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_GROUP       = 1U << 2,
 104         PERF_RECORD_CALLCHAIN   = 1U << 3,
 105         PERF_RECORD_TIME        = 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
 146                                 __reserved_1   : 53;
 147
 148         __u32                   extra_config_len;
 149         __u32                   wakeup_events;  /* wakeup every n events */
 150
 151         __u64                   __reserved_2;
 152         __u64                   __reserved_3;
 153 };
 154
 155 /*
 156  * Ioctls that can be done on a perf counter fd:
 157  */
 158 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
 159 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
 160 #define PERF_COUNTER_IOC_REFRESH        _IOW('$', 2, u32)
 161
 162 /*
 163  * Structure of the page that can be mapped via mmap
 164  */
 165 struct perf_counter_mmap_page {
 166         __u32   version;                /* version number of this structure */
 167         __u32   compat_version;         /* lowest version this is compat with */
 168
 169         /*
 170          * Bits needed to read the hw counters in user-space.
 171          *
 172          *   u32 seq;
 173          *   s64 count;
 174          *
 175          *   do {
 176          *     seq = pc->lock;
 177          *
 178          *     barrier()
 179          *     if (pc->index) {
 180          *       count = pmc_read(pc->index - 1);
 181          *       count += pc->offset;
 182          *     } else
 183          *       goto regular_read;
 184          *
 185          *     barrier();
 186          *   } while (pc->lock != seq);
 187          *
 188          * NOTE: for obvious reason this only works on self-monitoring
 189          *       processes.
 190          */
 191         __u32   lock;                   /* seqlock for synchronization */
 192         __u32   index;                  /* hardware counter identifier */
 193         __s64   offset;                 /* add to hardware counter value */
 194
 195         /*
 196          * Control data for the mmap() data buffer.
 197          *
 198          * User-space reading this value should issue an rmb(), on SMP capable
 199          * platforms, after reading this value -- see perf_counter_wakeup().
 200          */
 201         __u32   data_head;              /* head in the data section */
 202 };
 203
 204 struct perf_event_header {
 205         __u32   type;
 206         __u32   size;
 207 };
 208
 209 enum perf_event_type {
 210
 211         /*
 212          * The MMAP events record the PROT_EXEC mappings so that we can
 213          * correlate userspace IPs to code. They have the following structure:
 214          *
 215          * struct {
 216          *      struct perf_event_header        header;
 217          *
 218          *      u32                             pid, tid;
 219          *      u64                             addr;
 220          *      u64                             len;
 221          *      u64                             pgoff;
 222          *      char                            filename[];
 223          * };
 224          */
 225         PERF_EVENT_MMAP                 = 1,
 226         PERF_EVENT_MUNMAP               = 2,
 227
 228         /*
 229          * Half the event type space is reserved for the counter overflow
 230          * bitfields, as found in hw_event.record_type.
 231          *
 232          * These events will have types of the form:
 233          *   PERF_EVENT_COUNTER_OVERFLOW { | __PERF_EVENT_* } *
 234          *
 235          * struct {
 236          *      struct perf_event_header        header;
 237          *
 238          *      { u64                   ip;       } && __PERF_EVENT_IP
 239          *      { u32                   pid, tid; } && __PERF_EVENT_TID
 240          *
 241          *      { u64                   nr;
 242          *        { u64 event, val; }   cnt[nr];  } && __PERF_EVENT_GROUP
 243          *
 244          *      { u16                   nr,
 245          *                              hv,
 246          *                              kernel,
 247          *                              user;
 248          *        u64                   ips[nr];  } && __PERF_EVENT_CALLCHAIN
 249          *
 250          *      { u64                   time;     } && __PERF_EVENT_TIME
 251          * };
 252          */
 253         PERF_EVENT_COUNTER_OVERFLOW     = 1UL << 31,
 254         __PERF_EVENT_IP                 = PERF_RECORD_IP,
 255         __PERF_EVENT_TID                = PERF_RECORD_TID,
 256         __PERF_EVENT_GROUP              = PERF_RECORD_GROUP,
 257         __PERF_EVENT_CALLCHAIN          = PERF_RECORD_CALLCHAIN,
 258         __PERF_EVENT_TIME               = PERF_RECORD_TIME,
 259 };
 260
 261 #ifdef __KERNEL__
 262 /*
 263  * Kernel-internal data types and definitions:
 264  */
 265
 266 #ifdef CONFIG_PERF_COUNTERS
 267 # include <asm/perf_counter.h>
 268 #endif
 269
 270 #include <linux/list.h>
 271 #include <linux/mutex.h>
 272 #include <linux/rculist.h>
 273 #include <linux/rcupdate.h>
 274 #include <linux/spinlock.h>
 275 #include <linux/hrtimer.h>
 276 #include <linux/fs.h>
 277 #include <asm/atomic.h>
 278
 279 struct task_struct;
 280
 281 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 282 {
 283         return hw_event->config & PERF_COUNTER_RAW_MASK;
