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 #include <uapi/linux/bpf_perf_event.h>
  19
  20 /*
  21  * Kernel-internal data types and definitions:
  22  */
  23
  24 #ifdef CONFIG_PERF_EVENTS
  25 # include <asm/perf_event.h>
  26 # include <asm/local64.h>
  27 #endif
  28
  29 #define PERF_GUEST_ACTIVE       0x01
  30 #define PERF_GUEST_USER 0x02
  31
  32 struct perf_guest_info_callbacks {
  33         unsigned int                    (*state)(void);
  34         unsigned long                   (*get_ip)(void);
  35         unsigned int                    (*handle_intel_pt_intr)(void);
  36 };
  37
  38 #ifdef CONFIG_HAVE_HW_BREAKPOINT
  39 #include <linux/rhashtable-types.h>
  40 #include <asm/hw_breakpoint.h>
  41 #endif
  42
  43 #include <linux/list.h>
  44 #include <linux/mutex.h>
  45 #include <linux/rculist.h>
  46 #include <linux/rcupdate.h>
  47 #include <linux/spinlock.h>
  48 #include <linux/hrtimer.h>
  49 #include <linux/fs.h>
  50 #include <linux/pid_namespace.h>
  51 #include <linux/workqueue.h>
  52 #include <linux/ftrace.h>
  53 #include <linux/cpu.h>
  54 #include <linux/irq_work.h>
  55 #include <linux/static_key.h>
  56 #include <linux/jump_label_ratelimit.h>
  57 #include <linux/atomic.h>
  58 #include <linux/sysfs.h>
  59 #include <linux/perf_regs.h>
  60 #include <linux/cgroup.h>
  61 #include <linux/refcount.h>
  62 #include <linux/security.h>
  63 #include <linux/static_call.h>
  64 #include <linux/lockdep.h>
  65 #include <asm/local.h>
  66
  67 struct perf_callchain_entry {
  68         __u64                           nr;
  69         __u64                           ip[]; /* /proc/sys/kernel/perf_event_max_stack */
  70 };
  71
  72 struct perf_callchain_entry_ctx {
  73         struct perf_callchain_entry *entry;
  74         u32                         max_stack;
  75         u32                         nr;
  76         short                       contexts;
  77         bool                        contexts_maxed;
  78 };
  79
  80 typedef unsigned long (*perf_copy_f)(void *dst, const void *src,
  81                                      unsigned long off, unsigned long len);
  82
  83 struct perf_raw_frag {
  84         union {
  85                 struct perf_raw_frag    *next;
  86                 unsigned long           pad;
  87         };
  88         perf_copy_f                     copy;
  89         void                            *data;
  90         u32                             size;
  91 } __packed;
  92
  93 struct perf_raw_record {
  94         struct perf_raw_frag            frag;
  95         u32                             size;
  96 };
  97
  98 static __always_inline bool perf_raw_frag_last(const struct perf_raw_frag *frag)
  99 {
 100         return frag->pad < sizeof(u64);
 101 }
 102
 103 /*
 104  * branch stack layout:
 105  *  nr: number of taken branches stored in entries[]
 106  *  hw_idx: The low level index of raw branch records
 107  *          for the most recent branch.
 108  *          -1ULL means invalid/unknown.
 109  *
 110  * Note that nr can vary from sample to sample
 111  * branches (to, from) are stored from most recent
 112  * to least recent, i.e., entries[0] contains the most
 113  * recent branch.
 114  * The entries[] is an abstraction of raw branch records,
 115  * which may not be stored in age order in HW, e.g. Intel LBR.
 116  * The hw_idx is to expose the low level index of raw
 117  * branch record for the most recent branch aka entries[0].
 118  * The hw_idx index is between -1 (unknown) and max depth,
 119  * which can be retrieved in /sys/devices/cpu/caps/branches.
 120  * For the architectures whose raw branch records are
 121  * already stored in age order, the hw_idx should be 0.
 122  */
 123 struct perf_branch_stack {
 124         __u64                           nr;
 125         __u64                           hw_idx;
 126         struct perf_branch_entry        entries[];
 127 };
 128
 129 struct task_struct;
 130
 131 /*
 132  * extra PMU register associated with an event
 133  */
 134 struct hw_perf_event_extra {
 135         u64             config; /* register value */
 136         unsigned int    reg;    /* register address or index */
 137         int             alloc;  /* extra register already allocated */
 138         int             idx;    /* index in shared_regs->regs[] */
 139 };
 140
 141 /**
 142  * hw_perf_event::flag values
 143  *
 144  * PERF_EVENT_FLAG_ARCH bits are reserved for architecture-specific
 145  * usage.
 146  */
 147 #define PERF_EVENT_FLAG_ARCH                    0x000fffff
 148 #define PERF_EVENT_FLAG_USER_READ_CNT           0x80000000
 149
 150 static_assert((PERF_EVENT_FLAG_USER_READ_CNT & PERF_EVENT_FLAG_ARCH) == 0);
 151
 152 /**
 153  * struct hw_perf_event - performance event hardware details:
 154  */
 155 struct hw_perf_event {
 156 #ifdef CONFIG_PERF_EVENTS
 157         union {
 158                 struct { /* hardware */
 159                         u64             config;
 160                         u64             last_tag;
 161                         unsigned long   config_base;
 162                         unsigned long   event_base;
 163                         int             event_base_rdpmc;
 164                         int             idx;
 165                         int             last_cpu;
 166                         int             flags;
 167
 168                         struct hw_perf_event_extra extra_reg;
 169                         struct hw_perf_event_extra branch_reg;
 170                 };
 171                 struct { /* software */
 172                         struct hrtimer  hrtimer;
 173                 };
 174                 struct { /* tracepoint */
 175                         /* for tp_event->class */
 176                         struct list_head        tp_list;
 177                 };
 178                 struct { /* amd_power */
 179                         u64     pwr_acc;
 180                         u64     ptsc;
 181                 };
 182 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 183                 struct { /* breakpoint */
 184                         /*
 185                          * Crufty hack to avoid the chicken and egg
 186                          * problem hw_breakpoint has with context
 187                          * creation and event initalization.
 188                          */
 189                         struct arch_hw_breakpoint       info;
 190                         struct rhlist_head              bp_list;
 191                 };
 192 #endif
 193                 struct { /* amd_iommu */
 194                         u8      iommu_bank;
 195                         u8      iommu_cntr;
 196                         u16     padding;
 197                         u64     conf;
 198                         u64     conf1;
 199                 };
 200         };
 201         /*
 202          * If the event is a per task event, this will point to the task in
 203          * question. See the comment in perf_event_alloc().
 204          */
 205         struct task_struct              *target;
 206
 207         /*
 208          * PMU would store hardware filter configuration
 209          * here.
 210          */
 211         void                            *addr_filters;
 212
 213         /* Last sync'ed generation of filters */
 214         unsigned long                   addr_filters_gen;
 215
 216 /*
 217  * hw_perf_event::state flags; used to track the PERF_EF_* state.
 218  */
 219 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 220 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 221 #define PERF_HES_ARCH           0x04
 222
 223         int                             state;
 224
 225         /*
 226          * The last observed hardware counter value, updated with a
 227          * local64_cmpxchg() such that pmu::read() can be called nested.
 228          */
 229         local64_t                       prev_count;
 230
 231         /*
 232          * The period to start the next sample with.
 233          */
 234         u64                             sample_period;
 235
 236         union {
 237                 struct { /* Sampling */
 238                         /*
 239                          * The period we started this sample with.
 240                          */
 241                         u64                             last_period;
 242
 243                         /*
 244                          * However much is left of the current period;
 245                          * note that this is a full 64bit value and
 246                          * allows for generation of periods longer
 247                          * than hardware might allow.
 248                          */
 249                         local64_t                       period_left;
 250                 };
 251                 struct { /* Topdown events counting for context switch */
 252                         u64                             saved_metric;
 253                         u64                             saved_slots;
 254                 };
 255         };
 256
 257         /*
 258          * State for throttling the event, see __perf_event_overflow() and
 259          * perf_adjust_freq_unthr_context().
 260          */
 261         u64                             interrupts_seq;
 262         u64                             interrupts;
 263
 264         /*
 265          * State for freq target events, see __perf_event_overflow() and
 266          * perf_adjust_freq_unthr_context().
