SAFE public projects git trees. - safe/jmp/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 #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_TIME        = 1U << 2,
 104         PERF_RECORD_ADDR        = 1U << 3,
 105         PERF_RECORD_GROUP       = 1U << 4,
 106         PERF_RECORD_CALLCHAIN   = 1U << 5,
 107         PERF_RECORD_CONFIG      = 1U << 6,
 108 };
 109
 110 /*
 111  * Bits that can be set in hw_event.read_format to request that
 112  * reads on the counter should return the indicated quantities,
 113  * in increasing order of bit value, after the counter value.
 114  */
 115 enum perf_counter_read_format {
 116         PERF_FORMAT_TOTAL_TIME_ENABLED  =  1,
 117         PERF_FORMAT_TOTAL_TIME_RUNNING  =  2,
 118 };
 119
 120 /*
 121  * Hardware event to monitor via a performance monitoring counter:
 122  */
 123 struct perf_counter_hw_event {
 124         /*
 125          * The MSB of the config word signifies if the rest contains cpu
 126          * specific (raw) counter configuration data, if unset, the next
 127          * 7 bits are an event type and the rest of the bits are the event
 128          * identifier.
 129          */
 130         __u64                   config;
 131
 132         __u64                   irq_period;
 133         __u32                   record_type;
 134         __u32                   read_format;
 135
 136         __u64                   disabled       :  1, /* off by default        */
 137                                 nmi            :  1, /* NMI sampling          */
 138                                 inherit        :  1, /* children inherit it   */
 139                                 pinned         :  1, /* must always be on PMU */
 140                                 exclusive      :  1, /* only group on PMU     */
 141                                 exclude_user   :  1, /* don't count user      */
 142                                 exclude_kernel :  1, /* ditto kernel          */
 143                                 exclude_hv     :  1, /* ditto hypervisor      */
 144                                 exclude_idle   :  1, /* don't count when idle */
 145                                 mmap           :  1, /* include mmap data     */
 146                                 munmap         :  1, /* include munmap data   */
 147                                 comm           :  1, /* include comm data     */
 148
 149                                 __reserved_1   : 52;
 150
 151         __u32                   extra_config_len;
 152         __u32                   wakeup_events;  /* wakeup every n events */
 153
 154         __u64                   __reserved_2;
 155         __u64                   __reserved_3;
 156 };
 157
 158 /*
 159  * Ioctls that can be done on a perf counter fd:
 160  */
 161 #define PERF_COUNTER_IOC_ENABLE         _IOW('$', 0, u32)
 162 #define PERF_COUNTER_IOC_DISABLE        _IOW('$', 1, u32)
 163 #define PERF_COUNTER_IOC_REFRESH        _IOW('$', 2, u32)
 164 #define PERF_COUNTER_IOC_RESET          _IOW('$', 3, u32)
 165
 166 enum perf_counter_ioc_flags {
 167         PERF_IOC_FLAG_GROUP             = 1U << 0,
 168 };
 169
 170 /*
 171  * Structure of the page that can be mapped via mmap
 172  */
 173 struct perf_counter_mmap_page {
 174         __u32   version;                /* version number of this structure */
 175         __u32   compat_version;         /* lowest version this is compat with */
 176
 177         /*
 178          * Bits needed to read the hw counters in user-space.
 179          *
 180          *   u32 seq;
 181          *   s64 count;
 182          *
 183          *   do {
 184          *     seq = pc->lock;
 185          *
 186          *     barrier()
 187          *     if (pc->index) {
 188          *       count = pmc_read(pc->index - 1);
 189          *       count += pc->offset;
 190          *     } else
 191          *       goto regular_read;
 192          *
 193          *     barrier();
 194          *   } while (pc->lock != seq);
 195          *
 196          * NOTE: for obvious reason this only works on self-monitoring
 197          *       processes.
 198          */
 199         __u32   lock;                   /* seqlock for synchronization */
 200         __u32   index;                  /* hardware counter identifier */
 201         __s64   offset;                 /* add to hardware counter value */
 202
 203         /*
 204          * Control data for the mmap() data buffer.
 205          *
 206          * User-space reading this value should issue an rmb(), on SMP capable
 207          * platforms, after reading this value -- see perf_counter_wakeup().
