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_GROUP       = 1U << 3,
 105         PERF_RECORD_CALLCHAIN   = 1U << 4,
 106 };
 107
 108 /*
 109  * Bits that can be set in hw_event.read_format to request that
 110  * reads on the counter should return the indicated quantities,
 111  * in increasing order of bit value, after the counter value.
 112  */
 113 enum perf_counter_read_format {
 114         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
 115         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
 116 };
 117
 118 /*
 119  * Hardware event to monitor via a performance monitoring counter:
 120  */
 121 struct perf_counter_hw_event {
 122         /*
 123          * The MSB of the config word signifies if the rest contains cpu
 124          * specific (raw) counter configuration data, if unset, the next
 125          * 7 bits are an event type and the rest of the bits are the event
 126          * identifier.
 127          */
 128         __u64                   config;
 129
 130         __u64                   irq_period;
 131         __u32                   record_type;
 132         __u32                   read_format;
 133
 134         __u64                   disabled       :  1, /* off by default        */
 135                                 nmi            :  1, /* NMI sampling          */
 136                                 inherit        :  1, /* children inherit it   */
 137                                 pinned         :  1, /* must always be on PMU */
 138                                 exclusive      :  1, /* only group on PMU     */
 139                                 exclude_user   :  1, /* don't count user      */
 140                                 exclude_kernel :  1, /* ditto kernel          */
 141                                 exclude_hv     :  1, /* ditto hypervisor      */
 142                                 exclude_idle   :  1, /* don't count when idle */
 143                                 mmap           :  1, /* include mmap data     */
 144                                 munmap         :  1, /* include munmap data   */
 145                                 comm           :  1, /* include comm data     */
 146
 147                                 __reserved_1   : 52;
 148
 149         __u32                   extra_config_len;
 150         __u32                   wakeup_events;  /* wakeup every n events */
 151
 152         __u64                   __reserved_2;
 153         __u64                   __reserved_3;
 154 };
 155
 156 /*
 157  * Ioctls that can be done on a perf counter fd:
 158  */
 159 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
 160 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
 161 #define PERF_COUNTER_IOC_REFRESH        _IOW('$', 2, u32)
 162
 163 /*
 164  * Structure of the page that can be mapped via mmap
 165  */
 166 struct perf_counter_mmap_page {
 167         __u32   version;                /* version number of this structure */
 168         __u32   compat_version;         /* lowest version this is compat with */
 169
 170         /*
 171          * Bits needed to read the hw counters in user-space.
 172          *
 173          *   u32 seq;
 174          *   s64 count;
 175          *
 176          *   do {
 177          *     seq = pc->lock;
 178          *
 179          *     barrier()
 180          *     if (pc->index) {
 181          *       count = pmc_read(pc->index - 1);
 182          *       count += pc->offset;
 183          *     } else
 184          *       goto regular_read;
 185          *
 186          *     barrier();
 187          *   } while (pc->lock != seq);
 188          *
 189          * NOTE: for obvious reason this only works on self-monitoring
 190          *       processes.
 191          */
 192         __u32   lock;                   /* seqlock for synchronization */
 193         __u32   index;                  /* hardware counter identifier */
 194         __s64   offset;                 /* add to hardware counter value */
 195
 196         /*
 197          * Control data for the mmap() data buffer.
 198          *
 199          * User-space reading this value should issue an rmb(), on SMP capable
 200          * platforms, after reading this value -- see perf_counter_wakeup().