 267          */
 268         u64                             freq_time_stamp;
 269         u64                             freq_count_stamp;
 270 #endif
 271 };
 272
 273 struct perf_event;
 274 struct perf_event_pmu_context;
 275
 276 /*
 277  * Common implementation detail of pmu::{start,commit,cancel}_txn
 278  */
 279 #define PERF_PMU_TXN_ADD  0x1           /* txn to add/schedule event on PMU */
 280 #define PERF_PMU_TXN_READ 0x2           /* txn to read event group from PMU */
 281
 282 /**
 283  * pmu::capabilities flags
 284  */
 285 #define PERF_PMU_CAP_NO_INTERRUPT               0x0001
 286 #define PERF_PMU_CAP_NO_NMI                     0x0002
 287 #define PERF_PMU_CAP_AUX_NO_SG                  0x0004
 288 #define PERF_PMU_CAP_EXTENDED_REGS              0x0008
 289 #define PERF_PMU_CAP_EXCLUSIVE                  0x0010
 290 #define PERF_PMU_CAP_ITRACE                     0x0020
 291 #define PERF_PMU_CAP_HETEROGENEOUS_CPUS         0x0040
 292 #define PERF_PMU_CAP_NO_EXCLUDE                 0x0080
 293 #define PERF_PMU_CAP_AUX_OUTPUT                 0x0100
 294 #define PERF_PMU_CAP_EXTENDED_HW_TYPE           0x0200
 295
 296 struct perf_output_handle;
 297
 298 #define PMU_NULL_DEV    ((void *)(~0UL))
 299
 300 /**
 301  * struct pmu - generic performance monitoring unit
 302  */
 303 struct pmu {
 304         struct list_head                entry;
 305
 306         struct module                   *module;
 307         struct device                   *dev;
 308         struct device                   *parent;
 309         const struct attribute_group    **attr_groups;
 310         const struct attribute_group    **attr_update;
 311         const char                      *name;
 312         int                             type;
 313
 314         /*
 315          * various common per-pmu feature flags
 316          */
 317         int                             capabilities;
 318
 319         int __percpu                    *pmu_disable_count;
 320         struct perf_cpu_pmu_context __percpu *cpu_pmu_context;
 321         atomic_t                        exclusive_cnt; /* < 0: cpu; > 0: tsk */
 322         int                             task_ctx_nr;
 323         int                             hrtimer_interval_ms;
 324
 325         /* number of address filters this PMU can do */
 326         unsigned int                    nr_addr_filters;
 327
 328         /*
 329          * Fully disable/enable this PMU, can be used to protect from the PMI
 330          * as well as for lazy/batch writing of the MSRs.
 331          */
 332         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 333         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 334
 335         /*
 336          * Try and initialize the event for this PMU.
 337          *
 338          * Returns:
 339          *  -ENOENT     -- @event is not for this PMU
 340          *
 341          *  -ENODEV     -- @event is for this PMU but PMU not present
 342          *  -EBUSY      -- @event is for this PMU but PMU temporarily unavailable
 343          *  -EINVAL     -- @event is for this PMU but @event is not valid
 344          *  -EOPNOTSUPP -- @event is for this PMU, @event is valid, but not supported
 345          *  -EACCES     -- @event is for this PMU, @event is valid, but no privileges
 346          *
 347          *  0           -- @event is for this PMU and valid
 348          *
 349          * Other error return values are allowed.
 350          */
 351         int (*event_init)               (struct perf_event *event);
 352
 353         /*
 354          * Notification that the event was mapped or unmapped.  Called
 355          * in the context of the mapping task.
 356          */
 357         void (*event_mapped)            (struct perf_event *event, struct mm_struct *mm); /* optional */
 358         void (*event_unmapped)          (struct perf_event *event, struct mm_struct *mm); /* optional */
 359
 360         /*
 361          * Flags for ->add()/->del()/ ->start()/->stop(). There are
 362          * matching hw_perf_event::state flags.
 363          */
 364 #define PERF_EF_START   0x01            /* start the counter when adding    */
 365 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 366 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 367
 368         /*
 369          * Adds/Removes a counter to/from the PMU, can be done inside a
 370          * transaction, see the ->*_txn() methods.
 371          *
 372          * The add/del callbacks will reserve all hardware resources required
 373          * to service the event, this includes any counter constraint
 374          * scheduling etc.
 375          *
 376          * Called with IRQs disabled and the PMU disabled on the CPU the event
 377          * is on.
 378          *
 379          * ->add() called without PERF_EF_START should result in the same state
 380          *  as ->add() followed by ->stop().
 381          *
 382          * ->del() must always PERF_EF_UPDATE stop an event. If it calls
 383          *  ->stop() that must deal with already being stopped without
 384          *  PERF_EF_UPDATE.
 385          */
 386         int  (*add)                     (struct perf_event *event, int flags);
 387         void (*del)                     (struct perf_event *event, int flags);
 388
 389         /*
 390          * Starts/Stops a counter present on the PMU.
 391          *
 392          * The PMI handler should stop the counter when perf_event_overflow()
 393          * returns !0. ->start() will be used to continue.
 394          *
 395          * Also used to change the sample period.
 396          *
 397          * Called with IRQs disabled and the PMU disabled on the CPU the event
 398          * is on -- will be called from NMI context with the PMU generates
 399          * NMIs.
 400          *
 401          * ->stop() with PERF_EF_UPDATE will read the counter and update
 402          *  period/count values like ->read() would.
 403          *
 404          * ->start() with PERF_EF_RELOAD will reprogram the counter
 405          *  value, must be preceded by a ->stop() with PERF_EF_UPDATE.
 406          */
 407         void (*start)                   (struct perf_event *event, int flags);
 408         void (*stop)                    (struct perf_event *event, int flags);
 409
 410         /*
 411          * Updates the counter value of the event.
 412          *
 413          * For sampling capable PMUs this will also update the software period
 414          * hw_perf_event::period_left field.
 415          */
 416         void (*read)                    (struct perf_event *event);
 417
 418         /*
 419          * Group events scheduling is treated as a transaction, add
 420          * group events as a whole and perform one schedulability test.
 421          * If the test fails, roll back the whole group
 422          *
 423          * Start the transaction, after this ->add() doesn't need to
 424          * do schedulability tests.
 425          *
 426          * Optional.
 427          */
 428         void (*start_txn)               (struct pmu *pmu, unsigned int txn_flags);
 429         /*
 430          * If ->start_txn() disabled the ->add() schedulability test
 431          * then ->commit_txn() is required to perform one. On success
 432          * the transaction is closed. On error the transaction is kept
 433          * open until ->cancel_txn() is called.
 434          *
 435          * Optional.
 436          */
 437         int  (*commit_txn)              (struct pmu *pmu);
 438         /*
 439          * Will cancel the transaction, assumes ->del() is called
 440          * for each successful ->add() during the transaction.
 441          *
 442          * Optional.
 443          */
 444         void (*cancel_txn)              (struct pmu *pmu);
 445
 446         /*
 447          * Will return the value for perf_event_mmap_page::index for this event,
 448          * if no implementation is provided it will default to: event->hw.idx + 1.
 449          */
 450         int (*event_idx)                (struct perf_event *event); /*optional */
 451
 452         /*
 453          * context-switches callback
 454          */
 455         void (*sched_task)              (struct perf_event_pmu_context *pmu_ctx,
 456                                         bool sched_in);
 457
 458         /*
 459          * Kmem cache of PMU specific data
 460          */
 461         struct kmem_cache               *task_ctx_cache;
 462
 463         /*
 464          * PMU specific parts of task perf event context (i.e. ctx->task_ctx_data)
 465          * can be synchronized using this function. See Intel LBR callstack support
 466          * implementation and Perf core context switch handling callbacks for usage
 467          * examples.
 468          */
 469         void (*swap_task_ctx)           (struct perf_event_pmu_context *prev_epc,
 470                                          struct perf_event_pmu_context *next_epc);
 471                                         /* optional */
 472
 473         /*
 474          * Set up pmu-private data structures for an AUX area
 475          */
 476         void *(*setup_aux)              (struct perf_event *event, void **pages,
 477                                          int nr_pages, bool overwrite);
 478                                         /* optional */
 479
 480         /*
 481          * Free pmu-private AUX data structures
 482          */
 483         void (*free_aux)                (void *aux); /* optional */
 484
 485         /*
 486          * Take a snapshot of the AUX buffer without touching the event
 487          * state, so that preempting ->start()/->stop() callbacks does
 488          * not interfere with their logic. Called in PMI context.
 489          *
 490          * Returns the size of AUX data copied to the output handle.
 491          *
 492          * Optional.
 493          */
 494         long (*snapshot_aux)            (struct perf_event *event,
 495                                          struct perf_output_handle *handle,
 496                                          unsigned long size);
 497
 498         /*
 499          * Validate address range filters: make sure the HW supports the
 500          * requested configuration and number of filters; return 0 if the
 501          * supplied filters are valid, -errno otherwise.
 502          *
 503          * Runs in the context of the ioctl()ing process and is not serialized
 504          * with the rest of the PMU callbacks.
 505          */
 506         int (*addr_filters_validate)    (struct list_head *filters);
 507                                         /* optional */
 508
 509         /*
 510          * Synchronize address range filter configuration:
 511          * translate hw-agnostic filters into hardware configuration in
 512          * event::hw::addr_filters.
