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                   extra_config_len;
 158         __u32                   wakeup_events;  /* wakeup every n events */
 159
 160         __u64                   __reserved_2;
 161         __u64                   __reserved_3;
 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          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
 262          * will be PERF_RECORD_*
 263          *
 264          * struct {
 265          *      struct perf_event_header        header;
 266          *
 267          *      { u64                   ip;       } && PERF_RECORD_IP
 268          *      { u32                   pid, tid; } && PERF_RECORD_TID
 269          *      { u64                   time;     } && PERF_RECORD_TIME
 270          *      { u64                   addr;     } && PERF_RECORD_ADDR
 271          *      { u64                   config;   } && PERF_RECORD_CONFIG
 272          *      { u32                   cpu, res; } && PERF_RECORD_CPU
 273          *
 274          *      { u64                   nr;
 275          *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
 276          *
 277          *      { u16                   nr,
 278          *                              hv,
 279          *                              kernel,
 280          *                              user;
 281          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
 282          * };
 283          */
 284 };
 285
 286 #ifdef __KERNEL__
 287 /*
 288  * Kernel-internal data types and definitions:
 289  */
 290
 291 #ifdef CONFIG_PERF_COUNTERS
 292 # include <asm/perf_counter.h>
 293 #endif
 294
 295 #include <linux/list.h>
 296 #include <linux/mutex.h>
 297 #include <linux/rculist.h>
 298 #include <linux/rcupdate.h>
 299 #include <linux/spinlock.h>
 300 #include <linux/hrtimer.h>
 301 #include <linux/fs.h>
 302 #include <asm/atomic.h>
 303
 304 struct task_struct;
 305
 306 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 307 {
 308         return hw_event->config & PERF_COUNTER_RAW_MASK;
 309 }
 310
 311 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 312 {
 313         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 314 }
 315
 316 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 317 {
 318         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 319                 PERF_COUNTER_TYPE_SHIFT;
 320 }
 321
 322 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 323 {
 324         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 325 }
 326
 327 /**
 328  * struct hw_perf_counter - performance counter hardware details:
 329  */
 330 struct hw_perf_counter {
 331 #ifdef CONFIG_PERF_COUNTERS
 332         union {
 333                 struct { /* hardware */
 334                         u64                             config;
 335                         unsigned long                   config_base;
 336                         unsigned long                   counter_base;
 337                         int                             nmi;
 338                         int                             idx;
 339                 };
 340                 union { /* software */
 341                         atomic64_t                      count;
 342                         struct hrtimer                  hrtimer;
 343                 };
 344         };
 345         atomic64_t                      prev_count;
 346         u64                             irq_period;
 347         atomic64_t                      period_left;
 348         u64                             interrupts;
 349 #endif
 350 };
 351
 352 struct perf_counter;
 353
 354 /**
 355  * struct pmu - generic performance monitoring unit
 356  */
 357 struct pmu {
 358         int (*enable)                   (struct perf_counter *counter);
 359         void (*disable)                 (struct perf_counter *counter);
 360         void (*read)                    (struct perf_counter *counter);
 361 };
 362
 363 /**
 364  * enum perf_counter_active_state - the states of a counter
 365  */
 366 enum perf_counter_active_state {
 367         PERF_COUNTER_STATE_ERROR        = -2,
 368         PERF_COUNTER_STATE_OFF          = -1,
 369         PERF_COUNTER_STATE_INACTIVE     =  0,
 370         PERF_COUNTER_STATE_ACTIVE       =  1,
 371 };
 372
 373 struct file;
 374
 375 struct perf_mmap_data {
 376         struct rcu_head                 rcu_head;
 377         int                             nr_pages;       /* nr of data pages  */
 378         int                             nr_locked;      /* nr pages mlocked  */
 379
 380         atomic_t                        poll;           /* POLL_ for wakeups */
 381         atomic_t                        head;           /* write position    */
 382         atomic_t                        events;         /* event limit       */
 383
 384         atomic_t                        done_head;      /* completed head    */
 385         atomic_t                        lock;           /* concurrent writes */
 386
 387         atomic_t                        wakeup;         /* needs a wakeup    */
 388
 389         struct perf_counter_mmap_page   *user_page;
 390         void                            *data_pages[0];
 391 };
 392
 393 struct perf_pending_entry {
 394         struct perf_pending_entry *next;
 395         void (*func)(struct perf_pending_entry *);
 396 };
 397
 398 /**
 399  * struct perf_counter - performance counter kernel representation:
 400  */
 401 struct perf_counter {
 402 #ifdef CONFIG_PERF_COUNTERS
 403         struct list_head                list_entry;
 404         struct list_head                event_entry;
 405         struct list_head                sibling_list;
 406         int                             nr_siblings;
 407         struct perf_counter             *group_leader;
 408         const struct pmu                *pmu;
 409
 410         enum perf_counter_active_state  state;
 411         enum perf_counter_active_state  prev_state;
 412         atomic64_t                      count;
 413
 414         /*
 415          * These are the total time in nanoseconds that the counter
 416          * has been enabled (i.e. eligible to run, and the task has
 417          * been scheduled in, if this is a per-task counter)
 418          * and running (scheduled onto the CPU), respectively.
