include/linux/perf_counter.h

   1 /*
   2  *  Performance counters:
   3  *
   4  *   Copyright(C) 2008, Thomas Gleixner <tglx@linutronix.de>
   5  *   Copyright(C) 2008, Red Hat, Inc., Ingo Molnar
   6  *
   7  *  Data type definitions, declarations, prototypes.
   8  *
   9  *  Started by: Thomas Gleixner and Ingo Molnar
  10  *
  11  *  For licencing details see kernel-base/COPYING
  12  */
  13 #ifndef _LINUX_PERF_COUNTER_H
  14 #define _LINUX_PERF_COUNTER_H
  15
  16 #include <linux/types.h>
  17 #include <linux/ioctl.h>
  18 #include <asm/byteorder.h>
  19
  20 /*
  21  * User-space ABI bits:
  22  */
  23
  24 /*
  25  * hw_event.type
  26  */
  27 enum perf_event_types {
  28         PERF_TYPE_HARDWARE              = 0,
  29         PERF_TYPE_SOFTWARE              = 1,
  30         PERF_TYPE_TRACEPOINT            = 2,
  31
  32         /*
  33          * available TYPE space, raw is the max value.
  34          */
  35
  36         PERF_TYPE_RAW                   = 128,
  37 };
  38
  39 /*
  40  * Generalized performance counter event types, used by the hw_event.event_id
  41  * parameter of the sys_perf_counter_open() syscall:
  42  */
  43 enum hw_event_ids {
  44         /*
  45          * Common hardware events, generalized by the kernel:
  46          */
  47         PERF_COUNT_CPU_CYCLES           = 0,
  48         PERF_COUNT_INSTRUCTIONS         = 1,
  49         PERF_COUNT_CACHE_REFERENCES     = 2,
  50         PERF_COUNT_CACHE_MISSES         = 3,
  51         PERF_COUNT_BRANCH_INSTRUCTIONS  = 4,
  52         PERF_COUNT_BRANCH_MISSES        = 5,
  53         PERF_COUNT_BUS_CYCLES           = 6,
  54
  55         PERF_HW_EVENTS_MAX              = 7,
  56 };
  57
  58 /*
  59  * Special "software" counters provided by the kernel, even if the hardware
  60  * does not support performance counters. These counters measure various
  61  * physical and sw events of the kernel (and allow the profiling of them as
  62  * well):
  63  */
  64 enum sw_event_ids {
  65         PERF_COUNT_CPU_CLOCK            = 0,
  66         PERF_COUNT_TASK_CLOCK           = 1,
  67         PERF_COUNT_PAGE_FAULTS          = 2,
  68         PERF_COUNT_CONTEXT_SWITCHES     = 3,
  69         PERF_COUNT_CPU_MIGRATIONS       = 4,
  70         PERF_COUNT_PAGE_FAULTS_MIN      = 5,
  71         PERF_COUNT_PAGE_FAULTS_MAJ      = 6,
  72
  73         PERF_SW_EVENTS_MAX              = 7,
  74 };
  75
  76 #define __PERF_COUNTER_MASK(name)                       \
  77         (((1ULL << PERF_COUNTER_##name##_BITS) - 1) <<  \
  78          PERF_COUNTER_##name##_SHIFT)
  79
  80 #define PERF_COUNTER_RAW_BITS           1
  81 #define PERF_COUNTER_RAW_SHIFT          63
  82 #define PERF_COUNTER_RAW_MASK           __PERF_COUNTER_MASK(RAW)
  83
  84 #define PERF_COUNTER_CONFIG_BITS        63
  85 #define PERF_COUNTER_CONFIG_SHIFT       0
  86 #define PERF_COUNTER_CONFIG_MASK        __PERF_COUNTER_MASK(CONFIG)
  87
  88 #define PERF_COUNTER_TYPE_BITS          7
  89 #define PERF_COUNTER_TYPE_SHIFT         56
  90 #define PERF_COUNTER_TYPE_MASK          __PERF_COUNTER_MASK(TYPE)
  91
  92 #define PERF_COUNTER_EVENT_BITS         56
  93 #define PERF_COUNTER_EVENT_SHIFT        0
  94 #define PERF_COUNTER_EVENT_MASK         __PERF_COUNTER_MASK(EVENT)
  95
  96 /*
  97  * Bits that can be set in hw_event.record_type to request information
  98  * in the overflow packets.
