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                         /* for tp_event->class */
 123                         struct list_head        tp_list;
 124                 };
 125                 struct { /* intel_cqm */
 126                         int                     cqm_state;
 127                         int                     cqm_rmid;
 128                         struct list_head        cqm_events_entry;
 129                         struct list_head        cqm_groups_entry;
 130                         struct list_head        cqm_group_entry;
 131                 };
 132                 struct { /* itrace */
 133                         int                     itrace_started;
 134                 };
 135 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 136                 struct { /* breakpoint */
 137                         /*
 138                          * Crufty hack to avoid the chicken and egg
 139                          * problem hw_breakpoint has with context
 140                          * creation and event initalization.
 141                          */
 142                         struct arch_hw_breakpoint       info;
 143                         struct list_head                bp_list;
 144                 };
 145 #endif
 146         };
 147         struct task_struct              *target;
 148         int                             state;
 149         local64_t                       prev_count;
 150         u64                             sample_period;
 151         u64                             last_period;
 152         local64_t                       period_left;
 153         u64                             interrupts_seq;
 154         u64                             interrupts;
 155
 156         u64                             freq_time_stamp;
 157         u64                             freq_count_stamp;
 158 #endif
 159 };
 160
 161 /*
 162  * hw_perf_event::state flags
 163  */
 164 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 165 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 166 #define PERF_HES_ARCH           0x04
 167
 168 struct perf_event;
 169
 170 /*
 171  * Common implementation detail of pmu::{start,commit,cancel}_txn
 172  */
 173 #define PERF_EVENT_TXN 0x1
 174
 175 /**
 176  * pmu::capabilities flags
 177  */
 178 #define PERF_PMU_CAP_NO_INTERRUPT               0x01
 179 #define PERF_PMU_CAP_NO_NMI                     0x02
 180 #define PERF_PMU_CAP_AUX_NO_SG                  0x04
 181 #define PERF_PMU_CAP_AUX_SW_DOUBLEBUF           0x08
 182 #define PERF_PMU_CAP_EXCLUSIVE                  0x10
 183 #define PERF_PMU_CAP_ITRACE                     0x20
 184
 185 /**
 186  * struct pmu - generic performance monitoring unit
 187  */
 188 struct pmu {
 189         struct list_head                entry;
 190
 191         struct module                   *module;
 192         struct device                   *dev;
 193         const struct attribute_group    **attr_groups;
 194         const char                      *name;
 195         int                             type;
 196
 197         /*
 198          * various common per-pmu feature flags
 199          */
 200         int                             capabilities;
 201
 202         int * __percpu                  pmu_disable_count;
 203         struct perf_cpu_context * __percpu pmu_cpu_context;
 204         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 205         int                             task_ctx_nr;
 206         int                             hrtimer_interval_ms;
 207
 208         /*
 209          * Fully disable/enable this PMU, can be used to protect from the PMI
 210          * as well as for lazy/batch writing of the MSRs.
 211          */
 212         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 213         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 214
 215         /*
 216          * Try and initialize the event for this PMU.
 217          * Should return -ENOENT when the @event doesn't match this PMU.
 218          */
 219         int (*event_init)               (struct perf_event *event);
 220
 221         /*
 222          * Notification that the event was mapped or unmapped.  Called
 223          * in the context of the mapping task.
 224          */
 225         void (*event_mapped)            (struct perf_event *event); /*optional*/
 226         void (*event_unmapped)          (struct perf_event *event); /*optional*/
 227
 228 #define PERF_EF_START   0x01            /* start the counter when adding    */
 229 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 230 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 231
 232         /*
 233          * Adds/Removes a counter to/from the PMU, can be done inside
 234          * a transaction, see the ->*_txn() methods.
 235          */
 236         int  (*add)                     (struct perf_event *event, int flags);
 237         void (*del)                     (struct perf_event *event, int flags);
 238
 239         /*
 240          * Starts/Stops a counter present on the PMU. The PMI handler
 241          * should stop the counter when perf_event_overflow() returns
 242          * !0. ->start() will be used to continue.
