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 };
 108
 109 /*
 110  * Bits that can be set in hw_event.read_format to request that
 111  * reads on the counter should return the indicated quantities,
 112  * in increasing order of bit value, after the counter value.
 113  */
 114 enum perf_counter_read_format {
 115         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
 116         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
 117 };
 118
 119 /*
 120  * Hardware event to monitor via a performance monitoring counter:
 121  */
 122 struct perf_counter_hw_event {
 123         /*
 124          * The MSB of the config word signifies if the rest contains cpu
 125          * specific (raw) counter configuration data, if unset, the next
 126          * 7 bits are an event type and the rest of the bits are the event
 127          * identifier.
 128          */
 129         __u64                   config;
 130
 131         __u64                   irq_period;
 132         __u32                   record_type;
 133         __u32                   read_format;
 134
 135         __u64                   disabled       :  1, /* off by default        */
 136                                 nmi            :  1, /* NMI sampling          */
 137                                 inherit        :  1, /* children inherit it   */
 138                                 pinned         :  1, /* must always be on PMU */
 139                                 exclusive      :  1, /* only group on PMU     */
 140                                 exclude_user   :  1, /* don't count user      */
 141                                 exclude_kernel :  1, /* ditto kernel          */
 142                                 exclude_hv     :  1, /* ditto hypervisor      */
 143                                 exclude_idle   :  1, /* don't count when idle */
 144                                 mmap           :  1, /* include mmap data     */
 145                                 munmap         :  1, /* include munmap data   */
 146                                 comm           :  1, /* include comm data     */
 147
 148                                 __reserved_1   : 52;
 149
 150         __u32                   extra_config_len;
 151         __u32                   wakeup_events;  /* wakeup every n events */
 152
 153         __u64                   __reserved_2;
 154         __u64                   __reserved_3;
 155 };
 156
 157 /*
 158  * Ioctls that can be done on a perf counter fd:
 159  */
 160 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
 161 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
 162 #define PERF_COUNTER_IOC_REFRESH        _IOW('$', 2, u32)
 163
 164 /*
 165  * Structure of the page that can be mapped via mmap
 166  */
 167 struct perf_counter_mmap_page {
 168         __u32   version;                /* version number of this structure */
 169         __u32   compat_version;         /* lowest version this is compat with */
 170
 171         /*
 172          * Bits needed to read the hw counters in user-space.
 173          *
 174          *   u32 seq;
 175          *   s64 count;
 176          *
 177          *   do {
 178          *     seq = pc->lock;
 179          *
 180          *     barrier()
 181          *     if (pc->index) {
 182          *       count = pmc_read(pc->index - 1);
 183          *       count += pc->offset;
 184          *     } else
 185          *       goto regular_read;
 186          *
 187          *     barrier();
 188          *   } while (pc->lock != seq);
 189          *
 190          * NOTE: for obvious reason this only works on self-monitoring
 191          *       processes.
 192          */
 193         __u32   lock;                   /* seqlock for synchronization */
 194         __u32   index;                  /* hardware counter identifier */
 195         __s64   offset;                 /* add to hardware counter value */
 196
 197         /*
 198          * Control data for the mmap() data buffer.
 199          *
 200          * User-space reading this value should issue an rmb(), on SMP capable
 201          * platforms, after reading this value -- see perf_counter_wakeup().