  99  */
 100 enum perf_counter_record_format {
 101         PERF_RECORD_IP          = 1U << 0,
 102         PERF_RECORD_TID         = 1U << 1,
 103         PERF_RECORD_TIME        = 1U << 2,
 104         PERF_RECORD_ADDR        = 1U << 3,
 105         PERF_RECORD_GROUP       = 1U << 4,
 106         PERF_RECORD_CALLCHAIN   = 1U << 5,
 107         PERF_RECORD_CONFIG      = 1U << 6,
 108         PERF_RECORD_CPU         = 1U << 7,
 109 };
 110
 111 /*
 112  * Bits that can be set in hw_event.read_format to request that
 113  * reads on the counter should return the indicated quantities,
 114  * in increasing order of bit value, after the counter value.
 115  */
 116 enum perf_counter_read_format {
 117         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
 118         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
 119 };
 120
 121 /*
 122  * Hardware event to monitor via a performance monitoring counter:
 123  */
 124 struct perf_counter_hw_event {
 125         /*
 126          * The MSB of the config word signifies if the rest contains cpu
 127          * specific (raw) counter configuration data, if unset, the next
 128          * 7 bits are an event type and the rest of the bits are the event
 129          * identifier.
 130          */
 131         __u64                   config;
 132
 133         union {
 134                 __u64           irq_period;
 135                 __u64           irq_freq;
 136         };
 137
 138         __u32                   record_type;
 139         __u32                   read_format;
 140
 141         __u64                   disabled       :  1, /* off by default        */
 142                                 nmi            :  1, /* NMI sampling          */
 143                                 inherit        :  1, /* children inherit it   */
 144                                 pinned         :  1, /* must always be on PMU */
 145                                 exclusive      :  1, /* only group on PMU     */
 146                                 exclude_user   :  1, /* don't count user      */
 147                                 exclude_kernel :  1, /* ditto kernel          */
 148                                 exclude_hv     :  1, /* ditto hypervisor      */
 149                                 exclude_idle   :  1, /* don't count when idle */
 150                                 mmap           :  1, /* include mmap data     */
 151                                 munmap         :  1, /* include munmap data   */
 152                                 comm           :  1, /* include comm data     */
 153                                 freq           :  1, /* use freq, not period  */
 154
 155                                 __reserved_1   : 51;
 156
 157         __u32                   wakeup_events;  /* wakeup every n events */
 158         __u32                   __reserved_2;
 159
 160         __u64                   __reserved_3;
 161         __u64                   __reserved_4;
 162 };
 163
 164 /*
 165  * Ioctls that can be done on a perf counter fd:
 166  */
 167 #define PERF_COUNTER_IOC_ENABLE         _IOW('$', 0, u32)
 168 #define PERF_COUNTER_IOC_DISABLE        _IOW('$', 1, u32)
 169 #define PERF_COUNTER_IOC_REFRESH        _IOW('$', 2, u32)
 170 #define PERF_COUNTER_IOC_RESET          _IOW('$', 3, u32)
 171
 172 enum perf_counter_ioc_flags {
 173         PERF_IOC_FLAG_GROUP             = 1U << 0,
 174 };
 175
 176 /*
 177  * Structure of the page that can be mapped via mmap
 178  */
 179 struct perf_counter_mmap_page {
 180         __u32   version;                /* version number of this structure */
 181         __u32   compat_version;         /* lowest version this is compat with */
 182
 183         /*
 184          * Bits needed to read the hw counters in user-space.
 185          *
 186          *   u32 seq;
 187          *   s64 count;
 188          *
 189          *   do {
 190          *     seq = pc->lock;
 191          *
 192          *     barrier()
 193          *     if (pc->index) {
 194          *       count = pmc_read(pc->index - 1);
 195          *       count += pc->offset;
 196          *     } else
 197          *       goto regular_read;
 198          *
 199          *     barrier();
 200          *   } while (pc->lock != seq);
 201          *
 202          * NOTE: for obvious reason this only works on self-monitoring
 203          *       processes.
 204          */
 205         __u32   lock;                   /* seqlock for synchronization */
 206         __u32   index;                  /* hardware counter identifier */
 207         __s64   offset;                 /* add to hardware counter value */
 208
 209         /*
 210          * Control data for the mmap() data buffer.
 211          *
 212          * User-space reading this value should issue an rmb(), on SMP capable
 213          * platforms, after reading this value -- see perf_counter_wakeup().