 513          *
 514          * Runs as a part of filter sync sequence that is done in ->start()
 515          * callback by calling perf_event_addr_filters_sync().
 516          *
 517          * May (and should) traverse event::addr_filters::list, for which its
 518          * caller provides necessary serialization.
 519          */
 520         void (*addr_filters_sync)       (struct perf_event *event);
 521                                         /* optional */
 522
 523         /*
 524          * Check if event can be used for aux_output purposes for
 525          * events of this PMU.
 526          *
 527          * Runs from perf_event_open(). Should return 0 for "no match"
 528          * or non-zero for "match".
 529          */
 530         int (*aux_output_match)         (struct perf_event *event);
 531                                         /* optional */
 532
 533         /*
 534          * Skip programming this PMU on the given CPU. Typically needed for
 535          * big.LITTLE things.
 536          */
 537         bool (*filter)                  (struct pmu *pmu, int cpu); /* optional */
 538
 539         /*
 540          * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
 541          */
 542         int (*check_period)             (struct perf_event *event, u64 value); /* optional */
 543 };
 544
 545 enum perf_addr_filter_action_t {
 546         PERF_ADDR_FILTER_ACTION_STOP = 0,
 547         PERF_ADDR_FILTER_ACTION_START,
 548         PERF_ADDR_FILTER_ACTION_FILTER,
 549 };
 550
 551 /**
 552  * struct perf_addr_filter - address range filter definition
 553  * @entry:      event's filter list linkage
 554  * @path:       object file's path for file-based filters
 555  * @offset:     filter range offset
 556  * @size:       filter range size (size==0 means single address trigger)
 557  * @action:     filter/start/stop
 558  *
 559  * This is a hardware-agnostic filter configuration as specified by the user.
 560  */
 561 struct perf_addr_filter {
 562         struct list_head        entry;
 563         struct path             path;
 564         unsigned long           offset;
 565         unsigned long           size;
 566         enum perf_addr_filter_action_t  action;
 567 };
 568
 569 /**
 570  * struct perf_addr_filters_head - container for address range filters
 571  * @list:       list of filters for this event
 572  * @lock:       spinlock that serializes accesses to the @list and event's
 573  *              (and its children's) filter generations.
 574  * @nr_file_filters:    number of file-based filters
 575  *
 576  * A child event will use parent's @list (and therefore @lock), so they are
 577  * bundled together; see perf_event_addr_filters().
 578  */
 579 struct perf_addr_filters_head {
 580         struct list_head        list;
 581         raw_spinlock_t          lock;
 582         unsigned int            nr_file_filters;
 583 };
 584
 585 struct perf_addr_filter_range {
 586         unsigned long           start;
 587         unsigned long           size;
 588 };
 589
 590 /**
 591  * enum perf_event_state - the states of an event:
 592  */
 593 enum perf_event_state {
 594         PERF_EVENT_STATE_DEAD           = -4,
 595         PERF_EVENT_STATE_EXIT           = -3,
 596         PERF_EVENT_STATE_ERROR          = -2,
 597         PERF_EVENT_STATE_OFF            = -1,
 598         PERF_EVENT_STATE_INACTIVE       =  0,
 599         PERF_EVENT_STATE_ACTIVE         =  1,
 600 };
 601
 602 struct file;
 603 struct perf_sample_data;
 604
 605 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 606                                         struct perf_sample_data *,
 607                                         struct pt_regs *regs);
 608
 609 /*
 610  * Event capabilities. For event_caps and groups caps.
 611  *
 612  * PERF_EV_CAP_SOFTWARE: Is a software event.
 613  * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read
 614  * from any CPU in the package where it is active.
 615  * PERF_EV_CAP_SIBLING: An event with this flag must be a group sibling and
 616  * cannot be a group leader. If an event with this flag is detached from the
 617  * group it is scheduled out and moved into an unrecoverable ERROR state.
 618  */
 619 #define PERF_EV_CAP_SOFTWARE            BIT(0)
 620 #define PERF_EV_CAP_READ_ACTIVE_PKG     BIT(1)
 621 #define PERF_EV_CAP_SIBLING             BIT(2)
 622
 623 #define SWEVENT_HLIST_BITS              8
 624 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 625
 626 struct swevent_hlist {
 627         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 628         struct rcu_head                 rcu_head;
 629 };
 630
 631 #define PERF_ATTACH_CONTEXT     0x01
 632 #define PERF_ATTACH_GROUP       0x02
 633 #define PERF_ATTACH_TASK        0x04
 634 #define PERF_ATTACH_TASK_DATA   0x08
 635 #define PERF_ATTACH_ITRACE      0x10
 636 #define PERF_ATTACH_SCHED_CB    0x20
 637 #define PERF_ATTACH_CHILD       0x40
 638
 639 struct bpf_prog;
 640 struct perf_cgroup;
 641 struct perf_buffer;
 642
 643 struct pmu_event_list {
 644         raw_spinlock_t          lock;
 645         struct list_head        list;
 646 };
 647
 648 /*
 649  * event->sibling_list is modified whole holding both ctx->lock and ctx->mutex
 650  * as such iteration must hold either lock. However, since ctx->lock is an IRQ
 651  * safe lock, and is only held by the CPU doing the modification, having IRQs
 652  * disabled is sufficient since it will hold-off the IPIs.
 653  */
 654 #ifdef CONFIG_PROVE_LOCKING
 655 #define lockdep_assert_event_ctx(event)                         \
 656         WARN_ON_ONCE(__lockdep_enabled &&                       \
 657                      (this_cpu_read(hardirqs_enabled) &&        \
 658                       lockdep_is_held(&(event)->ctx->mutex) != LOCK_STATE_HELD))
 659 #else
 660 #define lockdep_assert_event_ctx(event)
 661 #endif
 662
 663 #define for_each_sibling_event(sibling, event)                  \
 664         lockdep_assert_event_ctx(event);                        \
 665         if ((event)->group_leader == (event))                   \
 666                 list_for_each_entry((sibling), &(event)->sibling_list, sibling_list)
 667
 668 /**
 669  * struct perf_event - performance event kernel representation:
 670  */
 671 struct perf_event {
 672 #ifdef CONFIG_PERF_EVENTS
 673         /*
 674          * entry onto perf_event_context::event_list;
 675          *   modifications require ctx->lock
 676          *   RCU safe iterations.
 677          */
 678         struct list_head                event_entry;
 679
 680         /*
 681          * Locked for modification by both ctx->mutex and ctx->lock; holding
 682          * either sufficies for read.
 683          */
 684         struct list_head                sibling_list;
 685         struct list_head                active_list;
 686         /*
 687          * Node on the pinned or flexible tree located at the event context;
 688          */
 689         struct rb_node                  group_node;
 690         u64                             group_index;
 691         /*
 692          * We need storage to track the entries in perf_pmu_migrate_context; we
 693          * cannot use the event_entry because of RCU and we want to keep the
 694          * group in tact which avoids us using the other two entries.
 695          */
 696         struct list_head                migrate_entry;
 697
 698         struct hlist_node               hlist_entry;
 699         struct list_head                active_entry;
 700         int                             nr_siblings;
 701
 702         /* Not serialized. Only written during event initialization. */
 703         int                             event_caps;
 704         /* The cumulative AND of all event_caps for events in this group. */
 705         int                             group_caps;
 706
 707         struct perf_event               *group_leader;
 708         /*
 709          * event->pmu will always point to pmu in which this event belongs.
 710          * Whereas event->pmu_ctx->pmu may point to other pmu when group of
 711          * different pmu events is created.
 712          */
 713         struct pmu                      *pmu;
 714         void                            *pmu_private;
 715
 716         enum perf_event_state           state;
 717         unsigned int                    attach_state;
 718         local64_t                       count;
 719         atomic64_t                      child_count;
 720
 721         /*
 722          * These are the total time in nanoseconds that the event
 723          * has been enabled (i.e. eligible to run, and the task has
 724          * been scheduled in, if this is a per-task event)
 725          * and running (scheduled onto the CPU), respectively.
 726          */
 727         u64                             total_time_enabled;
 728         u64                             total_time_running;
 729         u64                             tstamp;
 730
 731         struct perf_event_attr          attr;
 732         u16                             header_size;
 733         u16                             id_header_size;
 734         u16                             read_size;
 735         struct hw_perf_event            hw;
 736
 737         struct perf_event_context       *ctx;
 738         /*
 739          * event->pmu_ctx points to perf_event_pmu_context in which the event
 740          * is added. This pmu_ctx can be of other pmu for sw event when that
 741          * sw event is part of a group which also contains non-sw events.