 201          */
 202         __u32   data_head;              /* head in the data section */
 203 };
 204
 205 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
 206 #define PERF_EVENT_MISC_USER            (1 << 1)
 207 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
 208
 209 struct perf_event_header {
 210         __u32   type;
 211         __u16   misc;
 212         __u16   size;
 213 };
 214
 215 enum perf_event_type {
 216
 217         /*
 218          * The MMAP events record the PROT_EXEC mappings so that we can
 219          * correlate userspace IPs to code. They have the following structure:
 220          *
 221          * struct {
 222          *      struct perf_event_header        header;
 223          *
 224          *      u32                             pid, tid;
 225          *      u64                             addr;
 226          *      u64                             len;
 227          *      u64                             pgoff;
 228          *      char                            filename[];
 229          * };
 230          */
 231         PERF_EVENT_MMAP                 = 1,
 232         PERF_EVENT_MUNMAP               = 2,
 233
 234         /*
 235          * struct {
 236          *      struct perf_event_header        header;
 237          *
 238          *      u32                             pid, tid;
 239          *      char                            comm[];
 240          * };
 241          */
 242         PERF_EVENT_COMM                 = 3,
 243
 244         /*
 245          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
 246          * will be PERF_RECORD_*
 247          *
 248          * struct {
 249          *      struct perf_event_header        header;
 250          *
 251          *      { u64                   ip;       } && PERF_RECORD_IP
 252          *      { u32                   pid, tid; } && PERF_RECORD_TID
 253          *      { u64                   time;     } && PERF_RECORD_TIME
 254          *
 255          *      { u64                   nr;
 256          *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
 257          *
 258          *      { u16                   nr,
 259          *                              hv,
 260          *                              kernel,
 261          *                              user;
 262          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
 263          * };
 264          */
 265 };
 266
 267 #ifdef __KERNEL__
 268 /*
 269  * Kernel-internal data types and definitions:
 270  */
 271
 272 #ifdef CONFIG_PERF_COUNTERS
 273 # include <asm/perf_counter.h>
 274 #endif
 275
 276 #include <linux/list.h>
 277 #include <linux/mutex.h>
 278 #include <linux/rculist.h>
 279 #include <linux/rcupdate.h>
 280 #include <linux/spinlock.h>
 281 #include <linux/hrtimer.h>
 282 #include <linux/fs.h>
 283 #include <asm/atomic.h>
 284
 285 struct task_struct;
 286
 287 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 288 {
 289         return hw_event->config & PERF_COUNTER_RAW_MASK;
 290 }
 291
 292 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 293 {
 294         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 295 }
 296
 297 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 298 {
 299         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 300                 PERF_COUNTER_TYPE_SHIFT;
 301 }
 302
 303 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 304 {
 305         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 306 }
 307
 308 /**
 309  * struct hw_perf_counter - performance counter hardware details:
 310  */
 311 struct hw_perf_counter {
 312 #ifdef CONFIG_PERF_COUNTERS
 313         union {
 314                 struct { /* hardware */
 315                         u64                             config;
 316                         unsigned long                   config_base;
 317                         unsigned long                   counter_base;
 318                         int                             nmi;
 319                         unsigned int                    idx;
 320                 };
 321                 union { /* software */
 322                         atomic64_t                      count;
 323                         struct hrtimer                  hrtimer;
 324                 };
 325         };
 326         atomic64_t                      prev_count;
 327         u64                             irq_period;
 328         atomic64_t                      period_left;
 329 #endif
 330 };
 331
 332 struct perf_counter;
 333
 334 /**
 335  * struct hw_perf_counter_ops - performance counter hw ops
 336  */
 337 struct hw_perf_counter_ops {
 338         int (*enable)                   (struct perf_counter *counter);
 339         void (*disable)                 (struct perf_counter *counter);
 340         void (*read)                    (struct perf_counter *counter);
 341 };
 342
 343 /**
 344  * enum perf_counter_active_state - the states of a counter
 345  */
 346 enum perf_counter_active_state {
 347         PERF_COUNTER_STATE_ERROR        = -2,
 348         PERF_COUNTER_STATE_OFF          = -1,
 349         PERF_COUNTER_STATE_INACTIVE     =  0,
 350         PERF_COUNTER_STATE_ACTIVE       =  1,