 214          */
 215         __u32   data_head;              /* head in the data section */
 216 };
 217
 218 #define PERF_EVENT_MISC_CPUMODE_MASK    (3 << 0)
 219 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0)
 220 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
 221 #define PERF_EVENT_MISC_USER            (2 << 0)
 222 #define PERF_EVENT_MISC_HYPERVISOR      (3 << 0)
 223 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
 224
 225 struct perf_event_header {
 226         __u32   type;
 227         __u16   misc;
 228         __u16   size;
 229 };
 230
 231 enum perf_event_type {
 232
 233         /*
 234          * The MMAP events record the PROT_EXEC mappings so that we can
 235          * correlate userspace IPs to code. They have the following structure:
 236          *
 237          * struct {
 238          *      struct perf_event_header        header;
 239          *
 240          *      u32                             pid, tid;
 241          *      u64                             addr;
 242          *      u64                             len;
 243          *      u64                             pgoff;
 244          *      char                            filename[];
 245          * };
 246          */
 247         PERF_EVENT_MMAP                 = 1,
 248         PERF_EVENT_MUNMAP               = 2,
 249
 250         /*
 251          * struct {
 252          *      struct perf_event_header        header;
 253          *
 254          *      u32                             pid, tid;
 255          *      char                            comm[];
 256          * };
 257          */
 258         PERF_EVENT_COMM                 = 3,
 259
 260         /*
 261          * struct {
 262          *      struct perf_event_header        header;
 263          *      u64                             time;
 264          *      u64                             irq_period;
 265          * };
 266          */
 267         PERF_EVENT_PERIOD               = 4,
 268
 269         /*
 270          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
 271          * will be PERF_RECORD_*
 272          *
 273          * struct {
 274          *      struct perf_event_header        header;
 275          *
 276          *      { u64                   ip;       } && PERF_RECORD_IP
 277          *      { u32                   pid, tid; } && PERF_RECORD_TID
 278          *      { u64                   time;     } && PERF_RECORD_TIME
 279          *      { u64                   addr;     } && PERF_RECORD_ADDR
 280          *      { u64                   config;   } && PERF_RECORD_CONFIG
 281          *      { u32                   cpu, res; } && PERF_RECORD_CPU
 282          *
 283          *      { u64                   nr;
 284          *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
 285          *
 286          *      { u16                   nr,
 287          *                              hv,
 288          *                              kernel,
 289          *                              user;
 290          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
 291          * };
 292          */
 293 };
 294
 295 #ifdef __KERNEL__
 296 /*
 297  * Kernel-internal data types and definitions:
 298  */
 299
 300 #ifdef CONFIG_PERF_COUNTERS
 301 # include <asm/perf_counter.h>
 302 #endif
 303
 304 #include <linux/list.h>
 305 #include <linux/mutex.h>
 306 #include <linux/rculist.h>
 307 #include <linux/rcupdate.h>
 308 #include <linux/spinlock.h>
 309 #include <linux/hrtimer.h>
 310 #include <linux/fs.h>
 311 #include <asm/atomic.h>
 312
 313 struct task_struct;
 314
 315 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 316 {
 317         return hw_event->config & PERF_COUNTER_RAW_MASK;
 318 }
 319
 320 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 321 {
 322         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 323 }
 324
 325 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 326 {
 327         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 328                 PERF_COUNTER_TYPE_SHIFT;
 329 }
 330
 331 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 332 {
 333         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 334 }
 335
 336 /**
 337  * struct hw_perf_counter - performance counter hardware details:
 338  */
 339 struct hw_perf_counter {
 340 #ifdef CONFIG_PERF_COUNTERS
 341         union {
 342                 struct { /* hardware */
 343                         u64                             config;
 344                         unsigned long                   config_base;
 345                         unsigned long                   counter_base;
 346                         int                             nmi;
 347                         int                             idx;
 348                 };
 349                 union { /* software */
 350                         atomic64_t                      count;
 351                         struct hrtimer                  hrtimer;
 352                 };
 353         };
 354         atomic64_t                      prev_count;
 355         u64                             irq_period;
 356         atomic64_t                      period_left;
 357         u64                             interrupts;
 358 #endif
 359 };
 360
 361 struct perf_counter;
 362
 363 /**
 364  * struct pmu - generic performance monitoring unit
 365  */
 366 struct pmu {
 367         int (*enable)                   (struct perf_counter *counter);
 368         void (*disable)                 (struct perf_counter *counter);
 369         void (*read)                    (struct perf_counter *counter);
 370 };
 371
 372 /**
 373  * enum perf_counter_active_state - the states of a counter
 374  */
 375 enum perf_counter_active_state {
