time as it was when the event was last scheduled in.
 399          *
 400          * ctx_time already accounts for ctx->timestamp. Therefore to
 401          * compute ctx_time for a sample, simply add perf_clock().
 402          */
 403         u64                             shadow_ctx_time;
 404
 405         struct perf_event_attr          attr;
 406         u16                             header_size;
 407         u16                             id_header_size;
 408         u16                             read_size;
 409         struct hw_perf_event            hw;
 410
 411         struct perf_event_context       *ctx;
 412         atomic_long_t                   refcount;
 413
 414         /*
 415          * These accumulate total time (in nanoseconds) that children
 416          * events have been enabled and running, respectively.
 417          */
 418         atomic64_t                      child_total_time_enabled;
 419         atomic64_t                      child_total_time_running;
 420
 421         /*
 422          * Protect attach/detach and child_list:
 423          */
 424         struct mutex                    child_mutex;
 425         struct list_head                child_list;
 426         struct perf_event               *parent;
 427
 428         int                             oncpu;
 429         int                             cpu;
 430
 431         struct list_head                owner_entry;
 432         struct task_struct              *owner;
 433
 434         /* mmap bits */
 435         struct mutex                    mmap_mutex;
 436         atomic_t                        mmap_count;
 437
 438         struct ring_buffer              *rb;
 439         struct list_head                rb_entry;
 440         unsigned long                   rcu_batches;
 441         int                             rcu_pending;
 442
 443         /* poll related */
 444         wait_queue_head_t               waitq;
 445         struct fasync_struct            *fasync;
 446
 447         /* delayed work for NMIs and such */
 448         int                             pending_wakeup;
 449         int                             pending_kill;
 450         int                             pending_disable;
 451         struct irq_work                 pending;
 452
 453         atomic_t                        event_limit;
 454
 455         void (*destroy)(struct perf_event *);
 456         struct rcu_head                 rcu_head;
 457
 458         struct pid_namespace            *ns;
 459         u64                             id;
 460
 461         perf_overflow_handler_t         overflow_handler;
 462         void                            *overflow_handler_context;
 463
 464 #ifdef CONFIG_EVENT_TRACING
 465         struct ftrace_event_call        *tp_event;
 466         struct event_filter             *filter;
 467 #ifdef CONFIG_FUNCTION_TRACER
 468         struct ftrace_ops               ftrace_ops;
 469 #endif
 470 #endif
 471
 472 #ifdef CONFIG_CGROUP_PERF
 473         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 474         int                             cgrp_defer_enabled;
 475 #endif
 476
 477 #endif /* CONFIG_PERF_EVENTS */
 478 };
 479
 480 /**
 481  * struct perf_event_context - event context structure
 482  *
 483  * Used as a container for task events and CPU events as well:
 484  */
 485 struct perf_event_context {
 486         struct pmu                      *pmu;
 487         /*
 488          * Protect the states of the events in the list,
 489          * nr_active, and the list:
 490          */
 491         raw_spinlock_t                  lock;
 492         /*
 493          * Protect the list of events.  Locking either mutex or lock
 494          * is sufficient to ensure the list doesn't change; to change
 495          * the list you need to lock both the mutex and the spinlock.
 496          */
 497         struct mutex                    mutex;
 498
 499         struct list_head                active_ctx_list;
 500         struct list_head                pinned_groups;
 501         struct list_head                flexible_groups;
 502         struct list_head                event_list;
 503         int                             nr_events;
 504         int                             nr_active;
 505         int                             is_active;
 506         int                             nr_stat;
 507         int                             nr_freq;
 508         int                             rotate_disable;
 509         atomic_t                        refcount;
 510         struct task_struct              *task;
 511
 512         /*
 513          * Context clock, runs when context enabled.
 514          */
 515         u64                             time;
 516         u64                             timestamp;
 517
 518         /*
 519          * These fields let us detect when two contexts have both
 520          * been cloned (inherited) from a common ancestor.
