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 /*
  77  * IRQ-notification data record type:
  78  */
  79 enum perf_counter_record_type {
  80         PERF_RECORD_SIMPLE              = 0,
  81         PERF_RECORD_IRQ                 = 1,
  82         PERF_RECORD_GROUP               = 2,
  83 };
  84
  85 #define __PERF_COUNTER_MASK(name)                       \
  86         (((1ULL << PERF_COUNTER_##name##_BITS) - 1) <<  \
  87          PERF_COUNTER_##name##_SHIFT)
  88
  89 #define PERF_COUNTER_RAW_BITS           1
  90 #define PERF_COUNTER_RAW_SHIFT          63
  91 #define PERF_COUNTER_RAW_MASK           __PERF_COUNTER_MASK(RAW)
  92
  93 #define PERF_COUNTER_CONFIG_BITS        63
  94 #define PERF_COUNTER_CONFIG_SHIFT       0
  95 #define PERF_COUNTER_CONFIG_MASK        __PERF_COUNTER_MASK(CONFIG)
  96
  97 #define PERF_COUNTER_TYPE_BITS          7
  98 #define PERF_COUNTER_TYPE_SHIFT         56
  99 #define PERF_COUNTER_TYPE_MASK          __PERF_COUNTER_MASK(TYPE)
 100
 101 #define PERF_COUNTER_EVENT_BITS         56
 102 #define PERF_COUNTER_EVENT_SHIFT        0
 103 #define PERF_COUNTER_EVENT_MASK         __PERF_COUNTER_MASK(EVENT)
 104
 105 /*
 106  * Bits that can be set in hw_event.read_format to request that
 107  * reads on the counter should return the indicated quantities,
 108  * in increasing order of bit value, after the counter value.
 109  */
 110 enum perf_counter_read_format {
 111         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
 112         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
 113 };
 114
 115 /*
 116  * Hardware event to monitor via a performance monitoring counter:
 117  */
 118 struct perf_counter_hw_event {
 119         /*
 120          * The MSB of the config word signifies if the rest contains cpu
 121          * specific (raw) counter configuration data, if unset, the next
 122          * 7 bits are an event type and the rest of the bits are the event
 123          * identifier.
 124          */
 125         __u64                   config;
 126
 127         __u64                   irq_period;
 128         __u64                   record_type;
 129         __u64                   read_format;
 130
 131         __u64                   disabled       :  1, /* off by default        */
 132                                 nmi            :  1, /* NMI sampling          */
 133                                 inherit        :  1, /* children inherit it   */
 134                                 pinned         :  1, /* must always be on PMU */
 135                                 exclusive      :  1, /* only group on PMU     */
 136                                 exclude_user   :  1, /* don't count user      */
 137                                 exclude_kernel :  1, /* ditto kernel          */
 138                                 exclude_hv     :  1, /* ditto hypervisor      */
 139                                 exclude_idle   :  1, /* don't count when idle */
 140                                 include_tid    :  1, /* include the tid       */
 141                                 mmap           :  1, /* include mmap data     */
 142                                 munmap         :  1, /* include munmap data   */
 143                                 callchain      :  1, /* add callchain data    */
 144
 145                                 __reserved_1   : 51;
 146
 147         __u32                   extra_config_len;
 148         __u32                   __reserved_4;
 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_GROUP        = 1,
 216
 217         PERF_EVENT_MMAP         = 2,
 218         PERF_EVENT_MUNMAP       = 3,
 219
 220         PERF_EVENT_OVERFLOW     = 1UL << 31,
 221         __PERF_EVENT_IP         = 1UL << 30,
 222         __PERF_EVENT_TID        = 1UL << 29,
 223         __PERF_EVENT_CALLCHAIN  = 1UL << 28,
 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         struct perf_counter_mmap_page   *user_page;
 319         void                            *data_pages[0];
 320 };
 321
 322 struct perf_wakeup_entry {
 323         struct perf_wakeup_entry *next;
 324 };
 325
 326 /**
 327  * struct perf_counter - performance counter kernel representation:
 328  */
 329 struct perf_counter {
 330 #ifdef CONFIG_PERF_COUNTERS
 331         struct list_head                list_entry;
 332         struct list_head                event_entry;
 333         struct list_head                sibling_list;
 334         int                             nr_siblings;
 335         struct perf_counter             *group_leader;
 336         const struct hw_perf_counter_ops *hw_ops;
 337
 338         enum perf_counter_active_state  state;
 339         enum perf_counter_active_state  prev_state;
 340         atomic64_t                      count;
 341
 342         /*
 343          * These are the total time in nanoseconds that the counter
 344          * has been enabled (i.e. eligible to run, and the task has
 345          * been scheduled in, if this is a per-task counter)
 346          * and running (scheduled onto the CPU), respectively.
 347          *
 348          * They are computed from tstamp_enabled, tstamp_running and
 349          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 350          */
 351         u64                             total_time_enabled;
 352         u64                             total_time_running;
 353
 354         /*
 355          * These are timestamps used for computing total_time_enabled
 356          * and total_time_running when the counter is in INACTIVE or
 357          * ACTIVE state, measured in nanoseconds from an arbitrary point
 358          * in time.
 359          * tstamp_enabled: the notional time when the counter was enabled
 360          * tstamp_running: the notional time when the counter was scheduled on
 361          * tstamp_stopped: in INACTIVE state, the notional time when the
 362          *      counter was scheduled off.