 376         PERF_COUNTER_STATE_ERROR        = -2,
 377         PERF_COUNTER_STATE_OFF          = -1,
 378         PERF_COUNTER_STATE_INACTIVE     =  0,
 379         PERF_COUNTER_STATE_ACTIVE       =  1,
 380 };
 381
 382 struct file;
 383
 384 struct perf_mmap_data {
 385         struct rcu_head                 rcu_head;
 386         int                             nr_pages;       /* nr of data pages  */
 387         int                             nr_locked;      /* nr pages mlocked  */
 388
 389         atomic_t                        poll;           /* POLL_ for wakeups */
 390         atomic_t                        head;           /* write position    */
 391         atomic_t                        events;         /* event limit       */
 392
 393         atomic_t                        done_head;      /* completed head    */
 394         atomic_t                        lock;           /* concurrent writes */
 395
 396         atomic_t                        wakeup;         /* needs a wakeup    */
 397
 398         struct perf_counter_mmap_page   *user_page;
 399         void                            *data_pages[0];
 400 };
 401
 402 struct perf_pending_entry {
 403         struct perf_pending_entry *next;
 404         void (*func)(struct perf_pending_entry *);
 405 };
 406
 407 /**
 408  * struct perf_counter - performance counter kernel representation:
 409  */
 410 struct perf_counter {
 411 #ifdef CONFIG_PERF_COUNTERS
 412         struct list_head                list_entry;
 413         struct list_head                event_entry;
 414         struct list_head                sibling_list;
 415         int                             nr_siblings;
 416         struct perf_counter             *group_leader;
 417         const struct pmu                *pmu;
 418
 419         enum perf_counter_active_state  state;
 420         enum perf_counter_active_state  prev_state;
 421         atomic64_t                      count;
 422
 423         /*
 424          * These are the total time in nanoseconds that the counter
 425          * has been enabled (i.e. eligible to run, and the task has
 426          * been scheduled in, if this is a per-task counter)
 427          * and running (scheduled onto the CPU), respectively.
 428          *
 429          * They are computed from tstamp_enabled, tstamp_running and
 430          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 431          */
 432         u64                             total_time_enabled;
 433         u64                             total_time_running;
 434
 435         /*
 436          * These are timestamps used for computing total_time_enabled
 437          * and total_time_running when the counter is in INACTIVE or
 438          * ACTIVE state, measured in nanoseconds from an arbitrary point
 439          * in time.
 440          * tstamp_enabled: the notional time when the counter was enabled
 441          * tstamp_running: the notional time when the counter was scheduled on
 442          * tstamp_stopped: in INACTIVE state, the notional time when the
 443          *      counter was scheduled off.
 444          */
 445         u64                             tstamp_enabled;
 446         u64                             tstamp_running;
 447         u64                             tstamp_stopped;
 448
 449         struct perf_counter_hw_event    hw_event;
 450         struct hw_perf_counter          hw;
 451
 452         struct perf_counter_context     *ctx;
 453         struct file                     *filp;
 454
 455         /*
 456          * These accumulate total time (in nanoseconds) that children
 457          * counters have been enabled and running, respectively.
 458          */
 459         atomic64_t                      child_total_time_enabled;
 460         atomic64_t                      child_total_time_running;
 461
 462         /*
 463          * Protect attach/detach and child_list:
 464          */
 465         struct mutex                    child_mutex;
 466         struct list_head                child_list;
 467         struct perf_counter             *parent;
 468
 469         int                             oncpu;
 470         int                             cpu;
 471
 472         struct list_head                owner_entry;
 473         struct task_struct              *owner;
 474
 475         /* mmap bits */
 476         struct mutex                    mmap_mutex;
 477         atomic_t                        mmap_count;
 478         struct perf_mmap_data           *data;
 479
 480         /* poll related */
 481         wait_queue_head_t               waitq;
 482         struct fasync_struct            *fasync;
 483
 484         /* delayed work for NMIs and such */
 485         int                             pending_wakeup;
 486         int                             pending_kill;
 487         int                             pending_disable;
 488         struct perf_pending_entry       pending;
 489
 490         atomic_t                        event_limit;
 491
 492         void (*destroy)(struct perf_counter *);
 493         struct rcu_head                 rcu_head;
 494 #endif
 495 };
 496
 497 /**
 498  * struct perf_counter_context - counter context structure
 499  *
 500  * Used as a container for task counters and CPU counters as well:
 501  */
 502 struct perf_counter_context {
 503         /*
 504          * Protect the states of the counters in the list,
 505          * nr_active, and the list:
 506          */
 507         spinlock_t              lock;
 508         /*
 509          * Protect the list of counters.  Locking either mutex or lock
 510          * is sufficient to ensure the list doesn't change; to change
 511          * the list you need to lock both the mutex and the spinlock.