 742          */
 743         struct perf_event_pmu_context   *pmu_ctx;
 744         atomic_long_t                   refcount;
 745
 746         /*
 747          * These accumulate total time (in nanoseconds) that children
 748          * events have been enabled and running, respectively.
 749          */
 750         atomic64_t                      child_total_time_enabled;
 751         atomic64_t                      child_total_time_running;
 752
 753         /*
 754          * Protect attach/detach and child_list:
 755          */
 756         struct mutex                    child_mutex;
 757         struct list_head                child_list;
 758         struct perf_event               *parent;
 759
 760         int                             oncpu;
 761         int                             cpu;
 762
 763         struct list_head                owner_entry;
 764         struct task_struct              *owner;
 765
 766         /* mmap bits */
 767         struct mutex                    mmap_mutex;
 768         atomic_t                        mmap_count;
 769
 770         struct perf_buffer              *rb;
 771         struct list_head                rb_entry;
 772         unsigned long                   rcu_batches;
 773         int                             rcu_pending;
 774
 775         /* poll related */
 776         wait_queue_head_t               waitq;
 777         struct fasync_struct            *fasync;
 778
 779         /* delayed work for NMIs and such */
 780         unsigned int                    pending_wakeup;
 781         unsigned int                    pending_kill;
 782         unsigned int                    pending_disable;
 783         unsigned int                    pending_sigtrap;
 784         unsigned long                   pending_addr;   /* SIGTRAP */
 785         struct irq_work                 pending_irq;
 786         struct callback_head            pending_task;
 787         unsigned int                    pending_work;
 788
 789         atomic_t                        event_limit;
 790
 791         /* address range filters */
 792         struct perf_addr_filters_head   addr_filters;
 793         /* vma address array for file-based filders */
 794         struct perf_addr_filter_range   *addr_filter_ranges;
 795         unsigned long                   addr_filters_gen;
 796
 797         /* for aux_output events */
 798         struct perf_event               *aux_event;
 799
 800         void (*destroy)(struct perf_event *);
 801         struct rcu_head                 rcu_head;
 802
 803         struct pid_namespace            *ns;
 804         u64                             id;
 805
 806         atomic64_t                      lost_samples;
 807
 808         u64                             (*clock)(void);
 809         perf_overflow_handler_t         overflow_handler;
 810         void                            *overflow_handler_context;
 811 #ifdef CONFIG_BPF_SYSCALL
 812         perf_overflow_handler_t         orig_overflow_handler;
 813         struct bpf_prog                 *prog;
 814         u64                             bpf_cookie;
 815 #endif
 816
 817 #ifdef CONFIG_EVENT_TRACING
 818         struct trace_event_call         *tp_event;
 819         struct event_filter             *filter;
 820 #ifdef CONFIG_FUNCTION_TRACER
 821         struct ftrace_ops               ftrace_ops;
 822 #endif
 823 #endif
 824
 825 #ifdef CONFIG_CGROUP_PERF
 826         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 827 #endif
 828
 829 #ifdef CONFIG_SECURITY
 830         void *security;
 831 #endif
 832         struct list_head                sb_list;
 833
 834         /*
 835          * Certain events gets forwarded to another pmu internally by over-
 836          * writing kernel copy of event->attr.type without user being aware
 837          * of it. event->orig_type contains original 'type' requested by
 838          * user.
 839          */
 840         __u32                           orig_type;
 841 #endif /* CONFIG_PERF_EVENTS */
 842 };
 843
 844 /*
 845  *           ,-----------------------[1:n]----------------------.
 846  *           V                                                  V
 847  * perf_event_context <-[1:n]-> perf_event_pmu_context <--- perf_event
 848  *           ^                      ^     |                     |
 849  *           `--------[1:n]---------'     `-[n:1]-> pmu <-[1:n]-'
 850  *
 851  *
 852  * struct perf_event_pmu_context  lifetime is refcount based and RCU freed
 853  * (similar to perf_event_context). Locking is as if it were a member of
 854  * perf_event_context; specifically:
 855  *
 856  *   modification, both: ctx->mutex && ctx->lock
 857  *   reading, either:    ctx->mutex || ctx->lock
 858  *
 859  * There is one exception to this; namely put_pmu_ctx() isn't always called
 860  * with ctx->mutex held; this means that as long as we can guarantee the epc
 861  * has events the above rules hold.
 862  *
 863  * Specificially, sys_perf_event_open()'s group_leader case depends on
 864  * ctx->mutex pinning the configuration. Since we hold a reference on
 865  * group_leader (through the filedesc) it can't go away, therefore it's
 866  * associated pmu_ctx must exist and cannot change due to ctx->mutex.
 867  */
 868 struct perf_event_pmu_context {
 869         struct pmu                      *pmu;
 870         struct perf_event_context       *ctx;
 871
 872         struct list_head                pmu_ctx_entry;
 873
 874         struct list_head                pinned_active;
 875         struct list_head                flexible_active;
 876
 877         /* Used to avoid freeing per-cpu perf_event_pmu_context */
 878         unsigned int                    embedded : 1;
 879
 880         unsigned int                    nr_events;
 881
 882         atomic_t                        refcount; /* event <-> epc */
 883         struct rcu_head                 rcu_head;
 884
 885         void                            *task_ctx_data; /* pmu specific data */
 886         /*
 887          * Set when one or more (plausibly active) event can't be scheduled
 888          * due to pmu overcommit or pmu constraints, except tolerant to
 889          * events not necessary to be active due to scheduling constraints,
 890          * such as cgroups.
 891          */
 892         int                             rotate_necessary;
 893 };
 894
 895 struct perf_event_groups {
 896         struct rb_root  tree;
 897         u64             index;
 898 };
 899
 900
 901 /**
 902  * struct perf_event_context - event context structure
 903  *
 904  * Used as a container for task events and CPU events as well:
 905  */
 906 struct perf_event_context {
 907         /*
 908          * Protect the states of the events in the list,
 909          * nr_active, and the list:
 910          */
 911         raw_spinlock_t                  lock;
 912         /*
 913          * Protect the list of events.  Locking either mutex or lock
 914          * is sufficient to ensure the list doesn't change; to change
 915          * the list you need to lock both the mutex and the spinlock.
 916          */
 917         struct mutex                    mutex;
 918
 919         struct list_head                pmu_ctx_list;
 920         struct perf_event_groups        pinned_groups;
 921         struct perf_event_groups        flexible_groups;
 922         struct list_head                event_list;
 923
 924         int                             nr_events;
 925         int                             nr_user;
 926         int                             is_active;
 927
 928         int                             nr_task_data;
 929         int                             nr_stat;
 930         int                             nr_freq;
 931         int                             rotate_disable;
 932
 933         refcount_t                      refcount; /* event <-> ctx */
 934         struct task_struct              *task;
 935
 936         /*
 937          * Context clock, runs when context enabled.
 938          */
 939         u64                             time;
 940         u64                             timestamp;
 941         u64                             timeoffset;
 942
 943         /*
 944          * These fields let us detect when two contexts have both
 945          * been cloned (inherited) from a common ancestor.
 946          */
 947         struct perf_event_context       *parent_ctx;
 948         u64                             parent_gen;
 949         u64                             generation;
 950         int                             pin_count;
 951 #ifdef CONFIG_CGROUP_PERF
 952         int                             nr_cgroups;      /* cgroup evts */
 953 #endif
 954         struct rcu_head                 rcu_head;
 955
 956         /*
 957          * Sum (event->pending_sigtrap + event->pending_work)
 958          *
 959          * The SIGTRAP is targeted at ctx->task, as such it won't do changing
 960          * that until the signal is delivered.
 961          */
 962         local_t                         nr_pending;
 963 };
 964
 965 /*
 966  * Number of contexts where an event can trigger:
 967  *      task, softirq, hardirq, nmi.
 968  */
 969 #define PERF_NR_CONTEXTS        4
 970
 971 struct perf_cpu_pmu_context {
 972         struct perf_event_pmu_context   epc;
 973         struct perf_event_pmu_context   *task_epc;
 974
 975         struct list_head                sched_cb_entry;
 976         int                             sched_cb_usage;
 977
 978         int                             active_oncpu;
 979         int                             exclusive;
 980
 981         raw_spinlock_t                  hrtimer_lock;
 982         struct hrtimer                  hrtimer;
 983         ktime_t                         hrtimer_interval;
 984         unsigned int                    hrtimer_active;
 985 };
 986
 987 /**
 988  * struct perf_event_cpu_context - per cpu event context structure
 989  */
 990 struct perf_cpu_context {
 991         struct perf_event_context       ctx;
 992         struct perf_event_context       *task_ctx;
 993         int                             online;
 994
 995 #ifdef CONFIG_CGROUP_PERF
 996         struct perf_cgroup              *cgrp;
 997 #endif
 998
 999         /*
1000          * Per-CPU storage for iterators used in visit_groups_merge. The default
1001          * storage is of size 2 to hold the CPU and any CPU event iterators.