 202          */
 203         __u32   data_head;              /* head in the data section */
 204 };
 205
 206 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
 207 #define PERF_EVENT_MISC_USER            (1 << 1)
 208 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
 209
 210 struct perf_event_header {
 211         __u32   type;
 212         __u16   misc;
 213         __u16   size;
 214 };
 215
 216 enum perf_event_type {
 217
 218         /*
 219          * The MMAP events record the PROT_EXEC mappings so that we can
 220          * correlate userspace IPs to code. They have the following structure:
 221          *
 222          * struct {
 223          *      struct perf_event_header        header;
 224          *
 225          *      u32                             pid, tid;
 226          *      u64                             addr;
 227          *      u64                             len;
 228          *      u64                             pgoff;
 229          *      char                            filename[];
 230          * };
 231          */
 232         PERF_EVENT_MMAP                 = 1,
 233         PERF_EVENT_MUNMAP               = 2,
 234
 235         /*
 236          * struct {
 237          *      struct perf_event_header        header;
 238          *
 239          *      u32                             pid, tid;
 240          *      char                            comm[];
 241          * };
 242          */
 243         PERF_EVENT_COMM                 = 3,
 244
 245         /*
 246          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
 247          * will be PERF_RECORD_*
 248          *
 249          * struct {
 250          *      struct perf_event_header        header;
 251          *
 252          *      { u64                   ip;       } && PERF_RECORD_IP
 253          *      { u32                   pid, tid; } && PERF_RECORD_TID
 254          *      { u64                   time;     } && PERF_RECORD_TIME
 255          *      { u64                   addr;     } && PERF_RECORD_ADDR
 256          *
 257          *      { u64                   nr;
 258          *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
 259          *
 260          *      { u16                   nr,
 261          *                              hv,
 262          *                              kernel,
 263          *                              user;
 264          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
 265          * };
 266          */
 267 };
 268
 269 #ifdef __KERNEL__
 270 /*
 271  * Kernel-internal data types and definitions:
 272  */
 273
 274 #ifdef CONFIG_PERF_COUNTERS
 275 # include <asm/perf_counter.h>
 276 #endif
 277
 278 #include <linux/list.h>
 279 #include <linux/mutex.h>
 280 #include <linux/rculist.h>
 281 #include <linux/rcupdate.h>
 282 #include <linux/spinlock.h>
 283 #include <linux/hrtimer.h>
 284 #include <linux/fs.h>
 285 #include <asm/atomic.h>
 286
 287 struct task_struct;
 288
 289 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 290 {
 291         return hw_event->config & PERF_COUNTER_RAW_MASK;
 292 }
 293
 294 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 295 {
 296         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 297 }
 298
 299 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 300 {
 301         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 302                 PERF_COUNTER_TYPE_SHIFT;
 303 }
 304
 305 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 306 {
 307         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 308 }
 309
 310 /**
 311  * struct hw_perf_counter - performance counter hardware details:
 312  */
 313 struct hw_perf_counter {
 314 #ifdef CONFIG_PERF_COUNTERS
 315         union {
 316                 struct { /* hardware */
 317                         u64                             config;
 318                         unsigned long                   config_base;
 319                         unsigned long                   counter_base;
 320                         int                             nmi;
 321                         unsigned int                    idx;
 322                 };
 323                 union { /* software */
 324                         atomic64_t                      count;
 325                         struct hrtimer                  hrtimer;
 326                 };
 327         };
 328         atomic64_t                      prev_count;
 329         u64                             irq_period;
 330         atomic64_t                      period_left;
 331 #endif
 332 };
 333
 334 struct perf_counter;
 335
 336 /**
 337  * struct hw_perf_counter_ops - performance counter hw ops
 338  */
 339 struct hw_perf_counter_ops {
 340         int (*enable)                   (struct perf_counter *counter);
 341         void (*disable)                 (struct perf_counter *counter);
 342         void (*read)                    (struct perf_counter *counter);
 343 };
 344
 345 /**
 346  * enum perf_counter_active_state - the states of a counter
 347  */
 348 enum perf_counter_active_state {
 349         PERF_COUNTER_STATE_ERROR        = -2,
 350         PERF_COUNTER_STATE_OFF          = -1,
 351         PERF_COUNTER_STATE_INACTIVE     =  0,
 352         PERF_COUNTER_STATE_ACTIVE       =  1,
 353 };
 354
 355 struct file;
 356
 357 struct perf_mmap_data {
 358         struct rcu_head                 rcu_head;
 359         int                             nr_pages;       /* nr of data pages  */
 360
 361         atomic_t                        wakeup;         /* POLL_ for wakeups */
 362         atomic_t                        head;           /* write position    */
 363         atomic_t                        events;         /* event limit       */
 364
 365         struct perf_counter_mmap_page   *user_page;
 366         void                            *data_pages[0];
 367 };
 368
 369 struct perf_pending_entry {
 370         struct perf_pending_entry *next;
 371         void (*func)(struct perf_pending_entry *);
 372 };
 373
 374 /**
 375  * struct perf_counter - performance counter kernel representation:
 376  */
 377 struct perf_counter {
 378 #ifdef CONFIG_PERF_COUNTERS
 379         struct list_head                list_entry;
 380         struct list_head                event_entry;
 381         struct list_head                sibling_list;
 382         int                             nr_siblings;
 383         struct perf_counter             *group_leader;
 384         const struct hw_perf_counter_ops *hw_ops;
 385
 386         enum perf_counter_active_state  state;
 387         enum perf_counter_active_state  prev_state;
 388         atomic64_t                      count;
 389
 390         /*
 391          * These are the total time in nanoseconds that the counter
 392          * has been enabled (i.e. eligible to run, and the task has
 393          * been scheduled in, if this is a per-task counter)
 394          * and running (scheduled onto the CPU), respectively.