 521          */
 522         struct perf_event_context       *parent_ctx;
 523         u64                             parent_gen;
 524         u64                             generation;
 525         int                             pin_count;
 526         int                             nr_cgroups;      /* cgroup evts */
 527         void                            *task_ctx_data; /* pmu specific data */
 528         struct rcu_head                 rcu_head;
 529
 530         struct delayed_work             orphans_remove;
 531         bool                            orphans_remove_sched;
 532 };
 533
 534 /*
 535  * Number of contexts where an event can trigger:
 536  *      task, softirq, hardirq, nmi.
 537  */
 538 #define PERF_NR_CONTEXTS        4
 539
 540 /**
 541  * struct perf_event_cpu_context - per cpu event context structure
 542  */
 543 struct perf_cpu_context {
 544         struct perf_event_context       ctx;
 545         struct perf_event_context       *task_ctx;
 546         int                             active_oncpu;
 547         int                             exclusive;
 548         struct hrtimer                  hrtimer;
 549         ktime_t                         hrtimer_interval;
 550         struct pmu                      *unique_pmu;
 551         struct perf_cgroup              *cgrp;
 552 };
 553
 554 struct perf_output_handle {
 555         struct perf_event               *event;
 556         struct ring_buffer              *rb;
 557         unsigned long                   wakeup;
 558         unsigned long                   size;
 559         void                            *addr;
 560         int                             page;
 561 };
 562
 563 #ifdef CONFIG_CGROUP_PERF
 564
 565 /*
 566  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 567  * This is a per-cpu dynamically allocated data structure.
 568  */
 569 struct perf_cgroup_info {
 570         u64                             time;
 571         u64                             timestamp;
 572 };
 573
 574 struct perf_cgroup {
 575         struct cgroup_subsys_state      css;
 576         struct perf_cgroup_info __percpu *info;
 577 };
 578
 579 /*
 580  * Must ensure cgroup is pinned (css_get) before calling
 581  * this function. In other words, we cannot call this function
 582  * if there is no cgroup event for the current CPU context.
 583  */
 584 static inline struct perf_cgroup *
 585 perf_cgroup_from_task(struct task_struct *task)
 586 {
 587         return container_of(task_css(task, perf_event_cgrp_id),
 588                             struct perf_cgroup, css);
 589 }
 590 #endif /* CONFIG_CGROUP_PERF */
 591
 592 #ifdef CONFIG_PERF_EVENTS
 593
 594 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 595 extern void perf_pmu_unregister(struct pmu *pmu);
 596
 597 extern int perf_num_counters(void);
 598 extern const char *perf_pmu_name(void);
 599 extern void __perf_event_task_sched_in(struct task_struct *prev,
 600                                        struct task_struct *task);
 601 extern void __perf_event_task_sched_out(struct task_struct *prev,
 602                                         struct task_struct *next);
 603 extern int perf_event_init_task(struct task_struct *child);
 604 extern void perf_event_exit_task(struct task_struct *child);
 605 extern void perf_event_free_task(struct task_struct *task);
 606 extern void perf_event_delayed_put(struct task_struct *task);
 607 extern void perf_event_print_debug(void);
 608 extern void perf_pmu_disable(struct pmu *pmu);
 609 extern void perf_pmu_enable(struct pmu *pmu);
 610 extern void perf_sched_cb_dec(struct pmu *pmu);
 611 extern void perf_sched_cb_inc(struct pmu *pmu);
 612 extern int perf_event_task_disable(void);
 613 extern int perf_event_task_enable(void);
 614 extern int perf_event_refresh(struct perf_event *event, int refresh);
 615 extern void perf_event_update_userpage(struct perf_event *event);
 616 extern int perf_event_release_kernel(struct perf_event *event);
 617 extern struct perf_event *
 618 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 619                                 int cpu,
 620                                 struct task_struct *task,
 621                                 perf_overflow_handler_t callback,
 622                                 void *context);
 623 extern void perf_pmu_migrate_context(struct pmu *pmu,
 624                                 int src_cpu, int dst_cpu);
 625 extern u64 perf_event_read_value(struct perf_event *event,
 626                                  u64 *enabled, u64 *running);
 627
 628
 629 struct perf_sample_data {
 630         /*
 631          * Fields set by perf_sample_data_init(), group so as to
 632          * minimize the cachelines touched.