 243          */
 244         void (*start)                   (struct perf_event *event, int flags);
 245         void (*stop)                    (struct perf_event *event, int flags);
 246
 247         /*
 248          * Updates the counter value of the event.
 249          */
 250         void (*read)                    (struct perf_event *event);
 251
 252         /*
 253          * Group events scheduling is treated as a transaction, add
 254          * group events as a whole and perform one schedulability test.
 255          * If the test fails, roll back the whole group
 256          *
 257          * Start the transaction, after this ->add() doesn't need to
 258          * do schedulability tests.
 259          */
 260         void (*start_txn)               (struct pmu *pmu); /* optional */
 261         /*
 262          * If ->start_txn() disabled the ->add() schedulability test
 263          * then ->commit_txn() is required to perform one. On success
 264          * the transaction is closed. On error the transaction is kept
 265          * open until ->cancel_txn() is called.
 266          */
 267         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 268         /*
 269          * Will cancel the transaction, assumes ->del() is called
 270          * for each successful ->add() during the transaction.
 271          */
 272         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 273
 274         /*
 275          * Will return the value for perf_event_mmap_page::index for this event,
 276          * if no implementation is provided it will default to: event->hw.idx + 1.
 277          */
 278         int (*event_idx)                (struct perf_event *event); /*optional */
 279
 280         /*
 281          * context-switches callback
 282          */
 283         void (*sched_task)              (struct perf_event_context *ctx,
 284                                         bool sched_in);
 285         /*
 286          * PMU specific data size
 287          */
 288         size_t                          task_ctx_size;
 289
 290
 291         /*
 292          * Return the count value for a counter.
 293          */
 294         u64 (*count)                    (struct perf_event *event); /*optional*/
 295
 296         /*
 297          * Set up pmu-private data structures for an AUX area
 298          */
 299         void *(*setup_aux)              (int cpu, void **pages,
 300                                          int nr_pages, bool overwrite);
 301                                         /* optional */
 302
 303         /*
 304          * Free pmu-private AUX data structures
 305          */
 306         void (*free_aux)                (void *aux); /* optional */
 307 };
 308
 309 /**
 310  * enum perf_event_active_state - the states of a event
 311  */
 312 enum perf_event_active_state {
 313         PERF_EVENT_STATE_EXIT           = -3,
 314         PERF_EVENT_STATE_ERROR          = -2,
 315         PERF_EVENT_STATE_OFF            = -1,
 316         PERF_EVENT_STATE_INACTIVE       =  0,
 317         PERF_EVENT_STATE_ACTIVE         =  1,
 318 };
 319
 320 struct file;
 321 struct perf_sample_data;
 322
 323 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 324                                         struct perf_sample_data *,
 325                                         struct pt_regs *regs);
 326
 327 enum perf_group_flag {
 328         PERF_GROUP_SOFTWARE             = 0x1,
 329 };
 330
 331 #define SWEVENT_HLIST_BITS              8
 332 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 333
 334 struct swevent_hlist {
 335         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 336         struct rcu_head                 rcu_head;
 337 };
 338
 339 #define PERF_ATTACH_CONTEXT     0x01
 340 #define PERF_ATTACH_GROUP       0x02
 341 #define PERF_ATTACH_TASK        0x04
 342 #define PERF_ATTACH_TASK_DATA   0x08
 343
 344 struct perf_cgroup;
 345 struct ring_buffer;
 346
 347 /**
 348  * struct perf_event - performance event kernel representation:
 349  */
 350 struct perf_event {
 351 #ifdef CONFIG_PERF_EVENTS
 352         /*
 353          * entry onto perf_event_context::event_list;
 354          *   modifications require ctx->lock
 355          *   RCU safe iterations.
 356          */
 357         struct list_head                event_entry;
 358
 359         /*
 360          * XXX: group_entry and sibling_list should be mutually exclusive;
 361          * either you're a sibling on a group, or you're the group leader.