 351 };
 352
 353 struct file;
 354
 355 struct perf_mmap_data {
 356         struct rcu_head                 rcu_head;
 357         int                             nr_pages;       /* nr of data pages  */
 358
 359         atomic_t                        wakeup;         /* POLL_ for wakeups */
 360         atomic_t                        head;           /* write position    */
 361         atomic_t                        events;         /* event limit       */
 362
 363         struct perf_counter_mmap_page   *user_page;
 364         void                            *data_pages[0];
 365 };
 366
 367 struct perf_pending_entry {
 368         struct perf_pending_entry *next;
 369         void (*func)(struct perf_pending_entry *);
 370 };
 371
 372 /**
 373  * struct perf_counter - performance counter kernel representation:
 374  */
 375 struct perf_counter {
 376 #ifdef CONFIG_PERF_COUNTERS
 377         struct list_head                list_entry;
 378         struct list_head                event_entry;
 379         struct list_head                sibling_list;
 380         int                             nr_siblings;
 381         struct perf_counter             *group_leader;
 382         const struct hw_perf_counter_ops *hw_ops;
 383
 384         enum perf_counter_active_state  state;
 385         enum perf_counter_active_state  prev_state;
 386         atomic64_t                      count;
 387
 388         /*
 389          * These are the total time in nanoseconds that the counter
 390          * has been enabled (i.e. eligible to run, and the task has
 391          * been scheduled in, if this is a per-task counter)
 392          * and running (scheduled onto the CPU), respectively.
 393          *
 394          * They are computed from tstamp_enabled, tstamp_running and
 395          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 396          */
 397         u64                             total_time_enabled;
 398         u64                             total_time_running;
 399
 400         /*
 401          * These are timestamps used for computing total_time_enabled
 402          * and total_time_running when the counter is in INACTIVE or
 403          * ACTIVE state, measured in nanoseconds from an arbitrary point
 404          * in time.
 405          * tstamp_enabled: the notional time when the counter was enabled
 406          * tstamp_running: the notional time when the counter was scheduled on
 407          * tstamp_stopped: in INACTIVE state, the notional time when the
 408          *      counter was scheduled off.
 409          */
 410         u64                             tstamp_enabled;
 411         u64                             tstamp_running;
 412         u64                             tstamp_stopped;
 413
 414         struct perf_counter_hw_event    hw_event;
 415         struct hw_perf_counter          hw;
 416
 417         struct perf_counter_context     *ctx;
 418         struct task_struct              *task;
 419         struct file                     *filp;
 420
 421         struct perf_counter             *parent;
 422         struct list_head                child_list;
 423
 424         /*
 425          * These accumulate total time (in nanoseconds) that children
 426          * counters have been enabled and running, respectively.
 427          */
 428         atomic64_t                      child_total_time_enabled;
 429         atomic64_t                      child_total_time_running;
 430
 431         /*
 432          * Protect attach/detach and child_list:
 433          */
 434         struct mutex                    mutex;
 435
 436         int                             oncpu;
 437         int                             cpu;
 438
 439         /* mmap bits */
 440         struct mutex                    mmap_mutex;
 441         atomic_t                        mmap_count;
 442         struct perf_mmap_data           *data;
 443
 444         /* poll related */
 445         wait_queue_head_t               waitq;
 446         struct fasync_struct            *fasync;
 447
 448         /* delayed work for NMIs and such */
 449         int                             pending_wakeup;
 450         int                             pending_kill;
 451         int                             pending_disable;
 452         struct perf_pending_entry       pending;
 453
 454         atomic_t                        event_limit;
 455
 456         void (*destroy)(struct perf_counter *);
 457         struct rcu_head                 rcu_head;
 458 #endif
 459 };
 460
 461 /**
 462  * struct perf_counter_context - counter context structure
 463  *
 464  * Used as a container for task counters and CPU counters as well:
 465  */
 466 struct perf_counter_context {
 467 #ifdef CONFIG_PERF_COUNTERS
 468         /*
 469          * Protect the states of the counters in the list,
 470          * nr_active, and the list:
 471          */
 472         spinlock_t              lock;
 473         /*
 474          * Protect the list of counters.  Locking either mutex or lock
 475          * is sufficient to ensure the list doesn't change; to change
 476          * the list you need to lock both the mutex and the spinlock.
