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