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 };
 106
 107 /*
 108  * Bits that can be set in hw_event.read_format to request that
 109  * reads on the counter should return the indicated quantities,
 110  * in increasing order of bit value, after the counter value.
 111  */
 112 enum perf_counter_read_format {
 113         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
 114         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
 115 };
 116
 117 /*
 118  * Hardware event to monitor via a performance monitoring counter:
 119  */
 120 struct perf_counter_hw_event {
 121         /*
 122          * The MSB of the config word signifies if the rest contains cpu
 123          * specific (raw) counter configuration data, if unset, the next
 124          * 7 bits are an event type and the rest of the bits are the event
 125          * identifier.
 126          */
 127         __u64                   config;
 128
 129         __u64                   irq_period;
 130         __u32                   record_type;
 131         __u32                   read_format;
 132
 133         __u64                   disabled       :  1, /* off by default        */
 134                                 nmi            :  1, /* NMI sampling          */
 135                                 inherit        :  1, /* children inherit it   */
 136                                 pinned         :  1, /* must always be on PMU */
 137                                 exclusive      :  1, /* only group on PMU     */
 138                                 exclude_user   :  1, /* don't count user      */
 139                                 exclude_kernel :  1, /* ditto kernel          */
 140                                 exclude_hv     :  1, /* ditto hypervisor      */
 141                                 exclude_idle   :  1, /* don't count when idle */
 142                                 mmap           :  1, /* include mmap data     */
 143                                 munmap         :  1, /* include munmap data   */
 144
 145                                 __reserved_1   : 53;
 146
 147         __u32                   extra_config_len;
 148         __u32                   wakeup_events;  /* wakeup every n events */
 149
 150         __u64                   __reserved_2;
 151         __u64                   __reserved_3;
 152 };
 153
 154 /*
 155  * Ioctls that can be done on a perf counter fd:
 156  */
 157 #define PERF_COUNTER_IOC_ENABLE         _IO('$', 0)
 158 #define PERF_COUNTER_IOC_DISABLE        _IO('$', 1)
 159
 160 /*
 161  * Structure of the page that can be mapped via mmap
 162  */
 163 struct perf_counter_mmap_page {
 164         __u32   version;                /* version number of this structure */
 165         __u32   compat_version;         /* lowest version this is compat with */
 166
 167         /*
 168          * Bits needed to read the hw counters in user-space.
 169          *
 170          *   u32 seq;
 171          *   s64 count;
 172          *
 173          *   do {
 174          *     seq = pc->lock;
 175          *
 176          *     barrier()
 177          *     if (pc->index) {
 178          *       count = pmc_read(pc->index - 1);
 179          *       count += pc->offset;
 180          *     } else
 181          *       goto regular_read;
 182          *
 183          *     barrier();
 184          *   } while (pc->lock != seq);
 185          *
 186          * NOTE: for obvious reason this only works on self-monitoring
 187          *       processes.
 188          */
 189         __u32   lock;                   /* seqlock for synchronization */
 190         __u32   index;                  /* hardware counter identifier */
 191         __s64   offset;                 /* add to hardware counter value */
 192
 193         /*
 194          * Control data for the mmap() data buffer.
 195          *
 196          * User-space reading this value should issue an rmb(), on SMP capable
 197          * platforms, after reading this value -- see perf_counter_wakeup().
 198          */
 199         __u32   data_head;              /* head in the data section */
 200 };
 201
 202 struct perf_event_header {
 203         __u32   type;
 204         __u32   size;
 205 };
 206
 207 enum perf_event_type {
 208
 209         PERF_EVENT_MMAP                 = 1,
 210         PERF_EVENT_MUNMAP               = 2,
 211
 212         /*
 213          * Half the event type space is reserved for the counter overflow
 214          * bitfields, as found in hw_event.record_type.