1002          */
1003         int                             heap_size;
1004         struct perf_event               **heap;
1005         struct perf_event               *heap_default[2];
1006 };
1007
1008 struct perf_output_handle {
1009         struct perf_event               *event;
1010         struct perf_buffer              *rb;
1011         unsigned long                   wakeup;
1012         unsigned long                   size;
1013         u64                             aux_flags;
1014         union {
1015                 void                    *addr;
1016                 unsigned long           head;
1017         };
1018         int                             page;
1019 };
1020
1021 struct bpf_perf_event_data_kern {
1022         bpf_user_pt_regs_t *regs;
1023         struct perf_sample_data *data;
1024         struct perf_event *event;
1025 };
1026
1027 #ifdef CONFIG_CGROUP_PERF
1028
1029 /*
1030  * perf_cgroup_info keeps track of time_enabled for a cgroup.
1031  * This is a per-cpu dynamically allocated data structure.
1032  */
1033 struct perf_cgroup_info {
1034         u64                             time;
1035         u64                             timestamp;
1036         u64                             timeoffset;
1037         int                             active;
1038 };
1039
1040 struct perf_cgroup {
1041         struct cgroup_subsys_state      css;
1042         struct perf_cgroup_info __percpu *info;
1043 };
1044
1045 /*
1046  * Must ensure cgroup is pinned (css_get) before calling
1047  * this function. In other words, we cannot call this function
1048  * if there is no cgroup event for the current CPU context.
1049  */
1050 static inline struct perf_cgroup *
1051 perf_cgroup_from_task(struct task_struct *task, struct perf_event_context *ctx)
1052 {
1053         return container_of(task_css_check(task, perf_event_cgrp_id,
1054                                            ctx ? lockdep_is_held(&ctx->lock)
1055                                                : true),
1056                             struct perf_cgroup, css);
1057 }
1058 #endif /* CONFIG_CGROUP_PERF */
1059
1060 #ifdef CONFIG_PERF_EVENTS
1061
1062 extern struct perf_event_context *perf_cpu_task_ctx(void);
1063
1064 extern void *perf_aux_output_begin(struct perf_output_handle *handle,
1065                                    struct perf_event *event);
1066 extern void perf_aux_output_end(struct perf_output_handle *handle,
1067                                 unsigned long size);
1068 extern int perf_aux_output_skip(struct perf_output_handle *handle,
1069                                 unsigned long size);
1070 extern void *perf_get_aux(struct perf_output_handle *handle);
1071 extern void perf_aux_output_flag(struct perf_output_handle *handle, u64 flags);
1072 extern void perf_event_itrace_started(struct perf_event *event);
1073
1074 extern int perf_pmu_register(struct pmu *pmu, const char *name, int type);
1075 extern void perf_pmu_unregister(struct pmu *pmu);
1076
1077 extern void __perf_event_task_sched_in(struct task_struct *prev,
1078                                        struct task_struct *task);
1079 extern void __perf_event_task_sched_out(struct task_struct *prev,
1080                                         struct task_struct *next);
1081 extern int perf_event_init_task(struct task_struct *child, u64 clone_flags);
1082 extern void perf_event_exit_task(struct task_struct *child);
1083 extern void perf_event_free_task(struct task_struct *task);
1084 extern void perf_event_delayed_put(struct task_struct *task);
1085 extern struct file *perf_event_get(unsigned int fd);
1086 extern const struct perf_event *perf_get_event(struct file *file);
1087 extern const struct perf_event_attr *perf_event_attrs(struct perf_event *event);
1088 extern void perf_event_print_debug(void);
1089 extern void perf_pmu_disable(struct pmu *pmu);
1090 extern void perf_pmu_enable(struct pmu *pmu);
1091 extern void perf_sched_cb_dec(struct pmu *pmu);
1092 extern void perf_sched_cb_inc(struct pmu *pmu);
1093 extern int perf_event_task_disable(void);
1094 extern int perf_event_task_enable(void);
1095
1096 extern void perf_pmu_resched(struct pmu *pmu);
1097
1098 extern int perf_event_refresh(struct perf_event *event, int refresh);
1099 extern void perf_event_update_userpage(struct perf_event *event);
1100 extern int perf_event_release_kernel(struct perf_event *event);
1101 extern struct perf_event *
1102 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1103                                 int cpu,
1104                                 struct task_struct *task,
1105                                 perf_overflow_handler_t callback,
1106                                 void *context);
1107 extern void perf_pmu_migrate_context(struct pmu *pmu,
1108                                 int src_cpu, int dst_cpu);
1109 int perf_event_read_local(struct perf_event *event, u64 *value,
1110                           u64 *enabled, u64 *running);
1111 extern u64 perf_event_read_value(struct perf_event *event,
1112                                  u64 *enabled, u64 *running);
1113
1114 extern struct perf_callchain_entry *perf_callchain(struct perf_event *event, struct pt_regs *regs);
1115
1116 static inline bool branch_sample_no_flags(const struct perf_event *event)
1117 {
1118         return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_FLAGS;
1119 }
1120
1121 static inline bool branch_sample_no_cycles(const struct perf_event *event)
1122 {
1123         return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_NO_CYCLES;
1124 }
1125
1126 static inline bool branch_sample_type(const struct perf_event *event)
1127 {
1128         return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_TYPE_SAVE;
1129 }
1130
1131 static inline bool branch_sample_hw_index(const struct perf_event *event)
1132 {
1133         return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX;
1134 }
1135
1136 static inline bool branch_sample_priv(const struct perf_event *event)
1137 {
1138         return event->attr.branch_sample_type & PERF_SAMPLE_BRANCH_PRIV_SAVE;
1139 }
1140
1141
1142 struct perf_sample_data {
1143         /*
1144          * Fields set by perf_sample_data_init() unconditionally,
1145          * group so as to minimize the cachelines touched.
1146          */
1147         u64                             sample_flags;
1148         u64                             period;
1149         u64                             dyn_size;
1150
1151         /*
1152          * Fields commonly set by __perf_event_header__init_id(),
1153          * group so as to minimize the cachelines touched.
1154          */
1155         u64                             type;
1156         struct {
1157                 u32     pid;
1158                 u32     tid;
1159         }                               tid_entry;
1160         u64                             time;
1161         u64                             id;
1162         struct {
1163                 u32     cpu;
1164                 u32     reserved;
1165         }                               cpu_entry;
1166
1167         /*
1168          * The other fields, optionally {set,used} by
1169          * perf_{prepare,output}_sample().
1170          */
1171         u64                             ip;
1172         struct perf_callchain_entry     *callchain;
1173         struct perf_raw_record          *raw;
1174         struct perf_branch_stack        *br_stack;
1175         union perf_sample_weight        weight;
1176         union  perf_mem_data_src        data_src;
1177         u64                             txn;
1178
1179         struct perf_regs                regs_user;
1180         struct perf_regs                regs_intr;
1181         u64                             stack_user_size;
1182
1183         u64                             stream_id;
1184         u64                             cgroup;
1185         u64                             addr;
1186         u64                             phys_addr;
1187         u64                             data_page_size;
1188         u64                             code_page_size;
1189         u64                             aux_size;
1190 } ____cacheline_aligned;
1191
1192 /* default value for data source */
1193 #define PERF_MEM_NA (PERF_MEM_S(OP, NA)   |\
1194                     PERF_MEM_S(LVL, NA)   |\
1195                     PERF_MEM_S(SNOOP, NA) |\
1196                     PERF_MEM_S(LOCK, NA)  |\
1197                     PERF_MEM_S(TLB, NA))
1198
1199 static inline void perf_sample_data_init(struct perf_sample_data *data,
1200                                          u64 addr, u64 period)
1201 {
1202         /* remaining struct members initialized in perf_prepare_sample() */
1203         data->sample_flags = PERF_SAMPLE_PERIOD;
1204         data->period = period;
1205         data->dyn_size = 0;
1206
1207         if (addr) {
1208                 data->addr = addr;
1209                 data->sample_flags |= PERF_SAMPLE_ADDR;
1210         }
1211 }
1212
1213 static inline void perf_sample_save_callchain(struct perf_sample_data *data,
1214                                               struct perf_event *event,
1215                                               struct pt_regs *regs)
1216 {
1217         int size = 1;
1218
1219         data->callchain = perf_callchain(event, regs);
1220         size += data->callchain->nr;
1221
1222         data->dyn_size += size * sizeof(u64);
1223         data->sample_flags |= PERF_SAMPLE_CALLCHAIN;
1224 }
1225
1226 static inline void perf_sample_save_raw_data(struct perf_sample_data *data,
1227                                              struct perf_raw_record *raw)
1228 {
1229         struct perf_raw_frag *frag = &raw->frag;
1230         u32 sum = 0;
1231         int size;
1232
1233         do {
1234                 sum += frag->size;
1235                 if (perf_raw_frag_last(frag))
1236                         break;
1237                 frag = frag->next;
1238         } while (1);
1239
1240         size = round_up(sum + sizeof(u32), sizeof(u64));
1241         raw->size = size - sizeof(u32);
1242         frag->pad = raw->size - sum;
1243
1244         data->raw = raw;
1245         data->dyn_size += size;
1246         data->sample_flags |= PERF_SAMPLE_RAW;
1247 }
1248
1249 static inline void perf_sample_save_brstack(struct perf_sample_data *data,
1250                                             struct perf_event *event,
1251                                             struct perf_branch_stack *brs)
1252 {
1253         int size = sizeof(u64); /* nr */
1254
1255         if (branch_sample_hw_index(event))
1256                 size += sizeof(u64);
1257         size += brs->nr * sizeof(struct perf_branch_entry);
1258
1259         data->br_stack = brs;
1260         data->dyn_size += size;
1261         data->sample_flags |= PERF_SAMPLE_BRANCH_STACK;
1262 }
1263
1264 static inline u32 perf_sample_data_size(struct perf_sample_data *data,
1265                                         struct perf_event *event)
1266 {
1267         u32 size = sizeof(struct perf_event_header);
1268
1269         size += event->header_size + event->id_header_size;
1270         size += data->dyn_size;
1271
1272         return size;
1273 }
1274
1275 /*
1276  * Clear all bitfields in the perf_branch_entry.