 362          * Rework the code to always use the same list element.
 363          *
 364          * Locked for modification by both ctx->mutex and ctx->lock; holding
 365          * either sufficies for read.
 366          */
 367         struct list_head                group_entry;
 368         struct list_head                sibling_list;
 369
 370         /*
 371          * We need storage to track the entries in perf_pmu_migrate_context; we
 372          * cannot use the event_entry because of RCU and we want to keep the
 373          * group in tact which avoids us using the other two entries.
 374          */
 375         struct list_head                migrate_entry;
 376
 377         struct hlist_node               hlist_entry;
 378         struct list_head                active_entry;
 379         int                             nr_siblings;
 380         int                             group_flags;
 381         struct perf_event               *group_leader;
 382         struct pmu                      *pmu;
 383
 384         enum perf_event_active_state    state;
 385         unsigned int                    attach_state;
 386         local64_t                       count;
 387         atomic64_t                      child_count;
 388
 389         /*
 390          * These are the total time in nanoseconds that the event
 391          * has been enabled (i.e. eligible to run, and the task has
 392          * been scheduled in, if this is a per-task event)
 393          * and running (scheduled onto the CPU), respectively.
 394          *
 395          * They are computed from tstamp_enabled, tstamp_running and
 396          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 397          */
 398         u64                             total_time_enabled;
 399         u64                             total_time_running;
 400
 401         /*
 402          * These are timestamps used for computing total_time_enabled
 403          * and total_time_running when the event is in INACTIVE or
 404          * ACTIVE state, measured in nanoseconds from an arbitrary point
 405          * in time.
 406          * tstamp_enabled: the notional time when the event was enabled
 407          * tstamp_running: the notional time when the event was scheduled on
 408          * tstamp_stopped: in INACTIVE state, the notional time when the
 409          *      event was scheduled off.
 410          */
 411         u64                             tstamp_enabled;
 412         u64                             tstamp_running;
 413         u64                             tstamp_stopped;
 414
 415         /*
 416          * timestamp shadows the actual context timing but it can
 417          * be safely used in NMI interrupt context. It reflects the
 418          * context time as it was when the event was last scheduled in.
 419          *
 420          * ctx_time already accounts for ctx->timestamp. Therefore to
 421          * compute ctx_time for a sample, simply add perf_clock().
 422          */
 423         u64                             shadow_ctx_time;
 424
 425         struct perf_event_attr          attr;
 426         u16                             header_size;
 427         u16                             id_header_size;
 428         u16                             read_size;
 429         struct hw_perf_event            hw;
 430
 431         struct perf_event_context       *ctx;
 432         atomic_long_t                   refcount;
 433
 434         /*
 435          * These accumulate total time (in nanoseconds) that children
 436          * events have been enabled and running, respectively.
 437          */
 438         atomic64_t                      child_total_time_enabled;
 439         atomic64_t                      child_total_time_running;
 440
 441         /*
 442          * Protect attach/detach and child_list:
 443          */
 444         struct mutex                    child_mutex;
 445         struct list_head                child_list;
 446         struct perf_event               *parent;
 447
 448         int                             oncpu;
 449         int                             cpu;
 450
 451         struct list_head                owner_entry;
 452         struct task_struct              *owner;
 453
 454         /* mmap bits */
 455         struct mutex                    mmap_mutex;
 456         atomic_t                        mmap_count;
 457
 458         struct ring_buffer              *rb;
 459         struct list_head                rb_entry;
 460         unsigned long                   rcu_batches;
 461         int                             rcu_pending;
 462
 463         /* poll related */
 464         wait_queue_head_t               waitq;
 465         struct fasync_struct            *fasync;
 466
 467         /* delayed work for NMIs and such */
 468         int                             pending_wakeup;
 469         int                             pending_kill;
 470         int                             pending_disable;
 471         struct irq_work                 pending;
 472
 473         atomic_t                        event_limit;
