git.oblomov.eu Git - linux-2.6/blob - 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
 144                                 __reserved_1   : 52;
 145
 146         __u32                   extra_config_len;
 147         __u32                   __reserved_4;
 148
 149         __u64                   __reserved_2;
 150         __u64                   __reserved_3;
 151 };
 152
 153 /*
 154  * Ioctls that can be done on a perf counter fd:
 155  */
 156 #define PERF_COUNTER_IOC_ENABLE         _IO('$', 0)
 157 #define PERF_COUNTER_IOC_DISABLE        _IO('$', 1)
 158
 159 /*
 160  * Structure of the page that can be mapped via mmap
 161  */
 162 struct perf_counter_mmap_page {
 163         __u32   version;                /* version number of this structure */
 164         __u32   compat_version;         /* lowest version this is compat with */
 165
 166         /*
 167          * Bits needed to read the hw counters in user-space.
 168          *
 169          * The index and offset should be read atomically using the seqlock:
 170          *
 171          *   __u32 seq, index;
 172          *   __s64 offset;
 173          *
 174          * again:
 175          *   rmb();
 176          *   seq = pc->lock;
 177          *
 178          *   if (unlikely(seq & 1)) {
 179          *     cpu_relax();
 180          *     goto again;
 181          *   }
 182          *
 183          *   index = pc->index;
 184          *   offset = pc->offset;
 185          *
 186          *   rmb();
 187          *   if (pc->lock != seq)
 188          *     goto again;
 189          *
 190          * After this, index contains architecture specific counter index + 1,
 191          * so that 0 means unavailable, offset contains the value to be added
 192          * to the result of the raw timer read to obtain this counter's value.
 193          */
 194         __u32   lock;                   /* seqlock for synchronization */
 195         __u32   index;                  /* hardware counter identifier */
 196         __s64   offset;                 /* add to hardware counter value */
 197
 198         /*
 199          * Control data for the mmap() data buffer.
 200          *
 201          * User-space reading this value should issue an rmb(), on SMP capable
 202          * platforms, after reading this value -- see perf_counter_wakeup().
 203          */
 204         __u32   data_head;              /* head in the data section */
 205 };
 206
 207 struct perf_event_header {
 208         __u32   type;
 209         __u32   size;
 210 };
 211
 212 enum perf_event_type {
 213         PERF_EVENT_IP           = 0,
 214         PERF_EVENT_GROUP        = 1,
 215
 216         PERF_EVENT_MMAP         = 2,
 217         PERF_EVENT_MUNMAP       = 3,
 218
 219         __PERF_EVENT_TID        = 0x100,
 220 };
 221
 222 #ifdef __KERNEL__
 223 /*
 224  * Kernel-internal data types and definitions:
 225  */
 226
 227 #ifdef CONFIG_PERF_COUNTERS
 228 # include <asm/perf_counter.h>
 229 #endif
 230
 231 #include <linux/list.h>
 232 #include <linux/mutex.h>
 233 #include <linux/rculist.h>
 234 #include <linux/rcupdate.h>
 235 #include <linux/spinlock.h>
 236 #include <linux/hrtimer.h>
 237 #include <asm/atomic.h>
 238
 239 struct task_struct;
 240
 241 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 242 {
 243         return hw_event->config & PERF_COUNTER_RAW_MASK;
 244 }
 245
 246 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 247 {
 248         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 249 }
 250
 251 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 252 {
 253         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 254                 PERF_COUNTER_TYPE_SHIFT;
 255 }
 256
 257 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 258 {
 259         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 260 }
 261
 262 /**
 263  * struct hw_perf_counter - performance counter hardware details:
 264  */
 265 struct hw_perf_counter {
 266 #ifdef CONFIG_PERF_COUNTERS
 267         union {
 268                 struct { /* hardware */
 269                         u64                             config;
 270                         unsigned long                   config_base;
 271                         unsigned long                   counter_base;
 272                         int                             nmi;
 273                         unsigned int                    idx;
 274                 };
 275                 union { /* software */
 276                         atomic64_t                      count;
 277                         struct hrtimer                  hrtimer;
 278                 };
 279         };
 280         atomic64_t                      prev_count;
 281         u64                             irq_period;
 282         atomic64_t                      period_left;
 283 #endif
 284 };
 285
 286 struct perf_counter;
 287
 288 /**
 289  * struct hw_perf_counter_ops - performance counter hw ops
 290  */
 291 struct hw_perf_counter_ops {
 292         int (*enable)                   (struct perf_counter *counter);