 284 }
 285
 286 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 287 {
 288         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 289 }
 290
 291 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 292 {
 293         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 294                 PERF_COUNTER_TYPE_SHIFT;
 295 }
 296
 297 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 298 {
 299         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 300 }
 301
 302 /**
 303  * struct hw_perf_counter - performance counter hardware details:
 304  */
 305 struct hw_perf_counter {
 306 #ifdef CONFIG_PERF_COUNTERS
 307         union {
 308                 struct { /* hardware */
 309                         u64                             config;
 310                         unsigned long                   config_base;
 311                         unsigned long                   counter_base;
 312                         int                             nmi;
 313                         unsigned int                    idx;
 314                 };
 315                 union { /* software */
 316                         atomic64_t                      count;
 317                         struct hrtimer                  hrtimer;
 318                 };
 319         };
 320         atomic64_t                      prev_count;
 321         u64                             irq_period;
 322         atomic64_t                      period_left;
 323 #endif
 324 };
 325
 326 struct perf_counter;
 327
 328 /**
 329  * struct hw_perf_counter_ops - performance counter hw ops
 330  */
 331 struct hw_perf_counter_ops {
 332         int (*enable)                   (struct perf_counter *counter);
 333         void (*disable)                 (struct perf_counter *counter);
 334         void (*read)                    (struct perf_counter *counter);
 335 };
 336
 337 /**
 338  * enum perf_counter_active_state - the states of a counter
 339  */
 340 enum perf_counter_active_state {
 341         PERF_COUNTER_STATE_ERROR        = -2,
 342         PERF_COUNTER_STATE_OFF          = -1,
 343         PERF_COUNTER_STATE_INACTIVE     =  0,
 344         PERF_COUNTER_STATE_ACTIVE       =  1,
 345 };
 346
 347 struct file;
 348
 349 struct perf_mmap_data {
 350         struct rcu_head                 rcu_head;
 351         int                             nr_pages;
 352         atomic_t                        wakeup;
 353         atomic_t                        head;
 354         atomic_t                        events;
 355         struct perf_counter_mmap_page   *user_page;
 356         void                            *data_pages[0];
 357 };
 358
 359 struct perf_pending_entry {
 360         struct perf_pending_entry *next;
 361         void (*func)(struct perf_pending_entry *);
 362 };
 363
 364 /**
 365  * struct perf_counter - performance counter kernel representation:
 366  */
 367 struct perf_counter {
 368 #ifdef CONFIG_PERF_COUNTERS
 369         struct list_head                list_entry;
 370         struct list_head                event_entry;
 371         struct list_head                sibling_list;
 372         int                             nr_siblings;
 373         struct perf_counter             *group_leader;
 374         const struct hw_perf_counter_ops *hw_ops;
 375
 376         enum perf_counter_active_state  state;
 377         enum perf_counter_active_state  prev_state;
 378         atomic64_t                      count;
 379
 380         /*
 381          * These are the total time in nanoseconds that the counter
 382          * has been enabled (i.e. eligible to run, and the task has
 383          * been scheduled in, if this is a per-task counter)
 384          * and running (scheduled onto the CPU), respectively.
 385          *
 386          * They are computed from tstamp_enabled, tstamp_running and
 387          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 388          */
 389         u64                             total_time_enabled;
 390         u64                             total_time_running;
 391
 392         /*
 393          * These are timestamps used for computing total_time_enabled
 394          * and total_time_running when the counter is in INACTIVE or
 395          * ACTIVE state, measured in nanoseconds from an arbitrary point
 396          * in time.
 397          * tstamp_enabled: the notional time when the counter was enabled
 398          * tstamp_running: the notional time when the counter was scheduled on
 399          * tstamp_stopped: in INACTIVE state, the notional time when the
 400          *      counter was scheduled off.
 401          */
 402         u64                             tstamp_enabled;
 403         u64                             tstamp_running;
 404         u64                             tstamp_stopped;
 405
 406         struct perf_counter_hw_event    hw_event;
 407         struct hw_perf_counter          hw;
 408
 409         struct perf_counter_context     *ctx;
 410         struct task_struct              *task;
 411         struct file                     *filp;
 412
 413         struct perf_counter             *parent;
 414         struct list_head                child_list;
 415
 416         /*
 417          * These accumulate total time (in nanoseconds) that children
 418          * counters have been enabled and running, respectively.