1277  * The to and from fields are not cleared because they are
1278  * systematically modified by caller.
1279  */
1280 static inline void perf_clear_branch_entry_bitfields(struct perf_branch_entry *br)
1281 {
1282         br->mispred = 0;
1283         br->predicted = 0;
1284         br->in_tx = 0;
1285         br->abort = 0;
1286         br->cycles = 0;
1287         br->type = 0;
1288         br->spec = PERF_BR_SPEC_NA;
1289         br->reserved = 0;
1290 }
1291
1292 extern void perf_output_sample(struct perf_output_handle *handle,
1293                                struct perf_event_header *header,
1294                                struct perf_sample_data *data,
1295                                struct perf_event *event);
1296 extern void perf_prepare_sample(struct perf_sample_data *data,
1297                                 struct perf_event *event,
1298                                 struct pt_regs *regs);
1299 extern void perf_prepare_header(struct perf_event_header *header,
1300                                 struct perf_sample_data *data,
1301                                 struct perf_event *event,
1302                                 struct pt_regs *regs);
1303
1304 extern int perf_event_overflow(struct perf_event *event,
1305                                  struct perf_sample_data *data,
1306                                  struct pt_regs *regs);
1307
1308 extern void perf_event_output_forward(struct perf_event *event,
1309                                      struct perf_sample_data *data,
1310                                      struct pt_regs *regs);
1311 extern void perf_event_output_backward(struct perf_event *event,
1312                                        struct perf_sample_data *data,
1313                                        struct pt_regs *regs);
1314 extern int perf_event_output(struct perf_event *event,
1315                              struct perf_sample_data *data,
1316                              struct pt_regs *regs);
1317
1318 static inline bool
1319 is_default_overflow_handler(struct perf_event *event)
1320 {
1321         if (likely(event->overflow_handler == perf_event_output_forward))
1322                 return true;
1323         if (unlikely(event->overflow_handler == perf_event_output_backward))
1324                 return true;
1325         return false;
1326 }
1327
1328 extern void
1329 perf_event_header__init_id(struct perf_event_header *header,
1330                            struct perf_sample_data *data,
1331                            struct perf_event *event);
1332 extern void
1333 perf_event__output_id_sample(struct perf_event *event,
1334                              struct perf_output_handle *handle,
1335                              struct perf_sample_data *sample);
1336
1337 extern void
1338 perf_log_lost_samples(struct perf_event *event, u64 lost);
1339
1340 static inline bool event_has_any_exclude_flag(struct perf_event *event)
1341 {
1342         struct perf_event_attr *attr = &event->attr;
1343
1344         return attr->exclude_idle || attr->exclude_user ||
1345                attr->exclude_kernel || attr->exclude_hv ||
1346                attr->exclude_guest || attr->exclude_host;
1347 }
1348
1349 static inline bool is_sampling_event(struct perf_event *event)
1350 {
1351         return event->attr.sample_period != 0;
1352 }
1353
1354 /*
1355  * Return 1 for a software event, 0 for a hardware event
1356  */
1357 static inline int is_software_event(struct perf_event *event)
1358 {
1359         return event->event_caps & PERF_EV_CAP_SOFTWARE;
1360 }
1361
1362 /*
1363  * Return 1 for event in sw context, 0 for event in hw context
1364  */
1365 static inline int in_software_context(struct perf_event *event)
1366 {
1367         return event->pmu_ctx->pmu->task_ctx_nr == perf_sw_context;
1368 }
1369
1370 static inline int is_exclusive_pmu(struct pmu *pmu)
1371 {
1372         return pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE;
1373 }
1374
1375 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1376
1377 extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64);
1378 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1379
1380 #ifndef perf_arch_fetch_caller_regs
1381 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1382 #endif
1383
1384 /*
1385  * When generating a perf sample in-line, instead of from an interrupt /
1386  * exception, we lack a pt_regs. This is typically used from software events
1387  * like: SW_CONTEXT_SWITCHES, SW_MIGRATIONS and the tie-in with tracepoints.
1388  *
1389  * We typically don't need a full set, but (for x86) do require:
1390  * - ip for PERF_SAMPLE_IP
1391  * - cs for user_mode() tests
1392  * - sp for PERF_SAMPLE_CALLCHAIN
1393  * - eflags for MISC bits and CALLCHAIN (see: perf_hw_regs())
1394  *
1395  * NOTE: assumes @regs is otherwise already 0 filled; this is important for
1396  * things like PERF_SAMPLE_REGS_INTR.
1397  */
1398 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1399 {
1400         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1401 }
1402
1403 static __always_inline void
1404 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1405 {
1406         if (static_key_false(&perf_swevent_enabled[event_id]))
1407                 __perf_sw_event(event_id, nr, regs, addr);
1408 }
1409
1410 DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]);
1411
1412 /*
1413  * 'Special' version for the scheduler, it hard assumes no recursion,
1414  * which is guaranteed by us not actually scheduling inside other swevents
1415  * because those disable preemption.