 208          */
 209         __u32   data_head;              /* head in the data section */
 210 };
 211
 212 #define PERF_EVENT_MISC_KERNEL          (1 << 0)
 213 #define PERF_EVENT_MISC_USER            (1 << 1)
 214 #define PERF_EVENT_MISC_OVERFLOW        (1 << 2)
 215
 216 struct perf_event_header {
 217         __u32   type;
 218         __u16   misc;
 219         __u16   size;
 220 };
 221
 222 enum perf_event_type {
 223
 224         /*
 225          * The MMAP events record the PROT_EXEC mappings so that we can
 226          * correlate userspace IPs to code. They have the following structure:
 227          *
 228          * struct {
 229          *      struct perf_event_header        header;
 230          *
 231          *      u32                             pid, tid;
 232          *      u64                             addr;
 233          *      u64                             len;
 234          *      u64                             pgoff;
 235          *      char                            filename[];
 236          * };
 237          */
 238         PERF_EVENT_MMAP                 = 1,
 239         PERF_EVENT_MUNMAP               = 2,
 240
 241         /*
 242          * struct {
 243          *      struct perf_event_header        header;
 244          *
 245          *      u32                             pid, tid;
 246          *      char                            comm[];
 247          * };
 248          */
 249         PERF_EVENT_COMM                 = 3,
 250
 251         /*
 252          * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
 253          * will be PERF_RECORD_*
 254          *
 255          * struct {
 256          *      struct perf_event_header        header;
 257          *
 258          *      { u64                   ip;       } && PERF_RECORD_IP
 259          *      { u32                   pid, tid; } && PERF_RECORD_TID
 260          *      { u64                   time;     } && PERF_RECORD_TIME
 261          *      { u64                   addr;     } && PERF_RECORD_ADDR
 262          *      { u64                   config;   } && PERF_RECORD_CONFIG
 263          *
 264          *      { u64                   nr;
 265          *        { u64 event, val; }   cnt[nr];  } && PERF_RECORD_GROUP
 266          *
 267          *      { u16                   nr,
 268          *                              hv,
 269          *                              kernel,
 270          *                              user;
 271          *        u64                   ips[nr];  } && PERF_RECORD_CALLCHAIN
 272          * };
 273          */
 274 };
 275
 276 #ifdef __KERNEL__
 277 /*
 278  * Kernel-internal data types and definitions:
 279  */
 280
 281 #ifdef CONFIG_PERF_COUNTERS
 282 # include <asm/perf_counter.h>
 283 #endif
 284
 285 #include <linux/list.h>
 286 #include <linux/mutex.h>
 287 #include <linux/rculist.h>
 288 #include <linux/rcupdate.h>
 289 #include <linux/spinlock.h>
 290 #include <linux/hrtimer.h>
 291 #include <linux/fs.h>
 292 #include <asm/atomic.h>
 293
 294 struct task_struct;
 295
 296 static inline u64 perf_event_raw(struct perf_counter_hw_event *hw_event)
 297 {
 298         return hw_event->config & PERF_COUNTER_RAW_MASK;
 299 }
 300
 301 static inline u64 perf_event_config(struct perf_counter_hw_event *hw_event)
 302 {
 303         return hw_event->config & PERF_COUNTER_CONFIG_MASK;
 304 }
 305
 306 static inline u64 perf_event_type(struct perf_counter_hw_event *hw_event)
 307 {
 308         return (hw_event->config & PERF_COUNTER_TYPE_MASK) >>
 309                 PERF_COUNTER_TYPE_SHIFT;
 310 }
 311
 312 static inline u64 perf_event_id(struct perf_counter_hw_event *hw_event)
 313 {
 314         return hw_event->config & PERF_COUNTER_EVENT_MASK;
 315 }
 316
 317 /**
 318  * struct hw_perf_counter - performance counter hardware details:
 319  */
 320 struct hw_perf_counter {
 321 #ifdef CONFIG_PERF_COUNTERS
 322         union {
 323                 struct { /* hardware */
 324                         u64                             config;
 325                         unsigned long                   config_base;
 326                         unsigned long                   counter_base;
 327                         int                             nmi;
 328                         int                             idx;
 329                 };
 330                 union { /* software */
 331                         atomic64_t                      count;
 332                         struct hrtimer                  hrtimer;
 333                 };
 334         };
 335         atomic64_t                      prev_count;
 336         u64                             irq_period;
 337         atomic64_t                      period_left;
 338 #endif
 339 };
 340
 341 struct perf_counter;
 342
 343 /**
 344  * struct pmu - generic performance monitoring unit
 345  */
 346 struct pmu {
 347         int (*enable)                   (struct perf_counter *counter);
 348         void (*disable)                 (struct perf_counter *counter);
 349         void (*read)                    (struct perf_counter *counter);
 350 };
 351
 352 /**
 353  * enum perf_counter_active_state - the states of a counter
 354  */
 355 enum perf_counter_active_state {
 356         PERF_COUNTER_STATE_ERROR        = -2,
 357         PERF_COUNTER_STATE_OFF          = -1,
 358         PERF_COUNTER_STATE_INACTIVE     =  0,
 359         PERF_COUNTER_STATE_ACTIVE       =  1,
 360 };
 361
 362 struct file;
 363