 419          */
 420         atomic64_t                      child_total_time_enabled;
 421         atomic64_t                      child_total_time_running;
 422
 423         /*
 424          * Protect attach/detach and child_list:
 425          */
 426         struct mutex                    mutex;
 427
 428         int                             oncpu;
 429         int                             cpu;
 430
 431         /* mmap bits */
 432         struct mutex                    mmap_mutex;
 433         atomic_t                        mmap_count;
 434         struct perf_mmap_data           *data;
 435
 436         /* poll related */
 437         wait_queue_head_t               waitq;
 438         struct fasync_struct            *fasync;
 439
 440         /* delayed work for NMIs and such */
 441         int                             pending_wakeup;
 442         int                             pending_kill;
 443         int                             pending_disable;
 444         struct perf_pending_entry       pending;
 445
 446         atomic_t                        event_limit;
 447
 448         void (*destroy)(struct perf_counter *);
 449         struct rcu_head                 rcu_head;
 450 #endif
 451 };
 452
 453 /**
 454  * struct perf_counter_context - counter context structure
 455  *
 456  * Used as a container for task counters and CPU counters as well:
 457  */
 458 struct perf_counter_context {
 459 #ifdef CONFIG_PERF_COUNTERS
 460         /*
 461          * Protect the states of the counters in the list,
 462          * nr_active, and the list:
 463          */
 464         spinlock_t              lock;
 465         /*
 466          * Protect the list of counters.  Locking either mutex or lock
 467          * is sufficient to ensure the list doesn't change; to change
 468          * the list you need to lock both the mutex and the spinlock.
 469          */
 470         struct mutex            mutex;
 471
 472         struct list_head        counter_list;
 473         struct list_head        event_list;
 474         int                     nr_counters;
 475         int                     nr_active;
 476         int                     is_active;
 477         struct task_struct      *task;
 478
 479         /*
 480          * time_now is the current time in nanoseconds since an arbitrary
 481          * point in the past.  For per-task counters, this is based on the
 482          * task clock, and for per-cpu counters it is based on the cpu clock.
 483          * time_lost is an offset from the task/cpu clock, used to make it
 484          * appear that time only passes while the context is scheduled in.
 485          */
 486         u64                     time_now;
 487         u64                     time_lost;
 488 #endif
 489 };
 490
 491 /**
 492  * struct perf_counter_cpu_context - per cpu counter context structure
 493  */
 494 struct perf_cpu_context {
 495         struct perf_counter_context     ctx;
 496         struct perf_counter_context     *task_ctx;
 497         int                             active_oncpu;
 498         int                             max_pertask;
 499         int                             exclusive;
 500
 501         /*
 502          * Recursion avoidance:
 503          *
 504          * task, softirq, irq, nmi context
 505          */
 506         int                     recursion[4];
 507 };
 508
 509 /*
 510  * Set by architecture code:
 511  */
 512 extern int perf_max_counters;
 513
 514 #ifdef CONFIG_PERF_COUNTERS
 515 extern const struct hw_perf_counter_ops *
 516 hw_perf_counter_init(struct perf_counter *counter);
 517
 518 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 519 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 520 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 521 extern void perf_counter_init_task(struct task_struct *child);
 522 extern void perf_counter_exit_task(struct task_struct *child);
 523 extern void perf_counter_do_pending(void);
 524 extern void perf_counter_print_debug(void);
 525 extern void perf_counter_unthrottle(void);
 526 extern u64 hw_perf_save_disable(void);
 527 extern void hw_perf_restore(u64 ctrl);
 528 extern int perf_counter_task_disable(void);
 529 extern int perf_counter_task_enable(void);
 530 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 531                struct perf_cpu_context *cpuctx,
 532                struct perf_counter_context *ctx, int cpu);
 533 extern void perf_counter_update_userpage(struct perf_counter *counter);
 534
 535 extern int perf_counter_overflow(struct perf_counter *counter,
 536                                  int nmi, struct pt_regs *regs);
 537 /*
 538  * Return 1 for a software counter, 0 for a hardware counter
 539  */
 540 static inline int is_software_counter(struct perf_counter *counter)
 541 {
 542         return !perf_event_raw(&counter->hw_event) &&
 543                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 544 }
 545
 546 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 547
 548 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 549                               unsigned long pgoff, struct file *file);
 550
 551 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 552                                 unsigned long pgoff, struct file *file);
 553
 554 #define MAX_STACK_DEPTH         255
 555
 556 struct perf_callchain_entry {
 557         u16     nr, hv, kernel, user;
 558         u64     ip[MAX_STACK_DEPTH];
 559 };
 560
 561 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 562
 563 #else
 564 static inline void
 565 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 566 static inline void
 567 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 568 static inline void
 569 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 570 static inline void perf_counter_init_task(struct task_struct *child)    { }
 571 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 572 static inline void perf_counter_do_pending(void)                        { }
 573 static inline void perf_counter_print_debug(void)                       { }
 574 static inline void perf_counter_unthrottle(void)                        { }
 575 static inline void hw_perf_restore(u64 ctrl)                            { }
 576 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 577 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 578 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 579
 580 static inline void
 581 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)  { }
 582
 583
 584 static inline void
 585 perf_counter_mmap(unsigned long addr, unsigned long len,
 586                   unsigned long pgoff, struct file *file)               { }
 587
 588 static inline void
 589 perf_counter_munmap(unsigned long addr, unsigned long len,
 590                     unsigned long pgoff, struct file *file)             { }
 591
 592 #endif
 593
 594 #endif /* __KERNEL__ */
 595 #endif /* _LINUX_PERF_COUNTER_H */