 633          */
 634         u64                             addr;
 635         struct perf_raw_record          *raw;
 636         struct perf_branch_stack        *br_stack;
 637         u64                             period;
 638         u64                             weight;
 639         u64                             txn;
 640         union  perf_mem_data_src        data_src;
 641
 642         /*
 643          * The other fields, optionally {set,used} by
 644          * perf_{prepare,output}_sample().
 645          */
 646         u64                             type;
 647         u64                             ip;
 648         struct {
 649                 u32     pid;
 650                 u32     tid;
 651         }                               tid_entry;
 652         u64                             time;
 653         u64                             id;
 654         u64                             stream_id;
 655         struct {
 656                 u32     cpu;
 657                 u32     reserved;
 658         }                               cpu_entry;
 659         struct perf_callchain_entry     *callchain;
 660
 661         /*
 662          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 663          * on arch details.
 664          */
 665         struct perf_regs                regs_user;
 666         struct pt_regs                  regs_user_copy;
 667
 668         struct perf_regs                regs_intr;
 669         u64                             stack_user_size;
 670 } ____cacheline_aligned;
 671
 672 /* default value for data source */
 673 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 674                     PERF_MEM_S(LVL, NA)   |\
 675                     PERF_MEM_S(SNOOP, NA) |\
 676                     PERF_MEM_S(LOCK, NA)  |\
 677                     PERF_MEM_S(TLB, NA))
 678
 679 static inline void perf_sample_data_init(struct perf_sample_data *data,
 680                                          u64 addr, u64 period)
 681 {
 682         /* remaining struct members initialized in perf_prepare_sample() */
 683         data->addr = addr;
 684         data->raw  = NULL;
 685         data->br_stack = NULL;
 686         data->period = period;
 687         data->weight = 0;
 688         data->data_src.val = PERF_MEM_NA;
 689         data->txn = 0;
 690 }
 691
 692 extern void perf_output_sample(struct perf_output_handle *handle,
 693                                struct perf_event_header *header,
 694                                struct perf_sample_data *data,
 695                                struct perf_event *event);
 696 extern void perf_prepare_sample(struct perf_event_header *header,
 697                                 struct perf_sample_data *data,
 698                                 struct perf_event *event,
 699                                 struct pt_regs *regs);
 700
 701 extern int perf_event_overflow(struct perf_event *event,
 702                                  struct perf_sample_data *data,
 703                                  struct pt_regs *regs);
 704
 705 static inline bool is_sampling_event(struct perf_event *event)
 706 {
 707         return event->attr.sample_period != 0;
 708 }
 709
 710 /*
 711  * Return 1 for a software event, 0 for a hardware event
 712  */
 713 static inline int is_software_event(struct perf_event *event)
 714 {
 715         return event->pmu->task_ctx_nr == perf_sw_context;
 716 }
 717
 718 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 719
 720 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 721 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 722
 723 #ifndef perf_arch_fetch_caller_regs
 724 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 725 #endif
 726
 727 /*
 728  * Take a snapshot of the regs. Skip ip and frame pointer to
 729  * the nth caller. We only need a few of the regs:
 730  * - ip for PERF_SAMPLE_IP
 731  * - cs for user_mode() tests
 732  * - bp for callchains
 733  * - eflags, for future purposes, just in case
 734  */
 735 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 736 {
 737         memset(regs, 0, sizeof(*regs));
 738
 739         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 740 }
 741
 742 static __always_inline void
 743 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 744 {
 745         if (static_key_false(&perf_swevent_enabled[event_id]))
 746                 __perf_sw_event(event_id, nr, regs, addr);
 747 }
 748
 749 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 750
 751 /*
 752  * 'Special' version for the scheduler, it hard assumes no recursion,
 753  * which is guaranteed by us not actually scheduling inside other swevents
 754  * because those disable preemption.