 419          *
 420          * They are computed from tstamp_enabled, tstamp_running and
 421          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 422          */
 423         u64                             total_time_enabled;
 424         u64                             total_time_running;
 425
 426         /*
 427          * These are timestamps used for computing total_time_enabled
 428          * and total_time_running when the counter is in INACTIVE or
 429          * ACTIVE state, measured in nanoseconds from an arbitrary point
 430          * in time.
 431          * tstamp_enabled: the notional time when the counter was enabled
 432          * tstamp_running: the notional time when the counter was scheduled on
 433          * tstamp_stopped: in INACTIVE state, the notional time when the
 434          *      counter was scheduled off.
 435          */
 436         u64                             tstamp_enabled;
 437         u64                             tstamp_running;
 438         u64                             tstamp_stopped;
 439
 440         struct perf_counter_hw_event    hw_event;
 441         struct hw_perf_counter          hw;
 442
 443         struct perf_counter_context     *ctx;
 444         struct task_struct              *task;
 445         struct file                     *filp;
 446
 447         struct perf_counter             *parent;
 448         struct list_head                child_list;
 449
 450         /*
 451          * These accumulate total time (in nanoseconds) that children
 452          * counters have been enabled and running, respectively.
 453          */
 454         atomic64_t                      child_total_time_enabled;
 455         atomic64_t                      child_total_time_running;
 456
 457         /*
 458          * Protect attach/detach and child_list:
 459          */
 460         struct mutex                    mutex;
 461
 462         int                             oncpu;
 463         int                             cpu;
 464
 465         /* mmap bits */
 466         struct mutex                    mmap_mutex;
 467         atomic_t                        mmap_count;
 468         struct perf_mmap_data           *data;
 469
 470         /* poll related */
 471         wait_queue_head_t               waitq;
 472         struct fasync_struct            *fasync;
 473
 474         /* delayed work for NMIs and such */
 475         int                             pending_wakeup;
 476         int                             pending_kill;
 477         int                             pending_disable;
 478         struct perf_pending_entry       pending;
 479
 480         atomic_t                        event_limit;
 481
 482         void (*destroy)(struct perf_counter *);
 483         struct rcu_head                 rcu_head;
 484 #endif
 485 };
 486
 487 /**
 488  * struct perf_counter_context - counter context structure
 489  *
 490  * Used as a container for task counters and CPU counters as well:
 491  */
 492 struct perf_counter_context {
 493 #ifdef CONFIG_PERF_COUNTERS
 494         /*
 495          * Protect the states of the counters in the list,
 496          * nr_active, and the list:
 497          */
 498         spinlock_t              lock;
 499         /*
 500          * Protect the list of counters.  Locking either mutex or lock
 501          * is sufficient to ensure the list doesn't change; to change
 502          * the list you need to lock both the mutex and the spinlock.
 503          */
 504         struct mutex            mutex;
 505
 506         struct list_head        counter_list;
 507         struct list_head        event_list;
 508         int                     nr_counters;
 509         int                     nr_active;
 510         int                     is_active;
 511         struct task_struct      *task;
 512
 513         /*
 514          * Context clock, runs when context enabled.
