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