 364 struct perf_mmap_data {
 365         struct rcu_head                 rcu_head;
 366         int                             nr_pages;       /* nr of data pages  */
 367         int                             nr_locked;      /* nr pages mlocked  */
 368
 369         atomic_t                        poll;           /* POLL_ for wakeups */
 370         atomic_t                        head;           /* write position    */
 371         atomic_t                        events;         /* event limit       */
 372
 373         atomic_t                        done_head;      /* completed head    */
 374         atomic_t                        lock;           /* concurrent writes */
 375
 376         atomic_t                        wakeup;         /* needs a wakeup    */
 377
 378         struct perf_counter_mmap_page   *user_page;
 379         void                            *data_pages[0];
 380 };
 381
 382 struct perf_pending_entry {
 383         struct perf_pending_entry *next;
 384         void (*func)(struct perf_pending_entry *);
 385 };
 386
 387 /**
 388  * struct perf_counter - performance counter kernel representation:
 389  */
 390 struct perf_counter {
 391 #ifdef CONFIG_PERF_COUNTERS
 392         struct list_head                list_entry;
 393         struct list_head                event_entry;
 394         struct list_head                sibling_list;
 395         int                             nr_siblings;
 396         struct perf_counter             *group_leader;
 397         const struct pmu                *pmu;
 398
 399         enum perf_counter_active_state  state;
 400         enum perf_counter_active_state  prev_state;
 401         atomic64_t                      count;
 402
 403         /*
 404          * These are the total time in nanoseconds that the counter
 405          * has been enabled (i.e. eligible to run, and the task has
 406          * been scheduled in, if this is a per-task counter)
 407          * and running (scheduled onto the CPU), respectively.
 408          *
 409          * They are computed from tstamp_enabled, tstamp_running and
 410          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 411          */
 412         u64                             total_time_enabled;
 413         u64                             total_time_running;
 414
 415         /*
 416          * These are timestamps used for computing total_time_enabled
 417          * and total_time_running when the counter is in INACTIVE or
 418          * ACTIVE state, measured in nanoseconds from an arbitrary point
 419          * in time.
 420          * tstamp_enabled: the notional time when the counter was enabled
 421          * tstamp_running: the notional time when the counter was scheduled on
 422          * tstamp_stopped: in INACTIVE state, the notional time when the
 423          *      counter was scheduled off.
 424          */
 425         u64                             tstamp_enabled;
 426         u64                             tstamp_running;
 427         u64                             tstamp_stopped;
 428
 429         struct perf_counter_hw_event    hw_event;
 430         struct hw_perf_counter          hw;
 431
 432         struct perf_counter_context     *ctx;
 433         struct task_struct              *task;
 434         struct file                     *filp;
 435
 436         struct perf_counter             *parent;
 437         struct list_head                child_list;
 438
 439         /*
 440          * These accumulate total time (in nanoseconds) that children
 441          * counters have been enabled and running, respectively.
 442          */
 443         atomic64_t                      child_total_time_enabled;
 444         atomic64_t                      child_total_time_running;
 445
 446         /*
 447          * Protect attach/detach and child_list:
 448          */
 449         struct mutex                    mutex;
 450
 451         int                             oncpu;
 452         int                             cpu;
 453
 454         /* mmap bits */
 455         struct mutex                    mmap_mutex;
 456         atomic_t                        mmap_count;
 457         struct perf_mmap_data           *data;
 458
 459         /* poll related */
 460         wait_queue_head_t               waitq;
 461         struct fasync_struct            *fasync;
 462
 463         /* delayed work for NMIs and such */
 464         int                             pending_wakeup;
 465         int                             pending_kill;
 466         int                             pending_disable;
 467         struct perf_pending_entry       pending;
 468
 469         atomic_t                        event_limit;
 470
 471         void (*destroy)(struct perf_counter *);
 472         struct rcu_head                 rcu_head;
 473 #endif
 474 };
 475
 476 /**
 477  * struct perf_counter_context - counter context structure
 478  *
 479  * Used as a container for task counters and CPU counters as well:
 480  */
 481 struct perf_counter_context {
 482 #ifdef CONFIG_PERF_COUNTERS
 483         /*
 484          * Protect the states of the counters in the list,
 485          * nr_active, and the list:
 486          */
 487         spinlock_t              lock;
 488         /*
 489          * Protect the list of counters.  Locking either mutex or lock
 490          * is sufficient to ensure the list doesn't change; to change
 491          * the list you need to lock both the mutex and the spinlock.