 755  */
 756 static __always_inline void
 757 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 758 {
 759         if (static_key_false(&perf_swevent_enabled[event_id])) {
 760                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 761
 762                 perf_fetch_caller_regs(regs);
 763                 ___perf_sw_event(event_id, nr, regs, addr);
 764         }
 765 }
 766
 767 extern struct static_key_deferred perf_sched_events;
 768
 769 static inline void perf_event_task_sched_in(struct task_struct *prev,
 770                                             struct task_struct *task)
 771 {
 772         if (static_key_false(&perf_sched_events.key))
 773                 __perf_event_task_sched_in(prev, task);
 774 }
 775
 776 static inline void perf_event_task_sched_out(struct task_struct *prev,
 777                                              struct task_struct *next)
 778 {
 779         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 780
 781         if (static_key_false(&perf_sched_events.key))
 782                 __perf_event_task_sched_out(prev, next);
 783 }
 784
 785 static inline u64 __perf_event_count(struct perf_event *event)
 786 {
 787         return local64_read(&event->count) + atomic64_read(&event->child_count);
 788 }
 789
 790 extern void perf_event_mmap(struct vm_area_struct *vma);
 791 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 792 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 793 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 794
 795 extern void perf_event_exec(void);
 796 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 797 extern void perf_event_fork(struct task_struct *tsk);
 798
 799 /* Callchains */
 800 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 801
 802 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 803 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 804
 805 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 806 {
 807         if (entry->nr < PERF_MAX_STACK_DEPTH)
 808                 entry->ip[entry->nr++] = ip;
 809 }
 810
 811 extern int sysctl_perf_event_paranoid;
 812 extern int sysctl_perf_event_mlock;
 813 extern int sysctl_perf_event_sample_rate;
 814 extern int sysctl_perf_cpu_time_max_percent;
 815
 816 extern void perf_sample_event_took(u64 sample_len_ns);
 817
 818 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 819                 void __user *buffer, size_t *lenp,
 820                 loff_t *ppos);
 821 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 822                 void __user *buffer, size_t *lenp,
 823                 loff_t *ppos);
 824
 825
 826 static inline bool perf_paranoid_tracepoint_raw(void)
 827 {
 828         return sysctl_perf_event_paranoid > -1;
 829 }
 830
 831 static inline bool perf_paranoid_cpu(void)
 832 {
 833         return sysctl_perf_event_paranoid > 0;
 834 }
 835
 836 static inline bool perf_paranoid_kernel(void)
 837 {
 838         return sysctl_perf_event_paranoid > 1;
 839 }
 840
 841 extern void perf_event_init(void);
 842 extern void perf_tp_event(u64 addr, u64 count, void *record,
 843                           int entry_size, struct pt_regs *regs,
 844                           struct hlist_head *head, int rctx,
 845                           struct task_struct *task);
 846 extern void perf_bp_event(struct perf_event *event, void *data);
 847
 848 #ifndef perf_misc_flags
 849 # define perf_misc_flags(regs) \
 850                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 851 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 852 #endif
 853
 854 static inline bool has_branch_stack(struct perf_event *event)
 855 {
 856         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 857 }
 858
 859 static inline bool needs_branch_stack(struct perf_event *event)
 860 {
 861         return event->attr.branch_sample_type != 0;
 862 }
 863
 864 extern int perf_output_begin(struct perf_output_handle *handle,
 865                              struct perf_event *event, unsigned int size);
 866 extern void perf_output_end(struct perf_output_handle *handle);
 867 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 868                              const void *buf, unsigned int len);
 869 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 870                                      unsigned int len);
 871 extern int perf_swevent_get_recursion_context(void);
 872 extern void perf_swevent_put_recursion_context(int rctx);
 873 extern u64 perf_swevent_set_period(struct perf_event *event);
 874 extern void perf_event_enable(struct perf_event *event);
 875 extern void perf_event_disable(struct perf_event *event);
 876 extern int __perf_event_disable(void *info);
 877 extern void perf_event_task_tick(void);
 878 #else /* !CONFIG_PERF_EVENTS: */
 879 static inline void
 880 perf_event_task_sched_in(struct task_struct *prev,
 881                          struct task_struct *task)                      { }
 882 static inline void
 883 perf_event_task_sched_out(struct task_struct *prev,
 884                           struct task_struct *next)                     { }
 885 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 886 static inline void perf_event_exit_task(struct task_struct *child)      { }
 887 static inline void perf_event_free_task(struct task_struct *task)       { }
 888 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 889 static inline void perf_event_print_debug(void)                         { }