 293         void (*disable)                 (struct perf_counter *counter);
 294         void (*read)                    (struct perf_counter *counter);
 295 };
 296
 297 /**
 298  * enum perf_counter_active_state - the states of a counter
 299  */
 300 enum perf_counter_active_state {
 301         PERF_COUNTER_STATE_ERROR        = -2,
 302         PERF_COUNTER_STATE_OFF          = -1,
 303         PERF_COUNTER_STATE_INACTIVE     =  0,
 304         PERF_COUNTER_STATE_ACTIVE       =  1,
 305 };
 306
 307 struct file;
 308
 309 struct perf_mmap_data {
 310         struct rcu_head                 rcu_head;
 311         int                             nr_pages;
 312         atomic_t                        wakeup;
 313         atomic_t                        head;
 314         struct perf_counter_mmap_page   *user_page;
 315         void                            *data_pages[0];
 316 };
 317
 318 struct perf_wakeup_entry {
 319         struct perf_wakeup_entry *next;
 320 };
 321
 322 /**
 323  * struct perf_counter - performance counter kernel representation:
 324  */
 325 struct perf_counter {
 326 #ifdef CONFIG_PERF_COUNTERS
 327         struct list_head                list_entry;
 328         struct list_head                event_entry;
 329         struct list_head                sibling_list;
 330         int                             nr_siblings;
 331         struct perf_counter             *group_leader;
 332         const struct hw_perf_counter_ops *hw_ops;
 333
 334         enum perf_counter_active_state  state;
 335         enum perf_counter_active_state  prev_state;
 336         atomic64_t                      count;
 337
 338         /*
 339          * These are the total time in nanoseconds that the counter
 340          * has been enabled (i.e. eligible to run, and the task has
 341          * been scheduled in, if this is a per-task counter)
 342          * and running (scheduled onto the CPU), respectively.
 343          *
 344          * They are computed from tstamp_enabled, tstamp_running and
 345          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 346          */
 347         u64                             total_time_enabled;
 348         u64                             total_time_running;
 349
 350         /*
 351          * These are timestamps used for computing total_time_enabled
 352          * and total_time_running when the counter is in INACTIVE or
 353          * ACTIVE state, measured in nanoseconds from an arbitrary point
 354          * in time.
 355          * tstamp_enabled: the notional time when the counter was enabled
 356          * tstamp_running: the notional time when the counter was scheduled on
 357          * tstamp_stopped: in INACTIVE state, the notional time when the
 358          *      counter was scheduled off.
 359          */
 360         u64                             tstamp_enabled;
 361         u64                             tstamp_running;
 362         u64                             tstamp_stopped;
 363
 364         struct perf_counter_hw_event    hw_event;
 365         struct hw_perf_counter          hw;
 366
 367         struct perf_counter_context     *ctx;
 368         struct task_struct              *task;
 369         struct file                     *filp;
 370
 371         struct perf_counter             *parent;
 372         struct list_head                child_list;
 373
 374         /*
 375          * These accumulate total time (in nanoseconds) that children
 376          * counters have been enabled and running, respectively.
 377          */
 378         atomic64_t                      child_total_time_enabled;
 379         atomic64_t                      child_total_time_running;
 380
 381         /*
 382          * Protect attach/detach and child_list:
 383          */
 384         struct mutex                    mutex;
 385
 386         int                             oncpu;
 387         int                             cpu;
 388
 389         /* mmap bits */
 390         struct mutex                    mmap_mutex;
 391         atomic_t                        mmap_count;
 392         struct perf_mmap_data           *data;
 393
 394         /* poll related */
 395         wait_queue_head_t               waitq;
 396         /* optional: for NMIs */
 397         struct perf_wakeup_entry        wakeup;
 398
 399         void (*destroy)(struct perf_counter *);
 400         struct rcu_head                 rcu_head;
 401 #endif
 402 };
 403
 404 /**
 405  * struct perf_counter_context - counter context structure
 406  *
 407  * Used as a container for task counters and CPU counters as well:
 408  */
 409 struct perf_counter_context {
 410 #ifdef CONFIG_PERF_COUNTERS
 411         /*
 412          * Protect the states of the counters in the list,
 413          * nr_active, and the list:
 414          */
 415         spinlock_t              lock;
 416         /*
 417          * Protect the list of counters.  Locking either mutex or lock
 418          * is sufficient to ensure the list doesn't change; to change
 419          * the list you need to lock both the mutex and the spinlock.