 395          *
 396          * They are computed from tstamp_enabled, tstamp_running and
 397          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 398          */
 399         u64                             total_time_enabled;
 400         u64                             total_time_running;
 401
 402         /*
 403          * These are timestamps used for computing total_time_enabled
 404          * and total_time_running when the counter is in INACTIVE or
 405          * ACTIVE state, measured in nanoseconds from an arbitrary point
 406          * in time.
 407          * tstamp_enabled: the notional time when the counter was enabled
 408          * tstamp_running: the notional time when the counter was scheduled on
 409          * tstamp_stopped: in INACTIVE state, the notional time when the
 410          *      counter was scheduled off.
 411          */
 412         u64                             tstamp_enabled;
 413         u64                             tstamp_running;
 414         u64                             tstamp_stopped;
 415
 416         struct perf_counter_hw_event    hw_event;
 417         struct hw_perf_counter          hw;
 418
 419         struct perf_counter_context     *ctx;
 420         struct task_struct              *task;
 421         struct file                     *filp;
 422
 423         struct perf_counter             *parent;
 424         struct list_head                child_list;
 425
 426         /*
 427          * These accumulate total time (in nanoseconds) that children
 428          * counters have been enabled and running, respectively.
 429          */
 430         atomic64_t                      child_total_time_enabled;
 431         atomic64_t                      child_total_time_running;
 432
 433         /*
 434          * Protect attach/detach and child_list:
 435          */
 436         struct mutex                    mutex;
 437
 438         int                             oncpu;
 439         int                             cpu;
 440
 441         /* mmap bits */
 442         struct mutex                    mmap_mutex;
 443         atomic_t                        mmap_count;
 444         struct perf_mmap_data           *data;
 445
 446         /* poll related */
 447         wait_queue_head_t               waitq;
 448         struct fasync_struct            *fasync;
 449
 450         /* delayed work for NMIs and such */
 451         int                             pending_wakeup;
 452         int                             pending_kill;
 453         int                             pending_disable;
 454         struct perf_pending_entry       pending;
 455
 456         atomic_t                        event_limit;
 457
 458         void (*destroy)(struct perf_counter *);
 459         struct rcu_head                 rcu_head;
 460 #endif
 461 };
 462
 463 /**
 464  * struct perf_counter_context - counter context structure
 465  *
 466  * Used as a container for task counters and CPU counters as well:
 467  */
 468 struct perf_counter_context {
 469 #ifdef CONFIG_PERF_COUNTERS
 470         /*
 471          * Protect the states of the counters in the list,
 472          * nr_active, and the list:
 473          */
 474         spinlock_t              lock;
 475         /*
 476          * Protect the list of counters.  Locking either mutex or lock
 477          * is sufficient to ensure the list doesn't change; to change
 478          * the list you need to lock both the mutex and the spinlock.