 363          */
 364         u64                             tstamp_enabled;
 365         u64                             tstamp_running;
 366         u64                             tstamp_stopped;
 367
 368         struct perf_counter_hw_event    hw_event;
 369         struct hw_perf_counter          hw;
 370
 371         struct perf_counter_context     *ctx;
 372         struct task_struct              *task;
 373         struct file                     *filp;
 374
 375         struct perf_counter             *parent;
 376         struct list_head                child_list;
 377
 378         /*
 379          * These accumulate total time (in nanoseconds) that children
 380          * counters have been enabled and running, respectively.
 381          */
 382         atomic64_t                      child_total_time_enabled;
 383         atomic64_t                      child_total_time_running;
 384
 385         /*
 386          * Protect attach/detach and child_list:
 387          */
 388         struct mutex                    mutex;
 389
 390         int                             oncpu;
 391         int                             cpu;
 392
 393         /* mmap bits */
 394         struct mutex                    mmap_mutex;
 395         atomic_t                        mmap_count;
 396         struct perf_mmap_data           *data;
 397
 398         /* poll related */
 399         wait_queue_head_t               waitq;
 400         /* optional: for NMIs */
 401         struct perf_wakeup_entry        wakeup;
 402
 403         void (*destroy)(struct perf_counter *);
 404         struct rcu_head                 rcu_head;
 405 #endif
 406 };
 407
 408 /**
 409  * struct perf_counter_context - counter context structure
 410  *
 411  * Used as a container for task counters and CPU counters as well:
 412  */
 413 struct perf_counter_context {
 414 #ifdef CONFIG_PERF_COUNTERS
 415         /*
 416          * Protect the states of the counters in the list,
 417          * nr_active, and the list:
 418          */
 419         spinlock_t              lock;
 420         /*
 421          * Protect the list of counters.  Locking either mutex or lock
 422          * is sufficient to ensure the list doesn't change; to change
 423          * the list you need to lock both the mutex and the spinlock.
 424          */
 425         struct mutex            mutex;
 426
 427         struct list_head        counter_list;
 428         struct list_head        event_list;
 429         int                     nr_counters;
 430         int                     nr_active;
 431         int                     is_active;
 432         struct task_struct      *task;
 433
 434         /*
 435          * time_now is the current time in nanoseconds since an arbitrary
 436          * point in the past.  For per-task counters, this is based on the
 437          * task clock, and for per-cpu counters it is based on the cpu clock.
 438          * time_lost is an offset from the task/cpu clock, used to make it
 439          * appear that time only passes while the context is scheduled in.
 440          */
 441         u64                     time_now;
 442         u64                     time_lost;
 443 #endif
 444 };
 445
 446 /**
 447  * struct perf_counter_cpu_context - per cpu counter context structure
 448  */
 449 struct perf_cpu_context {
 450         struct perf_counter_context     ctx;
 451         struct perf_counter_context     *task_ctx;
 452         int                             active_oncpu;
 453         int                             max_pertask;
 454         int                             exclusive;
 455
 456         /*
 457          * Recursion avoidance:
 458          *
 459          * task, softirq, irq, nmi context
 460          */
 461         int                     recursion[4];
 462 };
 463
 464 /*
 465  * Set by architecture code:
 466  */
 467 extern int perf_max_counters;
 468
 469 #ifdef CONFIG_PERF_COUNTERS
 470 extern const struct hw_perf_counter_ops *
 471 hw_perf_counter_init(struct perf_counter *counter);
 472
 473 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 474 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 475 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 476 extern void perf_counter_init_task(struct task_struct *child);
 477 extern void perf_counter_exit_task(struct task_struct *child);
 478 extern void perf_counter_do_pending(void);
 479 extern void perf_counter_print_debug(void);
 480 extern void perf_counter_unthrottle(void);
 481 extern u64 hw_perf_save_disable(void);
 482 extern void hw_perf_restore(u64 ctrl);
 483 extern int perf_counter_task_disable(void);
 484 extern int perf_counter_task_enable(void);
 485 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 486                struct perf_cpu_context *cpuctx,
 487                struct perf_counter_context *ctx, int cpu);
 488 extern void perf_counter_update_userpage(struct perf_counter *counter);
 489
 490 extern void perf_counter_output(struct perf_counter *counter,
 491                                 int nmi, struct pt_regs *regs);
 492 /*
 493  * Return 1 for a software counter, 0 for a hardware counter
 494  */
 495 static inline int is_software_counter(struct perf_counter *counter)
 496 {
 497         return !perf_event_raw(&counter->hw_event) &&
 498                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 499 }
 500
 501 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 502
 503 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 504                               unsigned long pgoff, struct file *file);
 505
 506 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 507                                 unsigned long pgoff, struct file *file);
 508
 509 #define MAX_STACK_DEPTH         255
 510
 511 struct perf_callchain_entry {
 512         u64     nr;
 513         u64     ip[MAX_STACK_DEPTH];
 514 };
 515
 516 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 517
 518 #else
 519 static inline void
 520 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 521 static inline void
 522 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 523 static inline void
 524 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 525 static inline void perf_counter_init_task(struct task_struct *child)    { }
 526 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 527 static inline void perf_counter_do_pending(void)                        { }
 528 static inline void perf_counter_print_debug(void)                       { }
 529 static inline void perf_counter_unthrottle(void)                        { }
 530 static inline void hw_perf_restore(u64 ctrl)                            { }
 531 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 532 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 533 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 534
 535 static inline void
 536 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)  { }
 537
 538
 539 static inline void
 540 perf_counter_mmap(unsigned long addr, unsigned long len,
 541                   unsigned long pgoff, struct file *file)               { }
 542
 543 static inline void
 544 perf_counter_munmap(unsigned long addr, unsigned long len,
 545                     unsigned long pgoff, struct file *file)             { }
 546
 547 #endif
 548
 549 #endif /* __KERNEL__ */
 550 #endif /* _LINUX_PERF_COUNTER_H */