 215          *
 216          * These events will have types of the form:
 217          *   PERF_EVENT_COUNTER_OVERFLOW { | __PERF_EVENT_* } *
 218          */
 219         PERF_EVENT_COUNTER_OVERFLOW     = 1UL << 31,
 220         __PERF_EVENT_IP                 = PERF_RECORD_IP,
 221         __PERF_EVENT_TID                = PERF_RECORD_TID,
 222         __PERF_EVENT_GROUP              = PERF_RECORD_GROUP,
 223         __PERF_EVENT_CALLCHAIN          = PERF_RECORD_CALLCHAIN,
 224 };
 225
 226 #ifdef __KERNEL__
 227 /*
 228  * Kernel-internal data types and definitions:
 229  */
 230
 231 #ifdef CONFIG_PERF_COUNTERS
 232 # include <asm/perf_counter.h>
 233 #endif
 234
 235 #include <linux/list.h>
 236 #include <linux/mutex.h>
 237 #include <linux/rculist.h>
 238 #include <linux/rcupdate.h>
 239 #include <linux/spinlock.h>
 240 #include <linux/hrtimer.h>
 241 #include <asm/atomic.h>
 242
 243 struct task_struct;
 244
 245 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 246 {
 247         return hw_event->config & PERF_COUNTER_RAW_MASK;
 248 }
 249
 250 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 251 {
 252         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 253 }
 254
 255 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 256 {
 257         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 258                 PERF_COUNTER_TYPE_SHIFT;
 259 }
 260
 261 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 262 {
 263         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 264 }
 265
 266 /**
 267  * struct hw_perf_counter - performance counter hardware details:
 268  */
 269 struct hw_perf_counter {
 270 #ifdef CONFIG_PERF_COUNTERS
 271         union {
 272                 struct { /* hardware */
 273                         u64                             config;
 274                         unsigned long                   config_base;
 275                         unsigned long                   counter_base;
 276                         int                             nmi;
 277                         unsigned int                    idx;
 278                 };
 279                 union { /* software */
 280                         atomic64_t                      count;
 281                         struct hrtimer                  hrtimer;
 282                 };
 283         };
 284         atomic64_t                      prev_count;
 285         u64                             irq_period;
 286         atomic64_t                      period_left;
 287 #endif
 288 };
 289
 290 struct perf_counter;
 291
 292 /**
 293  * struct hw_perf_counter_ops - performance counter hw ops
 294  */
 295 struct hw_perf_counter_ops {
 296         int (*enable)                   (struct perf_counter *counter);
 297         void (*disable)                 (struct perf_counter *counter);
 298         void (*read)                    (struct perf_counter *counter);
 299 };
 300
 301 /**
 302  * enum perf_counter_active_state - the states of a counter
 303  */
 304 enum perf_counter_active_state {
 305         PERF_COUNTER_STATE_ERROR        = -2,
 306         PERF_COUNTER_STATE_OFF          = -1,
 307         PERF_COUNTER_STATE_INACTIVE     =  0,
 308         PERF_COUNTER_STATE_ACTIVE       =  1,
 309 };
 310
 311 struct file;
 312
 313 struct perf_mmap_data {
 314         struct rcu_head                 rcu_head;
 315         int                             nr_pages;
 316         atomic_t                        wakeup;
 317         atomic_t                        head;
 318         atomic_t                        events;
 319         struct perf_counter_mmap_page   *user_page;
 320         void                            *data_pages[0];
 321 };
 322
 323 struct perf_wakeup_entry {
 324         struct perf_wakeup_entry *next;
 325 };
 326
 327 /**
 328  * struct perf_counter - performance counter kernel representation:
 329  */
 330 struct perf_counter {
 331 #ifdef CONFIG_PERF_COUNTERS
 332         struct list_head                list_entry;
 333         struct list_head                event_entry;
 334         struct list_head                sibling_list;
 335         int                             nr_siblings;
 336         struct perf_counter             *group_leader;
 337         const struct hw_perf_counter_ops *hw_ops;
 338
 339         enum perf_counter_active_state  state;
 340         enum perf_counter_active_state  prev_state;
 341         atomic64_t                      count;
 342
 343         /*
 344          * These are the total time in nanoseconds that the counter
 345          * has been enabled (i.e. eligible to run, and the task has
 346          * been scheduled in, if this is a per-task counter)
 347          * and running (scheduled onto the CPU), respectively.
 348          *
 349          * They are computed from tstamp_enabled, tstamp_running and
 350          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 351          */
 352         u64                             total_time_enabled;
 353         u64                             total_time_running;
 354
 355         /*
 356          * These are timestamps used for computing total_time_enabled
 357          * and total_time_running when the counter is in INACTIVE or
 358          * ACTIVE state, measured in nanoseconds from an arbitrary point
 359          * in time.
 360          * tstamp_enabled: the notional time when the counter was enabled
 361          * tstamp_running: the notional time when the counter was scheduled on
 362          * tstamp_stopped: in INACTIVE state, the notional time when the
 363          *      counter was scheduled off.
 364          */
 365         u64                             tstamp_enabled;
 366         u64                             tstamp_running;
 367         u64                             tstamp_stopped;
 368
 369         struct perf_counter_hw_event    hw_event;
 370         struct hw_perf_counter          hw;
 371
 372         struct perf_counter_context     *ctx;
 373         struct task_struct              *task;
 374         struct file                     *filp;
 375
 376         struct perf_counter             *parent;
 377         struct list_head                child_list;
 378
 379         /*
 380          * These accumulate total time (in nanoseconds) that children
 381          * counters have been enabled and running, respectively.