 477          */
 478         struct mutex            mutex;
 479
 480         struct list_head        counter_list;
 481         struct list_head        event_list;
 482         int                     nr_counters;
 483         int                     nr_active;
 484         int                     is_active;
 485         struct task_struct      *task;
 486
 487         /*
 488          * Context clock, runs when context enabled.
 489          */
 490         u64                     time;
 491         u64                     timestamp;
 492 #endif
 493 };
 494
 495 /**
 496  * struct perf_counter_cpu_context - per cpu counter context structure
 497  */
 498 struct perf_cpu_context {
 499         struct perf_counter_context     ctx;
 500         struct perf_counter_context     *task_ctx;
 501         int                             active_oncpu;
 502         int                             max_pertask;
 503         int                             exclusive;
 504
 505         /*
 506          * Recursion avoidance:
 507          *
 508          * task, softirq, irq, nmi context
 509          */
 510         int                     recursion[4];
 511 };
 512
 513 /*
 514  * Set by architecture code:
 515  */
 516 extern int perf_max_counters;
 517
 518 #ifdef CONFIG_PERF_COUNTERS
 519 extern const struct hw_perf_counter_ops *
 520 hw_perf_counter_init(struct perf_counter *counter);
 521
 522 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 523 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 524 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 525 extern void perf_counter_init_task(struct task_struct *child);
 526 extern void perf_counter_exit_task(struct task_struct *child);
 527 extern void perf_counter_do_pending(void);
 528 extern void perf_counter_print_debug(void);
 529 extern void perf_counter_unthrottle(void);
 530 extern u64 hw_perf_save_disable(void);
 531 extern void hw_perf_restore(u64 ctrl);
 532 extern int perf_counter_task_disable(void);
 533 extern int perf_counter_task_enable(void);
 534 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 535                struct perf_cpu_context *cpuctx,
 536                struct perf_counter_context *ctx, int cpu);
 537 extern void perf_counter_update_userpage(struct perf_counter *counter);
 538
 539 extern int perf_counter_overflow(struct perf_counter *counter,
 540                                  int nmi, struct pt_regs *regs);
 541 /*
 542  * Return 1 for a software counter, 0 for a hardware counter
 543  */
 544 static inline int is_software_counter(struct perf_counter *counter)
 545 {
 546         return !perf_event_raw(&counter->hw_event) &&
 547                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 548 }
 549
 550 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 551
 552 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 553                               unsigned long pgoff, struct file *file);
 554
 555 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 556                                 unsigned long pgoff, struct file *file);
 557
 558 extern void perf_counter_comm(struct task_struct *tsk);
 559
 560 #define MAX_STACK_DEPTH         255
 561
 562 struct perf_callchain_entry {
 563         u16     nr, hv, kernel, user;
 564         u64     ip[MAX_STACK_DEPTH];
 565 };
 566
 567 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 568
 569 #else
 570 static inline void
 571 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 572 static inline void
 573 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 574 static inline void
 575 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 576 static inline void perf_counter_init_task(struct task_struct *child)    { }
 577 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 578 static inline void perf_counter_do_pending(void)                        { }
 579 static inline void perf_counter_print_debug(void)                       { }
 580 static inline void perf_counter_unthrottle(void)                        { }
 581 static inline void hw_perf_restore(u64 ctrl)                            { }
 582 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 583 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 584 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 585
 586 static inline void
 587 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)  { }
 588
 589
 590 static inline void
 591 perf_counter_mmap(unsigned long addr, unsigned long len,
 592                   unsigned long pgoff, struct file *file)               { }
 593
 594 static inline void
 595 perf_counter_munmap(unsigned long addr, unsigned long len,
 596                     unsigned long pgoff, struct file *file)             { }
 597
 598 static inline void perf_counter_comm(struct task_struct *tsk)           { }
 599 #endif
 600
 601 #endif /* __KERNEL__ */
 602 #endif /* _LINUX_PERF_COUNTER_H */