1416  */
1417 static __always_inline void __perf_sw_event_sched(u32 event_id, u64 nr, u64 addr)
1418 {
1419         struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]);
1420
1421         perf_fetch_caller_regs(regs);
1422         ___perf_sw_event(event_id, nr, regs, addr);
1423 }
1424
1425 extern struct static_key_false perf_sched_events;
1426
1427 static __always_inline bool __perf_sw_enabled(int swevt)
1428 {
1429         return static_key_false(&perf_swevent_enabled[swevt]);
1430 }
1431
1432 static inline void perf_event_task_migrate(struct task_struct *task)
1433 {
1434         if (__perf_sw_enabled(PERF_COUNT_SW_CPU_MIGRATIONS))
1435                 task->sched_migrated = 1;
1436 }
1437
1438 static inline void perf_event_task_sched_in(struct task_struct *prev,
1439                                             struct task_struct *task)
1440 {
1441         if (static_branch_unlikely(&perf_sched_events))
1442                 __perf_event_task_sched_in(prev, task);
1443
1444         if (__perf_sw_enabled(PERF_COUNT_SW_CPU_MIGRATIONS) &&
1445             task->sched_migrated) {
1446                 __perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1447                 task->sched_migrated = 0;
1448         }
1449 }
1450
1451 static inline void perf_event_task_sched_out(struct task_struct *prev,
1452                                              struct task_struct *next)
1453 {
1454         if (__perf_sw_enabled(PERF_COUNT_SW_CONTEXT_SWITCHES))
1455                 __perf_sw_event_sched(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 0);
1456
1457 #ifdef CONFIG_CGROUP_PERF
1458         if (__perf_sw_enabled(PERF_COUNT_SW_CGROUP_SWITCHES) &&
1459             perf_cgroup_from_task(prev, NULL) !=
1460             perf_cgroup_from_task(next, NULL))
1461                 __perf_sw_event_sched(PERF_COUNT_SW_CGROUP_SWITCHES, 1, 0);
1462 #endif
1463
1464         if (static_branch_unlikely(&perf_sched_events))
1465                 __perf_event_task_sched_out(prev, next);
1466 }
1467
1468 extern void perf_event_mmap(struct vm_area_struct *vma);
1469
1470 extern void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1471                                bool unregister, const char *sym);
1472 extern void perf_event_bpf_event(struct bpf_prog *prog,
1473                                  enum perf_bpf_event_type type,
1474                                  u16 flags);
1475
1476 #ifdef CONFIG_GUEST_PERF_EVENTS
1477 extern struct perf_guest_info_callbacks __rcu *perf_guest_cbs;
1478
1479 DECLARE_STATIC_CALL(__perf_guest_state, *perf_guest_cbs->state);
1480 DECLARE_STATIC_CALL(__perf_guest_get_ip, *perf_guest_cbs->get_ip);
1481 DECLARE_STATIC_CALL(__perf_guest_handle_intel_pt_intr, *perf_guest_cbs->handle_intel_pt_intr);
1482
1483 static inline unsigned int perf_guest_state(void)
1484 {
1485         return static_call(__perf_guest_state)();
1486 }
1487 static inline unsigned long perf_guest_get_ip(void)
1488 {
1489         return static_call(__perf_guest_get_ip)();
1490 }
1491 static inline unsigned int perf_guest_handle_intel_pt_intr(void)
1492 {
1493         return static_call(__perf_guest_handle_intel_pt_intr)();
1494 }
1495 extern void perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1496 extern void perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs);
1497 #else
1498 static inline unsigned int perf_guest_state(void)                { return 0; }
1499 static inline unsigned long perf_guest_get_ip(void)              { return 0; }
1500 static inline unsigned int perf_guest_handle_intel_pt_intr(void) { return 0; }
1501 #endif /* CONFIG_GUEST_PERF_EVENTS */
1502
1503 extern void perf_event_exec(void);
1504 extern void perf_event_comm(struct task_struct *tsk, bool exec);
1505 extern void perf_event_namespaces(struct task_struct *tsk);
1506 extern void perf_event_fork(struct task_struct *tsk);
1507 extern void perf_event_text_poke(const void *addr,
1508                                  const void *old_bytes, size_t old_len,
1509                                  const void *new_bytes, size_t new_len);
1510
1511 /* Callchains */
1512 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1513
1514 extern void perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1515 extern void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs);
1516 extern struct perf_callchain_entry *
1517 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
1518                    u32 max_stack, bool crosstask, bool add_mark);
1519 extern int get_callchain_buffers(int max_stack);
1520 extern void put_callchain_buffers(void);
1521 extern struct perf_callchain_entry *get_callchain_entry(int *rctx);
1522 extern void put_callchain_entry(int rctx);
1523
1524 extern int sysctl_perf_event_max_stack;
1525 extern int sysctl_perf_event_max_contexts_per_stack;
1526
1527 static inline int perf_callchain_store_context(struct perf_callchain_entry_ctx *ctx, u64 ip)
1528 {
1529         if (ctx->contexts < sysctl_perf_event_max_contexts_per_stack) {
1530                 struct perf_callchain_entry *entry = ctx->entry;
1531                 entry->ip[entry->nr++] = ip;
1532                 ++ctx->contexts;
1533                 return 0;
1534         } else {
1535                 ctx->contexts_maxed = true;
1536                 return -1; /* no more room, stop walking the stack */
1537         }
1538 }
1539
1540 static inline int perf_callchain_store(struct perf_callchain_entry_ctx *ctx, u64 ip)
1541 {
1542         if (ctx->nr < ctx->max_stack && !ctx->contexts_maxed) {
1543                 struct perf_callchain_entry *entry = ctx->entry;
1544                 entry->ip[entry->nr++] = ip;
1545                 ++ctx->nr;
1546                 return 0;
1547         } else {
1548                 return -1; /* no more room, stop walking the stack */
1549         }
1550 }
1551
1552 extern int sysctl_perf_event_paranoid;
1553 extern int sysctl_perf_event_mlock;
1554 extern int sysctl_perf_event_sample_rate;
1555 extern int sysctl_perf_cpu_time_max_percent;
1556
1557 extern void perf_sample_event_took(u64 sample_len_ns);
1558
1559 int perf_proc_update_handler(struct ctl_table *table, int write,
1560                 void *buffer, size_t *lenp, loff_t *ppos);
1561 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
1562                 void *buffer, size_t *lenp, loff_t *ppos);
1563 int perf_event_max_stack_handler(struct ctl_table *table, int write,
1564                 void *buffer, size_t *lenp, loff_t *ppos);
1565
1566 /* Access to perf_event_open(2) syscall. */
1567 #define PERF_SECURITY_OPEN              0
1568
1569 /* Finer grained perf_event_open(2) access control. */
1570 #define PERF_SECURITY_CPU               1
1571 #define PERF_SECURITY_KERNEL            2
1572 #define PERF_SECURITY_TRACEPOINT        3
1573
1574 static inline int perf_is_paranoid(void)
1575 {
1576         return sysctl_perf_event_paranoid > -1;
1577 }
1578
1579 static inline int perf_allow_kernel(struct perf_event_attr *attr)
1580 {
1581         if (sysctl_perf_event_paranoid > 1 && !perfmon_capable())
1582                 return -EACCES;
1583
1584         return security_perf_event_open(attr, PERF_SECURITY_KERNEL);
1585 }
1586
1587 static inline int perf_allow_cpu(struct perf_event_attr *attr)
1588 {
1589         if (sysctl_perf_event_paranoid > 0 && !perfmon_capable())
1590                 return -EACCES;
1591
1592         return security_perf_event_open(attr, PERF_SECURITY_CPU);
1593 }
1594
1595 static inline int perf_allow_tracepoint(struct perf_event_attr *attr)
1596 {
1597         if (sysctl_perf_event_paranoid > -1 && !perfmon_capable())
1598                 return -EPERM;
1599
1600         return security_perf_event_open(attr, PERF_SECURITY_TRACEPOINT);
1601 }
1602
1603 extern void perf_event_init(void);
1604 extern void perf_tp_event(u16 event_type, u64 count, void *record,
1605                           int entry_size, struct pt_regs *regs,
1606                           struct hlist_head *head, int rctx,
1607                           struct task_struct *task);
1608 extern void perf_bp_event(struct perf_event *event, void *data);
1609
1610 #ifndef perf_misc_flags
1611 # define perf_misc_flags(regs) \
1612                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1613 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1614 #endif
1615 #ifndef perf_arch_bpf_user_pt_regs
1616 # define perf_arch_bpf_user_pt_regs(regs) regs
1617 #endif
1618
1619 static inline bool has_branch_stack(struct perf_event *event)
1620 {
1621         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1622 }
1623
1624 static inline bool needs_branch_stack(struct perf_event *event)
1625 {
1626         return event->attr.branch_sample_type != 0;
1627 }
1628
1629 static inline bool has_aux(struct perf_event *event)
1630 {
1631         return event->pmu->setup_aux;
1632 }
1633
1634 static inline bool is_write_backward(struct perf_event *event)
1635 {
1636         return !!event->attr.write_backward;
1637 }
1638
1639 static inline bool has_addr_filter(struct perf_event *event)
1640 {
1641         return event->pmu->nr_addr_filters;
1642 }
1643
1644 /*
1645  * An inherited event uses parent's filters
1646  */
1647 static inline struct perf_addr_filters_head *
1648 perf_event_addr_filters(struct perf_event *event)
1649 {
1650         struct perf_addr_filters_head *ifh = &event->addr_filters;
1651
1652         if (event->parent)
1653                 ifh = &event->parent->addr_filters;
1654
1655         return ifh;
1656 }
1657
1658 extern void perf_event_addr_filters_sync(struct perf_event *event);
1659 extern void perf_report_aux_output_id(struct perf_event *event, u64 hw_id);
1660
1661 extern int perf_output_begin(struct perf_output_handle *handle,
1662                              struct perf_sample_data *data,
1663                              struct perf_event *event, unsigned int size);
1664 extern int perf_output_begin_forward(struct perf_output_handle *handle,
1665                                      struct perf_sample_data *data,
1666                                      struct perf_event *event,
1667                                      unsigned int size);
1668 extern int perf_output_begin_backward(struct perf_output_handle *handle,
1669                                       struct perf_sample_data *data,
1670                                       struct perf_event *event,
1671                                       unsigned int size);
1672
1673 extern void perf_output_end(struct perf_output_handle *handle);
1674 extern unsigned int perf_output_copy(struct perf_output_handle *handle,
1675                              const void *buf, unsigned int len);
1676 extern unsigned int perf_output_skip(struct perf_output_handle *handle,
1677                                      unsigned int len);
1678 extern long perf_output_copy_aux(struct perf_output_handle *aux_handle,