 474
 475         void (*destroy)(struct perf_event *);
 476         struct rcu_head                 rcu_head;
 477
 478         struct pid_namespace            *ns;
 479         u64                             id;
 480
 481         u64                             (*clock)(void);
 482         perf_overflow_handler_t         overflow_handler;
 483         void                            *overflow_handler_context;
 484
 485 #ifdef CONFIG_EVENT_TRACING
 486         struct ftrace_event_call        *tp_event;
 487         struct event_filter             *filter;
 488 #ifdef CONFIG_FUNCTION_TRACER
 489         struct ftrace_ops               ftrace_ops;
 490 #endif
 491 #endif
 492
 493 #ifdef CONFIG_CGROUP_PERF
 494         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 495         int                             cgrp_defer_enabled;
 496 #endif
 497
 498 #endif /* CONFIG_PERF_EVENTS */
 499 };
 500
 501 /**
 502  * struct perf_event_context - event context structure
 503  *
 504  * Used as a container for task events and CPU events as well:
 505  */
 506 struct perf_event_context {
 507         struct pmu                      *pmu;
 508         /*
 509          * Protect the states of the events in the list,
 510          * nr_active, and the list:
 511          */
 512         raw_spinlock_t                  lock;
 513         /*
 514          * Protect the list of events.  Locking either mutex or lock
 515          * is sufficient to ensure the list doesn't change; to change
 516          * the list you need to lock both the mutex and the spinlock.
 517          */
 518         struct mutex                    mutex;
 519
 520         struct list_head                active_ctx_list;
 521         struct list_head                pinned_groups;
 522         struct list_head                flexible_groups;
 523         struct list_head                event_list;
 524         int                             nr_events;
 525         int                             nr_active;
 526         int                             is_active;
 527         int                             nr_stat;
 528         int                             nr_freq;
 529         int                             rotate_disable;
 530         atomic_t                        refcount;
 531         struct task_struct              *task;
 532
 533         /*
 534          * Context clock, runs when context enabled.
 535          */
 536         u64                             time;
 537         u64                             timestamp;
 538
 539         /*
 540          * These fields let us detect when two contexts have both
 541          * been cloned (inherited) from a common ancestor.
 542          */
 543         struct perf_event_context       *parent_ctx;
 544         u64                             parent_gen;
 545         u64                             generation;
 546         int                             pin_count;
 547         int                             nr_cgroups;      /* cgroup evts */
 548         void                            *task_ctx_data; /* pmu specific data */
 549         struct rcu_head                 rcu_head;
 550
 551         struct delayed_work             orphans_remove;
 552         bool                            orphans_remove_sched;
 553 };
 554
 555 /*
 556  * Number of contexts where an event can trigger:
 557  *      task, softirq, hardirq, nmi.
 558  */
 559 #define PERF_NR_CONTEXTS        4
 560
 561 /**
 562  * struct perf_event_cpu_context - per cpu event context structure
 563  */
 564 struct perf_cpu_context {
 565         struct perf_event_context       ctx;
 566         struct perf_event_context       *task_ctx;
 567         int                             active_oncpu;
 568         int                             exclusive;
 569         struct hrtimer                  hrtimer;
 570         ktime_t                         hrtimer_interval;
 571         struct pmu                      *unique_pmu;
 572         struct perf_cgroup              *cgrp;
 573 };
 574
 575 struct perf_output_handle {
 576         struct perf_event               *event;
 577         struct ring_buffer              *rb;
 578         unsigned long                   wakeup;
 579         unsigned long                   size;
 580         union {
 581                 void                    *addr;
 582                 unsigned long           head;
 583         };
 584         int                             page;
 585 };
 586
 587 #ifdef CONFIG_CGROUP_PERF
 588
 589 /*
 590  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 591  * This is a per-cpu dynamically allocated data structure.
 592  */
 593 struct perf_cgroup_info {
 594         u64                             time;
 595         u64                             timestamp;
 596 };
 597
 598 struct perf_cgroup {
 599         struct cgroup_subsys_state      css;
 600         struct perf_cgroup_info __percpu *info;
 601 };
 602
 603 /*
 604  * Must ensure cgroup is pinned (css_get) before calling
 605  * this function. In other words, we cannot call this function
 606  * if there is no cgroup event for the current CPU context.