 512          */
 513         struct mutex            mutex;
 514
 515         struct list_head        counter_list;
 516         struct list_head        event_list;
 517         int                     nr_counters;
 518         int                     nr_active;
 519         int                     is_active;
 520         atomic_t                refcount;
 521         struct task_struct      *task;
 522
 523         /*
 524          * Context clock, runs when context enabled.
 525          */
 526         u64                     time;
 527         u64                     timestamp;
 528
 529         /*
 530          * These fields let us detect when two contexts have both
 531          * been cloned (inherited) from a common ancestor.
 532          */
 533         struct perf_counter_context *parent_ctx;
 534         u32                     parent_gen;
 535         u32                     generation;
 536 };
 537
 538 /**
 539  * struct perf_counter_cpu_context - per cpu counter context structure
 540  */
 541 struct perf_cpu_context {
 542         struct perf_counter_context     ctx;
 543         struct perf_counter_context     *task_ctx;
 544         int                             active_oncpu;
 545         int                             max_pertask;
 546         int                             exclusive;
 547
 548         /*
 549          * Recursion avoidance:
 550          *
 551          * task, softirq, irq, nmi context
 552          */
 553         int                     recursion[4];
 554 };
 555
 556 #ifdef CONFIG_PERF_COUNTERS
 557
 558 /*
 559  * Set by architecture code:
 560  */
 561 extern int perf_max_counters;
 562
 563 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
 564
 565 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 566 extern void perf_counter_task_sched_out(struct task_struct *task,
 567                                         struct task_struct *next, int cpu);
 568 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 569 extern int perf_counter_init_task(struct task_struct *child);
 570 extern void perf_counter_exit_task(struct task_struct *child);
 571 extern void perf_counter_do_pending(void);
 572 extern void perf_counter_print_debug(void);
 573 extern void __perf_disable(void);
 574 extern bool __perf_enable(void);
 575 extern void perf_disable(void);
 576 extern void perf_enable(void);
 577 extern int perf_counter_task_disable(void);
 578 extern int perf_counter_task_enable(void);
 579 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 580                struct perf_cpu_context *cpuctx,
 581                struct perf_counter_context *ctx, int cpu);
 582 extern void perf_counter_update_userpage(struct perf_counter *counter);
 583
 584 extern int perf_counter_overflow(struct perf_counter *counter,
 585                                  int nmi, struct pt_regs *regs, u64 addr);
 586 /*
 587  * Return 1 for a software counter, 0 for a hardware counter
 588  */
 589 static inline int is_software_counter(struct perf_counter *counter)
 590 {
 591         return !perf_event_raw(&counter->hw_event) &&
 592                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 593 }
 594
 595 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
 596
 597 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 598                               unsigned long pgoff, struct file *file);
 599
 600 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 601                                 unsigned long pgoff, struct file *file);
 602
 603 extern void perf_counter_comm(struct task_struct *tsk);
 604
 605 #define MAX_STACK_DEPTH         255
 606
 607 struct perf_callchain_entry {
 608         u16     nr, hv, kernel, user;
 609         u64     ip[MAX_STACK_DEPTH];
 610 };
 611
 612 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 613
 614 extern int sysctl_perf_counter_priv;
 615 extern int sysctl_perf_counter_mlock;
 616
 617 extern void perf_counter_init(void);
 618
 619 #ifndef perf_misc_flags
 620 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
 621                                  PERF_EVENT_MISC_KERNEL)
 622 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
 623 #endif
 624
 625 #else
 626 static inline void
 627 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 628 static inline void
 629 perf_counter_task_sched_out(struct task_struct *task,
 630                             struct task_struct *next, int cpu)          { }
 631 static inline void
 632 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 633 static inline int perf_counter_init_task(struct task_struct *child)     { }
 634 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 635 static inline void perf_counter_do_pending(void)                        { }
 636 static inline void perf_counter_print_debug(void)                       { }
 637 static inline void perf_disable(void)                                   { }
 638 static inline void perf_enable(void)                                    { }
 639 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 640 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 641
 642 static inline void
 643 perf_swcounter_event(u32 event, u64 nr, int nmi,
 644                      struct pt_regs *regs, u64 addr)                    { }
 645
 646 static inline void
 647 perf_counter_mmap(unsigned long addr, unsigned long len,
 648                   unsigned long pgoff, struct file *file)               { }
 649
 650 static inline void
 651 perf_counter_munmap(unsigned long addr, unsigned long len,
 652                     unsigned long pgoff, struct file *file)             { }
 653
 654 static inline void perf_counter_comm(struct task_struct *tsk)           { }
 655 static inline void perf_counter_init(void)                              { }
 656 #endif
 657
 658 #endif /* __KERNEL__ */
 659 #endif /* _LINUX_PERF_COUNTER_H */