1679                                  struct perf_output_handle *handle,
1680                                  unsigned long from, unsigned long to);
1681 extern int perf_swevent_get_recursion_context(void);
1682 extern void perf_swevent_put_recursion_context(int rctx);
1683 extern u64 perf_swevent_set_period(struct perf_event *event);
1684 extern void perf_event_enable(struct perf_event *event);
1685 extern void perf_event_disable(struct perf_event *event);
1686 extern void perf_event_disable_local(struct perf_event *event);
1687 extern void perf_event_disable_inatomic(struct perf_event *event);
1688 extern void perf_event_task_tick(void);
1689 extern int perf_event_account_interrupt(struct perf_event *event);
1690 extern int perf_event_period(struct perf_event *event, u64 value);
1691 extern u64 perf_event_pause(struct perf_event *event, bool reset);
1692 #else /* !CONFIG_PERF_EVENTS: */
1693 static inline void *
1694 perf_aux_output_begin(struct perf_output_handle *handle,
1695                       struct perf_event *event)                         { return NULL; }
1696 static inline void
1697 perf_aux_output_end(struct perf_output_handle *handle, unsigned long size)
1698                                                                         { }
1699 static inline int
1700 perf_aux_output_skip(struct perf_output_handle *handle,
1701                      unsigned long size)                                { return -EINVAL; }
1702 static inline void *
1703 perf_get_aux(struct perf_output_handle *handle)                         { return NULL; }
1704 static inline void
1705 perf_event_task_migrate(struct task_struct *task)                       { }
1706 static inline void
1707 perf_event_task_sched_in(struct task_struct *prev,
1708                          struct task_struct *task)                      { }
1709 static inline void
1710 perf_event_task_sched_out(struct task_struct *prev,
1711                           struct task_struct *next)                     { }
1712 static inline int perf_event_init_task(struct task_struct *child,
1713                                        u64 clone_flags)                 { return 0; }
1714 static inline void perf_event_exit_task(struct task_struct *child)      { }
1715 static inline void perf_event_free_task(struct task_struct *task)       { }
1716 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1717 static inline struct file *perf_event_get(unsigned int fd)      { return ERR_PTR(-EINVAL); }
1718 static inline const struct perf_event *perf_get_event(struct file *file)
1719 {
1720         return ERR_PTR(-EINVAL);
1721 }
1722 static inline const struct perf_event_attr *perf_event_attrs(struct perf_event *event)
1723 {
1724         return ERR_PTR(-EINVAL);
1725 }
1726 static inline int perf_event_read_local(struct perf_event *event, u64 *value,
1727                                         u64 *enabled, u64 *running)
1728 {
1729         return -EINVAL;
1730 }
1731 static inline void perf_event_print_debug(void)                         { }
1732 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1733 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1734 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1735 {
1736         return -EINVAL;
1737 }
1738
1739 static inline void
1740 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1741 static inline void
1742 perf_bp_event(struct perf_event *event, void *data)                     { }
1743
1744 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1745
1746 typedef int (perf_ksymbol_get_name_f)(char *name, int name_len, void *data);
1747 static inline void perf_event_ksymbol(u16 ksym_type, u64 addr, u32 len,
1748                                       bool unregister, const char *sym) { }
1749 static inline void perf_event_bpf_event(struct bpf_prog *prog,
1750                                         enum perf_bpf_event_type type,
1751                                         u16 flags)                      { }
1752 static inline void perf_event_exec(void)                                { }
1753 static inline void perf_event_comm(struct task_struct *tsk, bool exec)  { }
1754 static inline void perf_event_namespaces(struct task_struct *tsk)       { }
1755 static inline void perf_event_fork(struct task_struct *tsk)             { }
1756 static inline void perf_event_text_poke(const void *addr,
1757                                         const void *old_bytes,
1758                                         size_t old_len,
1759                                         const void *new_bytes,
1760                                         size_t new_len)                 { }
1761 static inline void perf_event_init(void)                                { }
1762 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1763 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1764 static inline u64 perf_swevent_set_period(struct perf_event *event)     { return 0; }
1765 static inline void perf_event_enable(struct perf_event *event)          { }
1766 static inline void perf_event_disable(struct perf_event *event)         { }
1767 static inline int __perf_event_disable(void *info)                      { return -1; }
1768 static inline void perf_event_task_tick(void)                           { }
1769 static inline int perf_event_release_kernel(struct perf_event *event)   { return 0; }
1770 static inline int perf_event_period(struct perf_event *event, u64 value)
1771 {
1772         return -EINVAL;
1773 }
1774 static inline u64 perf_event_pause(struct perf_event *event, bool reset)
1775 {
1776         return 0;
1777 }
1778 #endif
1779
1780 #if defined(CONFIG_PERF_EVENTS) && defined(CONFIG_CPU_SUP_INTEL)
1781 extern void perf_restore_debug_store(void);
1782 #else
1783 static inline void perf_restore_debug_store(void)                       { }
1784 #endif
1785
1786 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1787
1788 struct perf_pmu_events_attr {
1789         struct device_attribute attr;
1790         u64 id;
1791         const char *event_str;
1792 };
1793
1794 struct perf_pmu_events_ht_attr {
1795         struct device_attribute                 attr;
1796         u64                                     id;
1797         const char                              *event_str_ht;
1798         const char                              *event_str_noht;
1799 };
1800
1801 struct perf_pmu_events_hybrid_attr {
1802         struct device_attribute                 attr;
1803         u64                                     id;
1804         const char                              *event_str;
1805         u64                                     pmu_type;
1806 };
1807
1808 struct perf_pmu_format_hybrid_attr {
1809         struct device_attribute                 attr;
1810         u64                                     pmu_type;
1811 };
1812
1813 ssize_t perf_event_sysfs_show(struct device *dev, struct device_attribute *attr,
1814                               char *page);
1815
1816 #define PMU_EVENT_ATTR(_name, _var, _id, _show)                         \
1817 static struct perf_pmu_events_attr _var = {                             \
1818         .attr = __ATTR(_name, 0444, _show, NULL),                       \
1819         .id   =  _id,                                                   \
1820 };
1821
1822 #define PMU_EVENT_ATTR_STRING(_name, _var, _str)                            \
1823 static struct perf_pmu_events_attr _var = {                                 \
1824         .attr           = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
1825         .id             = 0,                                                \
1826         .event_str      = _str,                                             \
1827 };
1828
1829 #define PMU_EVENT_ATTR_ID(_name, _show, _id)                            \
1830         (&((struct perf_pmu_events_attr[]) {                            \
1831                 { .attr = __ATTR(_name, 0444, _show, NULL),             \
1832                   .id = _id, }                                          \
1833         })[0].attr.attr)
1834
1835 #define PMU_FORMAT_ATTR_SHOW(_name, _format)                            \
1836 static ssize_t                                                          \
1837 _name##_show(struct device *dev,                                        \
1838                                struct device_attribute *attr,           \
1839                                char *page)                              \
1840 {                                                                       \
1841         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1842         return sprintf(page, _format "\n");                             \
1843 }                                                                       \
1844
1845 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1846         PMU_FORMAT_ATTR_SHOW(_name, _format)                            \
1847                                                                         \
1848 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1849
1850 /* Performance counter hotplug functions */
1851 #ifdef CONFIG_PERF_EVENTS
1852 int perf_event_init_cpu(unsigned int cpu);
1853 int perf_event_exit_cpu(unsigned int cpu);
1854 #else
1855 #define perf_event_init_cpu     NULL
1856 #define perf_event_exit_cpu     NULL
1857 #endif
1858
1859 extern void arch_perf_update_userpage(struct perf_event *event,
1860                                       struct perf_event_mmap_page *userpg,
1861                                       u64 now);
1862
1863 #ifdef CONFIG_MMU
1864 extern __weak u64 arch_perf_get_page_size(struct mm_struct *mm, unsigned long addr);
1865 #endif
1866
1867 /*
1868  * Snapshot branch stack on software events.
1869  *
1870  * Branch stack can be very useful in understanding software events. For
1871  * example, when a long function, e.g. sys_perf_event_open, returns an
1872  * errno, it is not obvious why the function failed. Branch stack could
1873  * provide very helpful information in this type of scenarios.
1874  *
1875  * On software event, it is necessary to stop the hardware branch recorder
1876  * fast. Otherwise, the hardware register/buffer will be flushed with
1877  * entries of the triggering event. Therefore, static call is used to
1878  * stop the hardware recorder.
1879  */
1880
1881 /*
1882  * cnt is the number of entries allocated for entries.
1883  * Return number of entries copied to .
1884  */
1885 typedef int (perf_snapshot_branch_stack_t)(struct perf_branch_entry *entries,
1886                                            unsigned int cnt);
1887 DECLARE_STATIC_CALL(perf_snapshot_branch_stack, perf_snapshot_branch_stack_t);
1888
1889 #ifndef PERF_NEEDS_LOPWR_CB
1890 static inline void perf_lopwr_cb(bool mode)
1891 {
1892 }
1893 #endif
1894
1895 #endif /* _LINUX_PERF_EVENT_H */