 382          */
 383         atomic64_t                      child_total_time_enabled;
 384         atomic64_t                      child_total_time_running;
 385
 386         /*
 387          * Protect attach/detach and child_list:
 388          */
 389         struct mutex                    mutex;
 390
 391         int                             oncpu;
 392         int                             cpu;
 393
 394         /* mmap bits */
 395         struct mutex                    mmap_mutex;
 396         atomic_t                        mmap_count;
 397         struct perf_mmap_data           *data;
 398
 399         /* poll related */
 400         wait_queue_head_t               waitq;
 401         /* optional: for NMIs */
 402         struct perf_wakeup_entry        wakeup;
 403
 404         void (*destroy)(struct perf_counter *);
 405         struct rcu_head                 rcu_head;
 406 #endif
 407 };
 408
 409 /**
 410  * struct perf_counter_context - counter context structure
 411  *
 412  * Used as a container for task counters and CPU counters as well:
 413  */
 414 struct perf_counter_context {
 415 #ifdef CONFIG_PERF_COUNTERS
 416         /*
 417          * Protect the states of the counters in the list,
 418          * nr_active, and the list:
 419          */
 420         spinlock_t              lock;
 421         /*
 422          * Protect the list of counters.  Locking either mutex or lock
 423          * is sufficient to ensure the list doesn't change; to change
 424          * the list you need to lock both the mutex and the spinlock.
 425          */
 426         struct mutex            mutex;
 427
 428         struct list_head        counter_list;
 429         struct list_head        event_list;
 430         int                     nr_counters;
 431         int                     nr_active;
 432         int                     is_active;
 433         struct task_struct      *task;
 434
 435         /*
 436          * time_now is the current time in nanoseconds since an arbitrary
 437          * point in the past.  For per-task counters, this is based on the
 438          * task clock, and for per-cpu counters it is based on the cpu clock.
 439          * time_lost is an offset from the task/cpu clock, used to make it
 440          * appear that time only passes while the context is scheduled in.
 441          */
 442         u64                     time_now;
 443         u64                     time_lost;
 444 #endif
 445 };
 446
 447 /**
 448  * struct perf_counter_cpu_context - per cpu counter context structure
 449  */
 450 struct perf_cpu_context {
 451         struct perf_counter_context     ctx;
 452         struct perf_counter_context     *task_ctx;
 453         int                             active_oncpu;
 454         int                             max_pertask;
 455         int                             exclusive;
 456
 457         /*
 458          * Recursion avoidance:
 459          *
 460          * task, softirq, irq, nmi context
 461          */
 462         int                     recursion[4];
 463 };
 464
 465 /*
 466  * Set by architecture code:
 467  */
 468 extern int perf_max_counters;
 469
 470 #ifdef CONFIG_PERF_COUNTERS
 471 extern const struct hw_perf_counter_ops *
 472 hw_perf_counter_init(struct perf_counter *counter);
 473
 474 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 475 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 476 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 477 extern void perf_counter_init_task(struct task_struct *child);
 478 extern void perf_counter_exit_task(struct task_struct *child);
 479 extern void perf_counter_do_pending(void);
 480 extern void perf_counter_print_debug(void);
 481 extern void perf_counter_unthrottle(void);
 482 extern u64 hw_perf_save_disable(void);
 483 extern void hw_perf_restore(u64 ctrl);
 484 extern int perf_counter_task_disable(void);
 485 extern int perf_counter_task_enable(void);
 486 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 487                struct perf_cpu_context *cpuctx,
 488                struct perf_counter_context *ctx, int cpu);
 489 extern void perf_counter_update_userpage(struct perf_counter *counter);
 490
 491 extern void perf_counter_output(struct perf_counter *counter,
 492                                 int nmi, struct pt_regs *regs);
 493 /*
 494  * Return 1 for a software counter, 0 for a hardware counter
 495  */
 496 static inline int is_software_counter(struct perf_counter *counter)
 497 {
 498         return !perf_event_raw(&counter->hw_event) &&
 499                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 500 }
 501
 502 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 503
 504 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 505                               unsigned long pgoff, struct file *file);
 506
 507 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 508                                 unsigned long pgoff, struct file *file);
 509
 510 #define MAX_STACK_DEPTH         254
 511
 512 struct perf_callchain_entry {
 513         u32     nr, hv, kernel, user;
 514         u64     ip[MAX_STACK_DEPTH];
 515 };
 516
 517 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 518
 519 #else
 520 static inline void
 521 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 522 static inline void
 523 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 524 static inline void
 525 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 526 static inline void perf_counter_init_task(struct task_struct *child)    { }
 527 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 528 static inline void perf_counter_do_pending(void)                        { }
 529 static inline void perf_counter_print_debug(void)                       { }
 530 static inline void perf_counter_unthrottle(void)                        { }
 531 static inline void hw_perf_restore(u64 ctrl)                            { }
 532 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 533 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 534 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 535
 536 static inline void
 537 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)  { }
 538
 539
 540 static inline void
 541 perf_counter_mmap(unsigned long addr, unsigned long len,
 542                   unsigned long pgoff, struct file *file)               { }
 543
 544 static inline void
 545 perf_counter_munmap(unsigned long addr, unsigned long len,
 546                     unsigned long pgoff, struct file *file)             { }
 547
 548 #endif
 549
 550 #endif /* __KERNEL__ */
 551 #endif /* _LINUX_PERF_COUNTER_H */