 890 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 891 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 892 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 893 {
 894         return -EINVAL;
 895 }
 896
 897 static inline void
 898 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 899 static inline void
 900 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
 901 static inline void
 902 perf_bp_event(struct perf_event *event, void *data)                     { }
 903
 904 static inline int perf_register_guest_info_callbacks
 905 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 906 static inline int perf_unregister_guest_info_callbacks
 907 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 908
 909 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 910 static inline void perf_event_exec(void)                                { }
 911 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
 912 static inline void perf_event_fork(struct task_struct *tsk)             { }
 913 static inline void perf_event_init(void)                                { }
 914 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 915 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 916 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
 917 static inline void perf_event_enable(struct perf_event *event)          { }
 918 static inline void perf_event_disable(struct perf_event *event)         { }
 919 static inline int __perf_event_disable(void *info)                      { return -1; }
 920 static inline void perf_event_task_tick(void)                           { }
 921 #endif
 922
 923 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 924 extern bool perf_event_can_stop_tick(void);
 925 #else
 926 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 927 #endif
 928
 929 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 930 extern void perf_restore_debug_store(void);
 931 #else
 932 static inline void perf_restore_debug_store(void)                       { }
 933 #endif
 934
 935 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 936
 937 /*
 938  * This has to have a higher priority than migration_notifier in sched/core.c.
 939  */
 940 #define perf_cpu_notifier(fn)                                           \
 941 do {                                                                    \
 942         static struct notifier_block fn##_nb =                          \
 943                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 944         unsigned long cpu = smp_processor_id();                         \
 945         unsigned long flags;                                            \
 946                                                                         \
 947         cpu_notifier_register_begin();                                  \
 948         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 949                 (void *)(unsigned long)cpu);                            \
 950         local_irq_save(flags);                                          \
 951         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
 952                 (void *)(unsigned long)cpu);                            \
 953         local_irq_restore(flags);                                       \
 954         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
 955                 (void *)(unsigned long)cpu);                            \
 956         __register_cpu_notifier(&fn##_nb);                              \
 957         cpu_notifier_register_done();                                   \
 958 } while (0)
 959
 960 /*
 961  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
 962  * callback for already online CPUs.
 963  */
 964 #define __perf_cpu_notifier(fn)                                         \
 965 do {                                                                    \
 966         static struct notifier_block fn##_nb =                          \
 967                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 968                                                                         \
 969         __register_cpu_notifier(&fn##_nb);                              \
 970 } while (0)
 971
 972 struct perf_pmu_events_attr {
 973         struct device_attribute attr;
 974         u64 id;
 975         const char *event_str;
 976 };
 977
 978 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
 979 static struct perf_pmu_events_attr _var = {                             \
 980         .attr = __ATTR(_name, 0444, _show, NULL),                       \
 981         .id   =  _id,                                                   \
 982 };
 983
 984 #define PMU_FORMAT_ATTR(_name, _format)                                 \
 985 static ssize_t                                                          \
 986 _name##_show(struct device *dev,                                        \
 987                                struct device_attribute *attr,           \
 988                                char *page)                              \
 989 {                                                                       \
 990         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
 991         return sprintf(page, _format "\n");                             \
 992 }                                                                       \
 993                                                                         \
 994 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
 995
 996 #endif /* _LINUX_PERF_EVENT_H */