 479          */
 480         struct mutex            mutex;
 481
 482         struct list_head        counter_list;
 483         struct list_head        event_list;
 484         int                     nr_counters;
 485         int                     nr_active;
 486         int                     is_active;
 487         struct task_struct      *task;
 488
 489         /*
 490          * Context clock, runs when context enabled.
 491          */
 492         u64                     time;
 493         u64                     timestamp;
 494 #endif
 495 };
 496
 497 /**
 498  * struct perf_counter_cpu_context - per cpu counter context structure
 499  */
 500 struct perf_cpu_context {
 501         struct perf_counter_context     ctx;
 502         struct perf_counter_context     *task_ctx;
 503         int                             active_oncpu;
 504         int                             max_pertask;
 505         int                             exclusive;
 506
 507         /*
 508          * Recursion avoidance:
 509          *
 510          * task, softirq, irq, nmi context
 511          */
 512         int                     recursion[4];
 513 };
 514
 515 /*
 516  * Set by architecture code:
 517  */
 518 extern int perf_max_counters;
 519
 520 #ifdef CONFIG_PERF_COUNTERS
 521 extern const struct hw_perf_counter_ops *
 522 hw_perf_counter_init(struct perf_counter *counter);
 523
 524 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 525 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 526 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 527 extern void perf_counter_init_task(struct task_struct *child);
 528 extern void perf_counter_exit_task(struct task_struct *child);
 529 extern void perf_counter_do_pending(void);
 530 extern void perf_counter_print_debug(void);
 531 extern void perf_counter_unthrottle(void);
 532 extern u64 hw_perf_save_disable(void);
 533 extern void hw_perf_restore(u64 ctrl);
 534 extern int perf_counter_task_disable(void);
 535 extern int perf_counter_task_enable(void);
 536 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 537                struct perf_cpu_context *cpuctx,
 538                struct perf_counter_context *ctx, int cpu);
 539 extern void perf_counter_update_userpage(struct perf_counter *counter);
 540
 541 extern int perf_counter_overflow(struct perf_counter *counter,
 542                                  int nmi, struct pt_regs *regs, u64 addr);
 543 /*
 544  * Return 1 for a software counter, 0 for a hardware counter
 545  */
 546 static inline int is_software_counter(struct perf_counter *counter)
 547 {
 548         return !perf_event_raw(&counter->hw_event) &&
 549                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 550 }
 551
 552 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
 553
 554 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 555                               unsigned long pgoff, struct file *file);
 556
 557 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 558                                 unsigned long pgoff, struct file *file);
 559
 560 extern void perf_counter_comm(struct task_struct *tsk);
 561
 562 #define MAX_STACK_DEPTH         255
 563
 564 struct perf_callchain_entry {
 565         u16     nr, hv, kernel, user;
 566         u64     ip[MAX_STACK_DEPTH];
 567 };
 568
 569 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 570
 571 extern int sysctl_perf_counter_priv;
 572
 573 #else
 574 static inline void
 575 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 576 static inline void
 577 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 578 static inline void
 579 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 580 static inline void perf_counter_init_task(struct task_struct *child)    { }
 581 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 582 static inline void perf_counter_do_pending(void)                        { }
 583 static inline void perf_counter_print_debug(void)                       { }
 584 static inline void perf_counter_unthrottle(void)                        { }
 585 static inline void hw_perf_restore(u64 ctrl)                            { }
 586 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 587 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 588 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 589
 590 static inline void
 591 perf_swcounter_event(u32 event, u64 nr, int nmi,
 592                      struct pt_regs *regs, u64 addr)                    { }
 593
 594 static inline void
 595 perf_counter_mmap(unsigned long addr, unsigned long len,
 596                   unsigned long pgoff, struct file *file)               { }
 597
 598 static inline void
 599 perf_counter_munmap(unsigned long addr, unsigned long len,
 600                     unsigned long pgoff, struct file *file)             { }
 601
 602 static inline void perf_counter_comm(struct task_struct *tsk)           { }
 603 #endif
 604
 605 #endif /* __KERNEL__ */
 606 #endif /* _LINUX_PERF_COUNTER_H */