 607  */
 608 static inline struct perf_cgroup *
 609 perf_cgroup_from_task(struct task_struct *task)
 610 {
 611         return container_of(task_css(task, perf_event_cgrp_id),
 612                             struct perf_cgroup, css);
 613 }
 614 #endif /* CONFIG_CGROUP_PERF */
 615
 616 #ifdef CONFIG_PERF_EVENTS
 617
 618 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
 619                                    struct perf_event *event);
 620 extern void perf_aux_output_end(struct perf_output_handle *handle,
 621                                 unsigned long size, bool truncated);
 622 extern int perf_aux_output_skip(struct perf_output_handle *handle,
 623                                 unsigned long size);
 624 extern void *perf_get_aux(struct perf_output_handle *handle);
 625
 626 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
 627 extern void perf_pmu_unregister(struct pmu *pmu);
 628
 629 extern int perf_num_counters(void);
 630 extern const char *perf_pmu_name(void);
 631 extern void __perf_event_task_sched_in(struct task_struct *prev,
 632                                        struct task_struct *task);
 633 extern void __perf_event_task_sched_out(struct task_struct *prev,
 634                                         struct task_struct *next);
 635 extern int perf_event_init_task(struct task_struct *child);
 636 extern void perf_event_exit_task(struct task_struct *child);
 637 extern void perf_event_free_task(struct task_struct *task);
 638 extern void perf_event_delayed_put(struct task_struct *task);
 639 extern void perf_event_print_debug(void);
 640 extern void perf_pmu_disable(struct pmu *pmu);
 641 extern void perf_pmu_enable(struct pmu *pmu);
 642 extern void perf_sched_cb_dec(struct pmu *pmu);
 643 extern void perf_sched_cb_inc(struct pmu *pmu);
 644 extern int perf_event_task_disable(void);
 645 extern int perf_event_task_enable(void);
 646 extern int perf_event_refresh(struct perf_event *event, int refresh);
 647 extern void perf_event_update_userpage(struct perf_event *event);
 648 extern int perf_event_release_kernel(struct perf_event *event);
 649 extern struct perf_event *
 650 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 651                                 int cpu,
 652                                 struct task_struct *task,
 653                                 perf_overflow_handler_t callback,
 654                                 void *context);
 655 extern void perf_pmu_migrate_context(struct pmu *pmu,
 656                                 int src_cpu, int dst_cpu);
 657 extern u64 perf_event_read_value(struct perf_event *event,
 658                                  u64 *enabled, u64 *running);
 659
 660
 661 struct perf_sample_data {
 662         /*
 663          * Fields set by perf_sample_data_init(), group so as to
 664          * minimize the cachelines touched.
 665          */
 666         u64                             addr;
 667         struct perf_raw_record          *raw;
 668         struct perf_branch_stack        *br_stack;
 669         u64                             period;
 670         u64                             weight;
 671         u64                             txn;
 672         union  perf_mem_data_src        data_src;
 673
 674         /*
 675          * The other fields, optionally {set,used} by
 676          * perf_{prepare,output}_sample().
 677          */
 678         u64                             type;
 679         u64                             ip;
 680         struct {
 681                 u32     pid;
 682                 u32     tid;
 683         }                               tid_entry;
 684         u64                             time;
 685         u64                             id;
 686         u64                             stream_id;
 687         struct {
 688                 u32     cpu;
 689                 u32     reserved;
 690         }                               cpu_entry;
 691         struct perf_callchain_entry     *callchain;
 692
 693         /*
 694          * regs_user may point to task_pt_regs or to regs_user_copy, depending
 695          * on arch details.