 492          */
 493         struct mutex            mutex;
 494
 495         struct list_head        counter_list;
 496         struct list_head        event_list;
 497         int                     nr_counters;
 498         int                     nr_active;
 499         int                     is_active;
 500         struct task_struct      *task;
 501
 502         /*
 503          * Context clock, runs when context enabled.
 504          */
 505         u64                     time;
 506         u64                     timestamp;
 507 #endif
 508 };
 509
 510 /**
 511  * struct perf_counter_cpu_context - per cpu counter context structure
 512  */
 513 struct perf_cpu_context {
 514         struct perf_counter_context     ctx;
 515         struct perf_counter_context     *task_ctx;
 516         int                             active_oncpu;
 517         int                             max_pertask;
 518         int                             exclusive;
 519
 520         /*
 521          * Recursion avoidance:
 522          *
 523          * task, softirq, irq, nmi context
 524          */
 525         int                     recursion[4];
 526 };
 527
 528 #ifdef CONFIG_PERF_COUNTERS
 529
 530 /*
 531  * Set by architecture code:
 532  */
 533 extern int perf_max_counters;
 534
 535 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
 536
 537 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 538 extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 539 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 540 extern void perf_counter_init_task(struct task_struct *child);
 541 extern void perf_counter_exit_task(struct task_struct *child);
 542 extern void perf_counter_do_pending(void);
 543 extern void perf_counter_print_debug(void);
 544 extern void perf_counter_unthrottle(void);
 545 extern u64 hw_perf_save_disable(void);
 546 extern void hw_perf_restore(u64 ctrl);
 547 extern int perf_counter_task_disable(void);
 548 extern int perf_counter_task_enable(void);
 549 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 550                struct perf_cpu_context *cpuctx,
 551                struct perf_counter_context *ctx, int cpu);
 552 extern void perf_counter_update_userpage(struct perf_counter *counter);
 553
 554 extern int perf_counter_overflow(struct perf_counter *counter,
 555                                  int nmi, struct pt_regs *regs, u64 addr);
 556 /*
 557  * Return 1 for a software counter, 0 for a hardware counter
 558  */
 559 static inline int is_software_counter(struct perf_counter *counter)
 560 {
 561         return !perf_event_raw(&counter->hw_event) &&
 562                 perf_event_type(&counter->hw_event) != PERF_TYPE_HARDWARE;
 563 }
 564
 565 extern void perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
 566
 567 extern void perf_counter_mmap(unsigned long addr, unsigned long len,
 568                               unsigned long pgoff, struct file *file);
 569
 570 extern void perf_counter_munmap(unsigned long addr, unsigned long len,
 571                                 unsigned long pgoff, struct file *file);
 572
 573 extern void perf_counter_comm(struct task_struct *tsk);
 574
 575 #define MAX_STACK_DEPTH         255
 576
 577 struct perf_callchain_entry {
 578         u16     nr, hv, kernel, user;
 579         u64     ip[MAX_STACK_DEPTH];
 580 };
 581
 582 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 583
 584 extern int sysctl_perf_counter_priv;
 585 extern int sysctl_perf_counter_mlock;
 586
 587 extern void perf_counter_init(void);
 588
 589 #else
 590 static inline void
 591 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 592 static inline void
 593 perf_counter_task_sched_out(struct task_struct *task, int cpu)          { }
 594 static inline void
 595 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 596 static inline void perf_counter_init_task(struct task_struct *child)    { }
 597 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 598 static inline void perf_counter_do_pending(void)                        { }
 599 static inline void perf_counter_print_debug(void)                       { }
 600 static inline void perf_counter_unthrottle(void)                        { }
 601 static inline void hw_perf_restore(u64 ctrl)                            { }
 602 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 603 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 604 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 605
 606 static inline void
 607 perf_swcounter_event(u32 event, u64 nr, int nmi,
 608                      struct pt_regs *regs, u64 addr)                    { }
 609
 610 static inline void
 611 perf_counter_mmap(unsigned long addr, unsigned long len,
 612                   unsigned long pgoff, struct file *file)               { }
 613
 614 static inline void
 615 perf_counter_munmap(unsigned long addr, unsigned long len,
 616                     unsigned long pgoff, struct file *file)             { }
 617
 618 static inline void perf_counter_comm(struct task_struct *tsk)           { }
 619 static inline void perf_counter_init(void)                              { }
 620 #endif
 621
 622 #endif /* __KERNEL__ */
 623 #endif /* _LINUX_PERF_COUNTER_H */