 420          */
 421         struct mutex            mutex;
 422
 423         struct list_head        counter_list;
 424         struct list_head        event_list;
 425         int                     nr_counters;
 426         int                     nr_active;
 427         int                     is_active;
 428         struct task_struct      *task;
 429
 430         /*
 431          * time_now is the current time in nanoseconds since an arbitrary
 432          * point in the past.  For per-task counters, this is based on the
 433          * task clock, and for per-cpu counters it is based on the cpu clock.
 434          * time_lost is an offset from the task/cpu clock, used to make it
 435          * appear that time only passes while the context is scheduled in.
 436          */
 437         u64                     time_now;
 438         u64                     time_lost;
 439 #endif
 440 };
 441
 442 /**
 443  * struct perf_counter_cpu_context - per cpu counter context structure
 444  */
 445 struct perf_cpu_context {
 446         struct perf_counter_context     ctx;
 447         struct perf_counter_context     *task_ctx;
 448         int                             active_oncpu;
 449         int                             max_pertask;
 450         int                             exclusive;
 451
 452         /*
 453          * Recursion avoidance:
 454          *
 455          * task, softirq, irq, nmi context
 456          */
 457         int                     recursion[4];
 458 };
 459
 460 /*
 461  * Set by architecture code:
 462  */
 463 extern int perf_max_counters;
 464
 465 #ifdef CONFIG_PERF_COUNTERS
 466 extern const struct hw_perf_counter_ops *
 467 hw_perf_counter_init(struct perf_counter *counter);
 468
 469 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 470 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 471 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 472 extern void perf_counter_init_task(struct task_struct *child);
 473 extern void perf_counter_exit_task(struct task_struct *child);
 474 extern void perf_counter_do_pending(void);
 475 extern void perf_counter_print_debug(void);
 476 extern void perf_counter_unthrottle(void);
 477 extern u64 hw_perf_save_disable(void);
 478 extern void hw_perf_restore(u64 ctrl);
 479 extern int perf_counter_task_disable(void);
 480 extern int perf_counter_task_enable(void);
 481 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 482                struct perf_cpu_context *cpuctx,
 483                struct perf_counter_context *ctx, int cpu);
 484 extern void perf_counter_update_userpage(struct perf_counter *counter);
 485
 486 extern void perf_counter_output(struct perf_counter *counter,
 487                                 int nmi, struct pt_regs *regs);
 488 /*
 489  * Return 1 for a software counter, 0 for a hardware counter
 490  */
 491 static inline int is_software_counter(struct perf_counter *counter)
 492 {
 493         return !perf_event_raw(&counter->hw_event) &&
 494                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 495 }
 496
 497 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *);
 498
 499 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 500                               unsigned long pgoff, struct file *file);
 501
 502 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 503                                 unsigned long pgoff, struct file *file);
 504
 505 #else
 506 static inline void
 507 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 508 static inline void
 509 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 510 static inline void
 511 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 512 static inline void perf_counter_init_task(struct task_struct *child)    { }
 513 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 514 static inline void perf_counter_do_pending(void)                        { }
 515 static inline void perf_counter_print_debug(void)                       { }
 516 static inline void perf_counter_unthrottle(void)                        { }
 517 static inline void hw_perf_restore(u64 ctrl)                            { }
 518 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 519 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 520 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 521
 522 static inline void
 523 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)  { }
 524
 525
 526 static inline void
 527 perf_counter_mmap(unsigned long addr, unsigned long len,
 528                   unsigned long pgoff, struct file *file)               { }
 529
 530 static inline void
 531 perf_counter_munmap(unsigned long addr, unsigned long len,
 532                     unsigned long pgoff, struct file *file)             { }
 533
 534 #endif
 535
 536 #endif /* __KERNEL__ */
 537 #endif /* _LINUX_PERF_COUNTER_H */