 696          */
 697         struct perf_regs                regs_user;
 698         struct pt_regs                  regs_user_copy;
 699
 700         struct perf_regs                regs_intr;
 701         u64                             stack_user_size;
 702 } ____cacheline_aligned;
 703
 704 /* default value for data source */
 705 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
 706                     PERF_MEM_S(LVL, NA)   |\
 707                     PERF_MEM_S(SNOOP, NA) |\
 708                     PERF_MEM_S(LOCK, NA)  |\
 709                     PERF_MEM_S(TLB, NA))
 710
 711 static inline void perf_sample_data_init(struct perf_sample_data *data,
 712                                          u64 addr, u64 period)
 713 {
 714         /* remaining struct members initialized in perf_prepare_sample() */
 715         data->addr = addr;
 716         data->raw  = NULL;
 717         data->br_stack = NULL;
 718         data->period = period;
 719         data->weight = 0;
 720         data->data_src.val = PERF_MEM_NA;
 721         data->txn = 0;
 722 }
 723
 724 extern void perf_output_sample(struct perf_output_handle *handle,
 725                                struct perf_event_header *header,
 726                                struct perf_sample_data *data,
 727                                struct perf_event *event);
 728 extern void perf_prepare_sample(struct perf_event_header *header,
 729                                 struct perf_sample_data *data,
 730                                 struct perf_event *event,
 731                                 struct pt_regs *regs);
 732
 733 extern int perf_event_overflow(struct perf_event *event,
 734                                  struct perf_sample_data *data,
 735                                  struct pt_regs *regs);
 736
 737 static inline bool is_sampling_event(struct perf_event *event)
 738 {
 739         return event->attr.sample_period != 0;
 740 }
 741
 742 /*
 743  * Return 1 for a software event, 0 for a hardware event
 744  */
 745 static inline int is_software_event(struct perf_event *event)
 746 {
 747         return event->pmu->task_ctx_nr == perf_sw_context;
 748 }
 749
 750 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 751
 752 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
 753 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
 754
 755 #ifndef perf_arch_fetch_caller_regs
 756 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 757 #endif
 758
 759 /*
 760  * Take a snapshot of the regs. Skip ip and frame pointer to
 761  * the nth caller. We only need a few of the regs:
 762  * - ip for PERF_SAMPLE_IP
 763  * - cs for user_mode() tests
 764  * - bp for callchains
 765  * - eflags, for future purposes, just in case
 766  */
 767 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 768 {
 769         memset(regs, 0, sizeof(*regs));
 770
 771         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 772 }
 773
 774 static __always_inline void
 775 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
 776 {
 777         if (static_key_false(&perf_swevent_enabled[event_id]))
 778                 __perf_sw_event(event_id, nr, regs, addr);
 779 }
 780
 781 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
 782
 783 /*
 784  * 'Special' version for the scheduler, it hard assumes no recursion,
 785  * which is guaranteed by us not actually scheduling inside other swevents
 786  * because those disable preemption.
 787  */
 788 static __always_inline void
 789 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
 790 {
 791         if (static_key_false(&perf_swevent_enabled[event_id])) {
 792                 struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
 793
 794                 perf_fetch_caller_regs(regs);
 795                 ___perf_sw_event(event_id, nr, regs, addr);
 796         }
 797 }
 798
 799 extern struct static_key_deferred perf_sched_events;
 800
 801 static inline void perf_event_task_sched_in(struct task_struct *prev,
 802                                             struct task_struct *task)
 803 {
 804         if (static_key_false(&perf_sched_events.key))
 805                 __perf_event_task_sched_in(prev, task);
 806 }
 807
 808 static inline void perf_event_task_sched_out(struct task_struct *prev,
 809                                              struct task_struct *next)
 810 {
 811         perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
 812
 813         if (static_key_false(&perf_sched_events.key))
 814                 __perf_event_task_sched_out(prev, next);
 815 }
 816
 817 static inline u64 __perf_event_count(struct perf_event *event)
 818 {
 819         return local64_read(&event->count) + atomic64_read(&event->child_count);
 820 }
 821
 822 extern void perf_event_mmap(struct vm_area_struct *vma);
 823 extern struct perf_guest_info_callbacks *perf_guest_cbs;
 824 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 825 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
 826
 827 extern void perf_event_exec(void);
 828 extern void perf_event_comm(struct task_struct *tsk, bool exec);
 829 extern void perf_event_fork(struct task_struct *tsk);
 830
 831 /* Callchains */
 832 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
 833