 515          */
 516         u64                     time;
 517         u64                     timestamp;
 518 #endif
 519 };
 520
 521 /**
 522  * struct perf_counter_cpu_context - per cpu counter context structure
 523  */
 524 struct perf_cpu_context {
 525         struct perf_counter_context     ctx;
 526         struct perf_counter_context     *task_ctx;
 527         int                             active_oncpu;
 528         int                             max_pertask;
 529         int                             exclusive;
 530
 531         /*
 532          * Recursion avoidance:
 533          *
 534          * task, softirq, irq, nmi context
 535          */
 536         int                     recursion[4];
 537 };
 538
 539 #ifdef CONFIG_PERF_COUNTERS
 540
 541 /*
 542  * Set by architecture code:
 543  */
 544 extern int perf_max_counters;
 545
 546 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
 547
 548 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 549 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 550 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 551 extern void perf_counter_init_task(struct task_struct *child);
 552 extern void perf_counter_exit_task(struct task_struct *child);
 553 extern void perf_counter_do_pending(void);
 554 extern void perf_counter_print_debug(void);
 555 extern void perf_counter_unthrottle(void);
 556 extern void __perf_disable(void);
 557 extern bool __perf_enable(void);
 558 extern void perf_disable(void);
 559 extern void perf_enable(void);
 560 extern int perf_counter_task_disable(void);
 561 extern int perf_counter_task_enable(void);
 562 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 563                struct perf_cpu_context *cpuctx,
 564                struct perf_counter_context *ctx, int cpu);
 565 extern void perf_counter_update_userpage(struct perf_counter *counter);
 566
 567 extern int perf_counter_overflow(struct perf_counter *counter,
 568                                  int nmi, struct pt_regs *regs, u64 addr);
 569 /*
 570  * Return 1 for a software counter, 0 for a hardware counter
 571  */
 572 static inline int is_software_counter(struct perf_counter *counter)
 573 {
 574         return !perf_event_raw(&counter->hw_event) &&
 575                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 576 }
 577
 578 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
 579
 580 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 581                               unsigned long pgoff, struct file *file);
 582
 583 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 584                                 unsigned long pgoff, struct file *file);
 585
 586 extern void perf_counter_comm(struct task_struct *tsk);
 587
 588 #define MAX_STACK_DEPTH         255
 589
 590 struct perf_callchain_entry {
 591         u16     nr, hv, kernel, user;
 592         u64     ip[MAX_STACK_DEPTH];
 593 };
 594
 595 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 596
 597 extern int sysctl_perf_counter_priv;
 598 extern int sysctl_perf_counter_mlock;
 599
 600 extern void perf_counter_init(void);
 601
 602 #ifndef perf_misc_flags
 603 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
 604                                  PERF_EVENT_MISC_KERNEL)
 605 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
 606 #endif
 607
 608 #else
 609 static inline void
 610 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 611 static inline void
 612 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 613 static inline void
 614 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 615 static inline void perf_counter_init_task(struct task_struct *child)    { }
 616 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 617 static inline void perf_counter_do_pending(void)                        { }
 618 static inline void perf_counter_print_debug(void)                       { }
 619 static inline void perf_counter_unthrottle(void)                        { }
 620 static inline void perf_disable(void)                                   { }
 621 static inline void perf_enable(void)                                    { }
 622 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 623 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 624
 625 static inline void
 626 perf_swcounter_event(u32 event, u64 nr, int nmi,
 627                      struct pt_regs *regs, u64 addr)                    { }
 628
 629 static inline void
 630 perf_counter_mmap(unsigned long addr, unsigned long len,
 631                   unsigned long pgoff, struct file *file)               { }
 632
 633 static inline void
 634 perf_counter_munmap(unsigned long addr, unsigned long len,
 635                     unsigned long pgoff, struct file *file)             { }
 636
 637 static inline void perf_counter_comm(struct task_struct *tsk)           { }
 638 static inline void perf_counter_init(void)                              { }
 639 #endif
 640
 641 #endif /* __KERNEL__ */
 642 #endif /* _LINUX_PERF_COUNTER_H */