 834 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
 835 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
 836
 837 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
 838 {
 839         if (entry->nr < PERF_MAX_STACK_DEPTH)
 840                 entry->ip[entry->nr++] = ip;
 841 }
 842
 843 extern int sysctl_perf_event_paranoid;
 844 extern int sysctl_perf_event_mlock;
 845 extern int sysctl_perf_event_sample_rate;
 846 extern int sysctl_perf_cpu_time_max_percent;
 847
 848 extern void perf_sample_event_took(u64 sample_len_ns);
 849
 850 extern int perf_proc_update_handler(struct ctl_table *table, int write,
 851                 void __user *buffer, size_t *lenp,
 852                 loff_t *ppos);
 853 extern int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
 854                 void __user *buffer, size_t *lenp,
 855                 loff_t *ppos);
 856
 857
 858 static inline bool perf_paranoid_tracepoint_raw(void)
 859 {
 860         return sysctl_perf_event_paranoid > -1;
 861 }
 862
 863 static inline bool perf_paranoid_cpu(void)
 864 {
 865         return sysctl_perf_event_paranoid > 0;
 866 }
 867
 868 static inline bool perf_paranoid_kernel(void)
 869 {
 870         return sysctl_perf_event_paranoid > 1;
 871 }
 872
 873 extern void perf_event_init(void);
 874 extern void perf_tp_event(u64 addr, u64 count, void *record,
 875                           int entry_size, struct pt_regs *regs,
 876                           struct hlist_head *head, int rctx,
 877                           struct task_struct *task);
 878 extern void perf_bp_event(struct perf_event *event, void *data);
 879
 880 #ifndef perf_misc_flags
 881 # define perf_misc_flags(regs) \
 882                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
 883 # define perf_instruction_pointer(regs) instruction_pointer(regs)
 884 #endif
 885
 886 static inline bool has_branch_stack(struct perf_event *event)
 887 {
 888         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
 889 }
 890
 891 static inline bool needs_branch_stack(struct perf_event *event)
 892 {
 893         return event->attr.branch_sample_type != 0;
 894 }
 895
 896 static inline bool has_aux(struct perf_event *event)
 897 {
 898         return event->pmu->setup_aux;
 899 }
 900
 901 extern int perf_output_begin(struct perf_output_handle *handle,
 902                              struct perf_event *event, unsigned int size);
 903 extern void perf_output_end(struct perf_output_handle *handle);
 904 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
 905                              const void *buf, unsigned int len);
 906 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
 907                                      unsigned int len);
 908 extern int perf_swevent_get_recursion_context(void);
 909 extern void perf_swevent_put_recursion_context(int rctx);
 910 extern u64 perf_swevent_set_period(struct perf_event *event);
 911 extern void perf_event_enable(struct perf_event *event);
 912 extern void perf_event_disable(struct perf_event *event);
 913 extern int __perf_event_disable(void *info);
 914 extern void perf_event_task_tick(void);
 915 #else /* !CONFIG_PERF_EVENTS: */
 916 static inline void *
 917 perf_aux_output_begin(struct perf_output_handle *handle,
 918                       struct perf_event *event)                         { return NULL; }
 919 static inline void
 920 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size,
 921                     bool truncated)                                     { }
 922 static inline int
 923 perf_aux_output_skip(struct perf_output_handle *handle,
 924                      unsigned long size)                                { return -EINVAL; }
 925 static inline void *
 926 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
 927 static inline void
 928 perf_event_task_sched_in(struct task_struct *prev,
 929                          struct task_struct *task)                      { }
 930 static inline void
 931 perf_event_task_sched_out(struct task_struct *prev,
 932                           struct task_struct *next)                     { }
 933 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 934 static inline void perf_event_exit_task(struct task_struct *child)      { }
 935 static inline void perf_event_free_task(struct task_struct *task)       { }
 936 static inline void perf_event_delayed_put(struct task_struct *task)     { }
 937 static inline void perf_event_print_debug(void)                         { }
 938 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 939 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 940 static inline int perf_event_refresh(struct perf_event *event, int refresh)
 941 {
 942         return -EINVAL;
 943 }
 944
 945 static inline void
 946 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
 947 static inline void
 948 perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)                     { }
 949 static inline void
 950 perf_bp_event(struct perf_event *event, void *data)                     { }
 951
 952 static inline int perf_register_guest_info_callbacks
 953 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 954 static inline int perf_unregister_guest_info_callbacks
 955 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
 956
 957 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 958 static inline void perf_event_exec(void)                                { }
 959 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
 960 static inline void perf_event_fork(struct task_struct *tsk)             { }
 961 static inline void perf_event_init(void)                                { }
 962 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 963 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 964 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
 965 static inline void perf_event_enable(struct perf_event *event)          { }
 966 static inline void perf_event_disable(struct perf_event *event)         { }
 967 static inline int __perf_event_disable(void *info)                      { return -1; }
 968 static inline void perf_event_task_tick(void)                           { }
 969 #endif
 970
 971 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_NO_HZ_FULL)
 972 extern bool perf_event_can_stop_tick(void);
 973 #else
 974 static inline bool perf_event_can_stop_tick(void)                       { return true; }
 975 #endif
 976
 977 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
 978 extern void perf_restore_debug_store(void);
 979 #else
 980 static inline void perf_restore_debug_store(void)                       { }
 981 #endif
 982
 983 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
 984
 985 /*
 986  * This has to have a higher priority than migration_notifier in sched/core.c.
 987  */
 988 #define perf_cpu_notifier(fn)                                           \
 989 do {                                                                    \
 990         static struct notifier_block fn##_nb =                          \
 991                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
 992         unsigned long cpu = smp_processor_id();                         \
 993         unsigned long flags;                                            \
 994                                                                         \
 995         cpu_notifier_register_begin();                                  \
 996         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
 997                 (void *)(unsigned long)cpu);                            \
 998         local_irq_save(flags);                                          \
 999         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1000                 (void *)(unsigned long)cpu);                            \
1001         local_irq_restore(flags);                                       \
1002         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1003                 (void *)(unsigned long)cpu);                            \
1004         __register_cpu_notifier(&fn##_nb);                              \
1005         cpu_notifier_register_done();                                   \
1006 } while (0)
1007
1008 /*
1009  * Bare-bones version of perf_cpu_notifier(), which doesn't invoke the
1010  * callback for already online CPUs.
1011  */
1012 #define __perf_cpu_notifier(fn)                                         \
1013 do {                                                                    \
1014         static struct notifier_block fn##_nb =                          \
1015                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1016                                                                         \
1017         __register_cpu_notifier(&fn##_nb);                              \
1018 } while (0)
1019
1020 struct perf_pmu_events_attr {
1021         struct device_attribute attr;
1022         u64 id;
1023         const char *event_str;
1024 };
1025
1026 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1027                               char *page);
1028
1029 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1030 static struct perf_pmu_events_attr _var = {                             \
1031         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1032         .id   =  _id,                                                   \
1033 };
1034
1035 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1036 static struct perf_pmu_events_attr _var = {                                 \
1037         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1038         .id             = 0,                                                \
1039         .event_str      = _str,                                             \
1040 };
1041
1042 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1043 static ssize_t                                                          \
1044 _name##_show(struct device *dev,                                        \
1045                                struct device_attribute *attr,           \
1046                                char *page)                              \
1047 {                                                                       \
1048         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1049         return sprintf(page, _format "\n");                             \
1050 }                                                                       \
1051                                                                         \
1052 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1053
1054 #endif /* _LINUX_PERF_EVENT_H */